fix: pgvec

* feat: durability and idempotency keys

* feat: more durable frontend

* fix: tests

* fix: mini issues

* fix: better json validation

* fix: tests

.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

AGENTS.md

@@ -37,6 +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
@@ -99,7 +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

README.md

@@ -47,11 +47,13 @@
 </ul>
 
 ## Roadmap
-- [x] Add OAuth 2.0 authentication for MCP ( September 2025 )
-- [x] Deep Agents ( October 2025 )
-- [x] Prompt Templating ( October 2025 )
-- [x] Full api tooling ( Dec 2025 )
-- [ ] Agent scheduling ( Jan 2026 )
+- [x] Agent Workflow Builder with conditional nodes ( February 2026 )
+- [x] SharePoint & Confluence connectors ( March – April 2026 )
+- [x] Research mode ( March 2026 )
+- [x] Postgres migration for user data ( April 2026 )
+- [x] OpenTelemetry observability ( April 2026 )
+- [x] Bring Your Own Model (BYOM) ( April 2026 )
+- [ ] Agent scheduling (RedBeat-backed) ( Q2 2026 )
 
 You can find our full roadmap [here](https://github.com/orgs/arc53/projects/2). Please don't hesitate to contribute or create issues, it helps us improve DocsGPT!
 

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

@@ -1,18 +1,107 @@
 import logging
 import uuid
 from collections import Counter
-from typing import Dict, List, Optional, Tuple
+from typing import Any, Dict, List, Optional, Tuple
 
 from application.agents.tools.tool_action_parser import ToolActionParser
 from application.agents.tools.tool_manager import ToolManager
 from application.security.encryption import decrypt_credentials
+from application.storage.db.base_repository import looks_like_uuid
 from application.storage.db.repositories.agents import AgentsRepository
+from application.storage.db.repositories.tool_call_attempts import (
+    ToolCallAttemptsRepository,
+)
 from application.storage.db.repositories.user_tools import UserToolsRepository
-from application.storage.db.session import db_readonly
+from application.storage.db.session import db_readonly, db_session
 
 logger = logging.getLogger(__name__)
 
 
+def _record_proposed(
+    call_id: str,
+    tool_name: str,
+    action_name: str,
+    arguments: Any,
+    *,
+    tool_id: Optional[str] = None,
+) -> bool:
+    """Insert a ``proposed`` row; swallow infra failures so tool calls
+    still run when the journal is unreachable. Returns True iff the row
+    is now journaled (newly created or already present).
+    """
+    try:
+        with db_session() as conn:
+            inserted = ToolCallAttemptsRepository(conn).record_proposed(
+                call_id,
+                tool_name,
+                action_name,
+                arguments,
+                tool_id=tool_id if tool_id and looks_like_uuid(tool_id) else None,
+            )
+        if not inserted:
+            logger.warning(
+                "tool_call_attempts duplicate call_id=%s; existing row left in place",
+                call_id,
+                extra={"alert": "tool_call_id_collision", "call_id": call_id},
+            )
+        return True
+    except Exception:
+        logger.exception("tool_call_attempts proposed write failed for %s", call_id)
+        return False
+
+
+def _mark_executed(
+    call_id: str,
+    result: Any,
+    *,
+    message_id: Optional[str] = None,
+    artifact_id: Optional[str] = None,
+    proposed_ok: bool = True,
+    tool_name: Optional[str] = None,
+    action_name: Optional[str] = None,
+    arguments: Any = None,
+    tool_id: Optional[str] = None,
+) -> None:
+    """Flip the row to ``executed``. If ``proposed_ok`` is False (the
+    proposed write failed earlier), upsert a fresh row in ``executed`` so
+    the reconciler can still see the attempt — without this, the side
+    effect would be invisible to the journal.
+    """
+    try:
+        with db_session() as conn:
+            repo = ToolCallAttemptsRepository(conn)
+            if proposed_ok:
+                updated = repo.mark_executed(
+                    call_id,
+                    result,
+                    message_id=message_id,
+                    artifact_id=artifact_id,
+                )
+                if updated:
+                    return
+            # Fallback synthesizes the row so the journal isn't lost.
+            repo.upsert_executed(
+                call_id,
+                tool_name=tool_name or "unknown",
+                action_name=action_name or "",
+                arguments=arguments if arguments is not None else {},
+                result=result,
+                tool_id=tool_id if tool_id and looks_like_uuid(tool_id) else None,
+                message_id=message_id,
+                artifact_id=artifact_id,
+            )
+    except Exception:
+        logger.exception("tool_call_attempts executed write failed for %s", call_id)
+
+
+def _mark_failed(call_id: str, error: str) -> None:
+    try:
+        with db_session() as conn:
+            ToolCallAttemptsRepository(conn).mark_failed(call_id, error)
+    except Exception:
+        logger.exception("tool_call_attempts failed-write failed for %s", call_id)
+
+
 class ToolExecutor:
     """Handles tool discovery, preparation, and execution.
 
@@ -31,6 +120,7 @@ class ToolExecutor:
         self.tool_calls: List[Dict] = []
         self._loaded_tools: Dict[str, object] = {}
         self.conversation_id: Optional[str] = None
+        self.message_id: Optional[str] = None
         self.client_tools: Optional[List[Dict]] = None
         self._name_to_tool: Dict[str, Tuple[str, str]] = {}
         self._tool_to_name: Dict[Tuple[str, str], str] = {}
@@ -274,7 +364,14 @@ class ToolExecutor:
 
         if tool_id is None or action_name is None:
             error_message = f"Error: Failed to parse LLM tool call. Tool name: {llm_name}"
-            logger.error(error_message)
+            logger.error(
+                "tool_call_parse_failed",
+                extra={
+                    "llm_class_name": llm_class_name,
+                    "llm_tool_name": llm_name,
+                    "call_id": call_id,
+                },
+            )
 
             tool_call_data = {
                 "tool_name": "unknown",
@@ -289,7 +386,15 @@ class ToolExecutor:
 
         if tool_id not in tools_dict:
             error_message = f"Error: Tool ID '{tool_id}' extracted from LLM call not found in available tools_dict. Available IDs: {list(tools_dict.keys())}"
-            logger.error(error_message)
+            logger.error(
+                "tool_id_not_found",
+                extra={
+                    "tool_id": tool_id,
+                    "llm_tool_name": llm_name,
+                    "call_id": call_id,
+                    "available_tool_count": len(tools_dict),
+                },
+            )
 
             tool_call_data = {
                 "tool_name": "unknown",
@@ -308,9 +413,36 @@ class ToolExecutor:
             "action_name": llm_name,
             "arguments": call_args,
         }
-        yield {"type": "tool_call", "data": {**tool_call_data, "status": "pending"}}
-
         tool_data = tools_dict[tool_id]
+        # Journal first so the reconciler sees malformed calls and any
+        # subsequent ``_mark_failed`` actually updates a real row.
+        proposed_ok = _record_proposed(
+            call_id,
+            tool_data["name"],
+            action_name,
+            call_args if isinstance(call_args, dict) else {},
+            tool_id=tool_data.get("id"),
+        )
+        # Defensive guard: a non-dict ``call_args`` (e.g. malformed
+        # JSON on the resume path) would crash the param walk below
+        # with AttributeError on ``.items()``. Surface a clean error
+        # event and flip the journal row to ``failed`` instead of
+        # killing the stream.
+        if not isinstance(call_args, dict):
+            error_message = (
+                f"Tool call arguments must be a JSON object, got "
+                f"{type(call_args).__name__}."
+            )
+            tool_call_data["result"] = error_message
+            tool_call_data["arguments"] = {}
+            _mark_failed(call_id, error_message)
+            yield {
+                "type": "tool_call",
+                "data": {**tool_call_data, "status": "error"},
+            }
+            self.tool_calls.append(tool_call_data)
+            return error_message, call_id
+        yield {"type": "tool_call", "data": {**tool_call_data, "status": "pending"}}
         action_data = (
             tool_data["config"]["actions"][action_name]
             if tool_data["name"] == "api_tool"
@@ -356,8 +488,17 @@ class ToolExecutor:
                 f"Failed to load tool '{tool_data.get('name')}' (tool_id key={tool_id}): "
                 "missing 'id' on tool row."
             )
-            logger.error(error_message)
+            logger.error(
+                "tool_load_failed",
+                extra={
+                    "tool_name": tool_data.get("name"),
+                    "tool_id": tool_id,
+                    "action_name": action_name,
+                    "call_id": call_id,
+                },
+            )
             tool_call_data["result"] = error_message
+            _mark_failed(call_id, error_message)
             yield {"type": "tool_call", "data": {**tool_call_data, "status": "error"}}
             self.tool_calls.append(tool_call_data)
             return error_message, call_id
@@ -367,14 +508,18 @@ class ToolExecutor:
             if tool_data["name"] == "api_tool"
             else parameters
         )
-        if tool_data["name"] == "api_tool":
-            logger.debug(
-                f"Executing api: {action_name} with query_params: {query_params}, headers: {headers}, body: {body}"
-            )
-            result = tool.execute_action(action_name, **body)
-        else:
-            logger.debug(f"Executing tool: {action_name} with args: {call_args}")
-            result = tool.execute_action(action_name, **parameters)
+        try:
+            if tool_data["name"] == "api_tool":
+                logger.debug(
+                    f"Executing api: {action_name} with query_params: {query_params}, headers: {headers}, body: {body}"
+                )
+                result = tool.execute_action(action_name, **body)
+            else:
+                logger.debug(f"Executing tool: {action_name} with args: {call_args}")
+                result = tool.execute_action(action_name, **parameters)
+        except Exception as exc:
+            _mark_failed(call_id, str(exc))
+            raise
 
         get_artifact_id = (
             getattr(tool, "get_artifact_id", None)
@@ -403,6 +548,22 @@ class ToolExecutor:
             f"{result_full[:50]}..." if len(result_full) > 50 else result_full
         )
 
+        # Tool side effect has run; flip the journal row so the
+        # message-finalize path can later confirm it. If the proposed
+        # write failed (DB outage), upsert a fresh row in ``executed`` so
+        # the reconciler still sees the side effect.
+        _mark_executed(
+            call_id,
+            result_full,
+            message_id=self.message_id,
+            artifact_id=artifact_id or None,
+            proposed_ok=proposed_ok,
+            tool_name=tool_data["name"],
+            action_name=action_name,
+            arguments=call_args,
+            tool_id=tool_data.get("id"),
+        )
+
         stream_tool_call_data = {
             key: value
             for key, value in tool_call_data.items()
@@ -451,10 +612,12 @@ class ToolExecutor:
             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,
+                    "tool_missing_row_id",
+                    extra={
+                        "tool_name": tool_data.get("name"),
+                        "tool_id": tool_id,
+                        "action_name": action_name,
+                    },
                 )
                 return None
             tool_config["tool_id"] = str(row_id)

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),
@@ -435,6 +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,
@@ -449,6 +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/postgres.py

@@ -177,3 +177,4 @@ class PostgresTool(Tool):
                 "order": 1,
             },
         }
+

application/agents/tools/tool_action_parser.py

@@ -57,6 +57,29 @@ class ToolActionParser:
     def _parse_google_llm(self, call):
         try:
             call_args = call.arguments
+            # Gemini's SDK natively returns ``args`` as a dict, but the
+            # resume path (``gen_continuation``) stringifies it for the
+            # assistant message. Coerce a JSON string back into a dict;
+            # fall back to an empty dict on malformed input so downstream
+            # ``call_args.items()`` doesn't crash the stream.
+            if isinstance(call_args, str):
+                try:
+                    call_args = json.loads(call_args)
+                except (json.JSONDecodeError, TypeError):
+                    logger.warning(
+                        "Google call.arguments was not valid JSON; "
+                        "falling back to empty args for %s",
+                        getattr(call, "name", "<unknown>"),
+                    )
+                    call_args = {}
+            if not isinstance(call_args, dict):
+                logger.warning(
+                    "Google call.arguments has unexpected type %s; "
+                    "falling back to empty args for %s",
+                    type(call_args).__name__,
+                    getattr(call, "name", "<unknown>"),
+                )
+                call_args = {}
 
             resolved = self._resolve_via_mapping(call.name)
             if resolved:

application/agents/workflows/workflow_engine.py

@@ -211,15 +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
             ):
@@ -232,6 +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,

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/alembic/versions/0004_durability_foundation.py

@@ -0,0 +1,217 @@
+"""0004 durability foundation — idempotency, tool-call log, ingest checkpoint.
+
+Adds ``task_dedup``, ``webhook_dedup``, ``tool_call_attempts``,
+``ingest_chunk_progress``, and per-row status flags on
+``conversation_messages`` and ``pending_tool_state``. Also adds
+``token_usage.source`` and ``token_usage.request_id`` so per-channel
+cost attribution (``agent_stream`` / ``title`` / ``compression`` /
+``rag_condense`` / ``fallback``) is queryable and multi-call agent runs
+can be DISTINCT-collapsed into a single user request for rate limiting.
+
+Revision ID: 0004_durability_foundation
+Revises: 0003_user_custom_models
+"""
+
+from typing import Sequence, Union
+
+from alembic import op
+
+
+revision: str = "0004_durability_foundation"
+down_revision: Union[str, None] = "0003_user_custom_models"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # ------------------------------------------------------------------
+    # New tables
+    # ------------------------------------------------------------------
+    # ``attempt_count`` bounds the per-Celery-task idempotency wrapper's
+    # retry loop so a poison message can't run forever; default 0 means
+    # existing rows behave as if no attempts have run yet.
+    op.execute(
+        """
+        CREATE TABLE task_dedup (
+            idempotency_key TEXT PRIMARY KEY,
+            task_name       TEXT NOT NULL,
+            task_id         TEXT NOT NULL,
+            result_json     JSONB,
+            status          TEXT NOT NULL
+                            CHECK (status IN ('pending', 'completed', 'failed')),
+            attempt_count   INT  NOT NULL DEFAULT 0,
+            created_at      TIMESTAMPTZ NOT NULL DEFAULT now()
+        );
+        """
+    )
+
+    op.execute(
+        """
+        CREATE TABLE webhook_dedup (
+            idempotency_key TEXT PRIMARY KEY,
+            agent_id        UUID NOT NULL,
+            task_id         TEXT NOT NULL,
+            response_json   JSONB,
+            created_at      TIMESTAMPTZ NOT NULL DEFAULT now()
+        );
+        """
+    )
+
+    # FK on ``message_id`` uses ``ON DELETE SET NULL`` so the journal row
+    # survives parent-message deletion (compliance / cost-attribution).
+    op.execute(
+        """
+        CREATE TABLE tool_call_attempts (
+            call_id      TEXT PRIMARY KEY,
+            message_id   UUID
+                         REFERENCES conversation_messages (id)
+                         ON DELETE SET NULL,
+            tool_id      UUID,
+            tool_name    TEXT NOT NULL,
+            action_name  TEXT NOT NULL,
+            arguments    JSONB NOT NULL,
+            result       JSONB,
+            error        TEXT,
+            status       TEXT NOT NULL
+                         CHECK (status IN (
+                             'proposed', 'executed', 'confirmed',
+                             'compensated', 'failed'
+                         )),
+            attempted_at TIMESTAMPTZ NOT NULL DEFAULT now(),
+            updated_at   TIMESTAMPTZ NOT NULL DEFAULT now()
+        );
+        """
+    )
+
+    op.execute(
+        """
+        CREATE TABLE ingest_chunk_progress (
+            source_id        UUID PRIMARY KEY,
+            total_chunks     INT NOT NULL,
+            embedded_chunks  INT NOT NULL DEFAULT 0,
+            last_index       INT NOT NULL DEFAULT -1,
+            last_updated     TIMESTAMPTZ NOT NULL DEFAULT now()
+        );
+        """
+    )
+
+    # ------------------------------------------------------------------
+    # Column additions on existing tables
+    # ------------------------------------------------------------------
+    # DEFAULT 'complete' backfills existing rows — they're already done.
+    op.execute(
+        """
+        ALTER TABLE conversation_messages
+            ADD COLUMN status TEXT NOT NULL DEFAULT 'complete'
+                CHECK (status IN ('pending', 'streaming', 'complete', 'failed')),
+            ADD COLUMN request_id TEXT;
+        """
+    )
+
+    op.execute(
+        """
+        ALTER TABLE pending_tool_state
+            ADD COLUMN status TEXT NOT NULL DEFAULT 'pending'
+                CHECK (status IN ('pending', 'resuming')),
+            ADD COLUMN resumed_at TIMESTAMPTZ;
+        """
+    )
+
+    # Default ``agent_stream`` backfills historical rows under the
+    # assumption they were written from the primary path — pre-fix the
+    # only path that wrote was the error branch reading agent.llm.
+    # ``request_id`` is the stream-scoped UUID stamped by the route on
+    # ``agent.llm`` so multi-tool agent runs (which produce N rows)
+    # collapse to one request via DISTINCT in ``count_in_range``.
+    # Side-channel sources (``title`` / ``compression`` / ``rag_condense``
+    # / ``fallback``) leave it NULL and are excluded from the request
+    # count by source filter.
+    op.execute(
+        """
+        ALTER TABLE token_usage
+            ADD COLUMN source     TEXT NOT NULL DEFAULT 'agent_stream',
+            ADD COLUMN request_id TEXT;
+        """
+    )
+
+    # ------------------------------------------------------------------
+    # Indexes — partial where the predicate selects only non-terminal rows
+    # ------------------------------------------------------------------
+    op.execute(
+        "CREATE INDEX conversation_messages_pending_ts_idx "
+        "ON conversation_messages (timestamp) "
+        "WHERE status IN ('pending', 'streaming');"
+    )
+    op.execute(
+        "CREATE INDEX tool_call_attempts_pending_ts_idx "
+        "ON tool_call_attempts (attempted_at) "
+        "WHERE status IN ('proposed', 'executed');"
+    )
+    op.execute(
+        "CREATE INDEX tool_call_attempts_message_idx "
+        "ON tool_call_attempts (message_id) "
+        "WHERE message_id IS NOT NULL;"
+    )
+    op.execute(
+        "CREATE INDEX pending_tool_state_resuming_ts_idx "
+        "ON pending_tool_state (resumed_at) "
+        "WHERE status = 'resuming';"
+    )
+    op.execute(
+        "CREATE INDEX webhook_dedup_agent_idx "
+        "ON webhook_dedup (agent_id);"
+    )
+    op.execute(
+        "CREATE INDEX task_dedup_pending_attempts_idx "
+        "ON task_dedup (attempt_count) WHERE status = 'pending';"
+    )
+    # Cost-attribution dashboards filter ``token_usage`` by
+    # ``(timestamp, source)``; index the same shape so they stay cheap.
+    op.execute(
+        "CREATE INDEX token_usage_source_ts_idx "
+        "ON token_usage (source, timestamp);"
+    )
+    # Partial index — only rows with a stamped request_id participate
+    # in the DISTINCT count. NULL rows fall through to the COUNT(*)
+    # branch in the repository query.
+    op.execute(
+        "CREATE INDEX token_usage_request_id_idx "
+        "ON token_usage (request_id) "
+        "WHERE request_id IS NOT NULL;"
+    )
+
+    op.execute(
+        "CREATE TRIGGER tool_call_attempts_set_updated_at "
+        "BEFORE UPDATE ON tool_call_attempts "
+        "FOR EACH ROW WHEN (OLD.* IS DISTINCT FROM NEW.*) "
+        "EXECUTE FUNCTION set_updated_at();"
+    )
+
+
+def downgrade() -> None:
+    # CASCADE so the downgrade stays safe if later migrations FK into these.
+    for table in (
+        "ingest_chunk_progress",
+        "tool_call_attempts",
+        "webhook_dedup",
+        "task_dedup",
+    ):
+        op.execute(f"DROP TABLE IF EXISTS {table} CASCADE;")
+
+    op.execute(
+        "ALTER TABLE conversation_messages "
+        "DROP COLUMN IF EXISTS request_id, "
+        "DROP COLUMN IF EXISTS status;"
+    )
+    op.execute(
+        "ALTER TABLE pending_tool_state "
+        "DROP COLUMN IF EXISTS resumed_at, "
+        "DROP COLUMN IF EXISTS status;"
+    )
+    op.execute("DROP INDEX IF EXISTS token_usage_request_id_idx;")
+    op.execute("DROP INDEX IF EXISTS token_usage_source_ts_idx;")
+    op.execute(
+        "ALTER TABLE token_usage "
+        "DROP COLUMN IF EXISTS request_id, "
+        "DROP COLUMN IF EXISTS source;"
+    )

application/alembic/versions/0005_ingest_attempt_id.py

@@ -0,0 +1,44 @@
+"""0005 ingest_chunk_progress.attempt_id — per-attempt resume scoping.
+
+Without this column, a completed checkpoint row poisoned every later
+embed call on the same ``source_id``: a sync after an upload finished
+read the upload's terminal ``last_index`` and either embedded zero
+chunks (if new ``total_docs <= last_index + 1``) or stacked new chunks
+on top of the old vectors (if ``total_docs > last_index + 1``).
+
+``attempt_id`` is stamped from ``self.request.id`` (Celery's stable
+task id, which survives ``acks_late`` retries of the same task but
+differs across separate task invocations). The repository's
+``init_progress`` upsert resets ``last_index`` / ``embedded_chunks``
+when the incoming ``attempt_id`` differs from the stored one — so a
+fresh sync starts from chunk 0 while a retry of the same task resumes
+from the last checkpointed chunk.
+
+Revision ID: 0005_ingest_attempt_id
+Revises: 0004_durability_foundation
+"""
+
+from typing import Sequence, Union
+
+from alembic import op
+
+
+revision: str = "0005_ingest_attempt_id"
+down_revision: Union[str, None] = "0004_durability_foundation"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    op.execute(
+        """
+        ALTER TABLE ingest_chunk_progress
+            ADD COLUMN attempt_id TEXT;
+        """
+    )
+
+
+def downgrade() -> None:
+    op.execute(
+        "ALTER TABLE ingest_chunk_progress DROP COLUMN IF EXISTS attempt_id;"
+    )

application/alembic/versions/0006_idempotency_lease.py

@@ -0,0 +1,57 @@
+"""0006 task_dedup lease columns — running-lease for in-flight tasks.
+
+Without these, ``with_idempotency`` only short-circuits *completed*
+rows. A late-ack redelivery (Redis ``visibility_timeout`` exceeded by a
+long ingest, or a hung-but-alive worker) hands the same message to a
+second worker; ``_claim_or_bump`` only bumped the attempt counter and
+both workers ran the task body in parallel — duplicate vector writes,
+duplicate token spend, duplicate webhook side effects.
+
+``lease_owner_id`` + ``lease_expires_at`` turn that into an atomic
+compare-and-swap. The wrapper claims a lease at entry, refreshes it via
+a 30 s heartbeat thread, and finalises (which makes the lease moot via
+``status='completed'``). A second worker hitting the same key sees a
+fresh lease and ``self.retry(countdown=LEASE_TTL)``s instead of running.
+A crashed worker's lease expires after ``LEASE_TTL`` seconds and the
+next retry can claim it.
+
+Revision ID: 0006_idempotency_lease
+Revises: 0005_ingest_attempt_id
+"""
+
+from typing import Sequence, Union
+
+from alembic import op
+
+
+revision: str = "0006_idempotency_lease"
+down_revision: Union[str, None] = "0005_ingest_attempt_id"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    op.execute(
+        """
+        ALTER TABLE task_dedup
+            ADD COLUMN lease_owner_id   TEXT,
+            ADD COLUMN lease_expires_at TIMESTAMPTZ;
+        """
+    )
+    # Reconciler's stuck-pending sweep filters by
+    # ``(status='pending', lease_expires_at < now() - 60s, attempt_count >= 5)``.
+    # Partial index keeps the scan small even under heavy task throughput.
+    op.execute(
+        "CREATE INDEX task_dedup_pending_lease_idx "
+        "ON task_dedup (lease_expires_at) "
+        "WHERE status = 'pending';"
+    )
+
+
+def downgrade() -> None:
+    op.execute("DROP INDEX IF EXISTS task_dedup_pending_lease_idx;")
+    op.execute(
+        "ALTER TABLE task_dedup "
+        "DROP COLUMN IF EXISTS lease_expires_at, "
+        "DROP COLUMN IF EXISTS lease_owner_id;"
+    )

application/api/answer/routes/answer.py

@@ -102,6 +102,8 @@ class AnswerResource(Resource, BaseAnswerResource):
                         "tools_dict": tools_dict,
                         "pending_tool_calls": pending_tool_calls,
                         "tool_actions": tool_actions,
+                        "reserved_message_id": processor.reserved_message_id,
+                        "request_id": processor.request_id,
                     },
                 )
             else:

application/api/answer/routes/base.py

@@ -1,13 +1,18 @@
 import datetime
 import json
 import logging
+import time
+import uuid
 from typing import Any, Dict, Generator, List, Optional
 
 from flask import jsonify, make_response, Response
 from flask_restx import Namespace
 
 from application.api.answer.services.continuation_service import ContinuationService
-from application.api.answer.services.conversation_service import ConversationService
+from application.api.answer.services.conversation_service import (
+    ConversationService,
+    TERMINATED_RESPONSE_PLACEHOLDER,
+)
 from application.core.model_utils import (
     get_api_key_for_provider,
     get_default_model_id,
@@ -177,6 +182,7 @@ 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]:
         """
@@ -202,13 +208,118 @@ class BaseAnswerResource:
         Yields:
             Server-sent event strings
         """
+        response_full, thought, source_log_docs, tool_calls = "", "", [], []
+        is_structured = False
+        schema_info = None
+        structured_chunks = []
+        query_metadata: Dict[str, Any] = {}
+        paused = False
+
+        # One id shared across the WAL row, primary LLM (token_usage
+        # attribution), the SSE event, and resumed continuations.
+        request_id = (
+            _continuation.get("request_id") if _continuation else None
+        ) or str(uuid.uuid4())
+
+        # Reserve the placeholder row before the LLM call so a crash
+        # mid-stream still leaves the question queryable. Continuations
+        # reuse the original placeholder.
+        reserved_message_id: Optional[str] = None
+        wal_eligible = should_save_conversation and not _continuation
+        if wal_eligible:
+            try:
+                reservation = self.conversation_service.save_user_question(
+                    conversation_id=conversation_id,
+                    question=question,
+                    decoded_token=decoded_token,
+                    attachment_ids=attachment_ids,
+                    api_key=user_api_key,
+                    agent_id=agent_id,
+                    is_shared_usage=is_shared_usage,
+                    shared_token=shared_token,
+                    model_id=model_id or self.default_model_id,
+                    request_id=request_id,
+                    index=index,
+                )
+                conversation_id = reservation["conversation_id"]
+                reserved_message_id = reservation["message_id"]
+            except Exception as e:
+                logger.error(
+                    f"Failed to reserve message row before stream: {e}",
+                    exc_info=True,
+                )
+        elif _continuation and _continuation.get("reserved_message_id"):
+            reserved_message_id = _continuation["reserved_message_id"]
+
+        primary_llm = getattr(agent, "llm", None)
+        if primary_llm is not None:
+            primary_llm._request_id = request_id
+
+        # Flipped to ``streaming`` on first chunk; reconciler uses this
+        # to tell "never started" from "in flight".
+        streaming_marked = False
+        # Heartbeat goes into ``metadata.last_heartbeat_at`` (not
+        # ``updated_at``, which reconciler-side writes share) and uses
+        # ``time.monotonic`` so a blocked event loop can't fake fresh.
+        STREAM_HEARTBEAT_INTERVAL = 60
+        last_heartbeat_at = time.monotonic()
+
+        def _mark_streaming_once() -> None:
+            nonlocal streaming_marked, last_heartbeat_at
+            if streaming_marked or not reserved_message_id:
+                return
+            try:
+                self.conversation_service.update_message_status(
+                    reserved_message_id, "streaming",
+                )
+            except Exception:
+                logger.exception(
+                    "update_message_status streaming failed for %s",
+                    reserved_message_id,
+                )
+            streaming_marked = True
+            last_heartbeat_at = time.monotonic()
+
+        def _heartbeat_streaming() -> None:
+            nonlocal last_heartbeat_at
+            if not reserved_message_id or not streaming_marked:
+                return
+            now_mono = time.monotonic()
+            if now_mono - last_heartbeat_at < STREAM_HEARTBEAT_INTERVAL:
+                return
+            try:
+                self.conversation_service.heartbeat_message(
+                    reserved_message_id,
+                )
+            except Exception:
+                logger.exception(
+                    "stream heartbeat update failed for %s",
+                    reserved_message_id,
+                )
+            last_heartbeat_at = now_mono
+
+        # Correlates tool_call_attempts rows with this message.
+        if reserved_message_id and getattr(agent, "tool_executor", None):
+            try:
+                agent.tool_executor.message_id = reserved_message_id
+            except Exception:
+                pass
+
         try:
-            response_full, thought, source_log_docs, tool_calls = "", "", [], []
-            is_structured = False
-            schema_info = None
-            structured_chunks = []
-            query_metadata = {}
-            paused = False
+            # Surface the placeholder id before any LLM tokens so a
+            # mid-handshake disconnect still has a row to tail-poll.
+            if reserved_message_id:
+                early_event = json.dumps(
+                    {
+                        "type": "message_id",
+                        "message_id": reserved_message_id,
+                        "conversation_id": (
+                            str(conversation_id) if conversation_id else None
+                        ),
+                        "request_id": request_id,
+                    }
+                )
+                yield f"data: {early_event}\n\n"
 
             if _continuation:
                 gen_iter = agent.gen_continuation(
@@ -221,9 +332,13 @@ class BaseAnswerResource:
                 gen_iter = agent.gen(query=question)
 
             for line in gen_iter:
+                # Cheap closure check that only hits the DB when the
+                # heartbeat interval has elapsed.
+                _heartbeat_streaming()
                 if "metadata" in line:
                     query_metadata.update(line["metadata"])
                 elif "answer" in line:
+                    _mark_streaming_once()
                     response_full += str(line["answer"])
                     if line.get("structured"):
                         is_structured = True
@@ -233,6 +348,7 @@ class BaseAnswerResource:
                         data = json.dumps({"type": "answer", "answer": line["answer"]})
                         yield f"data: {data}\n\n"
                 elif "sources" in line:
+                    _mark_streaming_once()
                     truncated_sources = []
                     source_log_docs = line["sources"]
                     for source in line["sources"]:
@@ -285,12 +401,19 @@ class BaseAnswerResource:
             if paused:
                 continuation = getattr(agent, "_pending_continuation", None)
                 if continuation:
-                    # Ensure we have a conversation_id — create a partial
-                    # conversation if this is the first turn.
+                    # First-turn pause needs a conversation row to attach to.
                     if not conversation_id and should_save_conversation:
                         try:
                             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
                             )
@@ -304,6 +427,7 @@ 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(
@@ -340,6 +464,9 @@ class BaseAnswerResource:
                                 tool_schemas=getattr(agent, "tools", []),
                                 agent_config={
                                     "model_id": model_id or self.default_model_id,
+                                    # BYOM scope; without it resume falls
+                                    # 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,
@@ -348,6 +475,11 @@ class BaseAnswerResource:
                                     "prompt": getattr(agent, "prompt", ""),
                                     "json_schema": getattr(agent, "json_schema", None),
                                     "retriever_config": getattr(agent, "retriever_config", None),
+                                    # Reused on resume so the same WAL row
+                                    # is finalised and request_id stays
+                                    # consistent across token_usage rows.
+                                    "reserved_message_id": reserved_message_id,
+                                    "request_id": request_id,
                                 },
                                 client_tools=getattr(
                                     agent.tool_executor, "client_tools", None
@@ -370,8 +502,13 @@ class BaseAnswerResource:
             if isNoneDoc:
                 for doc in source_log_docs:
                     doc["source"] = "None"
+            # Model-owner scope so title-gen uses owner's BYOM 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
             )
@@ -384,27 +521,51 @@ class BaseAnswerResource:
                 decoded_token=decoded_token,
                 model_id=model_id,
                 agent_id=agent_id,
+                model_user_id=model_user_id,
             )
+            # Title-gen only; agent stream tokens live on ``agent.llm``.
+            llm._token_usage_source = "title"
 
             if should_save_conversation:
-                conversation_id = self.conversation_service.save_conversation(
-                    conversation_id,
-                    question,
-                    response_full,
-                    thought,
-                    source_log_docs,
-                    tool_calls,
-                    llm,
-                    model_id or self.default_model_id,
-                    decoded_token,
-                    index=index,
-                    api_key=user_api_key,
-                    agent_id=agent_id,
-                    is_shared_usage=is_shared_usage,
-                    shared_token=shared_token,
-                    attachment_ids=attachment_ids,
-                    metadata=query_metadata if query_metadata else None,
-                )
+                if reserved_message_id is not None:
+                    self.conversation_service.finalize_message(
+                        reserved_message_id,
+                        response_full,
+                        thought=thought,
+                        sources=source_log_docs,
+                        tool_calls=tool_calls,
+                        model_id=model_id or self.default_model_id,
+                        metadata=query_metadata if query_metadata else None,
+                        status="complete",
+                        title_inputs={
+                            "llm": llm,
+                            "question": question,
+                            "response": response_full,
+                            "model_id": model_id or self.default_model_id,
+                            "fallback_name": (
+                                question[:50] if question else "New Conversation"
+                            ),
+                        },
+                    )
+                else:
+                    conversation_id = self.conversation_service.save_conversation(
+                        conversation_id,
+                        question,
+                        response_full,
+                        thought,
+                        source_log_docs,
+                        tool_calls,
+                        llm,
+                        model_id or self.default_model_id,
+                        decoded_token,
+                        index=index,
+                        api_key=user_api_key,
+                        agent_id=agent_id,
+                        is_shared_usage=is_shared_usage,
+                        shared_token=shared_token,
+                        attachment_ids=attachment_ids,
+                        metadata=query_metadata if query_metadata else None,
+                    )
                 # Persist compression metadata/summary if it exists and wasn't saved mid-execution
                 compression_meta = getattr(agent, "compression_metadata", None)
                 compression_saved = getattr(agent, "compression_saved", False)
@@ -427,6 +588,21 @@ class BaseAnswerResource:
                         )
             else:
                 conversation_id = None
+            # Resume finished cleanly; drop the continuation row.
+            # Crash-paths leave it ``resuming`` for the janitor to revert.
+            if _continuation and conversation_id:
+                try:
+                    cont_service = ContinuationService()
+                    cont_service.delete_state(
+                        str(conversation_id),
+                        decoded_token.get("sub", "local"),
+                    )
+                except Exception as e:
+                    logger.error(
+                        f"Failed to delete continuation state on resume "
+                        f"completion: {e}",
+                        exc_info=True,
+                    )
             id_data = {"type": "id", "id": str(conversation_id)}
             data = json.dumps(id_data)
             yield f"data: {data}\n\n"
@@ -481,31 +657,73 @@ class BaseAnswerResource:
                     if isNoneDoc:
                         for doc in source_log_docs:
                             doc["source"] = "None"
+                    # Resolve under model-owner scope so shared-agent
+                    # title-gen uses owner BYOM, not deployment default.
+                    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,
-                        question,
-                        response_full,
-                        thought,
-                        source_log_docs,
-                        tool_calls,
-                        llm,
-                        model_id or self.default_model_id,
-                        decoded_token,
-                        index=index,
-                        api_key=user_api_key,
-                        agent_id=agent_id,
-                        is_shared_usage=is_shared_usage,
-                        shared_token=shared_token,
-                        attachment_ids=attachment_ids,
-                        metadata=query_metadata if query_metadata else None,
-                    )
+                    llm._token_usage_source = "title"
+                    if reserved_message_id is not None:
+                        self.conversation_service.finalize_message(
+                            reserved_message_id,
+                            response_full,
+                            thought=thought,
+                            sources=source_log_docs,
+                            tool_calls=tool_calls,
+                            model_id=model_id or self.default_model_id,
+                            metadata=query_metadata if query_metadata else None,
+                            status="complete",
+                            title_inputs={
+                                "llm": llm,
+                                "question": question,
+                                "response": response_full,
+                                "model_id": model_id or self.default_model_id,
+                                "fallback_name": (
+                                    question[:50] if question else "New Conversation"
+                                ),
+                            },
+                        )
+                    else:
+                        self.conversation_service.save_conversation(
+                            conversation_id,
+                            question,
+                            response_full,
+                            thought,
+                            source_log_docs,
+                            tool_calls,
+                            llm,
+                            model_id or self.default_model_id,
+                            decoded_token,
+                            index=index,
+                            api_key=user_api_key,
+                            agent_id=agent_id,
+                            is_shared_usage=is_shared_usage,
+                            shared_token=shared_token,
+                            attachment_ids=attachment_ids,
+                            metadata=query_metadata if query_metadata else None,
+                        )
                     compression_meta = getattr(agent, "compression_metadata", None)
                     compression_saved = getattr(agent, "compression_saved", False)
                     if conversation_id and compression_meta and not compression_saved:
@@ -532,6 +750,24 @@ class BaseAnswerResource:
             raise
         except Exception as e:
             logger.error(f"Error in stream: {str(e)}", exc_info=True)
+            if reserved_message_id is not None:
+                try:
+                    self.conversation_service.finalize_message(
+                        reserved_message_id,
+                        response_full or TERMINATED_RESPONSE_PLACEHOLDER,
+                        thought=thought,
+                        sources=source_log_docs,
+                        tool_calls=tool_calls,
+                        model_id=model_id or self.default_model_id,
+                        metadata=query_metadata if query_metadata else None,
+                        status="failed",
+                        error=e,
+                    )
+                except Exception as fin_err:
+                    logger.error(
+                        f"Failed to finalize errored message: {fin_err}",
+                        exc_info=True,
+                    )
             data = json.dumps(
                 {
                     "type": "error",

application/api/answer/routes/search.py

@@ -1,21 +1,21 @@
 import logging
-from typing import Any, Dict, List
 
 from flask import make_response, request
 from flask_restx import fields, Resource
 
 from application.api.answer.routes.base import answer_ns
-from application.core.settings import settings
-from application.storage.db.repositories.agents import AgentsRepository
-from application.storage.db.session import db_readonly
-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"""
+    """Fast search endpoint for retrieving relevant documents."""
 
     search_model = answer_ns.model(
         "SearchModel",
@@ -32,102 +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."""
-        with db_readonly() as conn:
-            agent_data = AgentsRepository(conn).find_by_key(api_key)
-        if not agent_data:
-            return []
-
-        source_ids: List[str] = []
-        # extra_source_ids is a PG ARRAY(UUID) of source UUIDs.
-        extra = agent_data.get("extra_source_ids") or []
-        for src in extra:
-            if src:
-                source_ids.append(str(src))
-
-        if not source_ids:
-            single = agent_data.get("source_id")
-            if single:
-                source_ids.append(str(single))
-
-        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")
@@ -135,32 +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
-        with db_readonly() as conn:
-            agent = AgentsRepository(conn).find_by_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

@@ -109,11 +109,14 @@ 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,
                             "pending_tool_calls": pending_tool_calls,
                             "tool_actions": tool_actions,
+                            "reserved_message_id": processor.reserved_message_id,
+                            "request_id": processor.request_id,
                         },
                     ),
                     mimetype="text/event-stream",
@@ -145,6 +148,7 @@ 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,
@@ -135,7 +158,11 @@ class CompressionOrchestrator:
                 decoded_token=decoded_token,
                 model_id=compression_model,
                 agent_id=conversation.get("agent_id"),
+                model_user_id=registry_user_id,
             )
+            # Side-channel LLM tag — distinguishes compression rows
+            # from primary stream rows for cost-attribution dashboards.
+            compression_llm._token_usage_source = "compression"
 
             # Create compression service with DB update capability
             compression_service = CompressionService(
@@ -167,9 +194,12 @@ 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 +222,21 @@ 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 +258,12 @@ 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/compression/token_counter.py

@@ -12,6 +12,12 @@ logger = logging.getLogger(__name__)
 class TokenCounter:
     """Centralized token counting for conversations and messages."""
 
+    # Per-image token estimate. Provider tokenizers vary widely
+    # (Gemini ~258, GPT-4o 85-1500, Claude ~1500) and the actual cost
+    # depends on resolution/detail we can't see here. Errs slightly high
+    # so the threshold check stays conservative.
+    _IMAGE_PART_TOKEN_ESTIMATE = 1500
+
     @staticmethod
     def count_message_tokens(messages: List[Dict]) -> int:
         """
@@ -29,12 +35,36 @@ class TokenCounter:
             if isinstance(content, str):
                 total_tokens += num_tokens_from_string(content)
             elif isinstance(content, list):
-                # Handle structured content (tool calls, etc.)
+                # Handle structured content (tool calls, image parts, etc.)
                 for item in content:
                     if isinstance(item, dict):
-                        total_tokens += num_tokens_from_string(str(item))
+                        total_tokens += TokenCounter._count_content_part(item)
         return total_tokens
 
+    @staticmethod
+    def _count_content_part(item: Dict) -> int:
+        # Image/file attachments are billed by the provider per image,
+        # not proportional to the inline bytes/base64 string.
+        # ``str(item)`` on a 1MB image inflates the count by ~10000x,
+        # which trips spurious compression and overflows downstream
+        # input limits.
+        item_type = item.get("type")
+
+        if "files" in item:
+            files = item.get("files")
+            count = len(files) if isinstance(files, list) and files else 1
+            return TokenCounter._IMAGE_PART_TOKEN_ESTIMATE * count
+
+        if "image_url" in item or item_type in {
+            "image",
+            "image_url",
+            "input_image",
+            "file",
+        }:
+            return TokenCounter._IMAGE_PART_TOKEN_ESTIMATE
+
+        return num_tokens_from_string(str(item))
+
     @staticmethod
     def count_query_tokens(
         queries: List[Dict[str, Any]], include_tool_calls: bool = True

application/api/answer/services/continuation_service.py

@@ -7,13 +7,13 @@ resume later by sending tool_actions.
 
 import logging
 from typing import Any, Dict, List, Optional
-from uuid import UUID
 
 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.serialization import coerce_pg_native as _make_serializable
 from application.storage.db.session import db_readonly, db_session
 
 logger = logging.getLogger(__name__)
@@ -21,23 +21,9 @@ logger = logging.getLogger(__name__)
 # TTL for pending states — auto-cleaned after this period
 PENDING_STATE_TTL_SECONDS = 30 * 60  # 30 minutes
 
-
-def _make_serializable(obj: Any) -> Any:
-    """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()}
-    if isinstance(obj, list):
-        return [_make_serializable(v) for v in obj]
-    if isinstance(obj, bytes):
-        return obj.decode("utf-8", errors="replace")
-    return obj
+# Re-export so the existing tests at tests/api/answer/services/test_continuation_service_pg.py
+# can keep importing ``_make_serializable`` from here.
+__all__ = ["_make_serializable", "ContinuationService", "PENDING_STATE_TTL_SECONDS"]
 
 
 class ContinuationService:
@@ -155,3 +141,23 @@ class ContinuationService:
                 f"Deleted continuation state for conversation {conversation_id}"
             )
         return deleted
+
+    def mark_resuming(self, conversation_id: str, user: str) -> bool:
+        """Flip the pending row to ``resuming`` so a crashed resume can be retried."""
+        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:
+                return False
+            flipped = PendingToolStateRepository(conn).mark_resuming(
+                pg_conv_id, user
+            )
+        if flipped:
+            logger.info(
+                f"Marked continuation state as resuming for conversation "
+                f"{conversation_id}"
+            )
+        return flipped

application/api/answer/services/conversation_service.py

@@ -6,6 +6,7 @@ than held for the duration of a stream.
 """
 
 import logging
+import uuid
 from datetime import datetime, timezone
 from typing import Any, Dict, List, Optional
 
@@ -21,6 +22,12 @@ from application.storage.db.session import db_readonly, db_session
 logger = logging.getLogger(__name__)
 
 
+# Shown to the user if the worker dies mid-stream and the response is never finalised.
+TERMINATED_RESPONSE_PLACEHOLDER = (
+    "Response was terminated prior to completion, try regenerating."
+)
+
+
 class ConversationService:
     def get_conversation(
         self, conversation_id: str, user_id: str
@@ -136,8 +143,14 @@ 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():
@@ -173,6 +186,236 @@ class ConversationService:
                 repo.append_message(conv_pg_id, append_payload)
             return conv_pg_id
 
+    def save_user_question(
+        self,
+        conversation_id: Optional[str],
+        question: str,
+        decoded_token: Dict[str, Any],
+        *,
+        attachment_ids: Optional[List[str]] = None,
+        api_key: Optional[str] = None,
+        agent_id: Optional[str] = None,
+        is_shared_usage: bool = False,
+        shared_token: Optional[str] = None,
+        model_id: Optional[str] = None,
+        request_id: Optional[str] = None,
+        status: str = "pending",
+        index: Optional[int] = None,
+    ) -> Dict[str, str]:
+        """Reserve the placeholder message row before the LLM call.
+
+        ``index`` triggers regenerate semantics: messages at
+        ``position >= index`` are truncated so the new placeholder
+        lands at ``position = index`` rather than appending.
+
+        Returns ``{"conversation_id", "message_id", "request_id"}``.
+        """
+        if decoded_token is None:
+            raise ValueError("Invalid or missing authentication token")
+        user_id = decoded_token.get("sub")
+        if not user_id:
+            raise ValueError("User ID not found in token")
+
+        request_id = request_id or str(uuid.uuid4())
+
+        resolved_api_key: Optional[str] = None
+        resolved_agent_id: Optional[str] = None
+        if api_key and not conversation_id:
+            with db_readonly() as conn:
+                agent = AgentsRepository(conn).find_by_key(api_key)
+            if agent:
+                resolved_api_key = agent.get("key")
+            if agent_id:
+                resolved_agent_id = agent_id
+
+        with db_session() as conn:
+            repo = ConversationsRepository(conn)
+            if conversation_id:
+                conv = repo.get_any(conversation_id, user_id)
+                if conv is None:
+                    raise ValueError("Conversation not found or unauthorized")
+                conv_pg_id = str(conv["id"])
+                # Regenerate / edit-prior-question: drop the message at
+                # ``index`` and everything after it so the new
+                # ``reserve_message`` lands at ``position=index`` rather
+                # than appending at the end of the conversation.
+                if isinstance(index, int) and index >= 0:
+                    repo.truncate_after(conv_pg_id, keep_up_to=index - 1)
+            else:
+                fallback_name = (question[:50] if question else "New Conversation")
+                conv = repo.create(
+                    user_id,
+                    fallback_name,
+                    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
+                    ),
+                )
+                conv_pg_id = str(conv["id"])
+
+            row = repo.reserve_message(
+                conv_pg_id,
+                prompt=question,
+                placeholder_response=TERMINATED_RESPONSE_PLACEHOLDER,
+                request_id=request_id,
+                status=status,
+                attachments=attachment_ids,
+                model_id=model_id,
+            )
+            message_id = str(row["id"])
+
+        return {
+            "conversation_id": conv_pg_id,
+            "message_id": message_id,
+            "request_id": request_id,
+        }
+
+    def update_message_status(self, message_id: str, status: str) -> bool:
+        """Cheap status-only transition (e.g. ``pending → streaming``)."""
+        if not message_id:
+            return False
+        with db_session() as conn:
+            return ConversationsRepository(conn).update_message_status(
+                message_id, status,
+            )
+
+    def heartbeat_message(self, message_id: str) -> bool:
+        """Bump ``message_metadata.last_heartbeat_at`` so the reconciler's
+        staleness sweep counts the row as alive. No-ops on terminal rows.
+        """
+        if not message_id:
+            return False
+        with db_session() as conn:
+            return ConversationsRepository(conn).heartbeat_message(message_id)
+
+    def finalize_message(
+        self,
+        message_id: str,
+        response: str,
+        *,
+        thought: str = "",
+        sources: Optional[List[Dict[str, Any]]] = None,
+        tool_calls: Optional[List[Dict[str, Any]]] = None,
+        model_id: Optional[str] = None,
+        metadata: Optional[Dict[str, Any]] = None,
+        status: str = "complete",
+        error: Optional[BaseException] = None,
+        title_inputs: Optional[Dict[str, Any]] = None,
+    ) -> bool:
+        """Commit the response and tool_call confirms in one transaction."""
+        if not message_id:
+            return False
+        sources = sources or []
+        for source in sources:
+            if "text" in source and isinstance(source["text"], str):
+                source["text"] = source["text"][:1000]
+
+        merged_metadata: Dict[str, Any] = dict(metadata or {})
+        if status == "failed" and error is not None:
+            merged_metadata.setdefault(
+                "error", f"{type(error).__name__}: {str(error)}"
+            )
+
+        update_fields: Dict[str, Any] = {
+            "response": response,
+            "status": status,
+            "thought": thought,
+            "sources": sources,
+            "tool_calls": tool_calls or [],
+            "metadata": merged_metadata,
+        }
+        if model_id is not None:
+            update_fields["model_id"] = model_id
+
+        # Atomic message update + tool_call_attempts confirm; the
+        # ``only_if_non_terminal`` guard prevents a late stream from
+        # retracting a row the reconciler already escalated.
+        with db_session() as conn:
+            repo = ConversationsRepository(conn)
+            ok = repo.update_message_by_id(
+                message_id, update_fields,
+                only_if_non_terminal=True,
+            )
+            if not ok:
+                logger.warning(
+                    f"finalize_message: no row updated for message_id={message_id} "
+                    f"(possibly already terminal — reconciler may have escalated)"
+                )
+                return False
+            repo.confirm_executed_tool_calls(message_id)
+
+        # Outside the txn — title-gen is a multi-second LLM round trip.
+        if title_inputs and status == "complete":
+            try:
+                with db_session() as conn:
+                    self._maybe_generate_title(conn, message_id, title_inputs)
+            except Exception as e:
+                logger.error(
+                    f"finalize_message title generation failed: {e}",
+                    exc_info=True,
+                )
+        return True
+
+    def _maybe_generate_title(
+        self,
+        conn,
+        message_id: str,
+        title_inputs: Dict[str, Any],
+    ) -> None:
+        """Generate an LLM-summarised conversation name if one isn't set yet."""
+        llm = title_inputs.get("llm")
+        question = title_inputs.get("question") or ""
+        response = title_inputs.get("response") or ""
+        fallback_name = title_inputs.get("fallback_name") or question[:50]
+        if llm is None:
+            return
+
+        row = conn.execute(
+            sql_text(
+                "SELECT c.id, c.name FROM conversation_messages m "
+                "JOIN conversations c ON c.id = m.conversation_id "
+                "WHERE m.id = CAST(:mid AS uuid)"
+            ),
+            {"mid": message_id},
+        ).fetchone()
+        if row is None:
+            return
+        conv_id, current_name = str(row[0]), row[1]
+        if current_name and current_name != fallback_name:
+            return
+
+        messages_summary = [
+            {
+                "role": "system",
+                "content": "You are a helpful assistant that creates concise conversation titles. "
+                "Summarize conversations in 3 words or less using the same language as the user.",
+            },
+            {
+                "role": "user",
+                "content": "Summarise following conversation in no more than 3 words, "
+                "respond ONLY with the summary, use the same language as the "
+                "user query \n\nUser: " + question + "\n\n" + "AI: " + response,
+            },
+        ]
+        completion = llm.gen(
+            model=getattr(llm, "model_id", None) or title_inputs.get("model_id"),
+            messages=messages_summary,
+            max_tokens=500,
+        )
+        if not completion or not completion.strip():
+            completion = fallback_name or "New Conversation"
+        conn.execute(
+            sql_text(
+                "UPDATE conversations SET name = :name, updated_at = now() "
+                "WHERE id = CAST(:id AS uuid)"
+            ),
+            {"id": conv_id, "name": completion.strip()},
+        )
+
     def update_compression_metadata(
         self, conversation_id: str, compression_metadata: Dict[str, Any]
     ) -> None:

application/api/answer/services/stream_processor.py

@@ -121,6 +121,12 @@ 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
+        # WAL placeholder id pulled from continuation state on resume.
+        self.reserved_message_id: Optional[str] = None
+        # Carried through resumes so multi-pause runs keep one request_id.
+        self.request_id: Optional[str] = None
         self.conversation_service = ConversationService()
         self.compression_orchestrator = CompressionOrchestrator(
             self.conversation_service
@@ -191,16 +197,23 @@ 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,
             )
@@ -284,11 +297,18 @@ 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])}"
@@ -299,12 +319,17 @@ 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."""
@@ -514,6 +539,10 @@ 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"),
                 }
             )
 
@@ -561,7 +590,13 @@ 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"
@@ -612,6 +647,7 @@ 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,
@@ -896,6 +932,20 @@ class StreamProcessor:
         if not state:
             raise ValueError("No pending tool state found for this conversation")
 
+        # Claim the resume up-front. ``mark_resuming`` only flips ``pending``
+        # → ``resuming``; if it returns False, another resume already
+        # claimed this row (status='resuming') — bail before any further
+        # LLM/tool work to avoid double-execution. The cleanup janitor
+        # reverts a stale ``resuming`` claim back to ``pending`` after the
+        # 10-minute grace window so the user can retry.
+        if not cont_service.mark_resuming(
+            conversation_id, self.initial_user_id,
+        ):
+            raise ValueError(
+                "Resume already in progress for this conversation; "
+                "retry after the grace window if it stalls."
+            )
+
         messages = state["messages"]
         pending_tool_calls = state["pending_tool_calls"]
         tools_dict = state["tools_dict"]
@@ -903,6 +953,11 @@ 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")
@@ -920,6 +975,7 @@ 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(
@@ -949,6 +1005,7 @@ 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,
@@ -971,12 +1028,22 @@ 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
-
-        # Delete state so it can't be replayed
-        cont_service.delete_state(conversation_id, self.initial_user_id)
+        # Reused on resume so the same WAL row gets finalised and
+        # request_id stays consistent across token_usage rows.
+        self.reserved_message_id = agent_config.get("reserved_message_id")
+        self.request_id = agent_config.get("request_id")
 
         return agent, messages, tools_dict, pending_tool_calls, tool_actions
 
@@ -1022,8 +1089,11 @@ 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
         )
@@ -1048,6 +1118,8 @@ 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"
@@ -1070,6 +1142,7 @@ 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"],
@@ -1097,6 +1170,7 @@ 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/user/agents/routes.py

@@ -46,7 +46,9 @@ AGENT_TYPE_SCHEMAS = {
             "prompt_id",
         ],
         "required_draft": ["name"],
-        "validate_published": ["name", "description", "prompt_id"],
+        # ``prompt_id`` intentionally omitted — the "default" sentinel
+        # is acceptable and maps to NULL downstream.
+        "validate_published": ["name", "description"],
         "validate_draft": [],
         "require_source": True,
         "fields": [
@@ -1009,12 +1011,16 @@ class UpdateAgent(Resource):
                                 400,
                             )
                     else:
+                        # ``prompt_id`` is intentionally omitted: the
+                        # frontend's "default" choice maps to NULL here
+                        # (see the prompt_id branch above), and NULL
+                        # means "use the built-in default prompt" which
+                        # is a valid published-agent state.
                         missing_published_fields = []
                         for req_field, field_label in (
                             ("name", "Agent name"),
                             ("description", "Agent description"),
                             ("chunks", "Chunks count"),
-                            ("prompt_id", "Prompt"),
                             ("agent_type", "Agent type"),
                         ):
                             final_value = update_fields.get(
@@ -1028,8 +1034,23 @@ class UpdateAgent(Resource):
                         extra_final = update_fields.get(
                             "extra_source_ids", existing_agent.get("extra_source_ids") or [],
                         )
-                        if not source_final and not extra_final:
-                            missing_published_fields.append("Source")
+                        # ``retriever`` carries the runtime identity for
+                        # agents that publish against the synthetic
+                        # "Default" source (frontend's auto-selected
+                        # ``{name: "Default", retriever: "classic"}``
+                        # entry has no ``id``, so ``source_id`` ends up
+                        # NULL even though the user picked something).
+                        # Without this fallback the most common new-agent
+                        # publish flow gets a 400.
+                        retriever_final = update_fields.get(
+                            "retriever", existing_agent.get("retriever"),
+                        )
+                        if (
+                            not source_final
+                            and not extra_final
+                            and not retriever_final
+                        ):
+                            missing_published_fields.append("Source or retriever")
                         if missing_published_fields:
                             return make_response(
                                 jsonify(

application/api/user/agents/webhooks.py

@@ -1,15 +1,19 @@
 """Agent management webhook handlers."""
 
 import secrets
+import uuid
 
 from flask import current_app, jsonify, make_response, request
 from flask_restx import Namespace, Resource
+from sqlalchemy import text as sql_text
 
 from application.api import api
 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.base_repository import looks_like_uuid
 from application.storage.db.repositories.agents import AgentsRepository
+from application.storage.db.repositories.idempotency import IdempotencyRepository
 from application.storage.db.session import db_readonly, db_session
 
 
@@ -18,6 +22,37 @@ agents_webhooks_ns = Namespace(
 )
 
 
+_IDEMPOTENCY_KEY_MAX_LEN = 256
+
+
+def _read_idempotency_key():
+    """Return (key, error_response). Empty header → (None, None); oversized → (None, 400)."""
+    key = request.headers.get("Idempotency-Key")
+    if not key:
+        return None, None
+    if len(key) > _IDEMPOTENCY_KEY_MAX_LEN:
+        return None, make_response(
+            jsonify(
+                {
+                    "success": False,
+                    "message": (
+                        f"Idempotency-Key exceeds maximum length of "
+                        f"{_IDEMPOTENCY_KEY_MAX_LEN} characters"
+                    ),
+                }
+            ),
+            400,
+        )
+    return key, None
+
+
+def _scoped_idempotency_key(idempotency_key, scope):
+    """``{scope}:{key}`` so different agents can't collide on the same key."""
+    if not idempotency_key or not scope:
+        return None
+    return f"{scope}:{idempotency_key}"
+
+
 @agents_webhooks_ns.route("/agent_webhook")
 class AgentWebhook(Resource):
     @api.doc(
@@ -68,7 +103,7 @@ class AgentWebhook(Resource):
 class AgentWebhookListener(Resource):
     method_decorators = [require_agent]
 
-    def _enqueue_webhook_task(self, agent_id_str, payload, source_method):
+    def _enqueue_webhook_task(self, agent_id_str, payload, source_method, agent=None):
         if not payload:
             current_app.logger.warning(
                 f"Webhook ({source_method}) received for agent {agent_id_str} with empty payload."
@@ -77,26 +112,94 @@ class AgentWebhookListener(Resource):
             f"Incoming {source_method} webhook for agent {agent_id_str}. Enqueuing task with payload: {payload}"
         )
 
-        try:
-            task = process_agent_webhook.delay(
-                agent_id=agent_id_str,
-                payload=payload,
+        idempotency_key, key_error = _read_idempotency_key()
+        if key_error is not None:
+            return key_error
+        # Resolve to PG UUID first so dedup writes don't crash on legacy ids.
+        agent_uuid = None
+        if agent is not None:
+            candidate = str(agent.get("id") or "")
+            if looks_like_uuid(candidate):
+                agent_uuid = candidate
+        if idempotency_key and agent_uuid is None:
+            current_app.logger.warning(
+                "Skipping webhook idempotency dedup: agent %s has non-UUID id",
+                agent_id_str,
             )
+            idempotency_key = None
+        # Agent-scoped (webhooks have no user_id).
+        scoped_key = _scoped_idempotency_key(idempotency_key, agent_uuid)
+        # Claim before enqueue; the loser returns the winner's task_id.
+        predetermined_task_id = None
+        if scoped_key:
+            predetermined_task_id = str(uuid.uuid4())
+            with db_session() as conn:
+                claimed = IdempotencyRepository(conn).record_webhook(
+                    key=scoped_key,
+                    agent_id=agent_uuid,
+                    task_id=predetermined_task_id,
+                    response_json={
+                        "success": True, "task_id": predetermined_task_id,
+                    },
+                )
+            if claimed is None:
+                with db_readonly() as conn:
+                    cached = IdempotencyRepository(conn).get_webhook(scoped_key)
+                if cached is not None:
+                    return make_response(jsonify(cached["response_json"]), 200)
+                return make_response(
+                    jsonify({"success": True, "task_id": "deduplicated"}), 200
+                )
+
+        try:
+            apply_kwargs = dict(
+                kwargs={
+                    "agent_id": agent_id_str,
+                    "payload": payload,
+                    # Scoped so the worker dedup row matches the HTTP claim.
+                    "idempotency_key": scoped_key or idempotency_key,
+                },
+            )
+            if predetermined_task_id is not None:
+                apply_kwargs["task_id"] = predetermined_task_id
+            task = process_agent_webhook.apply_async(**apply_kwargs)
             current_app.logger.info(
                 f"Task {task.id} enqueued for agent {agent_id_str} ({source_method})."
             )
-            return make_response(jsonify({"success": True, "task_id": task.id}), 200)
+            response_payload = {"success": True, "task_id": task.id}
+            return make_response(jsonify(response_payload), 200)
         except Exception as err:
             current_app.logger.error(
                 f"Error enqueuing webhook task ({source_method}) for agent {agent_id_str}: {err}",
                 exc_info=True,
             )
+            if scoped_key:
+                # Roll back the claim so a retry can succeed.
+                try:
+                    with db_session() as conn:
+                        conn.execute(
+                            sql_text(
+                                "DELETE FROM webhook_dedup "
+                                "WHERE idempotency_key = :k"
+                            ),
+                            {"k": scoped_key},
+                        )
+                except Exception:
+                    current_app.logger.exception(
+                        "Failed to release webhook_dedup claim for key=%s",
+                        scoped_key,
+                    )
             return make_response(
                 jsonify({"success": False, "message": "Error processing webhook"}), 500
             )
 
     @api.doc(
-        description="Webhook listener for agent events (POST). Expects JSON payload, which is used to trigger processing.",
+        description=(
+            "Webhook listener for agent events (POST). Expects JSON payload, which "
+            "is used to trigger processing. Honors an optional ``Idempotency-Key`` "
+            "header: a repeat request with the same key within 24h returns the "
+            "original cached response and does not re-enqueue the task."
+        ),
     )
     def post(self, webhook_token, agent, agent_id_str):
         payload = request.get_json()
@@ -110,11 +213,20 @@ class AgentWebhookListener(Resource):
                 ),
                 400,
             )
-        return self._enqueue_webhook_task(agent_id_str, payload, source_method="POST")
+        return self._enqueue_webhook_task(
+            agent_id_str, payload, source_method="POST", agent=agent,
+        )
 
     @api.doc(
-        description="Webhook listener for agent events (GET). Uses URL query parameters as payload to trigger processing.",
+        description=(
+            "Webhook listener for agent events (GET). Uses URL query parameters as "
+            "payload to trigger processing. Honors an optional ``Idempotency-Key`` "
+            "header: a repeat request with the same key within 24h returns the "
+            "original cached response and does not re-enqueue the task."
+        ),
     )
     def get(self, webhook_token, agent, agent_id_str):
         payload = request.args.to_dict(flat=True)
-        return self._enqueue_webhook_task(agent_id_str, payload, source_method="GET")
+        return self._enqueue_webhook_task(
+            agent_id_str, payload, source_method="GET", agent=agent,
+        )

application/api/user/conversations/routes.py

@@ -4,8 +4,10 @@ import datetime
 
 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.storage.db.base_repository import looks_like_uuid, row_to_dict
 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
@@ -133,6 +135,7 @@ class GetSingleConversation(Resource):
                 attachments_repo = AttachmentsRepository(conn)
                 queries = []
                 for msg in messages:
+                    metadata = msg.get("metadata") or {}
                     query = {
                         "prompt": msg.get("prompt"),
                         "response": msg.get("response"),
@@ -141,9 +144,15 @@ class GetSingleConversation(Resource):
                         "tool_calls": msg.get("tool_calls") or [],
                         "timestamp": msg.get("timestamp"),
                         "model_id": msg.get("model_id"),
+                        # Lets the client distinguish placeholder rows from
+                        # finalised answers and tail-poll in-flight ones.
+                        "message_id": str(msg["id"]) if msg.get("id") else None,
+                        "status": msg.get("status"),
+                        "request_id": msg.get("request_id"),
+                        "last_heartbeat_at": metadata.get("last_heartbeat_at"),
                     }
-                    if msg.get("metadata"):
-                        query["metadata"] = msg["metadata"]
+                    if metadata:
+                        query["metadata"] = 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
@@ -301,3 +310,61 @@ class SubmitFeedback(Resource):
             current_app.logger.error(f"Error submitting feedback: {err}", exc_info=True)
             return make_response(jsonify({"success": False}), 400)
         return make_response(jsonify({"success": True}), 200)
+
+
+@conversations_ns.route("/messages/<string:message_id>/tail")
+class GetMessageTail(Resource):
+    @api.doc(
+        description=(
+            "Current state of one conversation_messages row, scoped to the "
+            "authenticated user. Used to reconnect to an in-flight stream "
+            "after a refresh."
+        ),
+        params={"message_id": "Message UUID"},
+    )
+    def get(self, message_id):
+        decoded_token = request.decoded_token
+        if not decoded_token:
+            return make_response(jsonify({"success": False}), 401)
+        if not looks_like_uuid(message_id):
+            return make_response(
+                jsonify({"success": False, "message": "Invalid message id"}), 400
+            )
+        user_id = decoded_token.get("sub")
+        try:
+            with db_readonly() as conn:
+                # Owner-or-shared, matching ``ConversationsRepository.get``.
+                row = conn.execute(
+                    sql_text(
+                        "SELECT m.* FROM conversation_messages m "
+                        "JOIN conversations c ON c.id = m.conversation_id "
+                        "WHERE m.id = CAST(:mid AS uuid) "
+                        "AND (c.user_id = :uid OR :uid = ANY(c.shared_with))"
+                    ),
+                    {"mid": message_id, "uid": user_id},
+                ).fetchone()
+                if row is None:
+                    return make_response(jsonify({"status": "not found"}), 404)
+                msg = row_to_dict(row)
+        except Exception as err:
+            current_app.logger.error(
+                f"Error tailing message {message_id}: {err}", exc_info=True
+            )
+            return make_response(jsonify({"success": False}), 400)
+        metadata = msg.get("message_metadata") or {}
+        return make_response(
+            jsonify(
+                {
+                    "message_id": str(msg["id"]),
+                    "status": msg.get("status"),
+                    "response": msg.get("response"),
+                    "thought": msg.get("thought"),
+                    "sources": msg.get("sources") or [],
+                    "tool_calls": msg.get("tool_calls") or [],
+                    "request_id": msg.get("request_id"),
+                    "last_heartbeat_at": metadata.get("last_heartbeat_at"),
+                    "error": metadata.get("error"),
+                }
+            ),
+            200,
+        )

application/api/user/idempotency.py

@@ -0,0 +1,237 @@
+"""Per-Celery-task idempotency wrapper backed by ``task_dedup``."""
+
+from __future__ import annotations
+
+import functools
+import logging
+import threading
+import uuid
+from typing import Any, Callable, Optional
+
+from application.storage.db.repositories.idempotency import IdempotencyRepository
+from application.storage.db.session import db_readonly, db_session
+
+
+logger = logging.getLogger(__name__)
+
+
+# Poison-loop cap; transient-failure headroom without infinite retry.
+MAX_TASK_ATTEMPTS = 5
+
+# 30s heartbeat / 60s TTL → ~2 missed ticks of slack before reclaim.
+LEASE_TTL_SECONDS = 60
+LEASE_HEARTBEAT_INTERVAL = 30
+
+# 10 × 60s ≈ 5 min of deferral before giving up on a held lease.
+LEASE_RETRY_MAX = 10
+
+
+def with_idempotency(task_name: str) -> Callable[[Callable[..., Any]], Callable[..., Any]]:
+    """Short-circuit on completed key; gate concurrent runs via a lease.
+
+    Entry short-circuits:
+      - completed row → return cached result
+      - live lease held → retry(countdown=LEASE_TTL_SECONDS)
+      - attempt_count > MAX_TASK_ATTEMPTS → poison-loop alert
+    Success writes ``completed``; exceptions leave ``pending`` for
+    autoretry until the poison-loop guard trips.
+    """
+
+    def decorator(fn: Callable[..., Any]) -> Callable[..., Any]:
+        @functools.wraps(fn)
+        def wrapper(self, *args: Any, idempotency_key: Any = None, **kwargs: Any) -> Any:
+            key = idempotency_key if isinstance(idempotency_key, str) and idempotency_key else None
+            if key is None:
+                return fn(self, *args, idempotency_key=idempotency_key, **kwargs)
+
+            cached = _lookup_completed(key)
+            if cached is not None:
+                logger.info(
+                    "idempotency hit for task=%s key=%s — returning cached result",
+                    task_name, key,
+                )
+                return cached
+
+            owner_id = str(uuid.uuid4())
+            attempt = _try_claim_lease(
+                key, task_name, _safe_task_id(self), owner_id,
+            )
+            if attempt is None:
+                # Live lease held by another worker. Re-queue and bail
+                # quickly — by the time the retry fires (LEASE_TTL
+                # seconds), Worker 1 has either finalised (we'll hit
+                # ``_lookup_completed`` and return cached) or its lease
+                # has expired and we can claim.
+                logger.info(
+                    "idempotency: live lease held; deferring task=%s key=%s",
+                    task_name, key,
+                )
+                raise self.retry(
+                    countdown=LEASE_TTL_SECONDS,
+                    max_retries=LEASE_RETRY_MAX,
+                )
+
+            if attempt > MAX_TASK_ATTEMPTS:
+                logger.error(
+                    "idempotency poison-loop guard: task=%s key=%s attempts=%s",
+                    task_name, key, attempt,
+                    extra={
+                        "alert": "idempotency_poison_loop",
+                        "task_name": task_name,
+                        "idempotency_key": key,
+                        "attempts": attempt,
+                    },
+                )
+                poisoned = {
+                    "success": False,
+                    "error": "idempotency poison-loop guard tripped",
+                    "attempts": attempt,
+                }
+                _finalize(key, poisoned, status="failed")
+                return poisoned
+
+            heartbeat_thread, heartbeat_stop = _start_lease_heartbeat(
+                key, owner_id,
+            )
+            try:
+                result = fn(self, *args, idempotency_key=idempotency_key, **kwargs)
+                _finalize(key, result, status="completed")
+                return result
+            except Exception:
+                # Drop the lease so the next retry doesn't wait LEASE_TTL.
+                _release_lease(key, owner_id)
+                raise
+            finally:
+                _stop_lease_heartbeat(heartbeat_thread, heartbeat_stop)
+
+        return wrapper
+
+    return decorator
+
+
+def _lookup_completed(key: str) -> Any:
+    """Return cached ``result_json`` if a completed row exists for ``key``, else None."""
+    with db_readonly() as conn:
+        row = IdempotencyRepository(conn).get_task(key)
+    if row is None:
+        return None
+    if row.get("status") != "completed":
+        return None
+    return row.get("result_json")
+
+
+def _try_claim_lease(
+    key: str, task_name: str, task_id: str, owner_id: str,
+) -> Optional[int]:
+    """Atomic CAS; returns ``attempt_count`` or ``None`` when held.
+
+    DB outage → treated as ``attempt=1`` so transient failures don't
+    block all task execution; reconciler repairs the lease columns.
+    """
+    try:
+        with db_session() as conn:
+            return IdempotencyRepository(conn).try_claim_lease(
+                key=key,
+                task_name=task_name,
+                task_id=task_id,
+                owner_id=owner_id,
+                ttl_seconds=LEASE_TTL_SECONDS,
+            )
+    except Exception:
+        logger.exception(
+            "idempotency lease-claim failed for key=%s task=%s", key, task_name,
+        )
+        return 1
+
+
+def _finalize(key: str, result_json: Any, *, status: str) -> None:
+    """Best-effort terminal write. Never let DB outage fail the task."""
+    try:
+        with db_session() as conn:
+            IdempotencyRepository(conn).finalize_task(
+                key=key, result_json=result_json, status=status,
+            )
+    except Exception:
+        logger.exception(
+            "idempotency finalize failed for key=%s status=%s", key, status,
+        )
+
+
+def _release_lease(key: str, owner_id: str) -> None:
+    """Best-effort lease release on the wrapper's exception path."""
+    try:
+        with db_session() as conn:
+            IdempotencyRepository(conn).release_lease(key, owner_id)
+    except Exception:
+        logger.exception("idempotency release-lease failed for key=%s", key)
+
+
+def _start_lease_heartbeat(
+    key: str, owner_id: str,
+) -> tuple[threading.Thread, threading.Event]:
+    """Spawn a daemon thread that bumps ``lease_expires_at`` every
+    :data:`LEASE_HEARTBEAT_INTERVAL` seconds until ``stop_event`` fires.
+
+    Mirrors ``application.worker._start_ingest_heartbeat`` so the two
+    durability heartbeats share shape and cadence.
+    """
+    stop_event = threading.Event()
+    thread = threading.Thread(
+        target=_lease_heartbeat_loop,
+        args=(key, owner_id, stop_event, LEASE_HEARTBEAT_INTERVAL),
+        daemon=True,
+        name=f"idempotency-lease-heartbeat:{key[:32]}",
+    )
+    thread.start()
+    return thread, stop_event
+
+
+def _stop_lease_heartbeat(
+    thread: threading.Thread, stop_event: threading.Event,
+) -> None:
+    """Signal the heartbeat thread to exit and join with a short timeout."""
+    stop_event.set()
+    thread.join(timeout=10)
+
+
+def _lease_heartbeat_loop(
+    key: str,
+    owner_id: str,
+    stop_event: threading.Event,
+    interval: int,
+) -> None:
+    """Refresh the lease until ``stop_event`` is set or ownership is lost.
+
+    A failed refresh (rowcount 0) means another worker stole the lease
+    after expiry — at that point the damage is already possible, so we
+    log and keep ticking. Don't escalate to thread death; the main task
+    body needs to keep running so its outcome is at least *recorded*.
+    """
+    while not stop_event.wait(interval):
+        try:
+            with db_session() as conn:
+                still_owned = IdempotencyRepository(conn).refresh_lease(
+                    key=key, owner_id=owner_id, ttl_seconds=LEASE_TTL_SECONDS,
+                )
+            if not still_owned:
+                logger.warning(
+                    "idempotency lease lost mid-task for key=%s "
+                    "(another worker may have taken over)",
+                    key,
+                )
+        except Exception:
+            logger.exception(
+                "idempotency lease-heartbeat tick failed for key=%s", key,
+            )
+
+
+def _safe_task_id(task_self: Any) -> str:
+    """Best-effort extraction of ``self.request.id`` from a Celery task."""
+    try:
+        request = getattr(task_self, "request", None)
+        task_id: Optional[str] = (
+            getattr(request, "id", None) if request is not None else None
+        )
+    except Exception:
+        task_id = None
+    return task_id or "unknown"

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/reconciliation.py

@@ -0,0 +1,196 @@
+"""Reconciler tick: sweep stuck rows and escalate to terminal status + alert."""
+
+from __future__ import annotations
+
+import logging
+import uuid
+from typing import Any, Dict, Optional
+
+from sqlalchemy import Connection
+
+from application.api.user.idempotency import MAX_TASK_ATTEMPTS
+from application.core.settings import settings
+from application.storage.db.engine import get_engine
+from application.storage.db.repositories.reconciliation import (
+    ReconciliationRepository,
+)
+from application.storage.db.repositories.stack_logs import StackLogsRepository
+
+logger = logging.getLogger(__name__)
+
+
+MAX_MESSAGE_RECONCILE_ATTEMPTS = 3
+
+
+def run_reconciliation() -> Dict[str, Any]:
+    """Single tick of the reconciler. Five sweeps, FOR UPDATE SKIP LOCKED.
+
+    Stuck ``executed`` tool calls always flip to ``failed`` — operators
+    handle cleanup manually via the structured alert. The side effect is
+    assumed to have committed; no automated rollback is attempted.
+
+    Stuck ``task_dedup`` rows (lease expired AND attempts >= max)
+    promote to ``failed`` so a same-key retry can re-claim instead of
+    sitting in ``pending`` until 24 h TTL.
+    """
+    if not settings.POSTGRES_URI:
+        return {
+            "messages_failed": 0,
+            "tool_calls_failed": 0,
+            "skipped": "POSTGRES_URI not set",
+        }
+
+    engine = get_engine()
+    summary = {
+        "messages_failed": 0,
+        "tool_calls_failed": 0,
+        "ingests_stalled": 0,
+        "idempotency_pending_failed": 0,
+    }
+
+    with engine.begin() as conn:
+        repo = ReconciliationRepository(conn)
+        for msg in repo.find_and_lock_stuck_messages():
+            new_count = repo.increment_message_reconcile_attempts(msg["id"])
+            if new_count >= MAX_MESSAGE_RECONCILE_ATTEMPTS:
+                repo.mark_message_failed(
+                    msg["id"],
+                    error=(
+                        "reconciler: stuck in pending/streaming for >5 min "
+                        f"after {new_count} attempts"
+                    ),
+                )
+                summary["messages_failed"] += 1
+                _emit_alert(
+                    conn,
+                    name="reconciler_message_failed",
+                    user_id=msg.get("user_id"),
+                    detail={
+                        "message_id": str(msg["id"]),
+                        "attempts": new_count,
+                    },
+                )
+
+    with engine.begin() as conn:
+        repo = ReconciliationRepository(conn)
+        for row in repo.find_and_lock_proposed_tool_calls():
+            repo.mark_tool_call_failed(
+                row["call_id"],
+                error=(
+                    "reconciler: stuck in 'proposed' for >5 min; "
+                    "side effect status unknown"
+                ),
+            )
+            summary["tool_calls_failed"] += 1
+            _emit_alert(
+                conn,
+                name="reconciler_tool_call_failed_proposed",
+                user_id=None,
+                detail={
+                    "call_id": row["call_id"],
+                    "tool_name": row.get("tool_name"),
+                },
+            )
+
+    with engine.begin() as conn:
+        repo = ReconciliationRepository(conn)
+        for row in repo.find_and_lock_executed_tool_calls():
+            repo.mark_tool_call_failed(
+                row["call_id"],
+                error=(
+                    "reconciler: executed-not-confirmed; side effect "
+                    "assumed committed, manual cleanup required"
+                ),
+            )
+            summary["tool_calls_failed"] += 1
+            _emit_alert(
+                conn,
+                name="reconciler_tool_call_failed_executed",
+                user_id=None,
+                detail={
+                    "call_id": row["call_id"],
+                    "tool_name": row.get("tool_name"),
+                    "action_name": row.get("action_name"),
+                },
+            )
+
+    # Q4: ingest checkpoints whose heartbeat has gone silent. The
+    # reconciler only escalates (alerts) — it doesn't kill the worker
+    # or roll back the partial embed. The next dispatch resumes from
+    # ``last_index`` thanks to the per-chunk checkpoint, so this is an
+    # observability sweep, not a recovery action.
+    with engine.begin() as conn:
+        repo = ReconciliationRepository(conn)
+        for row in repo.find_and_lock_stalled_ingests():
+            summary["ingests_stalled"] += 1
+            _emit_alert(
+                conn,
+                name="reconciler_ingest_stalled",
+                user_id=None,
+                detail={
+                    "source_id": str(row.get("source_id")),
+                    "embedded_chunks": row.get("embedded_chunks"),
+                    "total_chunks": row.get("total_chunks"),
+                    "last_updated": str(row.get("last_updated")),
+                },
+            )
+            # Bump the heartbeat so we don't re-alert every tick.
+            repo.touch_ingest_progress(str(row["source_id"]))
+
+    # Q5: idempotency rows whose lease expired with attempts exhausted.
+    # The wrapper's poison-loop guard normally finalises these, but if
+    # the wrapper itself died mid-task (worker SIGKILL, OOM during
+    # heartbeat) the row sits in ``pending`` blocking same-key retries
+    # via ``_lookup_completed`` returning None for the whole 24 h TTL.
+    # Promote to ``failed`` so a retry can re-claim and either resume
+    # or fail loudly.
+    with engine.begin() as conn:
+        repo = ReconciliationRepository(conn)
+        for row in repo.find_stuck_idempotency_pending(
+            max_attempts=MAX_TASK_ATTEMPTS,
+        ):
+            error_msg = (
+                "reconciler: idempotency lease expired with attempts "
+                f"({row['attempt_count']}) >= {MAX_TASK_ATTEMPTS}; "
+                "task abandoned"
+            )
+            repo.mark_idempotency_pending_failed(
+                row["idempotency_key"], error=error_msg,
+            )
+            summary["idempotency_pending_failed"] += 1
+            _emit_alert(
+                conn,
+                name="reconciler_idempotency_pending_failed",
+                user_id=None,
+                detail={
+                    "idempotency_key": row["idempotency_key"],
+                    "task_name": row.get("task_name"),
+                    "task_id": row.get("task_id"),
+                    "attempts": row.get("attempt_count"),
+                },
+            )
+
+    return summary
+
+
+def _emit_alert(
+    conn: Connection,
+    *,
+    name: str,
+    user_id: Optional[str],
+    detail: Dict[str, Any],
+) -> None:
+    """Structured ``logger.error`` plus a ``stack_logs`` row for operators."""
+    extra = {"alert": name, **detail}
+    logger.error("reconciler alert: %s", name, extra=extra)
+    try:
+        StackLogsRepository(conn).insert(
+            activity_id=str(uuid.uuid4()),
+            endpoint="reconciliation_worker",
+            level="ERROR",
+            user_id=user_id,
+            query=name,
+            stacks=[extra],
+        )
+    except Exception:
+        logger.exception("reconciler: failed to write stack_logs row for %s", name)

application/api/user/sources/upload.py

@@ -3,16 +3,19 @@
 import json
 import os
 import tempfile
+import uuid
 import zipfile
 
 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.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.idempotency import IdempotencyRepository
 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
@@ -30,6 +33,79 @@ sources_upload_ns = Namespace(
 )
 
 
+_IDEMPOTENCY_KEY_MAX_LEN = 256
+
+
+def _read_idempotency_key():
+    """Return (key, error_response). Empty header → (None, None); oversized → (None, 400)."""
+    key = request.headers.get("Idempotency-Key")
+    if not key:
+        return None, None
+    if len(key) > _IDEMPOTENCY_KEY_MAX_LEN:
+        return None, make_response(
+            jsonify(
+                {
+                    "success": False,
+                    "message": (
+                        f"Idempotency-Key exceeds maximum length of "
+                        f"{_IDEMPOTENCY_KEY_MAX_LEN} characters"
+                    ),
+                }
+            ),
+            400,
+        )
+    return key, None
+
+
+def _scoped_idempotency_key(idempotency_key, scope):
+    """``{scope}:{key}`` so different users can't collide on the same key."""
+    if not idempotency_key or not scope:
+        return None
+    return f"{scope}:{idempotency_key}"
+
+
+def _claim_task_or_get_cached(key, task_name):
+    """Claim ``key`` for this request OR return the winner's cached payload.
+
+    Pre-generates the celery task_id so a losing writer sees the same
+    id immediately. Returns ``(task_id, cached_response)``; non-None
+    cached means the caller should return without enqueuing.
+    """
+    predetermined_id = str(uuid.uuid4())
+    with db_session() as conn:
+        claimed = IdempotencyRepository(conn).claim_task(
+            key=key, task_name=task_name, task_id=predetermined_id,
+        )
+    if claimed is not None:
+        return claimed["task_id"], None
+    with db_readonly() as conn:
+        existing = IdempotencyRepository(conn).get_task(key)
+    cached_id = existing.get("task_id") if existing else None
+    return None, {
+        "success": True,
+        "task_id": cached_id or "deduplicated",
+    }
+
+
+def _release_claim(key):
+    """Drop a pending claim so a client retry can re-claim it."""
+    try:
+        with db_session() as conn:
+            conn.execute(
+                sql_text(
+                    "DELETE FROM task_dedup WHERE idempotency_key = :k "
+                    "AND status = 'pending'"
+                ),
+                {"k": key},
+            )
+    except Exception:
+        current_app.logger.exception(
+            "Failed to release task_dedup claim for key=%s", key,
+        )
+
+
+
+
 def _enforce_audio_path_size_limit(file_path: str, filename: str) -> None:
     if not is_audio_filename(filename):
         return
@@ -49,17 +125,38 @@ class UploadFile(Resource):
         )
     )
     @api.doc(
-        description="Uploads a file to be vectorized and indexed",
+        description=(
+            "Uploads a file to be vectorized and indexed. Honors an optional "
+            "``Idempotency-Key`` header: a repeat request with the same key "
+            "within 24h returns the original cached response without re-enqueuing."
+        ),
     )
     def post(self):
         decoded_token = request.decoded_token
         if not decoded_token:
             return make_response(jsonify({"success": False}), 401)
+        user = decoded_token.get("sub")
+        idempotency_key, key_error = _read_idempotency_key()
+        if key_error is not None:
+            return key_error
+        # User-scoped to avoid cross-user collisions; also feeds
+        # ``_derive_source_id`` so uuid5 stays user-disjoint.
+        scoped_key = _scoped_idempotency_key(idempotency_key, user)
+        # Claim before enqueue; the loser returns the winner's task_id.
+        predetermined_task_id = None
+        if scoped_key:
+            predetermined_task_id, cached = _claim_task_or_get_cached(
+                scoped_key, "ingest",
+            )
+            if cached is not None:
+                return make_response(jsonify(cached), 200)
         data = request.form
         files = request.files.getlist("file")
         required_fields = ["user", "name"]
         missing_fields = check_required_fields(data, required_fields)
         if missing_fields or not files or all(file.filename == "" for file in files):
+            if scoped_key:
+                _release_claim(scoped_key)
             return make_response(
                 jsonify(
                     {
@@ -69,7 +166,6 @@ class UploadFile(Resource):
                 ),
                 400,
             )
-        user = decoded_token.get("sub")
         job_name = request.form["name"]
 
         # Create safe versions for filesystem operations
@@ -140,16 +236,27 @@ class UploadFile(Resource):
                         file_path = f"{base_path}/{safe_file}"
                         with open(temp_file_path, "rb") as f:
                             storage.save_file(f, file_path)
-            task = ingest.delay(
-                settings.UPLOAD_FOLDER,
-                list(SUPPORTED_SOURCE_EXTENSIONS),
-                job_name,
-                user,
-                file_path=base_path,
-                filename=dir_name,
-                file_name_map=file_name_map,
+            ingest_kwargs = dict(
+                args=(
+                    settings.UPLOAD_FOLDER,
+                    list(SUPPORTED_SOURCE_EXTENSIONS),
+                    job_name,
+                    user,
+                ),
+                kwargs={
+                    "file_path": base_path,
+                    "filename": dir_name,
+                    "file_name_map": file_name_map,
+                    # Scoped so the worker dedup row matches the HTTP claim.
+                    "idempotency_key": scoped_key or idempotency_key,
+                },
             )
+            if predetermined_task_id is not None:
+                ingest_kwargs["task_id"] = predetermined_task_id
+            task = ingest.apply_async(**ingest_kwargs)
         except AudioFileTooLargeError:
+            if scoped_key:
+                _release_claim(scoped_key)
             return make_response(
                 jsonify(
                     {
@@ -161,8 +268,13 @@ class UploadFile(Resource):
             )
         except Exception as err:
             current_app.logger.error(f"Error uploading file: {err}", exc_info=True)
+            if scoped_key:
+                _release_claim(scoped_key)
             return make_response(jsonify({"success": False}), 400)
-        return make_response(jsonify({"success": True, "task_id": task.id}), 200)
+        # Predetermined id matches the dedup-claim row; loser GET sees same.
+        response_task_id = predetermined_task_id or task.id
+        response_payload = {"success": True, "task_id": response_task_id}
+        return make_response(jsonify(response_payload), 200)
 
 
 @sources_upload_ns.route("/remote")
@@ -182,17 +294,38 @@ class UploadRemote(Resource):
         )
     )
     @api.doc(
-        description="Uploads remote source for vectorization",
+        description=(
+            "Uploads remote source for vectorization. Honors an optional "
+            "``Idempotency-Key`` header: a repeat request with the same key "
+            "within 24h returns the original cached response without re-enqueuing."
+        ),
     )
     def post(self):
         decoded_token = request.decoded_token
         if not decoded_token:
             return make_response(jsonify({"success": False}), 401)
+        user = decoded_token.get("sub")
+        idempotency_key, key_error = _read_idempotency_key()
+        if key_error is not None:
+            return key_error
+        scoped_key = _scoped_idempotency_key(idempotency_key, user)
         data = request.form
         required_fields = ["user", "source", "name", "data"]
         missing_fields = check_required_fields(data, required_fields)
         if missing_fields:
             return missing_fields
+        task_name_for_dedup = (
+            "ingest_connector_task"
+            if data.get("source") in ConnectorCreator.get_supported_connectors()
+            else "ingest_remote"
+        )
+        predetermined_task_id = None
+        if scoped_key:
+            predetermined_task_id, cached = _claim_task_or_get_cached(
+                scoped_key, task_name_for_dedup,
+            )
+            if cached is not None:
+                return make_response(jsonify(cached), 200)
         try:
             config = json.loads(data["data"])
             source_data = None
@@ -208,6 +341,8 @@ class UploadRemote(Resource):
             elif data["source"] in ConnectorCreator.get_supported_connectors():
                 session_token = config.get("session_token")
                 if not session_token:
+                    if scoped_key:
+                        _release_claim(scoped_key)
                     return make_response(
                         jsonify(
                             {
@@ -236,31 +371,47 @@ class UploadRemote(Resource):
                 config["file_ids"] = file_ids
                 config["folder_ids"] = folder_ids
 
-                task = ingest_connector_task.delay(
-                    job_name=data["name"],
-                    user=decoded_token.get("sub"),
-                    source_type=data["source"],
-                    session_token=session_token,
-                    file_ids=file_ids,
-                    folder_ids=folder_ids,
-                    recursive=config.get("recursive", False),
-                    retriever=config.get("retriever", "classic"),
-                )
-                return make_response(
-                    jsonify({"success": True, "task_id": task.id}), 200
-                )
-            task = ingest_remote.delay(
-                source_data=source_data,
-                job_name=data["name"],
-                user=decoded_token.get("sub"),
-                loader=data["source"],
-            )
+                connector_kwargs = {
+                    "kwargs": {
+                        "job_name": data["name"],
+                        "user": user,
+                        "source_type": data["source"],
+                        "session_token": session_token,
+                        "file_ids": file_ids,
+                        "folder_ids": folder_ids,
+                        "recursive": config.get("recursive", False),
+                        "retriever": config.get("retriever", "classic"),
+                        "idempotency_key": scoped_key or idempotency_key,
+                    },
+                }
+                if predetermined_task_id is not None:
+                    connector_kwargs["task_id"] = predetermined_task_id
+                task = ingest_connector_task.apply_async(**connector_kwargs)
+                response_task_id = predetermined_task_id or task.id
+                response_payload = {"success": True, "task_id": response_task_id}
+                return make_response(jsonify(response_payload), 200)
+            remote_kwargs = {
+                "kwargs": {
+                    "source_data": source_data,
+                    "job_name": data["name"],
+                    "user": user,
+                    "loader": data["source"],
+                    "idempotency_key": scoped_key or idempotency_key,
+                },
+            }
+            if predetermined_task_id is not None:
+                remote_kwargs["task_id"] = predetermined_task_id
+            task = ingest_remote.apply_async(**remote_kwargs)
         except Exception as err:
             current_app.logger.error(
                 f"Error uploading remote source: {err}", exc_info=True
             )
+            if scoped_key:
+                _release_claim(scoped_key)
             return make_response(jsonify({"success": False}), 400)
-        return make_response(jsonify({"success": True, "task_id": task.id}), 200)
+        response_task_id = predetermined_task_id or task.id
+        response_payload = {"success": True, "task_id": response_task_id}
+        return make_response(jsonify(response_payload), 200)
 
 
 @sources_upload_ns.route("/manage_source_files")
@@ -305,6 +456,10 @@ class ManageSourceFiles(Resource):
                 jsonify({"success": False, "message": "Unauthorized"}), 401
             )
         user = decoded_token.get("sub")
+        idempotency_key, key_error = _read_idempotency_key()
+        if key_error is not None:
+            return key_error
+        scoped_key = _scoped_idempotency_key(idempotency_key, user)
         source_id = request.form.get("source_id")
         operation = request.form.get("operation")
 
@@ -347,6 +502,12 @@ class ManageSourceFiles(Resource):
                 jsonify({"success": False, "message": "Database error"}), 500
             )
         resolved_source_id = str(source["id"])
+        # Flips to True after each branch's ``apply_async`` returns
+        # successfully — at that point the worker owns the predetermined
+        # task_id. The outer ``except`` only releases the claim while
+        # this is False, so a post-``apply_async`` failure (jsonify,
+        # make_response, etc.) doesn't double-enqueue on the next retry.
+        claim_transferred = False
         try:
             storage = StorageCreator.get_storage()
             source_file_path = source.get("file_path", "")
@@ -379,6 +540,21 @@ class ManageSourceFiles(Resource):
                         ),
                         400,
                     )
+
+                # Claim before any storage mutation so a duplicate request
+                # short-circuits without touching the filesystem. Mirrors
+                # the pattern in ``UploadFile.post`` / ``UploadRemote.post``
+                # — without it ``.delay()`` would enqueue twice for two
+                # racing same-key POSTs (the worker decorator only
+                # deduplicates *after* completion).
+                predetermined_task_id = None
+                if scoped_key:
+                    predetermined_task_id, cached = _claim_task_or_get_cached(
+                        scoped_key, "reingest_source_task",
+                    )
+                    if cached is not None:
+                        return make_response(jsonify(cached), 200)
+
                 added_files = []
                 map_updated = False
 
@@ -414,9 +590,15 @@ class ManageSourceFiles(Resource):
 
                 from application.api.user.tasks import reingest_source_task
 
-                task = reingest_source_task.delay(
-                    source_id=resolved_source_id, user=user
+                task = reingest_source_task.apply_async(
+                    kwargs={
+                        "source_id": resolved_source_id,
+                        "user": user,
+                        "idempotency_key": scoped_key or idempotency_key,
+                    },
+                    task_id=predetermined_task_id,
                 )
+                claim_transferred = True
 
                 return make_response(
                     jsonify(
@@ -455,10 +637,8 @@ class ManageSourceFiles(Resource):
                         ),
                         400,
                     )
-                # Remove files from storage and directory structure
-
-                removed_files = []
-                map_updated = False
+                # Path-traversal guard runs *before* the claim so a 400
+                # for an invalid path doesn't leave a pending dedup row.
                 for file_path in file_paths:
                     if ".." in str(file_path) or str(file_path).startswith("/"):
                         return make_response(
@@ -470,6 +650,22 @@ class ManageSourceFiles(Resource):
                             ),
                             400,
                         )
+
+                # Claim before any storage mutation. See ``add`` branch
+                # comment for rationale.
+                predetermined_task_id = None
+                if scoped_key:
+                    predetermined_task_id, cached = _claim_task_or_get_cached(
+                        scoped_key, "reingest_source_task",
+                    )
+                    if cached is not None:
+                        return make_response(jsonify(cached), 200)
+
+                # Remove files from storage and directory structure
+
+                removed_files = []
+                map_updated = False
+                for file_path in file_paths:
                     full_path = f"{source_file_path}/{file_path}"
 
                     # Remove from storage
@@ -491,9 +687,15 @@ class ManageSourceFiles(Resource):
 
                 from application.api.user.tasks import reingest_source_task
 
-                task = reingest_source_task.delay(
-                    source_id=resolved_source_id, user=user
+                task = reingest_source_task.apply_async(
+                    kwargs={
+                        "source_id": resolved_source_id,
+                        "user": user,
+                        "idempotency_key": scoped_key or idempotency_key,
+                    },
+                    task_id=predetermined_task_id,
                 )
+                claim_transferred = True
 
                 return make_response(
                     jsonify(
@@ -552,6 +754,16 @@ class ManageSourceFiles(Resource):
                         ),
                         404,
                     )
+
+                # Claim before mutation. See ``add`` branch for rationale.
+                predetermined_task_id = None
+                if scoped_key:
+                    predetermined_task_id, cached = _claim_task_or_get_cached(
+                        scoped_key, "reingest_source_task",
+                    )
+                    if cached is not None:
+                        return make_response(jsonify(cached), 200)
+
                 success = storage.remove_directory(full_directory_path)
 
                 if not success:
@@ -560,6 +772,11 @@ class ManageSourceFiles(Resource):
                         f"User: {user}, Source ID: {source_id}, Directory path: {directory_path}, "
                         f"Full path: {full_directory_path}"
                     )
+                    # Release so a client retry can reclaim — otherwise
+                    # the next request would silently 200-cache to the
+                    # task_id that never enqueued.
+                    if scoped_key:
+                        _release_claim(scoped_key)
                     return make_response(
                         jsonify(
                             {"success": False, "message": "Failed to remove directory"}
@@ -591,9 +808,15 @@ class ManageSourceFiles(Resource):
 
                 from application.api.user.tasks import reingest_source_task
 
-                task = reingest_source_task.delay(
-                    source_id=resolved_source_id, user=user
+                task = reingest_source_task.apply_async(
+                    kwargs={
+                        "source_id": resolved_source_id,
+                        "user": user,
+                        "idempotency_key": scoped_key or idempotency_key,
+                    },
+                    task_id=predetermined_task_id,
                 )
+                claim_transferred = True
 
                 return make_response(
                     jsonify(
@@ -607,6 +830,14 @@ class ManageSourceFiles(Resource):
                     200,
                 )
         except Exception as err:
+            # Release the dedup claim only if it wasn't transferred to
+            # a worker. Without this, a same-key retry within the 24h
+            # TTL would 200-cache to a predetermined task_id whose
+            # ``apply_async`` never ran (or ran but the response builder
+            # blew up afterward — only the first case matters in
+            # practice; the flag protects both).
+            if scoped_key and not claim_transferred:
+                _release_claim(scoped_key)
             error_context = f"operation={operation}, user={user}, source_id={source_id}"
             if operation == "remove_directory":
                 directory_path = request.form.get("directory_path", "")

application/api/user/tasks.py

@@ -1,5 +1,6 @@
 from datetime import timedelta
 
+from application.api.user.idempotency import with_idempotency
 from application.celery_init import celery
 from application.worker import (
     agent_webhook_worker,
@@ -13,9 +14,32 @@ from application.worker import (
 )
 
 
-@celery.task(bind=True)
+# Shared decorator config for long-running, side-effecting tasks. ``acks_late``
+# is also the celeryconfig default but stays explicit here so each task's
+# durability story is grep-able next to the body. Combined with
+# ``autoretry_for=(Exception,)`` and a bounded ``max_retries`` so a poison
+# message can't loop forever.
+DURABLE_TASK = dict(
+    bind=True,
+    acks_late=True,
+    autoretry_for=(Exception,),
+    retry_kwargs={"max_retries": 3, "countdown": 60},
+    retry_backoff=True,
+)
+
+
+@celery.task(**DURABLE_TASK)
+@with_idempotency(task_name="ingest")
 def ingest(
-    self, directory, formats, job_name, user, file_path, filename, file_name_map=None
+    self,
+    directory,
+    formats,
+    job_name,
+    user,
+    file_path,
+    filename,
+    file_name_map=None,
+    idempotency_key=None,
 ):
     resp = ingest_worker(
         self,
@@ -26,25 +50,35 @@ def ingest(
         filename,
         user,
         file_name_map=file_name_map,
+        idempotency_key=idempotency_key,
     )
     return resp
 
 
-@celery.task(bind=True)
-def ingest_remote(self, source_data, job_name, user, loader):
-    resp = remote_worker(self, source_data, job_name, user, loader)
+@celery.task(**DURABLE_TASK)
+@with_idempotency(task_name="ingest_remote")
+def ingest_remote(self, source_data, job_name, user, loader, idempotency_key=None):
+    resp = remote_worker(
+        self, source_data, job_name, user, loader,
+        idempotency_key=idempotency_key,
+    )
     return resp
 
 
-@celery.task(bind=True)
-def reingest_source_task(self, source_id, user):
+@celery.task(**DURABLE_TASK)
+@with_idempotency(task_name="reingest_source_task")
+def reingest_source_task(self, source_id, user, idempotency_key=None):
     from application.worker import reingest_source_worker
 
     resp = reingest_source_worker(self, source_id, user)
     return resp
 
 
-@celery.task(bind=True)
+# Beat-driven dispatch tasks default to ``acks_late=False``: a SIGKILL
+# of a beat tick is harmless to redeliver only if the dispatch itself is
+# idempotent. We keep these early-ACK so the broker doesn't replay a
+# dispatch that already enqueued downstream work.
+@celery.task(bind=True, acks_late=False)
 def schedule_syncs(self, frequency):
     resp = sync_worker(self, frequency)
     return resp
@@ -74,19 +108,22 @@ def sync_source(
     return resp
 
 
-@celery.task(bind=True)
-def store_attachment(self, file_info, user):
+@celery.task(**DURABLE_TASK)
+@with_idempotency(task_name="store_attachment")
+def store_attachment(self, file_info, user, idempotency_key=None):
     resp = attachment_worker(self, file_info, user)
     return resp
 
 
-@celery.task(bind=True)
-def process_agent_webhook(self, agent_id, payload):
+@celery.task(**DURABLE_TASK)
+@with_idempotency(task_name="process_agent_webhook")
+def process_agent_webhook(self, agent_id, payload, idempotency_key=None):
     resp = agent_webhook_worker(self, agent_id, payload)
     return resp
 
 
-@celery.task(bind=True)
+@celery.task(**DURABLE_TASK)
+@with_idempotency(task_name="ingest_connector_task")
 def ingest_connector_task(
     self,
     job_name,
@@ -100,6 +137,7 @@ def ingest_connector_task(
     operation_mode="upload",
     doc_id=None,
     sync_frequency="never",
+    idempotency_key=None,
 ):
     from application.worker import ingest_connector
 
@@ -116,6 +154,7 @@ def ingest_connector_task(
         operation_mode=operation_mode,
         doc_id=doc_id,
         sync_frequency=sync_frequency,
+        idempotency_key=idempotency_key,
     )
     return resp
 
@@ -140,6 +179,24 @@ def setup_periodic_tasks(sender, **kwargs):
         cleanup_pending_tool_state.s(),
         name="cleanup-pending-tool-state",
     )
+    # Pure housekeeping for ``task_dedup`` / ``webhook_dedup`` — the
+    # upsert paths already handle stale rows, so cadence only bounds
+    # table size. Hourly is plenty for typical traffic.
+    sender.add_periodic_task(
+        timedelta(hours=1),
+        cleanup_idempotency_dedup.s(),
+        name="cleanup-idempotency-dedup",
+    )
+    sender.add_periodic_task(
+        timedelta(seconds=30),
+        reconciliation_task.s(),
+        name="reconciliation",
+    )
+    sender.add_periodic_task(
+        timedelta(hours=7),
+        version_check_task.s(),
+        name="version-check",
+    )
 
 
 @celery.task(bind=True)
@@ -154,18 +211,12 @@ def mcp_oauth_status_task(self, task_id):
     return resp
 
 
-@celery.task(bind=True)
+@celery.task(bind=True, acks_late=False)
 def cleanup_pending_tool_state(self):
-    """Delete pending_tool_state rows past their TTL.
-
-    Replaces Mongo's ``expireAfterSeconds=0`` TTL index — Postgres has
-    no native TTL, so this task runs every 60 seconds to keep
-    ``pending_tool_state`` bounded. No-ops if ``POSTGRES_URI`` isn't
-    configured (keeps the task runnable in Mongo-only environments).
-    """
+    """Revert stale ``resuming`` rows, then delete TTL-expired rows."""
     from application.core.settings import settings
     if not settings.POSTGRES_URI:
-        return {"deleted": 0, "skipped": "POSTGRES_URI not set"}
+        return {"deleted": 0, "reverted": 0, "skipped": "POSTGRES_URI not set"}
 
     from application.storage.db.engine import get_engine
     from application.storage.db.repositories.pending_tool_state import (
@@ -174,5 +225,54 @@ def cleanup_pending_tool_state(self):
 
     engine = get_engine()
     with engine.begin() as conn:
-        deleted = PendingToolStateRepository(conn).cleanup_expired()
-    return {"deleted": deleted}
+        repo = PendingToolStateRepository(conn)
+        reverted = repo.revert_stale_resuming(grace_seconds=600)
+        deleted = repo.cleanup_expired()
+    return {"deleted": deleted, "reverted": reverted}
+
+
+@celery.task(bind=True, acks_late=False)
+def cleanup_idempotency_dedup(self):
+    """Delete TTL-expired rows from ``task_dedup`` and ``webhook_dedup``.
+
+    Pure housekeeping — the upsert paths already ignore stale rows
+    (TTL-aware ``ON CONFLICT DO UPDATE``), so this only bounds table
+    growth and keeps SELECT planning tight on large deployments.
+    """
+    from application.core.settings import settings
+    if not settings.POSTGRES_URI:
+        return {
+            "task_dedup_deleted": 0,
+            "webhook_dedup_deleted": 0,
+            "skipped": "POSTGRES_URI not set",
+        }
+
+    from application.storage.db.engine import get_engine
+    from application.storage.db.repositories.idempotency import (
+        IdempotencyRepository,
+    )
+
+    engine = get_engine()
+    with engine.begin() as conn:
+        return IdempotencyRepository(conn).cleanup_expired()
+
+
+@celery.task(bind=True, acks_late=False)
+def reconciliation_task(self):
+    """Sweep stuck durability rows and escalate them to terminal status + alert."""
+    from application.api.user.reconciliation import run_reconciliation
+
+    return run_reconciliation()
+
+
+@celery.task(bind=True, acks_late=False)
+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/workflows/routes.py

@@ -198,8 +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:
@@ -343,7 +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"
@@ -389,7 +395,9 @@ class WorkflowList(Resource):
         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
@@ -451,7 +459,9 @@ class WorkflowDetail(Resource):
         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

application/api/v1/routes.py

@@ -9,6 +9,7 @@ import json
 import logging
 import time
 import traceback
+from datetime import datetime
 from typing import Any, Dict, Generator, Optional
 
 from flask import Blueprint, jsonify, make_response, request, Response
@@ -213,6 +214,7 @@ 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,
     )
@@ -257,6 +259,7 @@ 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,
     )
@@ -304,7 +307,16 @@ def list_models():
                     401,
                 )
 
+        # Repository rows now go through ``coerce_pg_native`` at SELECT
+        # time, so timestamps arrive as ISO 8601 strings. Parse before
+        # taking ``.timestamp()``; fall back to ``time.time()`` only when
+        # the value is genuinely missing or unparseable.
         created = agent.get("created_at") or agent.get("createdAt")
+        if isinstance(created, str):
+            try:
+                created = datetime.fromisoformat(created)
+            except (ValueError, TypeError):
+                created = None
         created_ts = (
             int(created.timestamp()) if hasattr(created, "timestamp")
             else int(time.time())

application/app.py

@@ -4,11 +4,12 @@ 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
 
+from application.core import log_context
 from application.core.logging_config import setup_logging
 
 setup_logging()
@@ -112,6 +113,38 @@ def generate_token():
     return jsonify({"error": "Token generation not allowed in current auth mode"}), 400
 
 
+_LOG_CTX_TOKEN_ATTR = "_log_ctx_token"
+
+
+@app.before_request
+def _bind_log_context():
+    """Bind activity_id + endpoint for the duration of this request.
+
+    Runs before ``authenticate_request``; ``user_id`` is overlaid in a
+    follow-up handler once the JWT has been decoded.
+    """
+    if request.method == "OPTIONS":
+        return None
+    activity_id = str(uuid.uuid4())
+    request.activity_id = activity_id
+    token = log_context.bind(
+        activity_id=activity_id,
+        endpoint=request.endpoint,
+    )
+    setattr(request, _LOG_CTX_TOKEN_ATTR, token)
+    return None
+
+
+@app.teardown_request
+def _reset_log_context(_exc):
+    # SSE streams keep yielding after teardown fires, but a2wsgi runs each
+    # request inside ``copy_context().run(...)``, so this reset doesn't
+    # leak into the stream's view of the context.
+    token = getattr(request, _LOG_CTX_TOKEN_ATTR, None)
+    if token is not None:
+        log_context.reset(token)
+
+
 @app.before_request
 def enforce_stt_request_size_limits():
     if request.method == "OPTIONS":
@@ -148,13 +181,29 @@ def authenticate_request():
         request.decoded_token = decoded_token
 
 
+@app.before_request
+def _bind_user_id_to_log_context():
+    # Registered after ``authenticate_request`` (Flask runs before_request
+    # handlers in registration order), so ``request.decoded_token`` is
+    # populated by the time we read it. ``teardown_request`` unwinds the
+    # whole request-level bind, so no separate reset token is needed here.
+    if request.method == "OPTIONS":
+        return None
+    decoded_token = getattr(request, "decoded_token", None)
+    user_id = decoded_token.get("sub") if isinstance(decoded_token, dict) else None
+    if user_id:
+        log_context.bind(user_id=user_id)
+    return None
+
+
 @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, Idempotency-Key"
     )
+    response.headers["Access-Control-Allow-Methods"] = "GET, POST, PUT, PATCH, DELETE, OPTIONS"
     return response
 
 

application/asgi.py

@@ -0,0 +1,38 @@
+"""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",
+                "Idempotency-Key",
+            ],
+            expose_headers=["Mcp-Session-Id"],
+        ),
+    ],
+    lifespan=mcp_app.lifespan,
+)

application/cache.py

@@ -1,3 +1,4 @@
+import hashlib
 import json
 import logging
 import time
@@ -10,6 +11,14 @@ from application.utils import get_hash
 
 logger = logging.getLogger(__name__)
 
+
+def _cache_default(value):
+    # Image attachments arrive inline as bytes (see GoogleLLM.prepare_messages_with_attachments);
+    # hash so the cache key stays bounded in size and stable across identical content.
+    if isinstance(value, (bytes, bytearray, memoryview)):
+        return f"<bytes:sha256:{hashlib.sha256(bytes(value)).hexdigest()}>"
+    return repr(value)
+
 _redis_instance = None
 _redis_creation_failed = False
 _instance_lock = Lock()
@@ -36,7 +45,7 @@ def get_redis_instance():
 def gen_cache_key(messages, model="docgpt", tools=None):
     if not all(isinstance(msg, dict) for msg in messages):
         raise ValueError("All messages must be dictionaries.")
-    messages_str = json.dumps(messages)
+    messages_str = json.dumps(messages, default=_cache_default)
     tools_str = json.dumps(str(tools)) if tools else ""
     combined = f"{model}_{messages_str}_{tools_str}"
     cache_key = get_hash(combined)

application/celery_init.py

@@ -1,6 +1,17 @@
+import inspect
+import logging
+import threading
+
 from celery import Celery
+from application.core import log_context
 from application.core.settings import settings
-from celery.signals import setup_logging, worker_process_init
+from celery.signals import (
+    setup_logging,
+    task_postrun,
+    task_prerun,
+    worker_process_init,
+    worker_ready,
+)
 
 
 def make_celery(app_name=__name__):
@@ -39,5 +50,73 @@ def _dispose_db_engine_on_fork(*args, **kwargs):
     dispose_engine()
 
 
+# Most tasks in this repo accept ``user`` where the log context wants
+# ``user_id``; map task parameter names to context keys explicitly.
+_TASK_PARAM_TO_CTX_KEY: dict[str, str] = {
+    "user": "user_id",
+    "user_id": "user_id",
+    "agent_id": "agent_id",
+    "conversation_id": "conversation_id",
+}
+
+_task_log_tokens: dict[str, object] = {}
+
+
+@task_prerun.connect
+def _bind_task_log_context(task_id, task, args, kwargs, **_):
+    # Resolve task args by parameter name — nearly every task in this repo
+    # is called positionally, so ``kwargs.get('user')`` would bind nothing.
+    ctx = {"activity_id": task_id}
+    try:
+        sig = inspect.signature(task.run)
+        bound = sig.bind_partial(*args, **kwargs).arguments
+    except (TypeError, ValueError):
+        bound = dict(kwargs)
+    for param_name, value in bound.items():
+        ctx_key = _TASK_PARAM_TO_CTX_KEY.get(param_name)
+        if ctx_key and value:
+            ctx[ctx_key] = value
+    _task_log_tokens[task_id] = log_context.bind(**ctx)
+
+
+@task_postrun.connect
+def _unbind_task_log_context(task_id, **_):
+    # ``task_postrun`` fires on both success and failure. Required for
+    # Celery: unlike the Flask path, tasks aren't isolated in their own
+    # ``copy_context().run(...)``, so a missing reset would leak the
+    # bind onto the next task on the same worker.
+    token = _task_log_tokens.pop(task_id, None)
+    if token is None:
+        return
+    try:
+        log_context.reset(token)
+    except ValueError:
+        # task_prerun and task_postrun ran on different threads (non-default
+        # Celery pool); the token isn't valid in this context. Drop it.
+        logging.getLogger(__name__).debug(
+            "log_context reset skipped for task %s", task_id
+        )
+
+
+@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

@@ -1,7 +1,10 @@
-import os
+from application.core.settings import settings
 
-broker_url = os.getenv("CELERY_BROKER_URL")
-result_backend = os.getenv("CELERY_RESULT_BACKEND")
+# Pydantic loads .env into ``settings`` but does not inject values into
+# ``os.environ`` — read directly from settings so beat startup (which
+# imports this module before any explicit env load) sees a real URL.
+broker_url = settings.CELERY_BROKER_URL
+result_backend = settings.CELERY_RESULT_BACKEND
 
 task_serializer = 'json'
 result_serializer = 'json'
@@ -9,3 +12,22 @@ accept_content = ['json']
 
 # Autodiscover tasks
 imports = ('application.api.user.tasks',)
+
+# Project-scoped queue so a stray sibling worker on the same broker
+# (other repo, same default ``celery`` queue) can't grab DocsGPT tasks.
+task_default_queue = "docsgpt"
+task_default_exchange = "docsgpt"
+task_default_routing_key = "docsgpt"
+
+beat_scheduler = "redbeat.RedBeatScheduler"
+redbeat_redis_url = broker_url
+redbeat_key_prefix = "redbeat:docsgpt:"
+redbeat_lock_timeout = 90
+
+# Survive worker SIGKILL/OOM without silently dropping in-flight tasks.
+task_acks_late = True
+task_reject_on_worker_lost = True
+worker_prefetch_multiplier = settings.CELERY_WORKER_PREFETCH_MULTIPLIER
+broker_transport_options = {"visibility_timeout": settings.CELERY_VISIBILITY_TIMEOUT}
+result_expires = 86400 * 7
+task_track_started = True

application/core/log_context.py

@@ -0,0 +1,57 @@
+"""Per-activity logging context backed by ``contextvars``.
+
+The ``_ContextFilter`` installed by ``logging_config.setup_logging`` stamps
+every ``LogRecord`` emitted inside a ``bind`` block with the bound keys, so
+they land as first-class attributes on the OTLP log export rather than being
+buried inside formatted message bodies.
+
+A single ``ContextVar`` holds a dict so nested binds reset atomically (LIFO)
+via the token returned by ``bind``.
+"""
+
+from __future__ import annotations
+
+from contextvars import ContextVar, Token
+from typing import Mapping
+
+
+_CTX_KEYS: frozenset[str] = frozenset(
+    {
+        "activity_id",
+        "parent_activity_id",
+        "user_id",
+        "agent_id",
+        "conversation_id",
+        "endpoint",
+        "model",
+    }
+)
+
+_ctx: ContextVar[Mapping[str, str]] = ContextVar("log_ctx", default={})
+
+
+def bind(**kwargs: object) -> Token:
+    """Overlay the given keys onto the current context.
+
+    Returns a ``Token`` so the caller can ``reset`` in a ``finally`` block.
+    Keys outside :data:`_CTX_KEYS` are silently dropped (so a typo can't
+    stamp a stray field name onto every record), as are ``None`` values
+    (a missing attribute is more useful than the literal string ``"None"``).
+    """
+    overlay = {
+        k: str(v)
+        for k, v in kwargs.items()
+        if k in _CTX_KEYS and v is not None
+    }
+    new = {**_ctx.get(), **overlay}
+    return _ctx.set(new)
+
+
+def reset(token: Token) -> None:
+    """Restore the context to the snapshot captured by the matching ``bind``."""
+    _ctx.reset(token)
+
+
+def snapshot() -> Mapping[str, str]:
+    """Return the current context dict. Treat as read-only; use :func:`bind`."""
+    return _ctx.get()

application/core/logging_config.py

@@ -1,11 +1,75 @@
+import logging
+import os
 from logging.config import dictConfig
 
-def setup_logging():
+from application.core.log_context import snapshot as _ctx_snapshot
+
+
+# Loggers with ``propagate=False`` don't share root's handlers, so the
+# context filter has to be installed on their handlers directly.
+_NON_PROPAGATING_LOGGERS: tuple[str, ...] = (
+    "uvicorn",
+    "uvicorn.access",
+    "uvicorn.error",
+    "celery.app.trace",
+    "celery.worker.strategy",
+    "gunicorn.error",
+    "gunicorn.access",
+)
+
+
+class _ContextFilter(logging.Filter):
+    """Stamp the current ``log_context`` snapshot onto every ``LogRecord``.
+
+    Must be installed on **handlers**, not loggers: Python skips logger-level
+    filters when a child logger's record propagates up. The ``hasattr`` guard
+    keeps an explicit ``logger.info(..., extra={...})`` from being overwritten.
+    """
+
+    def filter(self, record: logging.LogRecord) -> bool:
+        for key, value in _ctx_snapshot().items():
+            if not hasattr(record, key):
+                setattr(record, key, value)
+        return True
+
+
+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 +79,34 @@ 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)
+
+    _install_context_filter()
+
+
+def _install_context_filter() -> None:
+    """Attach :class:`_ContextFilter` to root's handlers + every handler on
+    the known non-propagating loggers. Skipping handlers that already carry
+    one keeps repeat ``setup_logging`` calls from stacking filters.
+    """
+
+    def _has_ctx_filter(handler: logging.Handler) -> bool:
+        return any(isinstance(f, _ContextFilter) for f in handler.filters)
+
+    for handler in logging.getLogger().handlers:
+        if not _has_ctx_filter(handler):
+            handler.addFilter(_ContextFilter())
+    for name in _NON_PROPAGATING_LOGGERS:
+        for handler in logging.getLogger(name).handlers:
+            if not _has_ctx_filter(handler):
+                handler.addFilter(_ContextFilter())

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/settings.py

@@ -23,9 +23,19 @@ 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"
+    # Prefetch=1 caps SIGKILL loss to one task. Visibility timeout must exceed
+    # the longest legitimate task runtime (ingest, agent webhook) but stay
+    # short enough that SIGKILLed tasks redeliver promptly. 1h matches Onyx
+    # and Dify defaults; long ingests can override via env.
+    CELERY_WORKER_PREFETCH_MULTIPLIER: int = 1
+    CELERY_VISIBILITY_TIMEOUT: int = 3600
     # 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.
@@ -149,6 +159,9 @@ 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/anthropic.py

@@ -11,6 +11,7 @@ logger = logging.getLogger(__name__)
 
 
 class AnthropicLLM(BaseLLM):
+    provider_name = "anthropic"
 
     def __init__(self, api_key=None, user_api_key=None, base_url=None, *args, **kwargs):
 

application/llm/base.py

@@ -1,5 +1,6 @@
 import logging
 from abc import ABC, abstractmethod
+from typing import ClassVar
 
 from application.cache import gen_cache, stream_cache
 
@@ -10,6 +11,10 @@ logger = logging.getLogger(__name__)
 
 
 class BaseLLM(ABC):
+    # Stamped onto the ``llm_stream_start`` event so dashboards can group
+    # calls by vendor. Subclasses override.
+    provider_name: ClassVar[str] = "unknown"
+
     def __init__(
         self,
         decoded_token=None,
@@ -17,6 +22,8 @@ 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 +32,12 @@ 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 +52,19 @@ 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 +78,15 @@ 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,
+                )
+                # Tag the fallback LLM so its rows land as
+                # ``source='fallback'`` in cost-attribution dashboards.
+                # Propagate the parent's ``_request_id`` so a user
+                # request that ran fallback is still grouped under one id.
+                self._fallback_llm._token_usage_source = "fallback"
+                self._fallback_llm._request_id = getattr(
+                    self, "_request_id", None,
                 )
                 logger.info(
                     f"Fallback LLM initialized from agent backup model: "
@@ -68,7 +99,10 @@ 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 +112,12 @@ 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,
+                )
+                # Same rationale as the agent-backup branch.
+                self._fallback_llm._token_usage_source = "fallback"
+                self._fallback_llm._request_id = getattr(
+                    self, "_request_id", None,
                 )
                 logger.info(
                     f"Fallback LLM initialized from global settings: "
@@ -96,6 +136,26 @@ class BaseLLM(ABC):
             return args_dict
         return {k: v for k, v in args_dict.items() if v is not None}
 
+    @staticmethod
+    def _is_non_retriable_client_error(exc: BaseException) -> bool:
+        """4xx errors mean the request itself is malformed — retrying with
+        a different model fails identically and doubles the work. Only
+        transient/5xx/connection errors should trigger fallback."""
+        try:
+            from google.genai.errors import ClientError as _GenaiClientError
+
+            if isinstance(exc, _GenaiClientError):
+                return True
+        except ImportError:
+            pass
+        for attr in ("status_code", "code", "http_status"):
+            v = getattr(exc, attr, None)
+            if isinstance(v, int) and 400 <= v < 500:
+                return True
+        resp = getattr(exc, "response", None)
+        v = getattr(resp, "status_code", None)
+        return isinstance(v, int) and 400 <= v < 500
+
     def _execute_with_fallback(
         self, method_name: str, decorators: list, *args, **kwargs
     ):
@@ -119,12 +179,18 @@ class BaseLLM(ABC):
 
         if is_stream:
             return self._stream_with_fallback(
-                decorated_method, method_name, *args, **kwargs
+                decorated_method, method_name, decorators, *args, **kwargs
             )
 
         try:
             return decorated_method()
         except Exception as e:
+            if self._is_non_retriable_client_error(e):
+                logger.error(
+                    f"Primary LLM failed with non-retriable client error; "
+                    f"skipping fallback: {str(e)}"
+                )
+                raise
             if not self.fallback_llm:
                 logger.error(f"Primary LLM failed and no fallback configured: {str(e)}")
                 raise
@@ -134,14 +200,27 @@ class BaseLLM(ABC):
                 f"{fallback.model_id}. Error: {str(e)}"
             )
 
-            fallback_method = getattr(
-                fallback, method_name.replace("_raw_", "")
-            )
+            # Apply decorators to fallback's raw method directly — calling
+            # fallback.gen() would re-enter the orchestrator and recurse via
+            # fallback.fallback_llm.
+            fallback_method = getattr(fallback, method_name)
+            for decorator in decorators:
+                fallback_method = decorator(fallback_method)
             fallback_kwargs = {**kwargs, "model": fallback.model_id}
-            return fallback_method(*args, **fallback_kwargs)
+            try:
+                return fallback_method(fallback, *args, **fallback_kwargs)
+            except Exception as e2:
+                if self._is_non_retriable_client_error(e2):
+                    logger.error(
+                        f"Fallback LLM failed with non-retriable client "
+                        f"error; giving up: {str(e2)}"
+                    )
+                else:
+                    logger.error(f"Fallback LLM also failed; giving up: {str(e2)}")
+                raise
 
     def _stream_with_fallback(
-        self, decorated_method, method_name, *args, **kwargs
+        self, decorated_method, method_name, decorators, *args, **kwargs
     ):
         """
         Wrapper generator that catches mid-stream errors and falls back.
@@ -154,6 +233,12 @@ class BaseLLM(ABC):
         try:
             yield from decorated_method()
         except Exception as e:
+            if self._is_non_retriable_client_error(e):
+                logger.error(
+                    f"Primary LLM failed mid-stream with non-retriable client "
+                    f"error; skipping fallback: {str(e)}"
+                )
+                raise
             if not self.fallback_llm:
                 logger.error(
                     f"Primary LLM failed and no fallback configured: {str(e)}"
@@ -164,11 +249,37 @@ class BaseLLM(ABC):
                 f"Primary LLM failed mid-stream. Falling back to "
                 f"{fallback.model_id}. Error: {str(e)}"
             )
-            fallback_method = getattr(
-                fallback, method_name.replace("_raw_", "")
+            # Apply decorators to fallback's raw stream method directly —
+            # calling fallback.gen_stream() would re-enter the orchestrator
+            # and recurse via fallback.fallback_llm. Emit the stream-start
+            # event manually so dashboards still see the fallback's
+            # provider/model when the response actually comes from it.
+            fallback._emit_stream_start_log(
+                fallback.model_id,
+                kwargs.get("messages"),
+                kwargs.get("tools"),
+                bool(
+                    kwargs.get("_usage_attachments")
+                    or kwargs.get("attachments")
+                ),
             )
+            fallback_method = getattr(fallback, method_name)
+            for decorator in decorators:
+                fallback_method = decorator(fallback_method)
             fallback_kwargs = {**kwargs, "model": fallback.model_id}
-            yield from fallback_method(*args, **fallback_kwargs)
+            try:
+                yield from fallback_method(fallback, *args, **fallback_kwargs)
+            except Exception as e2:
+                if self._is_non_retriable_client_error(e2):
+                    logger.error(
+                        f"Fallback LLM failed mid-stream with non-retriable "
+                        f"client error; giving up: {str(e2)}"
+                    )
+                else:
+                    logger.error(
+                        f"Fallback LLM also failed mid-stream; giving up: {str(e2)}"
+                    )
+                raise
 
     def gen(self, model, messages, stream=False, tools=None, *args, **kwargs):
         decorators = [gen_token_usage, gen_cache]
@@ -183,7 +294,58 @@ class BaseLLM(ABC):
             **kwargs,
         )
 
+    def _emit_stream_start_log(self, model, messages, tools, has_attachments):
+        # Stamped with ``self.provider_name`` so dashboards can group calls
+        # by vendor; the fallback path emits its own copy on the fallback
+        # instance so the actual responding provider is recorded.
+        logging.info(
+            "llm_stream_start",
+            extra={
+                "model": model,
+                "provider": self.provider_name,
+                "message_count": len(messages) if messages is not None else 0,
+                "has_attachments": bool(has_attachments),
+                "has_tools": bool(tools),
+            },
+        )
+
+    def _emit_stream_finished_log(
+        self,
+        model,
+        *,
+        prompt_tokens,
+        completion_tokens,
+        latency_ms,
+        cached_tokens=None,
+        error=None,
+    ):
+        # Paired with ``llm_stream_start`` so cost dashboards can sum tokens
+        # by user/agent/provider. Token counts are client-side estimates
+        # from ``stream_token_usage``; vendor-reported counts (incl.
+        # ``cached_tokens`` for prompt caching) require per-provider
+        # extraction in each ``_raw_gen_stream`` and aren't wired yet.
+        extra = {
+            "model": model,
+            "provider": self.provider_name,
+            "prompt_tokens": int(prompt_tokens),
+            "completion_tokens": int(completion_tokens),
+            "latency_ms": int(latency_ms),
+            "status": "error" if error is not None else "ok",
+        }
+        if cached_tokens is not None:
+            extra["cached_tokens"] = int(cached_tokens)
+        if error is not None:
+            extra["error_class"] = type(error).__name__
+        logging.info("llm_stream_finished", extra=extra)
+
     def gen_stream(self, model, messages, stream=True, tools=None, *args, **kwargs):
+        # Attachments arrive as ``_usage_attachments`` from ``Agent._llm_gen``;
+        # the ``stream_token_usage`` decorator pops that key, but the log
+        # fires before the decorator runs so it's still in ``kwargs`` here.
+        has_attachments = bool(
+            kwargs.get("_usage_attachments") or kwargs.get("attachments")
+        )
+        self._emit_stream_start_log(model, messages, tools, has_attachments)
         decorators = [stream_cache, stream_token_usage]
         return self._execute_with_fallback(
             "_raw_gen_stream",

application/llm/docsgpt_provider.py

@@ -6,6 +6,8 @@ DOCSGPT_BASE_URL = "https://oai.arc53.com"
 DOCSGPT_MODEL = "docsgpt"
 
 class DocsGPTAPILLM(OpenAILLM):
+    provider_name = "docsgpt"
+
     def __init__(self, api_key=None, user_api_key=None, base_url=None, *args, **kwargs):
         super().__init__(
             api_key=DOCSGPT_API_KEY,

application/llm/google_ai.py

@@ -6,10 +6,13 @@ from google.genai import types
 from application.core.settings import settings
 
 from application.llm.base import BaseLLM
+from application.llm.handlers.google import _decode_thought_signature
 from application.storage.storage_creator import StorageCreator
 
 
 class GoogleLLM(BaseLLM):
+    provider_name = "google"
+
     def __init__(
         self, api_key=None, user_api_key=None, decoded_token=None, *args, **kwargs
     ):
@@ -79,24 +82,39 @@ class GoogleLLM(BaseLLM):
         for attachment in attachments:
             mime_type = attachment.get("mime_type")
 
-            if mime_type in self.get_supported_attachment_types():
-                try:
+            if mime_type not in self.get_supported_attachment_types():
+                continue
+            try:
+                # Images go inline as bytes per Google's guidance for
+                # requests under 20MB; the Files API can return before
+                # the upload reaches ACTIVE state and yield an empty URI.
+                if mime_type.startswith("image/"):
+                    file_bytes = self._read_attachment_bytes(attachment)
+                    files.append(
+                        {"file_bytes": file_bytes, "mime_type": mime_type}
+                    )
+                else:
                     file_uri = self._upload_file_to_google(attachment)
+                    if not file_uri:
+                        raise ValueError(
+                            f"Google Files API returned empty URI for "
+                            f"{attachment.get('path', 'unknown')}"
+                        )
                     logging.info(
                         f"GoogleLLM: Successfully uploaded file, got URI: {file_uri}"
                     )
                     files.append({"file_uri": file_uri, "mime_type": mime_type})
-                except Exception as e:
-                    logging.error(
-                        f"GoogleLLM: Error uploading file: {e}", exc_info=True
+            except Exception as e:
+                logging.error(
+                    f"GoogleLLM: Error processing attachment: {e}", exc_info=True
+                )
+                if "content" in attachment:
+                    prepared_messages[user_message_index]["content"].append(
+                        {
+                            "type": "text",
+                            "text": f"[File could not be processed: {attachment.get('path', 'unknown')}]",
+                        }
                     )
-                    if "content" in attachment:
-                        prepared_messages[user_message_index]["content"].append(
-                            {
-                                "type": "text",
-                                "text": f"[File could not be processed: {attachment.get('path', 'unknown')}]",
-                            }
-                        )
         if files:
             logging.info(f"GoogleLLM: Adding {len(files)} files to message")
             prepared_messages[user_message_index]["content"].append({"files": files})
@@ -112,7 +130,9 @@ class GoogleLLM(BaseLLM):
         Returns:
             str: Google AI file URI for the uploaded file.
         """
-        if "google_file_uri" in attachment:
+        # Truthy check, not membership: a poisoned cache row of "" or
+        # None must be treated as a miss and trigger a fresh upload.
+        if attachment.get("google_file_uri"):
             return attachment["google_file_uri"]
         file_path = attachment.get("path")
         if not file_path:
@@ -126,6 +146,10 @@ class GoogleLLM(BaseLLM):
                     file=local_path
                 ).uri,
             )
+            if not file_uri:
+                raise ValueError(
+                    f"Google Files API upload returned empty URI for {file_path}"
+                )
 
             # 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
@@ -159,6 +183,26 @@ class GoogleLLM(BaseLLM):
             logging.error(f"Error uploading file to Google AI: {e}", exc_info=True)
             raise
 
+    def _read_attachment_bytes(self, attachment):
+        """
+        Read attachment bytes from storage for inline transmission.
+
+        Args:
+            attachment (dict): Attachment dictionary with path and metadata.
+
+        Returns:
+            bytes: Raw file bytes.
+        """
+        file_path = attachment.get("path")
+        if not file_path:
+            raise ValueError("No file path provided in attachment")
+        if not self.storage.file_exists(file_path):
+            raise FileNotFoundError(f"File not found: {file_path}")
+        return self.storage.process_file(
+            file_path,
+            lambda local_path, **kwargs: open(local_path, "rb").read(),
+        )
+
     def _clean_messages_google(self, messages):
         """
         Convert OpenAI format messages to Google AI format and collect system prompts.
@@ -215,7 +259,7 @@ class GoogleLLM(BaseLLM):
                         except (_json.JSONDecodeError, TypeError):
                             args = {}
                     cleaned_args = self._remove_null_values(args)
-                    thought_sig = tc.get("thought_signature")
+                    thought_sig = _decode_thought_signature(tc.get("thought_signature"))
                     if thought_sig:
                         parts.append(
                             types.Part(
@@ -279,7 +323,9 @@ class GoogleLLM(BaseLLM):
                                             name=item["function_call"]["name"],
                                             args=cleaned_args,
                                         ),
-                                        thoughtSignature=item["thought_signature"],
+                                        thoughtSignature=_decode_thought_signature(
+                                            item["thought_signature"]
+                                        ),
                                     )
                                 )
                             else:
@@ -298,12 +344,24 @@ class GoogleLLM(BaseLLM):
                             )
                         elif "files" in item:
                             for file_data in item["files"]:
-                                parts.append(
-                                    types.Part.from_uri(
-                                        file_uri=file_data["file_uri"],
-                                        mime_type=file_data["mime_type"],
+                                if "file_bytes" in file_data:
+                                    parts.append(
+                                        types.Part.from_bytes(
+                                            data=file_data["file_bytes"],
+                                            mime_type=file_data["mime_type"],
+                                        )
+                                    )
+                                elif file_data.get("file_uri"):
+                                    parts.append(
+                                        types.Part.from_uri(
+                                            file_uri=file_data["file_uri"],
+                                            mime_type=file_data["mime_type"],
+                                        )
+                                    )
+                                else:
+                                    logging.warning(
+                                        "GoogleLLM: dropping file part with empty URI and no bytes"
                                     )
-                                )
                         else:
                             raise ValueError(
                                 f"Unexpected content dictionary format:{item}"
@@ -541,22 +599,6 @@ class GoogleLLM(BaseLLM):
             config.response_mime_type = "application/json"
         # Check if we have both tools and file attachments
 
-        has_attachments = False
-        for message in messages:
-            for part in message.parts:
-                if hasattr(part, "file_data") and part.file_data is not None:
-                    has_attachments = True
-                    break
-            if has_attachments:
-                break
-        messages_summary = self._summarize_messages_for_log(messages)
-        logging.info(
-            "GoogleLLM: Starting stream generation. Model: %s, Messages: %s, Has attachments: %s",
-            model,
-            messages_summary,
-            has_attachments,
-        )
-
         response = client.models.generate_content_stream(
             model=model,
             contents=messages,

application/llm/groq.py

@@ -5,6 +5,8 @@ GROQ_BASE_URL = "https://api.groq.com/openai/v1"
 
 
 class GroqLLM(OpenAILLM):
+    provider_name = "groq"
+
     def __init__(self, api_key=None, user_api_key=None, base_url=None, *args, **kwargs):
         super().__init__(
             api_key=api_key or settings.GROQ_API_KEY or settings.API_KEY,

application/llm/handlers/base.py

@@ -10,6 +10,18 @@ from application.logging import build_stack_data
 logger = logging.getLogger(__name__)
 
 
+# Cap the agent tool-call loop. Without this an LLM that keeps
+# requesting more tool calls (preview models, sparse tool results,
+# under-specified prompts) can chain searches indefinitely and the
+# stream never finalises. 25 mirrors Dify's default.
+MAX_TOOL_ITERATIONS = 25
+_FINALIZE_INSTRUCTION = (
+    f"You have made {MAX_TOOL_ITERATIONS} tool calls. Provide a final "
+    "response to the user based on what you have, without making any "
+    "additional tool calls."
+)
+
+
 @dataclass
 class ToolCall:
     """Represents a tool/function call from the LLM."""
@@ -280,7 +292,26 @@ class LLMHandler(ABC):
                         # Keep serialized function calls/responses so the compressor sees actions
                         parts_text.append(str(item))
                     elif "files" in item:
-                        parts_text.append(str(item))
+                        # Image attachments arrive with raw bytes / base64
+                        # inline (see GoogleLLM.prepare_messages_with_attachments).
+                        # ``str(item)`` would dump the whole byte/base64
+                        # blob into the compression prompt and bust the
+                        # compression LLM's input limit.
+                        files = item.get("files") or []
+                        descriptors = []
+                        if isinstance(files, list):
+                            for f in files:
+                                if isinstance(f, dict):
+                                    descriptors.append(
+                                        f.get("mime_type") or "file"
+                                    )
+                                elif isinstance(f, str):
+                                    descriptors.append(f)
+                        if not descriptors:
+                            descriptors = ["file"]
+                        parts_text.append(
+                            f"[attachment: {', '.join(descriptors)}]"
+                        )
             return "\n".join(parts_text)
         return ""
 
@@ -470,10 +501,14 @@ 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 +612,20 @@ 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,
@@ -586,7 +634,12 @@ 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,
             )
+            # Side-channel LLM tag — see ``orchestrator.py`` for rationale.
+            compression_llm._token_usage_source = "compression"
+            compression_llm._request_id = getattr(agent, "_request_id", None) \
+                or getattr(getattr(agent, "llm", None), "_request_id", None)
 
             # Create service without DB persistence capability
             compression_service = CompressionService(
@@ -897,7 +950,9 @@ class LLMHandler(ABC):
         parsed = self.parse_response(response)
         self.llm_calls.append(build_stack_data(agent.llm))
 
+        iteration = 0
         while parsed.requires_tool_call:
+            iteration += 1
             tool_handler_gen = self.handle_tool_calls(
                 agent, parsed.tool_calls, tools_dict, messages
             )
@@ -921,15 +976,46 @@ class LLMHandler(ABC):
                 }
                 return ""
 
+            # Cap reached: force one final tool-less call so the stream
+            # always ends with content rather than cutting off.
+            if iteration >= MAX_TOOL_ITERATIONS:
+                logger.warning(
+                    "agent tool loop hit cap (%d); forcing finalize",
+                    MAX_TOOL_ITERATIONS,
+                )
+                messages.append(
+                    {"role": "system", "content": _FINALIZE_INSTRUCTION},
+                )
+                response = agent.llm.gen(
+                    model=getattr(agent.llm, "model_id", None) or agent.model_id,
+                    messages=messages,
+                    tools=None,
+                )
+                parsed = self.parse_response(response)
+                self.llm_calls.append(build_stack_data(agent.llm))
+                break
+
+            # ``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))
         return parsed.content
 
     def handle_streaming(
-        self, agent, response: Any, tools_dict: Dict, messages: List[Dict]
+        self,
+        agent,
+        response: Any,
+        tools_dict: Dict,
+        messages: List[Dict],
+        _iteration: int = 0,
     ) -> Generator:
         """
         Handle streaming response flow.
@@ -998,6 +1084,9 @@ class LLMHandler(ABC):
                     }
                     return
 
+                next_iteration = _iteration + 1
+                cap_reached = next_iteration >= MAX_TOOL_ITERATIONS
+
                 # Check if context limit was reached during tool execution
                 if hasattr(agent, 'context_limit_reached') and agent.context_limit_reached:
                     # Add system message warning about context limit
@@ -1010,13 +1099,32 @@ class LLMHandler(ABC):
                         )
                     })
                     logger.info("Context limit reached - instructing agent to wrap up")
+                elif cap_reached:
+                    logger.warning(
+                        "agent tool loop hit cap (%d); forcing finalize",
+                        MAX_TOOL_ITERATIONS,
+                    )
+                    messages.append(
+                        {"role": "system", "content": _FINALIZE_INSTRUCTION},
+                    )
 
+                # 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=(
+                        None
+                        if cap_reached
+                        or getattr(agent, "context_limit_reached", False)
+                        else agent.tools
+                    ),
                 )
                 self.llm_calls.append(build_stack_data(agent.llm))
 
-                yield from self.handle_streaming(agent, response, tools_dict, messages)
+                yield from self.handle_streaming(
+                    agent, response, tools_dict, messages,
+                    _iteration=next_iteration,
+                )
                 return
             if parsed.content:
                 buffer += parsed.content

application/llm/handlers/google.py

@@ -1,9 +1,35 @@
+import base64
+import binascii
 import uuid
-from typing import Any, Dict, Generator
+from typing import Any, Dict, Generator, Optional, Union
 
 from application.llm.handlers.base import LLMHandler, LLMResponse, ToolCall
 
 
+def _encode_thought_signature(sig: Optional[Union[bytes, str]]) -> Optional[str]:
+    # Gemini's Python SDK returns thought_signature as raw bytes, but the
+    # field is typed Optional[str] downstream and gets json.dumps'd into
+    # SSE events. Encode once at ingress so callers only ever see a str.
+    if isinstance(sig, bytes):
+        return base64.b64encode(sig).decode("ascii")
+    return sig
+
+
+def _decode_thought_signature(
+    sig: Optional[Union[bytes, str]],
+) -> Optional[Union[bytes, str]]:
+    # Reverse of _encode_thought_signature — Gemini's SDK expects bytes
+    # back when we replay a tool call. ``validate=True`` keeps ASCII
+    # strings that happen to be loosely decodable from being silently
+    # turned into bytes; non-base64 inputs pass through unchanged.
+    if isinstance(sig, str):
+        try:
+            return base64.b64decode(sig.encode("ascii"), validate=True)
+        except (binascii.Error, ValueError):
+            return sig
+    return sig
+
+
 class GoogleLLMHandler(LLMHandler):
     """Handler for Google's GenAI API."""
 
@@ -23,7 +49,7 @@ class GoogleLLMHandler(LLMHandler):
             for idx, part in enumerate(parts):
                 if hasattr(part, "function_call") and part.function_call is not None:
                     has_sig = hasattr(part, "thought_signature") and part.thought_signature is not None
-                    thought_sig = part.thought_signature if has_sig else None
+                    thought_sig = _encode_thought_signature(part.thought_signature) if has_sig else None
                     tool_calls.append(
                         ToolCall(
                             id=str(uuid.uuid4()),
@@ -50,7 +76,7 @@ class GoogleLLMHandler(LLMHandler):
             tool_calls = []
             if hasattr(response, "function_call") and response.function_call is not None:
                 has_sig = hasattr(response, "thought_signature") and response.thought_signature is not None
-                thought_sig = response.thought_signature if has_sig else None
+                thought_sig = _encode_thought_signature(response.thought_signature) if has_sig else None
                 tool_calls.append(
                     ToolCall(
                         id=str(uuid.uuid4()),
@@ -70,8 +96,15 @@ class GoogleLLMHandler(LLMHandler):
         """Create a tool result message in the standard internal format."""
         import json as _json
 
+        from application.storage.db.serialization import PGNativeJSONEncoder
+
+        # PostgresTool results commonly include PG-native types
+        # (datetime / UUID / Decimal / bytea) when SELECT touches
+        # timestamptz / numeric / uuid / bytea columns. The shared
+        # encoder handles all five — bytes get base64 (lossless) instead
+        # of the ``str(b'...')`` repr that ``default=str`` would emit.
         content = (
-            _json.dumps(result)
+            _json.dumps(result, cls=PGNativeJSONEncoder)
             if not isinstance(result, str)
             else result
         )

application/llm/handlers/openai.py

@@ -40,8 +40,15 @@ class OpenAILLMHandler(LLMHandler):
         """Create a tool result message in the standard internal format."""
         import json as _json
 
+        from application.storage.db.serialization import PGNativeJSONEncoder
+
+        # PostgresTool results commonly include PG-native types
+        # (datetime / UUID / Decimal / bytea) when SELECT touches
+        # timestamptz / numeric / uuid / bytea columns. The shared
+        # encoder handles all five — bytes get base64 (lossless) instead
+        # of the ``str(b'...')`` repr that ``default=str`` would emit.
         content = (
-            _json.dumps(result)
+            _json.dumps(result, cls=PGNativeJSONEncoder)
             if not isinstance(result, str)
             else result
         )

application/llm/llama_cpp.py

@@ -26,6 +26,8 @@ class LlamaSingleton:
 
 
 class LlamaCpp(BaseLLM):
+    provider_name = "llama_cpp"
+
     def __init__(
         self,
         api_key=None,

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/novita.py

@@ -5,6 +5,8 @@ NOVITA_BASE_URL = "https://api.novita.ai/openai"
 
 
 class NovitaLLM(OpenAILLM):
+    provider_name = "novita"
+
     def __init__(self, api_key=None, user_api_key=None, base_url=None, *args, **kwargs):
         super().__init__(
             api_key=api_key or settings.NOVITA_API_KEY or settings.API_KEY,

application/llm/open_router.py

@@ -5,6 +5,8 @@ OPEN_ROUTER_BASE_URL = "https://openrouter.ai/api/v1"
 
 
 class OpenRouterLLM(OpenAILLM):
+    provider_name = "openrouter"
+
     def __init__(self, api_key=None, user_api_key=None, base_url=None, *args, **kwargs):
         super().__init__(
             api_key=api_key or settings.OPEN_ROUTER_API_KEY or settings.API_KEY,

application/llm/openai.py

@@ -61,8 +61,17 @@ def _truncate_base64_for_logging(messages):
 
 
 class OpenAILLM(BaseLLM):
+    provider_name = "openai"
 
-    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 +89,18 @@ 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 +263,13 @@ 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 +306,13 @@ 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 +354,17 @@ 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 +431,14 @@ 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):
         """

application/llm/premai.py

@@ -3,6 +3,7 @@ from application.core.settings import settings
 
 
 class PremAILLM(BaseLLM):
+    provider_name = "premai"
 
     def __init__(self, api_key=None, user_api_key=None, *args, **kwargs):
         from premai import Prem

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/llm/sagemaker.py

@@ -59,6 +59,7 @@ class LineIterator:
 
 
 class SagemakerAPILLM(BaseLLM):
+    provider_name = "sagemaker"
 
     def __init__(self, api_key=None, user_api_key=None, *args, **kwargs):
         import boto3

application/logging.py

@@ -1,11 +1,13 @@
 import datetime
 import functools
 import inspect
+import time
 
 import logging
 import uuid
 from typing import Any, Callable, Dict, Generator, List
 
+from application.core import log_context
 from application.storage.db.repositories.stack_logs import StackLogsRepository
 from application.storage.db.session import db_session
 
@@ -22,6 +24,15 @@ class LogContext:
         self.api_key = api_key
         self.query = query
         self.stacks = []
+        # Per-activity response aggregates populated by ``_consume_and_log``
+        # while it forwards stream items, then flushed onto the
+        # ``activity_finished`` event so every Flask request gets the
+        # same summary that ``run_agent_logic`` used to log only for the
+        # Celery webhook path.
+        self.answer_length = 0
+        self.thought_length = 0
+        self.source_count = 0
+        self.tool_call_count = 0
 
 
 def build_stack_data(
@@ -78,25 +89,125 @@ def log_activity() -> Callable:
             user = data.get("user", "local")
             api_key = data.get("user_api_key", "")
             query = kwargs.get("query", getattr(args[0], "query", ""))
+            agent_id = getattr(args[0], "agent_id", None) or kwargs.get("agent_id")
+            conversation_id = (
+                kwargs.get("conversation_id")
+                or getattr(args[0], "conversation_id", None)
+            )
+            model = getattr(args[0], "gpt_model", None) or getattr(args[0], "model", None)
+
+            # Capture the surrounding activity_id before overlaying ours,
+            # so nested activities record the parent → child link.
+            parent_activity_id = log_context.snapshot().get("activity_id")
 
             context = LogContext(endpoint, activity_id, user, api_key, query)
             kwargs["log_context"] = context
 
-            logging.info(
-                f"Starting activity: {endpoint} - {activity_id} - User: {user}"
+            ctx_token = log_context.bind(
+                activity_id=activity_id,
+                parent_activity_id=parent_activity_id,
+                user_id=user,
+                agent_id=agent_id,
+                conversation_id=conversation_id,
+                endpoint=endpoint,
+                model=model,
             )
 
-            generator = func(*args, **kwargs)
-            yield from _consume_and_log(generator, context)
+            started_at = time.monotonic()
+            logging.info(
+                "activity_started",
+                extra={
+                    "activity_id": activity_id,
+                    "parent_activity_id": parent_activity_id,
+                    "user_id": user,
+                    "agent_id": agent_id,
+                    "conversation_id": conversation_id,
+                    "endpoint": endpoint,
+                    "model": model,
+                },
+            )
+
+            error: BaseException | None = None
+            try:
+                generator = func(*args, **kwargs)
+                yield from _consume_and_log(generator, context)
+            except Exception as exc:
+                # Only ``Exception`` counts as an activity error; ``GeneratorExit``
+                # (consumer disconnected mid-stream) and ``KeyboardInterrupt``
+                # flow through the finally as ``status="ok"``, matching
+                # ``_consume_and_log``.
+                error = exc
+                raise
+            finally:
+                _emit_activity_finished(
+                    context=context,
+                    parent_activity_id=parent_activity_id,
+                    started_at=started_at,
+                    error=error,
+                )
+                log_context.reset(ctx_token)
 
         return wrapper
 
     return decorator
 
 
+def _emit_activity_finished(
+    *,
+    context: "LogContext",
+    parent_activity_id: str | None,
+    started_at: float,
+    error: BaseException | None,
+) -> None:
+    """Emit the paired ``activity_finished`` event with duration, outcome,
+    and per-activity response aggregates accumulated in ``_consume_and_log``.
+    """
+    duration_ms = int((time.monotonic() - started_at) * 1000)
+    logging.info(
+        "activity_finished",
+        extra={
+            "activity_id": context.activity_id,
+            "parent_activity_id": parent_activity_id,
+            "user_id": context.user,
+            "endpoint": context.endpoint,
+            "duration_ms": duration_ms,
+            "status": "error" if error is not None else "ok",
+            "error_class": type(error).__name__ if error is not None else None,
+            "answer_length": context.answer_length,
+            "thought_length": context.thought_length,
+            "source_count": context.source_count,
+            "tool_call_count": context.tool_call_count,
+        },
+    )
+
+
+def _accumulate_response_summary(item: Any, context: "LogContext") -> None:
+    """Mirror the per-line aggregation that ``run_agent_logic`` did for the
+    Celery webhook path, but at the generator-consumption layer so every
+    ``Agent.gen`` activity (Flask streaming, sub-agents, workflow agents)
+    gets the same summary.
+    """
+    if not isinstance(item, dict):
+        return
+    if "answer" in item:
+        context.answer_length += len(str(item["answer"]))
+        return
+    if "thought" in item:
+        context.thought_length += len(str(item["thought"]))
+        return
+    sources = item.get("sources") if "sources" in item else None
+    if isinstance(sources, list):
+        context.source_count += len(sources)
+        return
+    tool_calls = item.get("tool_calls") if "tool_calls" in item else None
+    if isinstance(tool_calls, list):
+        context.tool_call_count += len(tool_calls)
+
+
 def _consume_and_log(generator: Generator, context: "LogContext"):
     try:
         for item in generator:
+            _accumulate_response_summary(item, context)
             yield item
     except Exception as e:
         logging.exception(f"Error in {context.endpoint} - {context.activity_id}: {e}")

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/embedding_pipeline.py

@@ -1,12 +1,27 @@
 import os
 import logging
-from typing import List, Any
+from typing import Any, List, Optional
 from retry import retry
 from tqdm import tqdm
 from application.core.settings import settings
+from application.storage.db.repositories.ingest_chunk_progress import (
+    IngestChunkProgressRepository,
+)
+from application.storage.db.session import db_session
 from application.vectorstore.vector_creator import VectorCreator
 
 
+class EmbeddingPipelineError(Exception):
+    """Raised when the per-chunk embed loop produces a partial index.
+
+    Escapes into Celery's ``autoretry_for`` so a transient cause (rate
+    limit, network blip) gets another shot. The chunk-progress
+    checkpoint makes retries cheap — only the failed-and-after chunks
+    re-run. After ``MAX_TASK_ATTEMPTS`` the poison-loop guard in
+    ``with_idempotency`` finalises the row as ``failed``.
+    """
+
+
 def sanitize_content(content: str) -> str:
     """
     Remove NUL characters that can cause vector store ingestion to fail.
@@ -22,7 +37,11 @@ def sanitize_content(content: str) -> str:
     return content.replace('\x00', '')
 
 
-@retry(tries=10, delay=60)
+# Per-chunk inline retry. Aggressive defaults (tries=10, delay=60) blocked
+# the loop for up to 9 min per chunk and wedged the heartbeat: lower the
+# tail so a transient failure fails-fast and the chunk-progress checkpoint
+# resumes cleanly on next dispatch.
+@retry(tries=3, delay=5, backoff=2)
 def add_text_to_store_with_retry(store: Any, doc: Any, source_id: str) -> None:
     """Add a document's text and metadata to the vector store with retry logic.
     
@@ -45,21 +64,119 @@ def add_text_to_store_with_retry(store: Any, doc: Any, source_id: str) -> None:
         raise
 
 
-def embed_and_store_documents(docs: List[Any], folder_name: str, source_id: str, task_status: Any) -> None:
+def _init_progress_and_resume_index(
+    source_id: str, total_chunks: int, attempt_id: Optional[str],
+) -> int:
+    """Upsert the progress row and return the next chunk index to embed.
+
+    The repository's upsert preserves ``last_index`` only when the
+    incoming ``attempt_id`` matches the stored one (a Celery autoretry
+    of the same task). On a fresh attempt — including any caller that
+    doesn't pass an ``attempt_id``, e.g. legacy code or tests — the
+    row's checkpoint is reset so the loop starts from chunk 0. This
+    is what prevents a completed checkpoint from any prior run
+    silently no-op'ing the next sync/reingest.
+
+    Best-effort: a DB outage falls back to ``0`` (fresh run from
+    chunk 0). The embed loop's own re-raise still ensures partial
+    runs don't get cached as complete.
+    """
+    try:
+        with db_session() as conn:
+            progress = IngestChunkProgressRepository(conn).init_progress(
+                source_id, total_chunks, attempt_id,
+            )
+    except Exception as e:
+        logging.warning(
+            f"Could not init ingest progress for {source_id}: {e}",
+            exc_info=True,
+        )
+        return 0
+    if not progress:
+        return 0
+    last_index = progress.get("last_index", -1)
+    if last_index is None or last_index < 0:
+        return 0
+    return int(last_index) + 1
+
+
+def _record_progress(source_id: str, last_index: int, embedded_chunks: int) -> None:
+    """Best-effort checkpoint after each chunk; logged but never raised."""
+    try:
+        with db_session() as conn:
+            IngestChunkProgressRepository(conn).record_chunk(
+                source_id, last_index=last_index, embedded_chunks=embedded_chunks
+            )
+    except Exception as e:
+        logging.warning(
+            f"Could not record ingest progress for {source_id}: {e}", exc_info=True
+        )
+
+
+def assert_index_complete(source_id: str) -> None:
+    """Raise ``EmbeddingPipelineError`` if ``ingest_chunk_progress``
+    shows a partial embed for ``source_id``.
+
+    Defense-in-depth tripwire that workers run after
+    ``embed_and_store_documents`` to catch any future swallow path
+    that bypasses the function's own re-raise — the chunk-progress
+    row is the authoritative record of how many chunks landed.
+    No-op when no row exists (zero-doc validation raised before init,
+    or progress repo was unreachable).
+    """
+    try:
+        with db_session() as conn:
+            progress = IngestChunkProgressRepository(conn).get_progress(source_id)
+    except Exception as e:
+        logging.warning(
+            f"assert_index_complete: progress lookup failed for "
+            f"{source_id}: {e}",
+            exc_info=True,
+        )
+        return
+    if not progress:
+        return
+    embedded = int(progress.get("embedded_chunks") or 0)
+    total = int(progress.get("total_chunks") or 0)
+    if embedded < total:
+        raise EmbeddingPipelineError(
+            f"partial index for source {source_id}: "
+            f"{embedded}/{total} chunks embedded"
+        )
+
+
+def embed_and_store_documents(
+    docs: List[Any],
+    folder_name: str,
+    source_id: str,
+    task_status: Any,
+    *,
+    attempt_id: Optional[str] = None,
+) -> None:
     """Embeds documents and stores them in a vector store.
 
+    Resumable across Celery autoretries of the *same* task: when
+    ``attempt_id`` matches the stored checkpoint's ``attempt_id``,
+    the loop resumes from ``last_index + 1``. A different
+    ``attempt_id`` (a fresh sync / reingest invocation) resets the
+    checkpoint so the index is rebuilt from chunk 0 — this is what
+    keeps a completed checkpoint from poisoning the next sync.
+
     Args:
         docs: List of documents to be embedded and stored.
         folder_name: Directory to save the vector store.
         source_id: Unique identifier for the source.
         task_status: Task state manager for progress updates.
+        attempt_id: Stable id of the current task invocation,
+            typically ``self.request.id`` from the Celery task body.
+            ``None`` is treated as a fresh attempt every time.
 
     Returns:
         None
-        
+
     Raises:
         OSError: If unable to create folder or save vector store.
-        Exception: If vector store creation or document embedding fails.
+        EmbeddingPipelineError: If a chunk fails after retries.
     """
     # Ensure the folder exists
     if not os.path.exists(folder_name):
@@ -69,33 +186,77 @@ def embed_and_store_documents(docs: List[Any], folder_name: str, source_id: str,
     if not docs:
         raise ValueError("No documents to embed - check file format and extension")
 
+    total_docs = len(docs)
+    # Atomic upsert that preserves checkpoint state on attempt-id match
+    # (autoretry of same task) and resets it on mismatch (fresh sync /
+    # reingest). Returns the new resume index — 0 means "start fresh".
+    resume_index = _init_progress_and_resume_index(
+        source_id, total_docs, attempt_id,
+    )
+    is_resume = resume_index > 0
+
     # Initialize vector store
     if settings.VECTOR_STORE == "faiss":
-        docs_init = [docs.pop(0)]
-        store = VectorCreator.create_vectorstore(
-            settings.VECTOR_STORE,
-            docs_init=docs_init,
-            source_id=source_id,
-            embeddings_key=os.getenv("EMBEDDINGS_KEY"),
-        )
+        if is_resume:
+            # Load the existing FAISS index from storage so chunks
+            # already embedded by the prior attempt survive the
+            # save_local rewrite at the end of this run.
+            store = VectorCreator.create_vectorstore(
+                settings.VECTOR_STORE,
+                source_id=source_id,
+                embeddings_key=os.getenv("EMBEDDINGS_KEY"),
+            )
+            loop_start = resume_index
+        else:
+            # FAISS requires at least one doc to construct the store;
+            # seed with ``docs[0]`` and let the loop pick up at index 1.
+            store = VectorCreator.create_vectorstore(
+                settings.VECTOR_STORE,
+                docs_init=[docs[0]],
+                source_id=source_id,
+                embeddings_key=os.getenv("EMBEDDINGS_KEY"),
+            )
+            # Record the seeded chunk so single-doc ingests don't fail
+            # ``assert_index_complete`` — the loop never runs for
+            # ``total_docs == 1`` and would otherwise leave
+            # ``embedded_chunks`` at 0 / ``last_index`` at -1. The loop
+            # body's per-iteration ``_record_progress`` overshoots
+            # correctly for multi-chunk runs (counts seed + iterations),
+            # so writing this checkpoint up-front is a no-op for those.
+            _record_progress(source_id, last_index=0, embedded_chunks=1)
+            loop_start = 1
     else:
         store = VectorCreator.create_vectorstore(
             settings.VECTOR_STORE,
             source_id=source_id,
             embeddings_key=os.getenv("EMBEDDINGS_KEY"),
         )
-        store.delete_index()
+        # Only wipe the index on a fresh run — a resume must keep the
+        # chunks that earlier attempts already embedded.
+        if not is_resume:
+            store.delete_index()
+        loop_start = resume_index
 
-    total_docs = len(docs)
+    if is_resume and loop_start >= total_docs:
+        # Nothing left to do; the loop runs zero iterations and
+        # downstream finalize logic still executes. This is only
+        # reachable on a same-attempt retry of a task whose previous
+        # attempt finished — typically a Celery acks_late redelivery
+        # after the task already returned. The ``assert_index_complete``
+        # tripwire still validates ``embedded == total`` afterwards.
+        loop_start = total_docs
 
     # Process and embed documents
-    for idx, doc in tqdm(
-        enumerate(docs),
+    chunk_error: Exception | None = None
+    failed_idx: int | None = None
+    for idx in tqdm(
+        range(loop_start, total_docs),
         desc="Embedding 🦖",
         unit="docs",
-        total=total_docs,
+        total=total_docs - loop_start,
         bar_format="{l_bar}{bar}| Time Left: {remaining}",
     ):
+        doc = docs[idx]
         try:
             # Update task status for progress tracking
             progress = int(((idx + 1) / total_docs) * 100)
@@ -103,7 +264,10 @@ def embed_and_store_documents(docs: List[Any], folder_name: str, source_id: str,
 
             # Add document to vector store
             add_text_to_store_with_retry(store, doc, source_id)
+            _record_progress(source_id, last_index=idx, embedded_chunks=idx + 1)
         except Exception as e:
+            chunk_error = e
+            failed_idx = idx
             logging.error(f"Error embedding document {idx}: {e}", exc_info=True)
             logging.info(f"Saving progress at document {idx} out of {total_docs}")
             try:
@@ -124,3 +288,16 @@ def embed_and_store_documents(docs: List[Any], folder_name: str, source_id: str,
             raise OSError(f"Unable to save vector store to {folder_name}: {e}") from e
     else:
         logging.info("Vector store saved successfully.")
+
+    # Re-raise after the partial save: the chunks that *did* embed are
+    # flushed to disk and recorded in ``ingest_chunk_progress``, so a
+    # Celery autoretry resumes via ``_read_resume_index`` and only
+    # re-runs the failed-and-after chunks. Without the raise, the
+    # task body returns success and ``with_idempotency`` finalises
+    # ``task_dedup`` as ``completed`` for a partial index — poisoning
+    # the cache for 24h.
+    if chunk_error is not None:
+        raise EmbeddingPipelineError(
+            f"embed failure at chunk {failed_idx}/{total_docs} "
+            f"for source {source_id}"
+        ) from chunk_error

application/requirements.txt

@@ -1,9 +1,12 @@
+a2wsgi==1.10.10
 alembic>=1.13,<2
 anthropic==0.88.0
+asgiref>=3.11.1
 boto3==1.42.83
 beautifulsoup4==4.14.3
 cel-python==0.5.0
 celery==5.6.3
+celery-redbeat==2.3.3
 cryptography==46.0.7
 dataclasses-json==0.6.7
 defusedxml==0.7.1
@@ -14,7 +17,7 @@ docx2txt==0.9
 ddgs>=8.0.0
 fast-ebook
 elevenlabs==2.43.0
-Flask==3.1.3
+Flask==3.1.1
 faiss-cpu==1.13.2
 fastmcp==3.2.4
 flask-restx==1.3.2
@@ -35,7 +38,7 @@ langchain==1.2.3
 langchain-community==0.4.1
 langchain-core==1.2.29
 langchain-openai==1.1.12
-langchain-text-splitters==1.1.2
+langchain-text-splitters==1.1.1
 langsmith==0.7.31
 lazy-object-proxy==1.12.0
 lxml==6.0.2
@@ -49,6 +52,16 @@ networkx==3.6.1
 numpy==2.4.4
 openai==2.32.0
 openapi3-parser==1.1.22
+opentelemetry-distro>=0.50b0,<1
+opentelemetry-exporter-otlp>=1.29.0,<2
+opentelemetry-instrumentation-celery>=0.50b0,<1
+opentelemetry-instrumentation-flask>=0.50b0,<1
+opentelemetry-instrumentation-logging>=0.50b0,<1
+opentelemetry-instrumentation-psycopg>=0.50b0,<1
+opentelemetry-instrumentation-redis>=0.50b0,<1
+opentelemetry-instrumentation-requests>=0.50b0,<1
+opentelemetry-instrumentation-sqlalchemy>=0.50b0,<1
+opentelemetry-instrumentation-starlette>=0.50b0,<1
 orjson==3.11.7
 packaging==26.0
 pandas==3.0.2
@@ -58,7 +71,7 @@ pdf2image>=1.17.0
 pillow
 portalocker>=2.7.0,<4.0.0
 prompt-toolkit==3.0.52
-protobuf==7.34.1
+protobuf==6.33.6
 psycopg[binary,pool]>=3.1,<4
 py==1.11.0
 pydantic
@@ -69,6 +82,7 @@ python-dateutil==2.9.0.post0
 python-dotenv
 python-jose==3.5.0
 python-pptx==1.0.2
+PyYAML
 redis==7.4.0
 referencing>=0.28.0,<0.38.0
 regex==2026.4.4
@@ -76,6 +90,7 @@ requests==2.33.1
 retry==0.9.2
 sentence-transformers==5.3.0
 sqlalchemy>=2.0,<3
+starlette>=1.0,<2
 tiktoken==0.12.0
 tokenizers==0.22.2
 torch==2.11.0
@@ -85,6 +100,8 @@ typing-extensions==4.15.0
 typing-inspect==0.9.0
 tzdata==2026.1
 urllib3==2.6.3
+uvicorn[standard]>=0.30,<1
+uvicorn-worker>=0.4,<1
 vine==5.1.0
 wcwidth==0.6.0
 werkzeug>=3.1.0

application/retriever/classic_rag.py

@@ -22,6 +22,7 @@ class ClassicRAG(BaseRetriever):
         llm_name=settings.LLM_PROVIDER,
         api_key=settings.API_KEY,
         decoded_token=None,
+        model_user_id=None,
     ):
         self.original_question = source.get("question", "")
         self.chat_history = chat_history if chat_history is not None else []
@@ -42,18 +43,26 @@ class ClassicRAG(BaseRetriever):
             f"sources={'active_docs' in source and source['active_docs'] is not None}"
         )
         self.model_id = model_id
+        self.model_user_id = model_user_id
         self.doc_token_limit = doc_token_limit
         self.user_api_key = user_api_key
         self.agent_id = agent_id
         self.llm_name = llm_name
         self.api_key = api_key
+        # Forward model_id + model_user_id so LLMCreator resolves BYOM
+        # base_url / api_key / upstream id for the rephrase client.
         self.llm = LLMCreator.create_llm(
             self.llm_name,
             api_key=self.api_key,
             user_api_key=self.user_api_key,
             decoded_token=decoded_token,
+            model_id=self.model_id,
             agent_id=self.agent_id,
+            model_user_id=self.model_user_id,
         )
+        # Query-rephrase LLM is a side channel — tag it so its rows
+        # land as ``source='rag_condense'`` in cost-attribution.
+        self.llm._token_usage_source = "rag_condense"
 
         if "active_docs" in source and source["active_docs"] is not None:
             if isinstance(source["active_docs"], list):
@@ -103,7 +112,11 @@ class ClassicRAG(BaseRetriever):
         ]
 
         try:
-            rephrased_query = self.llm.gen(model=self.model_id, messages=messages)
+            # Send upstream id (resolved by LLMCreator), not registry UUID.
+            rephrased_query = self.llm.gen(
+                model=getattr(self.llm, "model_id", None) or self.model_id,
+                messages=messages,
+            )
             print(f"Rephrased query: {rephrased_query}")
             return rephrased_query if rephrased_query else self.original_question
         except Exception as e:

application/security/safe_url.py

@@ -0,0 +1,464 @@
+"""SSRF protection for user-supplied OpenAI-compatible base URLs.
+
+This module is the single chokepoint for validating any URL that a user
+provides as an OpenAI-compatible ``base_url`` ("Bring Your Own Model").
+The backend will later issue outbound HTTP requests to that URL on the
+user's behalf, so we must reject anything that could be used to reach
+internal-network resources (cloud metadata services, RFC 1918 ranges,
+loopback, link-local, etc.).
+
+Three entry points:
+
+* :func:`validate_user_base_url` — called at create/update time on REST
+  routes that persist the URL, to give the user immediate feedback.
+* :func:`pinned_post` — called at dispatch time when the caller drives
+  ``requests`` directly (e.g. the ``/api/models/test`` endpoint).
+  Resolves once, dials the IP literal, preserves the original hostname
+  in the ``Host`` header and via SNI / cert verification for HTTPS.
+* :func:`pinned_httpx_client` — called at dispatch time when the caller
+  hands an ``httpx.Client`` to a third-party SDK (e.g. the OpenAI
+  Python SDK via ``OpenAI(http_client=...)``). Same DNS-rebinding
+  closure on the httpx transport layer.
+
+Why all three: the OpenAI / httpx ecosystem performs its own DNS lookup
+inside ``socket.getaddrinfo`` when a connection opens, so a hostile DNS
+server can hand a public IP to the validator and a loopback / link-local
+address to the HTTP client. Validate-then-construct-SDK is unsafe; the
+pinned variants close that TOCTOU window by resolving exactly once and
+dialing the chosen IP literal directly.
+"""
+
+from __future__ import annotations
+
+import ipaddress
+import socket
+from typing import Any, Iterable
+from urllib.parse import urlsplit, urlunsplit
+
+import httpx
+import requests
+from requests.adapters import HTTPAdapter
+
+# Allowed URL schemes. Anything else (file, gopher, ftp, data, ...) is
+# rejected outright because it either bypasses HTTP entirely or enables
+# protocol smuggling against the proxy stack.
+_ALLOWED_SCHEMES: frozenset[str] = frozenset({"http", "https"})
+
+# Hostnames that resolve to a loopback / metadata / unspecified address
+# but which we want to reject *by name* as well, so the rejection
+# message is unambiguous and so we never accidentally call DNS on them.
+_BLOCKED_HOSTNAMES: frozenset[str] = frozenset(
+    {
+        "localhost",
+        "localhost.localdomain",
+        "0.0.0.0",
+        "::",
+        "::1",
+        "ip6-localhost",
+        "ip6-loopback",
+        # GCP metadata service. AWS/Azure use 169.254.169.254 which the
+        # IP-range check below already covers via the link-local range,
+        # but Google's hostname does not always resolve to a link-local
+        # IP from every VPC, so we hard-deny the string too.
+        "metadata.google.internal",
+    }
+)
+
+# Carrier-grade NAT (RFC 6598). Python's ``ipaddress`` module does NOT
+# classify this range as ``is_private``, so we must check it explicitly.
+_CGNAT_NETWORK_V4: ipaddress.IPv4Network = ipaddress.IPv4Network("100.64.0.0/10")
+
+
+class UnsafeUserUrlError(ValueError):
+    """Raised when a user-supplied URL fails SSRF validation.
+
+    Subclasses :class:`ValueError` so call sites that already treat
+    invalid input as a 400-class error continue to work. The string
+    message names the specific reason (scheme, hostname, resolved IP,
+    DNS failure, ...) so that it can be surfaced to the user verbatim.
+    """
+
+
+def _strip_ipv6_brackets(host: str) -> str:
+    """Return ``host`` with surrounding ``[`` / ``]`` removed if present."""
+
+    if host.startswith("[") and host.endswith("]"):
+        return host[1:-1]
+    return host
+
+
+def _is_blocked_ip(ip: ipaddress.IPv4Address | ipaddress.IPv6Address) -> bool:
+    """Return ``True`` if ``ip`` falls in any range we refuse to dial.
+
+    This is the single source of truth for the IP-range policy:
+
+    * loopback (``127.0.0.0/8``, ``::1``)
+    * private (RFC 1918, ULA ``fc00::/7``)
+    * link-local (``169.254.0.0/16``, ``fe80::/10``)
+    * multicast (``224.0.0.0/4``, ``ff00::/8``)
+    * unspecified (``0.0.0.0``, ``::``)
+    * reserved (``240.0.0.0/4``, etc.)
+    * carrier-grade NAT (``100.64.0.0/10``) — not covered by ``is_private``
+    """
+
+    if (
+        ip.is_loopback
+        or ip.is_private
+        or ip.is_link_local
+        or ip.is_multicast
+        or ip.is_unspecified
+        or ip.is_reserved
+    ):
+        return True
+    if isinstance(ip, ipaddress.IPv4Address) and ip in _CGNAT_NETWORK_V4:
+        return True
+    return False
+
+
+def _resolve(host: str) -> Iterable[ipaddress.IPv4Address | ipaddress.IPv6Address]:
+    """Resolve ``host`` to every A/AAAA record returned by the system.
+
+    Returning *all* addresses (rather than the first one) is critical:
+    a hostile DNS server can return a public IP first followed by a
+    private IP, and the underlying HTTP client may fail over to the
+    private one on connect. We treat the set as unsafe if any element
+    is unsafe.
+    """
+
+    try:
+        results = socket.getaddrinfo(host, None)
+    except socket.gaierror as exc:  # noqa: PERF203 — re-raise as our own type
+        raise UnsafeUserUrlError(f"could not resolve hostname {host!r}: {exc}") from exc
+
+    addresses: list[ipaddress.IPv4Address | ipaddress.IPv6Address] = []
+    for entry in results:
+        sockaddr = entry[4]
+        # IPv4 sockaddr: (host, port). IPv6 sockaddr: (host, port, flowinfo, scope_id).
+        ip_str = sockaddr[0]
+        # Strip IPv6 zone-id ("fe80::1%lo0") before parsing.
+        if "%" in ip_str:
+            ip_str = ip_str.split("%", 1)[0]
+        try:
+            addresses.append(ipaddress.ip_address(ip_str))
+        except ValueError:
+            # An entry we can't parse is itself suspicious; treat as unsafe.
+            raise UnsafeUserUrlError(
+                f"hostname {host!r} resolved to unparseable address {ip_str!r}"
+            ) from None
+    return addresses
+
+
+def _validate_and_pick_ip(
+    url: str,
+) -> tuple[str, ipaddress.IPv4Address | ipaddress.IPv6Address, "urlsplit"]:
+    """Run the SSRF guard and return the data needed to dial safely.
+
+    Performs every check :func:`validate_user_base_url` performs, but
+    additionally returns ``(hostname, ip, parts)`` where ``ip`` is one
+    of the validated addresses (the first record returned by the
+    resolver, or the literal itself if the URL already used an IP) and
+    ``parts`` is the :func:`urllib.parse.urlsplit` result so callers do
+    not have to re-parse the URL.
+
+    Raises :class:`UnsafeUserUrlError` on the same conditions as
+    :func:`validate_user_base_url`.
+    """
+
+    if not isinstance(url, str) or not url.strip():
+        raise UnsafeUserUrlError("url must be a non-empty string")
+
+    try:
+        parts = urlsplit(url)
+    except ValueError as exc:
+        raise UnsafeUserUrlError(f"could not parse url {url!r}: {exc}") from exc
+
+    scheme = parts.scheme.lower()
+    if scheme not in _ALLOWED_SCHEMES:
+        raise UnsafeUserUrlError(
+            f"scheme {scheme!r} is not allowed; only http and https are permitted"
+        )
+
+    # ``urlsplit`` returns the bracketed form for IPv6 in ``netloc`` but
+    # the bare form in ``hostname``. Normalize via lower() because
+    # hostnames are case-insensitive and we compare against a lowercase
+    # blocklist.
+    raw_host = parts.hostname
+    if not raw_host:
+        raise UnsafeUserUrlError(f"url {url!r} has no hostname")
+
+    host = raw_host.lower()
+
+    # Check the literal-string blocklist first. urlsplit().hostname strips
+    # IPv6 brackets, so we also test the bracketed form for completeness
+    # (matches the public-spec note about ``[::]``).
+    bracketed = f"[{host}]"
+    if host in _BLOCKED_HOSTNAMES or bracketed in _BLOCKED_HOSTNAMES:
+        raise UnsafeUserUrlError(
+            f"hostname {raw_host!r} is not allowed (matches internal-only name)"
+        )
+
+    # If the host is already an IP literal (with or without IPv6 brackets),
+    # check it directly without going to DNS — DNS for an IP literal is a
+    # no-op but it's clearer to short-circuit and gives a better message.
+    candidate = _strip_ipv6_brackets(host)
+    try:
+        literal = ipaddress.ip_address(candidate)
+    except ValueError:
+        literal = None
+
+    if literal is not None:
+        if _is_blocked_ip(literal):
+            raise UnsafeUserUrlError(
+                f"hostname {raw_host!r} resolves to blocked address {literal} "
+                f"(loopback/private/link-local/multicast/reserved/CGNAT)"
+            )
+        return host, literal, parts
+
+    # Hostname (not an IP literal) — resolve and validate every record.
+    addresses = list(_resolve(host))
+    for ip in addresses:
+        if _is_blocked_ip(ip):
+            raise UnsafeUserUrlError(
+                f"hostname {raw_host!r} resolves to blocked address {ip} "
+                f"(loopback/private/link-local/multicast/reserved/CGNAT)"
+            )
+    if not addresses:
+        # ``getaddrinfo`` would normally raise instead of returning an
+        # empty list, but treat the degenerate case as unsafe too — we
+        # have nothing to bind a connection to.
+        raise UnsafeUserUrlError(
+            f"hostname {raw_host!r} returned no addresses from DNS"
+        )
+    return host, addresses[0], parts
+
+
+def validate_user_base_url(url: str) -> None:
+    """Validate that ``url`` is safe to use as an outbound base URL.
+
+    Resolve the URL's hostname to one or more IPs and reject if any
+    resolved IP is private/loopback/link-local/multicast/reserved, or if
+    the URL uses a non-http(s) scheme, or if the hostname is one of the
+    known dangerous strings (``localhost``, ``0.0.0.0``, ``[::]``).
+
+    Raises :class:`UnsafeUserUrlError` on rejection. Returns ``None`` on
+    success.
+
+    This function is the create/update-time check. At dispatch time use
+    :func:`pinned_post` instead, which performs the same validation
+    *and* pins the outbound connection to the validated IP so a DNS
+    rebinder cannot flip the resolution between check and connect.
+
+    Args:
+        url: The user-supplied URL to validate. Expected to be an
+            absolute URL with an ``http`` or ``https`` scheme.
+
+    Raises:
+        UnsafeUserUrlError: If the URL fails to parse, uses a forbidden
+            scheme, has an empty/blocklisted hostname, fails DNS
+            resolution, or resolves to any IP in a blocked range.
+    """
+
+    _validate_and_pick_ip(url)
+
+
+class _PinnedHostAdapter(HTTPAdapter):
+    """HTTPS adapter that performs SNI and cert verification against a
+    fixed hostname even when the URL connects to an IP literal.
+
+    Used by :func:`pinned_post` so that resolving the user-supplied
+    hostname once and dialing the resolved IP doesn't break TLS.
+    Without this, ``urllib3`` would default ``server_hostname`` /
+    ``assert_hostname`` to the connect host (the IP) and either send the
+    wrong SNI or fail cert verification — the cert is for the original
+    hostname, not the IP literal.
+    """
+
+    def __init__(self, server_hostname: str, *args: Any, **kwargs: Any) -> None:
+        self._server_hostname = server_hostname
+        super().__init__(*args, **kwargs)
+
+    def init_poolmanager(self, *args: Any, **kwargs: Any) -> None:
+        kwargs["server_hostname"] = self._server_hostname
+        kwargs["assert_hostname"] = self._server_hostname
+        super().init_poolmanager(*args, **kwargs)
+
+
+def _ip_to_url_host(ip: ipaddress.IPv4Address | ipaddress.IPv6Address) -> str:
+    """Return ``ip`` formatted for use in a URL netloc (brackets for v6)."""
+
+    if isinstance(ip, ipaddress.IPv6Address):
+        return f"[{ip}]"
+    return str(ip)
+
+
+def pinned_post(
+    url: str,
+    *,
+    json: Any = None,
+    headers: dict[str, str] | None = None,
+    timeout: float = 5.0,
+    allow_redirects: bool = False,
+) -> requests.Response:
+    """POST to ``url`` with the outbound connection pinned to a single
+    validated IP, closing the DNS-rebinding TOCTOU window left by the
+    naive validate-then-``requests.post`` pattern.
+
+    The URL's hostname is resolved exactly once. Every returned address
+    must pass the same SSRF guard as :func:`validate_user_base_url`. The
+    outbound request is issued against the chosen IP literal (so
+    ``urllib3`` cannot ask the resolver again and receive a different
+    answer); the original hostname is preserved in the ``Host`` header
+    and, for HTTPS, via :class:`_PinnedHostAdapter` for SNI and cert
+    verification.
+
+    Args:
+        url: Absolute http(s) URL to POST to.
+        json: JSON-serializable payload — passed through to ``requests``.
+        headers: Caller-supplied headers. Any caller-supplied ``Host``
+            entry is overwritten so the in-flight request matches what
+            was validated.
+        timeout: Per-request timeout (seconds).
+        allow_redirects: Forwarded to ``requests``. Defaults to
+            ``False`` because the SSRF guard only inspects the supplied
+            URL — following redirects would let a hostile upstream
+            bounce the request to an internal address.
+
+    Raises:
+        UnsafeUserUrlError: If the URL fails the SSRF guard.
+        requests.RequestException: For network-level failures.
+    """
+
+    host, ip, parts = _validate_and_pick_ip(url)
+
+    netloc = _ip_to_url_host(ip)
+    if parts.port is not None:
+        netloc = f"{netloc}:{parts.port}"
+    pinned_url = urlunsplit(
+        (parts.scheme, netloc, parts.path, parts.query, parts.fragment)
+    )
+
+    request_headers = dict(headers or {})
+    host_header = host if parts.port is None else f"{host}:{parts.port}"
+    request_headers["Host"] = host_header
+
+    session = requests.Session()
+    if parts.scheme == "https":
+        session.mount("https://", _PinnedHostAdapter(host))
+    try:
+        return session.post(
+            pinned_url,
+            json=json,
+            headers=request_headers,
+            timeout=timeout,
+            allow_redirects=allow_redirects,
+        )
+    finally:
+        session.close()
+
+
+class _PinnedHTTPSTransport(httpx.HTTPTransport):
+    """``httpx`` transport pinned to a single validated IP literal.
+
+    Closes the DNS-rebinding TOCTOU window that
+    :func:`validate_user_base_url` cannot close on its own. The OpenAI
+    Python SDK (and any other SDK that uses ``httpx``) re-resolves the
+    hostname inside ``socket.getaddrinfo`` at request time, so a
+    hostile DNS server can return a public IP at validation time and a
+    private IP at request time. This transport rewrites every outgoing
+    request's URL host to the validated IP literal so ``httpcore``
+    dials that IP without a fresh lookup.
+
+    The original hostname is preserved in two places:
+
+    1. ``Host`` header — ``httpx.Request._prepare`` set it from the URL
+       netloc *before* this transport runs, so it carries the hostname
+       not the IP literal. We deliberately do not touch headers here.
+    2. TLS SNI / cert verification — set via the
+       ``request.extensions["sni_hostname"]`` extension which
+       ``httpcore`` feeds into ``start_tls``'s ``server_hostname``
+       parameter. Without this, ``urllib3``-equivalent code would use
+       the IP literal as SNI and cert verification would fail (the
+       cert is for the original hostname, not the IP).
+    """
+
+    def __init__(
+        self,
+        validated_host: str,
+        validated_ip: ipaddress.IPv4Address | ipaddress.IPv6Address,
+        **kwargs: Any,
+    ) -> None:
+        # http2=False (the httpx default) — defense in depth against
+        # HTTP/2 connection coalescing (RFC 7540 §9.1.1), where a
+        # client may reuse a TCP connection for any host whose cert
+        # covers it. Per-IP pinning never shares connections across
+        # hosts, but explicit is safer than relying on the default.
+        kwargs.setdefault("http2", False)
+        super().__init__(**kwargs)
+        self._host = validated_host
+        self._ip_netloc = _ip_to_url_host(validated_ip)
+
+    def handle_request(self, request: httpx.Request) -> httpx.Response:
+        # Defense in depth: refuse if the request URL's host doesn't
+        # match what we validated. Catches any future SDK regression
+        # that rewrites the URL between Request construction and dial,
+        # and any rare case where the SDK reuses our pinned client for
+        # a different host (which it shouldn't, but assert it anyway).
+        if request.url.host != self._host:
+            raise UnsafeUserUrlError(
+                f"pinned transport bound to {self._host!r}, refused "
+                f"request for {request.url.host!r}"
+            )
+        # SNI/server_hostname for TLS verification. httpcore reads this
+        # extension at _sync/connection.py and feeds it into
+        # start_tls's server_hostname argument. Set before the URL host
+        # is rewritten so cert validation continues to use the original
+        # hostname even though TCP dials the IP literal.
+        request.extensions = {
+            **request.extensions,
+            "sni_hostname": self._host.encode("ascii"),
+        }
+        request.url = request.url.copy_with(host=self._ip_netloc)
+        return super().handle_request(request)
+
+
+def pinned_httpx_client(
+    base_url: str,
+    *,
+    timeout: float = 600.0,
+) -> httpx.Client:
+    """Return an :class:`httpx.Client` whose connections are pinned to
+    one validated IP, closing the DNS-rebinding TOCTOU window the naive
+    ``OpenAI(base_url=...)`` flow leaves open.
+
+    The hostname in ``base_url`` is resolved exactly once. Every
+    returned address must pass :func:`_validate_and_pick_ip`'s SSRF
+    guard (loopback, RFC 1918, link-local, multicast, reserved, CGNAT,
+    cloud metadata names). The chosen IP becomes the URL host on every
+    outgoing request so ``httpcore`` cannot ask the resolver again.
+
+    Pass via ``OpenAI(http_client=pinned_httpx_client(base_url))`` (or
+    any other SDK that accepts an ``httpx.Client``) to make BYOM
+    dispatch immune to DNS-rebinding TOCTOU.
+
+    Args:
+        base_url: User-supplied http(s) URL. Validated through the same
+            SSRF guard as :func:`validate_user_base_url`.
+        timeout: Per-request timeout (seconds). Defaults to 600 to
+            match the OpenAI SDK's default; callers should override
+            for non-LLM workloads.
+
+    Raises:
+        UnsafeUserUrlError: If ``base_url`` fails the SSRF guard.
+    """
+
+    host, ip, _parts = _validate_and_pick_ip(base_url)
+    transport = _PinnedHTTPSTransport(host, ip)
+    # follow_redirects=False — the SSRF guard only inspects the
+    # supplied URL; following 3xx would let a hostile upstream bounce
+    # the in-network request to an internal address (cloud metadata,
+    # RFC1918, loopback) carrying whatever credentials the SDK adds.
+    return httpx.Client(
+        transport=transport,
+        timeout=timeout,
+        follow_redirects=False,
+    )