feat: BYOM

* feat: simplified model structure

* fix: test

* fix: mini docstring stuff

.env-template

@@ -35,8 +35,5 @@ MICROSOFT_TENANT_ID=your-azure-ad-tenant-id
 #Alternatively, use "https://login.microsoftonline.com/common" for multi-tenant app.
 MICROSOFT_AUTHORITY=https://{tenantId}.ciamlogin.com/{tenantId}
 
-# User-data Postgres DB (Phase 0 of the MongoDB→Postgres migration).
-# Standard Postgres URI — `postgres://` and `postgresql://` both work.
-# Leave unset while the migration is still being rolled out; the app will
-# fall back to MongoDB for user data until POSTGRES_URI is configured.
+
 # POSTGRES_URI=postgresql://docsgpt:docsgpt@localhost:5432/docsgpt

.github/workflows/vale.yml

@@ -11,7 +11,6 @@ on:
 
 permissions:
   contents: read
-  pull-requests: write
 
 jobs:
   vale:
@@ -20,11 +19,16 @@ jobs:
       - name: Checkout code
         uses: actions/checkout@v4
 
-      - name: Vale linter
-        uses: errata-ai/vale-action@v2
-        with:
-          files: docs
-          fail_on_error: false
-          version: 3.0.5
-        env:
-          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+      - name: Install Vale
+        run: |
+          curl -fsSL -o vale.tar.gz \
+            https://github.com/errata-ai/vale/releases/download/v3.0.5/vale_3.0.5_Linux_64-bit.tar.gz
+          tar -xzf vale.tar.gz
+          sudo mv vale /usr/local/bin/vale
+          vale --version
+
+      - name: Sync Vale packages
+        run: vale sync
+
+      - name: Run Vale
+        run: vale --minAlertLevel=error docs

.gitignore

@@ -186,3 +186,11 @@ node_modules/
 .vscode/sftp.json
 /models/
 model/
+
+# E2E test artifacts
+.e2e-tmp/
+/tmp/docsgpt-e2e/
+tests/e2e/node_modules/
+tests/e2e/playwright-report/
+tests/e2e/test-results/
+tests/e2e/.e2e-last-run.json

AGENTS.md

@@ -10,9 +10,15 @@
 For feature work, do **not** assume the environment needs to be recreated.
 
 - Check whether the user already has a Python virtual environment such as `venv/` or `.venv/`.
-- Check whether MongoDB is already running.
+- Check whether Postgres is already running and reachable via `POSTGRES_URI` (the canonical user-data store).
 - Check whether Redis is already running.
-- Reuse what is already working. Do not stop or recreate MongoDB, Redis, or the Python environment unless the task is environment setup or troubleshooting.
+- Reuse what is already working. Do not stop or recreate Postgres, Redis, or the Python environment unless the task is environment setup or troubleshooting.
+
+> MongoDB is **not** required for the default install. It is only needed if
+> the user opts into the Mongo vector-store backend (`VECTOR_STORE=mongodb`)
+> or is running the one-shot `scripts/db/backfill.py` to migrate existing
+> user data from the legacy Mongo-based install. In those cases, `pymongo`
+> is available as an optional extra, not a core dependency.
 
 ## Normal local development commands
 
@@ -31,6 +37,22 @@ Run the Flask API (if needed):
 flask --app application/app.py run --host=0.0.0.0 --port=7091
 ```
 
+That's the fast inner-loop option — quick startup, the Werkzeug interactive
+debugger still works, and it hot-reloads on source changes. It serves the
+Flask routes only (`/api/*`, `/stream`, etc.).
+
+If you need to exercise the full ASGI stack — the `/mcp` FastMCP endpoint,
+or to match the production runtime exactly — run the ASGI composition under
+uvicorn instead:
+
+```bash
+uvicorn application.asgi:asgi_app --host 0.0.0.0 --port 7091 --reload
+```
+
+Production uses `gunicorn -k uvicorn_worker.UvicornWorker` against the same
+`application.asgi:asgi_app` target; see `application/Dockerfile` for the
+full flag set.
+
 Run the Celery worker in a separate terminal (if needed):
 
 ```bash
@@ -93,7 +115,7 @@ vale .
 - `frontend/`: Vite + React + TypeScript application.
 - `frontend/src/`: main UI code, including `components`, `conversation`, `hooks`, `locale`, `settings`, `upload`, and Redux store wiring in `store.ts`.
 - `docs/`: separate documentation site built with Next.js/Nextra.
-- `extensions/`: integrations and widgets such as Chatwoot, Chrome, Discord, React widget, Slack bot, and web widget.
+- `extensions/`: integrations and widgets — currently the Chatwoot webhook bridge and the React widget (published to npm as `docsgpt`). The Discord bot, Slack bot, and Chrome extension have been moved to their own repos under `arc53/`.
 - `deployment/`: Docker Compose variants and Kubernetes manifests.
 
 ## Coding rules

application/Dockerfile

@@ -88,5 +88,15 @@ EXPOSE 7091
 # Switch to non-root user
 USER appuser
 
-# Start Gunicorn
-CMD ["gunicorn", "-w", "1", "--timeout", "120", "--bind", "0.0.0.0:7091", "--preload", "application.wsgi:app"]
+CMD ["gunicorn", \
+     "-w", "1", \
+     "-k", "uvicorn_worker.UvicornWorker", \
+     "--bind", "0.0.0.0:7091", \
+     "--timeout", "180", \
+     "--graceful-timeout", "120", \
+     "--keep-alive", "5", \
+     "--worker-tmp-dir", "/dev/shm", \
+     "--max-requests", "1000", \
+     "--max-requests-jitter", "100", \
+     "--config", "application/gunicorn_conf.py", \
+     "application.asgi:asgi_app"]

application/agents/base.py

@@ -42,6 +42,7 @@ class BaseAgent(ABC):
         llm_handler=None,
         tool_executor: Optional[ToolExecutor] = None,
         backup_models: Optional[List[str]] = None,
+        model_user_id: Optional[str] = None,
     ):
         self.endpoint = endpoint
         self.llm_name = llm_name
@@ -52,10 +53,13 @@ class BaseAgent(ABC):
         self.prompt = prompt
         self.decoded_token = decoded_token or {}
         self.user: str = self.decoded_token.get("sub")
+        # BYOM-resolution scope: owner for shared agents, caller for
+        # caller-owned BYOM, None for built-ins. Falls back to self.user
+        # for worker/legacy callers that don't thread model_user_id.
+        self.model_user_id = model_user_id
         self.tools: List[Dict] = []
         self.chat_history: List[Dict] = chat_history if chat_history is not None else []
 
-        # Dependency injection for LLM — fall back to creating if not provided
         if llm is not None:
             self.llm = llm
         else:
@@ -67,8 +71,16 @@ class BaseAgent(ABC):
                 model_id=model_id,
                 agent_id=agent_id,
                 backup_models=backup_models,
+                model_user_id=model_user_id,
             )
 
+        # For BYOM, registry id (UUID) differs from upstream model id
+        # (e.g. ``mistral-large-latest``). LLMCreator resolved this onto
+        # the LLM instance; cache it for subsequent gen calls.
+        self.upstream_model_id = (
+            getattr(self.llm, "model_id", None) or model_id
+        )
+
         self.retrieved_docs = retrieved_docs or []
 
         if llm_handler is not None:
@@ -306,7 +318,9 @@ class BaseAgent(ABC):
         try:
             current_tokens = self._calculate_current_context_tokens(messages)
             self.current_token_count = current_tokens
-            context_limit = get_token_limit(self.model_id)
+            context_limit = get_token_limit(
+                self.model_id, user_id=self.model_user_id or self.user
+            )
             threshold = int(context_limit * settings.COMPRESSION_THRESHOLD_PERCENTAGE)
 
             if current_tokens >= threshold:
@@ -325,7 +339,9 @@ class BaseAgent(ABC):
 
         current_tokens = self._calculate_current_context_tokens(messages)
         self.current_token_count = current_tokens
-        context_limit = get_token_limit(self.model_id)
+        context_limit = get_token_limit(
+            self.model_id, user_id=self.model_user_id or self.user
+        )
         percentage = (current_tokens / context_limit) * 100
 
         if current_tokens >= context_limit:
@@ -387,7 +403,9 @@ class BaseAgent(ABC):
             )
             system_prompt = system_prompt + compression_context
 
-        context_limit = get_token_limit(self.model_id)
+        context_limit = get_token_limit(
+            self.model_id, user_id=self.model_user_id or self.user
+        )
         system_tokens = num_tokens_from_string(system_prompt)
 
         safety_buffer = int(context_limit * 0.1)
@@ -497,7 +515,10 @@ class BaseAgent(ABC):
     def _llm_gen(self, messages: List[Dict], log_context: Optional[LogContext] = None):
         self._validate_context_size(messages)
 
-        gen_kwargs = {"model": self.model_id, "messages": messages}
+        # Use the upstream id resolved by LLMCreator (see __init__).
+        # Built-in models: same as self.model_id. BYOM: the user's
+        # typed model name, not the internal UUID.
+        gen_kwargs = {"model": self.upstream_model_id, "messages": messages}
         if self.attachments:
             gen_kwargs["_usage_attachments"] = self.attachments
 

application/agents/research_agent.py

@@ -312,7 +312,7 @@ class ResearchAgent(BaseAgent):
 
         try:
             response = self.llm.gen(
-                model=self.model_id,
+                model=self.upstream_model_id,
                 messages=messages,
                 tools=None,
                 response_format={"type": "json_object"},
@@ -390,7 +390,7 @@ class ResearchAgent(BaseAgent):
 
         try:
             response = self.llm.gen(
-                model=self.model_id,
+                model=self.upstream_model_id,
                 messages=messages,
                 tools=None,
                 response_format={"type": "json_object"},
@@ -506,7 +506,7 @@ class ResearchAgent(BaseAgent):
 
             try:
                 response = self.llm.gen(
-                    model=self.model_id,
+                    model=self.upstream_model_id,
                     messages=messages,
                     tools=self.tools if self.tools else None,
                 )
@@ -537,7 +537,7 @@ class ResearchAgent(BaseAgent):
         )
         try:
             response = self.llm.gen(
-                model=self.model_id, messages=messages, tools=None
+                model=self.upstream_model_id, messages=messages, tools=None
             )
             self._track_tokens(self._snapshot_llm_tokens())
             text = self._extract_text(response)
@@ -664,7 +664,7 @@ class ResearchAgent(BaseAgent):
         ]
 
         llm_response = self.llm.gen_stream(
-            model=self.model_id, messages=messages, tools=None
+            model=self.upstream_model_id, messages=messages, tools=None
         )
 
         if log_context:

application/agents/tool_executor.py

@@ -3,13 +3,12 @@ import uuid
 from collections import Counter
 from typing import Dict, List, Optional, Tuple
 
-from bson.objectid import ObjectId
-
 from application.agents.tools.tool_action_parser import ToolActionParser
 from application.agents.tools.tool_manager import ToolManager
-from application.core.mongo_db import MongoDB
-from application.core.settings import settings
 from application.security.encryption import decrypt_credentials
+from application.storage.db.repositories.agents import AgentsRepository
+from application.storage.db.repositories.user_tools import UserToolsRepository
+from application.storage.db.session import db_readonly
 
 logger = logging.getLogger(__name__)
 
@@ -51,30 +50,28 @@ class ToolExecutor:
         return tools
 
     def _get_tools_by_api_key(self, api_key: str) -> Dict[str, Dict]:
-        mongo = MongoDB.get_client()
-        db = mongo[settings.MONGO_DB_NAME]
-        agents_collection = db["agents"]
-        tools_collection = db["user_tools"]
-
-        agent_data = agents_collection.find_one({"key": api_key})
-        tool_ids = agent_data.get("tools", []) if agent_data else []
-
-        tools = (
-            tools_collection.find(
-                {"_id": {"$in": [ObjectId(tool_id) for tool_id in tool_ids]}}
-            )
-            if tool_ids
-            else []
-        )
-        tools = list(tools)
-        return {str(tool["_id"]): tool for tool in tools} if tools else {}
+        # Per-operation session: the answer pipeline spans a long-lived
+        # generator; wrapping it in a single connection would pin a PG
+        # conn for the whole stream. Open, fetch, close.
+        with db_readonly() as conn:
+            agent_data = AgentsRepository(conn).find_by_key(api_key)
+            tool_ids = agent_data.get("tools", []) if agent_data else []
+            if not tool_ids:
+                return {}
+            tools_repo = UserToolsRepository(conn)
+            tools: List[Dict] = []
+            owner = (agent_data.get("user_id") or agent_data.get("user")) if agent_data else None
+            for tid in tool_ids:
+                row = None
+                if owner:
+                    row = tools_repo.get_any(str(tid), owner)
+                if row is not None:
+                    tools.append(row)
+        return {str(tool["id"]): tool for tool in tools} if tools else {}
 
     def _get_user_tools(self, user: str = "local") -> Dict[str, Dict]:
-        mongo = MongoDB.get_client()
-        db = mongo[settings.MONGO_DB_NAME]
-        user_tools_collection = db["user_tools"]
-        user_tools = user_tools_collection.find({"user": user, "status": True})
-        user_tools = list(user_tools)
+        with db_readonly() as conn:
+            user_tools = UserToolsRepository(conn).list_active_for_user(user)
         return {str(i): tool for i, tool in enumerate(user_tools)}
 
     def merge_client_tools(
@@ -354,6 +351,17 @@ class ToolExecutor:
             headers=headers, query_params=query_params,
         )
 
+        if tool is None:
+            error_message = (
+                f"Failed to load tool '{tool_data.get('name')}' (tool_id key={tool_id}): "
+                "missing 'id' on tool row."
+            )
+            logger.error(error_message)
+            tool_call_data["result"] = error_message
+            yield {"type": "tool_call", "data": {**tool_call_data, "status": "error"}}
+            self.tool_calls.append(tool_call_data)
+            return error_message, call_id
+
         resolved_arguments = (
             {"query_params": query_params, "headers": headers, "body": body}
             if tool_data["name"] == "api_tool"
@@ -440,7 +448,16 @@ class ToolExecutor:
                 tool_config.update(decrypted)
                 tool_config["auth_credentials"] = decrypted
                 tool_config.pop("encrypted_credentials", None)
-            tool_config["tool_id"] = str(tool_data.get("_id", tool_id))
+            row_id = tool_data.get("id")
+            if not row_id:
+                logger.error(
+                    "Tool data missing 'id' for tool name=%s (enumerate-key tool_id=%s); "
+                    "skipping load to avoid binding a non-UUID downstream.",
+                    tool_data.get("name"),
+                    tool_id,
+                )
+                return None
+            tool_config["tool_id"] = str(row_id)
             if self.conversation_id:
                 tool_config["conversation_id"] = self.conversation_id
             if tool_data["name"] == "mcp_tool":

application/agents/tools/internal_search.py

@@ -39,6 +39,7 @@ class InternalSearchTool(Tool):
                 chunks=int(self.config.get("chunks", 2)),
                 doc_token_limit=int(self.config.get("doc_token_limit", 50000)),
                 model_id=self.config.get("model_id", "docsgpt-local"),
+                model_user_id=self.config.get("model_user_id"),
                 user_api_key=self.config.get("user_api_key"),
                 agent_id=self.config.get("agent_id"),
                 llm_name=self.config.get("llm_name", settings.LLM_PROVIDER),
@@ -48,7 +49,7 @@ class InternalSearchTool(Tool):
         return self._retriever
 
     def _get_directory_structure(self) -> Optional[Dict]:
-        """Load directory structure from MongoDB for the configured sources."""
+        """Load directory structure from Postgres for the configured sources."""
         if self._dir_structure_loaded:
             return self._directory_structure
 
@@ -59,35 +60,39 @@ class InternalSearchTool(Tool):
             return None
 
         try:
-            from bson.objectid import ObjectId
-            from application.core.mongo_db import MongoDB
-
-            mongo = MongoDB.get_client()
-            db = mongo[settings.MONGO_DB_NAME]
-            sources_collection = db["sources"]
+            # Per-operation session: this tool runs inside the answer
+            # generator hot path, so we open a short-lived read
+            # connection for the batch lookup and release immediately.
+            from application.storage.db.repositories.sources import (
+                SourcesRepository,
+            )
+            from application.storage.db.session import db_readonly
 
             if isinstance(active_docs, str):
                 active_docs = [active_docs]
 
+            decoded_token = self.config.get("decoded_token") or {}
+            user_id = decoded_token.get("sub") if decoded_token else None
+
             merged_structure = {}
-            for doc_id in active_docs:
-                try:
-                    source_doc = sources_collection.find_one(
-                        {"_id": ObjectId(doc_id)}
-                    )
-                    if not source_doc:
-                        continue
-                    dir_str = source_doc.get("directory_structure")
-                    if dir_str:
-                        if isinstance(dir_str, str):
-                            dir_str = json.loads(dir_str)
-                        source_name = source_doc.get("name", doc_id)
-                        if len(active_docs) > 1:
-                            merged_structure[source_name] = dir_str
-                        else:
-                            merged_structure = dir_str
-                except Exception as e:
-                    logger.debug(f"Could not load dir structure for {doc_id}: {e}")
+            with db_readonly() as conn:
+                repo = SourcesRepository(conn)
+                for doc_id in active_docs:
+                    try:
+                        source_doc = repo.get_any(str(doc_id), user_id) if user_id else None
+                        if not source_doc:
+                            continue
+                        dir_str = source_doc.get("directory_structure")
+                        if dir_str:
+                            if isinstance(dir_str, str):
+                                dir_str = json.loads(dir_str)
+                            source_name = source_doc.get("name", doc_id)
+                            if len(active_docs) > 1:
+                                merged_structure[source_name] = dir_str
+                            else:
+                                merged_structure = dir_str
+                    except Exception as e:
+                        logger.debug(f"Could not load dir structure for {doc_id}: {e}")
 
             self._directory_structure = merged_structure if merged_structure else None
         except Exception as e:
@@ -357,32 +362,48 @@ INTERNAL_TOOL_ENTRY = build_internal_tool_entry(has_directory_structure=False)
 
 
 def sources_have_directory_structure(source: Dict) -> bool:
-    """Check if any of the active sources have directory_structure in MongoDB."""
+    """Check if any of the active sources have a ``directory_structure`` row."""
     active_docs = source.get("active_docs", [])
     if not active_docs:
         return False
 
     try:
-        from bson.objectid import ObjectId
-        from application.core.mongo_db import MongoDB
+        # TODO(pg-cutover): SourcesRepository.get_any requires ``user_id``
+        # scoping, but callers in the agent build path don't always
+        # thread the decoded token through here. Use a direct
+        # short-lived SQL lookup instead of the repo until the call
+        # sites are updated to propagate user context.
+        from sqlalchemy import text as _text
 
-        mongo = MongoDB.get_client()
-        db = mongo[settings.MONGO_DB_NAME]
-        sources_collection = db["sources"]
+        from application.storage.db.session import db_readonly
 
         if isinstance(active_docs, str):
             active_docs = [active_docs]
 
-        for doc_id in active_docs:
-            try:
-                source_doc = sources_collection.find_one(
-                    {"_id": ObjectId(doc_id)},
-                    {"directory_structure": 1},
-                )
-                if source_doc and source_doc.get("directory_structure"):
-                    return True
-            except Exception:
-                continue
+        with db_readonly() as conn:
+            for doc_id in active_docs:
+                try:
+                    value = str(doc_id)
+                    if len(value) == 36 and "-" in value:
+                        row = conn.execute(
+                            _text(
+                                "SELECT directory_structure FROM sources "
+                                "WHERE id = CAST(:id AS uuid)"
+                            ),
+                            {"id": value},
+                        ).fetchone()
+                    else:
+                        row = conn.execute(
+                            _text(
+                                "SELECT directory_structure FROM sources "
+                                "WHERE legacy_mongo_id = :lid"
+                            ),
+                            {"lid": value},
+                        ).fetchone()
+                    if row is not None and row[0]:
+                        return True
+                except Exception:
+                    continue
     except Exception as e:
         logger.debug(f"Could not check directory structure: {e}")
 
@@ -415,6 +436,7 @@ def build_internal_tool_config(
     chunks: int = 2,
     doc_token_limit: int = 50000,
     model_id: str = "docsgpt-local",
+    model_user_id: Optional[str] = None,
     user_api_key: Optional[str] = None,
     agent_id: Optional[str] = None,
     llm_name: str = None,
@@ -429,6 +451,7 @@ def build_internal_tool_config(
         "chunks": chunks,
         "doc_token_limit": doc_token_limit,
         "model_id": model_id,
+        "model_user_id": model_user_id,
         "user_api_key": user_api_key,
         "agent_id": agent_id,
         "llm_name": llm_name or settings.LLM_PROVIDER,

application/agents/tools/mcp_tool.py

@@ -22,16 +22,12 @@ from redis import Redis
 from application.agents.tools.base import Tool
 from application.api.user.tasks import mcp_oauth_status_task, mcp_oauth_task
 from application.cache import get_redis_instance
-from application.core.mongo_db import MongoDB
 from application.core.settings import settings
 from application.core.url_validation import SSRFError, validate_url
 from application.security.encryption import decrypt_credentials
 
 logger = logging.getLogger(__name__)
 
-mongo = MongoDB.get_client()
-db = mongo[settings.MONGO_DB_NAME]
-
 _mcp_clients_cache = {}
 
 
@@ -161,7 +157,6 @@ class MCPTool(Tool):
                     scopes=self.oauth_scopes,
                     redis_client=redis_client,
                     redirect_uri=self.redirect_uri,
-                    db=db,
                     user_id=self.user_id,
                 )
             else:
@@ -171,7 +166,6 @@ class MCPTool(Tool):
                     redis_client=redis_client,
                     redirect_uri=self.redirect_uri,
                     task_id=self.oauth_task_id,
-                    db=db,
                     user_id=self.user_id,
                 )
         elif self.auth_type == "bearer":
@@ -491,7 +485,7 @@ class MCPTool(Tool):
 
     def _test_oauth_connection(self) -> Dict:
         storage = DBTokenStorage(
-            server_url=self.server_url, user_id=self.user_id, db_client=db
+            server_url=self.server_url, user_id=self.user_id,
         )
         loop = asyncio.new_event_loop()
         try:
@@ -683,7 +677,6 @@ class DocsGPTOAuth(OAuthClientProvider):
         scopes: str | list[str] | None = None,
         client_name: str = "DocsGPT-MCP",
         user_id=None,
-        db=None,
         additional_client_metadata: dict[str, Any] | None = None,
         skip_redirect_validation: bool = False,
     ):
@@ -692,7 +685,6 @@ class DocsGPTOAuth(OAuthClientProvider):
         self.redis_prefix = redis_prefix
         self.task_id = task_id
         self.user_id = user_id
-        self.db = db
 
         parsed_url = urlparse(mcp_url)
         self.server_base_url = f"{parsed_url.scheme}://{parsed_url.netloc}"
@@ -711,7 +703,6 @@ class DocsGPTOAuth(OAuthClientProvider):
         storage = DBTokenStorage(
             server_url=self.server_base_url,
             user_id=self.user_id,
-            db_client=self.db,
             expected_redirect_uri=None if skip_redirect_validation else redirect_uri,
         )
 
@@ -853,54 +844,95 @@ class DBTokenStorage(TokenStorage):
         self,
         server_url: str,
         user_id: str,
-        db_client,
         expected_redirect_uri: Optional[str] = None,
     ):
         self.server_url = server_url
         self.user_id = user_id
-        self.db_client = db_client
         self.expected_redirect_uri = expected_redirect_uri
-        self.collection = db_client["connector_sessions"]
 
     @staticmethod
     def get_base_url(url: str) -> str:
         parsed = urlparse(url)
         return f"{parsed.scheme}://{parsed.netloc}"
 
-    def get_db_key(self) -> dict:
-        return {
-            "server_url": self.get_base_url(self.server_url),
-            "user_id": self.user_id,
-        }
+    def _pg_provider(self) -> str:
+        return f"mcp:{self.get_base_url(self.server_url)}"
+
+    def _fetch_session_data(self) -> dict:
+        """Read the JSONB ``session_data`` blob for this MCP server row."""
+        from application.storage.db.repositories.connector_sessions import (
+            ConnectorSessionsRepository,
+        )
+        from application.storage.db.session import db_readonly
+
+        base_url = self.get_base_url(self.server_url)
+        with db_readonly() as conn:
+            row = ConnectorSessionsRepository(conn).get_by_user_and_server_url(
+                self.user_id, base_url,
+            )
+        if not row:
+            return {}
+        data = row.get("session_data") or {}
+        if isinstance(data, str):
+            try:
+                data = json.loads(data)
+            except ValueError:
+                return {}
+        return data if isinstance(data, dict) else {}
 
     async def get_tokens(self) -> OAuthToken | None:
-        doc = await asyncio.to_thread(self.collection.find_one, self.get_db_key())
-        if not doc or "tokens" not in doc:
+        data = await asyncio.to_thread(self._fetch_session_data)
+        if not data or "tokens" not in data:
             return None
         try:
-            return OAuthToken.model_validate(doc["tokens"])
+            return OAuthToken.model_validate(data["tokens"])
         except ValidationError as e:
             logger.error("Could not load tokens: %s", e)
             return None
 
-    async def set_tokens(self, tokens: OAuthToken) -> None:
-        await asyncio.to_thread(
-            self.collection.update_one,
-            self.get_db_key(),
-            {"$set": {"tokens": tokens.model_dump()}},
-            True,
+    def _merge(self, patch: dict) -> None:
+        """Shallow-merge ``patch`` into this row's ``session_data``.
+
+        Threads ``server_url`` through to the repository so it lands in
+        the scalar column — ``get_by_user_and_server_url`` needs that to
+        resolve the row (``NULL = 'https://...'`` is UNKNOWN in SQL).
+        """
+        from application.storage.db.repositories.connector_sessions import (
+            ConnectorSessionsRepository,
         )
-        logger.info("Saved tokens for %s", self.get_base_url(self.server_url))
+        from application.storage.db.session import db_session
+
+        base_url = self.get_base_url(self.server_url)
+        with db_session() as conn:
+            ConnectorSessionsRepository(conn).merge_session_data(
+                self.user_id, self._pg_provider(), base_url, patch,
+            )
+
+    def _delete(self) -> None:
+        from application.storage.db.repositories.connector_sessions import (
+            ConnectorSessionsRepository,
+        )
+        from application.storage.db.session import db_session
+
+        with db_session() as conn:
+            ConnectorSessionsRepository(conn).delete(
+                self.user_id, self._pg_provider(),
+            )
+
+    async def set_tokens(self, tokens: OAuthToken) -> None:
+        base_url = self.get_base_url(self.server_url)
+        token_dump = tokens.model_dump()
+        await asyncio.to_thread(self._merge, {"tokens": token_dump})
+        logger.info("Saved tokens for %s", base_url)
 
     async def get_client_info(self) -> OAuthClientInformationFull | None:
-        doc = await asyncio.to_thread(self.collection.find_one, self.get_db_key())
-        if not doc or "client_info" not in doc:
-            logger.debug(
-                "No client_info in DB for %s", self.get_base_url(self.server_url)
-            )
+        data = await asyncio.to_thread(self._fetch_session_data)
+        base_url = self.get_base_url(self.server_url)
+        if not data or "client_info" not in data:
+            logger.debug("No client_info in DB for %s", base_url)
             return None
         try:
-            client_info = OAuthClientInformationFull.model_validate(doc["client_info"])
+            client_info = OAuthClientInformationFull.model_validate(data["client_info"])
             if self.expected_redirect_uri:
                 stored_uris = [
                     str(uri).rstrip("/") for uri in client_info.redirect_uris
@@ -909,14 +941,16 @@ class DBTokenStorage(TokenStorage):
                 if expected_uri not in stored_uris:
                     logger.warning(
                         "Redirect URI mismatch for %s: expected=%s stored=%s — clearing.",
-                        self.get_base_url(self.server_url),
+                        base_url,
                         expected_uri,
                         stored_uris,
                     )
+                    # Drop ``tokens`` and ``client_info`` from the JSONB
+                    # blob via merge_session_data's ``None``-drops-key
+                    # semantics — preserves the row + any other keys.
                     await asyncio.to_thread(
-                        self.collection.update_one,
-                        self.get_db_key(),
-                        {"$unset": {"client_info": "", "tokens": ""}},
+                        self._merge,
+                        {"tokens": None, "client_info": None},
                     )
                     return None
             return client_info
@@ -931,22 +965,37 @@ class DBTokenStorage(TokenStorage):
 
     async def set_client_info(self, client_info: OAuthClientInformationFull) -> None:
         serialized_info = self._serialize_client_info(client_info.model_dump())
+        base_url = self.get_base_url(self.server_url)
         await asyncio.to_thread(
-            self.collection.update_one,
-            self.get_db_key(),
-            {"$set": {"client_info": serialized_info}},
-            True,
+            self._merge, {"client_info": serialized_info},
         )
-        logger.info("Saved client info for %s", self.get_base_url(self.server_url))
+        logger.info("Saved client info for %s", base_url)
 
     async def clear(self) -> None:
-        await asyncio.to_thread(self.collection.delete_one, self.get_db_key())
+        await asyncio.to_thread(self._delete)
         logger.info("Cleared OAuth cache for %s", self.get_base_url(self.server_url))
 
     @classmethod
-    async def clear_all(cls, db_client) -> None:
-        collection = db_client["connector_sessions"]
-        await asyncio.to_thread(collection.delete_many, {})
+    async def clear_all(cls, db_client=None) -> None:
+        """Delete every MCP-tagged connector session row.
+
+        ``db_client`` retained for call-site compatibility but unused —
+        storage is Postgres-only now.
+        """
+        from sqlalchemy import text
+
+        from application.storage.db.session import db_session
+
+        def _delete_all() -> None:
+            with db_session() as conn:
+                conn.execute(
+                    text(
+                        "DELETE FROM connector_sessions "
+                        "WHERE provider LIKE 'mcp:%'"
+                    )
+                )
+
+        await asyncio.to_thread(_delete_all)
         logger.info("Cleared all OAuth client cache data.")
 
 

application/agents/tools/memory.py

@@ -1,12 +1,14 @@
-from datetime import datetime
 from pathlib import Path
 from typing import Any, Dict, List, Optional
-import re
+import logging
 import uuid
 
 from .base import Tool
-from application.core.mongo_db import MongoDB
-from application.core.settings import settings
+from application.storage.db.repositories.memories import MemoriesRepository
+from application.storage.db.session import db_readonly, db_session
+
+
+logger = logging.getLogger(__name__)
 
 
 class MemoryTool(Tool):
@@ -27,7 +29,7 @@ class MemoryTool(Tool):
         self.user_id: Optional[str] = user_id
 
         # Get tool_id from configuration (passed from user_tools._id in production)
-        # In production, tool_id is the MongoDB ObjectId string from user_tools collection
+        # In production, tool_id is the UUID string from user_tools.id.
         if tool_config and "tool_id" in tool_config:
             self.tool_id = tool_config["tool_id"]
         elif user_id:
@@ -37,8 +39,35 @@ class MemoryTool(Tool):
             # Last resort fallback (shouldn't happen in normal use)
             self.tool_id = str(uuid.uuid4())
 
-        db = MongoDB.get_client()[settings.MONGO_DB_NAME]
-        self.collection = db["memories"]
+    def _pg_enabled(self) -> bool:
+        """Return True if this MemoryTool's tool_id is a real ``user_tools.id``.
+
+        The ``memories`` PG table has a UUID foreign key to ``user_tools``.
+        The sentinel ``default_{uid}`` fallback tool_id is not a UUID and
+        has no row in ``user_tools``, so any storage operation would fail
+        the foreign-key check. After the Postgres cutover Postgres is the
+        only store, so for the sentinel case there is nowhere to read or
+        write — operations become no-ops and the tool returns an
+        explanatory error to the caller.
+        """
+        tool_id = getattr(self, "tool_id", None)
+        if not tool_id or not isinstance(tool_id, str):
+            return False
+        if tool_id.startswith("default_"):
+            logger.debug(
+                "Skipping Postgres operation for MemoryTool with sentinel tool_id=%s",
+                tool_id,
+            )
+            return False
+        from application.storage.db.base_repository import looks_like_uuid
+
+        if not looks_like_uuid(tool_id):
+            logger.debug(
+                "Skipping Postgres operation for MemoryTool with non-UUID tool_id=%s",
+                tool_id,
+            )
+            return False
+        return True
 
     # -----------------------------
     # Action implementations
@@ -56,6 +85,12 @@ class MemoryTool(Tool):
         if not self.user_id:
             return "Error: MemoryTool requires a valid user_id."
 
+        if not self._pg_enabled():
+            return (
+                "Error: MemoryTool is not configured with a persistent tool_id; "
+                "memory storage is unavailable for this session."
+            )
+
         if action_name == "view":
             return self._view(
                 kwargs.get("path", "/"),
@@ -282,14 +317,10 @@ class MemoryTool(Tool):
         # Ensure path ends with / for proper prefix matching
         search_path = path if path.endswith("/") else path + "/"
 
-        # Find all files that start with this directory path
-        query = {
-            "user_id": self.user_id,
-            "tool_id": self.tool_id,
-            "path": {"$regex": f"^{re.escape(search_path)}"}
-        }
-
-        docs = list(self.collection.find(query, {"path": 1}))
+        with db_readonly() as conn:
+            docs = MemoriesRepository(conn).list_by_prefix(
+                self.user_id, self.tool_id, search_path
+            )
 
         if not docs:
             return f"Directory: {path}\n(empty)"
@@ -310,7 +341,10 @@ class MemoryTool(Tool):
 
     def _view_file(self, path: str, view_range: Optional[List[int]] = None) -> str:
         """View file contents with optional line range."""
-        doc = self.collection.find_one({"user_id": self.user_id, "tool_id": self.tool_id, "path": path})
+        with db_readonly() as conn:
+            doc = MemoriesRepository(conn).get_by_path(
+                self.user_id, self.tool_id, path
+            )
 
         if not doc or not doc.get("content"):
             return f"Error: File not found: {path}"
@@ -344,16 +378,10 @@ class MemoryTool(Tool):
         if validated_path == "/" or validated_path.endswith("/"):
             return "Error: Cannot create a file at directory path."
 
-        self.collection.update_one(
-            {"user_id": self.user_id, "tool_id": self.tool_id, "path": validated_path},
-            {
-                "$set": {
-                    "content": file_text,
-                    "updated_at": datetime.now()
-                }
-            },
-            upsert=True
-        )
+        with db_session() as conn:
+            MemoriesRepository(conn).upsert(
+                self.user_id, self.tool_id, validated_path, file_text
+            )
 
         return f"File created: {validated_path}"
 
@@ -366,30 +394,29 @@ class MemoryTool(Tool):
         if not old_str:
             return "Error: old_str is required."
 
-        doc = self.collection.find_one({"user_id": self.user_id, "tool_id": self.tool_id, "path": validated_path})
+        with db_session() as conn:
+            repo = MemoriesRepository(conn)
+            doc = repo.get_by_path(self.user_id, self.tool_id, validated_path)
 
-        if not doc or not doc.get("content"):
-            return f"Error: File not found: {validated_path}"
+            if not doc or not doc.get("content"):
+                return f"Error: File not found: {validated_path}"
 
-        current_content = str(doc["content"])
+            current_content = str(doc["content"])
 
-        # Check if old_str exists (case-insensitive)
-        if old_str.lower() not in current_content.lower():
-            return f"Error: String '{old_str}' not found in file."
+            # Check if old_str exists (case-insensitive)
+            if old_str.lower() not in current_content.lower():
+                return f"Error: String '{old_str}' not found in file."
 
-        # Replace the string (case-insensitive)
-        import re as regex_module
-        updated_content = regex_module.sub(regex_module.escape(old_str), new_str, current_content, flags=regex_module.IGNORECASE)
+            # Case-insensitive replace
+            import re as regex_module
+            updated_content = regex_module.sub(
+                regex_module.escape(old_str),
+                new_str,
+                current_content,
+                flags=regex_module.IGNORECASE,
+            )
 
-        self.collection.update_one(
-            {"user_id": self.user_id, "tool_id": self.tool_id, "path": validated_path},
-            {
-                "$set": {
-                    "content": updated_content,
-                    "updated_at": datetime.now()
-                }
-            }
-        )
+            repo.upsert(self.user_id, self.tool_id, validated_path, updated_content)
 
         return f"File updated: {validated_path}"
 
@@ -402,31 +429,25 @@ class MemoryTool(Tool):
         if not insert_text:
             return "Error: insert_text is required."
 
-        doc = self.collection.find_one({"user_id": self.user_id, "tool_id": self.tool_id, "path": validated_path})
+        with db_session() as conn:
+            repo = MemoriesRepository(conn)
+            doc = repo.get_by_path(self.user_id, self.tool_id, validated_path)
 
-        if not doc or not doc.get("content"):
-            return f"Error: File not found: {validated_path}"
+            if not doc or not doc.get("content"):
+                return f"Error: File not found: {validated_path}"
 
-        current_content = str(doc["content"])
-        lines = current_content.split("\n")
+            current_content = str(doc["content"])
+            lines = current_content.split("\n")
 
-        # Convert to 0-indexed
-        index = insert_line - 1
-        if index < 0 or index > len(lines):
-            return f"Error: Invalid line number. File has {len(lines)} lines."
+            # Convert to 0-indexed
+            index = insert_line - 1
+            if index < 0 or index > len(lines):
+                return f"Error: Invalid line number. File has {len(lines)} lines."
 
-        lines.insert(index, insert_text)
-        updated_content = "\n".join(lines)
+            lines.insert(index, insert_text)
+            updated_content = "\n".join(lines)
 
-        self.collection.update_one(
-            {"user_id": self.user_id, "tool_id": self.tool_id, "path": validated_path},
-            {
-                "$set": {
-                    "content": updated_content,
-                    "updated_at": datetime.now()
-                }
-            }
-        )
+            repo.upsert(self.user_id, self.tool_id, validated_path, updated_content)
 
         return f"Text inserted at line {insert_line} in {validated_path}"
 
@@ -438,39 +459,36 @@ class MemoryTool(Tool):
 
         if validated_path == "/":
             # Delete all files for this user and tool
-            result = self.collection.delete_many({"user_id": self.user_id, "tool_id": self.tool_id})
-            return f"Deleted {result.deleted_count} file(s) from memory."
+            with db_session() as conn:
+                deleted = MemoriesRepository(conn).delete_all(
+                    self.user_id, self.tool_id
+                )
+            return f"Deleted {deleted} file(s) from memory."
 
         # Check if it's a directory (ends with /)
         if validated_path.endswith("/"):
-            # Delete all files in directory
-            result = self.collection.delete_many({
-                "user_id": self.user_id,
-                "tool_id": self.tool_id,
-                "path": {"$regex": f"^{re.escape(validated_path)}"}
-            })
-            return f"Deleted directory and {result.deleted_count} file(s)."
+            with db_session() as conn:
+                deleted = MemoriesRepository(conn).delete_by_prefix(
+                    self.user_id, self.tool_id, validated_path
+                )
+            return f"Deleted directory and {deleted} file(s)."
 
-        # Try to delete as directory first (without trailing slash)
-        # Check if any files start with this path + /
+        # Try as directory first (without trailing slash)
         search_path = validated_path + "/"
-        directory_result = self.collection.delete_many({
-            "user_id": self.user_id,
-            "tool_id": self.tool_id,
-            "path": {"$regex": f"^{re.escape(search_path)}"}
-        })
+        with db_session() as conn:
+            repo = MemoriesRepository(conn)
+            directory_deleted = repo.delete_by_prefix(
+                self.user_id, self.tool_id, search_path
+            )
+            if directory_deleted > 0:
+                return f"Deleted directory and {directory_deleted} file(s)."
 
-        if directory_result.deleted_count > 0:
-            return f"Deleted directory and {directory_result.deleted_count} file(s)."
+            # Otherwise delete a single file
+            file_deleted = repo.delete_by_path(
+                self.user_id, self.tool_id, validated_path
+            )
 
-        # Delete single file
-        result = self.collection.delete_one({
-            "user_id": self.user_id,
-            "tool_id": self.tool_id,
-            "path": validated_path
-        })
-
-        if result.deleted_count:
+        if file_deleted:
             return f"Deleted: {validated_path}"
         return f"Error: File not found: {validated_path}"
 
@@ -485,62 +503,46 @@ class MemoryTool(Tool):
         if validated_old == "/" or validated_new == "/":
             return "Error: Cannot rename root directory."
 
-        # Check if renaming a directory
+        # Directory rename: do all path updates inside one transaction so
+        # the rename is atomic from the caller's perspective.
         if validated_old.endswith("/"):
             # Ensure validated_new also ends with / for proper path replacement
             if not validated_new.endswith("/"):
                 validated_new = validated_new + "/"
 
-            # Find all files in the old directory
-            docs = list(self.collection.find({
-                "user_id": self.user_id,
-                "tool_id": self.tool_id,
-                "path": {"$regex": f"^{re.escape(validated_old)}"}
-            }))
-
-            if not docs:
-                return f"Error: Directory not found: {validated_old}"
-
-            # Update paths for all files
-            for doc in docs:
-                old_file_path = doc["path"]
-                new_file_path = old_file_path.replace(validated_old, validated_new, 1)
-
-                self.collection.update_one(
-                    {"_id": doc["_id"]},
-                    {"$set": {"path": new_file_path, "updated_at": datetime.now()}}
+            with db_session() as conn:
+                repo = MemoriesRepository(conn)
+                docs = repo.list_by_prefix(
+                    self.user_id, self.tool_id, validated_old
                 )
 
+                if not docs:
+                    return f"Error: Directory not found: {validated_old}"
+
+                for doc in docs:
+                    old_file_path = doc["path"]
+                    new_file_path = old_file_path.replace(
+                        validated_old, validated_new, 1
+                    )
+                    repo.update_path(
+                        self.user_id, self.tool_id, old_file_path, new_file_path
+                    )
+
             return f"Renamed directory: {validated_old} -> {validated_new} ({len(docs)} files)"
 
-        # Rename single file
-        doc = self.collection.find_one({
-            "user_id": self.user_id,
-            "tool_id": self.tool_id,
-            "path": validated_old
-        })
+        # Single-file rename: lookup, collision check, and update in one txn.
+        with db_session() as conn:
+            repo = MemoriesRepository(conn)
+            doc = repo.get_by_path(self.user_id, self.tool_id, validated_old)
+            if not doc:
+                return f"Error: File not found: {validated_old}"
 
-        if not doc:
-            return f"Error: File not found: {validated_old}"
+            existing = repo.get_by_path(self.user_id, self.tool_id, validated_new)
+            if existing:
+                return f"Error: File already exists at {validated_new}"
 
-        # Check if new path already exists
-        existing = self.collection.find_one({
-            "user_id": self.user_id,
-            "tool_id": self.tool_id,
-            "path": validated_new
-        })
-
-        if existing:
-            return f"Error: File already exists at {validated_new}"
-
-        # Delete the old document and create a new one with the new path
-        self.collection.delete_one({"user_id": self.user_id, "tool_id": self.tool_id, "path": validated_old})
-        self.collection.insert_one({
-            "user_id": self.user_id,
-            "tool_id": self.tool_id,
-            "path": validated_new,
-            "content": doc.get("content", ""),
-            "updated_at": datetime.now()
-        })
+            repo.update_path(
+                self.user_id, self.tool_id, validated_old, validated_new
+            )
 
         return f"Renamed: {validated_old} -> {validated_new}"

application/agents/tools/notes.py

@@ -1,10 +1,16 @@
-from datetime import datetime
 from typing import Any, Dict, List, Optional
 import uuid
 
 from .base import Tool
-from application.core.mongo_db import MongoDB
-from application.core.settings import settings
+from application.storage.db.repositories.notes import NotesRepository
+from application.storage.db.session import db_readonly, db_session
+
+
+# Stable synthetic title used in the Postgres ``notes.title`` column.
+# The notes tool stores one note per (user_id, tool_id); there is no
+# user-facing title. PG requires ``title`` NOT NULL, so we write a stable
+# constant alongside the actual note body in ``content``.
+_NOTE_TITLE = "note"
 
 
 class NotesTool(Tool):
@@ -25,7 +31,6 @@ class NotesTool(Tool):
         self.user_id: Optional[str] = user_id
 
         # Get tool_id from configuration (passed from user_tools._id in production)
-        # In production, tool_id is the MongoDB ObjectId string from user_tools collection
         if tool_config and "tool_id" in tool_config:
             self.tool_id = tool_config["tool_id"]
         elif user_id:
@@ -35,11 +40,25 @@ class NotesTool(Tool):
             # Last resort fallback (shouldn't happen in normal use)
             self.tool_id = str(uuid.uuid4())
 
-        db = MongoDB.get_client()[settings.MONGO_DB_NAME]
-        self.collection = db["notes"]
-
         self._last_artifact_id: Optional[str] = None
 
+    def _pg_enabled(self) -> bool:
+        """Return True only when ``tool_id`` is a real ``user_tools.id`` UUID.
+
+        ``notes.tool_id`` is a UUID FK to ``user_tools``; repo queries
+        ``CAST(:tool_id AS uuid)``. The sentinel ``default_{uid}``
+        fallback is neither a UUID nor a ``user_tools`` row, so any DB
+        operation would crash. Mirror MemoryTool's guard and no-op.
+        """
+        tool_id = getattr(self, "tool_id", None)
+        if not tool_id or not isinstance(tool_id, str):
+            return False
+        if tool_id.startswith("default_"):
+            return False
+        from application.storage.db.base_repository import looks_like_uuid
+
+        return looks_like_uuid(tool_id)
+
     # -----------------------------
     # Action implementations
     # -----------------------------
@@ -54,7 +73,13 @@ class NotesTool(Tool):
             A human-readable string result.
         """
         if not self.user_id:
-             return "Error: NotesTool requires a valid user_id."
+            return "Error: NotesTool requires a valid user_id."
+
+        if not self._pg_enabled():
+            return (
+                "Error: NotesTool is not configured with a persistent "
+                "tool_id; note storage is unavailable for this session."
+            )
 
         self._last_artifact_id = None
 
@@ -135,37 +160,45 @@ class NotesTool(Tool):
     # -----------------------------
     # Internal helpers (single-note)
     # -----------------------------
+    def _fetch_note(self) -> Optional[dict]:
+        """Read the note row for this (user, tool) from Postgres."""
+        with db_readonly() as conn:
+            return NotesRepository(conn).get_for_user_tool(self.user_id, self.tool_id)
+
     def _get_note(self) -> str:
-        doc = self.collection.find_one({"user_id": self.user_id, "tool_id": self.tool_id})
-        if not doc or not doc.get("note"):
+        doc = self._fetch_note()
+        # ``content`` is the PG column; expose as ``note`` to callers via the
+        # textual return value. Frontends that read the artifact via the
+        # repo dict get ``content`` (PG-native) plus the artifact id below.
+        body = (doc or {}).get("content")
+        if not doc or not body:
             return "No note found."
-        if doc.get("_id") is not None:
-            self._last_artifact_id = str(doc.get("_id"))
-        return str(doc["note"])
+        if doc.get("id") is not None:
+            self._last_artifact_id = str(doc.get("id"))
+        return str(body)
 
     def _overwrite_note(self, content: str) -> str:
         content = (content or "").strip()
         if not content:
             return "Note content required."
-        result = self.collection.find_one_and_update(
-            {"user_id": self.user_id, "tool_id": self.tool_id},
-            {"$set": {"note": content, "updated_at": datetime.utcnow()}},
-            upsert=True,
-            return_document=True,
-        )
-        if result and result.get("_id") is not None:
-            self._last_artifact_id = str(result.get("_id"))
+        with db_session() as conn:
+            row = NotesRepository(conn).upsert(
+                self.user_id, self.tool_id, _NOTE_TITLE, content
+            )
+        if row and row.get("id") is not None:
+            self._last_artifact_id = str(row.get("id"))
         return "Note saved."
 
     def _str_replace(self, old_str: str, new_str: str) -> str:
         if not old_str:
             return "old_str is required."
 
-        doc = self.collection.find_one({"user_id": self.user_id, "tool_id": self.tool_id})
-        if not doc or not doc.get("note"):
+        doc = self._fetch_note()
+        existing = (doc or {}).get("content")
+        if not doc or not existing:
             return "No note found."
 
-        current_note = str(doc["note"])
+        current_note = str(existing)
 
         # Case-insensitive search
         if old_str.lower() not in current_note.lower():
@@ -175,24 +208,24 @@ class NotesTool(Tool):
         import re
         updated_note = re.sub(re.escape(old_str), new_str, current_note, flags=re.IGNORECASE)
 
-        result = self.collection.find_one_and_update(
-            {"user_id": self.user_id, "tool_id": self.tool_id},
-            {"$set": {"note": updated_note, "updated_at": datetime.utcnow()}},
-            return_document=True,
-        )
-        if result and result.get("_id") is not None:
-            self._last_artifact_id = str(result.get("_id"))
+        with db_session() as conn:
+            row = NotesRepository(conn).upsert(
+                self.user_id, self.tool_id, _NOTE_TITLE, updated_note
+            )
+        if row and row.get("id") is not None:
+            self._last_artifact_id = str(row.get("id"))
         return "Note updated."
 
     def _insert(self, line_number: int, text: str) -> str:
         if not text:
             return "Text is required."
 
-        doc = self.collection.find_one({"user_id": self.user_id, "tool_id": self.tool_id})
-        if not doc or not doc.get("note"):
+        doc = self._fetch_note()
+        existing = (doc or {}).get("content")
+        if not doc or not existing:
             return "No note found."
 
-        current_note = str(doc["note"])
+        current_note = str(existing)
         lines = current_note.split("\n")
 
         # Convert to 0-indexed and validate
@@ -203,21 +236,23 @@ class NotesTool(Tool):
         lines.insert(index, text)
         updated_note = "\n".join(lines)
 
-        result = self.collection.find_one_and_update(
-            {"user_id": self.user_id, "tool_id": self.tool_id},
-            {"$set": {"note": updated_note, "updated_at": datetime.utcnow()}},
-            return_document=True,
-        )
-        if result and result.get("_id") is not None:
-            self._last_artifact_id = str(result.get("_id"))
+        with db_session() as conn:
+            row = NotesRepository(conn).upsert(
+                self.user_id, self.tool_id, _NOTE_TITLE, updated_note
+            )
+        if row and row.get("id") is not None:
+            self._last_artifact_id = str(row.get("id"))
         return "Text inserted."
 
     def _delete_note(self) -> str:
-        doc = self.collection.find_one_and_delete(
-            {"user_id": self.user_id, "tool_id": self.tool_id}
-        )
-        if not doc:
+        # Capture the id (for artifact tracking) before deleting.
+        existing = self._fetch_note()
+        if not existing:
             return "No note found to delete."
-        if doc.get("_id") is not None:
-            self._last_artifact_id = str(doc.get("_id"))
+        with db_session() as conn:
+            deleted = NotesRepository(conn).delete(self.user_id, self.tool_id)
+        if not deleted:
+            return "No note found to delete."
+        if existing.get("id") is not None:
+            self._last_artifact_id = str(existing.get("id"))
         return "Note deleted."

application/agents/tools/todo_list.py

@@ -1,10 +1,19 @@
-from datetime import datetime
 from typing import Any, Dict, List, Optional
 import uuid
 
 from .base import Tool
-from application.core.mongo_db import MongoDB
-from application.core.settings import settings
+from application.storage.db.repositories.todos import TodosRepository
+from application.storage.db.session import db_readonly, db_session
+
+
+def _status_from_completed(completed: Any) -> str:
+    """Translate the PG ``completed`` boolean to the legacy status string.
+
+    The frontend (and prior LLM-facing tool output) expects
+    ``"open"`` / ``"completed"``. Keeping that contract at the tool
+    boundary insulates callers from the schema change.
+    """
+    return "completed" if bool(completed) else "open"
 
 
 class TodoListTool(Tool):
@@ -25,7 +34,6 @@ class TodoListTool(Tool):
         self.user_id: Optional[str] = user_id
 
         # Get tool_id from configuration (passed from user_tools._id in production)
-        # In production, tool_id is the MongoDB ObjectId string from user_tools collection
         if tool_config and "tool_id" in tool_config:
             self.tool_id = tool_config["tool_id"]
         elif user_id:
@@ -35,11 +43,27 @@ class TodoListTool(Tool):
             # Last resort fallback (shouldn't happen in normal use)
             self.tool_id = str(uuid.uuid4())
 
-        db = MongoDB.get_client()[settings.MONGO_DB_NAME]
-        self.collection = db["todos"]
-
         self._last_artifact_id: Optional[str] = None
 
+    def _pg_enabled(self) -> bool:
+        """Return True only when ``tool_id`` is a real ``user_tools.id`` UUID.
+
+        The ``todos`` PG table has a UUID foreign key to ``user_tools`` and
+        the repo queries ``CAST(:tool_id AS uuid)``. The sentinel
+        ``default_{uid}`` fallback is neither a UUID nor a row in
+        ``user_tools`` — binding it would crash ``invalid input syntax for
+        type uuid`` and even if it didn't the FK would reject it. Mirror
+        the MemoryTool guard and no-op in that case.
+        """
+        tool_id = getattr(self, "tool_id", None)
+        if not tool_id or not isinstance(tool_id, str):
+            return False
+        if tool_id.startswith("default_"):
+            return False
+        from application.storage.db.base_repository import looks_like_uuid
+
+        return looks_like_uuid(tool_id)
+
     # -----------------------------
     # Action implementations
     # -----------------------------
@@ -56,6 +80,12 @@ class TodoListTool(Tool):
         if not self.user_id:
             return "Error: TodoListTool requires a valid user_id."
 
+        if not self._pg_enabled():
+            return (
+                "Error: TodoListTool is not configured with a persistent "
+                "tool_id; todo storage is unavailable for this session."
+            )
+
         self._last_artifact_id = None
 
         if action_name == "list":
@@ -191,28 +221,10 @@ class TodoListTool(Tool):
 
         return None
 
-    def _get_next_todo_id(self) -> int:
-        """Get the next sequential todo_id for this user and tool.
-
-        Returns a simple integer (1, 2, 3, ...) scoped to this user/tool.
-        With 5-10 todos max, scanning is negligible.
-        """
-        query = {"user_id": self.user_id, "tool_id": self.tool_id}
-        todos = list(self.collection.find(query, {"todo_id": 1}))
-
-        # Find the maximum todo_id
-        max_id = 0
-        for todo in todos:
-            todo_id = self._coerce_todo_id(todo.get("todo_id"))
-            if todo_id is not None:
-                max_id = max(max_id, todo_id)
-
-        return max_id + 1
-
     def _list(self) -> str:
         """List all todos for the user."""
-        query = {"user_id": self.user_id, "tool_id": self.tool_id}
-        todos = list(self.collection.find(query))
+        with db_readonly() as conn:
+            todos = TodosRepository(conn).list_for_tool(self.user_id, self.tool_id)
 
         if not todos:
             return "No todos found."
@@ -221,7 +233,7 @@ class TodoListTool(Tool):
         for doc in todos:
             todo_id = doc.get("todo_id")
             title = doc.get("title", "Untitled")
-            status = doc.get("status", "open")
+            status = _status_from_completed(doc.get("completed"))
 
             line = f"[{todo_id}] {title} ({status})"
             result_lines.append(line)
@@ -229,27 +241,23 @@ class TodoListTool(Tool):
         return "\n".join(result_lines)
 
     def _create(self, title: str) -> str:
-        """Create a new todo item."""
+        """Create a new todo item.
+
+        ``TodosRepository.create`` allocates the per-tool monotonic
+        ``todo_id`` inside the same transaction (``COALESCE(MAX(todo_id),0)+1``
+        scoped to ``tool_id``), so we no longer need a separate read-then-
+        write step here.
+        """
         title = (title or "").strip()
         if not title:
             return "Error: Title is required."
 
-        now = datetime.now()
-        todo_id = self._get_next_todo_id()
+        with db_session() as conn:
+            row = TodosRepository(conn).create(self.user_id, self.tool_id, title)
 
-        doc = {
-            "todo_id": todo_id,
-            "user_id": self.user_id,
-            "tool_id": self.tool_id,
-            "title": title,
-            "status": "open",
-            "created_at": now,
-            "updated_at": now,
-        }
-        insert_result = self.collection.insert_one(doc)
-        inserted_id = getattr(insert_result, "inserted_id", None) or doc.get("_id")
-        if inserted_id is not None:
-            self._last_artifact_id = str(inserted_id)
+        todo_id = row.get("todo_id")
+        if row.get("id") is not None:
+            self._last_artifact_id = str(row.get("id"))
         return f"Todo created with ID {todo_id}: {title}"
 
     def _get(self, todo_id: Optional[Any]) -> str:
@@ -258,21 +266,21 @@ class TodoListTool(Tool):
         if parsed_todo_id is None:
             return "Error: todo_id must be a positive integer."
 
-        query = {"user_id": self.user_id, "tool_id": self.tool_id, "todo_id": parsed_todo_id}
-        doc = self.collection.find_one(query)
+        with db_readonly() as conn:
+            doc = TodosRepository(conn).get_by_tool_and_todo_id(
+                self.user_id, self.tool_id, parsed_todo_id
+            )
 
         if not doc:
             return f"Error: Todo with ID {parsed_todo_id} not found."
 
-        if doc.get("_id") is not None:
-            self._last_artifact_id = str(doc.get("_id"))
+        if doc.get("id") is not None:
+            self._last_artifact_id = str(doc.get("id"))
 
         title = doc.get("title", "Untitled")
-        status = doc.get("status", "open")
+        status = _status_from_completed(doc.get("completed"))
 
-        result = f"Todo [{parsed_todo_id}]:\nTitle: {title}\nStatus: {status}"
-
-        return result
+        return f"Todo [{parsed_todo_id}]:\nTitle: {title}\nStatus: {status}"
 
     def _update(self, todo_id: Optional[Any], title: str) -> str:
         """Update a todo's title by ID."""
@@ -284,16 +292,19 @@ class TodoListTool(Tool):
         if not title:
             return "Error: Title is required."
 
-        query = {"user_id": self.user_id, "tool_id": self.tool_id, "todo_id": parsed_todo_id}
-        doc = self.collection.find_one_and_update(
-            query,
-            {"$set": {"title": title, "updated_at": datetime.now()}},
-        )
-        if not doc:
-            return f"Error: Todo with ID {parsed_todo_id} not found."
+        with db_session() as conn:
+            repo = TodosRepository(conn)
+            existing = repo.get_by_tool_and_todo_id(
+                self.user_id, self.tool_id, parsed_todo_id
+            )
+            if not existing:
+                return f"Error: Todo with ID {parsed_todo_id} not found."
+            repo.update_title_by_tool_and_todo_id(
+                self.user_id, self.tool_id, parsed_todo_id, title
+            )
 
-        if doc.get("_id") is not None:
-            self._last_artifact_id = str(doc.get("_id"))
+        if existing.get("id") is not None:
+            self._last_artifact_id = str(existing.get("id"))
 
         return f"Todo {parsed_todo_id} updated to: {title}"
 
@@ -303,16 +314,17 @@ class TodoListTool(Tool):
         if parsed_todo_id is None:
             return "Error: todo_id must be a positive integer."
 
-        query = {"user_id": self.user_id, "tool_id": self.tool_id, "todo_id": parsed_todo_id}
-        doc = self.collection.find_one_and_update(
-            query,
-            {"$set": {"status": "completed", "updated_at": datetime.now()}},
-        )
-        if not doc:
-            return f"Error: Todo with ID {parsed_todo_id} not found."
+        with db_session() as conn:
+            repo = TodosRepository(conn)
+            existing = repo.get_by_tool_and_todo_id(
+                self.user_id, self.tool_id, parsed_todo_id
+            )
+            if not existing:
+                return f"Error: Todo with ID {parsed_todo_id} not found."
+            repo.set_completed(self.user_id, self.tool_id, parsed_todo_id, True)
 
-        if doc.get("_id") is not None:
-            self._last_artifact_id = str(doc.get("_id"))
+        if existing.get("id") is not None:
+            self._last_artifact_id = str(existing.get("id"))
 
         return f"Todo {parsed_todo_id} marked as completed."
 
@@ -322,12 +334,18 @@ class TodoListTool(Tool):
         if parsed_todo_id is None:
             return "Error: todo_id must be a positive integer."
 
-        query = {"user_id": self.user_id, "tool_id": self.tool_id, "todo_id": parsed_todo_id}
-        doc = self.collection.find_one_and_delete(query)
-        if not doc:
-            return f"Error: Todo with ID {parsed_todo_id} not found."
+        with db_session() as conn:
+            repo = TodosRepository(conn)
+            existing = repo.get_by_tool_and_todo_id(
+                self.user_id, self.tool_id, parsed_todo_id
+            )
+            if not existing:
+                return f"Error: Todo with ID {parsed_todo_id} not found."
+            repo.delete_by_tool_and_todo_id(
+                self.user_id, self.tool_id, parsed_todo_id
+            )
 
-        if doc.get("_id") is not None:
-            self._last_artifact_id = str(doc.get("_id"))
+        if existing.get("id") is not None:
+            self._last_artifact_id = str(existing.get("id"))
 
         return f"Todo {parsed_todo_id} deleted."

application/agents/workflow_agent.py

@@ -12,12 +12,13 @@ from application.agents.workflows.schemas import (
     WorkflowRun,
 )
 from application.agents.workflows.workflow_engine import WorkflowEngine
-from application.core.mongo_db import MongoDB
-from application.core.settings import settings
 from application.logging import log_activity, LogContext
-from application.storage.db.dual_write import dual_write
+from application.storage.db.base_repository import looks_like_uuid
+from application.storage.db.repositories.workflow_edges import WorkflowEdgesRepository
+from application.storage.db.repositories.workflow_nodes import WorkflowNodesRepository
 from application.storage.db.repositories.workflow_runs import WorkflowRunsRepository
 from application.storage.db.repositories.workflows import WorkflowsRepository
+from application.storage.db.session import db_readonly, db_session
 
 logger = logging.getLogger(__name__)
 
@@ -106,10 +107,8 @@ class WorkflowAgent(BaseAgent):
 
     def _load_from_database(self) -> Optional[WorkflowGraph]:
         try:
-            from bson.objectid import ObjectId
-
-            if not self.workflow_id or not ObjectId.is_valid(self.workflow_id):
-                logger.error(f"Invalid workflow ID: {self.workflow_id}")
+            if not self.workflow_id:
+                logger.error("Missing workflow ID for load")
                 return None
             owner_id = self.workflow_owner
             if not owner_id and isinstance(self.decoded_token, dict):
@@ -120,61 +119,61 @@ class WorkflowAgent(BaseAgent):
                 )
                 return None
 
-            mongo = MongoDB.get_client()
-            db = mongo[settings.MONGO_DB_NAME]
-
-            workflows_coll = db["workflows"]
-            workflow_nodes_coll = db["workflow_nodes"]
-            workflow_edges_coll = db["workflow_edges"]
-
-            workflow_doc = workflows_coll.find_one(
-                {"_id": ObjectId(self.workflow_id), "user": owner_id}
-            )
-            if not workflow_doc:
-                logger.error(
-                    f"Workflow {self.workflow_id} not found or inaccessible for user {owner_id}"
-                )
-                return None
-            workflow = Workflow(**workflow_doc)
-            graph_version = workflow_doc.get("current_graph_version", 1)
-            try:
-                graph_version = int(graph_version)
-                if graph_version <= 0:
+            with db_readonly() as conn:
+                wf_repo = WorkflowsRepository(conn)
+                if looks_like_uuid(self.workflow_id):
+                    workflow_row = wf_repo.get(self.workflow_id, owner_id)
+                else:
+                    workflow_row = wf_repo.get_by_legacy_id(self.workflow_id, owner_id)
+                if workflow_row is None:
+                    logger.error(
+                        f"Workflow {self.workflow_id} not found or inaccessible "
+                        f"for user {owner_id}"
+                    )
+                    return None
+                pg_workflow_id = str(workflow_row["id"])
+                graph_version = workflow_row.get("current_graph_version", 1)
+                try:
+                    graph_version = int(graph_version)
+                    if graph_version <= 0:
+                        graph_version = 1
+                except (ValueError, TypeError):
                     graph_version = 1
-            except (ValueError, TypeError):
-                graph_version = 1
 
-            nodes_docs = list(
-                workflow_nodes_coll.find(
-                    {"workflow_id": self.workflow_id, "graph_version": graph_version}
+                node_rows = WorkflowNodesRepository(conn).find_by_version(
+                    pg_workflow_id, graph_version,
                 )
-            )
-            if not nodes_docs and graph_version == 1:
-                nodes_docs = list(
-                    workflow_nodes_coll.find(
-                        {
-                            "workflow_id": self.workflow_id,
-                            "graph_version": {"$exists": False},
-                        }
-                    )
+                edge_rows = WorkflowEdgesRepository(conn).find_by_version(
+                    pg_workflow_id, graph_version,
                 )
-            nodes = [WorkflowNode(**doc) for doc in nodes_docs]
 
-            edges_docs = list(
-                workflow_edges_coll.find(
-                    {"workflow_id": self.workflow_id, "graph_version": graph_version}
-                )
+            workflow = Workflow(
+                name=workflow_row.get("name"),
+                description=workflow_row.get("description"),
             )
-            if not edges_docs and graph_version == 1:
-                edges_docs = list(
-                    workflow_edges_coll.find(
-                        {
-                            "workflow_id": self.workflow_id,
-                            "graph_version": {"$exists": False},
-                        }
-                    )
+            nodes = [
+                WorkflowNode(
+                    id=n["node_id"],
+                    workflow_id=pg_workflow_id,
+                    type=n["node_type"],
+                    title=n.get("title") or "Node",
+                    description=n.get("description"),
+                    position=n.get("position") or {"x": 0, "y": 0},
+                    config=n.get("config") or {},
                 )
-            edges = [WorkflowEdge(**doc) for doc in edges_docs]
+                for n in node_rows
+            ]
+            edges = [
+                WorkflowEdge(
+                    id=e["edge_id"],
+                    workflow_id=pg_workflow_id,
+                    source=e.get("source_id"),
+                    target=e.get("target_id"),
+                    sourceHandle=e.get("source_handle"),
+                    targetHandle=e.get("target_handle"),
+                )
+                for e in edge_rows
+            ]
 
             return WorkflowGraph(workflow=workflow, nodes=nodes, edges=edges)
         except Exception as e:
@@ -188,10 +187,6 @@ class WorkflowAgent(BaseAgent):
         if not owner_id and isinstance(self.decoded_token, dict):
             owner_id = self.decoded_token.get("sub")
         try:
-            mongo = MongoDB.get_client()
-            db = mongo[settings.MONGO_DB_NAME]
-            workflow_runs_coll = db["workflow_runs"]
-
             run = WorkflowRun(
                 workflow_id=self.workflow_id or "unknown",
                 user=owner_id,
@@ -203,23 +198,20 @@ class WorkflowAgent(BaseAgent):
                 completed_at=datetime.now(timezone.utc),
             )
 
-            result = workflow_runs_coll.insert_one(run.to_mongo_doc())
-            legacy_mongo_id = (
-                str(result.inserted_id)
-                if getattr(result, "inserted_id", None) is not None
-                else None
-            )
-
-            def _pg_write(repo: WorkflowRunsRepository) -> None:
-                if not self.workflow_id or not owner_id or not legacy_mongo_id:
+            if not self.workflow_id or not owner_id:
+                return
+            with db_session() as conn:
+                wf_repo = WorkflowsRepository(conn)
+                if looks_like_uuid(self.workflow_id):
+                    workflow_row = wf_repo.get(self.workflow_id, owner_id)
+                else:
+                    workflow_row = wf_repo.get_by_legacy_id(
+                        self.workflow_id, owner_id,
+                    )
+                if workflow_row is None:
                     return
-                workflow = WorkflowsRepository(repo._conn).get_by_legacy_id(
-                    self.workflow_id, owner_id,
-                )
-                if workflow is None:
-                    return
-                repo.create(
-                    workflow["id"],
+                WorkflowRunsRepository(conn).create(
+                    str(workflow_row["id"]),
                     owner_id,
                     run.status.value,
                     inputs=run.inputs,
@@ -227,10 +219,7 @@ class WorkflowAgent(BaseAgent):
                     steps=[step.model_dump(mode="json") for step in run.steps],
                     started_at=run.created_at,
                     ended_at=run.completed_at,
-                    legacy_mongo_id=legacy_mongo_id,
                 )
-
-            dual_write(WorkflowRunsRepository, _pg_write)
         except Exception as e:
             logger.error(f"Failed to save workflow run: {e}")
 

application/agents/workflows/schemas.py

@@ -2,7 +2,6 @@ from datetime import datetime, timezone
 from enum import Enum
 from typing import Any, Dict, List, Literal, Optional, Union
 
-from bson import ObjectId
 from pydantic import BaseModel, ConfigDict, Field, field_validator
 
 
@@ -81,24 +80,7 @@ class WorkflowEdgeCreate(BaseModel):
 
 
 class WorkflowEdge(WorkflowEdgeCreate):
-    mongo_id: Optional[str] = Field(None, alias="_id")
-
-    @field_validator("mongo_id", mode="before")
-    @classmethod
-    def convert_objectid(cls, v: Any) -> Optional[str]:
-        if isinstance(v, ObjectId):
-            return str(v)
-        return v
-
-    def to_mongo_doc(self) -> Dict[str, Any]:
-        return {
-            "id": self.id,
-            "workflow_id": self.workflow_id,
-            "source_id": self.source_id,
-            "target_id": self.target_id,
-            "source_handle": self.source_handle,
-            "target_handle": self.target_handle,
-        }
+    pass
 
 
 class WorkflowNodeCreate(BaseModel):
@@ -120,25 +102,7 @@ class WorkflowNodeCreate(BaseModel):
 
 
 class WorkflowNode(WorkflowNodeCreate):
-    mongo_id: Optional[str] = Field(None, alias="_id")
-
-    @field_validator("mongo_id", mode="before")
-    @classmethod
-    def convert_objectid(cls, v: Any) -> Optional[str]:
-        if isinstance(v, ObjectId):
-            return str(v)
-        return v
-
-    def to_mongo_doc(self) -> Dict[str, Any]:
-        return {
-            "id": self.id,
-            "workflow_id": self.workflow_id,
-            "type": self.type.value,
-            "title": self.title,
-            "description": self.description,
-            "position": self.position.model_dump(),
-            "config": self.config,
-        }
+    pass
 
 
 class WorkflowCreate(BaseModel):
@@ -149,26 +113,10 @@ class WorkflowCreate(BaseModel):
 
 
 class Workflow(WorkflowCreate):
-    id: Optional[str] = Field(None, alias="_id")
+    id: Optional[str] = None
     created_at: datetime = Field(default_factory=lambda: datetime.now(timezone.utc))
     updated_at: datetime = Field(default_factory=lambda: datetime.now(timezone.utc))
 
-    @field_validator("id", mode="before")
-    @classmethod
-    def convert_objectid(cls, v: Any) -> Optional[str]:
-        if isinstance(v, ObjectId):
-            return str(v)
-        return v
-
-    def to_mongo_doc(self) -> Dict[str, Any]:
-        return {
-            "name": self.name,
-            "description": self.description,
-            "user": self.user,
-            "created_at": self.created_at,
-            "updated_at": self.updated_at,
-        }
-
 
 class WorkflowGraph(BaseModel):
     workflow: Workflow
@@ -209,7 +157,7 @@ class WorkflowRunCreate(BaseModel):
 
 class WorkflowRun(BaseModel):
     model_config = ConfigDict(extra="allow")
-    id: Optional[str] = Field(None, alias="_id")
+    id: Optional[str] = None
     workflow_id: str
     user: Optional[str] = None
     status: ExecutionStatus = ExecutionStatus.PENDING
@@ -218,25 +166,3 @@ class WorkflowRun(BaseModel):
     steps: List[NodeExecutionLog] = Field(default_factory=list)
     created_at: datetime = Field(default_factory=lambda: datetime.now(timezone.utc))
     completed_at: Optional[datetime] = None
-
-    @field_validator("id", mode="before")
-    @classmethod
-    def convert_objectid(cls, v: Any) -> Optional[str]:
-        if isinstance(v, ObjectId):
-            return str(v)
-        return v
-
-    def to_mongo_doc(self) -> Dict[str, Any]:
-        doc = {
-            "workflow_id": self.workflow_id,
-            "status": self.status.value,
-            "inputs": self.inputs,
-            "outputs": self.outputs,
-            "steps": [step.model_dump() for step in self.steps],
-            "created_at": self.created_at,
-            "completed_at": self.completed_at,
-        }
-        if self.user:
-            doc["user"] = self.user
-            doc["user_id"] = self.user
-        return doc

application/agents/workflows/workflow_engine.py

@@ -200,6 +200,9 @@ class WorkflowEngine:
 
         node_config = AgentNodeConfig(**node.config.get("config", node.config))
 
+        if node_config.sources:
+            self._retrieve_node_sources(node_config)
+
         if node_config.prompt_template:
             formatted_prompt = self._format_template(node_config.prompt_template)
         else:
@@ -208,15 +211,26 @@ class WorkflowEngine:
             node_config.json_schema, node.title
         )
         node_model_id = node_config.model_id or self.agent.model_id
+        # Inherit BYOM scope from parent agent so owner-stored BYOM
+        # resolves on shared workflows.
+        node_user_id = getattr(self.agent, "model_user_id", None) or (
+            self.agent.decoded_token.get("sub")
+            if isinstance(self.agent.decoded_token, dict)
+            else None
+        )
         node_llm_name = (
             node_config.llm_name
-            or get_provider_from_model_id(node_model_id or "")
+            or get_provider_from_model_id(
+                node_model_id or "", user_id=node_user_id
+            )
             or self.agent.llm_name
         )
         node_api_key = get_api_key_for_provider(node_llm_name) or self.agent.api_key
 
         if node_json_schema and node_model_id:
-            model_capabilities = get_model_capabilities(node_model_id)
+            model_capabilities = get_model_capabilities(
+                node_model_id, user_id=node_user_id
+            )
             if model_capabilities and not model_capabilities.get(
                 "supports_structured_output", False
             ):
@@ -229,6 +243,7 @@ class WorkflowEngine:
             "endpoint": self.agent.endpoint,
             "llm_name": node_llm_name,
             "model_id": node_model_id,
+            "model_user_id": getattr(self.agent, "model_user_id", None),
             "api_key": node_api_key,
             "tool_ids": node_config.tools,
             "prompt": node_config.system_prompt,
@@ -455,6 +470,29 @@ class WorkflowEngine:
         docs_together = "\n\n".join(docs_together_parts) if docs_together_parts else None
         return docs, docs_together
 
+    def _retrieve_node_sources(self, node_config: AgentNodeConfig) -> None:
+        """Retrieve documents from the node's sources for template resolution."""
+        from application.retriever.retriever_creator import RetrieverCreator
+
+        query = self.state.get("query", "")
+        if not query:
+            return
+
+        try:
+            retriever = RetrieverCreator.create_retriever(
+                node_config.retriever or "classic",
+                source={"active_docs": node_config.sources},
+                chat_history=[],
+                prompt="",
+                chunks=int(node_config.chunks) if node_config.chunks else 2,
+                decoded_token=self.agent.decoded_token,
+            )
+            docs = retriever.search(query)
+            if docs:
+                self.agent.retrieved_docs = docs
+        except Exception:
+            logger.exception("Failed to retrieve docs for workflow node")
+
     def get_execution_summary(self) -> List[NodeExecutionLog]:
         return [
             NodeExecutionLog(

application/alembic/versions/0001_initial.py

@@ -167,14 +167,19 @@ def upgrade() -> None:
     op.execute(
         """
         CREATE TABLE user_tools (
-            id           UUID PRIMARY KEY DEFAULT gen_random_uuid(),
-            user_id      TEXT NOT NULL,
-            name         TEXT NOT NULL,
-            custom_name  TEXT,
-            display_name TEXT,
-            config       JSONB NOT NULL DEFAULT '{}'::jsonb,
-            created_at   TIMESTAMPTZ NOT NULL DEFAULT now(),
-            updated_at   TIMESTAMPTZ NOT NULL DEFAULT now()
+            id                  UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+            user_id             TEXT NOT NULL,
+            name                TEXT NOT NULL,
+            custom_name         TEXT,
+            display_name        TEXT,
+            description         TEXT,
+            config              JSONB NOT NULL DEFAULT '{}'::jsonb,
+            config_requirements JSONB NOT NULL DEFAULT '{}'::jsonb,
+            actions             JSONB NOT NULL DEFAULT '[]'::jsonb,
+            status              BOOLEAN NOT NULL DEFAULT true,
+            created_at          TIMESTAMPTZ NOT NULL DEFAULT now(),
+            updated_at          TIMESTAMPTZ NOT NULL DEFAULT now(),
+            legacy_mongo_id     TEXT
         );
         """
     )
@@ -188,7 +193,8 @@ def upgrade() -> None:
             agent_id         UUID,
             prompt_tokens    INTEGER NOT NULL DEFAULT 0,
             generated_tokens INTEGER NOT NULL DEFAULT 0,
-            timestamp        TIMESTAMPTZ NOT NULL DEFAULT now()
+            timestamp        TIMESTAMPTZ NOT NULL DEFAULT now(),
+            mongo_id         TEXT
         );
         """
     )
@@ -204,7 +210,8 @@ def upgrade() -> None:
             user_id   TEXT,
             endpoint  TEXT,
             timestamp TIMESTAMPTZ NOT NULL DEFAULT now(),
-            data      JSONB
+            data      JSONB,
+            mongo_id  TEXT
         );
         """
     )
@@ -220,7 +227,8 @@ def upgrade() -> None:
             api_key     TEXT,
             query       TEXT,
             stacks      JSONB NOT NULL DEFAULT '[]'::jsonb,
-            timestamp   TIMESTAMPTZ NOT NULL DEFAULT now()
+            timestamp   TIMESTAMPTZ NOT NULL DEFAULT now(),
+            mongo_id    TEXT
         );
         """
     )
@@ -228,12 +236,14 @@ def upgrade() -> None:
     op.execute(
         """
         CREATE TABLE agent_folders (
-            id          UUID PRIMARY KEY DEFAULT gen_random_uuid(),
-            user_id     TEXT NOT NULL,
-            name        TEXT NOT NULL,
-            description TEXT,
-            created_at  TIMESTAMPTZ NOT NULL DEFAULT now(),
-            updated_at  TIMESTAMPTZ NOT NULL DEFAULT now()
+            id              UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+            user_id         TEXT NOT NULL,
+            name            TEXT NOT NULL,
+            description     TEXT,
+            parent_id       UUID REFERENCES agent_folders(id) ON DELETE SET NULL,
+            created_at      TIMESTAMPTZ NOT NULL DEFAULT now(),
+            updated_at      TIMESTAMPTZ NOT NULL DEFAULT now(),
+            legacy_mongo_id TEXT
         );
         """
     )
@@ -241,13 +251,24 @@ def upgrade() -> None:
     op.execute(
         """
         CREATE TABLE sources (
-            id         UUID PRIMARY KEY DEFAULT gen_random_uuid(),
-            user_id    TEXT NOT NULL,
-            name       TEXT NOT NULL,
-            type       TEXT,
-            metadata   JSONB NOT NULL DEFAULT '{}'::jsonb,
-            created_at TIMESTAMPTZ NOT NULL DEFAULT now(),
-            updated_at TIMESTAMPTZ NOT NULL DEFAULT now()
+            id                  UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+            user_id             TEXT NOT NULL,
+            name                TEXT NOT NULL,
+            language            TEXT,
+            date                TIMESTAMPTZ NOT NULL DEFAULT now(),
+            model               TEXT,
+            type                TEXT,
+            metadata            JSONB NOT NULL DEFAULT '{}'::jsonb,
+            retriever           TEXT,
+            sync_frequency      TEXT,
+            tokens              TEXT,
+            file_path           TEXT,
+            remote_data         JSONB,
+            directory_structure JSONB,
+            file_name_map       JSONB,
+            created_at          TIMESTAMPTZ NOT NULL DEFAULT now(),
+            updated_at          TIMESTAMPTZ NOT NULL DEFAULT now(),
+            legacy_mongo_id     TEXT
         );
         """
     )
@@ -255,33 +276,38 @@ def upgrade() -> None:
     op.execute(
         """
         CREATE TABLE agents (
-            id                     UUID PRIMARY KEY DEFAULT gen_random_uuid(),
-            user_id                TEXT NOT NULL,
-            name                   TEXT NOT NULL,
-            description            TEXT,
-            agent_type             TEXT,
-            status                 TEXT NOT NULL,
-            key                    CITEXT UNIQUE,
-            source_id              UUID REFERENCES sources(id) ON DELETE SET NULL,
-            extra_source_ids       UUID[] NOT NULL DEFAULT '{}',
-            chunks                 INTEGER,
-            retriever              TEXT,
-            prompt_id              UUID REFERENCES prompts(id) ON DELETE SET NULL,
-            tools                  JSONB NOT NULL DEFAULT '[]'::jsonb,
-            json_schema            JSONB,
-            models                 JSONB,
-            default_model_id       TEXT,
-            folder_id              UUID REFERENCES agent_folders(id) ON DELETE SET NULL,
-            limited_token_mode     BOOLEAN NOT NULL DEFAULT false,
-            token_limit            INTEGER,
-            limited_request_mode   BOOLEAN NOT NULL DEFAULT false,
-            request_limit          INTEGER,
-            shared                 BOOLEAN NOT NULL DEFAULT false,
-            incoming_webhook_token CITEXT UNIQUE,
-            created_at             TIMESTAMPTZ NOT NULL DEFAULT now(),
-            updated_at             TIMESTAMPTZ NOT NULL DEFAULT now(),
-            last_used_at           TIMESTAMPTZ,
-            legacy_mongo_id        TEXT
+            id                           UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+            user_id                      TEXT NOT NULL,
+            name                         TEXT NOT NULL,
+            description                  TEXT,
+            agent_type                   TEXT,
+            status                       TEXT NOT NULL,
+            key                          CITEXT UNIQUE,
+            image                        TEXT,
+            source_id                    UUID REFERENCES sources(id) ON DELETE SET NULL,
+            extra_source_ids             UUID[] NOT NULL DEFAULT '{}',
+            chunks                       INTEGER,
+            retriever                    TEXT,
+            prompt_id                    UUID REFERENCES prompts(id) ON DELETE SET NULL,
+            tools                        JSONB NOT NULL DEFAULT '[]'::jsonb,
+            json_schema                  JSONB,
+            models                       JSONB,
+            default_model_id             TEXT,
+            folder_id                    UUID REFERENCES agent_folders(id) ON DELETE SET NULL,
+            workflow_id                  UUID,
+            limited_token_mode           BOOLEAN NOT NULL DEFAULT false,
+            token_limit                  INTEGER,
+            limited_request_mode         BOOLEAN NOT NULL DEFAULT false,
+            request_limit                INTEGER,
+            allow_system_prompt_override BOOLEAN NOT NULL DEFAULT false,
+            shared                       BOOLEAN NOT NULL DEFAULT false,
+            shared_token                 CITEXT UNIQUE,
+            shared_metadata              JSONB,
+            incoming_webhook_token       CITEXT UNIQUE,
+            created_at                   TIMESTAMPTZ NOT NULL DEFAULT now(),
+            updated_at                   TIMESTAMPTZ NOT NULL DEFAULT now(),
+            last_used_at                 TIMESTAMPTZ,
+            legacy_mongo_id              TEXT
         );
         """
     )
@@ -299,6 +325,11 @@ def upgrade() -> None:
             upload_path     TEXT NOT NULL,
             mime_type       TEXT,
             size            BIGINT,
+            content         TEXT,
+            token_count     INTEGER,
+            openai_file_id  TEXT,
+            google_file_uri TEXT,
+            metadata        JSONB,
             created_at      TIMESTAMPTZ NOT NULL DEFAULT now(),
             legacy_mongo_id TEXT
         );
@@ -313,6 +344,7 @@ def upgrade() -> None:
             tool_id    UUID REFERENCES user_tools(id) ON DELETE CASCADE,
             path       TEXT NOT NULL,
             content    TEXT NOT NULL,
+            created_at TIMESTAMPTZ NOT NULL DEFAULT now(),
             updated_at TIMESTAMPTZ NOT NULL DEFAULT now()
         );
         """
@@ -321,13 +353,16 @@ def upgrade() -> None:
     op.execute(
         """
         CREATE TABLE todos (
-            id         UUID PRIMARY KEY DEFAULT gen_random_uuid(),
-            user_id    TEXT NOT NULL,
-            tool_id    UUID REFERENCES user_tools(id) ON DELETE CASCADE,
-            title      TEXT NOT NULL,
-            completed  BOOLEAN NOT NULL DEFAULT false,
-            created_at TIMESTAMPTZ NOT NULL DEFAULT now(),
-            updated_at TIMESTAMPTZ NOT NULL DEFAULT now()
+            id              UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+            user_id         TEXT NOT NULL,
+            tool_id         UUID REFERENCES user_tools(id) ON DELETE CASCADE,
+            todo_id         INTEGER,
+            title           TEXT NOT NULL,
+            completed       BOOLEAN NOT NULL DEFAULT false,
+            metadata        JSONB NOT NULL DEFAULT '{}'::jsonb,
+            created_at      TIMESTAMPTZ NOT NULL DEFAULT now(),
+            updated_at      TIMESTAMPTZ NOT NULL DEFAULT now(),
+            legacy_mongo_id TEXT
         );
         """
     )
@@ -335,13 +370,15 @@ def upgrade() -> None:
     op.execute(
         """
         CREATE TABLE notes (
-            id         UUID PRIMARY KEY DEFAULT gen_random_uuid(),
-            user_id    TEXT NOT NULL,
-            tool_id    UUID REFERENCES user_tools(id) ON DELETE CASCADE,
-            title      TEXT NOT NULL,
-            content    TEXT NOT NULL,
-            created_at TIMESTAMPTZ NOT NULL DEFAULT now(),
-            updated_at TIMESTAMPTZ NOT NULL DEFAULT now()
+            id              UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+            user_id         TEXT NOT NULL,
+            tool_id         UUID REFERENCES user_tools(id) ON DELETE CASCADE,
+            title           TEXT NOT NULL,
+            content         TEXT NOT NULL,
+            metadata        JSONB NOT NULL DEFAULT '{}'::jsonb,
+            created_at      TIMESTAMPTZ NOT NULL DEFAULT now(),
+            updated_at      TIMESTAMPTZ NOT NULL DEFAULT now(),
+            legacy_mongo_id TEXT
         );
         """
     )
@@ -349,12 +386,18 @@ def upgrade() -> None:
     op.execute(
         """
         CREATE TABLE connector_sessions (
-            id           UUID PRIMARY KEY DEFAULT gen_random_uuid(),
-            user_id      TEXT NOT NULL,
-            provider     TEXT NOT NULL,
-            session_data JSONB NOT NULL,
-            expires_at   TIMESTAMPTZ,
-            created_at   TIMESTAMPTZ NOT NULL DEFAULT now()
+            id              UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+            user_id         TEXT NOT NULL,
+            provider        TEXT NOT NULL,
+            server_url      TEXT,
+            session_token   TEXT UNIQUE,
+            user_email      TEXT,
+            status          TEXT,
+            token_info      JSONB,
+            session_data    JSONB NOT NULL DEFAULT '{}'::jsonb,
+            expires_at      TIMESTAMPTZ,
+            created_at      TIMESTAMPTZ NOT NULL DEFAULT now(),
+            legacy_mongo_id TEXT
         );
         """
     )
@@ -454,6 +497,14 @@ def upgrade() -> None:
         );
         """
     )
+    # Backfill the agents.workflow_id FK now that workflows exists.
+    # The column was created without a FK (forward reference to a table
+    # that hadn't been declared yet); add the constraint here so workflow
+    # deletion still cascades through to agent unset.
+    op.execute(
+        "ALTER TABLE agents ADD CONSTRAINT agents_workflow_fk "
+        "FOREIGN KEY (workflow_id) REFERENCES workflows(id) ON DELETE SET NULL;"
+    )
 
     op.execute(
         """
@@ -539,13 +590,26 @@ def upgrade() -> None:
     )
 
     op.execute(
-        "CREATE UNIQUE INDEX connector_sessions_user_provider_uidx "
-        "ON connector_sessions (user_id, provider);"
+        # MCP and OAuth connectors share the ``provider`` slot, so the
+        # dedup key is ``(user_id, server_url, provider)``: MCP rows
+        # differentiate by server_url (one per MCP server), OAuth rows
+        # have server_url = NULL and differentiate by provider alone.
+        # COALESCE lets NULL server_url participate in the constraint.
+        "CREATE UNIQUE INDEX connector_sessions_user_endpoint_uidx "
+        "ON connector_sessions (user_id, COALESCE(server_url, ''), provider);"
     )
     op.execute(
         "CREATE INDEX connector_sessions_expiry_idx "
         "ON connector_sessions (expires_at) WHERE expires_at IS NOT NULL;"
     )
+    op.execute(
+        "CREATE INDEX connector_sessions_server_url_idx "
+        "ON connector_sessions (server_url) WHERE server_url IS NOT NULL;"
+    )
+    op.execute(
+        "CREATE UNIQUE INDEX connector_sessions_legacy_mongo_id_uidx "
+        "ON connector_sessions (legacy_mongo_id) WHERE legacy_mongo_id IS NOT NULL;"
+    )
 
     op.execute(
         "CREATE UNIQUE INDEX conversation_messages_conv_pos_uidx "
@@ -587,6 +651,10 @@ def upgrade() -> None:
 
     op.execute("CREATE UNIQUE INDEX notes_user_tool_uidx ON notes (user_id, tool_id);")
     op.execute("CREATE INDEX notes_tool_id_idx ON notes (tool_id);")
+    op.execute(
+        "CREATE UNIQUE INDEX notes_legacy_mongo_id_uidx "
+        "ON notes (legacy_mongo_id) WHERE legacy_mongo_id IS NOT NULL;"
+    )
 
     op.execute(
         "CREATE UNIQUE INDEX pending_tool_state_conv_user_uidx "
@@ -616,20 +684,54 @@ def upgrade() -> None:
     )
 
     op.execute("CREATE INDEX sources_user_idx ON sources (user_id);")
+    op.execute(
+        "CREATE UNIQUE INDEX sources_legacy_mongo_id_uidx "
+        "ON sources (legacy_mongo_id) WHERE legacy_mongo_id IS NOT NULL;"
+    )
+    op.execute(
+        "CREATE UNIQUE INDEX user_tools_legacy_mongo_id_uidx "
+        "ON user_tools (legacy_mongo_id) WHERE legacy_mongo_id IS NOT NULL;"
+    )
+    op.execute(
+        "CREATE UNIQUE INDEX agent_folders_legacy_mongo_id_uidx "
+        "ON agent_folders (legacy_mongo_id) WHERE legacy_mongo_id IS NOT NULL;"
+    )
+    op.execute("CREATE INDEX agent_folders_parent_idx ON agent_folders (parent_id);")
+    op.execute("CREATE INDEX agents_workflow_idx ON agents (workflow_id);")
 
     op.execute('CREATE INDEX stack_logs_timestamp_idx ON stack_logs ("timestamp" DESC);')
     op.execute('CREATE INDEX stack_logs_user_ts_idx   ON stack_logs (user_id, "timestamp" DESC);')
     op.execute('CREATE INDEX stack_logs_level_ts_idx  ON stack_logs (level, "timestamp" DESC);')
     op.execute("CREATE INDEX stack_logs_activity_idx  ON stack_logs (activity_id);")
+    op.execute(
+        "CREATE UNIQUE INDEX stack_logs_mongo_id_uidx "
+        "ON stack_logs (mongo_id) WHERE mongo_id IS NOT NULL;"
+    )
 
     op.execute("CREATE INDEX todos_user_tool_idx ON todos (user_id, tool_id);")
     op.execute("CREATE INDEX todos_tool_id_idx   ON todos (tool_id);")
+    op.execute(
+        "CREATE UNIQUE INDEX todos_legacy_mongo_id_uidx "
+        "ON todos (legacy_mongo_id) WHERE legacy_mongo_id IS NOT NULL;"
+    )
+    op.execute(
+        "CREATE UNIQUE INDEX todos_tool_todo_id_uidx "
+        "ON todos (tool_id, todo_id) WHERE todo_id IS NOT NULL;"
+    )
 
     op.execute('CREATE INDEX token_usage_user_ts_idx  ON token_usage (user_id, "timestamp" DESC);')
     op.execute('CREATE INDEX token_usage_key_ts_idx   ON token_usage (api_key, "timestamp" DESC);')
     op.execute('CREATE INDEX token_usage_agent_ts_idx ON token_usage (agent_id, "timestamp" DESC);')
+    op.execute(
+        "CREATE UNIQUE INDEX token_usage_mongo_id_uidx "
+        "ON token_usage (mongo_id) WHERE mongo_id IS NOT NULL;"
+    )
 
     op.execute('CREATE INDEX user_logs_user_ts_idx ON user_logs (user_id, "timestamp" DESC);')
+    op.execute(
+        "CREATE UNIQUE INDEX user_logs_mongo_id_uidx "
+        "ON user_logs (mongo_id) WHERE mongo_id IS NOT NULL;"
+    )
 
     op.execute("CREATE INDEX user_tools_user_id_idx ON user_tools (user_id);")
 

application/alembic/versions/0002_app_metadata.py

@@ -0,0 +1,37 @@
+"""0002 app_metadata — singleton key/value table for instance-wide state.
+
+Used by the startup version-check client to persist the anonymous
+instance UUID and a one-shot "notice shown" flag. Both values are tiny
+plain-text strings; this is a deliberate generic-config table rather
+than dedicated columns so future one-off settings (telemetry opt-in
+timestamps, feature-flag overrides, etc.) don't each need their own
+migration.
+
+Revision ID: 0002_app_metadata
+Revises: 0001_initial
+"""
+
+from typing import Sequence, Union
+
+from alembic import op
+
+
+revision: str = "0002_app_metadata"
+down_revision: Union[str, None] = "0001_initial"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    op.execute(
+        """
+        CREATE TABLE app_metadata (
+            key   TEXT PRIMARY KEY,
+            value TEXT NOT NULL
+        );
+        """
+    )
+
+
+def downgrade() -> None:
+    op.execute("DROP TABLE IF EXISTS app_metadata;")

application/alembic/versions/0003_user_custom_models.py

@@ -0,0 +1,65 @@
+"""0003 user_custom_models — per-user OpenAI-compatible model registrations.
+
+Revision ID: 0003_user_custom_models
+Revises: 0002_app_metadata
+"""
+
+from typing import Sequence, Union
+
+from alembic import op
+
+
+revision: str = "0003_user_custom_models"
+down_revision: Union[str, None] = "0002_app_metadata"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    op.execute(
+        """
+        CREATE TABLE user_custom_models (
+            id                  UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+            user_id             TEXT NOT NULL,
+            upstream_model_id   TEXT NOT NULL,
+            display_name        TEXT NOT NULL,
+            description         TEXT NOT NULL DEFAULT '',
+            base_url            TEXT NOT NULL,
+            api_key_encrypted   TEXT NOT NULL,
+            capabilities        JSONB NOT NULL DEFAULT '{}'::jsonb,
+            enabled             BOOLEAN NOT NULL DEFAULT true,
+            created_at          TIMESTAMPTZ NOT NULL DEFAULT now(),
+            updated_at          TIMESTAMPTZ NOT NULL DEFAULT now()
+        );
+        """
+    )
+    op.execute(
+        "CREATE INDEX user_custom_models_user_id_idx "
+        "ON user_custom_models (user_id);"
+    )
+
+    # Mirror the project-wide invariants set up in 0001_initial:
+    #   * user_id FK with ON DELETE RESTRICT (deferrable),
+    #   * ensure_user_exists() trigger so the parent users row autocreates,
+    #   * set_updated_at() trigger.
+    op.execute(
+        "ALTER TABLE user_custom_models "
+        "ADD CONSTRAINT user_custom_models_user_id_fk "
+        "FOREIGN KEY (user_id) REFERENCES users(user_id) "
+        "ON DELETE RESTRICT DEFERRABLE INITIALLY IMMEDIATE;"
+    )
+    op.execute(
+        "CREATE TRIGGER user_custom_models_ensure_user "
+        "BEFORE INSERT OR UPDATE OF user_id ON user_custom_models "
+        "FOR EACH ROW EXECUTE FUNCTION ensure_user_exists();"
+    )
+    op.execute(
+        "CREATE TRIGGER user_custom_models_set_updated_at "
+        "BEFORE UPDATE ON user_custom_models "
+        "FOR EACH ROW WHEN (OLD.* IS DISTINCT FROM NEW.*) "
+        "EXECUTE FUNCTION set_updated_at();"
+    )
+
+
+def downgrade() -> None:
+    op.execute("DROP TABLE IF EXISTS user_custom_models;")

application/api/answer/routes/base.py

@@ -14,10 +14,13 @@ from application.core.model_utils import (
     get_provider_from_model_id,
 )
 
-from application.core.mongo_db import MongoDB
 from application.core.settings import settings
 from application.error import sanitize_api_error
 from application.llm.llm_creator import LLMCreator
+from application.storage.db.repositories.agents import AgentsRepository
+from application.storage.db.repositories.token_usage import TokenUsageRepository
+from application.storage.db.repositories.user_logs import UserLogsRepository
+from application.storage.db.session import db_readonly, db_session
 from application.utils import check_required_fields
 
 logger = logging.getLogger(__name__)
@@ -30,10 +33,6 @@ class BaseAnswerResource:
     """Shared base class for answer endpoints"""
 
     def __init__(self):
-        mongo = MongoDB.get_client()
-        db = mongo[settings.MONGO_DB_NAME]
-        self.db = db
-        self.user_logs_collection = db["user_logs"]
         self.default_model_id = get_default_model_id()
         self.conversation_service = ConversationService()
 
@@ -91,8 +90,8 @@ class BaseAnswerResource:
         api_key = agent_config.get("user_api_key")
         if not api_key:
             return None
-        agents_collection = self.db["agents"]
-        agent = agents_collection.find_one({"key": api_key})
+        with db_readonly() as conn:
+            agent = AgentsRepository(conn).find_by_key(api_key)
 
         if not agent:
             return make_response(
@@ -113,41 +112,32 @@ class BaseAnswerResource:
         )
 
         token_limit = int(
-            agent.get("token_limit", settings.DEFAULT_AGENT_LIMITS["token_limit"])
+            agent.get("token_limit") or settings.DEFAULT_AGENT_LIMITS["token_limit"]
         )
         request_limit = int(
-            agent.get("request_limit", settings.DEFAULT_AGENT_LIMITS["request_limit"])
+            agent.get("request_limit") or settings.DEFAULT_AGENT_LIMITS["request_limit"]
         )
 
-        token_usage_collection = self.db["token_usage"]
-
-        end_date = datetime.datetime.now()
+        end_date = datetime.datetime.now(datetime.timezone.utc)
         start_date = end_date - datetime.timedelta(hours=24)
 
-        match_query = {
-            "timestamp": {"$gte": start_date, "$lte": end_date},
-            "api_key": api_key,
-        }
-
-        if limited_token_mode:
-            token_pipeline = [
-                {"$match": match_query},
-                {
-                    "$group": {
-                        "_id": None,
-                        "total_tokens": {
-                            "$sum": {"$add": ["$prompt_tokens", "$generated_tokens"]}
-                        },
-                    }
-                },
-            ]
-            token_result = list(token_usage_collection.aggregate(token_pipeline))
-            daily_token_usage = token_result[0]["total_tokens"] if token_result else 0
+        if limited_token_mode or limited_request_mode:
+            with db_readonly() as conn:
+                token_repo = TokenUsageRepository(conn)
+                if limited_token_mode:
+                    daily_token_usage = token_repo.sum_tokens_in_range(
+                        start=start_date, end=end_date, api_key=api_key,
+                    )
+                else:
+                    daily_token_usage = 0
+                if limited_request_mode:
+                    daily_request_usage = token_repo.count_in_range(
+                        start=start_date, end=end_date, api_key=api_key,
+                    )
+                else:
+                    daily_request_usage = 0
         else:
             daily_token_usage = 0
-        if limited_request_mode:
-            daily_request_usage = token_usage_collection.count_documents(match_query)
-        else:
             daily_request_usage = 0
         if not limited_token_mode and not limited_request_mode:
             return None
@@ -187,6 +177,7 @@ class BaseAnswerResource:
         is_shared_usage: bool = False,
         shared_token: Optional[str] = None,
         model_id: Optional[str] = None,
+        model_user_id: Optional[str] = None,
         _continuation: Optional[Dict] = None,
     ) -> Generator[str, None, None]:
         """
@@ -299,8 +290,18 @@ class BaseAnswerResource:
                     # conversation if this is the first turn.
                     if not conversation_id and should_save_conversation:
                         try:
+                            # Use model-owner scope so shared-agent
+                            # owner-BYOM resolves to its registered plugin.
                             provider = (
-                                get_provider_from_model_id(model_id)
+                                get_provider_from_model_id(
+                                    model_id,
+                                    user_id=model_user_id
+                                    or (
+                                        decoded_token.get("sub")
+                                        if decoded_token
+                                        else None
+                                    ),
+                                )
                                 if model_id
                                 else settings.LLM_PROVIDER
                             )
@@ -314,6 +315,7 @@ class BaseAnswerResource:
                                 decoded_token=decoded_token,
                                 model_id=model_id,
                                 agent_id=agent_id,
+                                model_user_id=model_user_id,
                             )
                             conversation_id = (
                                 self.conversation_service.save_conversation(
@@ -350,6 +352,9 @@ class BaseAnswerResource:
                                 tool_schemas=getattr(agent, "tools", []),
                                 agent_config={
                                     "model_id": model_id or self.default_model_id,
+                                    # Persist BYOM scope so resume doesn't
+                                    # fall back to caller's layer.
+                                    "model_user_id": model_user_id,
                                     "llm_name": getattr(agent, "llm_name", settings.LLM_PROVIDER),
                                     "api_key": getattr(agent, "api_key", None),
                                     "user_api_key": user_api_key,
@@ -380,8 +385,14 @@ class BaseAnswerResource:
             if isNoneDoc:
                 for doc in source_log_docs:
                     doc["source"] = "None"
+            # Run under model-owner scope so title-gen LLM inside
+            # save_conversation uses the owner's BYOM provider/key.
             provider = (
-                get_provider_from_model_id(model_id)
+                get_provider_from_model_id(
+                    model_id,
+                    user_id=model_user_id
+                    or (decoded_token.get("sub") if decoded_token else None),
+                )
                 if model_id
                 else settings.LLM_PROVIDER
             )
@@ -394,6 +405,7 @@ class BaseAnswerResource:
                 decoded_token=decoded_token,
                 model_id=model_id,
                 agent_id=agent_id,
+                model_user_id=model_user_id,
             )
 
             if should_save_conversation:
@@ -467,19 +479,18 @@ class BaseAnswerResource:
             for key, value in log_data.items():
                 if isinstance(value, str) and len(value) > 10000:
                     log_data[key] = value[:10000]
-            self.user_logs_collection.insert_one(log_data)
-
-            from application.storage.db.dual_write import dual_write
-            from application.storage.db.repositories.user_logs import UserLogsRepository
-
-            dual_write(
-                UserLogsRepository,
-                lambda repo, d=log_data: repo.insert(
-                    user_id=d.get("user"),
-                    endpoint="stream_answer",
-                    data=d,
-                ),
-            )
+            try:
+                with db_session() as conn:
+                    UserLogsRepository(conn).insert(
+                        user_id=log_data.get("user"),
+                        endpoint="stream_answer",
+                        data=log_data,
+                    )
+            except Exception as log_err:
+                logger.error(
+                    f"Failed to persist stream_answer user log: {log_err}",
+                    exc_info=True,
+                )
 
             data = json.dumps({"type": "end"})
             yield f"data: {data}\n\n"
@@ -492,12 +503,34 @@ class BaseAnswerResource:
                     if isNoneDoc:
                         for doc in source_log_docs:
                             doc["source"] = "None"
+                    # Mirror the normal-path provider resolution so the
+                    # partial-save title LLM uses the model-owner's BYOM
+                    # registration (shared-agent dispatch) rather than
+                    # the deployment default with the instance api key.
+                    provider = (
+                        get_provider_from_model_id(
+                            model_id,
+                            user_id=model_user_id
+                            or (
+                                decoded_token.get("sub")
+                                if decoded_token
+                                else None
+                            ),
+                        )
+                        if model_id
+                        else settings.LLM_PROVIDER
+                    )
+                    sys_api_key = get_api_key_for_provider(
+                        provider or settings.LLM_PROVIDER
+                    )
                     llm = LLMCreator.create_llm(
-                        settings.LLM_PROVIDER,
-                        api_key=settings.API_KEY,
+                        provider or settings.LLM_PROVIDER,
+                        api_key=sys_api_key,
                         user_api_key=user_api_key,
                         decoded_token=decoded_token,
+                        model_id=model_id,
                         agent_id=agent_id,
+                        model_user_id=model_user_id,
                     )
                     self.conversation_service.save_conversation(
                         conversation_id,

application/api/answer/routes/search.py

@@ -1,28 +1,21 @@
 import logging
-from typing import Any, Dict, List
 
 from flask import make_response, request
 from flask_restx import fields, Resource
 
-from bson.dbref import DBRef
-
 from application.api.answer.routes.base import answer_ns
-from application.core.mongo_db import MongoDB
-from application.core.settings import settings
-from application.vectorstore.vector_creator import VectorCreator
+from application.services.search_service import (
+    InvalidAPIKey,
+    SearchFailed,
+    search,
+)
 
 logger = logging.getLogger(__name__)
 
 
 @answer_ns.route("/api/search")
 class SearchResource(Resource):
-    """Fast search endpoint for retrieving relevant documents"""
-
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        mongo = MongoDB.get_client()
-        self.db = mongo[settings.MONGO_DB_NAME]
-        self.agents_collection = self.db["agents"]
+    """Fast search endpoint for retrieving relevant documents."""
 
     search_model = answer_ns.model(
         "SearchModel",
@@ -39,116 +32,10 @@ class SearchResource(Resource):
         },
     )
 
-    def _get_sources_from_api_key(self, api_key: str) -> List[str]:
-        """Get source IDs connected to the API key/agent.
-
-        """
-        agent_data = self.agents_collection.find_one({"key": api_key})
-        if not agent_data:
-            return []
-
-        source_ids = []
-
-        # Handle multiple sources (only if non-empty)
-        sources = agent_data.get("sources", [])
-        if sources and isinstance(sources, list) and len(sources) > 0:
-            for source_ref in sources:
-                # Skip "default" - it's a placeholder, not an actual vectorstore
-                if source_ref == "default":
-                    continue
-                elif isinstance(source_ref, DBRef):
-                    source_doc = self.db.dereference(source_ref)
-                    if source_doc:
-                        source_ids.append(str(source_doc["_id"]))
-
-        # Handle single source (legacy) - check if sources was empty or didn't yield results
-        if not source_ids:
-            source = agent_data.get("source")
-            if isinstance(source, DBRef):
-                source_doc = self.db.dereference(source)
-                if source_doc:
-                    source_ids.append(str(source_doc["_id"]))
-            # Skip "default" - it's a placeholder, not an actual vectorstore
-            elif source and source != "default":
-                source_ids.append(source)
-
-        return source_ids
-
-    def _search_vectorstores(
-        self, query: str, source_ids: List[str], chunks: int
-    ) -> List[Dict[str, Any]]:
-        """Search across vectorstores and return results"""
-        if not source_ids:
-            return []
-
-        results = []
-        chunks_per_source = max(1, chunks // len(source_ids))
-        seen_texts = set()
-
-        for source_id in source_ids:
-            if not source_id or not source_id.strip():
-                continue
-
-            try:
-                docsearch = VectorCreator.create_vectorstore(
-                    settings.VECTOR_STORE, source_id, settings.EMBEDDINGS_KEY
-                )
-                docs = docsearch.search(query, k=chunks_per_source * 2)
-
-                for doc in docs:
-                    if len(results) >= chunks:
-                        break
-
-                    if hasattr(doc, "page_content") and hasattr(doc, "metadata"):
-                        page_content = doc.page_content
-                        metadata = doc.metadata
-                    else:
-                        page_content = doc.get("text", doc.get("page_content", ""))
-                        metadata = doc.get("metadata", {})
-
-                    # Skip duplicates
-                    text_hash = hash(page_content[:200])
-                    if text_hash in seen_texts:
-                        continue
-                    seen_texts.add(text_hash)
-
-                    title = metadata.get(
-                        "title", metadata.get("post_title", "")
-                    )
-                    if not isinstance(title, str):
-                        title = str(title) if title else ""
-
-                    # Clean up title
-                    if title:
-                        title = title.split("/")[-1]
-                    else:
-                        # Use filename or first part of content as title
-                        title = metadata.get("filename", page_content[:50] + "...")
-
-                    source = metadata.get("source", source_id)
-
-                    results.append({
-                        "text": page_content,
-                        "title": title,
-                        "source": source,
-                    })
-
-                if len(results) >= chunks:
-                    break
-
-            except Exception as e:
-                logger.error(
-                    f"Error searching vectorstore {source_id}: {e}",
-                    exc_info=True,
-                )
-                continue
-
-        return results[:chunks]
-
     @answer_ns.expect(search_model)
     @answer_ns.doc(description="Search for relevant documents based on query")
     def post(self):
-        data = request.get_json()
+        data = request.get_json() or {}
 
         question = data.get("question")
         api_key = data.get("api_key")
@@ -156,31 +43,13 @@ class SearchResource(Resource):
 
         if not question:
             return make_response({"error": "question is required"}, 400)
-
         if not api_key:
             return make_response({"error": "api_key is required"}, 400)
 
-        # Validate API key
-        agent = self.agents_collection.find_one({"key": api_key})
-        if not agent:
-            return make_response({"error": "Invalid API key"}, 401)
-
         try:
-            # Get sources connected to this API key
-            source_ids = self._get_sources_from_api_key(api_key)
-
-            if not source_ids:
-                return make_response([], 200)
-
-            # Perform search
-            results = self._search_vectorstores(question, source_ids, chunks)
-
-            return make_response(results, 200)
-
-        except Exception as e:
-            logger.error(
-                f"/api/search - error: {str(e)}",
-                extra={"error": str(e)},
-                exc_info=True,
-            )
+            return make_response(search(api_key, question, chunks), 200)
+        except InvalidAPIKey:
+            return make_response({"error": "Invalid API key"}, 401)
+        except SearchFailed:
+            logger.exception("/api/search failed")
             return make_response({"error": "Search failed"}, 500)

application/api/answer/routes/stream.py

@@ -109,6 +109,7 @@ class StreamResource(Resource, BaseAnswerResource):
                         decoded_token=processor.decoded_token,
                         agent_id=processor.agent_id,
                         model_id=processor.model_id,
+                        model_user_id=processor.model_user_id,
                         _continuation={
                             "messages": messages,
                             "tools_dict": tools_dict,
@@ -145,6 +146,7 @@ class StreamResource(Resource, BaseAnswerResource):
                     is_shared_usage=processor.is_shared_usage,
                     shared_token=processor.shared_token,
                     model_id=processor.model_id,
+                    model_user_id=processor.model_user_id,
                 ),
                 mimetype="text/event-stream",
             )

application/api/answer/services/compression/orchestrator.py

@@ -49,6 +49,7 @@ class CompressionOrchestrator:
         model_id: str,
         decoded_token: Dict[str, Any],
         current_query_tokens: int = 500,
+        model_user_id: Optional[str] = None,
     ) -> CompressionResult:
         """
         Check if compression is needed and perform it if so.
@@ -57,16 +58,18 @@ class CompressionOrchestrator:
 
         Args:
             conversation_id: Conversation ID
-            user_id: User ID
+            user_id: Caller's user id — used for conversation access checks
             model_id: Model being used for conversation
             decoded_token: User's decoded JWT token
             current_query_tokens: Estimated tokens for current query
+            model_user_id: BYOM-resolution scope (model owner); defaults
+                to ``user_id`` for built-in / caller-owned models.
 
         Returns:
             CompressionResult with summary and recent queries
         """
         try:
-            # Load conversation
+            # Conversation row is owned by the caller, not the model owner.
             conversation = self.conversation_service.get_conversation(
                 conversation_id, user_id
             )
@@ -77,9 +80,14 @@ class CompressionOrchestrator:
                 )
                 return CompressionResult.failure("Conversation not found")
 
-            # Check if compression is needed
+            # Use model-owner scope so per-user BYOM context windows
+            # (e.g. 8k) compute the threshold against the right limit.
+            registry_user_id = model_user_id or user_id
             if not self.threshold_checker.should_compress(
-                conversation, model_id, current_query_tokens
+                conversation,
+                model_id,
+                current_query_tokens,
+                user_id=registry_user_id,
             ):
                 # No compression needed, return full history
                 queries = conversation.get("queries", [])
@@ -87,7 +95,12 @@ class CompressionOrchestrator:
 
             # Perform compression
             return self._perform_compression(
-                conversation_id, conversation, model_id, decoded_token
+                conversation_id,
+                conversation,
+                model_id,
+                decoded_token,
+                user_id=user_id,
+                model_user_id=model_user_id,
             )
 
         except Exception as e:
@@ -102,6 +115,8 @@ class CompressionOrchestrator:
         conversation: Dict[str, Any],
         model_id: str,
         decoded_token: Dict[str, Any],
+        user_id: Optional[str] = None,
+        model_user_id: Optional[str] = None,
     ) -> CompressionResult:
         """
         Perform the actual compression operation.
@@ -111,6 +126,8 @@ class CompressionOrchestrator:
             conversation: Conversation document
             model_id: Model ID for conversation
             decoded_token: User token
+            user_id: Caller's id (for conversation reload after compression)
+            model_user_id: BYOM-resolution scope (model owner)
 
         Returns:
             CompressionResult
@@ -123,11 +140,17 @@ class CompressionOrchestrator:
                 else model_id
             )
 
-            # Get provider and API key for compression model
-            provider = get_provider_from_model_id(compression_model)
+            # Use model-owner scope so provider/api_key resolves to the
+            # owner's BYOM record (shared-agent dispatch).
+            caller_user_id = user_id
+            if caller_user_id is None and isinstance(decoded_token, dict):
+                caller_user_id = decoded_token.get("sub")
+            registry_user_id = model_user_id or caller_user_id
+            provider = get_provider_from_model_id(
+                compression_model, user_id=registry_user_id
+            )
             api_key = get_api_key_for_provider(provider)
 
-            # Create compression LLM
             compression_llm = LLMCreator.create_llm(
                 provider,
                 api_key=api_key,
@@ -135,6 +158,7 @@ class CompressionOrchestrator:
                 decoded_token=decoded_token,
                 model_id=compression_model,
                 agent_id=conversation.get("agent_id"),
+                model_user_id=registry_user_id,
             )
 
             # Create compression service with DB update capability
@@ -167,9 +191,12 @@ class CompressionOrchestrator:
                 f"saved {metadata.original_token_count - metadata.compressed_token_count} tokens"
             )
 
-            # Reload conversation with updated metadata
+            # Reload under caller (conversation is owned by caller).
+            reload_user_id = caller_user_id
+            if reload_user_id is None and isinstance(decoded_token, dict):
+                reload_user_id = decoded_token.get("sub")
             conversation = self.conversation_service.get_conversation(
-                conversation_id, user_id=decoded_token.get("sub")
+                conversation_id, user_id=reload_user_id
             )
 
             # Get compressed context
@@ -192,16 +219,21 @@ class CompressionOrchestrator:
         model_id: str,
         decoded_token: Dict[str, Any],
         current_conversation: Optional[Dict[str, Any]] = None,
+        model_user_id: Optional[str] = None,
     ) -> CompressionResult:
         """
         Perform compression during tool execution.
 
         Args:
             conversation_id: Conversation ID
-            user_id: User ID
+            user_id: Caller's user id — used for conversation access checks
             model_id: Model ID
             decoded_token: User token
             current_conversation: Pre-loaded conversation (optional)
+            model_user_id: BYOM-resolution scope (model owner). For
+                shared-agent dispatch this is the agent owner; defaults
+                to ``user_id`` so built-in / caller-owned models are
+                unaffected.
 
         Returns:
             CompressionResult
@@ -223,7 +255,12 @@ class CompressionOrchestrator:
 
             # Perform compression
             return self._perform_compression(
-                conversation_id, conversation, model_id, decoded_token
+                conversation_id,
+                conversation,
+                model_id,
+                decoded_token,
+                user_id=user_id,
+                model_user_id=model_user_id,
             )
 
         except Exception as e:

application/api/answer/services/compression/service.py

@@ -106,8 +106,13 @@ class CompressionService:
                 f"using model {self.model_id}"
             )
 
+            # See note in conversation_service.py: ``self.model_id`` is
+            # the registry id (UUID for BYOM); the LLM's own model_id is
+            # what the provider's API actually expects.
             response = self.llm.gen(
-                model=self.model_id, messages=messages, max_tokens=4000
+                model=getattr(self.llm, "model_id", None) or self.model_id,
+                messages=messages,
+                max_tokens=4000,
             )
 
             # Extract summary from response

application/api/answer/services/compression/threshold_checker.py

@@ -30,6 +30,7 @@ class CompressionThresholdChecker:
         conversation: Dict[str, Any],
         model_id: str,
         current_query_tokens: int = 500,
+        user_id: str | None = None,
     ) -> bool:
         """
         Determine if compression is needed.
@@ -38,6 +39,8 @@ class CompressionThresholdChecker:
             conversation: Full conversation document
             model_id: Target model for this request
             current_query_tokens: Estimated tokens for current query
+            user_id: Owner — needed so per-user BYOM custom-model UUIDs
+                resolve when looking up the context window.
 
         Returns:
             True if tokens >= threshold% of context window
@@ -48,7 +51,7 @@ class CompressionThresholdChecker:
             total_tokens += current_query_tokens
 
             # Get context window limit for model
-            context_limit = get_token_limit(model_id)
+            context_limit = get_token_limit(model_id, user_id=user_id)
 
             # Calculate threshold
             threshold = int(context_limit * self.threshold_percentage)
@@ -73,20 +76,24 @@ class CompressionThresholdChecker:
             logger.error(f"Error checking compression need: {str(e)}", exc_info=True)
             return False
 
-    def check_message_tokens(self, messages: list, model_id: str) -> bool:
+    def check_message_tokens(
+        self, messages: list, model_id: str, user_id: str | None = None
+    ) -> bool:
         """
         Check if message list exceeds threshold.
 
         Args:
             messages: List of message dicts
             model_id: Target model
+            user_id: Owner — needed so per-user BYOM custom-model UUIDs
+                resolve when looking up the context window.
 
         Returns:
             True if at or above threshold
         """
         try:
             current_tokens = TokenCounter.count_message_tokens(messages)
-            context_limit = get_token_limit(model_id)
+            context_limit = get_token_limit(model_id, user_id=user_id)
             threshold = int(context_limit * self.threshold_percentage)
 
             if current_tokens >= threshold:

application/api/answer/services/continuation_service.py

@@ -1,23 +1,20 @@
 """Service for saving and restoring tool-call continuation state.
 
 When a stream pauses (tool needs approval or client-side execution),
-the full execution state is persisted to MongoDB so the client can
+the full execution state is persisted to Postgres so the client can
 resume later by sending tool_actions.
 """
 
-import datetime
 import logging
 from typing import Any, Dict, List, Optional
+from uuid import UUID
 
-from bson import ObjectId
-
-from application.core.mongo_db import MongoDB
-from application.core.settings import settings
-from application.storage.db.dual_write import dual_write
+from application.storage.db.base_repository import looks_like_uuid
 from application.storage.db.repositories.conversations import ConversationsRepository
 from application.storage.db.repositories.pending_tool_state import (
     PendingToolStateRepository,
 )
+from application.storage.db.session import db_readonly, db_session
 
 logger = logging.getLogger(__name__)
 
@@ -26,8 +23,13 @@ PENDING_STATE_TTL_SECONDS = 30 * 60  # 30 minutes
 
 
 def _make_serializable(obj: Any) -> Any:
-    """Recursively convert MongoDB ObjectIds and other non-JSON types."""
-    if isinstance(obj, ObjectId):
+    """Recursively coerce non-JSON values into JSON-safe forms.
+
+    Handles ``uuid.UUID`` (from PG columns), ``bytes``, and recurses into
+    dicts/lists. Post-Mongo-cutover the ObjectId branch is gone — none of
+    our writers produce them anymore.
+    """
+    if isinstance(obj, UUID):
         return str(obj)
     if isinstance(obj, dict):
         return {str(k): _make_serializable(v) for k, v in obj.items()}
@@ -39,25 +41,13 @@ def _make_serializable(obj: Any) -> Any:
 
 
 class ContinuationService:
-    """Manages pending tool-call state in MongoDB."""
+    """Manages pending tool-call state in Postgres."""
 
     def __init__(self):
-        mongo = MongoDB.get_client()
-        db = mongo[settings.MONGO_DB_NAME]
-        self.collection = db["pending_tool_state"]
-        self._ensure_indexes()
-
-    def _ensure_indexes(self):
-        try:
-            self.collection.create_index(
-                "expires_at", expireAfterSeconds=0
-            )
-            self.collection.create_index(
-                [("conversation_id", 1), ("user", 1)], unique=True
-            )
-        except Exception:
-            # Indexes may already exist or mongomock doesn't support TTL
-            pass
+        # No-op constructor retained for call-site compatibility. State
+        # lives in Postgres now; each operation opens its own short-lived
+        # session rather than holding a connection on the service.
+        pass
 
     def save_state(
         self,
@@ -72,6 +62,10 @@ class ContinuationService:
     ) -> str:
         """Save execution state for later continuation.
 
+        ``conversation_id`` may be a Postgres UUID or the legacy Mongo
+        ``ObjectId`` string — the latter is resolved via
+        ``conversations.legacy_mongo_id`` to find the matching row.
+
         Args:
             conversation_id: The conversation this state belongs to.
             user: Owner user ID.
@@ -83,45 +77,26 @@ class ContinuationService:
             client_tools: Client-provided tool schemas for client-side execution.
 
         Returns:
-            The string ID of the saved state document.
+            The string ID (conversation_id as provided) of the saved state.
         """
-        now = datetime.datetime.now(datetime.timezone.utc)
-        expires_at = now + datetime.timedelta(seconds=PENDING_STATE_TTL_SECONDS)
-
-        doc = {
-            "conversation_id": conversation_id,
-            "user": user,
-            "messages": _make_serializable(messages),
-            "pending_tool_calls": _make_serializable(pending_tool_calls),
-            "tools_dict": _make_serializable(tools_dict),
-            "tool_schemas": _make_serializable(tool_schemas),
-            "agent_config": _make_serializable(agent_config),
-            "client_tools": _make_serializable(client_tools) if client_tools else None,
-            "created_at": now,
-            "expires_at": expires_at,
-        }
-
-        # Upsert — only one pending state per conversation per user
-        result = self.collection.replace_one(
-            {"conversation_id": conversation_id, "user": user},
-            doc,
-            upsert=True,
-        )
-        state_id = str(result.upserted_id) if result.upserted_id else conversation_id
-        logger.info(
-            f"Saved continuation state for conversation {conversation_id} "
-            f"with {len(pending_tool_calls)} pending tool call(s)"
-        )
-
-        # Dual-write to Postgres — upsert against the same Mongo conversation
-        # by resolving its UUID via conversations.legacy_mongo_id.
-        def _pg_save(_: PendingToolStateRepository) -> None:
-            conn = _._conn  # reuse the existing transaction
+        with db_session() as conn:
             conv = ConversationsRepository(conn).get_by_legacy_id(conversation_id)
-            if conv is None:
-                return
-            _.save_state(
-                conv["id"],
+            if conv is not None:
+                pg_conv_id = conv["id"]
+            elif looks_like_uuid(conversation_id):
+                pg_conv_id = conversation_id
+            else:
+                # Unresolvable legacy ObjectId — downstream ``CAST AS uuid``
+                # would raise and poison the save. Surface the mismatch so
+                # the caller can decide (the stream loop in routes/base.py
+                # already wraps this in try/except).
+                raise ValueError(
+                    f"Cannot save continuation state: conversation_id "
+                    f"{conversation_id!r} is neither a PG UUID nor a "
+                    f"backfilled legacy Mongo id."
+                )
+            PendingToolStateRepository(conn).save_state(
+                pg_conv_id,
                 user,
                 messages=_make_serializable(messages),
                 pending_tool_calls=_make_serializable(pending_tool_calls),
@@ -131,8 +106,11 @@ class ContinuationService:
                 client_tools=_make_serializable(client_tools) if client_tools else None,
             )
 
-        dual_write(PendingToolStateRepository, _pg_save)
-        return state_id
+        logger.info(
+            f"Saved continuation state for conversation {conversation_id} "
+            f"with {len(pending_tool_calls)} pending tool call(s)"
+        )
+        return conversation_id
 
     def load_state(
         self, conversation_id: str, user: str
@@ -142,34 +120,38 @@ class ContinuationService:
         Returns:
             The state dict, or None if no pending state exists.
         """
-        doc = self.collection.find_one(
-            {"conversation_id": conversation_id, "user": user}
-        )
+        with db_readonly() as conn:
+            conv = ConversationsRepository(conn).get_by_legacy_id(conversation_id)
+            if conv is not None:
+                pg_conv_id = conv["id"]
+            elif looks_like_uuid(conversation_id):
+                pg_conv_id = conversation_id
+            else:
+                # Unresolvable legacy ObjectId → no state can exist for it.
+                return None
+            doc = PendingToolStateRepository(conn).load_state(pg_conv_id, user)
         if not doc:
             return None
-        doc["_id"] = str(doc["_id"])
         return doc
 
     def delete_state(self, conversation_id: str, user: str) -> bool:
         """Delete pending state after successful resumption.
 
         Returns:
-            True if a document was deleted.
+            True if a row was deleted.
         """
-        result = self.collection.delete_one(
-            {"conversation_id": conversation_id, "user": user}
-        )
-        if result.deleted_count:
+        with db_session() as conn:
+            conv = ConversationsRepository(conn).get_by_legacy_id(conversation_id)
+            if conv is not None:
+                pg_conv_id = conv["id"]
+            elif looks_like_uuid(conversation_id):
+                pg_conv_id = conversation_id
+            else:
+                # Unresolvable legacy ObjectId → nothing to delete.
+                return False
+            deleted = PendingToolStateRepository(conn).delete_state(pg_conv_id, user)
+        if deleted:
             logger.info(
                 f"Deleted continuation state for conversation {conversation_id}"
             )
-
-        # Dual-write to Postgres — delete the same row.
-        def _pg_delete(repo: PendingToolStateRepository) -> None:
-            conv = ConversationsRepository(repo._conn).get_by_legacy_id(conversation_id)
-            if conv is None:
-                return
-            repo.delete_state(conv["id"], user)
-
-        dual_write(PendingToolStateRepository, _pg_delete)
-        return result.deleted_count > 0
+        return deleted

application/api/answer/services/conversation_service.py

@@ -1,46 +1,51 @@
+"""Conversation persistence service backed by Postgres.
+
+Handles create / append / update / compression for conversations during
+the answer-streaming path. Connections are opened per-operation rather
+than held for the duration of a stream.
+"""
+
 import logging
 from datetime import datetime, timezone
 from typing import Any, Dict, List, Optional
 
-from application.core.mongo_db import MongoDB
+from sqlalchemy import text as sql_text
 
 from application.core.settings import settings
-from application.storage.db.dual_write import dual_write
+from application.storage.db.base_repository import looks_like_uuid
+from application.storage.db.repositories.agents import AgentsRepository
 from application.storage.db.repositories.conversations import ConversationsRepository
-from bson import ObjectId
+from application.storage.db.session import db_readonly, db_session
 
 
 logger = logging.getLogger(__name__)
 
 
 class ConversationService:
-    def __init__(self):
-        mongo = MongoDB.get_client()
-        db = mongo[settings.MONGO_DB_NAME]
-        self.conversations_collection = db["conversations"]
-        self.agents_collection = db["agents"]
-
     def get_conversation(
         self, conversation_id: str, user_id: str
     ) -> Optional[Dict[str, Any]]:
-        """Retrieve a conversation with proper access control"""
+        """Retrieve a conversation with owner-or-shared access control.
+
+        Returns a dict in the legacy Mongo shape — ``queries`` is a list
+        of message dicts (prompt/response/...) — for compatibility with
+        the streaming pipeline that consumes this shape.
+        """
         if not conversation_id or not user_id:
             return None
         try:
-            conversation = self.conversations_collection.find_one(
-                {
-                    "_id": ObjectId(conversation_id),
-                    "$or": [{"user": user_id}, {"shared_with": user_id}],
-                }
-            )
-
-            if not conversation:
-                logger.warning(
-                    f"Conversation not found or unauthorized - ID: {conversation_id}, User: {user_id}"
-                )
-                return None
-            conversation["_id"] = str(conversation["_id"])
-            return conversation
+            with db_readonly() as conn:
+                repo = ConversationsRepository(conn)
+                conv = repo.get_any(conversation_id, user_id)
+                if conv is None:
+                    logger.warning(
+                        f"Conversation not found or unauthorized - ID: {conversation_id}, User: {user_id}"
+                    )
+                    return None
+                messages = repo.get_messages(str(conv["id"]))
+            conv["queries"] = messages
+            conv["_id"] = str(conv["id"])
+            return conv
         except Exception as e:
             logger.error(f"Error fetching conversation: {str(e)}", exc_info=True)
             return None
@@ -64,7 +69,11 @@ class ConversationService:
         attachment_ids: Optional[List[str]] = None,
         metadata: Optional[Dict[str, Any]] = None,
     ) -> str:
-        """Save or update a conversation in the database"""
+        """Save or update a conversation in Postgres.
+
+        Returns the string conversation id (PG UUID as string, or the
+        caller-provided id if it was already a UUID).
+        """
         if decoded_token is None:
             raise ValueError("Invalid or missing authentication token")
         user_id = decoded_token.get("sub")
@@ -72,117 +81,47 @@ class ConversationService:
             raise ValueError("User ID not found in token")
         current_time = datetime.now(timezone.utc)
 
-        # clean up in sources array such that we save max 1k characters for text part
+        # Trim huge inline source text to a reasonable max before persist.
         for source in sources:
             if "text" in source and isinstance(source["text"], str):
                 source["text"] = source["text"][:1000]
 
+        message_payload = {
+            "prompt": question,
+            "response": response,
+            "thought": thought,
+            "sources": sources,
+            "tool_calls": tool_calls,
+            "attachments": attachment_ids,
+            "model_id": model_id,
+            "timestamp": current_time,
+        }
+        if metadata:
+            message_payload["metadata"] = metadata
+
         if conversation_id is not None and index is not None:
-            # Update existing conversation with new query
-
-            result = self.conversations_collection.update_one(
-                {
-                    "_id": ObjectId(conversation_id),
-                    "user": user_id,
-                    f"queries.{index}": {"$exists": True},
-                },
-                {
-                    "$set": {
-                        f"queries.{index}.prompt": question,
-                        f"queries.{index}.response": response,
-                        f"queries.{index}.thought": thought,
-                        f"queries.{index}.sources": sources,
-                        f"queries.{index}.tool_calls": tool_calls,
-                        f"queries.{index}.timestamp": current_time,
-                        f"queries.{index}.attachments": attachment_ids,
-                        f"queries.{index}.model_id": model_id,
-                        **(
-                            {f"queries.{index}.metadata": metadata}
-                            if metadata
-                            else {}
-                        ),
-                    }
-                },
-            )
-
-            if result.matched_count == 0:
-                raise ValueError("Conversation not found or unauthorized")
-            self.conversations_collection.update_one(
-                {
-                    "_id": ObjectId(conversation_id),
-                    "user": user_id,
-                    f"queries.{index}": {"$exists": True},
-                },
-                {"$push": {"queries": {"$each": [], "$slice": index + 1}}},
-            )
-            # Dual-write to Postgres: update the message at :index and
-            # truncate anything after it, mirroring Mongo's $set+$slice.
-            def _pg_update_at_index(repo: ConversationsRepository) -> None:
-                conv = repo.get_by_legacy_id(conversation_id)
+            with db_session() as conn:
+                repo = ConversationsRepository(conn)
+                conv = repo.get_any(conversation_id, user_id)
                 if conv is None:
-                    return
-                repo.update_message_at(conv["id"], index, {
-                    "prompt": question,
-                    "response": response,
-                    "thought": thought,
-                    "sources": sources,
-                    "tool_calls": tool_calls,
-                    "attachments": attachment_ids,
-                    "model_id": model_id,
-                    "timestamp": current_time,
-                    **({"metadata": metadata} if metadata else {}),
-                })
-                repo.truncate_after(conv["id"], index)
-
-            dual_write(ConversationsRepository, _pg_update_at_index)
+                    raise ValueError("Conversation not found or unauthorized")
+                conv_pg_id = str(conv["id"])
+                repo.update_message_at(conv_pg_id, index, message_payload)
+                repo.truncate_after(conv_pg_id, index)
             return conversation_id
         elif conversation_id:
-            # Append new message to existing conversation
-
-            result = self.conversations_collection.update_one(
-                {"_id": ObjectId(conversation_id), "user": user_id},
-                {
-                    "$push": {
-                        "queries": {
-                            "prompt": question,
-                            "response": response,
-                            "thought": thought,
-                            "sources": sources,
-                            "tool_calls": tool_calls,
-                            "timestamp": current_time,
-                            "attachments": attachment_ids,
-                            "model_id": model_id,
-                            **({"metadata": metadata} if metadata else {}),
-                        }
-                    }
-                },
-            )
-
-            if result.matched_count == 0:
-                raise ValueError("Conversation not found or unauthorized")
-
-            # Dual-write to Postgres: append the same message.
-            def _pg_append(repo: ConversationsRepository) -> None:
-                conv = repo.get_by_legacy_id(conversation_id)
+            with db_session() as conn:
+                repo = ConversationsRepository(conn)
+                conv = repo.get_any(conversation_id, user_id)
                 if conv is None:
-                    return
-                repo.append_message(conv["id"], {
-                    "prompt": question,
-                    "response": response,
-                    "thought": thought,
-                    "sources": sources,
-                    "tool_calls": tool_calls,
-                    "attachments": attachment_ids,
-                    "model_id": model_id,
-                    "timestamp": current_time,
-                    "metadata": metadata or {},
-                })
-
-            dual_write(ConversationsRepository, _pg_append)
+                    raise ValueError("Conversation not found or unauthorized")
+                conv_pg_id = str(conv["id"])
+                # append_message expects 'metadata' key either way; normalise.
+                append_payload = dict(message_payload)
+                append_payload.setdefault("metadata", metadata or {})
+                repo.append_message(conv_pg_id, append_payload)
             return conversation_id
         else:
-            # Create new conversation
-
             messages_summary = [
                 {
                     "role": "system",
@@ -197,125 +136,80 @@ class ConversationService:
                 },
             ]
 
+            # ``model_id`` here is the registry id (a UUID for BYOM
+            # records). The LLM's own ``model_id`` is the upstream name
+            # LLMCreator resolved at construction time — that's what
+            # the provider's API expects. Built-ins are unaffected.
             completion = llm.gen(
-                model=model_id, messages=messages_summary, max_tokens=500
+                model=getattr(llm, "model_id", None) or model_id,
+                messages=messages_summary,
+                max_tokens=500,
             )
 
             if not completion or not completion.strip():
                 completion = question[:50] if question else "New Conversation"
 
-            query_doc = {
-                "prompt": question,
-                "response": response,
-                "thought": thought,
-                "sources": sources,
-                "tool_calls": tool_calls,
-                "timestamp": current_time,
-                "attachments": attachment_ids,
-                "model_id": model_id,
-            }
-            if metadata:
-                query_doc["metadata"] = metadata
-
-            conversation_data = {
-                "user": user_id,
-                "date": current_time,
-                "name": completion,
-                "queries": [query_doc],
-            }
-
+            resolved_api_key: Optional[str] = None
+            resolved_agent_id: Optional[str] = None
             if api_key:
-                if agent_id:
-                    conversation_data["agent_id"] = agent_id
-                    if is_shared_usage:
-                        conversation_data["is_shared_usage"] = is_shared_usage
-                        conversation_data["shared_token"] = shared_token
-                agent = self.agents_collection.find_one({"key": api_key})
+                with db_readonly() as conn:
+                    agent = AgentsRepository(conn).find_by_key(api_key)
                 if agent:
-                    conversation_data["api_key"] = agent["key"]
-            result = self.conversations_collection.insert_one(conversation_data)
-            inserted_id = str(result.inserted_id)
+                    resolved_api_key = agent.get("key")
+                if agent_id:
+                    resolved_agent_id = agent_id
 
-            # Dual-write to Postgres: create the conversation row with
-            # legacy_mongo_id and append the first message.
-            def _pg_create(repo: ConversationsRepository) -> None:
+            with db_session() as conn:
+                repo = ConversationsRepository(conn)
                 conv = repo.create(
                     user_id,
                     completion,
-                    agent_id=conversation_data.get("agent_id"),
-                    api_key=conversation_data.get("api_key"),
-                    is_shared_usage=conversation_data.get("is_shared_usage", False),
-                    shared_token=conversation_data.get("shared_token"),
-                    legacy_mongo_id=inserted_id,
+                    agent_id=resolved_agent_id,
+                    api_key=resolved_api_key,
+                    is_shared_usage=bool(resolved_agent_id and is_shared_usage),
+                    shared_token=(
+                        shared_token
+                        if (resolved_agent_id and is_shared_usage)
+                        else None
+                    ),
                 )
-                repo.append_message(conv["id"], {
-                    "prompt": question,
-                    "response": response,
-                    "thought": thought,
-                    "sources": sources,
-                    "tool_calls": tool_calls,
-                    "attachments": attachment_ids,
-                    "model_id": model_id,
-                    "timestamp": current_time,
-                    "metadata": metadata or {},
-                })
-
-            dual_write(ConversationsRepository, _pg_create)
-            return inserted_id
+                conv_pg_id = str(conv["id"])
+                append_payload = dict(message_payload)
+                append_payload.setdefault("metadata", metadata or {})
+                repo.append_message(conv_pg_id, append_payload)
+            return conv_pg_id
 
     def update_compression_metadata(
         self, conversation_id: str, compression_metadata: Dict[str, Any]
     ) -> None:
-        """
-        Update conversation with compression metadata.
+        """Persist compression flags and append a compression point.
 
-        Uses $push with $slice to keep only the most recent compression points,
-        preventing unbounded array growth. Since each compression incorporates
-        previous compressions, older points become redundant.
-
-        Args:
-            conversation_id: Conversation ID
-            compression_metadata: Compression point data
+        Mirrors the Mongo-era ``$set`` + ``$push $slice`` on
+        ``compression_metadata`` but goes through the PG repo API.
         """
         try:
-            self.conversations_collection.update_one(
-                {"_id": ObjectId(conversation_id)},
-                {
-                    "$set": {
-                        "compression_metadata.is_compressed": True,
-                        "compression_metadata.last_compression_at": compression_metadata.get(
-                            "timestamp"
-                        ),
-                    },
-                    "$push": {
-                        "compression_metadata.compression_points": {
-                            "$each": [compression_metadata],
-                            "$slice": -settings.COMPRESSION_MAX_HISTORY_POINTS,
-                        }
-                    },
-                },
-            )
-            logger.info(
-                f"Updated compression metadata for conversation {conversation_id}"
-            )
-
-            # Dual-write to Postgres: mirror $set + $push $slice.
-            def _pg_compression(repo: ConversationsRepository) -> None:
+            with db_session() as conn:
+                repo = ConversationsRepository(conn)
+                # conversation_id here comes from the streaming pipeline
+                # which has already resolved it; accept either UUID or
+                # legacy id for safety.
                 conv = repo.get_by_legacy_id(conversation_id)
-                if conv is None:
-                    return
+                conv_pg_id = (
+                    str(conv["id"]) if conv is not None else conversation_id
+                )
                 repo.set_compression_flags(
-                    conv["id"],
+                    conv_pg_id,
                     is_compressed=True,
                     last_compression_at=compression_metadata.get("timestamp"),
                 )
                 repo.append_compression_point(
-                    conv["id"],
+                    conv_pg_id,
                     compression_metadata,
                     max_points=settings.COMPRESSION_MAX_HISTORY_POINTS,
                 )
-
-            dual_write(ConversationsRepository, _pg_compression)
+            logger.info(
+                f"Updated compression metadata for conversation {conversation_id}"
+            )
         except Exception as e:
             logger.error(
                 f"Error updating compression metadata: {str(e)}", exc_info=True
@@ -325,39 +219,22 @@ class ConversationService:
     def append_compression_message(
         self, conversation_id: str, compression_metadata: Dict[str, Any]
     ) -> None:
-        """
-        Append a synthetic compression summary entry into the conversation history.
-        This makes the summary visible in the DB alongside normal queries.
-        """
+        """Append a synthetic compression summary message to the conversation."""
         try:
             summary = compression_metadata.get("compressed_summary", "")
             if not summary:
                 return
-            timestamp = compression_metadata.get("timestamp", datetime.now(timezone.utc))
-
-            self.conversations_collection.update_one(
-                {"_id": ObjectId(conversation_id)},
-                {
-                    "$push": {
-                        "queries": {
-                            "prompt": "[Context Compression Summary]",
-                            "response": summary,
-                            "thought": "",
-                            "sources": [],
-                            "tool_calls": [],
-                            "timestamp": timestamp,
-                            "attachments": [],
-                            "model_id": compression_metadata.get("model_used"),
-                        }
-                    }
-                },
+            timestamp = compression_metadata.get(
+                "timestamp", datetime.now(timezone.utc)
             )
 
-            def _pg_append_summary(repo: ConversationsRepository) -> None:
+            with db_session() as conn:
+                repo = ConversationsRepository(conn)
                 conv = repo.get_by_legacy_id(conversation_id)
-                if conv is None:
-                    return
-                repo.append_message(conv["id"], {
+                conv_pg_id = (
+                    str(conv["id"]) if conv is not None else conversation_id
+                )
+                repo.append_message(conv_pg_id, {
                     "prompt": "[Context Compression Summary]",
                     "response": summary,
                     "thought": "",
@@ -367,9 +244,9 @@ class ConversationService:
                     "model_id": compression_metadata.get("model_used"),
                     "timestamp": timestamp,
                 })
-
-            dual_write(ConversationsRepository, _pg_append_summary)
-            logger.info(f"Appended compression summary to conversation {conversation_id}")
+            logger.info(
+                f"Appended compression summary to conversation {conversation_id}"
+            )
         except Exception as e:
             logger.error(
                 f"Error appending compression summary: {str(e)}", exc_info=True
@@ -378,20 +255,30 @@ class ConversationService:
     def get_compression_metadata(
         self, conversation_id: str
     ) -> Optional[Dict[str, Any]]:
-        """
-        Get compression metadata for a conversation.
-
-        Args:
-            conversation_id: Conversation ID
-
-        Returns:
-            Compression metadata dict or None
-        """
+        """Fetch the stored compression metadata JSONB blob for a conversation."""
         try:
-            conversation = self.conversations_collection.find_one(
-                {"_id": ObjectId(conversation_id)}, {"compression_metadata": 1}
-            )
-            return conversation.get("compression_metadata") if conversation else None
+            with db_readonly() as conn:
+                repo = ConversationsRepository(conn)
+                conv = repo.get_by_legacy_id(conversation_id)
+                if conv is None:
+                    # Fallback to UUID lookup without user scoping — the
+                    # caller already holds an authenticated conversation
+                    # id from the streaming path. Gate on id shape so a
+                    # non-UUID (legacy ObjectId that wasn't backfilled)
+                    # doesn't reach CAST — the cast raises and spams the
+                    # logs with a stack trace on every call.
+                    if not looks_like_uuid(conversation_id):
+                        return None
+                    result = conn.execute(
+                        sql_text(
+                            "SELECT compression_metadata FROM conversations "
+                            "WHERE id = CAST(:id AS uuid)"
+                        ),
+                        {"id": conversation_id},
+                    )
+                    row = result.fetchone()
+                    return row[0] if row is not None else None
+            return conv.get("compression_metadata") if conv else None
         except Exception as e:
             logger.error(
                 f"Error getting compression metadata: {str(e)}", exc_info=True

application/api/answer/services/stream_processor.py

@@ -5,10 +5,6 @@ import os
 from pathlib import Path
 from typing import Any, Dict, Optional, Set
 
-from bson.dbref import DBRef
-
-from bson.objectid import ObjectId
-
 from application.agents.agent_creator import AgentCreator
 from application.api.answer.services.compression import CompressionOrchestrator
 from application.api.answer.services.compression.token_counter import TokenCounter
@@ -20,8 +16,16 @@ from application.core.model_utils import (
     get_provider_from_model_id,
     validate_model_id,
 )
-from application.core.mongo_db import MongoDB
 from application.core.settings import settings
+from sqlalchemy import text as sql_text
+
+from application.storage.db.base_repository import looks_like_uuid, row_to_dict
+from application.storage.db.repositories.agents import AgentsRepository
+from application.storage.db.repositories.attachments import AttachmentsRepository
+from application.storage.db.repositories.prompts import PromptsRepository
+from application.storage.db.repositories.sources import SourcesRepository
+from application.storage.db.repositories.user_tools import UserToolsRepository
+from application.storage.db.session import db_readonly, db_session
 from application.retriever.retriever_creator import RetrieverCreator
 from application.utils import (
     calculate_doc_token_budget,
@@ -32,28 +36,41 @@ logger = logging.getLogger(__name__)
 
 
 def get_prompt(prompt_id: str, prompts_collection=None) -> str:
+    """Get a prompt by preset name or Postgres ID (UUID or legacy ObjectId).
+
+    The ``prompts_collection`` parameter is retained for backwards
+    compatibility with call sites that still pass it positionally; it is
+    ignored post-cutover.
     """
-    Get a prompt by preset name or MongoDB ID
-    """
+    del prompts_collection  # unused — retained for call-site compatibility
+    # Callers may pass a ``uuid.UUID`` (from a PG ``prompt_id`` column) or a
+    # plain string ("default"/"creative"/legacy ObjectId). Normalise to str
+    # so both the preset lookup and the UUID-vs-legacy branching work.
+    # ``None`` / empty means "use the default prompt" — agents that never
+    # set a custom prompt land here (PG ``agents.prompt_id`` is NULL).
+    if prompt_id is None or prompt_id == "":
+        prompt_id = "default"
+    elif not isinstance(prompt_id, str):
+        prompt_id = str(prompt_id)
     current_dir = Path(__file__).resolve().parents[3]
     prompts_dir = current_dir / "prompts"
 
-    # Maps for classic agent types
     CLASSIC_PRESETS = {
         "default": "chat_combine_default.txt",
         "creative": "chat_combine_creative.txt",
         "strict": "chat_combine_strict.txt",
         "reduce": "chat_reduce_prompt.txt",
     }
-
-    # Agentic counterparts — same styles, but with search tool instructions
     AGENTIC_PRESETS = {
         "default": "agentic/default.txt",
         "creative": "agentic/creative.txt",
         "strict": "agentic/strict.txt",
     }
 
-    preset_mapping = {**CLASSIC_PRESETS, **{f"agentic_{k}": v for k, v in AGENTIC_PRESETS.items()}}
+    preset_mapping = {
+        **CLASSIC_PRESETS,
+        **{f"agentic_{k}": v for k, v in AGENTIC_PRESETS.items()},
+    }
 
     if prompt_id in preset_mapping:
         file_path = os.path.join(prompts_dir, preset_mapping[prompt_id])
@@ -63,14 +80,18 @@ def get_prompt(prompt_id: str, prompts_collection=None) -> str:
         except FileNotFoundError:
             raise FileNotFoundError(f"Prompt file not found: {file_path}")
     try:
-        if prompts_collection is None:
-            mongo = MongoDB.get_client()
-            db = mongo[settings.MONGO_DB_NAME]
-            prompts_collection = db["prompts"]
-        prompt_doc = prompts_collection.find_one({"_id": ObjectId(prompt_id)})
+        with db_readonly() as conn:
+            repo = PromptsRepository(conn)
+            prompt_doc = None
+            if looks_like_uuid(prompt_id):
+                prompt_doc = repo.get_for_rendering(prompt_id)
+            if prompt_doc is None:
+                prompt_doc = repo.get_by_legacy_id(prompt_id)
         if not prompt_doc:
             raise ValueError(f"Prompt with ID {prompt_id} not found")
         return prompt_doc["content"]
+    except ValueError:
+        raise
     except Exception as e:
         raise ValueError(f"Invalid prompt ID: {prompt_id}") from e
 
@@ -79,12 +100,9 @@ class StreamProcessor:
     def __init__(
         self, request_data: Dict[str, Any], decoded_token: Optional[Dict[str, Any]]
     ):
-        mongo = MongoDB.get_client()
-        self.db = mongo[settings.MONGO_DB_NAME]
-        self.agents_collection = self.db["agents"]
-        self.attachments_collection = self.db["attachments"]
-        self.prompts_collection = self.db["prompts"]
-
+        # Legacy attribute retained as None for any external callers that
+        # introspect the processor; all DB access uses per-op connections.
+        self.prompts_collection = None
         self.data = request_data
         self.decoded_token = decoded_token
         self.initial_user_id = (
@@ -103,6 +121,8 @@ class StreamProcessor:
         self.agent_id = self.data.get("agent_id")
         self.agent_key = None
         self.model_id: Optional[str] = None
+        # BYOM-resolution scope, set by _validate_and_set_model.
+        self.model_user_id: Optional[str] = None
         self.conversation_service = ConversationService()
         self.compression_orchestrator = CompressionOrchestrator(
             self.conversation_service
@@ -173,16 +193,23 @@ class StreamProcessor:
                     for query in conversation.get("queries", [])
                 ]
         else:
+            # model_user_id keeps history trim aligned with the BYOM's
+            # actual context window instead of the default 128k.
             self.history = limit_chat_history(
-                json.loads(self.data.get("history", "[]")), model_id=self.model_id
+                json.loads(self.data.get("history", "[]")),
+                model_id=self.model_id,
+                user_id=self.model_user_id,
             )
 
     def _handle_compression(self, conversation: Dict[str, Any]):
         """Handle conversation compression logic using orchestrator."""
         try:
+            # initial_user_id for conversation access; model_user_id
+            # for BYOM context-window / provider lookups.
             result = self.compression_orchestrator.compress_if_needed(
                 conversation_id=self.conversation_id,
                 user_id=self.initial_user_id,
+                model_user_id=self.model_user_id,
                 model_id=self.model_id,
                 decoded_token=self.decoded_token,
             )
@@ -244,17 +271,21 @@ class StreamProcessor:
         if not attachment_ids:
             return []
         attachments = []
-        for attachment_id in attachment_ids:
-            try:
-                attachment_doc = self.attachments_collection.find_one(
-                    {"_id": ObjectId(attachment_id), "user": user_id}
-                )
-                if attachment_doc:
-                    attachments.append(attachment_doc)
-            except Exception as e:
-                logger.error(
-                    f"Error retrieving attachment {attachment_id}: {e}", exc_info=True
-                )
+        try:
+            with db_readonly() as conn:
+                repo = AttachmentsRepository(conn)
+                for attachment_id in attachment_ids:
+                    try:
+                        attachment_doc = repo.get_any(str(attachment_id), user_id)
+                        if attachment_doc:
+                            attachments.append(attachment_doc)
+                    except Exception as e:
+                        logger.error(
+                            f"Error retrieving attachment {attachment_id}: {e}",
+                            exc_info=True,
+                        )
+        except Exception as e:
+            logger.error(f"Error opening attachments connection: {e}", exc_info=True)
         return attachments
 
     def _validate_and_set_model(self):
@@ -262,11 +293,18 @@ class StreamProcessor:
         from application.core.model_settings import ModelRegistry
 
         requested_model = self.data.get("model_id")
+        # Caller picks from their own BYOM layer; agent defaults resolve
+        # under the owner's layer (shared agents have caller != owner).
+        caller_user_id = self.initial_user_id
+        owner_user_id = self.agent_config.get("user_id") or caller_user_id
 
         if requested_model:
-            if not validate_model_id(requested_model):
+            if not validate_model_id(requested_model, user_id=caller_user_id):
                 registry = ModelRegistry.get_instance()
-                available_models = [m.id for m in registry.get_enabled_models()]
+                available_models = [
+                    m.id
+                    for m in registry.get_enabled_models(user_id=caller_user_id)
+                ]
                 raise ValueError(
                     f"Invalid model_id '{requested_model}'. "
                     f"Available models: {', '.join(available_models[:5])}"
@@ -277,86 +315,114 @@ class StreamProcessor:
                     )
                 )
             self.model_id = requested_model
+            self.model_user_id = caller_user_id
         else:
             agent_default_model = self.agent_config.get("default_model_id", "")
-            if agent_default_model and validate_model_id(agent_default_model):
+            if agent_default_model and validate_model_id(
+                agent_default_model, user_id=owner_user_id
+            ):
                 self.model_id = agent_default_model
+                self.model_user_id = owner_user_id
             else:
                 self.model_id = get_default_model_id()
+                self.model_user_id = None
 
     def _get_agent_key(self, agent_id: Optional[str], user_id: Optional[str]) -> tuple:
-        """Get API key for agent with access control"""
+        """Get API key for agent with access control."""
         if not agent_id:
             return None, False, None
         try:
-            agent = self.agents_collection.find_one({"_id": ObjectId(agent_id)})
+            with db_readonly() as conn:
+                # Lookup without user scoping — access control is done
+                # against ``user_id`` / ``shared_with`` / ``shared`` flags
+                # right below, matching the legacy Mongo semantics.
+                repo = AgentsRepository(conn)
+                agent = None
+                if looks_like_uuid(str(agent_id)):
+                    result = conn.execute(
+                        sql_text(
+                            "SELECT * FROM agents WHERE id = CAST(:id AS uuid)"
+                        ),
+                        {"id": str(agent_id)},
+                    )
+                    row = result.fetchone()
+                    if row is not None:
+                        agent = row_to_dict(row)
+                if agent is None:
+                    agent = repo.get_by_legacy_id(str(agent_id))
             if agent is None:
                 raise Exception("Agent not found")
-            is_owner = agent.get("user") == user_id
-            is_shared_with_user = agent.get(
-                "shared_publicly", False
-            ) or user_id in agent.get("shared_with", [])
+            agent_owner = agent.get("user_id")
+            is_owner = agent_owner == user_id
+            is_shared_with_user = bool(agent.get("shared", False))
 
             if not (is_owner or is_shared_with_user):
                 raise Exception("Unauthorized access to the agent")
             if is_owner:
-                self.agents_collection.update_one(
-                    {"_id": ObjectId(agent_id)},
-                    {
-                        "$set": {
-                            "lastUsedAt": datetime.datetime.now(datetime.timezone.utc)
-                        }
-                    },
-                )
-            return str(agent["key"]), not is_owner, agent.get("shared_token")
+                now = datetime.datetime.now(datetime.timezone.utc)
+                try:
+                    with db_session() as conn:
+                        AgentsRepository(conn).update(
+                            str(agent["id"]), agent_owner,
+                            {"last_used_at": now},
+                        )
+                except Exception:
+                    logger.warning(
+                        "Failed to update last_used_at for agent",
+                        exc_info=True,
+                    )
+            return (
+                str(agent["key"]) if agent.get("key") else None,
+                not is_owner,
+                agent.get("shared_token"),
+            )
         except Exception as e:
             logger.error(f"Error in get_agent_key: {str(e)}", exc_info=True)
             raise
 
     def _get_data_from_api_key(self, api_key: str) -> Dict[str, Any]:
-        data = self.agents_collection.find_one({"key": api_key})
-        if not data:
-            raise Exception("Invalid API Key, please generate a new key", 401)
-        source = data.get("source")
-        if isinstance(source, DBRef):
-            source_doc = self.db.dereference(source)
-            if source_doc:
-                data["source"] = str(source_doc["_id"])
-                data["retriever"] = source_doc.get("retriever", data.get("retriever"))
-                data["chunks"] = source_doc.get("chunks", data.get("chunks"))
+        with db_readonly() as conn:
+            agent = AgentsRepository(conn).find_by_key(api_key)
+            if not agent:
+                raise Exception("Invalid API Key, please generate a new key", 401)
+            sources_repo = SourcesRepository(conn)
+            # The repo dict uses "user_id" — the streaming path expects
+            # a "user" key (legacy Mongo shape) for identity propagation.
+            data: Dict[str, Any] = dict(agent)
+            data["user"] = agent.get("user_id")
+
+            # Resolve the primary source row (if any) for retriever/chunks.
+            source_id = agent.get("source_id")
+            if source_id:
+                source_doc = sources_repo.get(str(source_id), agent.get("user_id"))
+                if source_doc:
+                    data["source"] = str(source_doc["id"])
+                    data["retriever"] = source_doc.get(
+                        "retriever", data.get("retriever")
+                    )
+                    data["chunks"] = source_doc.get("chunks", data.get("chunks"))
+                else:
+                    data["source"] = None
             else:
                 data["source"] = None
-        elif source == "default":
-            data["source"] = "default"
-        else:
-            data["source"] = None
 
-        sources = data.get("sources", [])
-        if sources and isinstance(sources, list):
             sources_list = []
-            for i, source_ref in enumerate(sources):
-                if source_ref == "default":
-                    processed_source = {
-                        "id": "default",
-                        "retriever": "classic",
-                        "chunks": data.get("chunks", "2"),
-                    }
-                    sources_list.append(processed_source)
-                elif isinstance(source_ref, DBRef):
-                    source_doc = self.db.dereference(source_ref)
+            extra = agent.get("extra_source_ids") or []
+            if extra:
+                for sid in extra:
+                    source_doc = sources_repo.get(str(sid), agent.get("user_id"))
                     if source_doc:
-                        processed_source = {
-                            "id": str(source_doc["_id"]),
-                            "retriever": source_doc.get("retriever", "classic"),
-                            "chunks": source_doc.get("chunks", data.get("chunks", "2")),
-                        }
-                        sources_list.append(processed_source)
-            data["sources"] = sources_list
-        else:
-            data["sources"] = []
-
+                        sources_list.append(
+                            {
+                                "id": str(source_doc["id"]),
+                                "retriever": source_doc.get("retriever", "classic"),
+                                "chunks": source_doc.get(
+                                    "chunks", data.get("chunks", "2")
+                                ),
+                            }
+                        )
+        data["sources"] = sources_list
         data["default_model_id"] = data.get("default_model_id", "")
-
         return data
 
     def _configure_source(self):
@@ -469,6 +535,10 @@ class StreamProcessor:
                     "allow_system_prompt_override": self._agent_data.get(
                         "allow_system_prompt_override", False
                     ),
+                    # Owner identity — _validate_and_set_model reads this to
+                    # resolve owner-stored BYOM default_model_id against the
+                    # owner's per-user model layer rather than the caller's.
+                    "user_id": self._agent_data.get("user"),
                 }
             )
 
@@ -484,8 +554,14 @@ class StreamProcessor:
                 # Owner using their own agent
                 self.decoded_token = {"sub": self._agent_data.get("user")}
 
-            if self._agent_data.get("workflow"):
-                self.agent_config["workflow"] = self._agent_data["workflow"]
+            # PG row exposes the workflow as ``workflow_id`` (UUID column);
+            # legacy Mongo shape used the key ``workflow``. Accept either so
+            # API-key-invoked workflow agents bind correctly downstream.
+            wf_ref = self._agent_data.get("workflow") or self._agent_data.get(
+                "workflow_id"
+            )
+            if wf_ref:
+                self.agent_config["workflow"] = str(wf_ref)
                 self.agent_config["workflow_owner"] = self._agent_data.get("user")
         else:
             # No API key — default/workflow configuration
@@ -510,7 +586,13 @@ class StreamProcessor:
 
     def _configure_retriever(self):
         """Assemble retriever config with precedence: request > agent > default."""
-        doc_token_limit = calculate_doc_token_budget(model_id=self.model_id)
+        # BYOM scope: owner for shared-agent BYOM, caller for own BYOM,
+        # None for built-ins. Without ``user_id`` here, the doc budget
+        # falls back to settings.DEFAULT_LLM_TOKEN_LIMIT and overfills
+        # the upstream context window for any small (e.g. 8k/32k) BYOM.
+        doc_token_limit = calculate_doc_token_budget(
+            model_id=self.model_id, user_id=self.model_user_id
+        )
 
         # Start with defaults
         retriever_name = "classic"
@@ -561,6 +643,7 @@ class StreamProcessor:
             chunks=self.retriever_config["chunks"],
             doc_token_limit=self.retriever_config.get("doc_token_limit", 50000),
             model_id=self.model_id,
+            model_user_id=self.model_user_id,
             user_api_key=self.agent_config["user_api_key"],
             agent_id=self.agent_id,
             decoded_token=self.decoded_token,
@@ -620,12 +703,9 @@ class StreamProcessor:
         filtering_enabled = required_tool_actions is not None
 
         try:
-            user_tools_collection = self.db["user_tools"]
             user_id = self.initial_user_id or "local"
-
-            user_tools = list(
-                user_tools_collection.find({"user": user_id, "status": True})
-            )
+            with db_readonly() as conn:
+                user_tools = UserToolsRepository(conn).list_active_for_user(user_id)
 
             if not user_tools:
                 return None
@@ -855,6 +935,11 @@ class StreamProcessor:
         agent_config = state["agent_config"]
 
         model_id = agent_config.get("model_id")
+        # BYOM scope captured at initial dispatch. None for built-ins or
+        # caller-owned BYOM where decoded_token['sub'] is already the
+        # right scope; non-None for shared-agent owner BYOM where the
+        # caller's identity differs from the model owner's.
+        model_user_id = agent_config.get("model_user_id")
         llm_name = agent_config.get("llm_name", settings.LLM_PROVIDER)
         api_key = agent_config.get("api_key")
         user_api_key = agent_config.get("user_api_key")
@@ -872,6 +957,7 @@ class StreamProcessor:
             decoded_token=self.decoded_token,
             model_id=model_id,
             agent_id=agent_id,
+            model_user_id=model_user_id,
         )
         llm_handler = LLMHandlerCreator.create_handler(llm_name or "default")
         tool_executor = ToolExecutor(
@@ -901,6 +987,7 @@ class StreamProcessor:
             "endpoint": "stream",
             "llm_name": llm_name,
             "model_id": model_id,
+            "model_user_id": model_user_id,
             "api_key": system_api_key,
             "agent_id": agent_id,
             "user_api_key": user_api_key,
@@ -923,6 +1010,15 @@ class StreamProcessor:
 
         # Store config for the route layer
         self.model_id = model_id
+        # Mirror ``model_user_id`` back onto the processor so the route
+        # layer (StreamResource) reads the owner scope captured at
+        # initial dispatch. Without this, ``processor.model_user_id``
+        # stays at the __init__ default (None) and complete_stream
+        # falls back to the caller's sub: the post-resume title-LLM
+        # save misses the owner's BYOM layer, and any second tool
+        # pause persists ``model_user_id=None`` — losing owner scope
+        # for every subsequent resume of this conversation.
+        self.model_user_id = model_user_id
         self.agent_id = agent_id
         self.agent_config["user_api_key"] = user_api_key
         self.conversation_id = conversation_id
@@ -974,8 +1070,11 @@ class StreamProcessor:
                 tools_data=tools_data,
             )
 
+        # Use the user_id that resolved the model so owner-scoped BYOM
+        # records dispatch correctly on shared-agent requests.
+        model_user_id = getattr(self, "model_user_id", self.initial_user_id)
         provider = (
-            get_provider_from_model_id(self.model_id)
+            get_provider_from_model_id(self.model_id, user_id=model_user_id)
             if self.model_id
             else settings.LLM_PROVIDER
         )
@@ -986,8 +1085,10 @@ class StreamProcessor:
         from application.llm.handlers.handler_creator import LLMHandlerCreator
         from application.agents.tool_executor import ToolExecutor
 
-        # Compute backup models: agent's configured models minus the active one
-        agent_models = self.agent_config.get("models", [])
+        # Compute backup models: agent's configured models minus the active one.
+        # PG agents may carry an explicit ``models: NULL`` (not absent), so
+        # ``.get("models", [])`` isn't enough — coerce None → [].
+        agent_models = self.agent_config.get("models") or []
         backup_models = [m for m in agent_models if m != self.model_id]
 
         llm = LLMCreator.create_llm(
@@ -998,6 +1099,8 @@ class StreamProcessor:
             model_id=self.model_id,
             agent_id=self.agent_id,
             backup_models=backup_models,
+            # Owner-scope on shared-agent BYOM dispatch.
+            model_user_id=model_user_id,
         )
         llm_handler = LLMHandlerCreator.create_handler(
             provider if provider else "default"
@@ -1020,6 +1123,7 @@ class StreamProcessor:
             "endpoint": "stream",
             "llm_name": provider or settings.LLM_PROVIDER,
             "model_id": self.model_id,
+            "model_user_id": self.model_user_id,
             "api_key": system_api_key,
             "agent_id": self.agent_id,
             "user_api_key": self.agent_config["user_api_key"],
@@ -1047,6 +1151,7 @@ class StreamProcessor:
                     "doc_token_limit", 50000
                 ),
                 "model_id": self.model_id,
+                "model_user_id": self.model_user_id,
                 "user_api_key": self.agent_config["user_api_key"],
                 "agent_id": self.agent_id,
                 "llm_name": provider or settings.LLM_PROVIDER,

application/api/connector/routes.py

@@ -1,12 +1,10 @@
 import base64
-import datetime
 import html
 import json
 import uuid
 from urllib.parse import urlencode
 
 
-from bson.objectid import ObjectId
 from flask import (
     Blueprint,
     current_app,
@@ -17,22 +15,18 @@ from flask import (
 from flask_restx import fields, Namespace, Resource
 
 
+from application.api import api
 from application.api.user.tasks import (
     ingest_connector_task,
 )
-from application.core.mongo_db import MongoDB
-from application.core.settings import settings
-from application.api import api
-
-
 from application.parser.connectors.connector_creator import ConnectorCreator
+from application.storage.db.repositories.connector_sessions import (
+    ConnectorSessionsRepository,
+)
+from application.storage.db.repositories.sources import SourcesRepository
+from application.storage.db.session import db_readonly, db_session
 
 
-mongo = MongoDB.get_client()
-db = mongo[settings.MONGO_DB_NAME]
-sources_collection = db["sources"]
-sessions_collection = db["connector_sessions"]
-
 connector = Blueprint("connector", __name__)
 connectors_ns = Namespace("connectors", description="Connector operations", path="/")
 api.add_namespace(connectors_ns)
@@ -68,16 +62,14 @@ class ConnectorAuth(Resource):
                 return make_response(jsonify({"success": False, "error": "Unauthorized"}), 401)
             user_id = decoded_token.get('sub')
 
-            now = datetime.datetime.now(datetime.timezone.utc)
-            result = sessions_collection.insert_one({
-                "provider": provider,
-                "user": user_id,
-                "status": "pending",
-                "created_at": now
-            })
+            with db_session() as conn:
+                session_row = ConnectorSessionsRepository(conn).upsert(
+                    user_id, provider, status="pending",
+                )
+            session_pg_id = str(session_row["id"])
             state_dict = {
                 "provider": provider,
-                "object_id": str(result.inserted_id)
+                "object_id": session_pg_id,
             }
             state = base64.urlsafe_b64encode(json.dumps(state_dict).encode()).decode()
 
@@ -160,17 +152,25 @@ class ConnectorsCallback(Resource):
 
                 sanitized_token_info = auth.sanitize_token_info(token_info)
 
-                sessions_collection.find_one_and_update(
-                    {"_id": ObjectId(state_object_id), "provider": provider},
-                    {
-                        "$set": {
-                            "session_token": session_token,
-                            "token_info": sanitized_token_info,
-                            "user_email": user_email,
-                            "status": "authorized"
-                        }
-                    }
-                )
+                # ``object_id`` in the OAuth state is the PG session row
+                # UUID (new flow) or a legacy Mongo ObjectId (pre-cutover
+                # issued state). Try UUID update first; fall back to
+                # legacy id path.
+                patch = {
+                    "session_token": session_token,
+                    "token_info": sanitized_token_info,
+                    "user_email": user_email,
+                    "status": "authorized",
+                }
+                with db_session() as conn:
+                    repo = ConnectorSessionsRepository(conn)
+                    if state_object_id:
+                        value = str(state_object_id)
+                        updated = False
+                        if len(value) == 36 and "-" in value:
+                            updated = repo.update(value, patch)
+                        if not updated:
+                            repo.update_by_legacy_id(value, patch)
 
                 # Redirect to success page with session token and user email
                 return redirect(build_callback_redirect({
@@ -222,8 +222,11 @@ class ConnectorFiles(Resource):
             if not decoded_token:
                 return make_response(jsonify({"success": False, "error": "Unauthorized"}), 401)
             user = decoded_token.get('sub')
-            session = sessions_collection.find_one({"session_token": session_token, "user": user})
-            if not session:
+            with db_readonly() as conn:
+                session = ConnectorSessionsRepository(conn).get_by_session_token(
+                    session_token,
+                )
+            if not session or session.get("user_id") != user:
                 return make_response(jsonify({"success": False, "error": "Invalid or unauthorized session"}), 401)
 
             loader = ConnectorCreator.create_connector(provider, session_token)
@@ -288,8 +291,11 @@ class ConnectorValidateSession(Resource):
                 return make_response(jsonify({"success": False, "error": "Unauthorized"}), 401)
             user = decoded_token.get('sub')
 
-            session = sessions_collection.find_one({"session_token": session_token, "user": user})
-            if not session or "token_info" not in session:
+            with db_readonly() as conn:
+                session = ConnectorSessionsRepository(conn).get_by_session_token(
+                    session_token,
+                )
+            if not session or session.get("user_id") != user or not session.get("token_info"):
                 return make_response(jsonify({"success": False, "error": "Invalid or expired session"}), 401)
 
             token_info = session["token_info"]
@@ -300,10 +306,11 @@ class ConnectorValidateSession(Resource):
                 try:
                     refreshed_token_info = auth.refresh_access_token(token_info.get('refresh_token'))
                     sanitized_token_info = auth.sanitize_token_info(refreshed_token_info)
-                    sessions_collection.update_one(
-                        {"session_token": session_token},
-                        {"$set": {"token_info": sanitized_token_info}}
-                    )
+                    with db_session() as conn:
+                        repo = ConnectorSessionsRepository(conn)
+                        row = repo.get_by_session_token(session_token)
+                        if row:
+                            repo.update(str(row["id"]), {"token_info": sanitized_token_info})
                     token_info = sanitized_token_info
                     is_expired = False
                 except Exception as refresh_error:
@@ -347,8 +354,11 @@ class ConnectorDisconnect(Resource):
 
 
             if session_token:
-                sessions_collection.delete_one({"session_token": session_token})
-            
+                with db_session() as conn:
+                    ConnectorSessionsRepository(conn).delete_by_session_token(
+                        session_token,
+                    )
+
             return make_response(jsonify({"success": True}), 200)
         except Exception as e:
             current_app.logger.error(f"Error disconnecting connector session: {e}", exc_info=True)
@@ -385,32 +395,28 @@ class ConnectorSync(Resource):
                     }), 
                     400
                 )
-            source = sources_collection.find_one({"_id": ObjectId(source_id)})
+            user_id = decoded_token.get('sub')
+            with db_readonly() as conn:
+                source = SourcesRepository(conn).get_any(source_id, user_id)
             if not source:
                 return make_response(
                     jsonify({
                         "success": False,
                         "error": "Source not found"
-                    }), 
+                    }),
                     404
                 )
 
-            if source.get('user') != decoded_token.get('sub'):
-                return make_response(
-                    jsonify({
-                        "success": False,
-                        "error": "Unauthorized access to source"
-                    }), 
-                    403
-                )
+            # ``get_any`` already scopes by ``user_id``; an extra guard
+            # here would be dead code.
 
-            remote_data = {}
-            try:
-                if source.get('remote_data'):
-                    remote_data = json.loads(source.get('remote_data'))
-            except json.JSONDecodeError:
-                current_app.logger.error(f"Invalid remote_data format for source {source_id}")
-                remote_data = {}
+            remote_data = source.get('remote_data') or {}
+            if isinstance(remote_data, str):
+                try:
+                    remote_data = json.loads(remote_data)
+                except json.JSONDecodeError:
+                    current_app.logger.error(f"Invalid remote_data format for source {source_id}")
+                    remote_data = {}
 
             source_type = remote_data.get('provider')
             if not source_type:
@@ -438,7 +444,7 @@ class ConnectorSync(Resource):
                 recursive=recursive,
                 retriever=source.get('retriever', 'classic'),
                 operation_mode="sync",
-                doc_id=source_id,
+                doc_id=str(source.get('id') or source_id),
                 sync_frequency=source.get('sync_frequency', 'never')
             )
 

application/api/internal/routes.py

@@ -3,18 +3,16 @@ import datetime
 import json
 from flask import Blueprint, request, send_from_directory, jsonify
 from werkzeug.utils import secure_filename
-from bson.objectid import ObjectId
 import logging
-from application.core.mongo_db import MongoDB
+
 from application.core.settings import settings
+from application.storage.db.base_repository import looks_like_uuid
+from application.storage.db.repositories.sources import SourcesRepository
+from application.storage.db.session import db_session
 from application.storage.storage_creator import StorageCreator
 
 
 logger = logging.getLogger(__name__)
-mongo = MongoDB.get_client()
-db = mongo[settings.MONGO_DB_NAME]
-conversations_collection = db["conversations"]
-sources_collection = db["sources"]
 
 current_dir = os.path.dirname(
     os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
@@ -56,21 +54,21 @@ def upload_index_files():
     """Upload two files(index.faiss, index.pkl) to the user's folder."""
     if "user" not in request.form:
         return {"status": "no user"}
-    user = request.form["user"] 
+    user = request.form["user"]
     if "name" not in request.form:
         return {"status": "no name"}
     job_name = request.form["name"]
     tokens = request.form["tokens"]
     retriever = request.form["retriever"]
-    id = request.form["id"]
+    source_id = request.form["id"]
     type = request.form["type"]
     remote_data = request.form["remote_data"] if "remote_data" in request.form else None
     sync_frequency = request.form["sync_frequency"] if "sync_frequency" in request.form else None
-    
+
     file_path = request.form.get("file_path")
     directory_structure = request.form.get("directory_structure")
     file_name_map = request.form.get("file_name_map")
-    
+
     if directory_structure:
         try:
             directory_structure = json.loads(directory_structure)
@@ -89,8 +87,8 @@ def upload_index_files():
         file_name_map = None
 
     storage = StorageCreator.get_storage()
-    index_base_path = f"indexes/{id}"
-    
+    index_base_path = f"indexes/{source_id}"
+
     if settings.VECTOR_STORE == "faiss":
         if "file_faiss" not in request.files:
             logger.error("No file_faiss part")
@@ -111,46 +109,48 @@ def upload_index_files():
         storage.save_file(file_faiss, faiss_storage_path)
         storage.save_file(file_pkl, pkl_storage_path)
 
+    now = datetime.datetime.now(datetime.timezone.utc)
+    update_fields = {
+        "name": job_name,
+        "type": type,
+        "language": job_name,
+        "date": now,
+        "model": settings.EMBEDDINGS_NAME,
+        "tokens": tokens,
+        "retriever": retriever,
+        "remote_data": remote_data,
+        "sync_frequency": sync_frequency,
+        "file_path": file_path,
+        "directory_structure": directory_structure,
+    }
+    if file_name_map is not None:
+        update_fields["file_name_map"] = file_name_map
 
-    existing_entry = sources_collection.find_one({"_id": ObjectId(id)})
-    if existing_entry:
-        update_fields = {
-            "user": user,
-            "name": job_name,
-            "language": job_name,
-            "date": datetime.datetime.now(),
-            "model": settings.EMBEDDINGS_NAME,
-            "type": type,
-            "tokens": tokens,
-            "retriever": retriever,
-            "remote_data": remote_data,
-            "sync_frequency": sync_frequency,
-            "file_path": file_path,
-            "directory_structure": directory_structure,
-        }
-        if file_name_map is not None:
-            update_fields["file_name_map"] = file_name_map
-        sources_collection.update_one(
-            {"_id": ObjectId(id)},
-            {"$set": update_fields},
-        )
-    else:
-        insert_doc = {
-            "_id": ObjectId(id),
-            "user": user,
-            "name": job_name,
-            "language": job_name,
-            "date": datetime.datetime.now(),
-            "model": settings.EMBEDDINGS_NAME,
-            "type": type,
-            "tokens": tokens,
-            "retriever": retriever,
-            "remote_data": remote_data,
-            "sync_frequency": sync_frequency,
-            "file_path": file_path,
-            "directory_structure": directory_structure,
-        }
-        if file_name_map is not None:
-            insert_doc["file_name_map"] = file_name_map
-        sources_collection.insert_one(insert_doc)
+    with db_session() as conn:
+        repo = SourcesRepository(conn)
+        existing = None
+        if looks_like_uuid(source_id):
+            existing = repo.get(source_id, user)
+        if existing is None:
+            existing = repo.get_by_legacy_id(source_id, user)
+        if existing is not None:
+            repo.update(str(existing["id"]), user, update_fields)
+        else:
+            repo.create(
+                job_name,
+                source_id=source_id if looks_like_uuid(source_id) else None,
+                user_id=user,
+                type=type,
+                tokens=tokens,
+                retriever=retriever,
+                remote_data=remote_data,
+                sync_frequency=sync_frequency,
+                file_path=file_path,
+                directory_structure=directory_structure,
+                file_name_map=file_name_map,
+                language=job_name,
+                model=settings.EMBEDDINGS_NAME,
+                date=now,
+                legacy_mongo_id=None if looks_like_uuid(source_id) else str(source_id),
+            )
     return {"status": "ok"}

application/api/user/agents/folders.py

@@ -3,29 +3,50 @@ Agent folders management routes.
 Provides virtual folder organization for agents (Google Drive-like structure).
 """
 
-import datetime
-from bson.objectid import ObjectId
 from flask import current_app, jsonify, make_response, request
 from flask_restx import Namespace, Resource, fields
+from sqlalchemy import text as _sql_text
 
 from application.api import api
-from application.api.user.base import (
-    agent_folders_collection,
-    agents_collection,
-)
-from application.storage.db.dual_write import dual_write
+from application.storage.db.base_repository import looks_like_uuid
 from application.storage.db.repositories.agent_folders import AgentFoldersRepository
+from application.storage.db.repositories.agents import AgentsRepository
+from application.storage.db.session import db_readonly, db_session
+
 
 agents_folders_ns = Namespace(
     "agents_folders", description="Agent folder management", path="/api/agents/folders"
 )
 
 
+def _resolve_folder_id(repo: AgentFoldersRepository, folder_id: str, user: str):
+    """Resolve a folder id that may be either a UUID or legacy Mongo ObjectId."""
+    if not folder_id:
+        return None
+    if looks_like_uuid(folder_id):
+        row = repo.get(folder_id, user)
+        if row is not None:
+            return row
+    return repo.get_by_legacy_id(folder_id, user)
+
+
 def _folder_error_response(message: str, err: Exception):
     current_app.logger.error(f"{message}: {err}", exc_info=True)
     return make_response(jsonify({"success": False, "message": message}), 400)
 
 
+def _serialize_folder(f: dict) -> dict:
+    created_at = f.get("created_at")
+    updated_at = f.get("updated_at")
+    return {
+        "id": str(f["id"]),
+        "name": f.get("name"),
+        "parent_id": str(f["parent_id"]) if f.get("parent_id") else None,
+        "created_at": created_at.isoformat() if hasattr(created_at, "isoformat") else created_at,
+        "updated_at": updated_at.isoformat() if hasattr(updated_at, "isoformat") else updated_at,
+    }
+
+
 @agents_folders_ns.route("/")
 class AgentFolders(Resource):
     @api.doc(description="Get all folders for the user")
@@ -35,17 +56,9 @@ class AgentFolders(Resource):
             return make_response(jsonify({"success": False}), 401)
         user = decoded_token.get("sub")
         try:
-            folders = list(agent_folders_collection.find({"user": user}))
-            result = [
-                {
-                    "id": str(f["_id"]),
-                    "name": f["name"],
-                    "parent_id": f.get("parent_id"),
-                    "created_at": f.get("created_at", "").isoformat() if f.get("created_at") else None,
-                    "updated_at": f.get("updated_at", "").isoformat() if f.get("updated_at") else None,
-                }
-                for f in folders
-            ]
+            with db_readonly() as conn:
+                folders = AgentFoldersRepository(conn).list_for_user(user)
+            result = [_serialize_folder(f) for f in folders]
             return make_response(jsonify({"folders": result}), 200)
         except Exception as err:
             return _folder_error_response("Failed to fetch folders", err)
@@ -69,28 +82,34 @@ class AgentFolders(Resource):
         if not data or not data.get("name"):
             return make_response(jsonify({"success": False, "message": "Folder name is required"}), 400)
 
-        parent_id = data.get("parent_id")
-        if parent_id:
-            parent = agent_folders_collection.find_one({"_id": ObjectId(parent_id), "user": user})
-            if not parent:
-                return make_response(jsonify({"success": False, "message": "Parent folder not found"}), 404)
+        parent_id_input = data.get("parent_id")
+        description = data.get("description")
 
         try:
-            now = datetime.datetime.now(datetime.timezone.utc)
-            folder = {
-                "user": user,
-                "name": data["name"],
-                "parent_id": parent_id,
-                "created_at": now,
-                "updated_at": now,
-            }
-            result = agent_folders_collection.insert_one(folder)
-            dual_write(
-                AgentFoldersRepository,
-                lambda repo, u=user, n=data["name"]: repo.create(u, n),
-            )
+            with db_session() as conn:
+                repo = AgentFoldersRepository(conn)
+                pg_parent_id = None
+                if parent_id_input:
+                    parent = _resolve_folder_id(repo, parent_id_input, user)
+                    if not parent:
+                        return make_response(
+                            jsonify({"success": False, "message": "Parent folder not found"}),
+                            404,
+                        )
+                    pg_parent_id = str(parent["id"])
+                folder = repo.create(
+                    user, data["name"],
+                    description=description,
+                    parent_id=pg_parent_id,
+                )
             return make_response(
-                jsonify({"id": str(result.inserted_id), "name": data["name"], "parent_id": parent_id}),
+                jsonify(
+                    {
+                        "id": str(folder["id"]),
+                        "name": folder["name"],
+                        "parent_id": pg_parent_id,
+                    }
+                ),
                 201,
             )
         except Exception as err:
@@ -106,26 +125,51 @@ class AgentFolder(Resource):
             return make_response(jsonify({"success": False}), 401)
         user = decoded_token.get("sub")
         try:
-            folder = agent_folders_collection.find_one({"_id": ObjectId(folder_id), "user": user})
-            if not folder:
-                return make_response(jsonify({"success": False, "message": "Folder not found"}), 404)
-            
-            agents = list(agents_collection.find({"user": user, "folder_id": folder_id}))
-            agents_list = [
-                {"id": str(a["_id"]), "name": a["name"], "description": a.get("description", "")}
-                for a in agents
-            ]
-            subfolders = list(agent_folders_collection.find({"user": user, "parent_id": folder_id}))
-            subfolders_list = [{"id": str(sf["_id"]), "name": sf["name"]} for sf in subfolders]
+            with db_readonly() as conn:
+                folders_repo = AgentFoldersRepository(conn)
+                folder = _resolve_folder_id(folders_repo, folder_id, user)
+                if not folder:
+                    return make_response(
+                        jsonify({"success": False, "message": "Folder not found"}),
+                        404,
+                    )
+                pg_folder_id = str(folder["id"])
+
+                agents_rows = conn.execute(
+                    _sql_text(
+                        "SELECT id, name, description FROM agents "
+                        "WHERE user_id = :user_id AND folder_id = CAST(:fid AS uuid) "
+                        "ORDER BY created_at DESC"
+                    ),
+                    {"user_id": user, "fid": pg_folder_id},
+                ).fetchall()
+                agents_list = [
+                    {
+                        "id": str(row._mapping["id"]),
+                        "name": row._mapping["name"],
+                        "description": row._mapping.get("description", "") or "",
+                    }
+                    for row in agents_rows
+                ]
+
+                subfolders = folders_repo.list_children(pg_folder_id, user)
+                subfolders_list = [
+                    {"id": str(sf["id"]), "name": sf["name"]}
+                    for sf in subfolders
+                ]
 
             return make_response(
-                jsonify({
-                    "id": str(folder["_id"]),
-                    "name": folder["name"],
-                    "parent_id": folder.get("parent_id"),
-                    "agents": agents_list,
-                    "subfolders": subfolders_list,
-                }),
+                jsonify(
+                    {
+                        "id": pg_folder_id,
+                        "name": folder["name"],
+                        "parent_id": (
+                            str(folder["parent_id"]) if folder.get("parent_id") else None
+                        ),
+                        "agents": agents_list,
+                        "subfolders": subfolders_list,
+                    }
+                ),
                 200,
             )
         except Exception as err:
@@ -142,19 +186,57 @@ class AgentFolder(Resource):
             return make_response(jsonify({"success": False, "message": "No data provided"}), 400)
 
         try:
-            update_fields = {"updated_at": datetime.datetime.now(datetime.timezone.utc)}
-            if "name" in data:
-                update_fields["name"] = data["name"]
-            if "parent_id" in data:
-                if data["parent_id"] == folder_id:
-                    return make_response(jsonify({"success": False, "message": "Cannot set folder as its own parent"}), 400)
-                update_fields["parent_id"] = data["parent_id"]
+            with db_session() as conn:
+                repo = AgentFoldersRepository(conn)
+                folder = _resolve_folder_id(repo, folder_id, user)
+                if not folder:
+                    return make_response(
+                        jsonify({"success": False, "message": "Folder not found"}),
+                        404,
+                    )
+                pg_folder_id = str(folder["id"])
+
+                update_fields: dict = {}
+                if "name" in data:
+                    update_fields["name"] = data["name"]
+                if "description" in data:
+                    update_fields["description"] = data["description"]
+                if "parent_id" in data:
+                    parent_input = data.get("parent_id")
+                    if parent_input:
+                        if parent_input == folder_id or parent_input == pg_folder_id:
+                            return make_response(
+                                jsonify(
+                                    {
+                                        "success": False,
+                                        "message": "Cannot set folder as its own parent",
+                                    }
+                                ),
+                                400,
+                            )
+                        parent = _resolve_folder_id(repo, parent_input, user)
+                        if not parent:
+                            return make_response(
+                                jsonify({"success": False, "message": "Parent folder not found"}),
+                                404,
+                            )
+                        if str(parent["id"]) == pg_folder_id:
+                            return make_response(
+                                jsonify(
+                                    {
+                                        "success": False,
+                                        "message": "Cannot set folder as its own parent",
+                                    }
+                                ),
+                                400,
+                            )
+                        update_fields["parent_id"] = str(parent["id"])
+                    else:
+                        update_fields["parent_id"] = None
+
+                if update_fields:
+                    repo.update(pg_folder_id, user, update_fields)
 
-            result = agent_folders_collection.update_one(
-                {"_id": ObjectId(folder_id), "user": user}, {"$set": update_fields}
-            )
-            if result.matched_count == 0:
-                return make_response(jsonify({"success": False, "message": "Folder not found"}), 404)
             return make_response(jsonify({"success": True}), 200)
         except Exception as err:
             return _folder_error_response("Failed to update folder", err)
@@ -166,19 +248,24 @@ class AgentFolder(Resource):
             return make_response(jsonify({"success": False}), 401)
         user = decoded_token.get("sub")
         try:
-            agents_collection.update_many(
-                {"user": user, "folder_id": folder_id}, {"$unset": {"folder_id": ""}}
-            )
-            agent_folders_collection.update_many(
-                {"user": user, "parent_id": folder_id}, {"$unset": {"parent_id": ""}}
-            )
-            result = agent_folders_collection.delete_one({"_id": ObjectId(folder_id), "user": user})
-            dual_write(
-                AgentFoldersRepository,
-                lambda repo, fid=folder_id, u=user: repo.delete(fid, u),
-            )
-            if result.deleted_count == 0:
-                return make_response(jsonify({"success": False, "message": "Folder not found"}), 404)
+            with db_session() as conn:
+                repo = AgentFoldersRepository(conn)
+                folder = _resolve_folder_id(repo, folder_id, user)
+                if not folder:
+                    return make_response(
+                        jsonify({"success": False, "message": "Folder not found"}),
+                        404,
+                    )
+                pg_folder_id = str(folder["id"])
+                # Clear folder assignments from agents; self-FK
+                # ``ON DELETE SET NULL`` handles child folders.
+                AgentsRepository(conn).clear_folder_for_all(pg_folder_id, user)
+                deleted = repo.delete(pg_folder_id, user)
+            if not deleted:
+                return make_response(
+                    jsonify({"success": False, "message": "Folder not found"}),
+                    404,
+                )
             return make_response(jsonify({"success": True}), 200)
         except Exception as err:
             return _folder_error_response("Failed to delete folder", err)
@@ -205,26 +292,29 @@ class MoveAgentToFolder(Resource):
         if not data or not data.get("agent_id"):
             return make_response(jsonify({"success": False, "message": "Agent ID is required"}), 400)
 
-        agent_id = data["agent_id"]
-        folder_id = data.get("folder_id")
+        agent_id_input = data["agent_id"]
+        folder_id_input = data.get("folder_id")
 
         try:
-            agent = agents_collection.find_one({"_id": ObjectId(agent_id), "user": user})
-            if not agent:
-                return make_response(jsonify({"success": False, "message": "Agent not found"}), 404)
-
-            if folder_id:
-                folder = agent_folders_collection.find_one({"_id": ObjectId(folder_id), "user": user})
-                if not folder:
-                    return make_response(jsonify({"success": False, "message": "Folder not found"}), 404)
-                agents_collection.update_one(
-                    {"_id": ObjectId(agent_id)}, {"$set": {"folder_id": folder_id}}
-                )
-            else:
-                agents_collection.update_one(
-                    {"_id": ObjectId(agent_id)}, {"$unset": {"folder_id": ""}}
-                )
-
+            with db_session() as conn:
+                agents_repo = AgentsRepository(conn)
+                agent = agents_repo.get_any(agent_id_input, user)
+                if not agent:
+                    return make_response(
+                        jsonify({"success": False, "message": "Agent not found"}),
+                        404,
+                    )
+                pg_folder_id = None
+                if folder_id_input:
+                    folders_repo = AgentFoldersRepository(conn)
+                    folder = _resolve_folder_id(folders_repo, folder_id_input, user)
+                    if not folder:
+                        return make_response(
+                            jsonify({"success": False, "message": "Folder not found"}),
+                            404,
+                        )
+                    pg_folder_id = str(folder["id"])
+                agents_repo.set_folder(str(agent["id"]), user, pg_folder_id)
             return make_response(jsonify({"success": True}), 200)
         except Exception as err:
             return _folder_error_response("Failed to move agent", err)
@@ -252,25 +342,25 @@ class BulkMoveAgents(Resource):
             return make_response(jsonify({"success": False, "message": "Agent IDs are required"}), 400)
 
         agent_ids = data["agent_ids"]
-        folder_id = data.get("folder_id")
+        folder_id_input = data.get("folder_id")
 
         try:
-            if folder_id:
-                folder = agent_folders_collection.find_one({"_id": ObjectId(folder_id), "user": user})
-                if not folder:
-                    return make_response(jsonify({"success": False, "message": "Folder not found"}), 404)
-
-            object_ids = [ObjectId(aid) for aid in agent_ids]
-            if folder_id:
-                agents_collection.update_many(
-                    {"_id": {"$in": object_ids}, "user": user},
-                    {"$set": {"folder_id": folder_id}},
-                )
-            else:
-                agents_collection.update_many(
-                    {"_id": {"$in": object_ids}, "user": user},
-                    {"$unset": {"folder_id": ""}},
-                )
+            with db_session() as conn:
+                agents_repo = AgentsRepository(conn)
+                pg_folder_id = None
+                if folder_id_input:
+                    folders_repo = AgentFoldersRepository(conn)
+                    folder = _resolve_folder_id(folders_repo, folder_id_input, user)
+                    if not folder:
+                        return make_response(
+                            jsonify({"success": False, "message": "Folder not found"}),
+                            404,
+                        )
+                    pg_folder_id = str(folder["id"])
+                for agent_id_input in agent_ids:
+                    agent = agents_repo.get_any(agent_id_input, user)
+                    if agent is not None:
+                        agents_repo.set_folder(str(agent["id"]), user, pg_folder_id)
             return make_response(jsonify({"success": True}), 200)
         except Exception as err:
             return _folder_error_response("Failed to move agents", err)

application/api/user/agents/sharing.py

@@ -3,23 +3,17 @@
 import datetime
 import secrets
 
-from bson import DBRef
-from bson.objectid import ObjectId
 from flask import current_app, jsonify, make_response, request
 from flask_restx import fields, Namespace, Resource
+from sqlalchemy import text as _sql_text
 
 from application.api import api
 from application.core.settings import settings
-from application.api.user.base import (
-    agents_collection,
-    db,
-    ensure_user_doc,
-    resolve_tool_details,
-    user_tools_collection,
-    users_collection,
-)
-from application.storage.db.dual_write import dual_write
+from application.api.user.base import resolve_tool_details
+from application.storage.db.base_repository import looks_like_uuid
+from application.storage.db.repositories.agents import AgentsRepository
 from application.storage.db.repositories.users import UsersRepository
+from application.storage.db.session import db_readonly, db_session
 from application.utils import generate_image_url
 
 agents_sharing_ns = Namespace(
@@ -27,6 +21,38 @@ agents_sharing_ns = Namespace(
 )
 
 
+def _serialize_agent_basic(agent: dict) -> dict:
+    """Shape a PG agent row into the API response dict."""
+    source_id = agent.get("source_id")
+    return {
+        "id": str(agent["id"]),
+        "user": agent.get("user_id", ""),
+        "name": agent.get("name", ""),
+        "image": (
+            generate_image_url(agent["image"]) if agent.get("image") else ""
+        ),
+        "description": agent.get("description", ""),
+        "source": str(source_id) if source_id else "",
+        "chunks": str(agent["chunks"]) if agent.get("chunks") is not None else "0",
+        "retriever": agent.get("retriever", "classic") or "classic",
+        "prompt_id": str(agent["prompt_id"]) if agent.get("prompt_id") else "default",
+        "tools": agent.get("tools", []) or [],
+        "tool_details": resolve_tool_details(agent.get("tools", []) or []),
+        "agent_type": agent.get("agent_type", "") or "",
+        "status": agent.get("status", "") or "",
+        "json_schema": agent.get("json_schema"),
+        "limited_token_mode": agent.get("limited_token_mode", False),
+        "token_limit": agent.get("token_limit") or settings.DEFAULT_AGENT_LIMITS["token_limit"],
+        "limited_request_mode": agent.get("limited_request_mode", False),
+        "request_limit": agent.get("request_limit") or settings.DEFAULT_AGENT_LIMITS["request_limit"],
+        "created_at": agent.get("created_at", ""),
+        "updated_at": agent.get("updated_at", ""),
+        "shared": bool(agent.get("shared", False)),
+        "shared_token": agent.get("shared_token", "") or "",
+        "shared_metadata": agent.get("shared_metadata", {}) or {},
+    }
+
+
 @agents_sharing_ns.route("/shared_agent")
 class SharedAgent(Resource):
     @api.doc(
@@ -43,73 +69,33 @@ class SharedAgent(Resource):
                 jsonify({"success": False, "message": "Token or ID is required"}), 400
             )
         try:
-            query = {
-                "shared_publicly": True,
-                "shared_token": shared_token,
-            }
-            shared_agent = agents_collection.find_one(query)
+            with db_readonly() as conn:
+                shared_agent = AgentsRepository(conn).find_by_shared_token(
+                    shared_token,
+                )
             if not shared_agent:
                 return make_response(
                     jsonify({"success": False, "message": "Shared agent not found"}),
                     404,
                 )
-            agent_id = str(shared_agent["_id"])
-            data = {
-                "id": agent_id,
-                "user": shared_agent.get("user", ""),
-                "name": shared_agent.get("name", ""),
-                "image": (
-                    generate_image_url(shared_agent["image"])
-                    if shared_agent.get("image")
-                    else ""
-                ),
-                "description": shared_agent.get("description", ""),
-                "source": (
-                    str(source_doc["_id"])
-                    if isinstance(shared_agent.get("source"), DBRef)
-                    and (source_doc := db.dereference(shared_agent.get("source")))
-                    else ""
-                ),
-                "chunks": shared_agent.get("chunks", "0"),
-                "retriever": shared_agent.get("retriever", "classic"),
-                "prompt_id": shared_agent.get("prompt_id", "default"),
-                "tools": shared_agent.get("tools", []),
-                "tool_details": resolve_tool_details(shared_agent.get("tools", [])),
-                "agent_type": shared_agent.get("agent_type", ""),
-                "status": shared_agent.get("status", ""),
-                "json_schema": shared_agent.get("json_schema"),
-                "limited_token_mode": shared_agent.get("limited_token_mode", False),
-                "token_limit": shared_agent.get("token_limit", settings.DEFAULT_AGENT_LIMITS["token_limit"]),
-                "limited_request_mode": shared_agent.get("limited_request_mode", False),
-                "request_limit": shared_agent.get("request_limit", settings.DEFAULT_AGENT_LIMITS["request_limit"]),
-                "created_at": shared_agent.get("createdAt", ""),
-                "updated_at": shared_agent.get("updatedAt", ""),
-                "shared": shared_agent.get("shared_publicly", False),
-                "shared_token": shared_agent.get("shared_token", ""),
-                "shared_metadata": shared_agent.get("shared_metadata", {}),
-            }
+            agent_id = str(shared_agent["id"])
+            data = _serialize_agent_basic(shared_agent)
 
             if data["tools"]:
                 enriched_tools = []
-                for tool in data["tools"]:
-                    tool_data = user_tools_collection.find_one({"_id": ObjectId(tool)})
-                    if tool_data:
-                        enriched_tools.append(tool_data.get("name", ""))
+                for detail in data["tool_details"]:
+                    enriched_tools.append(detail.get("name", ""))
                 data["tools"] = enriched_tools
             decoded_token = getattr(request, "decoded_token", None)
             if decoded_token:
                 user_id = decoded_token.get("sub")
-                owner_id = shared_agent.get("user")
+                owner_id = shared_agent.get("user_id")
 
                 if user_id != owner_id:
-                    ensure_user_doc(user_id)
-                    users_collection.update_one(
-                        {"user_id": user_id},
-                        {"$addToSet": {"agent_preferences.shared_with_me": agent_id}},
-                    )
-                    dual_write(UsersRepository,
-                        lambda repo, uid=user_id, aid=agent_id: repo.add_shared(uid, aid)
-                    )
+                    with db_session() as conn:
+                        users_repo = UsersRepository(conn)
+                        users_repo.upsert(user_id)
+                        users_repo.add_shared(user_id, agent_id)
             return make_response(jsonify(data), 200)
         except Exception as err:
             current_app.logger.error(f"Error retrieving shared agent: {err}")
@@ -126,55 +112,73 @@ class SharedAgents(Resource):
                 return make_response(jsonify({"success": False}), 401)
             user_id = decoded_token.get("sub")
 
-            user_doc = ensure_user_doc(user_id)
-            shared_with_ids = user_doc.get("agent_preferences", {}).get(
-                "shared_with_me", []
-            )
-            shared_object_ids = [ObjectId(id) for id in shared_with_ids]
-
-            shared_agents_cursor = agents_collection.find(
-                {"_id": {"$in": shared_object_ids}, "shared_publicly": True}
-            )
-            shared_agents = list(shared_agents_cursor)
-
-            found_ids_set = {str(agent["_id"]) for agent in shared_agents}
-            stale_ids = [id for id in shared_with_ids if id not in found_ids_set]
-            if stale_ids:
-                users_collection.update_one(
-                    {"user_id": user_id},
-                    {"$pullAll": {"agent_preferences.shared_with_me": stale_ids}},
+            with db_session() as conn:
+                users_repo = UsersRepository(conn)
+                user_doc = users_repo.upsert(user_id)
+                shared_with_ids = (
+                    user_doc.get("agent_preferences", {}).get("shared_with_me", [])
+                    if isinstance(user_doc.get("agent_preferences"), dict)
+                    else []
                 )
-                dual_write(UsersRepository,
-                    lambda repo, uid=user_id, ids=stale_ids: repo.remove_shared_bulk(uid, ids)
-                )
-            pinned_ids = set(user_doc.get("agent_preferences", {}).get("pinned", []))
+                # Keep only UUID-shaped ids; ObjectId leftovers are stripped below.
+                uuid_ids = [sid for sid in shared_with_ids if looks_like_uuid(sid)]
+                non_uuid_ids = [sid for sid in shared_with_ids if not looks_like_uuid(sid)]
 
-            list_shared_agents = [
-                {
-                    "id": str(agent["_id"]),
-                    "name": agent.get("name", ""),
-                    "description": agent.get("description", ""),
-                    "image": (
-                        generate_image_url(agent["image"]) if agent.get("image") else ""
-                    ),
-                    "tools": agent.get("tools", []),
-                    "tool_details": resolve_tool_details(agent.get("tools", [])),
-                    "agent_type": agent.get("agent_type", ""),
-                    "status": agent.get("status", ""),
-                    "json_schema": agent.get("json_schema"),
-                    "limited_token_mode": agent.get("limited_token_mode", False),
-                    "token_limit": agent.get("token_limit", settings.DEFAULT_AGENT_LIMITS["token_limit"]),
-                    "limited_request_mode": agent.get("limited_request_mode", False),
-                    "request_limit": agent.get("request_limit", settings.DEFAULT_AGENT_LIMITS["request_limit"]),
-                    "created_at": agent.get("createdAt", ""),
-                    "updated_at": agent.get("updatedAt", ""),
-                    "pinned": str(agent["_id"]) in pinned_ids,
-                    "shared": agent.get("shared_publicly", False),
-                    "shared_token": agent.get("shared_token", ""),
-                    "shared_metadata": agent.get("shared_metadata", {}),
-                }
-                for agent in shared_agents
-            ]
+                if uuid_ids:
+                    result = conn.execute(
+                        _sql_text(
+                            "SELECT * FROM agents "
+                            "WHERE id = ANY(CAST(:ids AS uuid[])) "
+                            "AND shared = true"
+                        ),
+                        {"ids": uuid_ids},
+                    )
+                    shared_agents = [dict(row._mapping) for row in result.fetchall()]
+                else:
+                    shared_agents = []
+
+                found_ids_set = {str(agent["id"]) for agent in shared_agents}
+                stale_ids = [sid for sid in uuid_ids if sid not in found_ids_set]
+                stale_ids.extend(non_uuid_ids)
+                if stale_ids:
+                    users_repo.remove_shared_bulk(user_id, stale_ids)
+
+                pinned_ids = set(
+                    user_doc.get("agent_preferences", {}).get("pinned", [])
+                    if isinstance(user_doc.get("agent_preferences"), dict)
+                    else []
+                )
+
+            list_shared_agents = []
+            for agent in shared_agents:
+                agent_id_str = str(agent["id"])
+                list_shared_agents.append(
+                    {
+                        "id": agent_id_str,
+                        "name": agent.get("name", ""),
+                        "description": agent.get("description", ""),
+                        "image": (
+                            generate_image_url(agent["image"]) if agent.get("image") else ""
+                        ),
+                        "tools": agent.get("tools", []) or [],
+                        "tool_details": resolve_tool_details(
+                            agent.get("tools", []) or []
+                        ),
+                        "agent_type": agent.get("agent_type", "") or "",
+                        "status": agent.get("status", "") or "",
+                        "json_schema": agent.get("json_schema"),
+                        "limited_token_mode": agent.get("limited_token_mode", False),
+                        "token_limit": agent.get("token_limit") or settings.DEFAULT_AGENT_LIMITS["token_limit"],
+                        "limited_request_mode": agent.get("limited_request_mode", False),
+                        "request_limit": agent.get("request_limit") or settings.DEFAULT_AGENT_LIMITS["request_limit"],
+                        "created_at": agent.get("created_at", ""),
+                        "updated_at": agent.get("updated_at", ""),
+                        "pinned": agent_id_str in pinned_ids,
+                        "shared": bool(agent.get("shared", False)),
+                        "shared_token": agent.get("shared_token", "") or "",
+                        "shared_metadata": agent.get("shared_metadata", {}) or {},
+                    }
+                )
 
             return make_response(jsonify(list_shared_agents), 200)
         except Exception as err:
@@ -228,44 +232,43 @@ class ShareAgent(Resource):
                 ),
                 400,
             )
+        shared_token = None
         try:
-            try:
-                agent_oid = ObjectId(agent_id)
-            except Exception:
-                return make_response(
-                    jsonify({"success": False, "message": "Invalid agent ID"}), 400
-                )
-            agent = agents_collection.find_one({"_id": agent_oid, "user": user})
-            if not agent:
-                return make_response(
-                    jsonify({"success": False, "message": "Agent not found"}), 404
-                )
-            if shared:
-                shared_metadata = {
-                    "shared_by": username,
-                    "shared_at": datetime.datetime.now(datetime.timezone.utc),
-                }
-                shared_token = secrets.token_urlsafe(32)
-                agents_collection.update_one(
-                    {"_id": agent_oid, "user": user},
-                    {
-                        "$set": {
-                            "shared_publicly": shared,
-                            "shared_metadata": shared_metadata,
+            with db_session() as conn:
+                repo = AgentsRepository(conn)
+                agent = repo.get_any(agent_id, user)
+                if not agent:
+                    return make_response(
+                        jsonify({"success": False, "message": "Agent not found"}), 404
+                    )
+                if shared:
+                    shared_metadata = {
+                        "shared_by": username,
+                        "shared_at": datetime.datetime.now(
+                            datetime.timezone.utc
+                        ).isoformat(),
+                    }
+                    shared_token = secrets.token_urlsafe(32)
+                    repo.update(
+                        str(agent["id"]), user,
+                        {
+                            "shared": True,
                             "shared_token": shared_token,
-                        }
-                    },
-                )
-            else:
-                agents_collection.update_one(
-                    {"_id": agent_oid, "user": user},
-                    {"$set": {"shared_publicly": shared, "shared_token": None}},
-                    {"$unset": {"shared_metadata": ""}},
-                )
+                            "shared_metadata": shared_metadata,
+                        },
+                    )
+                else:
+                    repo.update(
+                        str(agent["id"]), user,
+                        {
+                            "shared": False,
+                            "shared_token": None,
+                            "shared_metadata": None,
+                        },
+                    )
         except Exception as err:
             current_app.logger.error(f"Error sharing/unsharing agent: {err}", exc_info=True)
             return make_response(jsonify({"success": False, "error": "Failed to update agent sharing status"}), 400)
-        shared_token = shared_token if shared else None
         return make_response(
             jsonify({"success": True, "shared_token": shared_token}), 200
         )

application/api/user/agents/webhooks.py

@@ -2,14 +2,15 @@
 
 import secrets
 
-from bson.objectid import ObjectId
 from flask import current_app, jsonify, make_response, request
 from flask_restx import Namespace, Resource
 
 from application.api import api
-from application.api.user.base import agents_collection, require_agent
+from application.api.user.base import require_agent
 from application.api.user.tasks import process_agent_webhook
 from application.core.settings import settings
+from application.storage.db.repositories.agents import AgentsRepository
+from application.storage.db.session import db_readonly, db_session
 
 
 agents_webhooks_ns = Namespace(
@@ -34,9 +35,8 @@ class AgentWebhook(Resource):
                 jsonify({"success": False, "message": "ID is required"}), 400
             )
         try:
-            agent = agents_collection.find_one(
-                {"_id": ObjectId(agent_id), "user": user}
-            )
+            with db_readonly() as conn:
+                agent = AgentsRepository(conn).get_any(agent_id, user)
             if not agent:
                 return make_response(
                     jsonify({"success": False, "message": "Agent not found"}), 404
@@ -44,10 +44,11 @@ class AgentWebhook(Resource):
             webhook_token = agent.get("incoming_webhook_token")
             if not webhook_token:
                 webhook_token = secrets.token_urlsafe(32)
-                agents_collection.update_one(
-                    {"_id": ObjectId(agent_id), "user": user},
-                    {"$set": {"incoming_webhook_token": webhook_token}},
-                )
+                with db_session() as conn:
+                    AgentsRepository(conn).update(
+                        str(agent["id"]), user,
+                        {"incoming_webhook_token": webhook_token},
+                    )
             base_url = settings.API_URL.rstrip("/")
             full_webhook_url = f"{base_url}/api/webhooks/agents/{webhook_token}"
         except Exception as err:

application/api/user/analytics/routes.py

@@ -2,26 +2,84 @@
 
 import datetime
 
-from bson.objectid import ObjectId
 from flask import current_app, jsonify, make_response, request
 from flask_restx import fields, Namespace, Resource
+from sqlalchemy import text as _sql_text
 
 from application.api import api
 from application.api.user.base import (
-    agents_collection,
-    conversations_collection,
     generate_date_range,
     generate_hourly_range,
     generate_minute_range,
-    token_usage_collection,
-    user_logs_collection,
 )
+from application.storage.db.repositories.agents import AgentsRepository
+from application.storage.db.repositories.token_usage import TokenUsageRepository
+from application.storage.db.repositories.user_logs import UserLogsRepository
+from application.storage.db.session import db_readonly
+
 
 analytics_ns = Namespace(
     "analytics", description="Analytics and reporting operations", path="/api"
 )
 
 
+_FILTER_BUCKETS = {
+    "last_hour": ("minute", "%Y-%m-%d %H:%M:00", "YYYY-MM-DD HH24:MI:00"),
+    "last_24_hour": ("hour", "%Y-%m-%d %H:00", "YYYY-MM-DD HH24:00"),
+    "last_7_days": ("day", "%Y-%m-%d", "YYYY-MM-DD"),
+    "last_15_days": ("day", "%Y-%m-%d", "YYYY-MM-DD"),
+    "last_30_days": ("day", "%Y-%m-%d", "YYYY-MM-DD"),
+}
+
+
+def _range_for_filter(filter_option: str):
+    """Return ``(start_date, end_date, bucket_unit, pg_fmt)`` for the filter.
+
+    Returns ``None`` on invalid filter.
+    """
+    if filter_option not in _FILTER_BUCKETS:
+        return None
+    end_date = datetime.datetime.now(datetime.timezone.utc)
+    bucket_unit, _py_fmt, pg_fmt = _FILTER_BUCKETS[filter_option]
+
+    if filter_option == "last_hour":
+        start_date = end_date - datetime.timedelta(hours=1)
+    elif filter_option == "last_24_hour":
+        start_date = end_date - datetime.timedelta(hours=24)
+    else:
+        days = {
+            "last_7_days": 6,
+            "last_15_days": 14,
+            "last_30_days": 29,
+        }[filter_option]
+        start_date = end_date - datetime.timedelta(days=days)
+        start_date = start_date.replace(hour=0, minute=0, second=0, microsecond=0)
+        end_date = end_date.replace(
+            hour=23, minute=59, second=59, microsecond=999999
+        )
+    return start_date, end_date, bucket_unit, pg_fmt
+
+
+def _intervals_for_filter(filter_option, start_date, end_date):
+    if filter_option == "last_hour":
+        return generate_minute_range(start_date, end_date)
+    if filter_option == "last_24_hour":
+        return generate_hourly_range(start_date, end_date)
+    return generate_date_range(start_date, end_date)
+
+
+def _resolve_api_key(conn, api_key_id, user_id):
+    """Look up the ``agents.key`` value for a given agent id.
+
+    Scoped by ``user_id`` so an authenticated caller can't probe another
+    user's agents. Accepts either UUID or legacy Mongo ObjectId shape.
+    """
+    if not api_key_id:
+        return None
+    agent = AgentsRepository(conn).get_any(api_key_id, user_id)
+    return (agent or {}).get("key") if agent else None
+
+
 @analytics_ns.route("/get_message_analytics")
 class GetMessageAnalytics(Resource):
     get_message_analytics_model = api.model(
@@ -32,13 +90,7 @@ class GetMessageAnalytics(Resource):
                 required=False,
                 description="Filter option for analytics",
                 default="last_30_days",
-                enum=[
-                    "last_hour",
-                    "last_24_hour",
-                    "last_7_days",
-                    "last_15_days",
-                    "last_30_days",
-                ],
+                enum=list(_FILTER_BUCKETS.keys()),
             ),
         },
     )
@@ -50,88 +102,54 @@ class GetMessageAnalytics(Resource):
         if not decoded_token:
             return make_response(jsonify({"success": False}), 401)
         user = decoded_token.get("sub")
-        data = request.get_json()
+        data = request.get_json() or {}
         api_key_id = data.get("api_key_id")
         filter_option = data.get("filter_option", "last_30_days")
 
+        window = _range_for_filter(filter_option)
+        if window is None:
+            return make_response(
+                jsonify({"success": False, "message": "Invalid option"}), 400
+            )
+        start_date, end_date, _bucket_unit, pg_fmt = window
+
         try:
-            api_key = (
-                agents_collection.find_one({"_id": ObjectId(api_key_id), "user": user})[
-                    "key"
+            with db_readonly() as conn:
+                api_key = _resolve_api_key(conn, api_key_id, user)
+
+                # Count messages per bucket, filtered by the conversation's
+                # owner (user_id) and optionally the agent api_key. The
+                # ``user_id`` filter is always applied post-cutover to
+                # prevent cross-tenant leakage on admin dashboards.
+                clauses = [
+                    "c.user_id = :user_id",
+                    "m.timestamp >= :start",
+                    "m.timestamp <= :end",
                 ]
-                if api_key_id
-                else None
-            )
-        except Exception as err:
-            current_app.logger.error(f"Error getting API key: {err}", exc_info=True)
-            return make_response(jsonify({"success": False}), 400)
-        end_date = datetime.datetime.now(datetime.timezone.utc)
-
-        if filter_option == "last_hour":
-            start_date = end_date - datetime.timedelta(hours=1)
-            group_format = "%Y-%m-%d %H:%M:00"
-        elif filter_option == "last_24_hour":
-            start_date = end_date - datetime.timedelta(hours=24)
-            group_format = "%Y-%m-%d %H:00"
-        else:
-            if filter_option in ["last_7_days", "last_15_days", "last_30_days"]:
-                filter_days = (
-                    6
-                    if filter_option == "last_7_days"
-                    else 14 if filter_option == "last_15_days" else 29
-                )
-            else:
-                return make_response(
-                    jsonify({"success": False, "message": "Invalid option"}), 400
-                )
-            start_date = end_date - datetime.timedelta(days=filter_days)
-            start_date = start_date.replace(hour=0, minute=0, second=0, microsecond=0)
-            end_date = end_date.replace(
-                hour=23, minute=59, second=59, microsecond=999999
-            )
-            group_format = "%Y-%m-%d"
-        try:
-            match_stage = {
-                "$match": {
-                    "user": user,
+                params: dict = {
+                    "user_id": user,
+                    "start": start_date,
+                    "end": end_date,
+                    "fmt": pg_fmt,
                 }
-            }
-            if api_key:
-                match_stage["$match"]["api_key"] = api_key
-            pipeline = [
-                match_stage,
-                {"$unwind": "$queries"},
-                {
-                    "$match": {
-                        "queries.timestamp": {"$gte": start_date, "$lte": end_date}
-                    }
-                },
-                {
-                    "$group": {
-                        "_id": {
-                            "$dateToString": {
-                                "format": group_format,
-                                "date": "$queries.timestamp",
-                            }
-                        },
-                        "count": {"$sum": 1},
-                    }
-                },
-                {"$sort": {"_id": 1}},
-            ]
+                if api_key:
+                    clauses.append("c.api_key = :api_key")
+                    params["api_key"] = api_key
+                where = " AND ".join(clauses)
+                sql = (
+                    "SELECT to_char(m.timestamp AT TIME ZONE 'UTC', :fmt) AS bucket, "
+                    "COUNT(*) AS count "
+                    "FROM conversation_messages m "
+                    "JOIN conversations c ON c.id = m.conversation_id "
+                    f"WHERE {where} "
+                    "GROUP BY bucket ORDER BY bucket ASC"
+                )
+                rows = conn.execute(_sql_text(sql), params).fetchall()
 
-            message_data = conversations_collection.aggregate(pipeline)
-
-            if filter_option == "last_hour":
-                intervals = generate_minute_range(start_date, end_date)
-            elif filter_option == "last_24_hour":
-                intervals = generate_hourly_range(start_date, end_date)
-            else:
-                intervals = generate_date_range(start_date, end_date)
+            intervals = _intervals_for_filter(filter_option, start_date, end_date)
             daily_messages = {interval: 0 for interval in intervals}
-
-            for entry in message_data:
-                daily_messages[entry["_id"]] = entry["count"]
+            for row in rows:
+                daily_messages[row._mapping["bucket"]] = int(row._mapping["count"])
         except Exception as err:
             current_app.logger.error(
                 f"Error getting message analytics: {err}", exc_info=True
@@ -152,13 +170,7 @@ class GetTokenAnalytics(Resource):
                 required=False,
                 description="Filter option for analytics",
                 default="last_30_days",
-                enum=[
-                    "last_hour",
-                    "last_24_hour",
-                    "last_7_days",
-                    "last_15_days",
-                    "last_30_days",
-                ],
+                enum=list(_FILTER_BUCKETS.keys()),
             ),
         },
     )
@@ -170,123 +182,36 @@ class GetTokenAnalytics(Resource):
         if not decoded_token:
             return make_response(jsonify({"success": False}), 401)
         user = decoded_token.get("sub")
-        data = request.get_json()
+        data = request.get_json() or {}
         api_key_id = data.get("api_key_id")
         filter_option = data.get("filter_option", "last_30_days")
 
-        try:
-            api_key = (
-                agents_collection.find_one({"_id": ObjectId(api_key_id), "user": user})[
-                    "key"
-                ]
-                if api_key_id
-                else None
+        window = _range_for_filter(filter_option)
+        if window is None:
+            return make_response(
+                jsonify({"success": False, "message": "Invalid option"}), 400
             )
-        except Exception as err:
-            current_app.logger.error(f"Error getting API key: {err}", exc_info=True)
-            return make_response(jsonify({"success": False}), 400)
-        end_date = datetime.datetime.now(datetime.timezone.utc)
+        start_date, end_date, bucket_unit, _pg_fmt = window
 
-        if filter_option == "last_hour":
-            start_date = end_date - datetime.timedelta(hours=1)
-            group_format = "%Y-%m-%d %H:%M:00"
-            group_stage = {
-                "$group": {
-                    "_id": {
-                        "minute": {
-                            "$dateToString": {
-                                "format": group_format,
-                                "date": "$timestamp",
-                            }
-                        }
-                    },
-                    "total_tokens": {
-                        "$sum": {"$add": ["$prompt_tokens", "$generated_tokens"]}
-                    },
-                }
-            }
-        elif filter_option == "last_24_hour":
-            start_date = end_date - datetime.timedelta(hours=24)
-            group_format = "%Y-%m-%d %H:00"
-            group_stage = {
-                "$group": {
-                    "_id": {
-                        "hour": {
-                            "$dateToString": {
-                                "format": group_format,
-                                "date": "$timestamp",
-                            }
-                        }
-                    },
-                    "total_tokens": {
-                        "$sum": {"$add": ["$prompt_tokens", "$generated_tokens"]}
-                    },
-                }
-            }
-        else:
-            if filter_option in ["last_7_days", "last_15_days", "last_30_days"]:
-                filter_days = (
-                    6
-                    if filter_option == "last_7_days"
-                    else (14 if filter_option == "last_15_days" else 29)
-                )
-            else:
-                return make_response(
-                    jsonify({"success": False, "message": "Invalid option"}), 400
-                )
-            start_date = end_date - datetime.timedelta(days=filter_days)
-            start_date = start_date.replace(hour=0, minute=0, second=0, microsecond=0)
-            end_date = end_date.replace(
-                hour=23, minute=59, second=59, microsecond=999999
-            )
-            group_format = "%Y-%m-%d"
-            group_stage = {
-                "$group": {
-                    "_id": {
-                        "day": {
-                            "$dateToString": {
-                                "format": group_format,
-                                "date": "$timestamp",
-                            }
-                        }
-                    },
-                    "total_tokens": {
-                        "$sum": {"$add": ["$prompt_tokens", "$generated_tokens"]}
-                    },
-                }
-            }
         try:
-            match_stage = {
-                "$match": {
-                    "user_id": user,
-                    "timestamp": {"$gte": start_date, "$lte": end_date},
-                }
-            }
-            if api_key:
-                match_stage["$match"]["api_key"] = api_key
-            token_usage_data = token_usage_collection.aggregate(
-                [
-                    match_stage,
-                    group_stage,
-                    {"$sort": {"_id": 1}},
-                ]
-            )
+            with db_readonly() as conn:
+                api_key = _resolve_api_key(conn, api_key_id, user)
+                # ``bucketed_totals`` applies user_id / api_key filters
+                # directly — no need to reshape a Mongo pipeline.
+                rows = TokenUsageRepository(conn).bucketed_totals(
+                    bucket_unit=bucket_unit,
+                    user_id=user,
+                    api_key=api_key,
+                    timestamp_gte=start_date,
+                    timestamp_lt=end_date,
+                )
 
-            if filter_option == "last_hour":
-                intervals = generate_minute_range(start_date, end_date)
-            elif filter_option == "last_24_hour":
-                intervals = generate_hourly_range(start_date, end_date)
-            else:
-                intervals = generate_date_range(start_date, end_date)
+            intervals = _intervals_for_filter(filter_option, start_date, end_date)
             daily_token_usage = {interval: 0 for interval in intervals}
-
-            for entry in token_usage_data:
-                if filter_option == "last_hour":
-                    daily_token_usage[entry["_id"]["minute"]] = entry["total_tokens"]
-                elif filter_option == "last_24_hour":
-                    daily_token_usage[entry["_id"]["hour"]] = entry["total_tokens"]
-                else:
-                    daily_token_usage[entry["_id"]["day"]] = entry["total_tokens"]
+            for entry in rows:
+                daily_token_usage[entry["bucket"]] = int(
+                    entry["prompt_tokens"] + entry["generated_tokens"]
+                )
         except Exception as err:
             current_app.logger.error(
                 f"Error getting token analytics: {err}", exc_info=True
@@ -307,13 +232,7 @@ class GetFeedbackAnalytics(Resource):
                 required=False,
                 description="Filter option for analytics",
                 default="last_30_days",
-                enum=[
-                    "last_hour",
-                    "last_24_hour",
-                    "last_7_days",
-                    "last_15_days",
-                    "last_30_days",
-                ],
+                enum=list(_FILTER_BUCKETS.keys()),
             ),
         },
     )
@@ -325,128 +244,64 @@ class GetFeedbackAnalytics(Resource):
         if not decoded_token:
             return make_response(jsonify({"success": False}), 401)
         user = decoded_token.get("sub")
-        data = request.get_json()
+        data = request.get_json() or {}
         api_key_id = data.get("api_key_id")
         filter_option = data.get("filter_option", "last_30_days")
 
+        window = _range_for_filter(filter_option)
+        if window is None:
+            return make_response(
+                jsonify({"success": False, "message": "Invalid option"}), 400
+            )
+        start_date, end_date, _bucket_unit, pg_fmt = window
+
         try:
-            api_key = (
-                agents_collection.find_one({"_id": ObjectId(api_key_id), "user": user})[
-                    "key"
+            with db_readonly() as conn:
+                api_key = _resolve_api_key(conn, api_key_id, user)
+
+                # Feedback lives inside the ``conversation_messages.feedback``
+                # JSONB as ``{"text": "like"|"dislike", "timestamp": "..."}``.
+                # There is no scalar ``feedback_timestamp`` column — extract
+                # the timestamp from the JSONB and cast it to timestamptz for
+                # the range filter + bucket grouping.
+                clauses = [
+                    "c.user_id = :user_id",
+                    "m.feedback IS NOT NULL",
+                    "(m.feedback->>'timestamp')::timestamptz >= :start",
+                    "(m.feedback->>'timestamp')::timestamptz <= :end",
                 ]
-                if api_key_id
-                else None
-            )
-        except Exception as err:
-            current_app.logger.error(f"Error getting API key: {err}", exc_info=True)
-            return make_response(jsonify({"success": False}), 400)
-        end_date = datetime.datetime.now(datetime.timezone.utc)
-
-        if filter_option == "last_hour":
-            start_date = end_date - datetime.timedelta(hours=1)
-            group_format = "%Y-%m-%d %H:%M:00"
-            date_field = {
-                "$dateToString": {
-                    "format": group_format,
-                    "date": "$queries.feedback_timestamp",
+                params: dict = {
+                    "user_id": user,
+                    "start": start_date,
+                    "end": end_date,
+                    "fmt": pg_fmt,
                 }
-            }
-        elif filter_option == "last_24_hour":
-            start_date = end_date - datetime.timedelta(hours=24)
-            group_format = "%Y-%m-%d %H:00"
-            date_field = {
-                "$dateToString": {
-                    "format": group_format,
-                    "date": "$queries.feedback_timestamp",
-                }
-            }
-        else:
-            if filter_option in ["last_7_days", "last_15_days", "last_30_days"]:
-                filter_days = (
-                    6
-                    if filter_option == "last_7_days"
-                    else (14 if filter_option == "last_15_days" else 29)
+                if api_key:
+                    clauses.append("c.api_key = :api_key")
+                    params["api_key"] = api_key
+                where = " AND ".join(clauses)
+                sql = (
+                    "SELECT to_char("
+                    "(m.feedback->>'timestamp')::timestamptz AT TIME ZONE 'UTC', :fmt"
+                    ") AS bucket, "
+                    "SUM(CASE WHEN m.feedback->>'text' = 'like' THEN 1 ELSE 0 END) AS positive, "
+                    "SUM(CASE WHEN m.feedback->>'text' = 'dislike' THEN 1 ELSE 0 END) AS negative "
+                    "FROM conversation_messages m "
+                    "JOIN conversations c ON c.id = m.conversation_id "
+                    f"WHERE {where} "
+                    "GROUP BY bucket ORDER BY bucket ASC"
                 )
-            else:
-                return make_response(
-                    jsonify({"success": False, "message": "Invalid option"}), 400
-                )
-            start_date = end_date - datetime.timedelta(days=filter_days)
-            start_date = start_date.replace(hour=0, minute=0, second=0, microsecond=0)
-            end_date = end_date.replace(
-                hour=23, minute=59, second=59, microsecond=999999
-            )
-            group_format = "%Y-%m-%d"
-            date_field = {
-                "$dateToString": {
-                    "format": group_format,
-                    "date": "$queries.feedback_timestamp",
-                }
-            }
-        try:
-            match_stage = {
-                "$match": {
-                    "queries.feedback_timestamp": {
-                        "$gte": start_date,
-                        "$lte": end_date,
-                    },
-                    "queries.feedback": {"$exists": True},
-                }
-            }
-            if api_key:
-                match_stage["$match"]["api_key"] = api_key
-            pipeline = [
-                match_stage,
-                {"$unwind": "$queries"},
-                {"$match": {"queries.feedback": {"$exists": True}}},
-                {
-                    "$group": {
-                        "_id": {"time": date_field, "feedback": "$queries.feedback"},
-                        "count": {"$sum": 1},
-                    }
-                },
-                {
-                    "$group": {
-                        "_id": "$_id.time",
-                        "positive": {
-                            "$sum": {
-                                "$cond": [
-                                    {"$eq": ["$_id.feedback", "LIKE"]},
-                                    "$count",
-                                    0,
-                                ]
-                            }
-                        },
-                        "negative": {
-                            "$sum": {
-                                "$cond": [
-                                    {"$eq": ["$_id.feedback", "DISLIKE"]},
-                                    "$count",
-                                    0,
-                                ]
-                            }
-                        },
-                    }
-                },
-                {"$sort": {"_id": 1}},
-            ]
+                rows = conn.execute(_sql_text(sql), params).fetchall()
 
-            feedback_data = conversations_collection.aggregate(pipeline)
-
-            if filter_option == "last_hour":
-                intervals = generate_minute_range(start_date, end_date)
-            elif filter_option == "last_24_hour":
-                intervals = generate_hourly_range(start_date, end_date)
-            else:
-                intervals = generate_date_range(start_date, end_date)
+            intervals = _intervals_for_filter(filter_option, start_date, end_date)
             daily_feedback = {
                 interval: {"positive": 0, "negative": 0} for interval in intervals
             }
-
-            for entry in feedback_data:
-                daily_feedback[entry["_id"]] = {
-                    "positive": entry["positive"],
-                    "negative": entry["negative"],
+            for row in rows:
+                bucket = row._mapping["bucket"]
+                daily_feedback[bucket] = {
+                    "positive": int(row._mapping["positive"] or 0),
+                    "negative": int(row._mapping["negative"] or 0),
                 }
         except Exception as err:
             current_app.logger.error(
@@ -484,47 +339,89 @@ class GetUserLogs(Resource):
         if not decoded_token:
             return make_response(jsonify({"success": False}), 401)
         user = decoded_token.get("sub")
-        data = request.get_json()
+        data = request.get_json() or {}
         page = int(data.get("page", 1))
         api_key_id = data.get("api_key_id")
         page_size = int(data.get("page_size", 10))
-        skip = (page - 1) * page_size
 
         try:
-            api_key = (
-                agents_collection.find_one({"_id": ObjectId(api_key_id)})["key"]
-                if api_key_id
-                else None
-            )
+            with db_readonly() as conn:
+                api_key = _resolve_api_key(conn, api_key_id, user)
+                logs_repo = UserLogsRepository(conn)
+                if api_key:
+                    # ``find_by_api_key`` filters on ``data->>'api_key'``
+                    # — the PG shape of the legacy top-level ``api_key``
+                    # filter. Paginate client-side using offset/limit.
+                    all_rows = logs_repo.find_by_api_key(api_key)
+                    offset = (page - 1) * page_size
+                    window = all_rows[offset: offset + page_size + 1]
+                    items = window
+                else:
+                    items, has_more_flag = logs_repo.list_paginated(
+                        user_id=user,
+                        page=page,
+                        page_size=page_size,
+                    )
+                    # list_paginated already trims to page_size and
+                    # returns has_more separately.
+                    results = [
+                        {
+                            "id": str(item.get("id") or item.get("_id")),
+                            "action": (item.get("data") or {}).get("action"),
+                            "level": (item.get("data") or {}).get("level"),
+                            "user": item.get("user_id"),
+                            "question": (item.get("data") or {}).get("question"),
+                            "sources": (item.get("data") or {}).get("sources"),
+                            "retriever_params": (item.get("data") or {}).get(
+                                "retriever_params"
+                            ),
+                            "timestamp": (
+                                item["timestamp"].isoformat()
+                                if hasattr(item.get("timestamp"), "isoformat")
+                                else item.get("timestamp")
+                            ),
+                        }
+                        for item in items
+                    ]
+                    return make_response(
+                        jsonify(
+                            {
+                                "success": True,
+                                "logs": results,
+                                "page": page,
+                                "page_size": page_size,
+                                "has_more": has_more_flag,
+                            }
+                        ),
+                        200,
+                    )
+
+            has_more = len(items) > page_size
+            items = items[:page_size]
+            results = [
+                {
+                    "id": str(item.get("id") or item.get("_id")),
+                    "action": (item.get("data") or {}).get("action"),
+                    "level": (item.get("data") or {}).get("level"),
+                    "user": item.get("user_id"),
+                    "question": (item.get("data") or {}).get("question"),
+                    "sources": (item.get("data") or {}).get("sources"),
+                    "retriever_params": (item.get("data") or {}).get(
+                        "retriever_params"
+                    ),
+                    "timestamp": (
+                        item["timestamp"].isoformat()
+                        if hasattr(item.get("timestamp"), "isoformat")
+                        else item.get("timestamp")
+                    ),
+                }
+                for item in items
+            ]
         except Exception as err:
-            current_app.logger.error(f"Error getting API key: {err}", exc_info=True)
+            current_app.logger.error(
+                f"Error getting user logs: {err}", exc_info=True
+            )
             return make_response(jsonify({"success": False}), 400)
-        query = {"user": user}
-        if api_key:
-            query = {"api_key": api_key}
-        items_cursor = (
-            user_logs_collection.find(query)
-            .sort("timestamp", -1)
-            .skip(skip)
-            .limit(page_size + 1)
-        )
-        items = list(items_cursor)
-
-        results = [
-            {
-                "id": str(item.get("_id")),
-                "action": item.get("action"),
-                "level": item.get("level"),
-                "user": item.get("user"),
-                "question": item.get("question"),
-                "sources": item.get("sources"),
-                "retriever_params": item.get("retriever_params"),
-                "timestamp": item.get("timestamp"),
-            }
-            for item in items[:page_size]
-        ]
-
-        has_more = len(items) > page_size
 
         return make_response(
             jsonify(

application/api/user/attachments/routes.py

@@ -4,13 +4,16 @@ import os
 import tempfile
 from pathlib import Path
 
-from bson.objectid import ObjectId
+import uuid
+
 from flask import current_app, jsonify, make_response, request
 from flask_restx import fields, Namespace, Resource
 
 from application.api import api
 from application.cache import get_redis_instance
 from application.core.settings import settings
+from application.storage.db.repositories.agents import AgentsRepository
+from application.storage.db.session import db_readonly
 from application.stt.constants import (
     SUPPORTED_AUDIO_EXTENSIONS,
     SUPPORTED_AUDIO_MIME_TYPES,
@@ -48,14 +51,13 @@ def _resolve_authenticated_user():
         return safe_filename(decoded_token.get("sub"))
 
     if api_key:
-        from application.api.user.base import agents_collection
-
-        agent = agents_collection.find_one({"key": api_key})
+        with db_readonly() as conn:
+            agent = AgentsRepository(conn).find_by_key(api_key)
         if not agent:
             return make_response(
                 jsonify({"success": False, "message": "Invalid API key"}), 401
             )
-        return safe_filename(agent.get("user"))
+        return safe_filename(agent.get("user_id"))
 
     return None
 
@@ -157,7 +159,7 @@ class StoreAttachment(Resource):
             
             for idx, file in enumerate(files):
                 try:
-                    attachment_id = ObjectId()
+                    attachment_id = uuid.uuid4()
                     original_filename = safe_filename(os.path.basename(file.filename))
                     _enforce_uploaded_audio_size_limit(file, original_filename)
                     relative_path = f"{settings.UPLOAD_FOLDER}/{user}/attachments/{str(attachment_id)}/{original_filename}"

application/api/user/base.py

@@ -8,15 +8,15 @@ import uuid
 from functools import wraps
 from typing import Optional, Tuple
 
-from bson.objectid import ObjectId
 from flask import current_app, jsonify, make_response, Response
-from pymongo import ReturnDocument
 from werkzeug.utils import secure_filename
 
-from application.core.mongo_db import MongoDB
+from sqlalchemy import text as _sql_text
+
 from application.core.settings import settings
-from application.storage.db.dual_write import dual_write
+from application.storage.db.base_repository import looks_like_uuid, row_to_dict
 from application.storage.db.repositories.users import UsersRepository
+from application.storage.db.session import db_readonly, db_session
 from application.storage.storage_creator import StorageCreator
 from application.vectorstore.vector_creator import VectorCreator
 
@@ -24,56 +24,6 @@ from application.vectorstore.vector_creator import VectorCreator
 storage = StorageCreator.get_storage()
 
 
-mongo = MongoDB.get_client()
-db = mongo[settings.MONGO_DB_NAME]
-
-
-conversations_collection = db["conversations"]
-sources_collection = db["sources"]
-prompts_collection = db["prompts"]
-feedback_collection = db["feedback"]
-agents_collection = db["agents"]
-agent_folders_collection = db["agent_folders"]
-token_usage_collection = db["token_usage"]
-shared_conversations_collections = db["shared_conversations"]
-users_collection = db["users"]
-user_logs_collection = db["user_logs"]
-user_tools_collection = db["user_tools"]
-attachments_collection = db["attachments"]
-workflow_runs_collection = db["workflow_runs"]
-workflows_collection = db["workflows"]
-workflow_nodes_collection = db["workflow_nodes"]
-workflow_edges_collection = db["workflow_edges"]
-
-
-try:
-    agents_collection.create_index(
-        [("shared", 1)],
-        name="shared_index",
-        background=True,
-    )
-    users_collection.create_index("user_id", unique=True)
-    workflows_collection.create_index(
-        [("user", 1)], name="workflow_user_index", background=True
-    )
-    workflow_nodes_collection.create_index(
-        [("workflow_id", 1)], name="node_workflow_index", background=True
-    )
-    workflow_nodes_collection.create_index(
-        [("workflow_id", 1), ("graph_version", 1)],
-        name="node_workflow_graph_version_index",
-        background=True,
-    )
-    workflow_edges_collection.create_index(
-        [("workflow_id", 1)], name="edge_workflow_index", background=True
-    )
-    workflow_edges_collection.create_index(
-        [("workflow_id", 1), ("graph_version", 1)],
-        name="edge_workflow_graph_version_index",
-        background=True,
-    )
-except Exception as e:
-    print("Error creating indexes:", e)
 current_dir = os.path.dirname(
     os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
 )
@@ -105,69 +55,95 @@ def generate_date_range(start_date, end_date):
 
 def ensure_user_doc(user_id):
     """
-    Ensure user document exists with proper agent preferences structure.
+    Ensure a Postgres ``users`` row exists for ``user_id``.
+
+    Returns the row as a dict with the shape legacy callers expect — in
+    particular ``user_id`` and ``agent_preferences`` (with ``pinned`` and
+    ``shared_with_me`` list keys always present).
 
     Args:
         user_id: The user ID to ensure
 
     Returns:
-        The user document
+        The user document as a dict.
     """
-    default_prefs = {
-        "pinned": [],
-        "shared_with_me": [],
-    }
-
-    user_doc = users_collection.find_one_and_update(
-        {"user_id": user_id},
-        {"$setOnInsert": {"agent_preferences": default_prefs}},
-        upsert=True,
-        return_document=ReturnDocument.AFTER,
-    )
-
-    prefs = user_doc.get("agent_preferences", {})
-    updates = {}
-    if "pinned" not in prefs:
-        updates["agent_preferences.pinned"] = []
-    if "shared_with_me" not in prefs:
-        updates["agent_preferences.shared_with_me"] = []
-    if updates:
-        users_collection.update_one({"user_id": user_id}, {"$set": updates})
-        user_doc = users_collection.find_one({"user_id": user_id})
-
-    dual_write(UsersRepository, lambda repo: repo.upsert(user_id))
+    with db_session() as conn:
+        user_doc = UsersRepository(conn).upsert(user_id)
 
+    prefs = user_doc.get("agent_preferences") or {}
+    if not isinstance(prefs, dict):
+        prefs = {}
+    prefs.setdefault("pinned", [])
+    prefs.setdefault("shared_with_me", [])
+    user_doc["agent_preferences"] = prefs
     return user_doc
 
 
 def resolve_tool_details(tool_ids):
     """
-    Resolve tool IDs to their details.
+    Resolve tool IDs to their display details.
+
+    Accepts either Postgres UUIDs or legacy Mongo ObjectId strings (mixed
+    lists are supported — each id is looked up via ``get_any``, which
+    resolves to whichever column matches). Unknown ids are silently
+    skipped.
 
     Args:
-        tool_ids: List of tool IDs
+        tool_ids: List of tool IDs (UUIDs or legacy Mongo ObjectId strings).
 
     Returns:
-        List of tool details with id, name, and display_name
+        List of tool details with ``id``, ``name``, and ``display_name``.
     """
-    valid_ids = []
+    if not tool_ids:
+        return []
+
+    uuid_ids: list[str] = []
+    legacy_ids: list[str] = []
     for tid in tool_ids:
-        try:
-            valid_ids.append(ObjectId(tid))
-        except Exception:
+        if not tid:
             continue
-    tools = user_tools_collection.find(
-        {"_id": {"$in": valid_ids}}
-    ) if valid_ids else []
+        tid_str = str(tid)
+        if looks_like_uuid(tid_str):
+            uuid_ids.append(tid_str)
+        else:
+            legacy_ids.append(tid_str)
+
+    if not uuid_ids and not legacy_ids:
+        return []
+
+    rows: list[dict] = []
+    with db_readonly() as conn:
+        if uuid_ids:
+            result = conn.execute(
+                _sql_text(
+                    "SELECT * FROM user_tools "
+                    "WHERE id = ANY(CAST(:ids AS uuid[]))"
+                ),
+                {"ids": uuid_ids},
+            )
+            rows.extend(row_to_dict(r) for r in result.fetchall())
+        if legacy_ids:
+            result = conn.execute(
+                _sql_text(
+                    "SELECT * FROM user_tools "
+                    "WHERE legacy_mongo_id = ANY(:ids)"
+                ),
+                {"ids": legacy_ids},
+            )
+            rows.extend(row_to_dict(r) for r in result.fetchall())
+
     return [
         {
-            "id": str(tool["_id"]),
-            "name": tool.get("name", ""),
-            "display_name": tool.get("customName")
-            or tool.get("displayName")
-            or tool.get("name", ""),
+            "id": str(tool.get("id") or tool.get("legacy_mongo_id") or ""),
+            "name": tool.get("name", "") or "",
+            "display_name": (
+                tool.get("custom_name")
+                or tool.get("display_name")
+                or tool.get("name", "")
+                or ""
+            ),
         }
-        for tool in tools
+        for tool in rows
     ]
 
 
@@ -237,14 +213,15 @@ def require_agent(func):
 
     @wraps(func)
     def wrapper(*args, **kwargs):
+        from application.storage.db.repositories.agents import AgentsRepository
+
         webhook_token = kwargs.get("webhook_token")
         if not webhook_token:
             return make_response(
                 jsonify({"success": False, "message": "Webhook token missing"}), 400
             )
-        agent = agents_collection.find_one(
-            {"incoming_webhook_token": webhook_token}, {"_id": 1}
-        )
+        with db_readonly() as conn:
+            agent = AgentsRepository(conn).find_by_webhook_token(webhook_token)
         if not agent:
             current_app.logger.warning(
                 f"Webhook attempt with invalid token: {webhook_token}"
@@ -253,7 +230,7 @@ def require_agent(func):
                 jsonify({"success": False, "message": "Agent not found"}), 404
             )
         kwargs["agent"] = agent
-        kwargs["agent_id_str"] = str(agent["_id"])
+        kwargs["agent_id_str"] = str(agent["id"])
         return func(*args, **kwargs)
 
     return wrapper

application/api/user/conversations/routes.py

@@ -2,14 +2,13 @@
 
 import datetime
 
-from bson.objectid import ObjectId
 from flask import current_app, jsonify, make_response, request
 from flask_restx import fields, Namespace, Resource
 
 from application.api import api
-from application.api.user.base import attachments_collection, conversations_collection
-from application.storage.db.dual_write import dual_write
+from application.storage.db.repositories.attachments import AttachmentsRepository
 from application.storage.db.repositories.conversations import ConversationsRepository
+from application.storage.db.session import db_readonly, db_session
 from application.utils import check_required_fields
 
 conversations_ns = Namespace(
@@ -34,21 +33,16 @@ class DeleteConversation(Resource):
             )
         user_id = decoded_token["sub"]
         try:
-            conversations_collection.delete_one(
-                {"_id": ObjectId(conversation_id), "user": user_id}
-            )
+            with db_session() as conn:
+                repo = ConversationsRepository(conn)
+                conv = repo.get_any(conversation_id, user_id)
+                if conv is not None:
+                    repo.delete(str(conv["id"]), user_id)
         except Exception as err:
             current_app.logger.error(
                 f"Error deleting conversation: {err}", exc_info=True
             )
             return make_response(jsonify({"success": False}), 400)
-
-        def _pg_delete(repo: ConversationsRepository) -> None:
-            conv = repo.get_by_legacy_id(conversation_id)
-            if conv is not None:
-                repo.delete(conv["id"], user_id)
-
-        dual_write(ConversationsRepository, _pg_delete)
         return make_response(jsonify({"success": True}), 200)
 
 
@@ -63,17 +57,13 @@ class DeleteAllConversations(Resource):
             return make_response(jsonify({"success": False}), 401)
         user_id = decoded_token.get("sub")
         try:
-            conversations_collection.delete_many({"user": user_id})
+            with db_session() as conn:
+                ConversationsRepository(conn).delete_all_for_user(user_id)
         except Exception as err:
             current_app.logger.error(
                 f"Error deleting all conversations: {err}", exc_info=True
             )
             return make_response(jsonify({"success": False}), 400)
-
-        dual_write(
-            ConversationsRepository,
-            lambda r, uid=user_id: r.delete_all_for_user(uid),
-        )
         return make_response(jsonify({"success": True}), 200)
 
 
@@ -86,26 +76,21 @@ class GetConversations(Resource):
         decoded_token = request.decoded_token
         if not decoded_token:
             return make_response(jsonify({"success": False}), 401)
+        user_id = decoded_token.get("sub")
         try:
-            conversations = (
-                conversations_collection.find(
-                    {
-                        "$or": [
-                            {"api_key": {"$exists": False}},
-                            {"agent_id": {"$exists": True}},
-                        ],
-                        "user": decoded_token.get("sub"),
-                    }
+            with db_readonly() as conn:
+                conversations = ConversationsRepository(conn).list_for_user(
+                    user_id, limit=30
                 )
-                .sort("date", -1)
-                .limit(30)
-            )
-
             list_conversations = [
                 {
-                    "id": str(conversation["_id"]),
+                    "id": str(conversation["id"]),
                     "name": conversation["name"],
-                    "agent_id": conversation.get("agent_id", None),
+                    "agent_id": (
+                        str(conversation["agent_id"])
+                        if conversation.get("agent_id")
+                        else None
+                    ),
                     "is_shared_usage": conversation.get("is_shared_usage", False),
                     "shared_token": conversation.get("shared_token", None),
                 }
@@ -134,38 +119,67 @@ class GetSingleConversation(Resource):
             return make_response(
                 jsonify({"success": False, "message": "ID is required"}), 400
             )
+        user_id = decoded_token.get("sub")
         try:
-            conversation = conversations_collection.find_one(
-                {"_id": ObjectId(conversation_id), "user": decoded_token.get("sub")}
-            )
-            if not conversation:
-                return make_response(jsonify({"status": "not found"}), 404)
-            # Process queries to include attachment names
+            with db_readonly() as conn:
+                repo = ConversationsRepository(conn)
+                conversation = repo.get_any(conversation_id, user_id)
+                if not conversation:
+                    return make_response(jsonify({"status": "not found"}), 404)
+                conv_pg_id = str(conversation["id"])
+                messages = repo.get_messages(conv_pg_id)
 
-            queries = conversation["queries"]
-            for query in queries:
-                if "attachments" in query and query["attachments"]:
-                    attachment_details = []
-                    for attachment_id in query["attachments"]:
-                        try:
-                            attachment = attachments_collection.find_one(
-                                {"_id": ObjectId(attachment_id)}
-                            )
-                            if attachment:
-                                attachment_details.append(
-                                    {
-                                        "id": str(attachment["_id"]),
-                                        "fileName": attachment.get(
-                                            "filename", "Unknown file"
-                                        ),
-                                    }
+                # Resolve attachment details (id, fileName) for each message.
+                attachments_repo = AttachmentsRepository(conn)
+                queries = []
+                for msg in messages:
+                    query = {
+                        "prompt": msg.get("prompt"),
+                        "response": msg.get("response"),
+                        "thought": msg.get("thought"),
+                        "sources": msg.get("sources") or [],
+                        "tool_calls": msg.get("tool_calls") or [],
+                        "timestamp": msg.get("timestamp"),
+                        "model_id": msg.get("model_id"),
+                    }
+                    if msg.get("metadata"):
+                        query["metadata"] = msg["metadata"]
+                    # Feedback on conversation_messages is a JSONB blob with
+                    # shape {"text": <str>, "timestamp": <iso>}. The legacy
+                    # frontend consumed a flat scalar feedback string, so
+                    # unwrap the ``text`` field for compat.
+                    feedback = msg.get("feedback")
+                    if feedback is not None:
+                        if isinstance(feedback, dict):
+                            query["feedback"] = feedback.get("text")
+                            if feedback.get("timestamp"):
+                                query["feedback_timestamp"] = feedback["timestamp"]
+                        else:
+                            query["feedback"] = feedback
+                    attachments = msg.get("attachments") or []
+                    if attachments:
+                        attachment_details = []
+                        for attachment_id in attachments:
+                            try:
+                                att = attachments_repo.get_any(
+                                    str(attachment_id), user_id
                                 )
-                        except Exception as e:
-                            current_app.logger.error(
-                                f"Error retrieving attachment {attachment_id}: {e}",
-                                exc_info=True,
-                            )
-                    query["attachments"] = attachment_details
+                                if att:
+                                    attachment_details.append(
+                                        {
+                                            "id": str(att["id"]),
+                                            "fileName": att.get(
+                                                "filename", "Unknown file"
+                                            ),
+                                        }
+                                    )
+                            except Exception as e:
+                                current_app.logger.error(
+                                    f"Error retrieving attachment {attachment_id}: {e}",
+                                    exc_info=True,
+                                )
+                        query["attachments"] = attachment_details
+                    queries.append(query)
         except Exception as err:
             current_app.logger.error(
                 f"Error retrieving conversation: {err}", exc_info=True
@@ -173,7 +187,9 @@ class GetSingleConversation(Resource):
             return make_response(jsonify({"success": False}), 400)
         data = {
             "queries": queries,
-            "agent_id": conversation.get("agent_id"),
+            "agent_id": (
+                str(conversation["agent_id"]) if conversation.get("agent_id") else None
+            ),
             "is_shared_usage": conversation.get("is_shared_usage", False),
             "shared_token": conversation.get("shared_token", None),
         }
@@ -207,22 +223,16 @@ class UpdateConversationName(Resource):
             return missing_fields
         user_id = decoded_token.get("sub")
         try:
-            conversations_collection.update_one(
-                {"_id": ObjectId(data["id"]), "user": user_id},
-                {"$set": {"name": data["name"]}},
-            )
+            with db_session() as conn:
+                repo = ConversationsRepository(conn)
+                conv = repo.get_any(data["id"], user_id)
+                if conv is not None:
+                    repo.rename(str(conv["id"]), user_id, data["name"])
         except Exception as err:
             current_app.logger.error(
                 f"Error updating conversation name: {err}", exc_info=True
             )
             return make_response(jsonify({"success": False}), 400)
-
-        def _pg_rename(repo: ConversationsRepository) -> None:
-            conv = repo.get_by_legacy_id(data["id"])
-            if conv is not None:
-                repo.rename(conv["id"], user_id, data["name"])
-
-        dual_write(ConversationsRepository, _pg_rename)
         return make_response(jsonify({"success": True}), 200)
 
 
@@ -260,61 +270,34 @@ class SubmitFeedback(Resource):
         missing_fields = check_required_fields(data, required_fields)
         if missing_fields:
             return missing_fields
+        user_id = decoded_token.get("sub")
+        feedback_value = data["feedback"]
+        question_index = int(data["question_index"])
+        # Normalize string feedback to lowercase so analytics queries
+        # (which match 'like'/'dislike') count rows correctly. Tolerate
+        # legacy uppercase clients on ingest. Non-string values pass through.
+        if isinstance(feedback_value, str):
+            feedback_value = feedback_value.lower()
+        feedback_payload = (
+            None
+            if feedback_value is None
+            else {
+                "text": feedback_value,
+                "timestamp": datetime.datetime.now(
+                    datetime.timezone.utc
+                ).isoformat(),
+            }
+        )
         try:
-            if data["feedback"] is None:
-                # Remove feedback and feedback_timestamp if feedback is null
-
-                conversations_collection.update_one(
-                    {
-                        "_id": ObjectId(data["conversation_id"]),
-                        "user": decoded_token.get("sub"),
-                        f"queries.{data['question_index']}": {"$exists": True},
-                    },
-                    {
-                        "$unset": {
-                            f"queries.{data['question_index']}.feedback": "",
-                            f"queries.{data['question_index']}.feedback_timestamp": "",
-                        }
-                    },
-                )
-            else:
-                # Set feedback and feedback_timestamp if feedback has a value
-
-                conversations_collection.update_one(
-                    {
-                        "_id": ObjectId(data["conversation_id"]),
-                        "user": decoded_token.get("sub"),
-                        f"queries.{data['question_index']}": {"$exists": True},
-                    },
-                    {
-                        "$set": {
-                            f"queries.{data['question_index']}.feedback": data[
-                                "feedback"
-                            ],
-                            f"queries.{data['question_index']}.feedback_timestamp": datetime.datetime.now(
-                                datetime.timezone.utc
-                            ),
-                        }
-                    },
-                )
+            with db_session() as conn:
+                repo = ConversationsRepository(conn)
+                conv = repo.get_any(data["conversation_id"], user_id)
+                if conv is None:
+                    return make_response(
+                        jsonify({"success": False, "message": "Not found"}), 404
+                    )
+                repo.set_feedback(str(conv["id"]), question_index, feedback_payload)
         except Exception as err:
             current_app.logger.error(f"Error submitting feedback: {err}", exc_info=True)
             return make_response(jsonify({"success": False}), 400)
-
-        # Dual-write to Postgres: mirror the per-message feedback set/unset.
-        feedback_value = data["feedback"]
-        question_index = int(data["question_index"])
-        feedback_payload = (
-            None if feedback_value is None
-            else {"text": feedback_value, "timestamp": datetime.datetime.now(
-                datetime.timezone.utc
-            ).isoformat()}
-        )
-
-        def _pg_feedback(repo: ConversationsRepository) -> None:
-            conv = repo.get_by_legacy_id(data["conversation_id"])
-            if conv is not None:
-                repo.set_feedback(conv["id"], question_index, feedback_payload)
-
-        dual_write(ConversationsRepository, _pg_feedback)
         return make_response(jsonify({"success": True}), 200)

application/api/user/models/routes.py

@@ -1,18 +1,135 @@
-from flask import current_app, jsonify, make_response
+"""Model routes.
+
+- ``GET /api/models`` — list available models for the current user.
+  Combines the built-in catalog with the user's BYOM records.
+- ``GET/POST/PATCH/DELETE /api/user/models[/<id>]`` — CRUD for the
+  user's own OpenAI-compatible model registrations (BYOM).
+- ``POST /api/user/models/<id>/test`` — sanity-check the upstream
+  endpoint with a tiny request.
+
+Every BYOM endpoint is user-scoped at the repository layer
+(every query filters on ``user_id`` from ``request.decoded_token``).
+"""
+
+from __future__ import annotations
+
+import logging
+
+import requests
+from flask import current_app, jsonify, make_response, request
 from flask_restx import Namespace, Resource
 
-from application.core.model_settings import ModelRegistry
+from application.api import api
+from application.core.model_registry import ModelRegistry
+from application.security.safe_url import (
+    UnsafeUserUrlError,
+    pinned_post,
+    validate_user_base_url,
+)
+from application.storage.db.repositories.user_custom_models import (
+    UserCustomModelsRepository,
+)
+from application.storage.db.session import db_readonly, db_session
+from application.utils import check_required_fields
+
+
+logger = logging.getLogger(__name__)
+
 
 models_ns = Namespace("models", description="Available models", path="/api")
 
 
+_CONTEXT_WINDOW_MIN = 1_000
+_CONTEXT_WINDOW_MAX = 10_000_000
+
+
+def _user_id_or_401():
+    decoded_token = request.decoded_token
+    if not decoded_token:
+        return None, make_response(jsonify({"success": False}), 401)
+    user_id = decoded_token.get("sub")
+    if not user_id:
+        return None, make_response(jsonify({"success": False}), 401)
+    return user_id, None
+
+
+def _normalize_capabilities(raw) -> dict:
+    """Coerce + bound the user-supplied capabilities payload."""
+    raw = raw or {}
+    out = {}
+    if "supports_tools" in raw:
+        out["supports_tools"] = bool(raw["supports_tools"])
+    if "supports_structured_output" in raw:
+        out["supports_structured_output"] = bool(raw["supports_structured_output"])
+    if "supports_streaming" in raw:
+        out["supports_streaming"] = bool(raw["supports_streaming"])
+    if "attachments" in raw:
+        atts = raw["attachments"] or []
+        if not isinstance(atts, list):
+            raise ValueError("'capabilities.attachments' must be a list")
+        coerced = [str(a) for a in atts]
+        # Reject unknown aliases at the API boundary so bad payloads
+        # never reach the registry layer (where lenient expansion just
+        # drops them). Raw MIME types (containing ``/``) pass through
+        # unchanged for parity with the built-in YAML schema.
+        from application.core.model_yaml import builtin_attachment_aliases
+
+        aliases = builtin_attachment_aliases()
+        for entry in coerced:
+            if "/" in entry:
+                continue
+            if entry not in aliases:
+                valid = ", ".join(sorted(aliases.keys())) or "<none defined>"
+                raise ValueError(
+                    f"unknown attachment alias '{entry}' in "
+                    f"'capabilities.attachments'. Valid aliases: {valid}, "
+                    f"or use a raw MIME type like 'image/png'."
+                )
+        out["attachments"] = coerced
+    if "context_window" in raw:
+        try:
+            cw = int(raw["context_window"])
+        except (TypeError, ValueError):
+            raise ValueError("'capabilities.context_window' must be an integer")
+        if not (_CONTEXT_WINDOW_MIN <= cw <= _CONTEXT_WINDOW_MAX):
+            raise ValueError(
+                f"'capabilities.context_window' must be between "
+                f"{_CONTEXT_WINDOW_MIN} and {_CONTEXT_WINDOW_MAX}"
+            )
+        out["context_window"] = cw
+    return out
+
+
+def _row_to_response(row: dict) -> dict:
+    """Wire-format projection — never includes the API key."""
+    return {
+        "id": str(row["id"]),
+        "upstream_model_id": row["upstream_model_id"],
+        "display_name": row["display_name"],
+        "description": row.get("description") or "",
+        "base_url": row["base_url"],
+        "capabilities": row.get("capabilities") or {},
+        "enabled": bool(row.get("enabled", True)),
+        "source": "user",
+    }
+
+
 @models_ns.route("/models")
 class ModelsListResource(Resource):
     def get(self):
-        """Get list of available models with their capabilities."""
+        """Get list of available models with their capabilities.
+
+        When the request is authenticated, the response includes the
+        user's own BYOM registrations alongside the built-in catalog.
+        """
         try:
+            user_id = None
+            decoded_token = getattr(request, "decoded_token", None)
+            if decoded_token:
+                user_id = decoded_token.get("sub")
+
             registry = ModelRegistry.get_instance()
-            models = registry.get_enabled_models()
+            models = registry.get_enabled_models(user_id=user_id)
 
             response = {
                 "models": [model.to_dict() for model in models],
@@ -23,3 +140,382 @@ class ModelsListResource(Resource):
             current_app.logger.error(f"Error fetching models: {err}", exc_info=True)
             return make_response(jsonify({"success": False}), 500)
         return make_response(jsonify(response), 200)
+
+
+@models_ns.route("/user/models")
+class UserModelsCollectionResource(Resource):
+    @api.doc(description="List the current user's BYOM custom models")
+    def get(self):
+        user_id, err = _user_id_or_401()
+        if err:
+            return err
+        try:
+            with db_readonly() as conn:
+                rows = UserCustomModelsRepository(conn).list_for_user(user_id)
+            return make_response(
+                jsonify({"models": [_row_to_response(r) for r in rows]}), 200
+            )
+        except Exception as e:
+            current_app.logger.error(
+                f"Error listing user custom models: {e}", exc_info=True
+            )
+            return make_response(jsonify({"success": False}), 500)
+
+    @api.doc(description="Register a new BYOM custom model")
+    def post(self):
+        user_id, err = _user_id_or_401()
+        if err:
+            return err
+
+        data = request.get_json() or {}
+        missing = check_required_fields(
+            data,
+            ["upstream_model_id", "display_name", "base_url", "api_key"],
+        )
+        if missing:
+            return missing
+
+        # SECURITY: reject blank api_key — would leak instance API key
+        # to the user-supplied base_url via LLMCreator fallback.
+        for required_nonblank in (
+            "upstream_model_id",
+            "display_name",
+            "base_url",
+            "api_key",
+        ):
+            value = data.get(required_nonblank)
+            if not isinstance(value, str) or not value.strip():
+                return make_response(
+                    jsonify(
+                        {
+                            "success": False,
+                            "error": f"'{required_nonblank}' must be a non-empty string",
+                        }
+                    ),
+                    400,
+                )
+
+        # SSRF guard at create time. Re-runs at dispatch time (LLMCreator)
+        # as defense in depth against DNS rebinding and pre-guard rows.
+        try:
+            validate_user_base_url(data["base_url"])
+        except UnsafeUserUrlError as e:
+            return make_response(
+                jsonify({"success": False, "error": str(e)}), 400
+            )
+
+        try:
+            capabilities = _normalize_capabilities(data.get("capabilities"))
+        except ValueError as e:
+            return make_response(
+                jsonify({"success": False, "error": str(e)}), 400
+            )
+
+        try:
+            with db_session() as conn:
+                row = UserCustomModelsRepository(conn).create(
+                    user_id=user_id,
+                    upstream_model_id=data["upstream_model_id"],
+                    display_name=data["display_name"],
+                    description=data.get("description") or "",
+                    base_url=data["base_url"],
+                    api_key_plaintext=data["api_key"],
+                    capabilities=capabilities,
+                    enabled=bool(data.get("enabled", True)),
+                )
+        except Exception as e:
+            current_app.logger.error(
+                f"Error creating user custom model: {e}", exc_info=True
+            )
+            return make_response(jsonify({"success": False}), 500)
+
+        ModelRegistry.invalidate_user(user_id)
+        return make_response(jsonify(_row_to_response(row)), 201)
+
+
+@models_ns.route("/user/models/<string:model_id>")
+class UserModelResource(Resource):
+    @api.doc(description="Get one BYOM custom model")
+    def get(self, model_id):
+        user_id, err = _user_id_or_401()
+        if err:
+            return err
+        try:
+            with db_readonly() as conn:
+                row = UserCustomModelsRepository(conn).get(model_id, user_id)
+        except Exception as e:
+            current_app.logger.error(
+                f"Error fetching user custom model: {e}", exc_info=True
+            )
+            return make_response(jsonify({"success": False}), 500)
+        if row is None:
+            return make_response(jsonify({"success": False}), 404)
+        return make_response(jsonify(_row_to_response(row)), 200)
+
+    @api.doc(description="Update a BYOM custom model (partial)")
+    def patch(self, model_id):
+        user_id, err = _user_id_or_401()
+        if err:
+            return err
+
+        data = request.get_json() or {}
+
+        # Reject present-but-blank values for fields where blank doesn't
+        # mean "no change". (The api_key special case — blank means "keep
+        # existing" — is handled below.)
+        for required_nonblank in (
+            "upstream_model_id",
+            "display_name",
+            "base_url",
+        ):
+            if required_nonblank in data:
+                value = data[required_nonblank]
+                if not isinstance(value, str) or not value.strip():
+                    return make_response(
+                        jsonify(
+                            {
+                                "success": False,
+                                "error": f"'{required_nonblank}' cannot be blank",
+                            }
+                        ),
+                        400,
+                    )
+
+        if "base_url" in data and data["base_url"]:
+            try:
+                validate_user_base_url(data["base_url"])
+            except UnsafeUserUrlError as e:
+                return make_response(
+                    jsonify({"success": False, "error": str(e)}), 400
+                )
+
+        update_fields: dict = {}
+        for k in (
+            "upstream_model_id",
+            "display_name",
+            "description",
+            "base_url",
+            "enabled",
+        ):
+            if k in data:
+                update_fields[k] = data[k]
+
+        if "capabilities" in data:
+            try:
+                update_fields["capabilities"] = _normalize_capabilities(
+                    data["capabilities"]
+                )
+            except ValueError as e:
+                return make_response(
+                    jsonify({"success": False, "error": str(e)}), 400
+                )
+
+        # PATCH semantics: blank/missing api_key → keep the existing
+        # ciphertext; non-empty api_key → re-encrypt and replace.
+        if data.get("api_key"):
+            update_fields["api_key_plaintext"] = data["api_key"]
+
+        if not update_fields:
+            return make_response(
+                jsonify({"success": False, "error": "no updatable fields"}), 400
+            )
+
+        try:
+            with db_session() as conn:
+                ok = UserCustomModelsRepository(conn).update(
+                    model_id, user_id, update_fields
+                )
+        except Exception as e:
+            current_app.logger.error(
+                f"Error updating user custom model: {e}", exc_info=True
+            )
+            return make_response(jsonify({"success": False}), 500)
+
+        if not ok:
+            return make_response(jsonify({"success": False}), 404)
+
+        ModelRegistry.invalidate_user(user_id)
+        with db_readonly() as conn:
+            row = UserCustomModelsRepository(conn).get(model_id, user_id)
+        return make_response(jsonify(_row_to_response(row)), 200)
+
+    @api.doc(description="Delete a BYOM custom model")
+    def delete(self, model_id):
+        user_id, err = _user_id_or_401()
+        if err:
+            return err
+        try:
+            with db_session() as conn:
+                ok = UserCustomModelsRepository(conn).delete(model_id, user_id)
+        except Exception as e:
+            current_app.logger.error(
+                f"Error deleting user custom model: {e}", exc_info=True
+            )
+            return make_response(jsonify({"success": False}), 500)
+        if not ok:
+            return make_response(jsonify({"success": False}), 404)
+
+        ModelRegistry.invalidate_user(user_id)
+        return make_response(jsonify({"success": True}), 200)
+
+
+def _run_connection_test(
+    base_url: str, api_key: str, upstream_model_id: str
+):
+    """Send a 1-token chat-completion to verify a BYOM endpoint.
+
+    Returns ``(body, http_status)``. Upstream errors return 200 with
+    ``ok=False`` so the UI can render inline errors; only local SSRF
+    rejection returns 400.
+    """
+    url = base_url.rstrip("/") + "/chat/completions"
+    payload = {
+        "model": upstream_model_id,
+        "messages": [{"role": "user", "content": "hi"}],
+        "max_tokens": 1,
+        "stream": False,
+    }
+    headers = {
+        "Authorization": f"Bearer {api_key}",
+        "Content-Type": "application/json",
+    }
+    try:
+        # pinned_post closes the DNS-rebinding window. Redirects off
+        # because 3xx could bounce to an internal address (the SSRF
+        # guard only validates the supplied URL).
+        resp = pinned_post(
+            url,
+            json=payload,
+            headers=headers,
+            timeout=5,
+            allow_redirects=False,
+        )
+    except UnsafeUserUrlError as e:
+        return {"ok": False, "error": str(e)}, 400
+    except requests.RequestException as e:
+        return {"ok": False, "error": f"connection error: {e}"}, 200
+
+    if 300 <= resp.status_code < 400:
+        return (
+            {
+                "ok": False,
+                "error": (
+                    f"upstream returned HTTP {resp.status_code} "
+                    "redirect; refusing to follow"
+                ),
+            },
+            200,
+        )
+
+    if resp.status_code >= 400:
+        # Cap and only reflect JSON to avoid body-exfil via non-API responses.
+        content_type = (resp.headers.get("Content-Type") or "").lower()
+        if "application/json" in content_type:
+            text = (resp.text or "")[:500]
+            error_msg = f"upstream returned HTTP {resp.status_code}: {text}"
+        else:
+            error_msg = f"upstream returned HTTP {resp.status_code}"
+        return {"ok": False, "error": error_msg}, 200
+
+    return {"ok": True}, 200
+
+
+@models_ns.route("/user/models/test")
+class UserModelTestPayloadResource(Resource):
+    @api.doc(
+        description=(
+            "Test an arbitrary BYOM payload (display_name / model id / "
+            "base_url / api_key) without saving. Used by the UI's 'Test "
+            "connection' button so the user can validate before they "
+            "Save. Same SSRF guard, same 1-token request, same 5s "
+            "timeout as the by-id variant."
+        )
+    )
+    def post(self):
+        user_id, err = _user_id_or_401()
+        if err:
+            return err
+
+        data = request.get_json() or {}
+        missing = check_required_fields(
+            data, ["base_url", "api_key", "upstream_model_id"]
+        )
+        if missing:
+            return missing
+
+        body, status = _run_connection_test(
+            data["base_url"], data["api_key"], data["upstream_model_id"]
+        )
+        return make_response(jsonify(body), status)
+
+
+@models_ns.route("/user/models/<string:model_id>/test")
+class UserModelTestResource(Resource):
+    @api.doc(
+        description=(
+            "Test a saved BYOM record. Defaults to the stored "
+            "base_url / upstream_model_id / encrypted api_key, but "
+            "any of those can be overridden via the request body so "
+            "the UI can test in-flight edits before saving. Used by "
+            "the 'Test connection' button in edit mode."
+        )
+    )
+    def post(self, model_id):
+        user_id, err = _user_id_or_401()
+        if err:
+            return err
+
+        data = request.get_json() or {}
+        # Per-field overrides; blank/missing falls back to stored value.
+        override_base_url = (data.get("base_url") or "").strip() or None
+        override_upstream_model_id = (
+            data.get("upstream_model_id") or ""
+        ).strip() or None
+        override_api_key = (data.get("api_key") or "").strip() or None
+
+        try:
+            with db_readonly() as conn:
+                repo = UserCustomModelsRepository(conn)
+                row = repo.get(model_id, user_id)
+                if row is None:
+                    return make_response(jsonify({"success": False}), 404)
+                stored_api_key = (
+                    repo._decrypt_api_key(
+                        row.get("api_key_encrypted", ""), user_id
+                    )
+                    if not override_api_key
+                    else None
+                )
+        except Exception as e:
+            current_app.logger.error(
+                f"Error loading user custom model for test: {e}", exc_info=True
+            )
+            return make_response(
+                jsonify({"ok": False, "error": "internal error loading model"}),
+                500,
+            )
+
+        api_key = override_api_key or stored_api_key
+        if not api_key:
+            return make_response(
+                jsonify(
+                    {
+                        "ok": False,
+                        "error": (
+                            "Stored API key could not be decrypted. The "
+                            "encryption secret may have rotated. Re-save "
+                            "the model with the API key to recover."
+                        ),
+                    }
+                ),
+                400,
+            )
+
+        base_url = override_base_url or row["base_url"]
+        upstream_model_id = (
+            override_upstream_model_id or row["upstream_model_id"]
+        )
+        body, status = _run_connection_test(
+            base_url, api_key, upstream_model_id
+        )
+        return make_response(jsonify(body), status)

application/api/user/prompts/routes.py

@@ -2,14 +2,13 @@
 
 import os
 
-from bson.objectid import ObjectId
 from flask import current_app, jsonify, make_response, request
 from flask_restx import fields, Namespace, Resource
 
 from application.api import api
-from application.api.user.base import current_dir, prompts_collection
-from application.storage.db.dual_write import dual_write
+from application.api.user.base import current_dir
 from application.storage.db.repositories.prompts import PromptsRepository
+from application.storage.db.session import db_readonly, db_session
 from application.utils import check_required_fields
 
 prompts_ns = Namespace(
@@ -42,21 +41,9 @@ class CreatePrompt(Resource):
             return missing_fields
         user = decoded_token.get("sub")
         try:
-
-            resp = prompts_collection.insert_one(
-                {
-                    "name": data["name"],
-                    "content": data["content"],
-                    "user": user,
-                }
-            )
-            new_id = str(resp.inserted_id)
-            dual_write(
-                PromptsRepository,
-                lambda repo, u=user, n=data["name"], c=data["content"], mid=new_id: repo.create(
-                    u, n, c, legacy_mongo_id=mid,
-                ),
-            )
+            with db_session() as conn:
+                prompt = PromptsRepository(conn).create(user, data["name"], data["content"])
+            new_id = str(prompt["id"])
         except Exception as err:
             current_app.logger.error(f"Error creating prompt: {err}", exc_info=True)
             return make_response(jsonify({"success": False}), 400)
@@ -72,17 +59,17 @@ class GetPrompts(Resource):
             return make_response(jsonify({"success": False}), 401)
         user = decoded_token.get("sub")
         try:
-            prompts = prompts_collection.find({"user": user})
+            with db_readonly() as conn:
+                prompts = PromptsRepository(conn).list_for_user(user)
             list_prompts = [
                 {"id": "default", "name": "default", "type": "public"},
                 {"id": "creative", "name": "creative", "type": "public"},
                 {"id": "strict", "name": "strict", "type": "public"},
             ]
-
             for prompt in prompts:
                 list_prompts.append(
                     {
-                        "id": str(prompt["_id"]),
+                        "id": str(prompt["id"]),
                         "name": prompt["name"],
                         "type": "private",
                     }
@@ -127,9 +114,12 @@ class GetSinglePrompt(Resource):
                 ) as f:
                     chat_reduce_strict = f.read()
                 return make_response(jsonify({"content": chat_reduce_strict}), 200)
-            prompt = prompts_collection.find_one(
-                {"_id": ObjectId(prompt_id), "user": user}
-            )
+            with db_readonly() as conn:
+                prompt = PromptsRepository(conn).get_any(prompt_id, user)
+            if not prompt:
+                return make_response(
+                    jsonify({"success": False, "message": "Prompt not found"}), 404
+                )
         except Exception as err:
             current_app.logger.error(f"Error retrieving prompt: {err}", exc_info=True)
             return make_response(jsonify({"success": False}), 400)
@@ -156,11 +146,15 @@ class DeletePrompt(Resource):
         if missing_fields:
             return missing_fields
         try:
-            prompts_collection.delete_one({"_id": ObjectId(data["id"]), "user": user})
-            dual_write(
-                PromptsRepository,
-                lambda repo, pid=data["id"], u=user: repo.delete_by_legacy_id(pid, u),
-            )
+            with db_session() as conn:
+                repo = PromptsRepository(conn)
+                prompt = repo.get_any(data["id"], user)
+                if not prompt:
+                    return make_response(
+                        jsonify({"success": False, "message": "Prompt not found"}),
+                        404,
+                    )
+                repo.delete(str(prompt["id"]), user)
         except Exception as err:
             current_app.logger.error(f"Error deleting prompt: {err}", exc_info=True)
             return make_response(jsonify({"success": False}), 400)
@@ -193,16 +187,15 @@ class UpdatePrompt(Resource):
         if missing_fields:
             return missing_fields
         try:
-            prompts_collection.update_one(
-                {"_id": ObjectId(data["id"]), "user": user},
-                {"$set": {"name": data["name"], "content": data["content"]}},
-            )
-            dual_write(
-                PromptsRepository,
-                lambda repo, pid=data["id"], u=user, n=data["name"], c=data["content"]: repo.update_by_legacy_id(
-                    pid, u, n, c,
-                ),
-            )
+            with db_session() as conn:
+                repo = PromptsRepository(conn)
+                prompt = repo.get_any(data["id"], user)
+                if not prompt:
+                    return make_response(
+                        jsonify({"success": False, "message": "Prompt not found"}),
+                        404,
+                    )
+                repo.update(str(prompt["id"]), user, data["name"], data["content"])
         except Exception as err:
             current_app.logger.error(f"Error updating prompt: {err}", exc_info=True)
             return make_response(jsonify({"success": False}), 400)

application/api/user/sharing/routes.py

@@ -2,89 +2,126 @@
 
 import uuid
 
-from bson.binary import Binary, UuidRepresentation
-from bson.dbref import DBRef
-from bson.objectid import ObjectId
 from flask import current_app, jsonify, make_response, request
 from flask_restx import fields, inputs, Namespace, Resource
+from sqlalchemy import text as _sql_text
 
 from application.api import api
-from application.api.user.base import (
-    agents_collection,
-    attachments_collection,
-    conversations_collection,
-    shared_conversations_collections,
-)
-from application.storage.db.dual_write import dual_write
+from application.storage.db.base_repository import looks_like_uuid
+from application.storage.db.repositories.agents import AgentsRepository
+from application.storage.db.repositories.attachments import AttachmentsRepository
 from application.storage.db.repositories.conversations import ConversationsRepository
 from application.storage.db.repositories.shared_conversations import (
     SharedConversationsRepository,
 )
+from application.storage.db.session import db_readonly, db_session
 from application.utils import check_required_fields
 
 
-def _dual_write_share(
-    mongo_conv_id: str,
-    share_uuid: str,
-    user: str,
-    *,
-    is_promptable: bool,
-    first_n_queries: int,
-    api_key: str | None,
-    prompt_id: str | None = None,
-    chunks: int | None = None,
-) -> None:
-    """Mirror a Mongo share-record insert into Postgres.
-
-    Preserves the Mongo-generated UUID so public ``/shared/{uuid}`` URLs
-    resolve from both stores during cutover.
-    """
-    def _write(repo: SharedConversationsRepository) -> None:
-        conv = ConversationsRepository(repo._conn).get_by_legacy_id(
-            mongo_conv_id, user_id=user,
-        )
-        if conv is None:
-            return
-        # prompt_id / chunks are only meaningful for promptable shares;
-        # prompt_id is often the string "default" or an ObjectId that
-        # hasn't been migrated — pass as-is and let the repo drop
-        # non-UUID values. Scope the prompt lookup by user_id so an
-        # authenticated caller can't link another user's prompt into
-        # their share record.
-        resolved_prompt_id = None
-        if prompt_id and len(str(prompt_id)) == 24:
-            from sqlalchemy import text as _text
-            row = repo._conn.execute(
-                _text(
-                    "SELECT id FROM prompts "
-                    "WHERE legacy_mongo_id = :legacy_id AND user_id = :user_id"
-                ),
-                {"legacy_id": str(prompt_id), "user_id": user},
-            ).fetchone()
-            if row:
-                resolved_prompt_id = str(row[0])
-        # get_or_create is race-free on the PG side thanks to the
-        # composite partial unique index on the dedup tuple
-        # (migration 0008). It converges concurrent share requests to
-        # a single row.
-        repo.get_or_create(
-            conv["id"],
-            user,
-            is_promptable=is_promptable,
-            first_n_queries=first_n_queries,
-            api_key=api_key,
-            prompt_id=resolved_prompt_id,
-            chunks=chunks,
-            share_uuid=share_uuid,
-        )
-
-    dual_write(SharedConversationsRepository, _write)
-
 sharing_ns = Namespace(
     "sharing", description="Conversation sharing operations", path="/api"
 )
 
 
+def _resolve_prompt_pg_id(conn, prompt_id_raw, user_id):
+    """Translate an incoming prompt id (UUID or legacy Mongo ObjectId) to a PG UUID.
+
+    Scoped by ``user_id`` so a caller can't link another user's prompt
+    into their share record. Returns ``None`` for sentinel values
+    (``"default"``) or unresolved ids.
+    """
+    if not prompt_id_raw or prompt_id_raw == "default":
+        return None
+    value = str(prompt_id_raw)
+    # Already UUID — trust it but still require ownership. A shape-gate
+    # (rather than a loose ``len == 36 and '-' in value`` check) keeps
+    # non-UUID input out of ``CAST(:pid AS uuid)``; the cast would raise
+    # and poison the readonly transaction otherwise.
+    if looks_like_uuid(value):
+        row = conn.execute(
+            _sql_text(
+                "SELECT id FROM prompts WHERE id = CAST(:pid AS uuid) "
+                "AND user_id = :uid"
+            ),
+            {"pid": value, "uid": user_id},
+        ).fetchone()
+        return str(row[0]) if row else None
+    # Legacy Mongo ObjectId fallback.
+    row = conn.execute(
+        _sql_text(
+            "SELECT id FROM prompts WHERE legacy_mongo_id = :pid "
+            "AND user_id = :uid"
+        ),
+        {"pid": value, "uid": user_id},
+    ).fetchone()
+    return str(row[0]) if row else None
+
+
+def _resolve_source_pg_id(conn, source_raw):
+    """Translate a source id (UUID or legacy Mongo ObjectId) to a PG UUID."""
+    if not source_raw:
+        return None
+    value = str(source_raw)
+    # See ``_resolve_prompt_pg_id`` for the shape-gate rationale.
+    if looks_like_uuid(value):
+        row = conn.execute(
+            _sql_text(
+                "SELECT id FROM sources WHERE id = CAST(:sid AS uuid)"
+            ),
+            {"sid": value},
+        ).fetchone()
+        return str(row[0]) if row else None
+    row = conn.execute(
+        _sql_text("SELECT id FROM sources WHERE legacy_mongo_id = :sid"),
+        {"sid": value},
+    ).fetchone()
+    return str(row[0]) if row else None
+
+
+def _find_reusable_share_agent(
+    conn, user_id, *, prompt_pg_id, chunks, source_pg_id, retriever,
+):
+    """Find an existing share-as-agent key row matching these parameters.
+
+    Mirrors the legacy Mongo ``agents_collection.find_one`` pre-existence
+    check. Used to reuse an api key across repeated shares of the same
+    conversation with the same prompt/chunks/source/retriever.
+    """
+    clauses = ["user_id = :uid", "key IS NOT NULL"]
+    params: dict = {"uid": user_id}
+    if prompt_pg_id is None:
+        clauses.append("prompt_id IS NULL")
+    else:
+        clauses.append("prompt_id = CAST(:pid AS uuid)")
+        params["pid"] = prompt_pg_id
+    if chunks is None:
+        clauses.append("chunks IS NULL")
+    else:
+        clauses.append("chunks = :chunks")
+        params["chunks"] = int(chunks)
+    if source_pg_id is None:
+        clauses.append("source_id IS NULL")
+    else:
+        clauses.append("source_id = CAST(:sid AS uuid)")
+        params["sid"] = source_pg_id
+    if retriever is None:
+        clauses.append("retriever IS NULL")
+    else:
+        clauses.append("retriever = :retr")
+        params["retr"] = retriever
+    sql = (
+        "SELECT * FROM agents WHERE "
+        + " AND ".join(clauses)
+        + " LIMIT 1"
+    )
+    row = conn.execute(_sql_text(sql), params).fetchone()
+    if row is None:
+        return None
+    mapping = dict(row._mapping)
+    mapping["id"] = str(mapping["id"]) if mapping.get("id") else None
+    return mapping
+
+
 @sharing_ns.route("/share")
 class ShareConversation(Resource):
     share_conversation_model = api.model(
@@ -119,173 +156,93 @@ class ShareConversation(Resource):
         conversation_id = data["conversation_id"]
 
         try:
-            conversation = conversations_collection.find_one(
-                {"_id": ObjectId(conversation_id), "user": user}
-            )
-            if conversation is None:
-                return make_response(
-                    jsonify(
-                        {
-                            "status": "error",
-                            "message": "Conversation does not exist",
-                        }
-                    ),
-                    404,
-                )
-            current_n_queries = len(conversation["queries"])
-            explicit_binary = Binary.from_uuid(
-                uuid.uuid4(), UuidRepresentation.STANDARD
-            )
+            with db_session() as conn:
+                conv_repo = ConversationsRepository(conn)
+                shared_repo = SharedConversationsRepository(conn)
+                agents_repo = AgentsRepository(conn)
 
-            if is_promptable:
-                prompt_id = data.get("prompt_id", "default")
-                chunks = data.get("chunks", "2")
-
-                name = conversation["name"] + "(shared)"
-                new_api_key_data = {
-                    "prompt_id": prompt_id,
-                    "chunks": chunks,
-                    "user": user,
-                }
-
-                if "source" in data and ObjectId.is_valid(data["source"]):
-                    new_api_key_data["source"] = DBRef(
-                        "sources", ObjectId(data["source"])
-                    )
-                if "retriever" in data:
-                    new_api_key_data["retriever"] = data["retriever"]
-                pre_existing_api_document = agents_collection.find_one(new_api_key_data)
-                if pre_existing_api_document:
-                    api_uuid = pre_existing_api_document["key"]
-                    pre_existing = shared_conversations_collections.find_one(
-                        {
-                            "conversation_id": ObjectId(conversation_id),
-                            "isPromptable": is_promptable,
-                            "first_n_queries": current_n_queries,
-                            "user": user,
-                            "api_key": api_uuid,
-                        }
-                    )
-                    if pre_existing is not None:
-                        return make_response(
-                            jsonify(
-                                {
-                                    "success": True,
-                                    "identifier": str(pre_existing["uuid"].as_uuid()),
-                                }
-                            ),
-                            200,
-                        )
-                    else:
-                        shared_conversations_collections.insert_one(
+                conversation = conv_repo.get_any(conversation_id, user)
+                if conversation is None:
+                    return make_response(
+                        jsonify(
                             {
-                                "uuid": explicit_binary,
-                                "conversation_id": ObjectId(conversation_id),
-                                "isPromptable": is_promptable,
-                                "first_n_queries": current_n_queries,
-                                "user": user,
-                                "api_key": api_uuid,
+                                "status": "error",
+                                "message": "Conversation does not exist",
                             }
-                        )
-                        _dual_write_share(
-                            conversation_id,
-                            str(explicit_binary.as_uuid()),
-                            user,
-                            is_promptable=is_promptable,
-                            first_n_queries=current_n_queries,
-                            api_key=api_uuid,
-                            prompt_id=prompt_id,
-                            chunks=int(chunks) if chunks else None,
-                        )
-                        return make_response(
-                            jsonify(
-                                {
-                                    "success": True,
-                                    "identifier": str(explicit_binary.as_uuid()),
-                                }
-                            ),
-                            201,
-                        )
-                else:
-                    api_uuid = str(uuid.uuid4())
-                    new_api_key_data["key"] = api_uuid
-                    new_api_key_data["name"] = name
-
-                    if "source" in data and ObjectId.is_valid(data["source"]):
-                        new_api_key_data["source"] = DBRef(
-                            "sources", ObjectId(data["source"])
-                        )
-                    if "retriever" in data:
-                        new_api_key_data["retriever"] = data["retriever"]
-                    agents_collection.insert_one(new_api_key_data)
-                    shared_conversations_collections.insert_one(
-                        {
-                            "uuid": explicit_binary,
-                            "conversation_id": ObjectId(conversation_id),
-                            "isPromptable": is_promptable,
-                            "first_n_queries": current_n_queries,
-                            "user": user,
-                            "api_key": api_uuid,
-                        }
+                        ),
+                        404,
                     )
-                    _dual_write_share(
-                        conversation_id,
-                        str(explicit_binary.as_uuid()),
+                conv_pg_id = str(conversation["id"])
+                current_n_queries = conv_repo.message_count(conv_pg_id)
+
+                if is_promptable:
+                    prompt_id_raw = data.get("prompt_id", "default")
+                    chunks_raw = data.get("chunks", "2")
+                    try:
+                        chunks_int = int(chunks_raw) if chunks_raw not in (None, "") else None
+                    except (TypeError, ValueError):
+                        chunks_int = None
+
+                    prompt_pg_id = _resolve_prompt_pg_id(conn, prompt_id_raw, user)
+                    source_pg_id = _resolve_source_pg_id(conn, data.get("source"))
+                    retriever = data.get("retriever")
+
+                    reusable = _find_reusable_share_agent(
+                        conn, user,
+                        prompt_pg_id=prompt_pg_id,
+                        chunks=chunks_int,
+                        source_pg_id=source_pg_id,
+                        retriever=retriever,
+                    )
+                    if reusable:
+                        api_uuid = reusable.get("key")
+                    else:
+                        api_uuid = str(uuid.uuid4())
+                        name = (conversation.get("name") or "") + "(shared)"
+                        agents_repo.create(
+                            user,
+                            name,
+                            "published",
+                            key=api_uuid,
+                            retriever=retriever,
+                            chunks=chunks_int,
+                            prompt_id=prompt_pg_id,
+                            source_id=source_pg_id,
+                        )
+
+                    share = shared_repo.get_or_create(
+                        conv_pg_id,
                         user,
-                        is_promptable=is_promptable,
+                        is_promptable=True,
                         first_n_queries=current_n_queries,
                         api_key=api_uuid,
-                        prompt_id=prompt_id,
-                        chunks=int(chunks) if chunks else None,
+                        prompt_id=prompt_pg_id,
+                        chunks=chunks_int,
                     )
                     return make_response(
                         jsonify(
                             {
                                 "success": True,
-                                "identifier": str(explicit_binary.as_uuid()),
+                                "identifier": str(share["uuid"]),
                             }
                         ),
-                        201,
+                        201 if reusable is None else 200,
                     )
-            pre_existing = shared_conversations_collections.find_one(
-                {
-                    "conversation_id": ObjectId(conversation_id),
-                    "isPromptable": is_promptable,
-                    "first_n_queries": current_n_queries,
-                    "user": user,
-                }
-            )
-            if pre_existing is not None:
-                return make_response(
-                    jsonify(
-                        {
-                            "success": True,
-                            "identifier": str(pre_existing["uuid"].as_uuid()),
-                        }
-                    ),
-                    200,
-                )
-            else:
-                shared_conversations_collections.insert_one(
-                    {
-                        "uuid": explicit_binary,
-                        "conversation_id": ObjectId(conversation_id),
-                        "isPromptable": is_promptable,
-                        "first_n_queries": current_n_queries,
-                        "user": user,
-                    }
-                )
-                _dual_write_share(
-                    conversation_id,
-                    str(explicit_binary.as_uuid()),
+
+                # Non-promptable share path.
+                share = shared_repo.get_or_create(
+                    conv_pg_id,
                     user,
-                    is_promptable=is_promptable,
+                    is_promptable=False,
                     first_n_queries=current_n_queries,
                     api_key=None,
                 )
                 return make_response(
                     jsonify(
-                        {"success": True, "identifier": str(explicit_binary.as_uuid())}
+                        {
+                            "success": True,
+                            "identifier": str(share["uuid"]),
+                        }
                     ),
                     201,
                 )
@@ -301,37 +258,13 @@ class GetPubliclySharedConversations(Resource):
     @api.doc(description="Get publicly shared conversations by identifier")
     def get(self, identifier: str):
         try:
-            query_uuid = Binary.from_uuid(
-                uuid.UUID(identifier), UuidRepresentation.STANDARD
-            )
-            shared = shared_conversations_collections.find_one({"uuid": query_uuid})
-            conversation_queries = []
+            with db_readonly() as conn:
+                shared_repo = SharedConversationsRepository(conn)
+                conv_repo = ConversationsRepository(conn)
+                attach_repo = AttachmentsRepository(conn)
 
-            if (
-                shared
-                and "conversation_id" in shared
-            ):
-                # Handle DBRef (legacy), ObjectId, dict, and string formats for conversation_id
-                conversation_id = shared["conversation_id"]
-                if isinstance(conversation_id, DBRef):
-                    conversation_id = conversation_id.id
-                elif isinstance(conversation_id, dict):
-                    # Handle dict representation of DBRef (e.g., {"$ref": "...", "$id": "..."})
-                    if "$id" in conversation_id:
-                        conv_id = conversation_id["$id"]
-                        # $id might be a dict like {"$oid": "..."} or a string
-                        if isinstance(conv_id, dict) and "$oid" in conv_id:
-                            conversation_id = ObjectId(conv_id["$oid"])
-                        else:
-                            conversation_id = ObjectId(conv_id)
-                    elif "_id" in conversation_id:
-                        conversation_id = ObjectId(conversation_id["_id"])
-                elif isinstance(conversation_id, str):
-                    conversation_id = ObjectId(conversation_id)
-                conversation = conversations_collection.find_one(
-                    {"_id": conversation_id}
-                )
-                if conversation is None:
+                shared = shared_repo.find_by_uuid(identifier)
+                if not shared or not shared.get("conversation_id"):
                     return make_response(
                         jsonify(
                             {
@@ -341,22 +274,60 @@ class GetPubliclySharedConversations(Resource):
                         ),
                         404,
                     )
-                conversation_queries = conversation["queries"][
-                    : (shared["first_n_queries"])
-                ]
+                conv_pg_id = str(shared["conversation_id"])
+                owner_user = shared.get("user_id")
 
-                for query in conversation_queries:
-                    if "attachments" in query and query["attachments"]:
+                conversation = conv_repo.get_owned(conv_pg_id, owner_user) if owner_user else None
+                if conversation is None:
+                    # Fall back to any-user lookup in case shared row's
+                    # user_id is missing — still keyed by PG UUID.
+                    row = conn.execute(
+                        _sql_text(
+                            "SELECT * FROM conversations WHERE id = CAST(:id AS uuid)"
+                        ),
+                        {"id": conv_pg_id},
+                    ).fetchone()
+                    if row is None:
+                        return make_response(
+                            jsonify(
+                                {
+                                    "success": False,
+                                    "error": "might have broken url or the conversation does not exist",
+                                }
+                            ),
+                            404,
+                        )
+                    conversation = dict(row._mapping)
+
+                messages = conv_repo.get_messages(conv_pg_id)
+                first_n = shared.get("first_n_queries") or 0
+                conversation_queries = []
+                for msg in messages[:first_n]:
+                    query = {
+                        "prompt": msg.get("prompt"),
+                        "response": msg.get("response"),
+                        "thought": msg.get("thought"),
+                        "sources": msg.get("sources") or [],
+                        "tool_calls": msg.get("tool_calls") or [],
+                        "timestamp": (
+                            msg["timestamp"].isoformat()
+                            if hasattr(msg.get("timestamp"), "isoformat")
+                            else msg.get("timestamp")
+                        ),
+                        "feedback": msg.get("feedback"),
+                    }
+                    attachments = msg.get("attachments") or []
+                    if attachments:
                         attachment_details = []
-                        for attachment_id in query["attachments"]:
+                        for attachment_id in attachments:
                             try:
-                                attachment = attachments_collection.find_one(
-                                    {"_id": ObjectId(attachment_id)}
-                                )
+                                attachment = attach_repo.get_any(
+                                    str(attachment_id), owner_user,
+                                ) if owner_user else None
                                 if attachment:
                                     attachment_details.append(
                                         {
-                                            "id": str(attachment["_id"]),
+                                            "id": str(attachment["id"]),
                                             "fileName": attachment.get(
                                                 "filename", "Unknown file"
                                             ),
@@ -368,26 +339,23 @@ class GetPubliclySharedConversations(Resource):
                                     exc_info=True,
                                 )
                         query["attachments"] = attachment_details
-            else:
-                return make_response(
-                    jsonify(
-                        {
-                            "success": False,
-                            "error": "might have broken url or the conversation does not exist",
-                        }
-                    ),
-                    404,
+                    conversation_queries.append(query)
+
+                created = conversation.get("created_at") or conversation.get("date")
+                date_iso = (
+                    created.isoformat()
+                    if hasattr(created, "isoformat")
+                    else (str(created) if created is not None else None)
                 )
-            date = conversation["_id"].generation_time.isoformat()
-            res = {
-                "success": True,
-                "queries": conversation_queries,
-                "title": conversation["name"],
-                "timestamp": date,
-            }
-            if shared["isPromptable"] and "api_key" in shared:
-                res["api_key"] = shared["api_key"]
-            return make_response(jsonify(res), 200)
+                res = {
+                    "success": True,
+                    "queries": conversation_queries,
+                    "title": conversation.get("name"),
+                    "timestamp": date_iso,
+                }
+                if shared.get("is_promptable") and shared.get("api_key"):
+                    res["api_key"] = shared["api_key"]
+                return make_response(jsonify(res), 200)
         except Exception as err:
             current_app.logger.error(
                 f"Error getting shared conversation: {err}", exc_info=True

application/api/user/sources/chunks.py

@@ -1,11 +1,12 @@
 """Source document management chunk management."""
 
-from bson.objectid import ObjectId
 from flask import current_app, jsonify, make_response, request
 from flask_restx import fields, Namespace, Resource
 
 from application.api import api
-from application.api.user.base import get_vector_store, sources_collection
+from application.api.user.base import get_vector_store
+from application.storage.db.repositories.sources import SourcesRepository
+from application.storage.db.session import db_readonly
 from application.utils import check_required_fields, num_tokens_from_string
 
 sources_chunks_ns = Namespace(
@@ -13,6 +14,15 @@ sources_chunks_ns = Namespace(
 )
 
 
+def _resolve_source(doc_id: str, user: str):
+    """Resolve a source (UUID or legacy ObjectId) for the caller.
+
+    Returns the row dict (with PG UUID in ``id``) or ``None`` if missing.
+    """
+    with db_readonly() as conn:
+        return SourcesRepository(conn).get_any(doc_id, user)
+
+
 @sources_chunks_ns.route("/get_chunks")
 class GetChunks(Resource):
     @api.doc(
@@ -36,36 +46,34 @@ class GetChunks(Resource):
         path = request.args.get("path")
         search_term = request.args.get("search", "").strip().lower()
 
-        if not ObjectId.is_valid(doc_id):
+        if not doc_id:
+            return make_response(jsonify({"error": "Invalid doc_id"}), 400)
+        try:
+            doc = _resolve_source(doc_id, user)
+        except Exception as e:
+            current_app.logger.error(f"Error resolving source: {e}", exc_info=True)
             return make_response(jsonify({"error": "Invalid doc_id"}), 400)
-        doc = sources_collection.find_one({"_id": ObjectId(doc_id), "user": user})
         if not doc:
             return make_response(
                 jsonify({"error": "Document not found or access denied"}), 404
             )
+        resolved_id = str(doc["id"])
         try:
-            store = get_vector_store(doc_id)
+            store = get_vector_store(resolved_id)
             chunks = store.get_chunks()
 
             filtered_chunks = []
             for chunk in chunks:
                 metadata = chunk.get("metadata", {})
 
-                # Filter by path if provided
-
                 if path:
                     chunk_source = metadata.get("source", "")
                     chunk_file_path = metadata.get("file_path", "")
-                    # Check if the chunk matches the requested path
-                    # For file uploads: source ends with path (e.g., "inputs/.../file.pdf" ends with "file.pdf")
-                    # For crawlers: file_path ends with path (e.g., "guides/setup.md" ends with "setup.md")
                     source_match = chunk_source and chunk_source.endswith(path)
                     file_path_match = chunk_file_path and chunk_file_path.endswith(path)
 
                     if not (source_match or file_path_match):
                         continue
-                # Filter by search term if provided
-
                 if search_term:
                     text_match = search_term in chunk.get("text", "").lower()
                     title_match = search_term in metadata.get("title", "").lower()
@@ -132,15 +140,17 @@ class AddChunk(Resource):
         token_count = num_tokens_from_string(text)
         metadata["token_count"] = token_count
 
-        if not ObjectId.is_valid(doc_id):
+        try:
+            doc = _resolve_source(doc_id, user)
+        except Exception as e:
+            current_app.logger.error(f"Error resolving source: {e}", exc_info=True)
             return make_response(jsonify({"error": "Invalid doc_id"}), 400)
-        doc = sources_collection.find_one({"_id": ObjectId(doc_id), "user": user})
         if not doc:
             return make_response(
                 jsonify({"error": "Document not found or access denied"}), 404
             )
         try:
-            store = get_vector_store(doc_id)
+            store = get_vector_store(str(doc["id"]))
             chunk_id = store.add_chunk(text, metadata)
             return make_response(
                 jsonify({"message": "Chunk added successfully", "chunk_id": chunk_id}),
@@ -165,15 +175,17 @@ class DeleteChunk(Resource):
         doc_id = request.args.get("id")
         chunk_id = request.args.get("chunk_id")
 
-        if not ObjectId.is_valid(doc_id):
+        try:
+            doc = _resolve_source(doc_id, user)
+        except Exception as e:
+            current_app.logger.error(f"Error resolving source: {e}", exc_info=True)
             return make_response(jsonify({"error": "Invalid doc_id"}), 400)
-        doc = sources_collection.find_one({"_id": ObjectId(doc_id), "user": user})
         if not doc:
             return make_response(
                 jsonify({"error": "Document not found or access denied"}), 404
             )
         try:
-            store = get_vector_store(doc_id)
+            store = get_vector_store(str(doc["id"]))
             deleted = store.delete_chunk(chunk_id)
             if deleted:
                 return make_response(
@@ -232,15 +244,17 @@ class UpdateChunk(Resource):
             if metadata is None:
                 metadata = {}
             metadata["token_count"] = token_count
-        if not ObjectId.is_valid(doc_id):
+        try:
+            doc = _resolve_source(doc_id, user)
+        except Exception as e:
+            current_app.logger.error(f"Error resolving source: {e}", exc_info=True)
             return make_response(jsonify({"error": "Invalid doc_id"}), 400)
-        doc = sources_collection.find_one({"_id": ObjectId(doc_id), "user": user})
         if not doc:
             return make_response(
                 jsonify({"error": "Document not found or access denied"}), 404
             )
         try:
-            store = get_vector_store(doc_id)
+            store = get_vector_store(str(doc["id"]))
 
             chunks = store.get_chunks()
             existing_chunk = next((c for c in chunks if c["doc_id"] == chunk_id), None)

application/api/user/sources/routes.py

@@ -3,14 +3,14 @@
 import json
 import math
 
-from bson.objectid import ObjectId
 from flask import current_app, jsonify, make_response, redirect, request
 from flask_restx import fields, Namespace, Resource
 
 from application.api import api
-from application.api.user.base import sources_collection
 from application.api.user.tasks import sync_source
 from application.core.settings import settings
+from application.storage.db.repositories.sources import SourcesRepository
+from application.storage.db.session import db_readonly, db_session
 from application.storage.storage_creator import StorageCreator
 from application.utils import check_required_fields
 from application.vectorstore.vector_creator import VectorCreator
@@ -56,11 +56,20 @@ class CombinedJson(Resource):
         ]
 
         try:
-            for index in sources_collection.find({"user": user}).sort("date", -1):
+            with db_readonly() as conn:
+                indexes = SourcesRepository(conn).list_for_user(user)
+            # list_for_user sorts by created_at DESC; legacy shape sorted by
+            # "date" DESC. Both are monotonic on creation so the ordering is
+            # equivalent for dev; re-sort defensively.
+            indexes = sorted(
+                indexes, key=lambda r: r.get("date") or r.get("created_at") or "",
+                reverse=True,
+            )
+            for index in indexes:
                 provider = _get_provider_from_remote_data(index.get("remote_data"))
                 data.append(
                     {
-                        "id": str(index["_id"]),
+                        "id": str(index["id"]),
                         "name": index.get("name"),
                         "date": index.get("date"),
                         "model": settings.EMBEDDINGS_NAME,
@@ -70,9 +79,7 @@ class CombinedJson(Resource):
                         "syncFrequency": index.get("sync_frequency", ""),
                         "provider": provider,
                         "is_nested": bool(index.get("directory_structure")),
-                        "type": index.get(
-                            "type", "file"
-                        ),  # Add type field with default "file"
+                        "type": index.get("type", "file"),
                     }
                 )
         except Exception as err:
@@ -89,61 +96,55 @@ class PaginatedSources(Resource):
         if not decoded_token:
             return make_response(jsonify({"success": False}), 401)
         user = decoded_token.get("sub")
-        sort_field = request.args.get("sort", "date")  # Default to 'date'
-        sort_order = request.args.get("order", "desc")  # Default to 'desc'
-        page = int(request.args.get("page", 1))  # Default to 1
-        rows_per_page = int(request.args.get("rows", 10))  # Default to 10
-        # add .strip() to remove leading and trailing whitespaces
-
-        search_term = request.args.get(
-            "search", ""
-        ).strip()  # add search for filter documents
-
-        # Prepare query for filtering
-
-        query = {"user": user}
-        if search_term:
-            query["name"] = {
-                "$regex": search_term,
-                "$options": "i",  # using case-insensitive search
-            }
-        total_documents = sources_collection.count_documents(query)
-        total_pages = max(1, math.ceil(total_documents / rows_per_page))
-        page = min(
-            max(1, page), total_pages
-        )  # add this to make sure page inbound is within the range
-        sort_order = 1 if sort_order == "asc" else -1
-        skip = (page - 1) * rows_per_page
+        sort_field = request.args.get("sort", "date")
+        sort_order = request.args.get("order", "desc")
+        page = max(1, int(request.args.get("page", 1)))
+        rows_per_page = max(1, int(request.args.get("rows", 10)))
+        search_term = request.args.get("search", "").strip() or None
 
         try:
-            documents = (
-                sources_collection.find(query)
-                .sort(sort_field, sort_order)
-                .skip(skip)
-                .limit(rows_per_page)
-            )
+            with db_readonly() as conn:
+                repo = SourcesRepository(conn)
+                total_documents = repo.count_for_user(
+                    user, search_term=search_term,
+                )
+                # Prior in-Python implementation returned ``totalPages = 1``
+                # for empty result sets (``max(1, ceil(0/rows))``); we
+                # preserve that contract so the frontend pager stays stable.
+                total_pages = max(1, math.ceil(total_documents / rows_per_page))
+                effective_page = min(page, total_pages)
+                offset = (effective_page - 1) * rows_per_page
+                window = repo.list_for_user(
+                    user,
+                    limit=rows_per_page,
+                    offset=offset,
+                    search_term=search_term,
+                    sort_field=sort_field,
+                    sort_order=sort_order,
+                )
 
             paginated_docs = []
-            for doc in documents:
+            for doc in window:
                 provider = _get_provider_from_remote_data(doc.get("remote_data"))
-                doc_data = {
-                    "id": str(doc["_id"]),
-                    "name": doc.get("name", ""),
-                    "date": doc.get("date", ""),
-                    "model": settings.EMBEDDINGS_NAME,
-                    "location": "local",
-                    "tokens": doc.get("tokens", ""),
-                    "retriever": doc.get("retriever", "classic"),
-                    "syncFrequency": doc.get("sync_frequency", ""),
-                    "provider": provider,
-                    "isNested": bool(doc.get("directory_structure")),
-                    "type": doc.get("type", "file"),
-                }
-                paginated_docs.append(doc_data)
+                paginated_docs.append(
+                    {
+                        "id": str(doc["id"]),
+                        "name": doc.get("name", ""),
+                        "date": doc.get("date", ""),
+                        "model": settings.EMBEDDINGS_NAME,
+                        "location": "local",
+                        "tokens": doc.get("tokens", ""),
+                        "retriever": doc.get("retriever", "classic"),
+                        "syncFrequency": doc.get("sync_frequency", ""),
+                        "provider": provider,
+                        "isNested": bool(doc.get("directory_structure")),
+                        "type": doc.get("type", "file"),
+                    }
+                )
             response = {
                 "total": total_documents,
                 "totalPages": total_pages,
-                "currentPage": page,
+                "currentPage": effective_page,
                 "paginated": paginated_docs,
             }
             return make_response(jsonify(response), 200)
@@ -154,28 +155,6 @@ class PaginatedSources(Resource):
             return make_response(jsonify({"success": False}), 400)
 
 
-@sources_ns.route("/delete_by_ids")
-class DeleteByIds(Resource):
-    @api.doc(
-        description="Deletes documents from the vector store by IDs",
-        params={"path": "Comma-separated list of IDs"},
-    )
-    def get(self):
-        ids = request.args.get("path")
-        if not ids:
-            return make_response(
-                jsonify({"success": False, "message": "Missing required fields"}), 400
-            )
-        try:
-            result = sources_collection.delete_index(ids=ids)
-            if result:
-                return make_response(jsonify({"success": True}), 200)
-        except Exception as err:
-            current_app.logger.error(f"Error deleting indexes: {err}", exc_info=True)
-            return make_response(jsonify({"success": False}), 400)
-        return make_response(jsonify({"success": False}), 400)
-
-
 @sources_ns.route("/delete_old")
 class DeleteOldIndexes(Resource):
     @api.doc(
@@ -186,30 +165,33 @@ class DeleteOldIndexes(Resource):
         decoded_token = request.decoded_token
         if not decoded_token:
             return make_response(jsonify({"success": False}), 401)
+        user = decoded_token.get("sub")
         source_id = request.args.get("source_id")
         if not source_id:
             return make_response(
                 jsonify({"success": False, "message": "Missing required fields"}), 400
             )
-        doc = sources_collection.find_one(
-            {"_id": ObjectId(source_id), "user": decoded_token.get("sub")}
-        )
+        try:
+            with db_readonly() as conn:
+                doc = SourcesRepository(conn).get_any(source_id, user)
+        except Exception as err:
+            current_app.logger.error(f"Error looking up source: {err}", exc_info=True)
+            return make_response(jsonify({"success": False}), 400)
         if not doc:
             return make_response(jsonify({"status": "not found"}), 404)
         storage = StorageCreator.get_storage()
+        resolved_id = str(doc["id"])
 
         try:
-            # Delete vector index
-
             if settings.VECTOR_STORE == "faiss":
-                index_path = f"indexes/{str(doc['_id'])}"
+                index_path = f"indexes/{resolved_id}"
                 if storage.file_exists(f"{index_path}/index.faiss"):
                     storage.delete_file(f"{index_path}/index.faiss")
                 if storage.file_exists(f"{index_path}/index.pkl"):
                     storage.delete_file(f"{index_path}/index.pkl")
             else:
                 vectorstore = VectorCreator.create_vectorstore(
-                    settings.VECTOR_STORE, source_id=str(doc["_id"])
+                    settings.VECTOR_STORE, source_id=resolved_id
                 )
                 vectorstore.delete_index()
             if "file_path" in doc and doc["file_path"]:
@@ -227,7 +209,14 @@ class DeleteOldIndexes(Resource):
                 f"Error deleting files and indexes: {err}", exc_info=True
             )
             return make_response(jsonify({"success": False}), 400)
-        sources_collection.delete_one({"_id": ObjectId(source_id)})
+        try:
+            with db_session() as conn:
+                SourcesRepository(conn).delete(resolved_id, user)
+        except Exception as err:
+            current_app.logger.error(
+                f"Error deleting source row: {err}", exc_info=True
+            )
+            return make_response(jsonify({"success": False}), 400)
         return make_response(jsonify({"success": True}), 200)
 
 
@@ -272,15 +261,16 @@ class ManageSync(Resource):
             return make_response(
                 jsonify({"success": False, "message": "Invalid frequency"}), 400
             )
-        update_data = {"$set": {"sync_frequency": sync_frequency}}
         try:
-            sources_collection.update_one(
-                {
-                    "_id": ObjectId(source_id),
-                    "user": user,
-                },
-                update_data,
-            )
+            with db_session() as conn:
+                repo = SourcesRepository(conn)
+                doc = repo.get_any(source_id, user)
+                if doc is None:
+                    return make_response(
+                        jsonify({"success": False, "message": "Source not found"}),
+                        404,
+                    )
+                repo.update(str(doc["id"]), user, {"sync_frequency": sync_frequency})
         except Exception as err:
             current_app.logger.error(
                 f"Error updating sync frequency: {err}", exc_info=True
@@ -309,19 +299,20 @@ class SyncSource(Resource):
         if missing_fields:
             return missing_fields
         source_id = data["source_id"]
-        if not ObjectId.is_valid(source_id):
+        try:
+            with db_readonly() as conn:
+                doc = SourcesRepository(conn).get_any(source_id, user)
+        except Exception as err:
+            current_app.logger.error(f"Error looking up source: {err}", exc_info=True)
             return make_response(
                 jsonify({"success": False, "message": "Invalid source ID"}), 400
             )
-        doc = sources_collection.find_one(
-            {"_id": ObjectId(source_id), "user": user}
-        )
         if not doc:
             return make_response(
                 jsonify({"success": False, "message": "Source not found"}), 404
             )
         source_type = doc.get("type", "")
-        if source_type.startswith("connector"):
+        if source_type and source_type.startswith("connector"):
             return make_response(
                 jsonify(
                     {
@@ -344,7 +335,7 @@ class SyncSource(Resource):
                 loader=source_type,
                 sync_frequency=doc.get("sync_frequency", "never"),
                 retriever=doc.get("retriever", "classic"),
-                doc_id=source_id,
+                doc_id=str(doc["id"]),
             )
         except Exception as err:
             current_app.logger.error(
@@ -370,10 +361,9 @@ class DirectoryStructure(Resource):
 
         if not doc_id:
             return make_response(jsonify({"error": "Document ID is required"}), 400)
-        if not ObjectId.is_valid(doc_id):
-            return make_response(jsonify({"error": "Invalid document ID"}), 400)
         try:
-            doc = sources_collection.find_one({"_id": ObjectId(doc_id), "user": user})
+            with db_readonly() as conn:
+                doc = SourcesRepository(conn).get_any(doc_id, user)
             if not doc:
                 return make_response(
                     jsonify({"error": "Document not found or access denied"}), 404
@@ -387,6 +377,8 @@ class DirectoryStructure(Resource):
                 if isinstance(remote_data, str) and remote_data:
                     remote_data_obj = json.loads(remote_data)
                     provider = remote_data_obj.get("provider")
+                elif isinstance(remote_data, dict):
+                    provider = remote_data.get("provider")
             except Exception as e:
                 current_app.logger.warning(
                     f"Failed to parse remote_data for doc {doc_id}: {e}"
@@ -406,4 +398,7 @@ class DirectoryStructure(Resource):
             current_app.logger.error(
                 f"Error retrieving directory structure: {e}", exc_info=True
             )
-            return make_response(jsonify({"success": False, "error": "Failed to retrieve directory structure"}), 500)
+            return make_response(
+                jsonify({"success": False, "error": "Failed to retrieve directory structure"}),
+                500,
+            )

application/api/user/sources/upload.py

@@ -5,16 +5,16 @@ import os
 import tempfile
 import zipfile
 
-from bson.objectid import ObjectId
 from flask import current_app, jsonify, make_response, request
 from flask_restx import fields, Namespace, Resource
 
 from application.api import api
-from application.api.user.base import sources_collection
 from application.api.user.tasks import ingest, ingest_connector_task, ingest_remote
 from application.core.settings import settings
 from application.parser.connectors.connector_creator import ConnectorCreator
 from application.parser.file.constants import SUPPORTED_SOURCE_EXTENSIONS
+from application.storage.db.repositories.sources import SourcesRepository
+from application.storage.db.session import db_readonly, db_session
 from application.storage.storage_creator import StorageCreator
 from application.stt.upload_limits import (
     AudioFileTooLargeError,
@@ -329,15 +329,8 @@ class ManageSourceFiles(Resource):
                 400,
             )
         try:
-            ObjectId(source_id)
-        except Exception:
-            return make_response(
-                jsonify({"success": False, "message": "Invalid source ID format"}), 400
-            )
-        try:
-            source = sources_collection.find_one(
-                {"_id": ObjectId(source_id), "user": user}
-            )
+            with db_readonly() as conn:
+                source = SourcesRepository(conn).get_any(source_id, user)
             if not source:
                 return make_response(
                     jsonify(
@@ -353,6 +346,7 @@ class ManageSourceFiles(Resource):
             return make_response(
                 jsonify({"success": False, "message": "Database error"}), 500
             )
+        resolved_source_id = str(source["id"])
         try:
             storage = StorageCreator.get_storage()
             source_file_path = source.get("file_path", "")
@@ -411,15 +405,18 @@ class ManageSourceFiles(Resource):
                             map_updated = True
 
                 if map_updated:
-                    sources_collection.update_one(
-                        {"_id": ObjectId(source_id)},
-                        {"$set": {"file_name_map": file_name_map}},
-                    )
+                    with db_session() as conn:
+                        SourcesRepository(conn).update(
+                            resolved_source_id, user,
+                            {"file_name_map": dict(file_name_map)},
+                        )
                 # Trigger re-ingestion pipeline
 
                 from application.api.user.tasks import reingest_source_task
 
-                task = reingest_source_task.delay(source_id=source_id, user=user)
+                task = reingest_source_task.delay(
+                    source_id=resolved_source_id, user=user
+                )
 
                 return make_response(
                     jsonify(
@@ -485,15 +482,18 @@ class ManageSourceFiles(Resource):
                         map_updated = True
 
                 if map_updated and isinstance(file_name_map, dict):
-                    sources_collection.update_one(
-                        {"_id": ObjectId(source_id)},
-                        {"$set": {"file_name_map": file_name_map}},
-                    )
+                    with db_session() as conn:
+                        SourcesRepository(conn).update(
+                            resolved_source_id, user,
+                            {"file_name_map": dict(file_name_map)},
+                        )
                 # Trigger re-ingestion pipeline
 
                 from application.api.user.tasks import reingest_source_task
 
-                task = reingest_source_task.delay(source_id=source_id, user=user)
+                task = reingest_source_task.delay(
+                    source_id=resolved_source_id, user=user
+                )
 
                 return make_response(
                     jsonify(
@@ -581,16 +581,19 @@ class ManageSourceFiles(Resource):
                     if keys_to_remove:
                         for key in keys_to_remove:
                             file_name_map.pop(key, None)
-                        sources_collection.update_one(
-                            {"_id": ObjectId(source_id)},
-                            {"$set": {"file_name_map": file_name_map}},
-                        )
+                        with db_session() as conn:
+                            SourcesRepository(conn).update(
+                                resolved_source_id, user,
+                                {"file_name_map": dict(file_name_map)},
+                            )
 
                 # Trigger re-ingestion pipeline
 
                 from application.api.user.tasks import reingest_source_task
 
-                task = reingest_source_task.delay(source_id=source_id, user=user)
+                task = reingest_source_task.delay(
+                    source_id=resolved_source_id, user=user
+                )
 
                 return make_response(
                     jsonify(

application/api/user/tasks.py

@@ -140,6 +140,11 @@ def setup_periodic_tasks(sender, **kwargs):
         cleanup_pending_tool_state.s(),
         name="cleanup-pending-tool-state",
     )
+    sender.add_periodic_task(
+        timedelta(hours=7),
+        version_check_task.s(),
+        name="version-check",
+    )
 
 
 @celery.task(bind=True)
@@ -176,3 +181,16 @@ def cleanup_pending_tool_state(self):
     with engine.begin() as conn:
         deleted = PendingToolStateRepository(conn).cleanup_expired()
     return {"deleted": deleted}
+
+
+@celery.task(bind=True)
+def version_check_task(self):
+    """Periodic anonymous version check.
+
+    Complements the ``worker_ready`` boot trigger so long-running
+    deployments (>6h cache TTL) still refresh advisories. ``run_check``
+    is fail-silent and coordinates across replicas via Redis lock +
+    cache (see ``application.updates.version_check``).
+    """
+    from application.updates.version_check import run_check
+    run_check()

application/api/user/tools/mcp.py

@@ -3,27 +3,24 @@
 import json
 from urllib.parse import urlencode, urlparse
 
-from bson.objectid import ObjectId
 from flask import current_app, jsonify, make_response, redirect, request
 from flask_restx import Namespace, Resource, fields
 
 from application.agents.tools.mcp_tool import MCPOAuthManager, MCPTool
 from application.api import api
-from application.api.user.base import user_tools_collection
 from application.api.user.tools.routes import transform_actions
 from application.cache import get_redis_instance
-from application.core.mongo_db import MongoDB
-from application.core.settings import settings
 from application.core.url_validation import SSRFError, validate_url
 from application.security.encryption import decrypt_credentials, encrypt_credentials
+from application.storage.db.repositories.connector_sessions import (
+    ConnectorSessionsRepository,
+)
+from application.storage.db.repositories.user_tools import UserToolsRepository
+from application.storage.db.session import db_readonly, db_session
 from application.utils import check_required_fields
 
 tools_mcp_ns = Namespace("tools", description="Tool management operations", path="/api")
 
-_mongo = MongoDB.get_client()
-_db = _mongo[settings.MONGO_DB_NAME]
-_connector_sessions = _db["connector_sessions"]
-
 _ALLOWED_TRANSPORTS = {"auto", "sse", "http"}
 
 
@@ -252,15 +249,18 @@ class MCPServerSave(Resource):
             storage_config = config.copy()
 
             tool_id = data.get("id")
+            existing_doc = None
             existing_encrypted = None
             if tool_id:
-                existing_doc = user_tools_collection.find_one(
-                    {"_id": ObjectId(tool_id), "user": user, "name": "mcp_tool"}
-                )
-                if existing_doc:
-                    existing_encrypted = existing_doc.get("config", {}).get(
+                with db_readonly() as conn:
+                    repo = UserToolsRepository(conn)
+                    existing_doc = repo.get_any(tool_id, user)
+                if existing_doc and existing_doc.get("name") == "mcp_tool":
+                    existing_encrypted = (existing_doc.get("config") or {}).get(
                         "encrypted_credentials"
                     )
+                else:
+                    existing_doc = None
 
             if auth_credentials:
                 if existing_encrypted:
@@ -283,47 +283,88 @@ class MCPServerSave(Resource):
             ]:
                 storage_config.pop(field, None)
             transformed_actions = transform_actions(actions_metadata)
-            tool_data = {
-                "name": "mcp_tool",
-                "displayName": data["displayName"],
-                "customName": data["displayName"],
-                "description": f"MCP Server: {storage_config.get('server_url', 'Unknown')}",
-                "config": storage_config,
-                "actions": transformed_actions,
-                "status": data.get("status", True),
-                "user": user,
-            }
 
-            if tool_id:
-                result = user_tools_collection.update_one(
-                    {"_id": ObjectId(tool_id), "user": user, "name": "mcp_tool"},
-                    {"$set": {k: v for k, v in tool_data.items() if k != "user"}},
-                )
-                if result.matched_count == 0:
-                    return make_response(
-                        jsonify(
-                            {
-                                "success": False,
-                                "error": "Tool not found or access denied",
-                            }
-                        ),
-                        404,
+            display_name = data["displayName"]
+            description = f"MCP Server: {storage_config.get('server_url', 'Unknown')}"
+            status_bool = bool(data.get("status", True))
+
+            with db_session() as conn:
+                repo = UserToolsRepository(conn)
+                if existing_doc:
+                    repo.update(
+                        str(existing_doc["id"]), user,
+                        {
+                            "display_name": display_name,
+                            "custom_name": display_name,
+                            "description": description,
+                            "config": storage_config,
+                            "actions": transformed_actions,
+                            "status": status_bool,
+                        },
                     )
-                response_data = {
-                    "success": True,
-                    "id": tool_id,
-                    "message": f"MCP server updated successfully! Discovered {len(transformed_actions)} tools.",
-                    "tools_count": len(transformed_actions),
-                }
-            else:
-                result = user_tools_collection.insert_one(tool_data)
-                tool_id = str(result.inserted_id)
-                response_data = {
-                    "success": True,
-                    "id": tool_id,
-                    "message": f"MCP server created successfully! Discovered {len(transformed_actions)} tools.",
-                    "tools_count": len(transformed_actions),
-                }
+                    saved_id = str(existing_doc["id"])
+                    response_data = {
+                        "success": True,
+                        "id": saved_id,
+                        "message": f"MCP server updated successfully! Discovered {len(transformed_actions)} tools.",
+                        "tools_count": len(transformed_actions),
+                    }
+                else:
+                    # Fall back to find_by_user_and_name — the original
+                    # dual-write path also ran an existence check before
+                    # deciding between insert and update.
+                    existing_by_name = repo.find_by_user_and_name(user, "mcp_tool")
+                    if tool_id is None and existing_by_name and (
+                        (existing_by_name.get("config") or {}).get("server_url")
+                        == storage_config.get("server_url")
+                    ):
+                        repo.update(
+                            str(existing_by_name["id"]), user,
+                            {
+                                "display_name": display_name,
+                                "custom_name": display_name,
+                                "description": description,
+                                "config": storage_config,
+                                "actions": transformed_actions,
+                                "status": status_bool,
+                            },
+                        )
+                        saved_id = str(existing_by_name["id"])
+                        response_data = {
+                            "success": True,
+                            "id": saved_id,
+                            "message": f"MCP server updated successfully! Discovered {len(transformed_actions)} tools.",
+                            "tools_count": len(transformed_actions),
+                        }
+                    else:
+                        created = repo.create(
+                            user, "mcp_tool",
+                            config=storage_config,
+                            custom_name=display_name,
+                            display_name=display_name,
+                            description=description,
+                            config_requirements={},
+                            actions=transformed_actions,
+                            status=status_bool,
+                        )
+                        saved_id = str(created["id"])
+                        response_data = {
+                            "success": True,
+                            "id": saved_id,
+                            "message": f"MCP server created successfully! Discovered {len(transformed_actions)} tools.",
+                            "tools_count": len(transformed_actions),
+                        }
+                    if tool_id and existing_doc is None:
+                        # Client requested update on a non-existent tool id.
+                        return make_response(
+                            jsonify(
+                                {
+                                    "success": False,
+                                    "error": "Tool not found or access denied",
+                                }
+                            ),
+                            404,
+                        )
             return make_response(jsonify(response_data), 200)
         except ValueError as e:
             current_app.logger.warning(f"Invalid MCP server save request: {e}")
@@ -459,49 +500,59 @@ class MCPAuthStatus(Resource):
             return make_response(jsonify({"success": False}), 401)
         user = decoded_token.get("sub")
         try:
-            mcp_tools = list(
-                user_tools_collection.find(
-                    {"user": user, "name": "mcp_tool"},
-                    {"_id": 1, "config": 1},
-                )
-            )
-            if not mcp_tools:
-                return make_response(jsonify({"success": True, "statuses": {}}), 200)
-
-            oauth_server_urls = {}
-            statuses = {}
-            for tool in mcp_tools:
-                tool_id = str(tool["_id"])
-                config = tool.get("config", {})
-                auth_type = config.get("auth_type", "none")
-                if auth_type == "oauth":
-                    server_url = config.get("server_url", "")
-                    if server_url:
-                        parsed = urlparse(server_url)
-                        base_url = f"{parsed.scheme}://{parsed.netloc}"
-                        oauth_server_urls[tool_id] = base_url
-                    else:
-                        statuses[tool_id] = "needs_auth"
-                else:
-                    statuses[tool_id] = "configured"
-
-            if oauth_server_urls:
-                unique_urls = list(set(oauth_server_urls.values()))
-                sessions = list(
-                    _connector_sessions.find(
-                        {"user_id": user, "server_url": {"$in": unique_urls}},
-                        {"server_url": 1, "tokens": 1},
+            with db_readonly() as conn:
+                tools_repo = UserToolsRepository(conn)
+                sessions_repo = ConnectorSessionsRepository(conn)
+                all_tools = tools_repo.list_for_user(user)
+                mcp_tools = [t for t in all_tools if t.get("name") == "mcp_tool"]
+                if not mcp_tools:
+                    return make_response(
+                        jsonify({"success": True, "statuses": {}}), 200
                     )
-                )
-                url_has_tokens = {
-                    doc["server_url"]: bool(doc.get("tokens", {}).get("access_token"))
-                    for doc in sessions
-                }
-                for tool_id, base_url in oauth_server_urls.items():
-                    if url_has_tokens.get(base_url):
-                        statuses[tool_id] = "connected"
+
+                oauth_server_urls: dict = {}
+                statuses: dict = {}
+                for tool in mcp_tools:
+                    tool_id = str(tool["id"])
+                    config = tool.get("config") or {}
+                    auth_type = config.get("auth_type", "none")
+                    if auth_type == "oauth":
+                        server_url = config.get("server_url", "")
+                        if server_url:
+                            parsed = urlparse(server_url)
+                            base_url = f"{parsed.scheme}://{parsed.netloc}"
+                            oauth_server_urls[tool_id] = base_url
+                        else:
+                            statuses[tool_id] = "needs_auth"
                     else:
-                        statuses[tool_id] = "needs_auth"
+                        statuses[tool_id] = "configured"
+
+                if oauth_server_urls:
+                    # Look up a session per distinct base URL. MCP sessions
+                    # are stored with ``provider = "mcp:<server_url>"``
+                    # and the URL in ``server_url``; reuse the repo's
+                    # per-URL accessor rather than an ad-hoc $in query.
+                    url_has_tokens: dict = {}
+                    for base_url in set(oauth_server_urls.values()):
+                        session = sessions_repo.get_by_user_and_server_url(
+                            user, base_url,
+                        )
+                        tokens = (
+                            (session or {}).get("session_data", {}) or {}
+                        ).get("tokens", {}) or {}
+                        # MCP code also stashes tokens into token_info on
+                        # the row; consider either present as "connected".
+                        token_info = (session or {}).get("token_info") or {}
+                        url_has_tokens[base_url] = bool(
+                            tokens.get("access_token")
+                            or token_info.get("access_token")
+                        )
+
+                    for tool_id, base_url in oauth_server_urls.items():
+                        if url_has_tokens.get(base_url):
+                            statuses[tool_id] = "connected"
+                        else:
+                            statuses[tool_id] = "needs_auth"
 
             return make_response(jsonify({"success": True, "statuses": statuses}), 200)
         except Exception as e:

application/api/user/tools/routes.py

@@ -1,23 +1,59 @@
 """Tool management routes."""
 
-from bson.objectid import ObjectId
 from flask import current_app, jsonify, make_response, request
 from flask_restx import fields, Namespace, Resource
 
 from application.agents.tools.spec_parser import parse_spec
 from application.agents.tools.tool_manager import ToolManager
 from application.api import api
-from application.api.user.base import user_tools_collection
 from application.core.url_validation import SSRFError, validate_url
-from application.storage.db.dual_write import dual_write
-from application.storage.db.repositories.user_tools import UserToolsRepository
 from application.security.encryption import decrypt_credentials, encrypt_credentials
+from application.storage.db.repositories.notes import NotesRepository
+from application.storage.db.repositories.todos import TodosRepository
+from application.storage.db.repositories.user_tools import UserToolsRepository
+from application.storage.db.session import db_readonly, db_session
 from application.utils import check_required_fields, validate_function_name
 
 tool_config = {}
 tool_manager = ToolManager(config=tool_config)
 
 
+# ---------------------------------------------------------------------------
+# Shape translation helpers
+# ---------------------------------------------------------------------------
+# The frontend speaks camelCase (``displayName`` / ``customName`` /
+# ``configRequirements``). The PG ``user_tools`` table stores snake_case
+# (``display_name`` / ``custom_name`` / ``config_requirements``). Keep the
+# translation localized to this module so repositories stay pure.
+
+_CAMEL_TO_SNAKE = {
+    "displayName": "display_name",
+    "customName": "custom_name",
+    "configRequirements": "config_requirements",
+}
+_SNAKE_TO_CAMEL = {v: k for k, v in _CAMEL_TO_SNAKE.items()}
+
+
+def _row_to_api(row: dict) -> dict:
+    """Rename DB-native snake_case keys to the camelCase shape the frontend expects."""
+    out = dict(row)
+    for snake, camel in _SNAKE_TO_CAMEL.items():
+        if snake in out:
+            out[camel] = out.pop(snake)
+    # ``user_id`` is exposed as ``user`` in the legacy API shape.
+    if "user_id" in out:
+        out["user"] = out.pop("user_id")
+    return out
+
+
+def _api_to_update_fields(data: dict) -> dict:
+    """Rename incoming camelCase update keys to the repo's snake_case columns."""
+    fields_out: dict = {}
+    for key, value in data.items():
+        fields_out[_CAMEL_TO_SNAKE.get(key, key)] = value
+    return fields_out
+
+
 def _encrypt_secret_fields(config, config_requirements, user_id):
     secret_keys = [
         key for key, spec in config_requirements.items()
@@ -170,12 +206,11 @@ class GetTools(Resource):
             if not decoded_token:
                 return make_response(jsonify({"success": False}), 401)
             user = decoded_token.get("sub")
-            tools = user_tools_collection.find({"user": user})
+            with db_readonly() as conn:
+                rows = UserToolsRepository(conn).list_for_user(user)
             user_tools = []
-            for tool in tools:
-                tool_copy = {**tool}
-                tool_copy["id"] = str(tool["_id"])
-                tool_copy.pop("_id", None)
+            for row in rows:
+                tool_copy = _row_to_api(row)
 
                 config_req = tool_copy.get("configRequirements", {})
                 if not config_req:
@@ -283,26 +318,19 @@ class CreateTool(Resource):
             storage_config = _encrypt_secret_fields(
                 data["config"], config_requirements, user
             )
-            new_tool = {
-                "user": user,
-                "name": data["name"],
-                "displayName": data["displayName"],
-                "description": data["description"],
-                "customName": data.get("customName", ""),
-                "actions": transformed_actions,
-                "config": storage_config,
-                "configRequirements": config_requirements,
-                "status": data["status"],
-            }
-            resp = user_tools_collection.insert_one(new_tool)
-            new_id = str(resp.inserted_id)
-            dual_write(
-                UserToolsRepository,
-                lambda repo, u=user, t=new_tool: repo.create(
-                    u, t["name"], config=t.get("config"),
-                    custom_name=t.get("customName"), display_name=t.get("displayName"),
-                ),
-            )
+            with db_session() as conn:
+                created = UserToolsRepository(conn).create(
+                    user,
+                    data["name"],
+                    config=storage_config,
+                    custom_name=data.get("customName", ""),
+                    display_name=data["displayName"],
+                    description=data["description"],
+                    config_requirements=config_requirements,
+                    actions=transformed_actions,
+                    status=bool(data.get("status", True)),
+                )
+            new_id = str(created["id"])
         except Exception as err:
             current_app.logger.error(f"Error creating tool: {err}", exc_info=True)
             return make_response(jsonify({"success": False}), 400)
@@ -340,17 +368,10 @@ class UpdateTool(Resource):
         if missing_fields:
             return missing_fields
         try:
-            update_data = {}
-            if "name" in data:
-                update_data["name"] = data["name"]
-            if "displayName" in data:
-                update_data["displayName"] = data["displayName"]
-            if "customName" in data:
-                update_data["customName"] = data["customName"]
-            if "description" in data:
-                update_data["description"] = data["description"]
-            if "actions" in data:
-                update_data["actions"] = data["actions"]
+            update_data: dict = {}
+            for key in ("name", "displayName", "customName", "description", "actions"):
+                if key in data:
+                    update_data[key] = data[key]
             if "config" in data:
                 if "actions" in data["config"]:
                     for action_name in list(data["config"]["actions"].keys()):
@@ -365,46 +386,61 @@ class UpdateTool(Resource):
                                 ),
                                 400,
                             )
-                tool_doc = user_tools_collection.find_one(
-                    {"_id": ObjectId(data["id"]), "user": user}
-                )
-                if not tool_doc:
-                    return make_response(
-                        jsonify({"success": False, "message": "Tool not found"}),
-                        404,
-                    )
-                tool_name = tool_doc.get("name", data.get("name"))
-                tool_instance = tool_manager.tools.get(tool_name)
-                config_requirements = (
-                    tool_instance.get_config_requirements() if tool_instance else {}
-                )
-                existing_config = tool_doc.get("config", {})
-                has_existing_secrets = "encrypted_credentials" in existing_config
-
-                if config_requirements:
-                    validation_errors = _validate_config(
-                        data["config"], config_requirements,
-                        has_existing_secrets=has_existing_secrets,
-                    )
-                    if validation_errors:
+                with db_session() as conn:
+                    repo = UserToolsRepository(conn)
+                    tool_doc = repo.get_any(data["id"], user)
+                    if not tool_doc:
                         return make_response(
-                            jsonify({
-                                "success": False,
-                                "message": "Validation failed",
-                                "errors": validation_errors,
-                            }),
-                            400,
+                            jsonify({"success": False, "message": "Tool not found"}),
+                            404,
                         )
+                    tool_name = tool_doc.get("name", data.get("name"))
+                    tool_instance = tool_manager.tools.get(tool_name)
+                    config_requirements = (
+                        tool_instance.get_config_requirements()
+                        if tool_instance
+                        else {}
+                    )
+                    existing_config = tool_doc.get("config", {}) or {}
+                    has_existing_secrets = "encrypted_credentials" in existing_config
 
-                update_data["config"] = _merge_secrets_on_update(
-                    data["config"], existing_config, config_requirements, user
-                )
-            if "status" in data:
-                update_data["status"] = data["status"]
-            user_tools_collection.update_one(
-                {"_id": ObjectId(data["id"]), "user": user},
-                {"$set": update_data},
-            )
+                    if config_requirements:
+                        validation_errors = _validate_config(
+                            data["config"], config_requirements,
+                            has_existing_secrets=has_existing_secrets,
+                        )
+                        if validation_errors:
+                            return make_response(
+                                jsonify({
+                                    "success": False,
+                                    "message": "Validation failed",
+                                    "errors": validation_errors,
+                                }),
+                                400,
+                            )
+
+                    update_data["config"] = _merge_secrets_on_update(
+                        data["config"], existing_config, config_requirements, user
+                    )
+                    if "status" in data:
+                        update_data["status"] = bool(data["status"])
+                    repo.update(
+                        str(tool_doc["id"]), user, _api_to_update_fields(update_data),
+                    )
+            else:
+                if "status" in data:
+                    update_data["status"] = bool(data["status"])
+                with db_session() as conn:
+                    repo = UserToolsRepository(conn)
+                    tool_doc = repo.get_any(data["id"], user)
+                    if not tool_doc:
+                        return make_response(
+                            jsonify({"success": False, "message": "Tool not found"}),
+                            404,
+                        )
+                    repo.update(
+                        str(tool_doc["id"]), user, _api_to_update_fields(update_data),
+                    )
         except Exception as err:
             current_app.logger.error(f"Error updating tool: {err}", exc_info=True)
             return make_response(jsonify({"success": False}), 400)
@@ -436,53 +472,50 @@ class UpdateToolConfig(Resource):
         if missing_fields:
             return missing_fields
         try:
-            tool_doc = user_tools_collection.find_one(
-                {"_id": ObjectId(data["id"]), "user": user}
-            )
-            if not tool_doc:
-                return make_response(jsonify({"success": False}), 404)
+            with db_session() as conn:
+                repo = UserToolsRepository(conn)
+                tool_doc = repo.get_any(data["id"], user)
+                if not tool_doc:
+                    return make_response(jsonify({"success": False}), 404)
 
-            tool_name = tool_doc.get("name")
-            if tool_name == "mcp_tool":
-                server_url = (data["config"].get("server_url") or "").strip()
-                if server_url:
-                    try:
-                        validate_url(server_url)
-                    except SSRFError:
+                tool_name = tool_doc.get("name")
+                if tool_name == "mcp_tool":
+                    server_url = (data["config"].get("server_url") or "").strip()
+                    if server_url:
+                        try:
+                            validate_url(server_url)
+                        except SSRFError:
+                            return make_response(
+                                jsonify({"success": False, "message": "Invalid server URL"}),
+                                400,
+                            )
+                tool_instance = tool_manager.tools.get(tool_name)
+                config_requirements = (
+                    tool_instance.get_config_requirements() if tool_instance else {}
+                )
+                existing_config = tool_doc.get("config", {}) or {}
+                has_existing_secrets = "encrypted_credentials" in existing_config
+
+                if config_requirements:
+                    validation_errors = _validate_config(
+                        data["config"], config_requirements,
+                        has_existing_secrets=has_existing_secrets,
+                    )
+                    if validation_errors:
                         return make_response(
-                            jsonify({"success": False, "message": "Invalid server URL"}),
+                            jsonify({
+                                "success": False,
+                                "message": "Validation failed",
+                                "errors": validation_errors,
+                            }),
                             400,
                         )
-            tool_instance = tool_manager.tools.get(tool_name)
-            config_requirements = (
-                tool_instance.get_config_requirements() if tool_instance else {}
-            )
-            existing_config = tool_doc.get("config", {})
-            has_existing_secrets = "encrypted_credentials" in existing_config
 
-            if config_requirements:
-                validation_errors = _validate_config(
-                    data["config"], config_requirements,
-                    has_existing_secrets=has_existing_secrets,
+                final_config = _merge_secrets_on_update(
+                    data["config"], existing_config, config_requirements, user
                 )
-                if validation_errors:
-                    return make_response(
-                        jsonify({
-                            "success": False,
-                            "message": "Validation failed",
-                            "errors": validation_errors,
-                        }),
-                        400,
-                    )
 
-            final_config = _merge_secrets_on_update(
-                data["config"], existing_config, config_requirements, user
-            )
-
-            user_tools_collection.update_one(
-                {"_id": ObjectId(data["id"]), "user": user},
-                {"$set": {"config": final_config}},
-            )
+                repo.update(str(tool_doc["id"]), user, {"config": final_config})
         except Exception as err:
             current_app.logger.error(
                 f"Error updating tool config: {err}", exc_info=True
@@ -518,10 +551,17 @@ class UpdateToolActions(Resource):
         if missing_fields:
             return missing_fields
         try:
-            user_tools_collection.update_one(
-                {"_id": ObjectId(data["id"]), "user": user},
-                {"$set": {"actions": data["actions"]}},
-            )
+            with db_session() as conn:
+                repo = UserToolsRepository(conn)
+                tool_doc = repo.get_any(data["id"], user)
+                if not tool_doc:
+                    return make_response(
+                        jsonify({"success": False, "message": "Tool not found"}),
+                        404,
+                    )
+                repo.update(
+                    str(tool_doc["id"]), user, {"actions": data["actions"]},
+                )
         except Exception as err:
             current_app.logger.error(
                 f"Error updating tool actions: {err}", exc_info=True
@@ -555,10 +595,17 @@ class UpdateToolStatus(Resource):
         if missing_fields:
             return missing_fields
         try:
-            user_tools_collection.update_one(
-                {"_id": ObjectId(data["id"]), "user": user},
-                {"$set": {"status": data["status"]}},
-            )
+            with db_session() as conn:
+                repo = UserToolsRepository(conn)
+                tool_doc = repo.get_any(data["id"], user)
+                if not tool_doc:
+                    return make_response(
+                        jsonify({"success": False, "message": "Tool not found"}),
+                        404,
+                    )
+                repo.update(
+                    str(tool_doc["id"]), user, {"status": bool(data["status"])},
+                )
         except Exception as err:
             current_app.logger.error(
                 f"Error updating tool status: {err}", exc_info=True
@@ -587,17 +634,14 @@ class DeleteTool(Resource):
         if missing_fields:
             return missing_fields
         try:
-            result = user_tools_collection.delete_one(
-                {"_id": ObjectId(data["id"]), "user": user}
-            )
-            dual_write(
-                UserToolsRepository,
-                lambda repo, tid=data["id"], u=user: repo.delete(tid, u),
-            )
-            if result.deleted_count == 0:
-                return make_response(
-                    jsonify({"success": False, "message": "Tool not found"}), 404
-                )
+            with db_session() as conn:
+                repo = UserToolsRepository(conn)
+                tool_doc = repo.get_any(data["id"], user)
+                if not tool_doc:
+                    return make_response(
+                        jsonify({"success": False, "message": "Tool not found"}), 404
+                    )
+                repo.delete(str(tool_doc["id"]), user)
         except Exception as err:
             current_app.logger.error(f"Error deleting tool: {err}", exc_info=True)
             return make_response(jsonify({"success": False}), 400)
@@ -666,70 +710,88 @@ class GetArtifact(Resource):
         user_id = decoded_token.get("sub")
 
         try:
-            obj_id = ObjectId(artifact_id)
-        except Exception:
-            return make_response(
-                jsonify({"success": False, "message": "Invalid artifact ID"}), 400
+            with db_readonly() as conn:
+                notes_repo = NotesRepository(conn)
+                todos_repo = TodosRepository(conn)
+
+                # Artifact IDs may be PG UUIDs (post-cutover) or legacy
+                # Mongo ObjectIds embedded in older conversation history.
+                # Both repos' ``get_any`` handles the id-shape branching
+                # internally so a non-UUID input never reaches
+                # ``CAST(:id AS uuid)`` (which would poison the readonly
+                # transaction and break the fallback below).
+                note_doc = notes_repo.get_any(artifact_id, user_id)
+
+                if note_doc:
+                    content = note_doc.get("note", "") or note_doc.get("content", "")
+                    line_count = len(content.split("\n")) if content else 0
+                    updated = note_doc.get("updated_at")
+                    artifact = {
+                        "artifact_type": "note",
+                        "data": {
+                            "content": content,
+                            "line_count": line_count,
+                            "updated_at": (
+                                updated.isoformat()
+                                if hasattr(updated, "isoformat")
+                                else updated
+                            ),
+                        },
+                    }
+                    return make_response(
+                        jsonify({"success": True, "artifact": artifact}), 200
+                    )
+
+                todo_doc = todos_repo.get_any(artifact_id, user_id)
+                if todo_doc:
+                    tool_id = todo_doc.get("tool_id")
+                    all_todos = todos_repo.list_for_tool(user_id, tool_id) if tool_id else []
+                    items = []
+                    open_count = 0
+                    completed_count = 0
+                    for t in all_todos:
+                        # PG ``todos`` stores a ``completed BOOLEAN`` column;
+                        # the legacy Mongo shape used a ``status`` string.
+                        # Keep the response shape stable by translating here.
+                        status = "completed" if t.get("completed") else "open"
+                        if status == "open":
+                            open_count += 1
+                        else:
+                            completed_count += 1
+                        created = t.get("created_at")
+                        updated = t.get("updated_at")
+                        items.append({
+                            "todo_id": t.get("todo_id"),
+                            "title": t.get("title", ""),
+                            "status": status,
+                            "created_at": (
+                                created.isoformat()
+                                if hasattr(created, "isoformat")
+                                else created
+                            ),
+                            "updated_at": (
+                                updated.isoformat()
+                                if hasattr(updated, "isoformat")
+                                else updated
+                            ),
+                        })
+                    artifact = {
+                        "artifact_type": "todo_list",
+                        "data": {
+                            "items": items,
+                            "total_count": len(items),
+                            "open_count": open_count,
+                            "completed_count": completed_count,
+                        },
+                    }
+                    return make_response(
+                        jsonify({"success": True, "artifact": artifact}), 200
+                    )
+        except Exception as err:
+            current_app.logger.error(
+                f"Error retrieving artifact: {err}", exc_info=True
             )
-
-        from application.core.mongo_db import MongoDB
-        from application.core.settings import settings
-
-        db = MongoDB.get_client()[settings.MONGO_DB_NAME]
-
-        note_doc = db["notes"].find_one({"_id": obj_id, "user_id": user_id})
-        if note_doc:
-            content = note_doc.get("note", "")
-            line_count = len(content.split("\n")) if content else 0
-            artifact = {
-                "artifact_type": "note",
-                "data": {
-                    "content": content,
-                    "line_count": line_count,
-                    "updated_at": (
-                        note_doc["updated_at"].isoformat()
-                        if note_doc.get("updated_at")
-                        else None
-                    ),
-                },
-            }
-            return make_response(jsonify({"success": True, "artifact": artifact}), 200)
-
-        todo_doc = db["todos"].find_one({"_id": obj_id, "user_id": user_id})
-        if todo_doc:
-            tool_id = todo_doc.get("tool_id")
-            query = {"user_id": user_id, "tool_id": tool_id}
-            all_todos = list(db["todos"].find(query))
-            items = []
-            open_count = 0
-            completed_count = 0
-            for t in all_todos:
-                status = t.get("status", "open")
-                if status == "open":
-                    open_count += 1
-                elif status == "completed":
-                    completed_count += 1
-                items.append({
-                    "todo_id": t.get("todo_id"),
-                    "title": t.get("title", ""),
-                    "status": status,
-                    "created_at": (
-                        t["created_at"].isoformat() if t.get("created_at") else None
-                    ),
-                    "updated_at": (
-                        t["updated_at"].isoformat() if t.get("updated_at") else None
-                    ),
-                })
-            artifact = {
-                "artifact_type": "todo_list",
-                "data": {
-                    "items": items,
-                    "total_count": len(items),
-                    "open_count": open_count,
-                    "completed_count": completed_count,
-                },
-            }
-            return make_response(jsonify({"success": True, "artifact": artifact}), 200)
+            return make_response(jsonify({"success": False}), 400)
 
         return make_response(
             jsonify({"success": False, "message": "Artifact not found"}), 404

application/api/user/utils.py

@@ -1,290 +1,61 @@
-"""Centralized utilities for API routes."""
+"""Centralized utilities for API routes.
+
+Post-Mongo-cutover slim: the old Mongo-shaped helpers (``validate_object_id``,
+``check_resource_ownership``, ``paginated_response``, ``serialize_object_id``,
+``safe_db_operation``, ``validate_enum``, ``extract_sort_params``) have been
+removed — they carried ``bson`` / ``pymongo`` imports and had zero callers.
+"""
 
 from functools import wraps
-from typing import Any, Callable, Dict, List, Optional, Tuple
+from typing import Callable, Optional
 
-from bson.errors import InvalidId
-from bson.objectid import ObjectId
 from flask import (
     Response,
-    current_app,
-    has_app_context,
     jsonify,
     make_response,
     request,
 )
-from pymongo.collection import Collection
 
 
 def get_user_id() -> Optional[str]:
-    """
-    Extract user ID from decoded JWT token.
-
-    Returns:
-        User ID string or None if not authenticated
-    """
+    """Extract user ID from decoded JWT token, or None if unauthenticated."""
     decoded_token = getattr(request, "decoded_token", None)
     return decoded_token.get("sub") if decoded_token else None
 
 
 def require_auth(func: Callable) -> Callable:
-    """
-    Decorator to require authentication for route handlers.
-
-    Usage:
-        @require_auth
-        def get(self):
-            user_id = get_user_id()
-            ...
-    """
+    """Decorator to require authentication. Returns 401 when absent."""
 
     @wraps(func)
     def wrapper(*args, **kwargs):
         user_id = get_user_id()
         if not user_id:
-            return error_response("Unauthorized", 401)
+            return make_response(jsonify({"success": False, "error": "Unauthorized"}), 401)
         return func(*args, **kwargs)
 
     return wrapper
 
 
 def success_response(
-    data: Optional[Dict[str, Any]] = None, status: int = 200
+    data=None, message: Optional[str] = None, status: int = 200
 ) -> Response:
-    """
-    Create a standardized success response.
-
-    Args:
-        data: Optional data dictionary to include in response
-        status: HTTP status code (default: 200)
-
-    Returns:
-        Flask Response object
-
-    Example:
-        return success_response({"users": [...], "total": 10})
-    """
-    response = {"success": True}
-    if data:
-        response.update(data)
-    return make_response(jsonify(response), status)
+    """Shape a successful JSON response."""
+    body = {"success": True}
+    if data is not None:
+        body["data"] = data
+    if message is not None:
+        body["message"] = message
+    return make_response(jsonify(body), status)
 
 
 def error_response(message: str, status: int = 400, **kwargs) -> Response:
-    """
-    Create a standardized error response.
-
-    Args:
-        message: Error message string
-        status: HTTP status code (default: 400)
-        **kwargs: Additional fields to include in response
-
-    Returns:
-        Flask Response object
-
-    Example:
-        return error_response("Resource not found", 404)
-        return error_response("Invalid input", 400, errors=["field1", "field2"])
-    """
-    response = {"success": False, "message": message}
-    response.update(kwargs)
-    return make_response(jsonify(response), status)
+    """Shape an error JSON response; any kwargs are merged into the body."""
+    body = {"success": False, "error": message, **kwargs}
+    return make_response(jsonify(body), status)
 
 
-def validate_object_id(
-    id_string: str, resource_name: str = "Resource"
-) -> Tuple[Optional[ObjectId], Optional[Response]]:
-    """
-    Validate and convert string to ObjectId.
-
-    Args:
-        id_string: String to convert
-        resource_name: Name of resource for error message
-
-    Returns:
-        Tuple of (ObjectId or None, error_response or None)
-
-    Example:
-        obj_id, error = validate_object_id(workflow_id, "Workflow")
-        if error:
-            return error
-    """
-    try:
-        return ObjectId(id_string), None
-    except (InvalidId, TypeError):
-        return None, error_response(f"Invalid {resource_name} ID format")
-
-
-def validate_pagination(
-    default_limit: int = 20, max_limit: int = 100
-) -> Tuple[int, int, Optional[Response]]:
-    """
-    Extract and validate pagination parameters from request.
-
-    Args:
-        default_limit: Default items per page
-        max_limit: Maximum allowed items per page
-
-    Returns:
-        Tuple of (limit, skip, error_response or None)
-
-    Example:
-        limit, skip, error = validate_pagination()
-        if error:
-            return error
-    """
-    try:
-        limit = min(int(request.args.get("limit", default_limit)), max_limit)
-        skip = int(request.args.get("skip", 0))
-        if limit < 1 or skip < 0:
-            return 0, 0, error_response("Invalid pagination parameters")
-        return limit, skip, None
-    except ValueError:
-        return 0, 0, error_response("Invalid pagination parameters")
-
-
-def check_resource_ownership(
-    collection: Collection,
-    resource_id: ObjectId,
-    user_id: str,
-    resource_name: str = "Resource",
-) -> Tuple[Optional[Dict], Optional[Response]]:
-    """
-    Check if resource exists and belongs to user.
-
-    Args:
-        collection: MongoDB collection
-        resource_id: Resource ObjectId
-        user_id: User ID string
-        resource_name: Name of resource for error messages
-
-    Returns:
-        Tuple of (resource_dict or None, error_response or None)
-
-    Example:
-        workflow, error = check_resource_ownership(
-            workflows_collection,
-            workflow_id,
-            user_id,
-            "Workflow"
-        )
-        if error:
-            return error
-    """
-    resource = collection.find_one({"_id": resource_id, "user": user_id})
-    if not resource:
-        return None, error_response(f"{resource_name} not found", 404)
-    return resource, None
-
-
-def serialize_object_id(
-    obj: Dict[str, Any], id_field: str = "_id", new_field: str = "id"
-) -> Dict[str, Any]:
-    """
-    Convert ObjectId to string in a dictionary.
-
-    Args:
-        obj: Dictionary containing ObjectId
-        id_field: Field name containing ObjectId
-        new_field: New field name for string ID
-
-    Returns:
-        Modified dictionary
-
-    Example:
-        user = serialize_object_id(user_doc)
-        # user["id"] = "507f1f77bcf86cd799439011"
-    """
-    if id_field in obj:
-        obj[new_field] = str(obj[id_field])
-        if id_field != new_field:
-            obj.pop(id_field, None)
-    return obj
-
-
-def serialize_list(items: List[Dict], serializer: Callable[[Dict], Dict]) -> List[Dict]:
-    """
-    Apply serializer function to list of items.
-
-    Args:
-        items: List of dictionaries
-        serializer: Function to apply to each item
-
-    Returns:
-        List of serialized items
-
-    Example:
-        workflows = serialize_list(workflow_docs, serialize_workflow)
-    """
-    return [serializer(item) for item in items]
-
-
-def paginated_response(
-    collection: Collection,
-    query: Dict[str, Any],
-    serializer: Callable[[Dict], Dict],
-    limit: int,
-    skip: int,
-    sort_field: str = "created_at",
-    sort_order: int = -1,
-    response_key: str = "items",
-) -> Response:
-    """
-    Create paginated response for collection query.
-
-    Args:
-        collection: MongoDB collection
-        query: Query dictionary
-        serializer: Function to serialize each item
-        limit: Items per page
-        skip: Number of items to skip
-        sort_field: Field to sort by
-        sort_order: Sort order (1=asc, -1=desc)
-        response_key: Key name for items in response
-
-    Returns:
-        Flask Response with paginated data
-
-    Example:
-        return paginated_response(
-            workflows_collection,
-            {"user": user_id},
-            serialize_workflow,
-            limit, skip,
-            response_key="workflows"
-        )
-    """
-    items = list(
-        collection.find(query).sort(sort_field, sort_order).skip(skip).limit(limit)
-    )
-    total = collection.count_documents(query)
-
-    return success_response(
-        {
-            response_key: serialize_list(items, serializer),
-            "total": total,
-            "limit": limit,
-            "skip": skip,
-        }
-    )
-
-
-def require_fields(required: List[str]) -> Callable:
-    """
-    Decorator to validate required fields in request JSON.
-
-    Args:
-        required: List of required field names
-
-    Returns:
-        Decorator function
-
-    Example:
-        @require_fields(["name", "description"])
-        def post(self):
-            data = request.get_json()
-            ...
-    """
+def require_fields(required: list) -> Callable:
+    """Decorator: return 400 if any listed field is missing/falsy in the JSON body."""
 
     def decorator(func: Callable) -> Callable:
         @wraps(func)
@@ -294,94 +65,11 @@ def require_fields(required: List[str]) -> Callable:
                 return error_response("Request body required")
             missing = [field for field in required if not data.get(field)]
             if missing:
-                return error_response(f"Missing required fields: {', '.join(missing)}")
+                return error_response(
+                    f"Missing required fields: {', '.join(missing)}"
+                )
             return func(*args, **kwargs)
 
         return wrapper
 
     return decorator
-
-
-def safe_db_operation(
-    operation: Callable, error_message: str = "Database operation failed"
-) -> Tuple[Any, Optional[Response]]:
-    """
-    Safely execute database operation with error handling.
-
-    Args:
-        operation: Function to execute
-        error_message: Error message if operation fails
-
-    Returns:
-        Tuple of (result or None, error_response or None)
-
-    Example:
-        result, error = safe_db_operation(
-            lambda: collection.insert_one(doc),
-            "Failed to create resource"
-        )
-        if error:
-            return error
-    """
-    try:
-        result = operation()
-        return result, None
-    except Exception as err:
-        if has_app_context():
-            current_app.logger.error(f"{error_message}: {err}", exc_info=True)
-        return None, error_response(error_message)
-
-
-def validate_enum(
-    value: Any, allowed: List[Any], field_name: str
-) -> Optional[Response]:
-    """
-    Validate that value is in allowed list.
-
-    Args:
-        value: Value to validate
-        allowed: List of allowed values
-        field_name: Field name for error message
-
-    Returns:
-        error_response if invalid, None if valid
-
-    Example:
-        error = validate_enum(status, ["draft", "published"], "status")
-        if error:
-            return error
-    """
-    if value not in allowed:
-        allowed_str = ", ".join(f"'{v}'" for v in allowed)
-        return error_response(f"Invalid {field_name}. Must be one of: {allowed_str}")
-    return None
-
-
-def extract_sort_params(
-    default_field: str = "created_at",
-    default_order: str = "desc",
-    allowed_fields: Optional[List[str]] = None,
-) -> Tuple[str, int]:
-    """
-    Extract and validate sort parameters from request.
-
-    Args:
-        default_field: Default sort field
-        default_order: Default sort order ("asc" or "desc")
-        allowed_fields: List of allowed sort fields (None = no validation)
-
-    Returns:
-        Tuple of (sort_field, sort_order)
-
-    Example:
-        sort_field, sort_order = extract_sort_params(
-            allowed_fields=["name", "date", "status"]
-        )
-    """
-    sort_field = request.args.get("sort", default_field)
-    sort_order_str = request.args.get("order", default_order).lower()
-
-    if allowed_fields and sort_field not in allowed_fields:
-        sort_field = default_field
-    sort_order = -1 if sort_order_str == "desc" else 1
-    return sort_field, sort_order

application/api/user/workflows/routes.py

@@ -1,34 +1,26 @@
 """Workflow management routes."""
 
-from datetime import datetime, timezone
 from typing import Any, Dict, List, Optional, Set
 
 from flask import current_app, request
 from flask_restx import Namespace, Resource
 
-from application.api.user.base import (
-    workflow_edges_collection,
-    workflow_nodes_collection,
-    workflows_collection,
-)
-from application.storage.db.dual_write import dual_write
+from application.storage.db.base_repository import looks_like_uuid
 from application.storage.db.repositories.workflow_edges import WorkflowEdgesRepository
 from application.storage.db.repositories.workflow_nodes import WorkflowNodesRepository
 from application.storage.db.repositories.workflows import WorkflowsRepository
+from application.storage.db.session import db_readonly, db_session
 from application.core.json_schema_utils import (
     JsonSchemaValidationError,
     normalize_json_schema_payload,
 )
 from application.core.model_utils import get_model_capabilities
 from application.api.user.utils import (
-    check_resource_ownership,
     error_response,
     get_user_id,
     require_auth,
     require_fields,
-    safe_db_operation,
     success_response,
-    validate_object_id,
 )
 
 workflows_ns = Namespace("workflows", path="/api")
@@ -39,109 +31,15 @@ def _workflow_error_response(message: str, err: Exception):
     return error_response(message)
 
 
-# ---------------------------------------------------------------------------
-# Postgres dual-write helpers
-#
-# Workflows are unusual relative to other Phase 3 tables: a single user
-# action (create / update) writes to three collections in concert
-# (workflows + workflow_nodes + workflow_edges) and the edges reference
-# nodes by user-provided string ids. The Postgres mirror needs to:
-#
-# 1. Run all three writes inside one PG transaction (so the just-created
-#    nodes are visible when we resolve their UUIDs for the edge insert).
-# 2. Translate edge source_id/target_id strings → workflow_nodes.id UUIDs
-#    after the bulk_create returns them.
-#
-# Each helper opens exactly one ``dual_write`` call (one PG txn) and uses
-# the connection from whichever repo it was instantiated with to spin up
-# any sibling repos it needs.
-# ---------------------------------------------------------------------------
-
-
-def _dual_write_workflow_create(
-    mongo_workflow_id: str,
-    user_id: str,
-    name: str,
-    description: str,
-    nodes_data: List[Dict],
-    edges_data: List[Dict],
-    graph_version: int = 1,
-) -> None:
-    """Mirror a Mongo workflow create into Postgres."""
-
-    def _do(repo: WorkflowsRepository) -> None:
-        conn = repo._conn
-        wf = repo.create(
-            user_id,
-            name,
-            description=description,
-            legacy_mongo_id=mongo_workflow_id,
-        )
-        _write_graph(conn, wf["id"], graph_version, nodes_data, edges_data)
-
-    dual_write(WorkflowsRepository, _do)
-
-
-def _dual_write_workflow_update(
-    mongo_workflow_id: str,
-    user_id: str,
-    name: str,
-    description: str,
-    nodes_data: List[Dict],
-    edges_data: List[Dict],
-    next_graph_version: int,
-) -> None:
-    """Mirror a Mongo workflow update into Postgres.
-
-    Mirrors the Mongo route: insert the new graph_version's nodes/edges,
-    bump the workflow's name/description/current_graph_version, then drop
-    every other graph_version's nodes/edges.
-    """
-
-    def _do(repo: WorkflowsRepository) -> None:
-        conn = repo._conn
-        wf = _resolve_pg_workflow(conn, mongo_workflow_id)
-        if wf is None:
-            return
-        _write_graph(conn, wf["id"], next_graph_version, nodes_data, edges_data)
-        repo.update(wf["id"], user_id, {
-            "name": name,
-            "description": description,
-            "current_graph_version": next_graph_version,
-        })
-        WorkflowNodesRepository(conn).delete_other_versions(
-            wf["id"], next_graph_version,
-        )
-        WorkflowEdgesRepository(conn).delete_other_versions(
-            wf["id"], next_graph_version,
-        )
-
-    dual_write(WorkflowsRepository, _do)
-
-
-def _dual_write_workflow_delete(mongo_workflow_id: str, user_id: str) -> None:
-    """Mirror a Mongo workflow delete into Postgres.
-
-    The CASCADE on workflows.id → workflow_nodes/workflow_edges takes
-    care of the children automatically.
-    """
-
-    def _do(repo: WorkflowsRepository) -> None:
-        wf = _resolve_pg_workflow(repo._conn, mongo_workflow_id)
-        if wf is not None:
-            repo.delete(wf["id"], user_id)
-
-    dual_write(WorkflowsRepository, _do)
-
-
-def _resolve_pg_workflow(conn, mongo_workflow_id: str) -> Optional[Dict]:
-    """Look up a Postgres workflow by its Mongo ObjectId string."""
-    from sqlalchemy import text as _text
-    row = conn.execute(
-        _text("SELECT id FROM workflows WHERE legacy_mongo_id = :legacy_id"),
-        {"legacy_id": mongo_workflow_id},
-    ).fetchone()
-    return {"id": str(row[0])} if row else None
+def _resolve_workflow(repo: WorkflowsRepository, workflow_id: str, user_id: str):
+    """Resolve a workflow by UUID or legacy Mongo id, scoped to user."""
+    if not workflow_id:
+        return None
+    if looks_like_uuid(workflow_id):
+        row = repo.get(workflow_id, user_id)
+        if row is not None:
+            return row
+    return repo.get_by_legacy_id(workflow_id, user_id)
 
 
 def _write_graph(
@@ -150,14 +48,13 @@ def _write_graph(
     graph_version: int,
     nodes_data: List[Dict],
     edges_data: List[Dict],
-) -> None:
-    """Bulk-create nodes + edges for one graph version inside one txn.
+) -> List[Dict]:
+    """Bulk-create nodes + edges for one graph version. Uses ON CONFLICT upsert.
 
-    Edges arrive with source/target as user-provided node-id strings
-    (the same shape the Mongo route stores). We bulk-insert nodes first,
-    capture their ``node_id → UUID`` map from the returned rows, then
-    translate edge source/target strings to those UUIDs before the edge
-    bulk insert. Edges referencing missing nodes are dropped (logged).
+    Edges arrive with source/target as user-provided node-id strings. We
+    insert nodes first, capture their ``node_id → UUID`` map, then
+    translate edges before insertion. Edges referencing missing nodes are
+    dropped with a warning.
     """
     nodes_repo = WorkflowNodesRepository(conn)
     edges_repo = WorkflowEdgesRepository(conn)
@@ -173,13 +70,13 @@ def _write_graph(
                     "description": n.get("description", ""),
                     "position": n.get("position", {"x": 0, "y": 0}),
                     "config": n.get("data", {}),
-                    "legacy_mongo_id": n.get("legacy_mongo_id"),
                 }
                 for n in nodes_data
             ],
         )
         node_uuid_by_str = {n["node_id"]: n["id"] for n in created_nodes}
     else:
+        created_nodes = []
         node_uuid_by_str = {}
 
     if edges_data:
@@ -191,7 +88,7 @@ def _write_graph(
             to_uuid = node_uuid_by_str.get(tgt)
             if not from_uuid or not to_uuid:
                 current_app.logger.warning(
-                    "PG dual-write: dropping edge %s; node refs unresolved "
+                    "Workflow graph write: dropping edge %s; node refs unresolved "
                     "(source=%s, target=%s)",
                     e.get("id"), src, tgt,
                 )
@@ -204,36 +101,42 @@ def _write_graph(
                 "target_handle": e.get("targetHandle"),
             })
         if translated_edges:
-            edges_repo.bulk_create(pg_workflow_id, graph_version, translated_edges)
+            edges_repo.bulk_create(
+                pg_workflow_id, graph_version, translated_edges,
+            )
+
+    return created_nodes
 
 
 def serialize_workflow(w: Dict) -> Dict:
-    """Serialize workflow document to API response format."""
+    """Serialize workflow row to API response format."""
+    created_at = w.get("created_at")
+    updated_at = w.get("updated_at")
     return {
-        "id": str(w["_id"]),
+        "id": str(w["id"]),
         "name": w.get("name"),
         "description": w.get("description"),
-        "created_at": w["created_at"].isoformat() if w.get("created_at") else None,
-        "updated_at": w["updated_at"].isoformat() if w.get("updated_at") else None,
+        "created_at": created_at.isoformat() if hasattr(created_at, "isoformat") else created_at,
+        "updated_at": updated_at.isoformat() if hasattr(updated_at, "isoformat") else updated_at,
     }
 
 
 def serialize_node(n: Dict) -> Dict:
-    """Serialize workflow node document to API response format."""
+    """Serialize workflow node row to API response format."""
     return {
-        "id": n["id"],
-        "type": n["type"],
+        "id": n["node_id"],
+        "type": n["node_type"],
         "title": n.get("title"),
         "description": n.get("description"),
         "position": n.get("position"),
-        "data": n.get("config", {}),
+        "data": n.get("config", {}) or {},
     }
 
 
 def serialize_edge(e: Dict) -> Dict:
-    """Serialize workflow edge document to API response format."""
+    """Serialize workflow edge row to API response format."""
     return {
-        "id": e["id"],
+        "id": e["edge_id"],
         "source": e.get("source_id"),
         "target": e.get("target_id"),
         "sourceHandle": e.get("source_handle"),
@@ -242,7 +145,7 @@ def serialize_edge(e: Dict) -> Dict:
 
 
 def get_workflow_graph_version(workflow: Dict) -> int:
-    """Get current graph version with legacy fallback."""
+    """Get current graph version with fallback."""
     raw_version = workflow.get("current_graph_version", 1)
     try:
         version = int(raw_version)
@@ -251,22 +154,6 @@ def get_workflow_graph_version(workflow: Dict) -> int:
         return 1
 
 
-def fetch_graph_documents(collection, workflow_id: str, graph_version: int) -> List[Dict]:
-    """Fetch graph docs for active version, with fallback for legacy unversioned data."""
-    docs = list(
-        collection.find({"workflow_id": workflow_id, "graph_version": graph_version})
-    )
-    if docs:
-        return docs
-    if graph_version == 1:
-        return list(
-            collection.find(
-                {"workflow_id": workflow_id, "graph_version": {"$exists": False}}
-            )
-        )
-    return docs
-
-
 def validate_json_schema_payload(
     json_schema: Any,
 ) -> tuple[Optional[Dict[str, Any]], Optional[str]]:
@@ -311,8 +198,14 @@ def normalize_agent_node_json_schemas(nodes: List[Dict]) -> List[Dict]:
     return normalized_nodes
 
 
-def validate_workflow_structure(nodes: List[Dict], edges: List[Dict]) -> List[str]:
-    """Validate workflow graph structure."""
+def validate_workflow_structure(
+    nodes: List[Dict], edges: List[Dict], user_id: str | None = None
+) -> List[str]:
+    """Validate workflow graph structure.
+
+    ``user_id`` is required so per-user BYOM custom-model UUIDs resolve
+    when checking each agent node's structured-output capability.
+    """
     errors = []
 
     if not nodes:
@@ -456,7 +349,7 @@ def validate_workflow_structure(nodes: List[Dict], edges: List[Dict]) -> List[st
 
         model_id = raw_config.get("model_id")
         if has_json_schema and isinstance(model_id, str) and model_id.strip():
-            capabilities = get_model_capabilities(model_id.strip())
+            capabilities = get_model_capabilities(model_id.strip(), user_id=user_id)
             if capabilities and not capabilities.get("supports_structured_output", False):
                 errors.append(
                     f"Agent node '{agent_title}' selected model does not support structured output"
@@ -487,53 +380,6 @@ def _can_reach_end(
     return any(_can_reach_end(t, edges, node_map, end_ids, visited) for t in outgoing if t)
 
 
-def create_workflow_nodes(
-    workflow_id: str, nodes_data: List[Dict], graph_version: int
-) -> List[Dict]:
-    """Insert workflow nodes into Mongo and return rows with Mongo ids."""
-    if nodes_data:
-        mongo_nodes = [
-            {
-                "id": n["id"],
-                "workflow_id": workflow_id,
-                "graph_version": graph_version,
-                "type": n["type"],
-                "title": n.get("title", ""),
-                "description": n.get("description", ""),
-                "position": n.get("position", {"x": 0, "y": 0}),
-                "config": n.get("data", {}),
-            }
-            for n in nodes_data
-        ]
-        result = workflow_nodes_collection.insert_many(mongo_nodes)
-        return [
-            {**node, "legacy_mongo_id": str(inserted_id)}
-            for node, inserted_id in zip(nodes_data, result.inserted_ids)
-        ]
-    return []
-
-
-def create_workflow_edges(
-    workflow_id: str, edges_data: List[Dict], graph_version: int
-) -> None:
-    """Insert workflow edges into database."""
-    if edges_data:
-        workflow_edges_collection.insert_many(
-            [
-                {
-                    "id": e["id"],
-                    "workflow_id": workflow_id,
-                    "graph_version": graph_version,
-                    "source_id": e.get("source"),
-                    "target_id": e.get("target"),
-                    "source_handle": e.get("sourceHandle"),
-                    "target_handle": e.get("targetHandle"),
-                }
-                for e in edges_data
-            ]
-        )
-
-
 @workflows_ns.route("/workflows")
 class WorkflowList(Resource):
 
@@ -545,54 +391,29 @@ class WorkflowList(Resource):
         data = request.get_json()
 
         name = data.get("name", "").strip()
+        description = data.get("description", "")
         nodes_data = data.get("nodes", [])
         edges_data = data.get("edges", [])
 
-        validation_errors = validate_workflow_structure(nodes_data, edges_data)
+        validation_errors = validate_workflow_structure(
+            nodes_data, edges_data, user_id=user_id
+        )
         if validation_errors:
             return error_response(
                 "Workflow validation failed", errors=validation_errors
             )
         nodes_data = normalize_agent_node_json_schemas(nodes_data)
 
-        now = datetime.now(timezone.utc)
-        workflow_doc = {
-            "name": name,
-            "description": data.get("description", ""),
-            "user": user_id,
-            "created_at": now,
-            "updated_at": now,
-            "current_graph_version": 1,
-        }
-
-        result, error = safe_db_operation(
-            lambda: workflows_collection.insert_one(workflow_doc),
-            "Failed to create workflow",
-        )
-        if error:
-            return error
-
-        workflow_id = str(result.inserted_id)
-
         try:
-            created_nodes = create_workflow_nodes(workflow_id, nodes_data, 1)
-            create_workflow_edges(workflow_id, edges_data, 1)
+            with db_session() as conn:
+                repo = WorkflowsRepository(conn)
+                workflow = repo.create(user_id, name, description=description)
+                pg_workflow_id = str(workflow["id"])
+                _write_graph(conn, pg_workflow_id, 1, nodes_data, edges_data)
         except Exception as err:
-            workflow_nodes_collection.delete_many({"workflow_id": workflow_id})
-            workflow_edges_collection.delete_many({"workflow_id": workflow_id})
-            workflows_collection.delete_one({"_id": result.inserted_id})
-            return _workflow_error_response("Failed to create workflow structure", err)
+            return _workflow_error_response("Failed to create workflow", err)
 
-        _dual_write_workflow_create(
-            workflow_id,
-            user_id,
-            name,
-            data.get("description", ""),
-            created_nodes,
-            edges_data,
-        )
-
-        return success_response({"id": workflow_id}, 201)
+        return success_response({"id": pg_workflow_id}, 201)
 
 
 @workflows_ns.route("/workflows/<string:workflow_id>")
@@ -602,23 +423,22 @@ class WorkflowDetail(Resource):
     def get(self, workflow_id: str):
         """Get workflow details with nodes and edges."""
         user_id = get_user_id()
-        obj_id, error = validate_object_id(workflow_id, "Workflow")
-        if error:
-            return error
-
-        workflow, error = check_resource_ownership(
-            workflows_collection, obj_id, user_id, "Workflow"
-        )
-        if error:
-            return error
-
-        graph_version = get_workflow_graph_version(workflow)
-        nodes = fetch_graph_documents(
-            workflow_nodes_collection, workflow_id, graph_version
-        )
-        edges = fetch_graph_documents(
-            workflow_edges_collection, workflow_id, graph_version
-        )
+        try:
+            with db_readonly() as conn:
+                repo = WorkflowsRepository(conn)
+                workflow = _resolve_workflow(repo, workflow_id, user_id)
+                if workflow is None:
+                    return error_response("Workflow not found", 404)
+                pg_workflow_id = str(workflow["id"])
+                graph_version = get_workflow_graph_version(workflow)
+                nodes = WorkflowNodesRepository(conn).find_by_version(
+                    pg_workflow_id, graph_version,
+                )
+                edges = WorkflowEdgesRepository(conn).find_by_version(
+                    pg_workflow_id, graph_version,
+                )
+        except Exception as err:
+            return _workflow_error_response("Failed to fetch workflow", err)
 
         return success_response(
             {
@@ -633,89 +453,51 @@ class WorkflowDetail(Resource):
     def put(self, workflow_id: str):
         """Update workflow and replace nodes/edges."""
         user_id = get_user_id()
-        obj_id, error = validate_object_id(workflow_id, "Workflow")
-        if error:
-            return error
-
-        workflow, error = check_resource_ownership(
-            workflows_collection, obj_id, user_id, "Workflow"
-        )
-        if error:
-            return error
-
         data = request.get_json()
         name = data.get("name", "").strip()
+        description = data.get("description", "")
         nodes_data = data.get("nodes", [])
         edges_data = data.get("edges", [])
 
-        validation_errors = validate_workflow_structure(nodes_data, edges_data)
+        validation_errors = validate_workflow_structure(
+            nodes_data, edges_data, user_id=user_id
+        )
         if validation_errors:
             return error_response(
                 "Workflow validation failed", errors=validation_errors
             )
         nodes_data = normalize_agent_node_json_schemas(nodes_data)
 
-        current_graph_version = get_workflow_graph_version(workflow)
-        next_graph_version = current_graph_version + 1
         try:
-            created_nodes = create_workflow_nodes(
-                workflow_id, nodes_data, next_graph_version,
-            )
-            create_workflow_edges(workflow_id, edges_data, next_graph_version)
-        except Exception as err:
-            workflow_nodes_collection.delete_many(
-                {"workflow_id": workflow_id, "graph_version": next_graph_version}
-            )
-            workflow_edges_collection.delete_many(
-                {"workflow_id": workflow_id, "graph_version": next_graph_version}
-            )
-            return _workflow_error_response("Failed to update workflow structure", err)
+            with db_session() as conn:
+                repo = WorkflowsRepository(conn)
+                workflow = _resolve_workflow(repo, workflow_id, user_id)
+                if workflow is None:
+                    return error_response("Workflow not found", 404)
+                pg_workflow_id = str(workflow["id"])
+                current_graph_version = get_workflow_graph_version(workflow)
+                next_graph_version = current_graph_version + 1
 
-        now = datetime.now(timezone.utc)
-        _, error = safe_db_operation(
-            lambda: workflows_collection.update_one(
-                {"_id": obj_id},
-                {
-                    "$set": {
+                _write_graph(
+                    conn, pg_workflow_id, next_graph_version,
+                    nodes_data, edges_data,
+                )
+                repo.update(
+                    pg_workflow_id, user_id,
+                    {
                         "name": name,
-                        "description": data.get("description", ""),
-                        "updated_at": now,
+                        "description": description,
                         "current_graph_version": next_graph_version,
-                    }
-                },
-            ),
-            "Failed to update workflow",
-        )
-        if error:
-            workflow_nodes_collection.delete_many(
-                {"workflow_id": workflow_id, "graph_version": next_graph_version}
-            )
-            workflow_edges_collection.delete_many(
-                {"workflow_id": workflow_id, "graph_version": next_graph_version}
-            )
-            return error
-
-        try:
-            workflow_nodes_collection.delete_many(
-                {"workflow_id": workflow_id, "graph_version": {"$ne": next_graph_version}}
-            )
-            workflow_edges_collection.delete_many(
-                {"workflow_id": workflow_id, "graph_version": {"$ne": next_graph_version}}
-            )
-        except Exception as cleanup_err:
-            current_app.logger.warning(
-                f"Failed to clean old workflow graph versions for {workflow_id}: {cleanup_err}"
-            )
-
-        _dual_write_workflow_update(
-            workflow_id,
-            user_id,
-            name,
-            data.get("description", ""),
-            created_nodes,
-            edges_data,
-            next_graph_version,
-        )
+                    },
+                )
+                WorkflowNodesRepository(conn).delete_other_versions(
+                    pg_workflow_id, next_graph_version,
+                )
+                WorkflowEdgesRepository(conn).delete_other_versions(
+                    pg_workflow_id, next_graph_version,
+                )
+        except Exception as err:
+            return _workflow_error_response("Failed to update workflow", err)
 
         return success_response()
 
@@ -723,23 +505,15 @@ class WorkflowDetail(Resource):
     def delete(self, workflow_id: str):
         """Delete workflow and its graph."""
         user_id = get_user_id()
-        obj_id, error = validate_object_id(workflow_id, "Workflow")
-        if error:
-            return error
-
-        workflow, error = check_resource_ownership(
-            workflows_collection, obj_id, user_id, "Workflow"
-        )
-        if error:
-            return error
-
         try:
-            workflow_nodes_collection.delete_many({"workflow_id": workflow_id})
-            workflow_edges_collection.delete_many({"workflow_id": workflow_id})
-            workflows_collection.delete_one({"_id": workflow["_id"], "user": user_id})
+            with db_session() as conn:
+                repo = WorkflowsRepository(conn)
+                workflow = _resolve_workflow(repo, workflow_id, user_id)
+                if workflow is None:
+                    return error_response("Workflow not found", 404)
+                # ON DELETE CASCADE on workflow_nodes/edges cleans children.
+                repo.delete(str(workflow["id"]), user_id)
         except Exception as err:
             return _workflow_error_response("Failed to delete workflow", err)
 
-        _dual_write_workflow_delete(workflow_id, user_id)
-
         return success_response()

application/api/v1/routes.py

@@ -20,8 +20,8 @@ from application.api.v1.translator import (
     translate_response,
     translate_stream_event,
 )
-from application.core.mongo_db import MongoDB
-from application.core.settings import settings
+from application.storage.db.repositories.agents import AgentsRepository
+from application.storage.db.session import db_readonly
 
 logger = logging.getLogger(__name__)
 
@@ -39,9 +39,8 @@ def _extract_bearer_token() -> Optional[str]:
 def _lookup_agent(api_key: str) -> Optional[Dict]:
     """Look up the agent document for this API key."""
     try:
-        mongo = MongoDB.get_client()
-        db = mongo[settings.MONGO_DB_NAME]
-        return db["agents"].find_one({"key": api_key})
+        with db_readonly() as conn:
+            return AgentsRepository(conn).find_by_key(api_key)
     except Exception:
         logger.warning("Failed to look up agent for API key", exc_info=True)
         return None
@@ -90,8 +89,14 @@ def chat_completions():
         )
 
     # Link decoded_token to the agent's owner so continuation state,
-    # logs, and tool execution use the correct user identity.
-    agent_user = agent_doc.get("user") if agent_doc else None
+    # logs, and tool execution use the correct user identity. The PG
+    # ``agents`` row exposes the owner via ``user_id`` (``user`` is the
+    # legacy Mongo field name kept in ``row_to_dict`` only for the
+    # mapping ``id``/``_id``).
+    agent_user = (
+        (agent_doc.get("user_id") or agent_doc.get("user"))
+        if agent_doc else None
+    )
     decoded_token = {"sub": agent_user or "api_key_user"}
 
     try:
@@ -208,6 +213,7 @@ def _stream_response(
         decoded_token=processor.decoded_token,
         agent_id=processor.agent_id,
         model_id=processor.model_id,
+        model_user_id=processor.model_user_id,
         should_save_conversation=should_save_conversation,
         _continuation=continuation,
     )
@@ -252,6 +258,7 @@ def _non_stream_response(
         decoded_token=processor.decoded_token,
         agent_id=processor.agent_id,
         model_id=processor.model_id,
+        model_user_id=processor.model_user_id,
         should_save_conversation=should_save_conversation,
         _continuation=continuation,
     )
@@ -290,39 +297,32 @@ def list_models():
         )
 
     try:
-        mongo = MongoDB.get_client()
-        db = mongo[settings.MONGO_DB_NAME]
-        agents_collection = db["agents"]
+        with db_readonly() as conn:
+            agents_repo = AgentsRepository(conn)
+            agent = agents_repo.find_by_key(api_key)
+            if not agent:
+                return make_response(
+                    jsonify({"error": {"message": "Invalid API key", "type": "auth_error"}}),
+                    401,
+                )
 
-        # Find the agent for this api_key
-        agent = agents_collection.find_one({"key": api_key})
-        if not agent:
-            return make_response(
-                jsonify({"error": {"message": "Invalid API key", "type": "auth_error"}}),
-                401,
-            )
-
-        user = agent.get("user")
-
-        # Return all agents belonging to this user
-        user_agents = list(agents_collection.find({"user": user}))
-
-        models = []
-        for ag in user_agents:
-            created = ag.get("createdAt")
-            created_ts = int(created.timestamp()) if created else int(time.time())
-            model_id = str(ag.get("_id") or ag.get("id") or "")
-            models.append({
-                "id": model_id,
-                "object": "model",
-                "created": created_ts,
-                "owned_by": "docsgpt",
-                "name": ag.get("name", ""),
-                "description": ag.get("description", ""),
-            })
+        created = agent.get("created_at") or agent.get("createdAt")
+        created_ts = (
+            int(created.timestamp()) if hasattr(created, "timestamp")
+            else int(time.time())
+        )
+        model_id = str(agent.get("id") or agent.get("_id") or "")
+        model = {
+            "id": model_id,
+            "object": "model",
+            "created": created_ts,
+            "owned_by": "docsgpt",
+            "name": agent.get("name", ""),
+            "description": agent.get("description", ""),
+        }
 
         return make_response(
-            jsonify({"object": "list", "data": models}),
+            jsonify({"object": "list", "data": [model]}),
             200,
         )
     except Exception as e:

application/app.py

@@ -1,9 +1,10 @@
+import logging
 import os
 import platform
 import uuid
 
 import dotenv
-from flask import Flask, jsonify, redirect, request
+from flask import Flask, Response, jsonify, redirect, request
 from jose import jwt
 
 from application.auth import handle_auth
@@ -20,6 +21,7 @@ from application.api.connector.routes import connector  # noqa: E402
 from application.api.v1 import v1_bp  # noqa: E402
 from application.celery_init import celery  # noqa: E402
 from application.core.settings import settings  # noqa: E402
+from application.storage.db.bootstrap import ensure_database_ready  # noqa: E402
 from application.stt.upload_limits import (  # noqa: E402
     build_stt_file_size_limit_message,
     should_reject_stt_request,
@@ -32,6 +34,17 @@ if platform.system() == "Windows":
     pathlib.PosixPath = pathlib.WindowsPath
 dotenv.load_dotenv()
 
+# Self-bootstrap the user-data Postgres DB. Runs before any blueprint or
+# repository touches the engine, so the first request can't race the
+# schema being created. Gated by AUTO_CREATE_DB / AUTO_MIGRATE settings
+# (default ON for dev; disable in prod if schema is managed out-of-band).
+ensure_database_ready(
+    settings.POSTGRES_URI,
+    create_db=settings.AUTO_CREATE_DB,
+    migrate=settings.AUTO_MIGRATE,
+    logger=logging.getLogger("application.app"),
+)
+
 app = Flask(__name__)
 app.register_blueprint(user)
 app.register_blueprint(answer)
@@ -120,6 +133,12 @@ def enforce_stt_request_size_limits():
 def authenticate_request():
     if request.method == "OPTIONS":
         return "", 200
+    # OpenAI-compatible routes authenticate via opaque agent API keys in the
+    # Authorization header, which the JWT decoder below would reject. Defer
+    # auth to the route handlers (see application/api/v1/routes.py).
+    if request.path.startswith("/v1/"):
+        request.decoded_token = None
+        return None
     decoded_token = handle_auth(request)
     if not decoded_token:
         request.decoded_token = None
@@ -130,12 +149,11 @@ def authenticate_request():
 
 
 @app.after_request
-def after_request(response):
-    response.headers.add("Access-Control-Allow-Origin", "*")
-    response.headers.add("Access-Control-Allow-Headers", "Content-Type, Authorization")
-    response.headers.add(
-        "Access-Control-Allow-Methods", "GET, POST, PUT, DELETE, OPTIONS"
-    )
+def after_request(response: Response) -> Response:
+    """Add CORS headers for the pure Flask development entrypoint."""
+    response.headers["Access-Control-Allow-Origin"] = "*"
+    response.headers["Access-Control-Allow-Headers"] = "Content-Type, Authorization"
+    response.headers["Access-Control-Allow-Methods"] = "GET, POST, PUT, PATCH, DELETE, OPTIONS"
     return response
 
 

application/asgi.py

@@ -0,0 +1,33 @@
+"""ASGI entrypoint: Flask (WSGI) + FastMCP on the same process."""
+
+from __future__ import annotations
+
+from a2wsgi import WSGIMiddleware
+from starlette.applications import Starlette
+from starlette.middleware import Middleware
+from starlette.middleware.cors import CORSMiddleware
+from starlette.routing import Mount
+
+from application.app import app as flask_app
+from application.mcp_server import mcp
+
+_WSGI_THREADPOOL = 32
+
+mcp_app = mcp.http_app(path="/")
+
+asgi_app = Starlette(
+    routes=[
+        Mount("/mcp", app=mcp_app),
+        Mount("/", app=WSGIMiddleware(flask_app, workers=_WSGI_THREADPOOL)),
+    ],
+    middleware=[
+        Middleware(
+            CORSMiddleware,
+            allow_origins=["*"],
+            allow_methods=["GET", "POST", "PUT", "PATCH", "DELETE", "OPTIONS"],
+            allow_headers=["Content-Type", "Authorization", "Mcp-Session-Id"],
+            expose_headers=["Mcp-Session-Id"],
+        ),
+    ],
+    lifespan=mcp_app.lifespan,
+)

application/celery_init.py

@@ -1,6 +1,8 @@
+import threading
+
 from celery import Celery
 from application.core.settings import settings
-from celery.signals import setup_logging, worker_process_init
+from celery.signals import setup_logging, worker_process_init, worker_ready
 
 
 def make_celery(app_name=__name__):
@@ -39,5 +41,25 @@ def _dispose_db_engine_on_fork(*args, **kwargs):
     dispose_engine()
 
 
+@worker_ready.connect
+def _run_version_check(*args, **kwargs):
+    """Kick off the anonymous version check on worker startup.
+
+    Runs in a daemon thread so a slow endpoint or bad DNS never holds
+    up the worker becoming ready for tasks. The check itself is
+    fail-silent (see ``application.updates.version_check.run_check``);
+    this handler's only job is to launch it and get out of the way.
+
+    Import is lazy so the symbol resolution never fires at module
+    import time — consistent with the ``_dispose_db_engine_on_fork``
+    pattern above.
+    """
+    try:
+        from application.updates.version_check import run_check
+    except Exception:
+        return
+    threading.Thread(target=run_check, name="version-check", daemon=True).start()
+
+
 celery = make_celery()
 celery.config_from_object("application.celeryconfig")

application/celeryconfig.py

@@ -9,3 +9,8 @@ accept_content = ['json']
 
 # Autodiscover tasks
 imports = ('application.api.user.tasks',)
+
+beat_scheduler = "redbeat.RedBeatScheduler"
+redbeat_redis_url = broker_url
+redbeat_key_prefix = "redbeat:docsgpt:"
+redbeat_lock_timeout = 90

application/core/logging_config.py

@@ -1,11 +1,45 @@
+import logging
+import os
 from logging.config import dictConfig
 
-def setup_logging():
+
+def _otlp_logs_enabled() -> bool:
+    """Return True when the user has opted in to OTLP log export.
+
+    Gated by the standard OTEL env vars so no project-specific knob is needed:
+    set ``OTEL_LOGS_EXPORTER=otlp`` (and leave ``OTEL_SDK_DISABLED`` unset or
+    false) to flip it on. When false, ``setup_logging`` keeps its original
+    console-only behavior.
+    """
+    exporter = os.getenv("OTEL_LOGS_EXPORTER", "").strip().lower()
+    disabled = os.getenv("OTEL_SDK_DISABLED", "false").strip().lower() == "true"
+    return exporter == "otlp" and not disabled
+
+
+def setup_logging() -> None:
+    """Configure the root logger with a stdout console handler.
+
+    When OTLP log export is enabled, ``opentelemetry-instrument`` attaches a
+    ``LoggingHandler`` to the root logger before this function runs. The
+    ``dictConfig`` call below replaces ``root.handlers`` with the console
+    handler, which would silently drop the OTEL handler. To make OTLP log
+    export work without forcing every contributor to opt in, snapshot the
+    OTEL handlers up front and re-attach them after ``dictConfig``.
+    """
+    preserved_handlers: list[logging.Handler] = []
+    if _otlp_logs_enabled():
+        preserved_handlers = [
+            h
+            for h in logging.getLogger().handlers
+            if h.__class__.__module__.startswith("opentelemetry")
+        ]
+
     dictConfig({
-        'version': 1,
-        'formatters': {
-            'default': {
-                'format': '[%(asctime)s] %(levelname)s in %(module)s: %(message)s',
+        "version": 1,
+        "disable_existing_loggers": False,
+        "formatters": {
+            "default": {
+                "format": "[%(asctime)s] %(levelname)s in %(module)s: %(message)s",
             }
         },
         "handlers": {
@@ -15,8 +49,14 @@ def setup_logging():
                 "formatter": "default",
             }
         },
-        'root': {
-            'level': 'INFO',
-            'handlers': ['console'],
+        "root": {
+            "level": "INFO",
+            "handlers": ["console"],
         },
-    })
+    })
+
+    if preserved_handlers:
+        root = logging.getLogger()
+        for handler in preserved_handlers:
+            if handler not in root.handlers:
+                root.addHandler(handler)

application/core/model_configs.py

@@ -1,266 +0,0 @@
-"""
-Model configurations for all supported LLM providers.
-"""
-
-from application.core.model_settings import (
-    AvailableModel,
-    ModelCapabilities,
-    ModelProvider,
-)
-
-# Base image attachment types supported by most vision-capable LLMs
-IMAGE_ATTACHMENTS = [
-    "image/png",
-    "image/jpeg",
-    "image/jpg",
-    "image/webp",
-    "image/gif",
-]
-
-# PDF excluded: most OpenAI-compatible endpoints don't support native PDF uploads.
-# When excluded, PDFs are synthetically processed by converting pages to images.
-OPENAI_ATTACHMENTS = IMAGE_ATTACHMENTS
-
-GOOGLE_ATTACHMENTS = ["application/pdf"] + IMAGE_ATTACHMENTS
-
-ANTHROPIC_ATTACHMENTS = IMAGE_ATTACHMENTS
-
-OPENROUTER_ATTACHMENTS = IMAGE_ATTACHMENTS
-
-NOVITA_ATTACHMENTS = IMAGE_ATTACHMENTS
-
-
-OPENAI_MODELS = [
-    AvailableModel(
-        id="gpt-5.1",
-        provider=ModelProvider.OPENAI,
-        display_name="GPT-5.1",
-        description="Flagship model with enhanced reasoning, coding, and agentic capabilities",
-        capabilities=ModelCapabilities(
-            supports_tools=True,
-            supports_structured_output=True,
-            supported_attachment_types=OPENAI_ATTACHMENTS,
-            context_window=200000,
-        ),
-    ),
-    AvailableModel(
-        id="gpt-5-mini",
-        provider=ModelProvider.OPENAI,
-        display_name="GPT-5 Mini",
-        description="Faster, cost-effective variant of GPT-5.1",
-        capabilities=ModelCapabilities(
-            supports_tools=True,
-            supports_structured_output=True,
-            supported_attachment_types=OPENAI_ATTACHMENTS,
-            context_window=200000,
-        ),
-    )
-]
-
-
-ANTHROPIC_MODELS = [
-    AvailableModel(
-        id="claude-3-5-sonnet-20241022",
-        provider=ModelProvider.ANTHROPIC,
-        display_name="Claude 3.5 Sonnet (Latest)",
-        description="Latest Claude 3.5 Sonnet with enhanced capabilities",
-        capabilities=ModelCapabilities(
-            supports_tools=True,
-            supported_attachment_types=ANTHROPIC_ATTACHMENTS,
-            context_window=200000,
-        ),
-    ),
-    AvailableModel(
-        id="claude-3-5-sonnet",
-        provider=ModelProvider.ANTHROPIC,
-        display_name="Claude 3.5 Sonnet",
-        description="Balanced performance and capability",
-        capabilities=ModelCapabilities(
-            supports_tools=True,
-            supported_attachment_types=ANTHROPIC_ATTACHMENTS,
-            context_window=200000,
-        ),
-    ),
-    AvailableModel(
-        id="claude-3-opus",
-        provider=ModelProvider.ANTHROPIC,
-        display_name="Claude 3 Opus",
-        description="Most capable Claude model",
-        capabilities=ModelCapabilities(
-            supports_tools=True,
-            supported_attachment_types=ANTHROPIC_ATTACHMENTS,
-            context_window=200000,
-        ),
-    ),
-    AvailableModel(
-        id="claude-3-haiku",
-        provider=ModelProvider.ANTHROPIC,
-        display_name="Claude 3 Haiku",
-        description="Fastest Claude model",
-        capabilities=ModelCapabilities(
-            supports_tools=True,
-            supported_attachment_types=ANTHROPIC_ATTACHMENTS,
-            context_window=200000,
-        ),
-    ),
-]
-
-
-GOOGLE_MODELS = [
-    AvailableModel(
-        id="gemini-flash-latest",
-        provider=ModelProvider.GOOGLE,
-        display_name="Gemini Flash (Latest)",
-        description="Latest experimental Gemini model",
-        capabilities=ModelCapabilities(
-            supports_tools=True,
-            supports_structured_output=True,
-            supported_attachment_types=GOOGLE_ATTACHMENTS,
-            context_window=int(1e6),
-        ),
-    ),
-    AvailableModel(
-        id="gemini-flash-lite-latest",
-        provider=ModelProvider.GOOGLE,
-        display_name="Gemini Flash Lite (Latest)",
-        description="Fast with huge context window",
-        capabilities=ModelCapabilities(
-            supports_tools=True,
-            supports_structured_output=True,
-            supported_attachment_types=GOOGLE_ATTACHMENTS,
-            context_window=int(1e6),
-        ),
-    ),
-    AvailableModel(
-        id="gemini-3-pro-preview",
-        provider=ModelProvider.GOOGLE,
-        display_name="Gemini 3 Pro",
-        description="Most capable Gemini model",
-        capabilities=ModelCapabilities(
-            supports_tools=True,
-            supports_structured_output=True,
-            supported_attachment_types=GOOGLE_ATTACHMENTS,
-            context_window=2000000,
-        ),
-    ),
-]
-
-
-GROQ_MODELS = [
-    AvailableModel(
-        id="llama-3.3-70b-versatile",
-        provider=ModelProvider.GROQ,
-        display_name="Llama 3.3 70B",
-        description="Latest Llama model with high-speed inference",
-        capabilities=ModelCapabilities(
-            supports_tools=True,
-            context_window=128000,
-        ),
-    ),
-    AvailableModel(
-        id="openai/gpt-oss-120b",
-        provider=ModelProvider.GROQ,
-        display_name="GPT-OSS 120B",
-        description="Open-source GPT model optimized for speed",
-        capabilities=ModelCapabilities(
-            supports_tools=True,
-            context_window=128000,
-        ),
-    ),
-]
-
-
-OPENROUTER_MODELS = [
-    AvailableModel(
-        id="qwen/qwen3-coder:free",
-        provider=ModelProvider.OPENROUTER,
-        display_name="Qwen 3 Coder",
-        description="Latest Qwen model with high-speed inference",
-        capabilities=ModelCapabilities(
-            supports_tools=True,
-            context_window=128000,
-            supported_attachment_types=OPENROUTER_ATTACHMENTS
-        ),
-    ),
-    AvailableModel(
-        id="google/gemma-3-27b-it:free",
-        provider=ModelProvider.OPENROUTER,
-        display_name="Gemma 3 27B",
-        description="Latest Gemma model with high-speed inference",
-        capabilities=ModelCapabilities(
-            supports_tools=True,
-            context_window=128000,
-            supported_attachment_types=OPENROUTER_ATTACHMENTS
-        ),
-    ),
-]
-
-NOVITA_MODELS = [
-    AvailableModel(
-        id="moonshotai/kimi-k2.5",
-        provider=ModelProvider.NOVITA,
-        display_name="Kimi K2.5",
-        description="MoE model with function calling, structured output, reasoning, and vision",
-        capabilities=ModelCapabilities(
-            supports_tools=True,
-            supports_structured_output=True,
-            supported_attachment_types=NOVITA_ATTACHMENTS,
-            context_window=262144,
-        ),
-    ),
-    AvailableModel(
-        id="zai-org/glm-5",
-        provider=ModelProvider.NOVITA,
-        display_name="GLM-5",
-        description="MoE model with function calling, structured output, and reasoning",
-        capabilities=ModelCapabilities(
-            supports_tools=True,
-            supports_structured_output=True,
-            supported_attachment_types=[],
-            context_window=202800,
-        ),
-    ),
-    AvailableModel(
-        id="minimax/minimax-m2.5",
-        provider=ModelProvider.NOVITA,
-        display_name="MiniMax M2.5",
-        description="MoE model with function calling, structured output, and reasoning",
-        capabilities=ModelCapabilities(
-            supports_tools=True,
-            supports_structured_output=True,
-            supported_attachment_types=[],
-            context_window=204800,
-        ),
-    ),
-]
-
-
-AZURE_OPENAI_MODELS = [
-    AvailableModel(
-        id="azure-gpt-4",
-        provider=ModelProvider.AZURE_OPENAI,
-        display_name="Azure OpenAI GPT-4",
-        description="Azure-hosted GPT model",
-        capabilities=ModelCapabilities(
-            supports_tools=True,
-            supports_structured_output=True,
-            supported_attachment_types=OPENAI_ATTACHMENTS,
-            context_window=8192,
-        ),
-    ),
-]
-
-
-def create_custom_openai_model(model_name: str, base_url: str) -> AvailableModel:
-    """Create a custom OpenAI-compatible model (e.g., LM Studio, Ollama)."""
-    return AvailableModel(
-        id=model_name,
-        provider=ModelProvider.OPENAI,
-        display_name=model_name,
-        description=f"Custom OpenAI-compatible model at {base_url}",
-        base_url=base_url,
-        capabilities=ModelCapabilities(
-            supports_tools=True,
-            supported_attachment_types=OPENAI_ATTACHMENTS,
-        ),
-    )

application/core/model_registry.py

@@ -0,0 +1,385 @@
+"""Layered model registry.
+
+Loads model catalogs from YAML files (built-in + operator-supplied),
+groups them by provider name, then for each registered provider plugin
+calls ``get_models`` to produce the final per-provider model list.
+
+End-user BYOM (per-user model records in Postgres) is layered on top:
+when a lookup arrives with a ``user_id``, the registry consults a
+per-user cache first (loaded from the ``user_custom_models`` table on
+miss) and falls through to the built-in catalog.
+
+Cross-process invalidation: ``ModelRegistry`` is a per-process
+singleton, so a CRUD write only evicts the cache in the process that
+served it. Other gunicorn workers and Celery workers would otherwise
+keep using a deleted/disabled/key-rotated BYOM record indefinitely.
+``invalidate_user`` therefore both drops the local layer *and* bumps a
+Redis-side version counter; other processes notice the bump on their
+next access (after the local TTL window) and reload from Postgres. If
+Redis is unreachable the per-process TTL still bounds staleness — pure
+TTL semantics, no regression.
+"""
+
+from __future__ import annotations
+
+import logging
+import time
+from collections import defaultdict
+from typing import Dict, List, Optional, Tuple
+
+from application.core.model_settings import AvailableModel
+from application.core.model_yaml import (
+    BUILTIN_MODELS_DIR,
+    ProviderCatalog,
+    load_model_yamls,
+)
+
+logger = logging.getLogger(__name__)
+
+_USER_CACHE_TTL_SECONDS = 60.0
+_USER_VERSION_KEY_PREFIX = "byom:registry_version:"
+
+
+class ModelRegistry:
+    """Singleton registry of available models."""
+
+    _instance: Optional["ModelRegistry"] = None
+    _initialized: bool = False
+
+    def __new__(cls):
+        if cls._instance is None:
+            cls._instance = super().__new__(cls)
+        return cls._instance
+
+    def __init__(self):
+        if not ModelRegistry._initialized:
+            self.models: Dict[str, AvailableModel] = {}
+            self.default_model_id: Optional[str] = None
+            # Per-user BYOM cache. Each entry is
+            # ``(layer, version_at_load, loaded_at_monotonic)``:
+            #   * ``layer`` — {model_id: AvailableModel}
+            #   * ``version_at_load`` — Redis-side counter snapshot at
+            #     reload time, or ``None`` if Redis was unreachable
+            #   * ``loaded_at_monotonic`` — for TTL bookkeeping
+            # Populated lazily, evicted by TTL + cross-process
+            # invalidation (see ``invalidate_user``).
+            self._user_models: Dict[
+                str,
+                Tuple[Dict[str, AvailableModel], Optional[int], float],
+            ] = {}
+            self._load_models()
+            ModelRegistry._initialized = True
+
+    @classmethod
+    def get_instance(cls) -> "ModelRegistry":
+        return cls()
+
+    @classmethod
+    def reset(cls) -> None:
+        """Clear the singleton. Intended for test fixtures."""
+        cls._instance = None
+        cls._initialized = False
+
+    @classmethod
+    def invalidate_user(cls, user_id: str) -> None:
+        """Drop the cached per-user model layer for ``user_id``.
+
+        Called by the BYOM REST routes after every create/update/delete.
+        Two effects:
+
+        * Local: pop the entry from this process's cache so the next
+          lookup re-reads from Postgres immediately.
+        * Cross-process: ``INCR`` a Redis-side version counter for this
+          user. Other gunicorn/Celery processes notice the counter
+          changed on their next TTL-driven recheck (see
+          ``_user_models_for``) and reload. If Redis is unreachable we
+          log and continue — local invalidation still happened, and
+          peers fall back to TTL-only staleness bounds.
+        """
+        if cls._instance is not None:
+            cls._instance._user_models.pop(user_id, None)
+        try:
+            from application.cache import get_redis_instance
+
+            client = get_redis_instance()
+            if client is not None:
+                client.incr(_USER_VERSION_KEY_PREFIX + user_id)
+        except Exception as e:
+            logger.warning(
+                "BYOM invalidate: failed to publish version bump for "
+                "user %s (Redis unreachable?): %s",
+                user_id,
+                e,
+            )
+
+    @classmethod
+    def _read_user_version(cls, user_id: str) -> Optional[int]:
+        """Return the Redis-side invalidation counter for ``user_id``.
+
+        ``0`` if the key has never been bumped; ``None`` if Redis is
+        unreachable or the read failed (callers fall back to TTL-only
+        staleness in that case).
+        """
+        try:
+            from application.cache import get_redis_instance
+
+            client = get_redis_instance()
+            if client is None:
+                return None
+            raw = client.get(_USER_VERSION_KEY_PREFIX + user_id)
+            if raw is None:
+                return 0
+            return int(raw)
+        except Exception:
+            return None
+
+    def _load_models(self) -> None:
+        from pathlib import Path
+
+        from application.core.settings import settings
+        from application.llm.providers import ALL_PROVIDERS
+
+        directories = [BUILTIN_MODELS_DIR]
+        operator_dir = getattr(settings, "MODELS_CONFIG_DIR", None)
+        if operator_dir:
+            op_path = Path(operator_dir)
+            if not op_path.exists():
+                logger.warning(
+                    "MODELS_CONFIG_DIR=%s does not exist; no operator "
+                    "model YAMLs will be loaded.",
+                    operator_dir,
+                )
+            elif not op_path.is_dir():
+                logger.warning(
+                    "MODELS_CONFIG_DIR=%s is not a directory; no operator "
+                    "model YAMLs will be loaded.",
+                    operator_dir,
+                )
+            else:
+                directories.append(op_path)
+
+        catalogs = load_model_yamls(directories)
+
+        # Validate every catalog targets a known plugin before doing any
+        # registry work, so an unknown provider name in YAML aborts boot
+        # with a clear error.
+        plugin_names = {p.name for p in ALL_PROVIDERS}
+        for c in catalogs:
+            if c.provider not in plugin_names:
+                raise ValueError(
+                    f"{c.source_path}: YAML declares unknown provider "
+                    f"{c.provider!r}; no Provider plugin is registered "
+                    f"under that name. Known: {sorted(plugin_names)}"
+                )
+
+        catalogs_by_provider: Dict[str, List[ProviderCatalog]] = defaultdict(list)
+        for c in catalogs:
+            catalogs_by_provider[c.provider].append(c)
+
+        self.models.clear()
+        for provider in ALL_PROVIDERS:
+            if not provider.is_enabled(settings):
+                continue
+            for model in provider.get_models(
+                settings, catalogs_by_provider.get(provider.name, [])
+            ):
+                self.models[model.id] = model
+
+        self.default_model_id = self._resolve_default(settings)
+
+        logger.info(
+            "ModelRegistry loaded %d models, default: %s",
+            len(self.models),
+            self.default_model_id,
+        )
+
+    def _resolve_default(self, settings) -> Optional[str]:
+        if settings.LLM_NAME:
+            for name in self._parse_model_names(settings.LLM_NAME):
+                if name in self.models:
+                    return name
+            if settings.LLM_NAME in self.models:
+                return settings.LLM_NAME
+
+        if settings.LLM_PROVIDER and settings.API_KEY:
+            for model_id, model in self.models.items():
+                if model.provider.value == settings.LLM_PROVIDER:
+                    return model_id
+
+        if self.models:
+            return next(iter(self.models.keys()))
+        return None
+
+    @staticmethod
+    def _parse_model_names(llm_name: str) -> List[str]:
+        if not llm_name:
+            return []
+        return [name.strip() for name in llm_name.split(",") if name.strip()]
+
+    # Per-user (BYOM) layer
+
+    def _user_models_for(self, user_id: str) -> Dict[str, AvailableModel]:
+        """Return the user's BYOM models keyed by registry id (UUID).
+
+        Loaded lazily from Postgres on first access; cached subject to
+        a per-process TTL (``_USER_CACHE_TTL_SECONDS``) and a Redis-
+        backed version counter for cross-process invalidation. The TTL
+        bounds staleness even when Redis is unreachable, while the
+        version stamp lets peers refresh without a DB read on the
+        common case (no invalidation since last load). Decryption
+        failures and DB errors yield an empty layer (logged) — the
+        user simply doesn't see their custom models on this request,
+        never a 500.
+        """
+        cached = self._user_models.get(user_id)
+        now = time.monotonic()
+
+        if cached is not None:
+            layer, cached_version, loaded_at = cached
+            if (now - loaded_at) < _USER_CACHE_TTL_SECONDS:
+                return layer
+            # TTL elapsed: peek at the cross-process counter. If it
+            # matches what we saw at load time, no invalidation has
+            # happened — extend the TTL without touching Postgres. If
+            # Redis is unreachable (``current_version is None``) we
+            # fall through to a real reload, which keeps staleness
+            # bounded to the TTL.
+            current_version = self._read_user_version(user_id)
+            if (
+                current_version is not None
+                and cached_version is not None
+                and current_version == cached_version
+            ):
+                self._user_models[user_id] = (layer, cached_version, now)
+                return layer
+
+        # Capture the counter *before* the DB read so a CRUD that lands
+        # mid-reload doesn't get masked: the next access will see a
+        # newer version and reload again.
+        version_before_read = self._read_user_version(user_id)
+
+        layer: Dict[str, AvailableModel] = {}
+        try:
+            from application.core.model_settings import (
+                ModelCapabilities,
+                ModelProvider,
+            )
+            from application.storage.db.repositories.user_custom_models import (
+                UserCustomModelsRepository,
+            )
+            from application.storage.db.session import db_readonly
+
+            with db_readonly() as conn:
+                repo = UserCustomModelsRepository(conn)
+                rows = repo.list_for_user(user_id)
+                for row in rows:
+                    api_key = repo._decrypt_api_key(
+                        row.get("api_key_encrypted", ""), user_id
+                    )
+                    if not api_key:
+                        # SECURITY: do NOT register an unroutable BYOM
+                        # record. If we did, LLMCreator would fall back
+                        # to the caller-passed api_key (settings.API_KEY
+                        # for openai_compatible) and POST it to the
+                        # user-supplied base_url — leaking the instance
+                        # credential to the user's chosen endpoint.
+                        # Most likely cause is ENCRYPTION_SECRET_KEY
+                        # having rotated; user must re-save the model.
+                        logger.warning(
+                            "user_custom_models: skipping model %s for "
+                            "user %s — api_key could not be decrypted "
+                            "(rotated ENCRYPTION_SECRET_KEY?). Re-save "
+                            "the model to recover.",
+                            row.get("id"),
+                            user_id,
+                        )
+                        continue
+                    caps_raw = row.get("capabilities") or {}
+                    # Stored attachments may be aliases (``image``) or
+                    # raw MIME types. Built-in YAML models expand at
+                    # load time; mirror that here so downstream MIME-
+                    # type comparisons (handlers/base.prepare_messages)
+                    # match concrete types like ``image/png`` rather
+                    # than the bare alias.
+                    from application.core.model_yaml import (
+                        expand_attachments_lenient,
+                    )
+
+                    raw_attachments = caps_raw.get("attachments", []) or []
+                    expanded_attachments = expand_attachments_lenient(
+                        raw_attachments,
+                        f"user_custom_models[user={user_id}, model={row.get('id')}]",
+                    )
+                    caps = ModelCapabilities(
+                        supports_tools=bool(caps_raw.get("supports_tools", False)),
+                        supports_structured_output=bool(
+                            caps_raw.get("supports_structured_output", False)
+                        ),
+                        supports_streaming=bool(
+                            caps_raw.get("supports_streaming", True)
+                        ),
+                        supported_attachment_types=expanded_attachments,
+                        context_window=int(
+                            caps_raw.get("context_window") or 128000
+                        ),
+                    )
+                    model_id = str(row["id"])
+                    layer[model_id] = AvailableModel(
+                        id=model_id,
+                        provider=ModelProvider.OPENAI_COMPATIBLE,
+                        display_name=row["display_name"],
+                        description=row.get("description") or "",
+                        capabilities=caps,
+                        enabled=bool(row.get("enabled", True)),
+                        base_url=row["base_url"],
+                        upstream_model_id=row["upstream_model_id"],
+                        source="user",
+                        api_key=api_key,
+                    )
+        except Exception as e:
+            logger.warning(
+                "user_custom_models: failed to load layer for user %s: %s",
+                user_id,
+                e,
+            )
+            layer = {}
+
+        self._user_models[user_id] = (layer, version_before_read, now)
+        return layer
+
+    # Lookup API. ``user_id`` enables the BYOM per-user layer; without
+    # it, callers see only the built-in + operator catalog.
+
+    def get_model(
+        self, model_id: str, user_id: Optional[str] = None
+    ) -> Optional[AvailableModel]:
+        if user_id:
+            user_layer = self._user_models_for(user_id)
+            if model_id in user_layer:
+                return user_layer[model_id]
+        return self.models.get(model_id)
+
+    def get_all_models(
+        self, user_id: Optional[str] = None
+    ) -> List[AvailableModel]:
+        out = list(self.models.values())
+        if user_id:
+            out.extend(self._user_models_for(user_id).values())
+        return out
+
+    def get_enabled_models(
+        self, user_id: Optional[str] = None
+    ) -> List[AvailableModel]:
+        out = [m for m in self.models.values() if m.enabled]
+        if user_id:
+            out.extend(
+                m for m in self._user_models_for(user_id).values() if m.enabled
+            )
+        return out
+
+    def model_exists(
+        self, model_id: str, user_id: Optional[str] = None
+    ) -> bool:
+        if user_id and model_id in self._user_models_for(user_id):
+            return True
+        return model_id in self.models

application/core/model_settings.py

@@ -5,9 +5,16 @@ from typing import Dict, List, Optional
 
 logger = logging.getLogger(__name__)
 
+# Re-exported here so existing call sites (and tests) that do
+# ``from application.core.model_settings import ModelRegistry`` keep
+# working. The implementation lives in ``application/core/model_registry.py``.
+# Imported lazily inside ``__getattr__`` to avoid an import cycle with
+# ``model_yaml`` → ``model_settings`` (this file).
+
 
 class ModelProvider(str, Enum):
     OPENAI = "openai"
+    OPENAI_COMPATIBLE = "openai_compatible"
     OPENROUTER = "openrouter"
     AZURE_OPENAI = "azure_openai"
     ANTHROPIC = "anthropic"
@@ -41,11 +48,21 @@ class AvailableModel:
     capabilities: ModelCapabilities = field(default_factory=ModelCapabilities)
     enabled: bool = True
     base_url: Optional[str] = None
+    # User-facing label distinct from dispatch provider (e.g. mistral
+    # routed through openai_compatible).
+    display_provider: Optional[str] = None
+    # Sent in the API call's ``model`` field; falls back to ``self.id``
+    # for built-ins where id IS the upstream name.
+    upstream_model_id: Optional[str] = None
+    # "builtin" for catalog YAMLs, "user" for BYOM records.
+    source: str = "builtin"
+    # Decrypted/resolved at registry-merge time. Never serialized.
+    api_key: Optional[str] = field(default=None, repr=False, compare=False)
 
     def to_dict(self) -> Dict:
         result = {
             "id": self.id,
-            "provider": self.provider.value,
+            "provider": self.display_provider or self.provider.value,
             "display_name": self.display_name,
             "description": self.description,
             "supported_attachment_types": self.capabilities.supported_attachment_types,
@@ -54,261 +71,21 @@ class AvailableModel:
             "supports_streaming": self.capabilities.supports_streaming,
             "context_window": self.capabilities.context_window,
             "enabled": self.enabled,
+            "source": self.source,
         }
         if self.base_url:
             result["base_url"] = self.base_url
         return result
 
 
-class ModelRegistry:
-    _instance = None
-    _initialized = False
+def __getattr__(name):
+    """Lazy re-export of ``ModelRegistry`` from ``model_registry.py``.
 
-    def __new__(cls):
-        if cls._instance is None:
-            cls._instance = super().__new__(cls)
-        return cls._instance
+    Done lazily to avoid an import cycle: ``model_registry`` imports
+    ``model_yaml`` which imports the dataclasses from this file.
+    """
+    if name == "ModelRegistry":
+        from application.core.model_registry import ModelRegistry as _MR
 
-    def __init__(self):
-        if not ModelRegistry._initialized:
-            self.models: Dict[str, AvailableModel] = {}
-            self.default_model_id: Optional[str] = None
-            self._load_models()
-            ModelRegistry._initialized = True
-
-    @classmethod
-    def get_instance(cls) -> "ModelRegistry":
-        return cls()
-
-    def _load_models(self):
-        from application.core.settings import settings
-
-        self.models.clear()
-
-        # Skip DocsGPT model if using custom OpenAI-compatible endpoint
-        if not settings.OPENAI_BASE_URL:
-            self._add_docsgpt_models(settings)
-        if (
-            settings.OPENAI_API_KEY
-            or (settings.LLM_PROVIDER == "openai" and settings.API_KEY)
-            or settings.OPENAI_BASE_URL
-        ):
-            self._add_openai_models(settings)
-        if settings.OPENAI_API_BASE or (
-            settings.LLM_PROVIDER == "azure_openai" and settings.API_KEY
-        ):
-            self._add_azure_openai_models(settings)
-        if settings.ANTHROPIC_API_KEY or (
-            settings.LLM_PROVIDER == "anthropic" and settings.API_KEY
-        ):
-            self._add_anthropic_models(settings)
-        if settings.GOOGLE_API_KEY or (
-            settings.LLM_PROVIDER == "google" and settings.API_KEY
-        ):
-            self._add_google_models(settings)
-        if settings.GROQ_API_KEY or (
-            settings.LLM_PROVIDER == "groq" and settings.API_KEY
-        ):
-            self._add_groq_models(settings)
-        if settings.OPEN_ROUTER_API_KEY or (
-            settings.LLM_PROVIDER == "openrouter" and settings.API_KEY
-        ):
-            self._add_openrouter_models(settings)
-        if settings.NOVITA_API_KEY or (
-            settings.LLM_PROVIDER == "novita" and settings.API_KEY
-        ):
-            self._add_novita_models(settings)
-        if settings.HUGGINGFACE_API_KEY or (
-            settings.LLM_PROVIDER == "huggingface" and settings.API_KEY
-        ):
-            self._add_huggingface_models(settings)
-        # Default model selection
-        if settings.LLM_NAME:
-            # Parse LLM_NAME (may be comma-separated)
-            model_names = self._parse_model_names(settings.LLM_NAME)
-            # First model in the list becomes default
-            for model_name in model_names:
-                if model_name in self.models:
-                    self.default_model_id = model_name
-                    break
-            # Backward compat: try exact match if no parsed model found
-            if not self.default_model_id and settings.LLM_NAME in self.models:
-                self.default_model_id = settings.LLM_NAME
-
-        if not self.default_model_id:
-            if settings.LLM_PROVIDER and settings.API_KEY:
-                for model_id, model in self.models.items():
-                    if model.provider.value == settings.LLM_PROVIDER:
-                        self.default_model_id = model_id
-                        break
-
-        if not self.default_model_id and self.models:
-            self.default_model_id = next(iter(self.models.keys()))
-        logger.info(
-            f"ModelRegistry loaded {len(self.models)} models, default: {self.default_model_id}"
-        )
-
-    def _add_openai_models(self, settings):
-        from application.core.model_configs import (
-            OPENAI_MODELS,
-            create_custom_openai_model,
-        )
-
-        # Check if using local OpenAI-compatible endpoint (Ollama, LM Studio, etc.)
-        using_local_endpoint = bool(
-            settings.OPENAI_BASE_URL and settings.OPENAI_BASE_URL.strip()
-        )
-
-        if using_local_endpoint:
-            # When OPENAI_BASE_URL is set, ONLY register custom models from LLM_NAME
-            # Do NOT add standard OpenAI models (gpt-5.1, etc.)
-            if settings.LLM_NAME:
-                model_names = self._parse_model_names(settings.LLM_NAME)
-                for model_name in model_names:
-                    custom_model = create_custom_openai_model(
-                        model_name, settings.OPENAI_BASE_URL
-                    )
-                    self.models[model_name] = custom_model
-                    logger.info(
-                        f"Registered custom OpenAI model: {model_name} at {settings.OPENAI_BASE_URL}"
-                    )
-        else:
-            # Standard OpenAI API usage - add standard models if API key is valid
-            if settings.OPENAI_API_KEY:
-                for model in OPENAI_MODELS:
-                    self.models[model.id] = model
-
-    def _add_azure_openai_models(self, settings):
-        from application.core.model_configs import AZURE_OPENAI_MODELS
-
-        if settings.LLM_PROVIDER == "azure_openai" and settings.LLM_NAME:
-            for model in AZURE_OPENAI_MODELS:
-                if model.id == settings.LLM_NAME:
-                    self.models[model.id] = model
-                    return
-        for model in AZURE_OPENAI_MODELS:
-            self.models[model.id] = model
-
-    def _add_anthropic_models(self, settings):
-        from application.core.model_configs import ANTHROPIC_MODELS
-
-        if settings.ANTHROPIC_API_KEY:
-            for model in ANTHROPIC_MODELS:
-                self.models[model.id] = model
-            return
-        if settings.LLM_PROVIDER == "anthropic" and settings.LLM_NAME:
-            for model in ANTHROPIC_MODELS:
-                if model.id == settings.LLM_NAME:
-                    self.models[model.id] = model
-                    return
-        for model in ANTHROPIC_MODELS:
-            self.models[model.id] = model
-
-    def _add_google_models(self, settings):
-        from application.core.model_configs import GOOGLE_MODELS
-
-        if settings.GOOGLE_API_KEY:
-            for model in GOOGLE_MODELS:
-                self.models[model.id] = model
-            return
-        if settings.LLM_PROVIDER == "google" and settings.LLM_NAME:
-            for model in GOOGLE_MODELS:
-                if model.id == settings.LLM_NAME:
-                    self.models[model.id] = model
-                    return
-        for model in GOOGLE_MODELS:
-            self.models[model.id] = model
-
-    def _add_groq_models(self, settings):
-        from application.core.model_configs import GROQ_MODELS
-
-        if settings.GROQ_API_KEY:
-            for model in GROQ_MODELS:
-                self.models[model.id] = model
-            return
-        if settings.LLM_PROVIDER == "groq" and settings.LLM_NAME:
-            for model in GROQ_MODELS:
-                if model.id == settings.LLM_NAME:
-                    self.models[model.id] = model
-                    return
-        for model in GROQ_MODELS:
-            self.models[model.id] = model
-    
-    def _add_openrouter_models(self, settings):
-        from application.core.model_configs import OPENROUTER_MODELS
-
-        if settings.OPEN_ROUTER_API_KEY:
-            for model in OPENROUTER_MODELS:
-                self.models[model.id] = model
-            return
-        if settings.LLM_PROVIDER == "openrouter" and settings.LLM_NAME:
-            for model in OPENROUTER_MODELS:
-                if model.id == settings.LLM_NAME:
-                    self.models[model.id] = model
-                    return
-        for model in OPENROUTER_MODELS:
-            self.models[model.id] = model
-
-    def _add_novita_models(self, settings):
-        from application.core.model_configs import NOVITA_MODELS
-
-        if settings.NOVITA_API_KEY:
-            for model in NOVITA_MODELS:
-                self.models[model.id] = model
-            return
-        if settings.LLM_PROVIDER == "novita" and settings.LLM_NAME:
-            for model in NOVITA_MODELS:
-                if model.id == settings.LLM_NAME:
-                    self.models[model.id] = model
-                    return
-        for model in NOVITA_MODELS:
-            self.models[model.id] = model
-
-    def _add_docsgpt_models(self, settings):
-        model_id = "docsgpt-local"
-        model = AvailableModel(
-            id=model_id,
-            provider=ModelProvider.DOCSGPT,
-            display_name="DocsGPT Model",
-            description="Local model",
-            capabilities=ModelCapabilities(
-                supports_tools=False,
-                supported_attachment_types=[],
-            ),
-        )
-        self.models[model_id] = model
-
-    def _add_huggingface_models(self, settings):
-        model_id = "huggingface-local"
-        model = AvailableModel(
-            id=model_id,
-            provider=ModelProvider.HUGGINGFACE,
-            display_name="Hugging Face Model",
-            description="Local Hugging Face model",
-            capabilities=ModelCapabilities(
-                supports_tools=False,
-                supported_attachment_types=[],
-            ),
-        )
-        self.models[model_id] = model
-
-    def _parse_model_names(self, llm_name: str) -> List[str]:
-        """
-        Parse LLM_NAME which may contain comma-separated model names.
-        E.g., 'deepseek-r1:1.5b,gemma:2b' -> ['deepseek-r1:1.5b', 'gemma:2b']
-        """
-        if not llm_name:
-            return []
-        return [name.strip() for name in llm_name.split(",") if name.strip()]
-
-    def get_model(self, model_id: str) -> Optional[AvailableModel]:
-        return self.models.get(model_id)
-
-    def get_all_models(self) -> List[AvailableModel]:
-        return list(self.models.values())
-
-    def get_enabled_models(self) -> List[AvailableModel]:
-        return [m for m in self.models.values() if m.enabled]
-
-    def model_exists(self, model_id: str) -> bool:
-        return model_id in self.models
+        return _MR
+    raise AttributeError(f"module {__name__!r} has no attribute {name!r}")

application/core/model_utils.py

@@ -1,47 +1,59 @@
 from typing import Any, Dict, Optional
 
-from application.core.model_settings import ModelRegistry
+from application.core.model_registry import ModelRegistry
 
 
 def get_api_key_for_provider(provider: str) -> Optional[str]:
-    """Get the appropriate API key for a provider"""
+    """Get the appropriate API key for a provider.
+
+    Delegates to the provider plugin's ``get_api_key``. Falls back to the
+    generic ``settings.API_KEY`` for unknown providers.
+    """
     from application.core.settings import settings
+    from application.llm.providers import PROVIDERS_BY_NAME
 
-    provider_key_map = {
-        "openai": settings.OPENAI_API_KEY,
-        "openrouter": settings.OPEN_ROUTER_API_KEY,
-        "novita": settings.NOVITA_API_KEY,
-        "anthropic": settings.ANTHROPIC_API_KEY,
-        "google": settings.GOOGLE_API_KEY,
-        "groq": settings.GROQ_API_KEY,
-        "huggingface": settings.HUGGINGFACE_API_KEY,
-        "azure_openai": settings.API_KEY,
-        "docsgpt": None,
-        "llama.cpp": None,
-    }
-
-    provider_key = provider_key_map.get(provider)
-    if provider_key:
-        return provider_key
+    plugin = PROVIDERS_BY_NAME.get(provider)
+    if plugin is not None:
+        key = plugin.get_api_key(settings)
+        if key:
+            return key
     return settings.API_KEY
 
 
-def get_all_available_models() -> Dict[str, Dict[str, Any]]:
-    """Get all available models with metadata for API response"""
+def get_all_available_models(
+    user_id: Optional[str] = None,
+) -> Dict[str, Dict[str, Any]]:
+    """Get all available models with metadata for API response.
+
+    When ``user_id`` is supplied, the user's BYOM custom-model records
+    are merged into the result alongside the built-in catalog.
+    """
     registry = ModelRegistry.get_instance()
-    return {model.id: model.to_dict() for model in registry.get_enabled_models()}
+    return {
+        model.id: model.to_dict()
+        for model in registry.get_enabled_models(user_id=user_id)
+    }
 
 
-def validate_model_id(model_id: str) -> bool:
-    """Check if a model ID exists in registry"""
+def validate_model_id(model_id: str, user_id: Optional[str] = None) -> bool:
+    """Check if a model ID exists in registry.
+
+    ``user_id`` enables resolution of per-user BYOM records (UUIDs).
+    Without it, only built-in catalog ids resolve.
+    """
     registry = ModelRegistry.get_instance()
-    return registry.model_exists(model_id)
+    return registry.model_exists(model_id, user_id=user_id)
 
 
-def get_model_capabilities(model_id: str) -> Optional[Dict[str, Any]]:
-    """Get capabilities for a specific model"""
+def get_model_capabilities(
+    model_id: str, user_id: Optional[str] = None
+) -> Optional[Dict[str, Any]]:
+    """Get capabilities for a specific model.
+
+    ``user_id`` enables resolution of per-user BYOM records.
+    """
     registry = ModelRegistry.get_instance()
-    model = registry.get_model(model_id)
+    model = registry.get_model(model_id, user_id=user_id)
     if model:
         return {
             "supported_attachment_types": model.capabilities.supported_attachment_types,
@@ -58,36 +70,68 @@ def get_default_model_id() -> str:
     return registry.default_model_id
 
 
-def get_provider_from_model_id(model_id: str) -> Optional[str]:
-    """Get the provider name for a given model_id"""
+def get_provider_from_model_id(
+    model_id: str, user_id: Optional[str] = None
+) -> Optional[str]:
+    """Get the provider name for a given model_id.
+
+    ``user_id`` enables resolution of per-user BYOM records (UUIDs).
+    Without it, BYOM model ids return ``None`` and the caller falls
+    back to the deployment default.
+    """
     registry = ModelRegistry.get_instance()
-    model = registry.get_model(model_id)
+    model = registry.get_model(model_id, user_id=user_id)
     if model:
         return model.provider.value
     return None
 
 
-def get_token_limit(model_id: str) -> int:
-    """
-    Get context window (token limit) for a model.
-    Returns model's context_window or default 128000 if model not found.
+def get_token_limit(model_id: str, user_id: Optional[str] = None) -> int:
+    """Get context window (token limit) for a model.
+
+    Returns the model's ``context_window`` or ``DEFAULT_LLM_TOKEN_LIMIT``
+    if not found. ``user_id`` enables resolution of per-user BYOM records.
     """
     from application.core.settings import settings
 
     registry = ModelRegistry.get_instance()
-    model = registry.get_model(model_id)
+    model = registry.get_model(model_id, user_id=user_id)
     if model:
         return model.capabilities.context_window
     return settings.DEFAULT_LLM_TOKEN_LIMIT
 
 
-def get_base_url_for_model(model_id: str) -> Optional[str]:
-    """
-    Get the custom base_url for a specific model if configured.
-    Returns None if no custom base_url is set.
+def get_base_url_for_model(
+    model_id: str, user_id: Optional[str] = None
+) -> Optional[str]:
+    """Get the custom base_url for a specific model if configured.
+
+    Returns ``None`` if no custom base_url is set. ``user_id`` enables
+    resolution of per-user BYOM records.
     """
     registry = ModelRegistry.get_instance()
-    model = registry.get_model(model_id)
+    model = registry.get_model(model_id, user_id=user_id)
     if model:
         return model.base_url
     return None
+
+
+def get_api_key_for_model(
+    model_id: str, user_id: Optional[str] = None
+) -> Optional[str]:
+    """Resolve the API key to use when invoking ``model_id``.
+
+    Priority:
+      1. The model record's own ``api_key`` (BYOM records and
+         ``openai_compatible`` YAMLs populate this).
+      2. The provider plugin's settings-based key.
+
+    ``user_id`` enables resolution of per-user BYOM records.
+    """
+    registry = ModelRegistry.get_instance()
+    model = registry.get_model(model_id, user_id=user_id)
+    if model is not None and model.api_key:
+        return model.api_key
+    if model is not None:
+        return get_api_key_for_provider(model.provider.value)
+    return None

application/core/model_yaml.py

@@ -0,0 +1,358 @@
+"""YAML loader for model catalog files under ``application/core/models/``.
+
+Each ``*.yaml`` file declares one provider's static model catalog. Files
+are validated with Pydantic at load time; any parse, schema, or alias
+error aborts startup with the offending file path in the message.
+
+For most providers, one YAML maps to one catalog. The
+``openai_compatible`` provider is special: each YAML file represents a
+distinct logical endpoint (Mistral, Together, Ollama, ...) with its own
+``api_key_env`` and ``base_url``. The loader returns a flat list so the
+registry can distinguish multiple files with the same ``provider:`` value.
+"""
+
+from __future__ import annotations
+
+import logging
+from pathlib import Path
+from typing import Dict, List, Optional, Sequence
+
+import yaml
+from pydantic import BaseModel, ConfigDict, Field, field_validator
+
+from application.core.model_settings import (
+    AvailableModel,
+    ModelCapabilities,
+    ModelProvider,
+)
+
+logger = logging.getLogger(__name__)
+
+BUILTIN_MODELS_DIR = Path(__file__).parent / "models"
+DEFAULTS_FILENAME = "_defaults.yaml"
+
+
+class _DefaultsFile(BaseModel):
+    """Schema for ``_defaults.yaml``. Currently just attachment aliases."""
+
+    model_config = ConfigDict(extra="forbid")
+
+    attachment_aliases: Dict[str, List[str]] = Field(default_factory=dict)
+
+
+class _CapabilityFields(BaseModel):
+    """Capability fields shared between provider ``defaults:`` and per-model overrides.
+
+    All fields are optional so a per-model override can selectively replace
+    a single field from the provider-level defaults.
+    """
+
+    model_config = ConfigDict(extra="forbid")
+
+    supports_tools: Optional[bool] = None
+    supports_structured_output: Optional[bool] = None
+    supports_streaming: Optional[bool] = None
+    attachments: Optional[List[str]] = None
+    context_window: Optional[int] = None
+    input_cost_per_token: Optional[float] = None
+    output_cost_per_token: Optional[float] = None
+
+
+class _ModelEntry(_CapabilityFields):
+    """Schema for one model row inside a YAML's ``models:`` list."""
+
+    id: str
+    display_name: Optional[str] = None
+    description: str = ""
+    enabled: bool = True
+    base_url: Optional[str] = None
+    aliases: List[str] = Field(default_factory=list)
+
+    @field_validator("id")
+    @classmethod
+    def _id_nonempty(cls, v: str) -> str:
+        if not v or not v.strip():
+            raise ValueError("model id must be a non-empty string")
+        return v
+
+
+class _ProviderFile(BaseModel):
+    """Schema for one ``<provider>.yaml`` catalog file."""
+
+    model_config = ConfigDict(extra="forbid")
+
+    provider: str
+    defaults: _CapabilityFields = Field(default_factory=_CapabilityFields)
+    models: List[_ModelEntry] = Field(default_factory=list)
+    # openai_compatible metadata. Optional for other providers.
+    display_provider: Optional[str] = None
+    api_key_env: Optional[str] = None
+    base_url: Optional[str] = None
+
+
+class ProviderCatalog(BaseModel):
+    """One YAML file's parsed contents, ready for the registry.
+
+    For most providers, multiple catalogs with the same ``provider`` get
+    merged later by the registry. The ``openai_compatible`` provider is
+    the exception: each catalog is treated as a distinct endpoint, with
+    its own ``api_key_env`` and ``base_url``.
+    """
+
+    provider: str
+    models: List[AvailableModel]
+    source_path: Optional[Path] = None
+    display_provider: Optional[str] = None
+    api_key_env: Optional[str] = None
+    base_url: Optional[str] = None
+
+    model_config = ConfigDict(arbitrary_types_allowed=True)
+
+
+class ModelYAMLError(ValueError):
+    """Raised when a model YAML fails parsing, schema, or alias validation."""
+
+
+def _expand_attachments(
+    attachments: Sequence[str], aliases: Dict[str, List[str]], source: str
+) -> List[str]:
+    """Resolve attachment shorthands (``image``, ``pdf``) to MIME types.
+
+    Raw MIME-typed entries (containing ``/``) pass through unchanged.
+    Unknown aliases raise ``ModelYAMLError``.
+    """
+    expanded: List[str] = []
+    seen: set = set()
+    for entry in attachments:
+        if "/" in entry:
+            if entry not in seen:
+                expanded.append(entry)
+                seen.add(entry)
+            continue
+        if entry not in aliases:
+            valid = ", ".join(sorted(aliases.keys())) or "<none defined>"
+            raise ModelYAMLError(
+                f"{source}: unknown attachment alias '{entry}'. "
+                f"Valid aliases: {valid}. "
+                "(Or use a raw MIME type like 'image/png'.)"
+            )
+        for mime in aliases[entry]:
+            if mime not in seen:
+                expanded.append(mime)
+                seen.add(mime)
+    return expanded
+
+
+def _load_defaults(directory: Path) -> Dict[str, List[str]]:
+    """Load ``_defaults.yaml`` from ``directory`` if it exists."""
+    path = directory / DEFAULTS_FILENAME
+    if not path.exists():
+        return {}
+    try:
+        raw = yaml.safe_load(path.read_text(encoding="utf-8")) or {}
+    except yaml.YAMLError as e:
+        raise ModelYAMLError(f"{path}: invalid YAML: {e}") from e
+    try:
+        parsed = _DefaultsFile.model_validate(raw)
+    except Exception as e:
+        raise ModelYAMLError(f"{path}: schema error: {e}") from e
+    return parsed.attachment_aliases
+
+
+def _resolve_provider_enum(name: str, source: Path) -> ModelProvider:
+    try:
+        return ModelProvider(name)
+    except ValueError as e:
+        valid = ", ".join(p.value for p in ModelProvider)
+        raise ModelYAMLError(
+            f"{source}: unknown provider '{name}'. Valid: {valid}"
+        ) from e
+
+
+def _build_model(
+    entry: _ModelEntry,
+    defaults: _CapabilityFields,
+    provider: ModelProvider,
+    aliases: Dict[str, List[str]],
+    source: Path,
+    display_provider: Optional[str] = None,
+) -> AvailableModel:
+    """Merge defaults + per-model overrides into a final ``AvailableModel``."""
+
+    def pick(field_name: str, fallback):
+        v = getattr(entry, field_name)
+        if v is not None:
+            return v
+        d = getattr(defaults, field_name)
+        if d is not None:
+            return d
+        return fallback
+
+    raw_attachments = entry.attachments
+    if raw_attachments is None:
+        raw_attachments = defaults.attachments
+    if raw_attachments is None:
+        raw_attachments = []
+    expanded = _expand_attachments(
+        raw_attachments, aliases, f"{source} [model={entry.id}]"
+    )
+
+    caps = ModelCapabilities(
+        supports_tools=pick("supports_tools", False),
+        supports_structured_output=pick("supports_structured_output", False),
+        supports_streaming=pick("supports_streaming", True),
+        supported_attachment_types=expanded,
+        context_window=pick("context_window", 128000),
+        input_cost_per_token=pick("input_cost_per_token", None),
+        output_cost_per_token=pick("output_cost_per_token", None),
+    )
+
+    return AvailableModel(
+        id=entry.id,
+        provider=provider,
+        display_name=entry.display_name or entry.id,
+        description=entry.description,
+        capabilities=caps,
+        enabled=entry.enabled,
+        base_url=entry.base_url,
+        display_provider=display_provider,
+    )
+
+
+def _load_one_yaml(
+    path: Path, aliases: Dict[str, List[str]]
+) -> ProviderCatalog:
+    try:
+        raw = yaml.safe_load(path.read_text(encoding="utf-8")) or {}
+    except yaml.YAMLError as e:
+        raise ModelYAMLError(f"{path}: invalid YAML: {e}") from e
+    try:
+        parsed = _ProviderFile.model_validate(raw)
+    except Exception as e:
+        raise ModelYAMLError(f"{path}: schema error: {e}") from e
+
+    provider_enum = _resolve_provider_enum(parsed.provider, path)
+    models = [
+        _build_model(
+            entry,
+            parsed.defaults,
+            provider_enum,
+            aliases,
+            path,
+            display_provider=parsed.display_provider,
+        )
+        for entry in parsed.models
+    ]
+
+    return ProviderCatalog(
+        provider=parsed.provider,
+        models=models,
+        source_path=path,
+        display_provider=parsed.display_provider,
+        api_key_env=parsed.api_key_env,
+        base_url=parsed.base_url,
+    )
+
+
+_BUILTIN_ALIASES_CACHE: Optional[Dict[str, List[str]]] = None
+
+
+def builtin_attachment_aliases() -> Dict[str, List[str]]:
+    """Return the built-in attachment alias map from ``_defaults.yaml``.
+
+    Cached after first read so repeat calls are cheap.
+    """
+    global _BUILTIN_ALIASES_CACHE
+    if _BUILTIN_ALIASES_CACHE is None:
+        _BUILTIN_ALIASES_CACHE = _load_defaults(BUILTIN_MODELS_DIR)
+    return _BUILTIN_ALIASES_CACHE
+
+
+def resolve_attachment_alias(alias: str) -> List[str]:
+    """Resolve a single attachment alias (e.g. ``"image"``) to its
+    canonical MIME-type list. Raises ``ModelYAMLError`` if unknown.
+    """
+    aliases = builtin_attachment_aliases()
+    if alias not in aliases:
+        valid = ", ".join(sorted(aliases.keys())) or "<none defined>"
+        raise ModelYAMLError(
+            f"Unknown attachment alias '{alias}'. Valid: {valid}"
+        )
+    return list(aliases[alias])
+
+
+def expand_attachments_lenient(
+    attachments: Sequence[str], source: str
+) -> List[str]:
+    """Expand attachment aliases to MIME types, tolerating unknowns.
+
+    Mirrors ``_expand_attachments`` but logs+skips unknown aliases
+    rather than raising. Used for runtime call sites (BYOM registry
+    load) where an operator-side alias-map edit must not drop the
+    entire user's BYOM layer; the strict raise still happens at the
+    API validation boundary.
+    """
+    aliases = builtin_attachment_aliases()
+    expanded: List[str] = []
+    seen: set = set()
+    for entry in attachments:
+        if "/" in entry:
+            if entry not in seen:
+                expanded.append(entry)
+                seen.add(entry)
+            continue
+        mime_list = aliases.get(entry)
+        if mime_list is None:
+            logger.warning(
+                "%s: skipping unknown attachment alias %r", source, entry,
+            )
+            continue
+        for mime in mime_list:
+            if mime not in seen:
+                expanded.append(mime)
+                seen.add(mime)
+    return expanded
+
+
+def load_model_yamls(directories: Sequence[Path]) -> List[ProviderCatalog]:
+    """Load every ``*.yaml`` file (excluding ``_defaults.yaml``) under each
+    directory in order and return a flat list of catalogs.
+
+    Caller is responsible for merging multiple catalogs that target the
+    same provider plugin. The flat-list shape lets ``openai_compatible``
+    keep each file separate (one logical endpoint per file).
+
+    When the same model ``id`` appears in more than one YAML across the
+    directory list, a warning is logged. Order in the returned list
+    preserves load order, so the registry's "later wins" merge gives the
+    later directory's definition.
+    """
+    catalogs: List[ProviderCatalog] = []
+    seen_ids: Dict[str, Path] = {}
+
+    aliases: Dict[str, List[str]] = {}
+    for d in directories:
+        if not d or not d.exists():
+            continue
+        aliases.update(_load_defaults(d))
+
+    for d in directories:
+        if not d or not d.exists():
+            continue
+        for path in sorted(d.glob("*.yaml")):
+            if path.name == DEFAULTS_FILENAME:
+                continue
+            catalog = _load_one_yaml(path, aliases)
+            catalogs.append(catalog)
+            for m in catalog.models:
+                prior = seen_ids.get(m.id)
+                if prior is not None and prior != path:
+                    logger.warning(
+                        "Model id %r redefined: %s overrides %s (later wins)",
+                        m.id,
+                        path,
+                        prior,
+                    )
+                seen_ids[m.id] = path
+
+    return catalogs

application/core/models/README.md

@@ -0,0 +1,213 @@
+# Model catalogs
+
+Each `*.yaml` file in this directory declares one provider's model
+catalog. The registry loads every YAML at boot and joins it to the
+matching provider plugin under `application/llm/providers/`.
+
+To add or edit models, you almost always only touch a YAML here — no
+Python code required.
+
+## Add a model to an existing provider
+
+Open the provider's YAML (e.g. `anthropic.yaml`) and append two lines
+under `models:`:
+
+```yaml
+models:
+  - id: claude-3-7-sonnet
+    display_name: Claude 3.7 Sonnet
+```
+
+Capabilities default to the provider's `defaults:` block. Override
+per-model only when needed:
+
+```yaml
+  - id: claude-3-7-sonnet
+    display_name: Claude 3.7 Sonnet
+    context_window: 500000
+```
+
+Restart the app. The new model appears in `/api/models`.
+
+> The model `id` is what gets stored in agent / workflow records. Once
+> users start picking the model, **don't rename it** — agent and
+> workflow rows reference it as a free-form string and silently fall
+> back to the system default if the id disappears.
+
+## Add an OpenAI-compatible provider (zero Python)
+
+Drop a YAML in this directory (or in your `MODELS_CONFIG_DIR`) that uses
+the `openai_compatible` plugin. Set the env var named in `api_key_env`
+and you're done — no Python, no settings.py edit, no LLMCreator change:
+
+```yaml
+# mistral.yaml
+provider: openai_compatible
+display_provider: mistral             # shown in /api/models response
+api_key_env: MISTRAL_API_KEY          # env var the plugin reads at boot
+base_url: https://api.mistral.ai/v1
+defaults:
+  supports_tools: true
+  context_window: 128000
+models:
+  - id: mistral-large-latest
+    display_name: Mistral Large
+  - id: mistral-small-latest
+    display_name: Mistral Small
+```
+
+`MISTRAL_API_KEY=sk-... ; restart` — Mistral models appear in
+`/api/models` with `provider: "mistral"`. They route through the OpenAI
+wire format (it's `OpenAILLM` under the hood) but with Mistral's
+endpoint and key.
+
+Multiple `openai_compatible` YAMLs coexist: each file is one logical
+endpoint with its own `api_key_env` and `base_url`. Drop in
+`together.yaml`, `fireworks.yaml`, etc. side by side. If an env var
+isn't set, that catalog is silently skipped at boot (logged at INFO) —
+no error.
+
+Working example: `examples/mistral.yaml.example`. Files inside
+`examples/` aren't loaded by the registry; the glob only picks up
+`*.yaml` at the top level.
+
+## Add a provider with its own SDK
+
+For a provider that doesn't speak OpenAI's wire format, add one Python
+file to `application/llm/providers/<name>.py`:
+
+```python
+from application.llm.providers.base import Provider
+from application.llm.my_provider import MyLLM
+
+class MyProvider(Provider):
+    name = "my_provider"
+    llm_class = MyLLM
+
+    def get_api_key(self, settings):
+        return settings.MY_PROVIDER_API_KEY
+```
+
+Register it in `application/llm/providers/__init__.py` (one line in
+`ALL_PROVIDERS`), add `MY_PROVIDER_API_KEY` to `settings.py`, and create
+`my_provider.yaml` here with the model catalog.
+
+## Schema reference
+
+```yaml
+provider: <string, required>          # matches the Provider plugin's `name`
+
+# openai_compatible only — required for that provider, ignored for others
+display_provider: <string>            # label shown in /api/models response
+api_key_env: <string>                 # name of the env var carrying the key
+base_url: <string>                    # endpoint URL
+
+defaults:                              # optional, applied to every model below
+  supports_tools: bool                 # default false
+  supports_structured_output: bool     # default false
+  supports_streaming: bool             # default true
+  attachments: [<alias-or-mime>, ...]  # default []
+  context_window: int                  # default 128000
+  input_cost_per_token: float          # default null
+  output_cost_per_token: float         # default null
+
+models:                                # required
+  - id: <string, required>             # the value persisted in agent records
+    display_name: <string>             # default: id
+    description: <string>              # default: ""
+    enabled: bool                      # default true; false hides from /api/models
+    base_url: <string>                 # optional custom endpoint for this model
+    # All `defaults:` fields above can be overridden here per-model.
+```
+
+### Attachment aliases
+
+The `attachments:` list can mix human-readable aliases with raw MIME
+types. Aliases are defined in `_defaults.yaml`:
+
+| Alias | Expands to |
+|---|---|
+| `image` | `image/png`, `image/jpeg`, `image/jpg`, `image/webp`, `image/gif` |
+| `pdf` | `application/pdf` |
+| `audio` | `audio/mpeg`, `audio/wav`, `audio/ogg` |
+
+Use raw MIME types when you need surgical control:
+
+```yaml
+attachments: [image/png, image/webp]   # only these two
+```
+
+## Operator-supplied YAMLs (`MODELS_CONFIG_DIR`)
+
+Set the `MODELS_CONFIG_DIR` env var (or `.env` entry) to a directory
+path. Every `*.yaml` in that directory is loaded **after** the built-in
+catalog under `application/core/models/`. Operators use this to:
+
+- Add new `openai_compatible` providers (Mistral, Together, Fireworks,
+  Ollama, ...) without forking the repo.
+- Extend an existing provider's catalog with extra models — append
+  models under `provider: anthropic` and they show up alongside the
+  built-ins.
+- Override a built-in model's capabilities — declare the same `id`
+  with different fields (e.g. a higher `context_window`). Later wins;
+  the override is logged as a `WARNING` so you can audit it.
+
+Things you cannot do via `MODELS_CONFIG_DIR`:
+
+- Add a brand-new non-OpenAI provider — that needs a Python plugin
+  under `application/llm/providers/` (see "Add a provider with its own
+  SDK" above). Operator YAMLs may only target a `provider:` value that
+  already has a registered plugin.
+
+### Example: Docker
+
+Mount your model YAMLs into the container and point the env var at the
+mount path:
+
+```yaml
+# docker-compose.yml
+services:
+  app:
+    image: arc53/docsgpt
+    environment:
+      MODELS_CONFIG_DIR: /etc/docsgpt/models
+      MISTRAL_API_KEY: ${MISTRAL_API_KEY}
+    volumes:
+      - ./my-models:/etc/docsgpt/models:ro
+```
+
+Then `./my-models/mistral.yaml` (the file from
+`examples/mistral.yaml.example`) gets picked up at boot.
+
+### Example: Kubernetes
+
+Mount a `ConfigMap` containing your YAMLs at a known path and set
+`MODELS_CONFIG_DIR` on the deployment. The same `examples/mistral.yaml.example`
+becomes a key in the ConfigMap.
+
+### Misconfiguration
+
+If `MODELS_CONFIG_DIR` is set but the path doesn't exist (or isn't a
+directory), the app logs a `WARNING` at boot and continues with just
+the built-in catalog. The app does *not* fail to start — operators can
+ship config drift without taking down the service — but the warning is
+loud enough to surface in any reasonable log aggregator.
+
+## Validation
+
+YAMLs are parsed with Pydantic at boot. The app fails to start with a
+clear error message if:
+
+- a top-level key is unknown
+- a model is missing `id`
+- an attachment alias isn't defined
+- the `provider:` value isn't registered as a plugin
+
+This is intentional — silent fallbacks would mean users don't notice
+their model picks broke until they hit the API.
+
+## Reserved fields (not yet implemented)
+
+- `aliases:` on a model — old IDs that resolve to this model. Reserved
+  for future renames; the schema accepts the field but it is not yet
+  acted on.

application/core/models/_defaults.yaml

@@ -0,0 +1,18 @@
+# Global defaults applied across every model YAML in this directory.
+# Keep this file sparse — per-provider `defaults:` blocks are clearer
+# than a deep global default chain. This file is for things that
+# genuinely never vary, like the meaning of "image".
+
+attachment_aliases:
+  image:
+    - image/png
+    - image/jpeg
+    - image/jpg
+    - image/webp
+    - image/gif
+  pdf:
+    - application/pdf
+  audio:
+    - audio/mpeg
+    - audio/wav
+    - audio/ogg

application/core/models/anthropic.yaml

@@ -0,0 +1,23 @@
+provider: anthropic
+defaults:
+  supports_tools: true
+  attachments: [image]
+  context_window: 200000
+
+models:
+  - id: claude-opus-4-7
+    display_name: Claude Opus 4.7
+    description: Most capable Claude model for complex reasoning and agentic coding
+    context_window: 1000000
+    supports_structured_output: true
+
+  - id: claude-sonnet-4-6
+    display_name: Claude Sonnet 4.6
+    description: Best balance of speed and intelligence with extended thinking
+    context_window: 1000000
+    supports_structured_output: true
+
+  - id: claude-haiku-4-5
+    display_name: Claude Haiku 4.5
+    description: Fastest Claude model with near-frontier intelligence
+    supports_structured_output: true

application/core/models/azure_openai.yaml

@@ -0,0 +1,31 @@
+# Azure OpenAI catalog.
+#
+# IMPORTANT: For Azure OpenAI, the `id` field is the **deployment name**, not
+# a model name. Deployment names are arbitrary strings the operator chooses
+# in Azure portal (or via ARM/Bicep/Terraform) when they create a deployment
+# for a given underlying model + version.
+#
+# The IDs below are sensible defaults that mirror the underlying OpenAI
+# model name (prefixed with `azure-`). Operators almost always need to
+# override them via `MODELS_CONFIG_DIR` to match the deployment names that
+# actually exist in their Azure resource. The `display_name`, capability
+# flags, and `context_window` reflect the underlying OpenAI model.
+provider: azure_openai
+
+defaults:
+  supports_tools: true
+  supports_structured_output: true
+  attachments: [image]
+  context_window: 400000
+
+models:
+  - id: azure-gpt-5.5
+    display_name: Azure OpenAI GPT-5.5
+    description: Azure-hosted flagship frontier model for complex reasoning, coding, and agentic work with a 1M-token context window
+    context_window: 1050000
+  - id: azure-gpt-5.4-mini
+    display_name: Azure OpenAI GPT-5.4 Mini
+    description: Azure-hosted cost-efficient GPT-5.4-class model for high-volume coding, computer use, and subagent workloads
+  - id: azure-gpt-5.4-nano
+    display_name: Azure OpenAI GPT-5.4 Nano
+    description: Azure-hosted cheapest GPT-5.4-class model, optimized for simple high-volume tasks where speed and cost matter most

application/core/models/docsgpt.yaml

@@ -0,0 +1,7 @@
+provider: docsgpt
+
+models:
+  - id: docsgpt-local
+    display_name: DocsGPT Model
+    description: Local model
+    supports_tools: false

application/core/models/examples/mistral.yaml.example

@@ -0,0 +1,31 @@
+# EXAMPLE — copy this file to ../mistral.yaml (or to your
+# MODELS_CONFIG_DIR) and set MISTRAL_API_KEY in your environment.
+#
+# This is the entire integration. No Python required: the
+# `openai_compatible` plugin reads `api_key_env` and `base_url` from
+# the file and routes calls through the OpenAI wire format.
+#
+# Files in this `examples/` directory are NOT loaded by the registry
+# (the loader globs *.yaml at the top level only).
+
+provider: openai_compatible
+display_provider: mistral             # shown in /api/models response
+api_key_env: MISTRAL_API_KEY          # env var the plugin reads
+base_url: https://api.mistral.ai/v1   # OpenAI-compatible endpoint
+
+defaults:
+  supports_tools: true
+  context_window: 128000
+
+models:
+  - id: mistral-large-latest
+    display_name: Mistral Large
+    description: Top-tier reasoning model
+
+  - id: mistral-small-latest
+    display_name: Mistral Small
+    description: Fast, cost-efficient
+
+  - id: codestral-latest
+    display_name: Codestral
+    description: Code-specialized model

application/core/models/google.yaml

@@ -0,0 +1,17 @@
+provider: google
+defaults:
+  supports_tools: true
+  supports_structured_output: true
+  attachments: [pdf, image]
+  context_window: 1048576
+
+models:
+  - id: gemini-3.1-pro-preview
+    display_name: Gemini 3.1 Pro
+    description: Most capable Gemini 3 model with advanced reasoning and agentic coding (preview)
+  - id: gemini-3-flash-preview
+    display_name: Gemini 3 Flash
+    description: Frontier-class performance for low-latency, high-volume tasks (preview)
+  - id: gemini-3.1-flash-lite-preview
+    display_name: Gemini 3.1 Flash-Lite
+    description: Cost-efficient frontier-class multimodal model for high-throughput workloads (preview)

application/core/models/groq.yaml

@@ -0,0 +1,16 @@
+provider: groq
+defaults:
+  supports_tools: true
+  context_window: 131072
+
+models:
+  - id: openai/gpt-oss-120b
+    display_name: GPT-OSS 120B
+    description: OpenAI's open-weight 120B flagship served on Groq's LPU hardware; strong general reasoning with strict structured output support
+    supports_structured_output: true
+  - id: llama-3.3-70b-versatile
+    display_name: Llama 3.3 70B Versatile
+    description: Meta's Llama 3.3 70B for general-purpose chat with parallel tool use
+  - id: llama-3.1-8b-instant
+    display_name: Llama 3.1 8B Instant
+    description: Small, very low-latency Llama model (~560 tok/s) with parallel tool use

application/core/models/huggingface.yaml

@@ -0,0 +1,7 @@
+provider: huggingface
+
+models:
+  - id: huggingface-local
+    display_name: Hugging Face Model
+    description: Local Hugging Face model
+    supports_tools: false

application/core/models/novita.yaml

@@ -0,0 +1,21 @@
+provider: novita
+defaults:
+  supports_tools: true
+  supports_structured_output: true
+
+models:
+  - id: deepseek/deepseek-v4-pro
+    display_name: DeepSeek V4 Pro
+    description: 1.6T MoE (49B active) with 1M context, hybrid CSA/HCA attention, top-tier reasoning and agentic coding
+    context_window: 1048576
+
+  - id: moonshotai/kimi-k2.6
+    display_name: Kimi K2.6
+    description: 1T-parameter open-weight MoE with native vision/video, multi-step tool calling, and agentic long-horizon execution
+    attachments: [image]
+    context_window: 262144
+
+  - id: zai-org/glm-5
+    display_name: GLM-5
+    description: Z.AI 754B-parameter MoE with strong general reasoning, function calling, and structured output
+    context_window: 202800

application/core/models/openai.yaml

@@ -0,0 +1,18 @@
+provider: openai
+defaults:
+  supports_tools: true
+  supports_structured_output: true
+  attachments: [image]
+  context_window: 400000
+
+models:
+  - id: gpt-5.5
+    display_name: GPT-5.5
+    description: Flagship frontier model for complex reasoning, coding, and agentic work with a 1M-token context window
+    context_window: 1050000
+  - id: gpt-5.4-mini
+    display_name: GPT-5.4 Mini
+    description: Cost-efficient GPT-5.4-class model for high-volume coding, computer use, and subagent workloads
+  - id: gpt-5.4-nano
+    display_name: GPT-5.4 Nano
+    description: Cheapest GPT-5.4-class model, optimized for simple high-volume tasks where speed and cost matter most

application/core/models/openrouter.yaml

@@ -0,0 +1,25 @@
+provider: openrouter
+defaults:
+  supports_tools: true
+  attachments: [image]
+  context_window: 128000
+
+models:
+  - id: qwen/qwen3-coder:free
+    display_name: Qwen3 Coder (free)
+    description: Free-tier 480B MoE coder model with strong agentic tool use; rate-limited
+    context_window: 262000
+    attachments: []
+
+  - id: deepseek/deepseek-v3.2
+    display_name: DeepSeek V3.2
+    description: Open-weights reasoning model, very low cost (~$0.25 in / $0.38 out per 1M)
+    context_window: 131072
+    attachments: []
+    supports_structured_output: true
+
+  - id: anthropic/claude-sonnet-4.6
+    display_name: Claude Sonnet 4.6 (via OpenRouter)
+    description: Frontier Sonnet-class model with 1M context, vision, and extended thinking
+    context_window: 1000000
+    supports_structured_output: true

application/core/mongo_db.py

@@ -1,24 +0,0 @@
-from application.core.settings import settings
-from pymongo import MongoClient
-
-
-class MongoDB:
-    _client = None
-
-    @classmethod
-    def get_client(cls):
-        """
-        Get the MongoDB client instance, creating it if necessary.
-        """
-        if cls._client is None:
-            cls._client = MongoClient(settings.MONGO_URI)
-        return cls._client
-
-    @classmethod
-    def close_client(cls):
-        """
-        Close the MongoDB client connection.
-        """
-        if cls._client is not None:
-            cls._client.close()
-            cls._client = None

application/core/settings.py

@@ -23,16 +23,21 @@ class Settings(BaseSettings):
     EMBEDDINGS_NAME: str = "huggingface_sentence-transformers/all-mpnet-base-v2"
     EMBEDDINGS_BASE_URL: Optional[str] = None  # Remote embeddings API URL (OpenAI-compatible)
     EMBEDDINGS_KEY: Optional[str] = None  # api key for embeddings (if using openai, just copy API_KEY)
+    # Optional directory of operator-supplied model YAMLs, loaded after the
+    # built-in catalog under application/core/models/. Later wins on
+    # duplicate model id. See application/core/models/README.md.
+    MODELS_CONFIG_DIR: Optional[str] = None
 
     CELERY_BROKER_URL: str = "redis://localhost:6379/0"
     CELERY_RESULT_BACKEND: str = "redis://localhost:6379/1"
-    MONGO_URI: str = "mongodb://localhost:27017/docsgpt"
-    MONGO_DB_NAME: str = "docsgpt"
+    # Only consulted when VECTOR_STORE=mongodb or when running scripts/db/backfill.py; user data lives in Postgres.
+    MONGO_URI: Optional[str] = None
     # User-data Postgres DB.
     POSTGRES_URI: Optional[str] = None
-
-    # MongoDB→Postgres migration: dual-write to Postgres (Mongo stays source of truth)
-    USE_POSTGRES: bool = False
+    # On app startup, apply pending Alembic migrations. Default ON for dev; disable in prod if you manage schema out-of-band.
+    AUTO_MIGRATE: bool = True
+    # On app startup, create the target Postgres database if it's missing (requires CREATEDB privilege). Dev-friendly default.
+    AUTO_CREATE_DB: bool = True
     LLM_PATH: str = os.path.join(current_dir, "models/docsgpt-7b-f16.gguf")
     DEFAULT_MAX_HISTORY: int = 150
     DEFAULT_LLM_TOKEN_LIMIT: int = 128000  # Fallback when model not found in registry
@@ -148,6 +153,9 @@ class Settings(BaseSettings):
 
     FLASK_DEBUG_MODE: bool = False
     STORAGE_TYPE: str = "local"  # local or s3
+
+    # Anonymous startup version check for security issues.
+    VERSION_CHECK: bool = True
     URL_STRATEGY: str = "backend"  # backend or s3
 
     JWT_SECRET_KEY: str = ""

application/gunicorn_conf.py

@@ -0,0 +1,72 @@
+"""Gunicorn config — keeps uvicorn's access log in NCSA format."""
+
+from __future__ import annotations
+
+import logging
+import logging.config
+
+# NCSA common log format:
+#   %(h)s %(l)s %(u)s %(t)s "%(r)s" %(s)s %(b)s "%(f)s" "%(a)s"
+# Uvicorn's access formatter exposes a ``client_addr``/``request_line``/
+# ``status_code`` trio but not the full NCSA field set, so we re-derive
+# what we can.
+_NCSA_FMT = (
+    '%(client_addr)s - - [%(asctime)s] "%(request_line)s" %(status_code)s'
+)
+
+logconfig_dict = {
+    "version": 1,
+    "disable_existing_loggers": False,
+    "formatters": {
+        "ncsa_access": {
+            "()": "uvicorn.logging.AccessFormatter",
+            "fmt": _NCSA_FMT,
+            "datefmt": "%d/%b/%Y:%H:%M:%S %z",
+            "use_colors": False,
+        },
+        "default": {
+            "format": "[%(asctime)s] [%(process)d] [%(levelname)s] %(name)s: %(message)s",
+        },
+    },
+    "handlers": {
+        "access": {
+            "class": "logging.StreamHandler",
+            "formatter": "ncsa_access",
+            "stream": "ext://sys.stdout",
+        },
+        "default": {
+            "class": "logging.StreamHandler",
+            "formatter": "default",
+            "stream": "ext://sys.stderr",
+        },
+    },
+    "loggers": {
+        "uvicorn": {"handlers": ["default"], "level": "INFO", "propagate": False},
+        "uvicorn.error": {
+            "handlers": ["default"],
+            "level": "INFO",
+            "propagate": False,
+        },
+        "uvicorn.access": {
+            "handlers": ["access"],
+            "level": "INFO",
+            "propagate": False,
+        },
+        "gunicorn.error": {
+            "handlers": ["default"],
+            "level": "INFO",
+            "propagate": False,
+        },
+        "gunicorn.access": {
+            "handlers": ["access"],
+            "level": "INFO",
+            "propagate": False,
+        },
+    },
+    "root": {"handlers": ["default"], "level": "INFO"},
+}
+
+
+def on_starting(server):  # pragma: no cover — gunicorn hook
+    """Ensure gunicorn's own loggers use the configured handlers."""
+    logging.config.dictConfig(logconfig_dict)

application/llm/base.py

@@ -17,6 +17,8 @@ class BaseLLM(ABC):
         model_id=None,
         base_url=None,
         backup_models=None,
+        model_user_id=None,
+        capabilities=None,
     ):
         self.decoded_token = decoded_token
         self.agent_id = str(agent_id) if agent_id else None
@@ -25,6 +27,12 @@ class BaseLLM(ABC):
         self.token_usage = {"prompt_tokens": 0, "generated_tokens": 0}
         self._backup_models = backup_models or []
         self._fallback_llm = None
+        # Registry-resolved per-model capability overrides (BYOM caps,
+        # operator YAML). None falls back to provider-class defaults.
+        self.capabilities = capabilities
+        # BYOM-resolution scope captured at LLM creation time so backup
+        # / fallback lookups hit the same per-user layer as the primary.
+        self.model_user_id = model_user_id
 
     @property
     def fallback_llm(self):
@@ -39,10 +47,19 @@ class BaseLLM(ABC):
             get_api_key_for_provider,
         )
 
-        # Try per-agent backup models first
+        # model_user_id (BYOM scope) takes precedence over the caller's
+        # sub so shared-agent backups resolve under the owner's layer.
+        caller_sub = (
+            self.decoded_token.get("sub")
+            if isinstance(self.decoded_token, dict)
+            else None
+        )
+        backup_user_id = self.model_user_id or caller_sub
         for backup_model_id in self._backup_models:
             try:
-                provider = get_provider_from_model_id(backup_model_id)
+                provider = get_provider_from_model_id(
+                    backup_model_id, user_id=backup_user_id
+                )
                 if not provider:
                     logger.warning(
                         f"Could not resolve provider for backup model: {backup_model_id}"
@@ -56,6 +73,7 @@ class BaseLLM(ABC):
                     decoded_token=self.decoded_token,
                     model_id=backup_model_id,
                     agent_id=self.agent_id,
+                    model_user_id=self.model_user_id,
                 )
                 logger.info(
                     f"Fallback LLM initialized from agent backup model: "
@@ -68,7 +86,10 @@ class BaseLLM(ABC):
                 )
                 continue
 
-        # Fall back to global FALLBACK_* settings
+        # Fall back to global FALLBACK_* settings. Forward
+        # ``model_user_id`` here too: deployments can configure
+        # ``FALLBACK_LLM_NAME`` to a BYOM UUID, and that UUID is owned
+        # by the same user the primary model was resolved under.
         if settings.FALLBACK_LLM_PROVIDER:
             try:
                 self._fallback_llm = LLMCreator.create_llm(
@@ -78,6 +99,7 @@ class BaseLLM(ABC):
                     decoded_token=self.decoded_token,
                     model_id=settings.FALLBACK_LLM_NAME,
                     agent_id=self.agent_id,
+                    model_user_id=self.model_user_id,
                 )
                 logger.info(
                     f"Fallback LLM initialized from global settings: "

application/llm/google_ai.py

@@ -127,15 +127,33 @@ class GoogleLLM(BaseLLM):
                 ).uri,
             )
 
-            from application.core.mongo_db import MongoDB
-
-            mongo = MongoDB.get_client()
-            db = mongo[settings.MONGO_DB_NAME]
-            attachments_collection = db["attachments"]
-            if "_id" in attachment:
-                attachments_collection.update_one(
-                    {"_id": attachment["_id"]}, {"$set": {"google_file_uri": file_uri}}
+            # Cache the Google file URI on the attachment row so we don't
+            # re-upload on the next LLM call. Accept either a PG UUID
+            # (``id``) or a legacy Mongo ObjectId (``_id``). Opened per
+            # write — this runs mid-LLM-call, so we don't wrap the
+            # surrounding generator in a long-lived session.
+            attachment_id = attachment.get("id") or attachment.get("_id")
+            if attachment_id:
+                user_id = None
+                decoded = getattr(self, "decoded_token", None)
+                if isinstance(decoded, dict):
+                    user_id = decoded.get("sub")
+                from application.storage.db.repositories.attachments import (
+                    AttachmentsRepository,
                 )
+                from application.storage.db.session import db_session
+
+                try:
+                    with db_session() as conn:
+                        AttachmentsRepository(conn).update_any(
+                            str(attachment_id),
+                            user_id,
+                            {"google_file_uri": file_uri},
+                        )
+                except Exception as cache_err:
+                    logging.warning(
+                        f"Failed to cache google_file_uri on attachment {attachment_id}: {cache_err}"
+                    )
             return file_uri
         except Exception as e:
             logging.error(f"Error uploading file to Google AI: {e}", exc_info=True)

application/llm/handlers/base.py

@@ -470,10 +470,14 @@ class LLMHandler(ABC):
                 )
                 return self._perform_in_memory_compression(agent, messages)
 
-            # Use orchestrator to perform compression
+            # Use orchestrator to perform compression. ``model_user_id``
+            # keeps BYOM registry resolution scoped to the model owner
+            # (shared-agent dispatch) while ``user_id`` stays the caller
+            # for the conversation access check.
             result = orchestrator.compress_mid_execution(
                 conversation_id=agent.conversation_id,
                 user_id=agent.initial_user_id,
+                model_user_id=getattr(agent, "model_user_id", None),
                 model_id=agent.model_id,
                 decoded_token=getattr(agent, "decoded_token", {}),
                 current_conversation=conversation,
@@ -577,7 +581,20 @@ class LLMHandler(ABC):
                 if settings.COMPRESSION_MODEL_OVERRIDE
                 else agent.model_id
             )
-            provider = get_provider_from_model_id(compression_model)
+            agent_decoded = getattr(agent, "decoded_token", None)
+            caller_sub = (
+                agent_decoded.get("sub")
+                if isinstance(agent_decoded, dict)
+                else None
+            )
+            # Use model-owner scope (mirrors orchestrator path) so
+            # shared-agent owner-BYOM resolves under the owner's layer.
+            compression_user_id = (
+                getattr(agent, "model_user_id", None) or caller_sub
+            )
+            provider = get_provider_from_model_id(
+                compression_model, user_id=compression_user_id
+            )
             api_key = get_api_key_for_provider(provider)
             compression_llm = LLMCreator.create_llm(
                 provider,
@@ -586,6 +603,7 @@ class LLMHandler(ABC):
                 getattr(agent, "decoded_token", None),
                 model_id=compression_model,
                 agent_id=getattr(agent, "agent_id", None),
+                model_user_id=compression_user_id,
             )
 
             # Create service without DB persistence capability
@@ -921,8 +939,15 @@ class LLMHandler(ABC):
                 }
                 return ""
 
+            # ``agent.model_id`` is the registry id (a UUID for BYOM
+            # records). Use the LLM's own model_id, which LLMCreator
+            # already resolved to the upstream model name. Built-ins:
+            # the two are equal; BYOM: the upstream name like
+            # "mistral-large-latest" instead of the UUID.
             response = agent.llm.gen(
-                model=agent.model_id, messages=messages, tools=agent.tools
+                model=getattr(agent.llm, "model_id", None) or agent.model_id,
+                messages=messages,
+                tools=agent.tools,
             )
             parsed = self.parse_response(response)
             self.llm_calls.append(build_stack_data(agent.llm))
@@ -1011,8 +1036,11 @@ class LLMHandler(ABC):
                     })
                     logger.info("Context limit reached - instructing agent to wrap up")
 
+                # See note above on agent.model_id vs llm.model_id.
                 response = agent.llm.gen_stream(
-                    model=agent.model_id, messages=messages, tools=agent.tools if not agent.context_limit_reached else None
+                    model=getattr(agent.llm, "model_id", None) or agent.model_id,
+                    messages=messages,
+                    tools=agent.tools if not agent.context_limit_reached else None,
                 )
                 self.llm_calls.append(build_stack_data(agent.llm))
 

application/llm/llm_creator.py

@@ -1,34 +1,11 @@
 import logging
 
-from application.llm.anthropic import AnthropicLLM
-from application.llm.docsgpt_provider import DocsGPTAPILLM
-from application.llm.google_ai import GoogleLLM
-from application.llm.groq import GroqLLM
-from application.llm.llama_cpp import LlamaCpp
-from application.llm.novita import NovitaLLM
-from application.llm.openai import AzureOpenAILLM, OpenAILLM
-from application.llm.premai import PremAILLM
-from application.llm.sagemaker import SagemakerAPILLM
-from application.llm.open_router import OpenRouterLLM
+from application.llm.providers import PROVIDERS_BY_NAME
 
 logger = logging.getLogger(__name__)
 
 
 class LLMCreator:
-    llms = {
-        "openai": OpenAILLM,
-        "azure_openai": AzureOpenAILLM,
-        "sagemaker": SagemakerAPILLM,
-        "llama.cpp": LlamaCpp,
-        "anthropic": AnthropicLLM,
-        "docsgpt": DocsGPTAPILLM,
-        "premai": PremAILLM,
-        "groq": GroqLLM,
-        "google": GoogleLLM,
-        "novita": NovitaLLM,
-        "openrouter": OpenRouterLLM,
-    }
-
     @classmethod
     def create_llm(
         cls,
@@ -39,28 +16,111 @@ class LLMCreator:
         model_id=None,
         agent_id=None,
         backup_models=None,
+        model_user_id=None,
         *args,
         **kwargs,
     ):
-        from application.core.model_utils import get_base_url_for_model
+        """Construct an LLM for the given provider ``type``.
 
-        llm_class = cls.llms.get(type.lower())
-        if not llm_class:
+        ``model_user_id`` is the BYOM-resolution scope. Defaults to
+        ``decoded_token['sub']`` (the caller). Pass it explicitly when
+        the model record belongs to a *different* user — most notably
+        for shared-agent dispatch, where the agent's stored
+        ``default_model_id`` is the owner's BYOM UUID but
+        ``decoded_token`` represents the caller.
+        """
+        from application.core.model_registry import ModelRegistry
+        from application.security.safe_url import (
+            UnsafeUserUrlError,
+            pinned_httpx_client,
+            validate_user_base_url,
+        )
+
+        plugin = PROVIDERS_BY_NAME.get(type.lower())
+        if plugin is None or plugin.llm_class is None:
             raise ValueError(f"No LLM class found for type {type}")
 
-        # Extract base_url from model configuration if model_id is provided
+        # Prefer per-model endpoint config from the registry. This is what
+        # makes openai_compatible AND end-user BYOM work without changing
+        # every call site: if the registered AvailableModel carries its
+        # own api_key / base_url, they win over whatever the caller
+        # resolved via the provider plugin.
+        #
+        # End-user BYOM lookups need the user_id from decoded_token to
+        # find the user's per-user models layer (built-in models resolve
+        # without it, so this stays back-compat).
         base_url = None
+        upstream_model_id = model_id
+        capabilities = None
         if model_id:
-            base_url = get_base_url_for_model(model_id)
+            user_id = model_user_id
+            if user_id is None:
+                user_id = (
+                    (decoded_token or {}).get("sub") if decoded_token else None
+                )
+            model = ModelRegistry.get_instance().get_model(model_id, user_id=user_id)
+            if model is not None:
+                # Forward registry caps so the LLM enforces them at
+                # dispatch (built-in classes hard-code True otherwise).
+                capabilities = getattr(model, "capabilities", None)
+                # SECURITY: refuse user-source dispatch without its own
+                # api_key (would leak settings.API_KEY to base_url).
+                if (
+                    getattr(model, "source", "builtin") == "user"
+                    and not model.api_key
+                ):
+                    raise ValueError(
+                        f"Custom model {model_id!r} has no usable API key "
+                        "(decryption may have failed). Re-save the model "
+                        "in settings to dispatch it."
+                    )
+                if model.api_key:
+                    api_key = model.api_key
+                if model.base_url:
+                    base_url = model.base_url
+                # For BYOM the registry id is a UUID; the upstream API
+                # call needs the user's typed model name instead.
+                if model.upstream_model_id:
+                    upstream_model_id = model.upstream_model_id
 
-        return llm_class(
+                # SECURITY: re-validate at dispatch (defense in depth
+                # for pre-guard rows / YAML-supplied entries). The
+                # pinned httpx.Client below is what actually closes the
+                # DNS-rebinding TOCTOU window.
+                if base_url and getattr(model, "source", "builtin") == "user":
+                    try:
+                        validate_user_base_url(base_url)
+                    except UnsafeUserUrlError as e:
+                        raise ValueError(
+                            f"Refusing to dispatch model {model_id!r}: {e}"
+                        ) from e
+                    # Pinned httpx.Client: resolves once, validates, and
+                    # binds the SDK's outbound socket to the validated IP
+                    # (preserves Host / SNI). Future BYOM providers must
+                    # opt in explicitly — only openai_compatible takes
+                    # http_client today.
+                    if plugin.name == "openai_compatible":
+                        try:
+                            kwargs["http_client"] = pinned_httpx_client(
+                                base_url
+                            )
+                        except UnsafeUserUrlError as e:
+                            raise ValueError(
+                                f"Refusing to dispatch model {model_id!r}: {e}"
+                            ) from e
+
+        # Forward model_user_id so backup/fallback resolves under the
+        # owner's scope on shared-agent dispatch.
+        return plugin.llm_class(
             api_key,
             user_api_key,
             decoded_token=decoded_token,
-            model_id=model_id,
+            model_id=upstream_model_id,
             agent_id=agent_id,
             base_url=base_url,
             backup_models=backup_models,
+            model_user_id=model_user_id,
+            capabilities=capabilities,
             *args,
             **kwargs,
         )

application/llm/openai.py

@@ -62,7 +62,15 @@ def _truncate_base64_for_logging(messages):
 
 class OpenAILLM(BaseLLM):
 
-    def __init__(self, api_key=None, user_api_key=None, base_url=None, *args, **kwargs):
+    def __init__(
+        self,
+        api_key=None,
+        user_api_key=None,
+        base_url=None,
+        http_client=None,
+        *args,
+        **kwargs,
+    ):
 
         super().__init__(*args, **kwargs)
         self.api_key = api_key or settings.OPENAI_API_KEY or settings.API_KEY
@@ -80,7 +88,18 @@ class OpenAILLM(BaseLLM):
         else:
             effective_base_url = "https://api.openai.com/v1"
 
-        self.client = OpenAI(api_key=self.api_key, base_url=effective_base_url)
+        # http_client (set by LLMCreator for BYOM) is a DNS-rebinding-safe
+        # httpx.Client; without it the SDK re-resolves DNS per request.
+        if http_client is not None:
+            self.client = OpenAI(
+                api_key=self.api_key,
+                base_url=effective_base_url,
+                http_client=http_client,
+            )
+        else:
+            self.client = OpenAI(
+                api_key=self.api_key, base_url=effective_base_url
+            )
         self.storage = StorageCreator.get_storage()
 
     def _clean_messages_openai(self, messages):
@@ -243,6 +262,13 @@ class OpenAILLM(BaseLLM):
         if "max_tokens" in kwargs:
             kwargs["max_completion_tokens"] = kwargs.pop("max_tokens")
 
+        # Defense-in-depth: drop tools / response_format if the
+        # registry's capability flags deny them.
+        if tools and not self._supports_tools():
+            tools = None
+        if response_format and not self._supports_structured_output():
+            response_format = None
+
         request_params = {
             "model": model,
             "messages": messages,
@@ -279,6 +305,13 @@ class OpenAILLM(BaseLLM):
         if "max_tokens" in kwargs:
             kwargs["max_completion_tokens"] = kwargs.pop("max_tokens")
 
+        # See _raw_gen for rationale — drop tools/response_format when the
+        # registry-provided capabilities say the model doesn't support them.
+        if tools and not self._supports_tools():
+            tools = None
+        if response_format and not self._supports_structured_output():
+            response_format = None
+
         request_params = {
             "model": model,
             "messages": messages,
@@ -320,9 +353,17 @@ class OpenAILLM(BaseLLM):
                 response.close()
 
     def _supports_tools(self):
+        # When the LLM was constructed via LLMCreator with a registered
+        # AvailableModel, ``self.capabilities`` is the per-model record.
+        # BYOM users can disable tool support; respect that. Otherwise
+        # OpenAI's API supports tools by default.
+        if self.capabilities is not None:
+            return bool(self.capabilities.supports_tools)
         return True
 
     def _supports_structured_output(self):
+        if self.capabilities is not None:
+            return bool(self.capabilities.supports_structured_output)
         return True
 
     def prepare_structured_output_format(self, json_schema):
@@ -389,8 +430,14 @@ class OpenAILLM(BaseLLM):
         Returns:
             list: List of supported MIME types
         """
-        from application.core.model_configs import OPENAI_ATTACHMENTS
-        return OPENAI_ATTACHMENTS
+        # Per-model caps from the registry win when present — a BYOM
+        # endpoint that doesn't accept images would otherwise still be
+        # sent base64 image parts because the OpenAI default below
+        # advertises the image alias unconditionally.
+        if self.capabilities is not None:
+            return list(self.capabilities.supported_attachment_types or [])
+        from application.core.model_yaml import resolve_attachment_alias
+        return resolve_attachment_alias("image")
 
     def prepare_messages_with_attachments(self, messages, attachments=None):
         """
@@ -527,15 +574,34 @@ class OpenAILLM(BaseLLM):
                 ).id,
             )
 
-            from application.core.mongo_db import MongoDB
-
-            mongo = MongoDB.get_client()
-            db = mongo[settings.MONGO_DB_NAME]
-            attachments_collection = db["attachments"]
-            if "_id" in attachment:
-                attachments_collection.update_one(
-                    {"_id": attachment["_id"]}, {"$set": {"openai_file_id": file_id}}
+            # Cache the OpenAI file id on the attachment row so we don't
+            # re-upload the same blob on the next LLM call. Prefer the PG
+            # UUID (``id``) when present; fall back to the legacy Mongo
+            # ObjectId string (``_id``). Opened per-write — this runs
+            # inside the hot LLM path, so we don't want a long-lived
+            # session wrapping the generator.
+            attachment_id = attachment.get("id") or attachment.get("_id")
+            if attachment_id:
+                user_id = None
+                decoded = getattr(self, "decoded_token", None)
+                if isinstance(decoded, dict):
+                    user_id = decoded.get("sub")
+                from application.storage.db.repositories.attachments import (
+                    AttachmentsRepository,
                 )
+                from application.storage.db.session import db_session
+
+                try:
+                    with db_session() as conn:
+                        AttachmentsRepository(conn).update_any(
+                            str(attachment_id),
+                            user_id,
+                            {"openai_file_id": file_id},
+                        )
+                except Exception as cache_err:
+                    logging.warning(
+                        f"Failed to cache openai_file_id on attachment {attachment_id}: {cache_err}"
+                    )
             return file_id
         except Exception as e:
             logging.error(f"Error uploading file to OpenAI: {e}", exc_info=True)

application/llm/providers/__init__.py

@@ -0,0 +1,51 @@
+"""Provider plugin registry.
+
+Plugins are imported eagerly so import errors surface at app boot rather
+than at first request. ``ALL_PROVIDERS`` is the canonical ordered list;
+``PROVIDERS_BY_NAME`` is a name-keyed lookup for LLMCreator and the
+model registry.
+"""
+
+from __future__ import annotations
+
+from typing import Dict, List
+
+from application.llm.providers.anthropic import AnthropicProvider
+from application.llm.providers.azure_openai import AzureOpenAIProvider
+from application.llm.providers.base import Provider
+from application.llm.providers.docsgpt import DocsGPTProvider
+from application.llm.providers.google import GoogleProvider
+from application.llm.providers.groq import GroqProvider
+from application.llm.providers.huggingface import HuggingFaceProvider
+from application.llm.providers.llama_cpp import LlamaCppProvider
+from application.llm.providers.novita import NovitaProvider
+from application.llm.providers.openai import OpenAIProvider
+from application.llm.providers.openai_compatible import OpenAICompatibleProvider
+from application.llm.providers.openrouter import OpenRouterProvider
+from application.llm.providers.premai import PremAIProvider
+from application.llm.providers.sagemaker import SagemakerProvider
+
+# Order here is the order the registry iterates providers (and therefore
+# the order ``/api/models`` reports them). Match the historical order
+# from the old ModelRegistry._load_models for byte-stable output during
+# the migration. ``openai_compatible`` slots in right after ``openai``
+# so legacy ``OPENAI_BASE_URL`` models keep landing in the same place.
+ALL_PROVIDERS: List[Provider] = [
+    DocsGPTProvider(),
+    OpenAIProvider(),
+    OpenAICompatibleProvider(),
+    AzureOpenAIProvider(),
+    AnthropicProvider(),
+    GoogleProvider(),
+    GroqProvider(),
+    OpenRouterProvider(),
+    NovitaProvider(),
+    HuggingFaceProvider(),
+    LlamaCppProvider(),
+    PremAIProvider(),
+    SagemakerProvider(),
+]
+
+PROVIDERS_BY_NAME: Dict[str, Provider] = {p.name: p for p in ALL_PROVIDERS}
+
+__all__ = ["ALL_PROVIDERS", "PROVIDERS_BY_NAME", "Provider"]

application/llm/providers/_apikey_or_llm_name.py

@@ -0,0 +1,51 @@
+"""Shared helper for providers that follow the
+``<X>_API_KEY or (LLM_PROVIDER==X and API_KEY)`` pattern.
+
+This is the dominant pattern across Anthropic, Google, Groq, OpenRouter,
+and Novita. Extracted here so each plugin stays a few lines long.
+"""
+
+from __future__ import annotations
+
+from typing import List, Optional
+
+from application.core.model_settings import AvailableModel
+
+
+def get_api_key(
+    settings,
+    provider_name: str,
+    provider_specific_key: Optional[str],
+) -> Optional[str]:
+    if provider_specific_key:
+        return provider_specific_key
+    if settings.LLM_PROVIDER == provider_name and settings.API_KEY:
+        return settings.API_KEY
+    return None
+
+
+def filter_models_by_llm_name(
+    settings,
+    provider_name: str,
+    provider_specific_key: Optional[str],
+    models: List[AvailableModel],
+) -> List[AvailableModel]:
+    """Mirrors the historical ``_add_<X>_models`` selection logic.
+
+    Behavior:
+    - If the provider-specific API key is set → load all models.
+    - Else if ``LLM_PROVIDER`` matches and ``LLM_NAME`` matches a known
+      model → load just that model.
+    - Otherwise → load all models (preserved "load anyway" branch from
+      the original methods).
+    """
+    if provider_specific_key:
+        return models
+    if (
+        settings.LLM_PROVIDER == provider_name
+        and settings.LLM_NAME
+    ):
+        named = [m for m in models if m.id == settings.LLM_NAME]
+        if named:
+            return named
+    return models

application/llm/providers/anthropic.py

@@ -0,0 +1,23 @@
+from __future__ import annotations
+
+from typing import Optional
+
+from application.llm.anthropic import AnthropicLLM
+from application.llm.providers._apikey_or_llm_name import (
+    filter_models_by_llm_name,
+    get_api_key,
+)
+from application.llm.providers.base import Provider
+
+
+class AnthropicProvider(Provider):
+    name = "anthropic"
+    llm_class = AnthropicLLM
+
+    def get_api_key(self, settings) -> Optional[str]:
+        return get_api_key(settings, self.name, settings.ANTHROPIC_API_KEY)
+
+    def filter_yaml_models(self, settings, models):
+        return filter_models_by_llm_name(
+            settings, self.name, settings.ANTHROPIC_API_KEY, models
+        )

application/llm/providers/azure_openai.py

@@ -0,0 +1,30 @@
+from __future__ import annotations
+
+from typing import Optional
+
+from application.llm.openai import AzureOpenAILLM
+from application.llm.providers.base import Provider
+
+
+class AzureOpenAIProvider(Provider):
+    name = "azure_openai"
+    llm_class = AzureOpenAILLM
+
+    def get_api_key(self, settings) -> Optional[str]:
+        # Azure historically uses the generic API_KEY field.
+        return settings.API_KEY
+
+    def is_enabled(self, settings) -> bool:
+        if settings.OPENAI_API_BASE:
+            return True
+        return settings.LLM_PROVIDER == self.name and bool(settings.API_KEY)
+
+    def filter_yaml_models(self, settings, models):
+        # Mirrors _add_azure_openai_models: when LLM_PROVIDER==azure_openai
+        # and LLM_NAME matches a known model, narrow to that one model.
+        # Otherwise load the entire catalog.
+        if settings.LLM_PROVIDER == self.name and settings.LLM_NAME:
+            named = [m for m in models if m.id == settings.LLM_NAME]
+            if named:
+                return named
+        return models

application/llm/providers/base.py

@@ -0,0 +1,74 @@
+from __future__ import annotations
+
+from abc import ABC, abstractmethod
+from typing import TYPE_CHECKING, ClassVar, List, Optional, Type
+
+if TYPE_CHECKING:
+    from application.core.model_settings import AvailableModel
+    from application.core.model_yaml import ProviderCatalog
+    from application.core.settings import Settings
+    from application.llm.base import BaseLLM
+
+
+class Provider(ABC):
+    """Owns the *behavior* of an LLM provider.
+
+    Concrete providers declare their name, the LLM class to instantiate,
+    and how to resolve credentials from settings. Static model catalogs
+    live in YAML under ``application/core/models/`` and are joined to the
+    provider by name at registry load time.
+
+    Most plugins receive zero or one catalog at registry-build time. The
+    ``openai_compatible`` plugin is the exception: it receives one catalog
+    per matching YAML file, each with its own ``api_key_env`` and
+    ``base_url``. Plugins that need per-catalog metadata override
+    ``get_models``; the default implementation merges catalogs and routes
+    through ``filter_yaml_models`` + ``extra_models``.
+    """
+
+    name: ClassVar[str]
+    # ``None`` means the provider appears in the catalog but isn't
+    # dispatchable through LLMCreator (e.g. Hugging Face today, where the
+    # original LLMCreator dict had no entry).
+    llm_class: ClassVar[Optional[Type["BaseLLM"]]] = None
+
+    @abstractmethod
+    def get_api_key(self, settings: "Settings") -> Optional[str]:
+        """Return the API key for this provider, or None if unavailable."""
+
+    def is_enabled(self, settings: "Settings") -> bool:
+        """Whether this provider should contribute models to the registry."""
+        return bool(self.get_api_key(settings))
+
+    def filter_yaml_models(
+        self, settings: "Settings", models: List["AvailableModel"]
+    ) -> List["AvailableModel"]:
+        """Hook to filter YAML-loaded models. Default: return all."""
+        return models
+
+    def extra_models(self, settings: "Settings") -> List["AvailableModel"]:
+        """Hook to add dynamic models not declared in YAML. Default: none."""
+        return []
+
+    def get_models(
+        self,
+        settings: "Settings",
+        catalogs: List["ProviderCatalog"],
+    ) -> List["AvailableModel"]:
+        """Final list of models this plugin contributes.
+
+        Default: merge the models across all matched catalogs (later
+        catalog wins on duplicate id), filter via ``filter_yaml_models``,
+        then append ``extra_models``. Override when per-catalog metadata
+        matters (see ``OpenAICompatibleProvider``).
+        """
+        merged: List["AvailableModel"] = []
+        seen: dict = {}
+        for c in catalogs:
+            for m in c.models:
+                if m.id in seen:
+                    merged[seen[m.id]] = m
+                else:
+                    seen[m.id] = len(merged)
+                    merged.append(m)
+        return self.filter_yaml_models(settings, merged) + self.extra_models(settings)

application/llm/providers/docsgpt.py

@@ -0,0 +1,22 @@
+from __future__ import annotations
+
+from typing import Optional
+
+from application.llm.docsgpt_provider import DocsGPTAPILLM
+from application.llm.providers.base import Provider
+
+
+class DocsGPTProvider(Provider):
+    name = "docsgpt"
+    llm_class = DocsGPTAPILLM
+
+    def get_api_key(self, settings) -> Optional[str]:
+        # No provider-specific key; the LLM class can use the generic
+        # API_KEY fallback if it needs one. Mirrors model_utils' historical
+        # behavior of returning settings.API_KEY when no specific key exists.
+        return settings.API_KEY
+
+    def is_enabled(self, settings) -> bool:
+        # The hosted DocsGPT model is hidden when the deployment is
+        # pointed at a custom OpenAI-compatible endpoint.
+        return not settings.OPENAI_BASE_URL

application/llm/providers/google.py

@@ -0,0 +1,23 @@
+from __future__ import annotations
+
+from typing import Optional
+
+from application.llm.google_ai import GoogleLLM
+from application.llm.providers._apikey_or_llm_name import (
+    filter_models_by_llm_name,
+    get_api_key,
+)
+from application.llm.providers.base import Provider
+
+
+class GoogleProvider(Provider):
+    name = "google"
+    llm_class = GoogleLLM
+
+    def get_api_key(self, settings) -> Optional[str]:
+        return get_api_key(settings, self.name, settings.GOOGLE_API_KEY)
+
+    def filter_yaml_models(self, settings, models):
+        return filter_models_by_llm_name(
+            settings, self.name, settings.GOOGLE_API_KEY, models
+        )

application/llm/providers/groq.py

@@ -0,0 +1,23 @@
+from __future__ import annotations
+
+from typing import Optional
+
+from application.llm.groq import GroqLLM
+from application.llm.providers._apikey_or_llm_name import (
+    filter_models_by_llm_name,
+    get_api_key,
+)
+from application.llm.providers.base import Provider
+
+
+class GroqProvider(Provider):
+    name = "groq"
+    llm_class = GroqLLM
+
+    def get_api_key(self, settings) -> Optional[str]:
+        return get_api_key(settings, self.name, settings.GROQ_API_KEY)
+
+    def filter_yaml_models(self, settings, models):
+        return filter_models_by_llm_name(
+            settings, self.name, settings.GROQ_API_KEY, models
+        )

application/llm/providers/huggingface.py

@@ -0,0 +1,25 @@
+from __future__ import annotations
+
+from typing import Optional
+
+from application.llm.providers._apikey_or_llm_name import (
+    get_api_key as shared_get_api_key,
+)
+from application.llm.providers.base import Provider
+
+
+class HuggingFaceProvider(Provider):
+    """Surfaces ``huggingface-local`` to the model catalog.
+
+    Not dispatchable through LLMCreator — historically there was no
+    HuggingFaceLLM entry in ``LLMCreator.llms``, and calling ``create_llm``
+    with ``"huggingface"`` raised ``ValueError``. We preserve that
+    behavior: the model appears in ``/api/models`` but selecting it
+    surfaces the same error it always did.
+    """
+
+    name = "huggingface"
+    llm_class = None  # not dispatchable
+
+    def get_api_key(self, settings) -> Optional[str]:
+        return shared_get_api_key(settings, self.name, settings.HUGGINGFACE_API_KEY)

application/llm/providers/llama_cpp.py

@@ -0,0 +1,19 @@
+from __future__ import annotations
+
+from typing import Optional
+
+from application.llm.llama_cpp import LlamaCpp
+from application.llm.providers.base import Provider
+
+
+class LlamaCppProvider(Provider):
+    """LLMCreator-only plugin: invocable via LLM_PROVIDER but not in the catalog."""
+
+    name = "llama.cpp"
+    llm_class = LlamaCpp
+
+    def get_api_key(self, settings) -> Optional[str]:
+        return settings.API_KEY
+
+    def is_enabled(self, settings) -> bool:
+        return False

application/llm/providers/novita.py

@@ -0,0 +1,23 @@
+from __future__ import annotations
+
+from typing import Optional
+
+from application.llm.novita import NovitaLLM
+from application.llm.providers._apikey_or_llm_name import (
+    filter_models_by_llm_name,
+    get_api_key,
+)
+from application.llm.providers.base import Provider
+
+
+class NovitaProvider(Provider):
+    name = "novita"
+    llm_class = NovitaLLM
+
+    def get_api_key(self, settings) -> Optional[str]:
+        return get_api_key(settings, self.name, settings.NOVITA_API_KEY)
+
+    def filter_yaml_models(self, settings, models):
+        return filter_models_by_llm_name(
+            settings, self.name, settings.NOVITA_API_KEY, models
+        )

application/llm/providers/openai.py

@@ -0,0 +1,37 @@
+from __future__ import annotations
+
+from typing import Optional
+
+from application.llm.openai import OpenAILLM
+from application.llm.providers.base import Provider
+
+
+class OpenAIProvider(Provider):
+    name = "openai"
+    llm_class = OpenAILLM
+
+    def get_api_key(self, settings) -> Optional[str]:
+        if settings.OPENAI_API_KEY:
+            return settings.OPENAI_API_KEY
+        if settings.LLM_PROVIDER == self.name and settings.API_KEY:
+            return settings.API_KEY
+        return None
+
+    def is_enabled(self, settings) -> bool:
+        # When the deployment is pointed at a custom OpenAI-compatible
+        # endpoint (Ollama, LM Studio, ...), the cloud-OpenAI catalog is
+        # suppressed but ``is_enabled`` stays True — necessary so the
+        # filter below still gets to drop the catalog (rather than the
+        # registry skipping the provider entirely and missing the rule).
+        if settings.OPENAI_BASE_URL:
+            return True
+        return bool(self.get_api_key(settings))
+
+    def filter_yaml_models(self, settings, models):
+        # Legacy local-endpoint mode hides the cloud catalog. The
+        # corresponding dynamic models live in OpenAICompatibleProvider.
+        if settings.OPENAI_BASE_URL:
+            return []
+        if not settings.OPENAI_API_KEY:
+            return []
+        return models

application/llm/providers/openai_compatible.py

@@ -0,0 +1,149 @@
+"""Generic provider for OpenAI-wire-compatible endpoints.
+
+Each ``openai_compatible`` YAML file describes one logical endpoint
+(Mistral, Together, Fireworks, Ollama, ...) with its own
+``api_key_env`` and ``base_url``. Multiple files can coexist; the
+plugin produces one set of models per file, each pre-configured with
+the right credentials and URL.
+
+The plugin also handles the **legacy** ``OPENAI_BASE_URL`` + ``LLM_NAME``
+local-endpoint pattern that previously lived in ``OpenAIProvider``. That
+path generates models dynamically from ``LLM_NAME``, using
+``OPENAI_BASE_URL`` and ``OPENAI_API_KEY`` as the endpoint config.
+"""
+
+from __future__ import annotations
+
+import logging
+import os
+from typing import List, Optional
+
+from application.core.model_settings import (
+    AvailableModel,
+    ModelCapabilities,
+    ModelProvider,
+)
+from application.llm.openai import OpenAILLM
+from application.llm.providers.base import Provider
+
+logger = logging.getLogger(__name__)
+
+
+def _parse_model_names(llm_name: Optional[str]) -> List[str]:
+    if not llm_name:
+        return []
+    return [name.strip() for name in llm_name.split(",") if name.strip()]
+
+
+class OpenAICompatibleProvider(Provider):
+    name = "openai_compatible"
+    llm_class = OpenAILLM
+
+    def get_api_key(self, settings) -> Optional[str]:
+        # Per-model: each catalog supplies its own ``api_key_env``. There
+        # is no single plugin-wide key. LLMCreator reads the per-model
+        # ``api_key`` set during catalog materialization.
+        return None
+
+    def is_enabled(self, settings) -> bool:
+        # Concrete enablement happens per catalog (in ``get_models``).
+        # Returning True lets the registry call ``get_models`` so we can
+        # decide per-file whether to contribute models.
+        return True
+
+    def get_models(self, settings, catalogs) -> List[AvailableModel]:
+        out: List[AvailableModel] = []
+
+        for catalog in catalogs:
+            out.extend(self._materialize_yaml_catalog(catalog))
+
+        if settings.OPENAI_BASE_URL and settings.LLM_NAME:
+            out.extend(self._materialize_legacy_local_endpoint(settings))
+
+        return out
+
+    def _materialize_yaml_catalog(self, catalog) -> List[AvailableModel]:
+        """Resolve one openai_compatible YAML into ready-to-dispatch models.
+
+        Skipped (with an INFO-level log) if ``api_key_env`` resolves to
+        nothing — no point publishing models the user can't actually
+        call. INFO rather than WARNING because operators may legitimately
+        drop multiple provider YAMLs as templates and only set the env
+        vars for the ones they actually use; a missing key is ambiguous,
+        not necessarily a misconfig.
+        """
+        if not catalog.base_url:
+            raise ValueError(
+                f"{catalog.source_path}: openai_compatible YAML must set "
+                "'base_url'."
+            )
+        if not catalog.api_key_env:
+            raise ValueError(
+                f"{catalog.source_path}: openai_compatible YAML must set "
+                "'api_key_env'."
+            )
+
+        api_key = os.environ.get(catalog.api_key_env)
+        if not api_key:
+            logger.info(
+                "openai_compatible catalog %s skipped: env var %s is not set",
+                catalog.source_path,
+                catalog.api_key_env,
+            )
+            return []
+
+        out: List[AvailableModel] = []
+        for m in catalog.models:
+            out.append(self._with_endpoint(m, catalog.base_url, api_key))
+        return out
+
+    def _materialize_legacy_local_endpoint(self, settings) -> List[AvailableModel]:
+        """Generate AvailableModels from ``LLM_NAME`` for the legacy
+        ``OPENAI_BASE_URL`` deployment pattern (Ollama, LM Studio, ...).
+
+        Preserves the historical ``provider="openai"`` display behavior
+        by setting ``display_provider="openai"``.
+        """
+        from application.core.model_yaml import resolve_attachment_alias
+
+        attachments = resolve_attachment_alias("image")
+        api_key = settings.OPENAI_API_KEY or settings.API_KEY
+        out: List[AvailableModel] = []
+        for model_name in _parse_model_names(settings.LLM_NAME):
+            out.append(
+                AvailableModel(
+                    id=model_name,
+                    provider=ModelProvider.OPENAI_COMPATIBLE,
+                    display_name=model_name,
+                    description=f"Custom OpenAI-compatible model at {settings.OPENAI_BASE_URL}",
+                    base_url=settings.OPENAI_BASE_URL,
+                    capabilities=ModelCapabilities(
+                        supports_tools=True,
+                        supported_attachment_types=attachments,
+                    ),
+                    api_key=api_key,
+                    display_provider="openai",
+                )
+            )
+        return out
+
+    @staticmethod
+    def _with_endpoint(
+        model: AvailableModel, base_url: str, api_key: str
+    ) -> AvailableModel:
+        """Return a copy of ``model`` carrying the catalog's endpoint config.
+
+        The catalog-level ``base_url`` is the default; an explicit
+        per-model ``base_url`` in the YAML wins.
+        """
+        return AvailableModel(
+            id=model.id,
+            provider=model.provider,
+            display_name=model.display_name,
+            description=model.description,
+            capabilities=model.capabilities,
+            enabled=model.enabled,
+            base_url=model.base_url or base_url,
+            display_provider=model.display_provider,
+            api_key=api_key,
+        )

application/llm/providers/openrouter.py

@@ -0,0 +1,23 @@
+from __future__ import annotations
+
+from typing import Optional
+
+from application.llm.open_router import OpenRouterLLM
+from application.llm.providers._apikey_or_llm_name import (
+    filter_models_by_llm_name,
+    get_api_key,
+)
+from application.llm.providers.base import Provider
+
+
+class OpenRouterProvider(Provider):
+    name = "openrouter"
+    llm_class = OpenRouterLLM
+
+    def get_api_key(self, settings) -> Optional[str]:
+        return get_api_key(settings, self.name, settings.OPEN_ROUTER_API_KEY)
+
+    def filter_yaml_models(self, settings, models):
+        return filter_models_by_llm_name(
+            settings, self.name, settings.OPEN_ROUTER_API_KEY, models
+        )

application/llm/providers/premai.py

@@ -0,0 +1,19 @@
+from __future__ import annotations
+
+from typing import Optional
+
+from application.llm.premai import PremAILLM
+from application.llm.providers.base import Provider
+
+
+class PremAIProvider(Provider):
+    """LLMCreator-only plugin: invocable via LLM_PROVIDER but not in the catalog."""
+
+    name = "premai"
+    llm_class = PremAILLM
+
+    def get_api_key(self, settings) -> Optional[str]:
+        return settings.API_KEY
+
+    def is_enabled(self, settings) -> bool:
+        return False

application/llm/providers/sagemaker.py

@@ -0,0 +1,24 @@
+from __future__ import annotations
+
+from typing import Optional
+
+from application.llm.sagemaker import SagemakerAPILLM
+from application.llm.providers.base import Provider
+
+
+class SagemakerProvider(Provider):
+    """LLMCreator-only plugin: invocable via LLM_PROVIDER but not in the catalog.
+
+    SageMaker reads its credentials from ``SAGEMAKER_*`` settings inside
+    the LLM class itself; this plugin's ``get_api_key`` exists only for
+    LLMCreator's symmetry.
+    """
+
+    name = "sagemaker"
+    llm_class = SagemakerAPILLM
+
+    def get_api_key(self, settings) -> Optional[str]:
+        return settings.API_KEY
+
+    def is_enabled(self, settings) -> bool:
+        return False

application/logging.py

@@ -6,8 +6,8 @@ import logging
 import uuid
 from typing import Any, Callable, Dict, Generator, List
 
-from application.core.mongo_db import MongoDB
-from application.core.settings import settings
+from application.storage.db.repositories.stack_logs import StackLogsRepository
+from application.storage.db.session import db_session
 
 logging.basicConfig(
     level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s"
@@ -101,7 +101,7 @@ def _consume_and_log(generator: Generator, context: "LogContext"):
     except Exception as e:
         logging.exception(f"Error in {context.endpoint} - {context.activity_id}: {e}")
         context.stacks.append({"component": "error", "data": {"message": str(e)}})
-        _log_to_mongodb(
+        _log_activity_to_db(
             endpoint=context.endpoint,
             activity_id=context.activity_id,
             user=context.user,
@@ -112,7 +112,7 @@ def _consume_and_log(generator: Generator, context: "LogContext"):
         )
         raise
     finally:
-        _log_to_mongodb(
+        _log_activity_to_db(
             endpoint=context.endpoint,
             activity_id=context.activity_id,
             user=context.user,
@@ -123,7 +123,7 @@ def _consume_and_log(generator: Generator, context: "LogContext"):
         )
 
 
-def _log_to_mongodb(
+def _log_activity_to_db(
     endpoint: str,
     activity_id: str,
     user: str,
@@ -132,46 +132,26 @@ def _log_to_mongodb(
     stacks: List[Dict],
     level: str,
 ) -> None:
+    """Append a per-request activity log row to Postgres (``stack_logs``)."""
     try:
-        mongo = MongoDB.get_client()
-        db = mongo[settings.MONGO_DB_NAME]
-        user_logs_collection = db["stack_logs"]
-        
-
-
-        log_entry = {
-            "endpoint": endpoint,
-            "id": activity_id,
-            "level": level,
-            "user": user,
-            "api_key": api_key,
-            "query": query,
-            "stacks": stacks,
-            "timestamp": datetime.datetime.now(datetime.timezone.utc),
-        }
-        # clean up text fields to be no longer than 10000 characters
-        for key, value in log_entry.items():
-            if isinstance(value, str) and len(value) > 10000:
-                log_entry[key] = value[:10000]
-    
-        user_logs_collection.insert_one(log_entry)
-        logging.debug(f"Logged activity to MongoDB: {activity_id}")
-
-        from application.storage.db.dual_write import dual_write
-        from application.storage.db.repositories.stack_logs import StackLogsRepository
-
-        dual_write(
-            StackLogsRepository,
-            lambda repo, e=log_entry: repo.insert(
-                activity_id=e["id"],
-                endpoint=e.get("endpoint"),
-                level=e.get("level"),
-                user_id=e.get("user"),
-                api_key=e.get("api_key"),
-                query=e.get("query"),
-                stacks=e.get("stacks"),
-            ),
-        )
+        # Clean up text fields to be no longer than 10000 characters so a
+        # runaway payload can't blow up the insert.
+        def _truncate(val):
+            if isinstance(val, str) and len(val) > 10000:
+                return val[:10000]
+            return val
 
+        with db_session() as conn:
+            StackLogsRepository(conn).insert(
+                activity_id=activity_id,
+                endpoint=_truncate(endpoint),
+                level=_truncate(level),
+                user_id=_truncate(user),
+                api_key=_truncate(api_key),
+                query=_truncate(query),
+                stacks=stacks,
+                timestamp=datetime.datetime.now(datetime.timezone.utc),
+            )
+        logging.debug(f"Logged activity to Postgres: {activity_id}")
     except Exception as e:
-        logging.error(f"Failed to log to MongoDB: {e}", exc_info=True)
+        logging.error(f"Failed to log activity to Postgres: {e}", exc_info=True)

application/mcp_server.py

@@ -0,0 +1,59 @@
+"""FastMCP server exposing DocsGPT retrieval over streamable HTTP.
+
+Mounted at ``/mcp`` by ``application/asgi.py``. Bearer tokens are the
+existing DocsGPT agent API keys — no new credential surface.
+
+The tool reads the ``Authorization`` header directly via
+``get_http_headers(include={"authorization"})``. The ``include`` kwarg
+is required: by default ``get_http_headers`` strips ``authorization``
+(and a handful of other hop-by-hop headers) so they aren't forwarded
+to downstream services — since we deliberately want the caller's
+token, we opt it back in.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import logging
+
+from fastmcp import FastMCP
+from fastmcp.server.dependencies import get_http_headers
+
+from application.services.search_service import (
+    InvalidAPIKey,
+    SearchFailed,
+    search,
+)
+
+logger = logging.getLogger(__name__)
+
+mcp = FastMCP("docsgpt")
+
+
+def _extract_bearer_token() -> str | None:
+    auth = get_http_headers(include={"authorization"}).get("authorization", "")
+    parts = auth.split(None, 1)
+    if len(parts) != 2 or parts[0].lower() != "bearer" or not parts[1]:
+        return None
+    return parts[1]
+
+
+@mcp.tool
+async def search_docs(query: str, chunks: int = 5) -> list[dict]:
+    """Search the caller's DocsGPT knowledge base.
+
+    Authentication is via ``Authorization: Bearer <agent-api-key>`` on
+    the MCP request — the same opaque key that ``/api/search`` accepts
+    in its JSON body. Returns at most ``chunks`` hits, each a dict with
+    ``text``, ``title``, ``source`` keys.
+    """
+    api_key = _extract_bearer_token()
+    if not api_key:
+        raise PermissionError("Missing Bearer token")
+    try:
+        return await asyncio.to_thread(search, api_key, query, chunks)
+    except InvalidAPIKey as exc:
+        raise PermissionError("Invalid API key") from exc
+    except SearchFailed:
+        logger.exception("search_docs failed")
+        raise

application/parser/connectors/_auth_utils.py

@@ -0,0 +1,37 @@
+"""Shared helpers for connector auth modules.
+
+These helpers exist so that sensitive values (session tokens, bearer
+credentials) never end up interpolated into exception messages or log
+lines. Exception messages frequently flow into ``stack_logs`` (Postgres)
+and Sentry via ``exc_info=True``, so the raw value must never be the
+thing we format.
+"""
+
+from __future__ import annotations
+
+import hashlib
+
+
+def session_token_fingerprint(session_token: str) -> str:
+    """Return a short, irreversible fingerprint for a session token.
+
+    The returned string is safe to embed in exception messages and log
+    lines: it is a prefix of a SHA-256 digest, clearly tagged so an
+    operator reading the log knows it is a hash and not the token
+    itself. It is stable for a given input, which lets operators
+    correlate "which token failed" across log lines without exposing
+    the credential.
+
+    Args:
+        session_token: The raw session token. Accepts ``None`` or the
+            empty string for defensive callers; both yield a distinct
+            sentinel rather than raising.
+
+    Returns:
+        A string of the form ``"sha256:<6 hex chars>"``, or
+        ``"sha256:<empty>"`` when the input is falsy.
+    """
+    if not session_token:
+        return "sha256:<empty>"
+    digest = hashlib.sha256(session_token.encode("utf-8")).hexdigest()
+    return f"sha256:{digest[:6]}"

application/parser/connectors/confluence/auth.py

@@ -6,6 +6,7 @@ from urllib.parse import urlencode
 import requests
 
 from application.core.settings import settings
+from application.parser.connectors._auth_utils import session_token_fingerprint
 from application.parser.connectors.base import BaseConnectorAuth
 
 logger = logging.getLogger(__name__)
@@ -152,15 +153,19 @@ class ConfluenceAuth(BaseConnectorAuth):
             return True
 
     def get_token_info_from_session(self, session_token: str) -> Dict[str, Any]:
-        from application.core.mongo_db import MongoDB
-        from application.core.settings import settings as app_settings
+        from application.storage.db.repositories.connector_sessions import (
+            ConnectorSessionsRepository,
+        )
+        from application.storage.db.session import db_readonly
 
-        mongo = MongoDB.get_client()
-        db = mongo[app_settings.MONGO_DB_NAME]
-
-        session = db["connector_sessions"].find_one({"session_token": session_token})
+        with db_readonly() as conn:
+            session = ConnectorSessionsRepository(conn).get_by_session_token(
+                session_token
+            )
         if not session:
-            raise ValueError(f"Invalid session token: {session_token}")
+            raise ValueError(
+                f"Invalid session token ({session_token_fingerprint(session_token)})"
+            )
 
         token_info = session.get("token_info")
         if not token_info: