Apply SanitizeFunctionName on request and RestoreSanitizedToolName on
response for: gemini/claude, gemini/openai/chat-completions,
gemini/openai/responses, antigravity/openai/chat-completions,
gemini-cli/openai/chat-completions.
Also update SanitizedToolNameMap to handle OpenAI format
(tools[].function.name) in addition to Claude format (tools[].name).
Claude Code and MCP clients may send tool names containing characters
invalid for Gemini's function_declarations (e.g. '/', '@', spaces).
Sanitize on request via SanitizeFunctionName and restore original names
on response for both antigravity/claude and gemini-cli/claude translators.
- Introduced a new method `buildRecord` in `usageReporter` to encapsulate record creation, improving code readability and maintainability.
- Added latency tracking to usage records, ensuring accurate reporting of request latencies.
- Updated tests to validate the inclusion of latency in usage records and ensure proper functionality of the new reporting structure.
When Claude sends redacted thinking with empty text, the translator
was omitting the "text" field from thought parts. Google Antigravity
API requires this field, causing 500 "Unknown Error" responses.
Verified: 129/129 error logs with empty thought → 500, 0/97 success
logs had empty thought. After fix: 0 new "Unknown Error" 500s.
- Log a warning instead of silently ignoring sjson.SetBytes errors in the TranslateRequest fallback path
- Add registry_test.go with tests covering the fallback model normalization and verifying registered transforms take precedence
When no request translator is registered for a format pair (e.g.
openai-response → openai-response), TranslateRequest returned the raw
payload unchanged. This caused client-side model prefixes (e.g.
"copilot/gpt-5-mini") to leak into upstream requests, resulting in
"The requested model is not supported" errors from providers.
The fallback path now updates the "model" field in the payload to
match the resolved model name before returning.
When enabled, the server uses only the embedded models.json loaded at
init() and skips registry.StartModelsUpdater(), disabling the initial
remote fetch and 3-hour periodic refresh. The management panel
auto-updater (managementasset.StartAutoUpdater) is unaffected.
gjson.Parse() marks any string starting with { or [ as gjson.JSON type,
even when the content is not valid JSON (e.g. macOS plist format, truncated
tool results). This caused sjson.SetRaw to embed non-JSON content directly
into the Gemini API request payload, producing 400 errors.
Add json.Valid() check before using SetRaw to ensure only actually valid
JSON is embedded raw. Non-JSON content now falls through to sjson.Set
which properly escapes it as a JSON string.
Fixes#2161
The Claude-to-Gemini translator derived function names by splitting
tool_use_id on "-", which produced empty strings for IDs with exactly
2 segments (e.g. toolu_tool-<uuid>). Replace the string-splitting
heuristic with a lookup map built from tool_use blocks during the
main processing loop, with fallback to the raw ID on miss.
Only emit note in listAuthFilesFromDisk when the JSON value is actually
a string (gjson.String), matching the synthesizer/buildAuthFileEntry
behavior. Non-string values like numbers or booleans are now ignored
instead of being coerced via gjson.String().
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
buildAuthFileEntry now falls back to reading priority/note from
auth.Metadata when Attributes lacks them. This covers auths registered
via UploadAuthFile which bypass the synthesizer and only populate
Metadata from the raw JSON.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Read note from Attributes (consistent with priority) in buildAuthFileEntry,
fixing missing note on Gemini multi-project virtual auth cards.
- Propagate note from primary to virtual auths in SynthesizeGeminiVirtualAuths,
mirroring existing priority propagation.
- Sync note/priority writes to both Metadata and Attributes in PatchAuthFileFields,
with refactored nil-check to reduce duplication (review feedback).
- Validate priority type in fallback disk-read path instead of coercing all values
to 0 via gjson.Int(), aligning with the auth-manager code path.
- Add regression tests for note synthesis, virtual-auth note propagation, and
end-to-end multi-project Gemini note inheritance.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Gemini API rejects uniqueItems in tool schemas with 400. Add it to
unsupportedConstraints alongside minItems/maxItems where it belongs.
Same class of fix as #1424 and #1531.
When assistant has tool_calls but no text content, the translator
emitted an empty message into the Responses API input array before
function_call items. The API then couldn't match function_call_output
to its function_call by call_id, returning:
No tool output found for function call ...
Only emit assistant messages that have content parts. Tool-call-only
messages now produce function_call items directly.
Added 9 tests for tool calling translation covering single/parallel
calls, multi-turn conversations, name shortening, empty content
edge cases, and call_id integrity.
The list endpoint previously omitted priority and note, which are stored
inside each auth file's JSON content. This adds them to both the normal
(auth-manager) and fallback (disk-read) code paths, and extends
PATCH /auth-files/fields to support writing the note field.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
1. Claude translator: add CleanJSONSchemaForGemini() to sanitize tool
input schemas (removes $schema, anyOf, const, format, etc.) and
delete eager_input_streaming from tool declarations. Remove fragile
bytes.Replace for format:"uri" now covered by schema cleaner.
2. Gemini native translator: filter out content entries with empty or
missing parts arrays to prevent Gemini API 400 error "required
oneof field 'data' must have one initialized field".
Both fixes align gemini-cli with protections already present in the
antigravity translator.
when Amp or Claude Code sends functionResponse with an empty name in Gemini
conversation history, the Gemini API rejects the request with 400
"Name cannot be empty". this fix backfills empty names from the
corresponding preceding functionCall parts using positional matching.
covers all three Gemini translator paths:
- gemini/gemini (direct API key)
- antigravity/gemini (OAuth)
- gemini-cli/gemini (Gemini CLI)
also switches fixCLIToolResponse pending group matching from LIFO to
FIFO to correctly handle multiple sequential tool call groups.
fixes#1903
feat(proxy): centralize proxy handling with `proxyutil` package and enhance test coverage
- Added `proxyutil` package to simplify proxy handling across the codebase.
- Refactored various components (`executor`, `cliproxy`, `auth`, etc.) to use `proxyutil` for consistent and reusable proxy logic.
- Introduced support for "direct" proxy mode to explicitly bypass all proxies.
- Updated tests to validate proxy behavior (e.g., `direct`, HTTP/HTTPS, and SOCKS5).
- Enhanced YAML configuration documentation for proxy options.
test(auth-scheduler): add benchmarks and priority-based scheduling improvements
- Added `BenchmarkManagerPickNextMixedPriority500` for mixed-priority performance assessment.
- Updated `pickNextMixed` to prioritize highest ready priority tiers.
- Introduced `highestReadyPriorityLocked` and `pickReadyAtPriorityLocked` for better scheduling logic.
- Added unit test to validate selection of highest priority tiers in mixed provider scenarios.
When new OAuth auth files are added while the service is running,
`applyCoreAuthAddOrUpdate` calls `coreManager.Register()` (which upserts
into the scheduler) BEFORE `registerModelsForAuth()`. At upsert time,
`buildScheduledAuthMeta` snapshots `supportedModelSetForAuth` from the
global model registry — but models haven't been registered yet, so the
set is empty. With an empty `supportedModelSet`, `supportsModel()`
always returns false and the new auth is never added to any model shard.
Additionally, when all existing accounts are in cooldown, the scheduler
returns `modelCooldownError`, but `shouldRetrySchedulerPick` only
handles `*Error` types — so the `syncScheduler` safety-net rebuild
never triggers and the new accounts remain invisible.
Fix:
1. Add `RefreshSchedulerEntry()` to re-upsert a single auth after its
models are registered, rebuilding `supportedModelSet` from the
now-populated registry.
2. Call it from `applyCoreAuthAddOrUpdate` after `registerModelsForAuth`.
3. Make `shouldRetrySchedulerPick` also match `*modelCooldownError` so
the full scheduler rebuild triggers when all credentials are cooling
down — catching any similar stale-snapshot edge cases.
The Gemini v1internal (cloudcode-pa) and Antigravity Manager endpoints
require camelCase "systemInstruction" in request JSON. The current
snake_case "system_instruction" causes system prompts to be silently
ignored when routing through these endpoints.
Replace all "system_instruction" JSON keys with "systemInstruction" in
chat-completions and responses request translators.
Keep the OpenAI Responses computer tool intact when normalizing requests for the GitHub Copilot executor.
This change preserves built-in computer tool definitions instead of dropping them as non-function tools, keeps explicit computer tool_choice selections unchanged, and classifies computer_call / computer_call_output items as assistant and tool turns when deriving the initiator header.
Together these adjustments allow Responses requests that use the computer tool to reach the upstream executor without losing tool metadata or switching turn ownership unexpectedly.
- Added comprehensive unit tests for `authScheduler` and related components.
- Implemented `authScheduler` with support for Round Robin, Fill First, and custom selector strategies.
- Improved tracking of auth states, cooldowns, and recovery logic in scheduler.
test(websocket): add tests for incremental input and prewarm handling logic
- Added test cases for incremental input support based on upstream capabilities.
- Introduced validation for prewarm handling of `response.create` messages locally.
- Enhanced test coverage for websocket executor behavior, including payload forwarding checks.
- Updated websocket implementation with prewarm and incremental input logic for better testability.
test(translator): add unit tests for OpenAI to Claude requests and tool result handling
- Introduced tests for converting OpenAI requests to Claude with text, base64 images, and URL images in tool results.
- Refactored `convertClaudeToolResultContent` and related functionality to properly handle raw content with images and text.
- Updated conversion logic to streamline image handling for both base64 and URL formats.
- Simplified connection logic by removing `connCreateSent` and related state handling.
- Updated `buildCodexWebsocketRequestBody` to always use `response.create`.
- Added unit tests to validate `response.create` behavior and beta header preservation.
- Dropped unsupported `response.append` and outdated `response.done` event types.
- Add model to GetGeminiModels()
- Add model to GetGeminiVertexModels()
- Add model to GetGeminiCLIModels()
- Add model to GetAIStudioModels()
- Add to AntigravityModelConfig with thinking levels
- Update gemini-3-flash-preview description
Registers the new lightweight Gemini model across all provider
endpoints for cost-effective high-volume usage scenarios.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
SetOAuthSessionError previously sent generic messages to the management
panel (e.g. "Failed to complete Gemini CLI onboarding"), hiding the
actual error returned by Google APIs. The specific error was only
written to the server log via log.Errorf, which is often inaccessible
in headless/Docker deployments.
Include the upstream error in all 8 OAuth error paths so the
management panel shows actionable messages like "no Google Cloud
projects available for this account" instead of a generic failure.
normalizeCacheControlTTL unconditionally re-serializes the entire request
body through json.Unmarshal/json.Marshal even when no TTL normalization
is needed. Go's json.Marshal randomizes map key order and HTML-escapes
<, >, & characters (to \u003c, \u003e, \u0026), producing different raw
bytes on every call.
Anthropic's prompt caching uses byte-prefix matching, so any byte-level
difference causes a cache miss. This means the ~119K system prompt and
tools are re-processed on every request when routed through CPA.
The fix adds a bool return to normalizeTTLForBlock to indicate whether
it actually modified anything, and skips the marshal step in
normalizeCacheControlTTL when no blocks were changed.
Add util.SanitizeClaudeToolID() to replace non-conforming characters in
tool_use.id fields across all five response translators (gemini, codex,
openai, antigravity, gemini-cli).
Upstream tool names may contain dots or other special characters
(e.g. "fs.readFile") that violate Claude's ID validation regex.
The sanitizer replaces such characters with underscores and provides
a generated fallback for empty IDs.
Fixes#1872, Fixes#1849
Made-with: Cursor
- Add `TestTriggerServerUpdateCancelsPendingTimerOnImmediate` to verify proper handling of server update debounce and timer cancellation.
- Fix logic in `triggerServerUpdate` to prevent duplicate timers and ensure proper cleanup of pending state.
The instructions restore logic was originally needed when the proxy
injected custom instructions (per-model system prompts) into requests.
Since ac802a46 removed the injection system, the proxy no longer
modifies instructions before forwarding. The upstream response's
instructions field now matches the client's original value, making
the restore a no-op.
Also removes unused sjson import.
Closesrouter-for-me/CLIProxyAPI#1868
Tools like TaskList, TaskGet have no required parameters, so empty input
is valid. Previously, the truncation detector flagged all empty inputs as
truncated, causing these tools to be skipped and breaking the tool loop.
Now only flag empty input as truncation when the tool has required fields
defined in RequiredFieldsByTool.
fix(gemini): add `deprecated` to unsupported schema keywords
Add `deprecated` to the list of unsupported schema metadata fields in Gemini and update tests to verify its removal.
- Set Accept-Encoding: identity for SSE streams; upstream must not compress
line-delimited SSE bodies that bufio.Scanner reads directly
- Re-enforce identity after ApplyCustomHeadersFromAttrs to prevent auth
attribute injection from re-enabling compression on the stream path
- Add peekableBody type wrapping bufio.Reader for non-consuming magic-byte
inspection of the first 4 bytes without affecting downstream readers
- Detect gzip (0x1f 0x8b) and zstd (0x28 0xb5 0x2f 0xfd) by magic bytes
when Content-Encoding header is absent, covering misbehaving upstreams
- Remove if-Content-Encoding guard on all three error paths (Execute,
ExecuteStream, CountTokens); unconditionally delegate to decodeResponseBody
so magic-byte detection applies consistently to all response paths
- Add 10 tests covering stream identity enforcement, compressed success bodies,
magic-byte detection without headers, error path decoding, and
auth attribute override prevention
- Fix SSRF: validate API endpoint host against allowlist before use
- Limit /models response body to 2MB to prevent memory exhaustion (DoS)
- Use MakeAuthenticatedRequest for consistent headers across API calls
- Trim trailing slash on API endpoint to prevent double-slash URLs
- Use ListModelsWithGitHubToken to simplify token exchange + listing
- Deduplicate model IDs to prevent incorrect registry reference counting
- Remove dead capabilities enrichment code block
- Remove unused ModelExtra field with misleading json:"-" tag
- Extract magic numbers to named constants (defaultCopilotContextLength)
- Remove redundant hyphenID == id check (already filtered by Contains)
- Use defer cancel() for context timeout in service.go
- Add ListModels/ListModelsWithGitHubToken to CopilotAuth for querying
the /models endpoint at api.githubcopilot.com
- Add FetchGitHubCopilotModels in executor with static fallback on failure
- Update service.go to use dynamic fetching (15s timeout) instead of
hardcoded GetGitHubCopilotModels()
- Add GitHubCopilotAliasesFromModels for auto-generating dot-to-hyphen
model aliases from dynamic model lists
Claude's Tool Search feature (advanced-tool-use-2025-11-20 beta) adds
defer_loading field to tool definitions. When proxying Claude requests
to Codex or Gemini, this unknown field causes 400 errors upstream.
Strip defer_loading (and cache_control where missing) in all three
Claude-to-upstream translation paths:
- codex/claude: defer_loading + cache_control
- gemini-cli/claude: defer_loading
- gemini/claude: defer_loading
Fixes#1725, Fixes#1375
Add support for Claude's "adaptive" and "auto" thinking modes using `output_config.effort`. Introduce support for new effort level "max" in adaptive thinking. Update thinking logic, validate model capabilities, and extend converters and handling to ensure compatibility with adaptive modes. Adjust static model data with supported levels and refine handling across translators and executors.
## Problem
When using Antigravity Claude models through CLIProxyAPI, the thinking
chain (reasoning content) does not display in the Amp client.
## Root Cause
The Amp client sends `thinking: {"type": "auto"}` in its requests,
but `ConvertClaudeRequestToAntigravity` only handled `"enabled"` and
`"adaptive"` types in its switch statement. The `"auto"` type was
silently ignored, resulting in no `thinkingConfig` being set in the
translated Gemini request. Without `thinkingConfig`, the Antigravity
API returns responses without any thinking content.
Additionally, the Antigravity API for Claude models does not support
`thinkingBudget: -1` (auto mode sentinel). It requires a concrete
positive budget value. The fix uses 128000 as the budget for "auto"
mode, which `ApplyThinking` will then normalize to stay within the
model's actual limits (e.g., capped to `maxOutputTokens - 1`).
## Changes
### internal/translator/antigravity/claude/antigravity_claude_request.go
1. **Add "auto" case** to the thinking type switch statement.
Sets `thinkingBudget: 128000` and `includeThoughts: true`.
The budget is subsequently normalized by `ApplyThinking` based
on model-specific limits.
2. **Add "auto" to hasThinking check** so that interleaved thinking
hints are injected for tool-use scenarios when Amp sends
`thinking.type="auto"`.
### internal/registry/model_definitions_static_data.go
3. **Add Thinking configuration** for `claude-sonnet-4-6`,
`claude-sonnet-4-5`, and `claude-opus-4-6` in
`GetAntigravityModelConfig()` -- these were previously missing,
causing `ApplyThinking` to skip thinking config entirely.
## Testing
- Deployed to Railway test instance (cpa-thinking-test)
- Verified via debug logging that:
- Amp sends `thinking: {"type": "auto"}`
- CPA now translates this to `thinkingConfig: {thinkingBudget: 128000, includeThoughts: true}`
- `ApplyThinking` normalizes the budget to model-specific limits
- Antigravity API receives the correct thinkingConfig
Amp-Thread-ID: https://ampcode.com/threads/T-019ca511-710d-776d-a07c-4b750f871a93
Co-authored-by: Amp <amp@ampcode.com>
Captured and compared outgoing requests from CLIProxyAPI against real
Claude Code 2.1.63 and fixed all detectable differences:
Headers:
- Update anthropic-beta to match 2.1.63: replace fine-grained-tool-streaming
and prompt-caching-2024-07-31 with context-management-2025-06-27 and
prompt-caching-scope-2026-01-05
- Remove X-Stainless-Helper-Method header (real Claude Code does not send it)
- Update default User-Agent from "claude-cli/2.1.44 (external, sdk-cli)" to
"claude-cli/2.1.63 (external, cli)"
- Force Claude Code User-Agent for non-Claude clients to avoid leaking
real client identity (e.g. curl, OpenAI SDKs) during cloaking
Body:
- Inject x-anthropic-billing-header as system[0] (matches real format)
- Change system prompt identifier from "You are Claude Code..." to
"You are a Claude agent, built on Anthropic's Claude Agent SDK."
- Add cache_control with ttl:"1h" to match real request format
- Fix user_id format: user_[64hex]_account_[uuid]_session_[uuid]
(was missing account UUID)
- Disable tool name prefix (set claudeToolPrefix to empty string)
TLS:
- Switch utls fingerprint from HelloFirefox_Auto to HelloChrome_Auto
(closer to Node.js/OpenSSL used by real Claude Code)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add IAM Identity Center (IDC) authentication with CLI flags (--kiro-idc-login, --kiro-idc-start-url, --kiro-idc-region) and login flow
- Add ProfileArn auto-fetching in Execute/ExecuteStream for imported IDC accounts
- Simplify endpoint preference with map-based alias lookup and getAuthValue helper
- Redesign fingerprint as global singleton with external config and per-account deterministic generation
- Add StartURL and FingerprintConfig fields to Kiro config
- Add AgentContinuationID/AgentTaskType support in Kiro translators
- Add comprehensive tests for executor, fingerprint, SSO OIDC, and AWS helpers
- Add CLI login documentation to README
- Add 'github-user' fallback in WaitForAuthorization when FetchUserInfo
returns empty Login (fixes malformed 'github-copilot-.json' filenames)
- Standardize Web API file name to 'github-copilot-<user>.json' to match
CLI path convention (was 'github-<user>.json')
Addresses Gemini Code Assist review comments on PR #291.
- Expand OAuth scope to include read:user for full profile access
- Add GitHubUserInfo struct with Login, Email, Name fields
- Update FetchUserInfo to return complete user profile
- Add Email and Name fields to CopilotTokenStorage and CopilotAuthBundle
- Fix provider string bug: 'github' -> 'github-copilot' in auth_files.go
- Fix semantic bug: email field was storing username
- Update Label to prefer email over username in both CLI and Web API paths
- Add 9 unit tests covering new functionality
- Change RequiredFieldsByTool value type from []string to [][]string
- Outer slice = AND (all groups required); inner slice = OR (any one satisfies)
- Fix Bash entry to accept "cmd" or "command", resolving soft-truncation loop
- Update findMissingRequiredFields logic and inline docs accordingly
Guard the openai-response streaming path against truncated/invalid SSE data payloads by validating data: JSON before forwarding; surface a 502 terminal error instead of letting clients crash with JSON parse errors.
Move base64 image data from Claude tool_result into functionResponse.parts
as inlineData instead of outer sibling parts, preventing context bloat.
Unify all inlineData field naming to camelCase mimeType across Claude,
OpenAI, and Gemini translators. Add comprehensive edge case tests and
Gemini-side regression test for functionResponse.parts preservation.
CI blocks PRs that modify internal/translator. Revert translator edits and keep only the /v1/responses streaming error-chunk fix; file an issue for translator conformance work.
Ensure response.created and response.completed chunks produced by the OpenAI/Gemini/Claude translators always include required fields (response.model and response.usage) so clients validating Responses SSE do not fail schema validation.
When /v1/responses streaming fails after headers are sent, we now emit a type=error chunk instead of an HTTP-style {error:{...}} payload, preventing AI SDK chunk validation errors.
- Add 60 requests/minute rate limiting per credential using sliding window
- Detect insufficient_quota errors and set cooldown until next day (Beijing time)
- Map quota errors (HTTP 403/429) to 429 with retryAfter for conductor integration
- Cache Beijing timezone at package level to avoid repeated syscalls
- Add redactAuthID function to protect credentials in logs
- Extract wrapQwenError helper to consolidate error handling
- force http/1.1 instead of http/2
- explicit connection close
- strip proxy headers X-Forwarded-For and X-Real-IP
- add project id to fetch models payload
Changes the RoundRobinSelector to use two-level round-robin when
gemini-cli virtual auths are detected (via gemini_virtual_parent attr):
- Level 1: cycle across credential groups (parent accounts)
- Level 2: cycle within each group's project auths
Credentials start from a random offset (rand.IntN) for fair distribution.
Non-virtual auths and single-credential scenarios fall back to flat RR.
Adds 3 test cases covering multi-credential grouping, single-parent
fallback, and mixed virtual/non-virtual fallback.
Default to generating a fresh random user_id per request instead of
reusing cached IDs. Add cache-user-id config option to opt in to the
previous caching behavior.
- Add CacheUserID field to CloakConfig
- Extract user_id cache logic to dedicated file
- Generate fresh user_id by default, cache only when enabled
- Add tests for both paths
console.anthropic.com is now protected by a Cloudflare managed challenge
that blocks all non-browser POST requests to /v1/oauth/token, causing
`-claude-login` to fail with a 403 error.
Switch to api.anthropic.com which hosts the same OAuth token endpoint
without the Cloudflare managed challenge.
Fixes#1659
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Non-spark codex models (gpt-5.3-codex, gpt-5.2-codex) stream function call
arguments via multiple delta events followed by a done event. The done handler
unconditionally emitted the full arguments, duplicating what deltas already
streamed. This produced invalid double JSON that Claude Code couldn't parse,
causing tool calls to fail with missing parameters and infinite retry loops.
Add HasReceivedArgumentsDelta flag to track whether delta events were received.
The done handler now only emits arguments when no deltas preceded it (spark
models), while delta-based streaming continues to work for non-spark models.
Some Codex models (e.g. gpt-5.3-codex-spark) send function call arguments
in a single "done" event without preceding "delta" events. The streaming
translator only handled "delta" events, causing tool call arguments to be
lost — resulting in empty tool inputs and infinite retry loops in clients
like Claude Code.
Emit the full arguments from the "done" event as a single input_json_delta
so downstream clients receive the complete tool input.
- Updated `ExecuteStream` functions in executors to use `StreamResult` instead of channels.
- Enhanced upstream header handling in OpenAI handlers.
- Improved maintainability and alignment across executors and handlers.
- Introduced `passthrough-headers` option in configuration to control forwarding of upstream response headers.
- Updated handlers to respect the passthrough headers setting.
- Added tests to verify behavior when passthrough is enabled or disabled.
sjson.SetRaw with an empty string produces malformed JSON (e.g. "result":}).
This happens when a Claude tool_result block has no content field, causing
functionResponseResult.Raw to be "". Guard against this by falling back to
sjson.Set with an empty string only when .Raw is empty.
- Implemented `--standalone` mode to launch an embedded server for TUI.
- Enhanced TUI client to support API-based log polling when log hooks are unavailable.
- Added authentication gate for password input and connection handling.
- Improved localization and UX for logs, authentication, and status bar rendering.
- Implemented `--standalone` mode to launch an embedded server for TUI.
- Enhanced TUI client to support API-based log polling when log hooks are unavailable.
- Added authentication gate for password input and connection handling.
- Improved localization and UX for logs, authentication, and status bar rendering.
- Introduced unit tests for request logging middleware to enhance coverage.
- Added WebSocket-based Codex executor to support Responses API upgrade.
- Updated middleware logic to selectively capture request bodies for memory efficiency.
- Enhanced Codex configuration handling with new WebSocket attributes.
The OpenAI Chat Completions translator was silently dropping
response.function_call_arguments.delta and
response.function_call_arguments.done Codex SSE events, meaning
tool call arguments were never streamed incrementally to clients.
Add proper handling mirroring the proven Claude translator pattern:
- response.output_item.added: announce tool call (id, name, empty args)
- response.function_call_arguments.delta: stream argument chunks
- response.function_call_arguments.done: emit full args if no deltas
- response.output_item.done: defensive fallback for backward compat
State tracking via HasReceivedArgumentsDelta and HasToolCallAnnounced
ensures no duplicate argument emission and correct behavior for models
like codex-spark that skip delta events entirely.
Update hardcoded X-Stainless-* and User-Agent defaults to match
Claude Code 2.1.44 / @anthropic-ai/sdk 0.74.0 (verified via
diagnostic proxy capture 2026-02-17).
Changes:
- X-Stainless-Os/Arch: dynamic via runtime.GOOS/GOARCH
- X-Stainless-Package-Version: 0.55.1 → 0.74.0
- X-Stainless-Timeout: 60 → 600
- User-Agent: claude-cli/1.0.83 (external, cli) → claude-cli/2.1.44 (external, sdk-cli)
Add claude-header-defaults config section so values can be updated
without recompilation when Claude Code releases new versions.
CPA-internal thinking levels like 'xhigh' and 'minimal' are not accepted
by OpenAI-format providers (MiniMax, etc.). The OpenAI applier now maps
non-standard levels to the nearest valid reasoning_effort value before
writing to the request body:
xhigh → high
minimal → low
auto → medium
Allow disabling the proxy_ tool name prefix on a per-account basis.
Users who route their own Anthropic account through CPA can set
"tool_prefix_disabled": true in their OAuth auth JSON to send tool
names unchanged to Anthropic.
Default behavior is fully preserved — prefix is applied unless
explicitly disabled.
Changes:
- Add ToolPrefixDisabled() accessor to Auth (reads metadata key
"tool_prefix_disabled" or "tool-prefix-disabled")
- Gate all 6 prefix apply/strip points with the new flag
- Add unit tests for the accessor
The proxy_ prefix logic correctly skips built-in tools (those with a
non-empty "type" field) in tools[] definitions but does not skip them
in messages[].content[] tool_use blocks or tool_choice. This causes
web_search in conversation history to become proxy_web_search, which
Anthropic does not recognize.
Fix: collect built-in tool names from tools[] into a set and also
maintain a hardcoded fallback set (web_search, code_execution,
text_editor, computer) for cases where the built-in tool appears in
history but not in the current request's tools[] array. Skip prefixing
in messages and tool_choice when name matches a built-in.
- Fix X-Initiator detection: check for any assistant/tool role
in messages instead of only the last message role, matching
the correct agent detection for multi-turn tool conversations
- Add x-github-api-version: 2025-04-01 header for API compatibility
- Support Business/Enterprise accounts by using Endpoints.API from
the Copilot token response instead of hardcoded base URL
- Fix Responses API vision detection: detect vision content before
input normalization removes the messages array
- Add 8 test cases covering the above fixes
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Change Openai-Intent header from "conversation-edits" to
"conversation-panel" to avoid triggering GitHub's premium
execution path, which caused included models (0x multiplier)
to be billed as premium requests.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Collect built-in tool names (those with a "type" field like
web_search, code_execution) and skip prefixing tool_reference
blocks that reference them, preventing name mismatch.
- Refactor if-else if chains to switch statements in all three
prefix functions for idiomatic Go style.
tool_reference blocks can appear nested inside tool_result.content[]
arrays, not just at the top level of messages[].content[]. The prefix
logic now iterates into tool_result blocks with array content to find
and prefix/strip nested tool_reference.tool_name fields.
applyClaudeToolPrefix, stripClaudeToolPrefixFromResponse, and
stripClaudeToolPrefixFromStreamLine now handle "tool_reference" blocks
(field "tool_name") in addition to "tool_use" blocks (field "name").
Without this fix, tool_reference blocks in conversation history retain
their original unprefixed names while tool definitions carry the proxy_
prefix, causing Anthropic API 400 errors: "Tool reference 'X' not found
in available tools."
Co-authored-by: Kirill Turanskiy <kt@novamedia.ru>
- Add response.function_call_arguments.delta handler for tool call parameters
- Rewrite normalizeGitHubCopilotResponsesInput to produce structured input
array (message/function_call/function_call_output) instead of flattened
text, fixing infinite loop in multi-turn tool-use conversations
- Skip flattenAssistantContent for messages containing tool_use blocks,
preventing function_call items from being destroyed
- Add reasoning/thinking stream & non-stream support
- Fix stop_reason mapping (max_tokens/stop) and cached token reporting
- Update test to match new array-based input format
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The normalizeGitHubCopilotResponsesTools filter required type="function",
which dropped Claude-format tools (no type field, uses input_schema).
Relax the filter to accept tools without a type field and map input_schema
to parameters so tools are correctly sent to the upstream API.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The delete handler now sets the channel value to nil instead of removing
the map key, and the sanitization loop preserves nil/empty channel entries
as 'disabled' markers. This prevents SanitizeOAuthModelAlias from
re-injecting default kiro aliases after a user explicitly deletes them
through the management API.
Change GenerateSearchIndicatorEvents to return [][]byte instead of []sseEvent
for consistency with BuildFallbackTextEvents and other event building functions.
Benefits:
- Consistent API across all event generation functions
- Eliminates intermediate sseEvent type conversion in caller
- Simplifies usage by returning ready-to-send SSE byte slices
This addresses the code quality feedback from PR #226 review.
Align Kiro executor with all other executors (Claude, Gemini, OpenAI,
etc.) by using payloadRequestedModel(opts, req.Model) instead of
req.Model when constructing response model names.
This ensures model aliases are correctly reflected in responses:
- Execute: BuildClaudeResponse + TranslateNonStream
- ExecuteStream: streamToChannel
- handleWebSearchStream: BuildClaudeMessageStartEvent
- handleWebSearch: via executeNonStreamFallback (automatic)
Previously Kiro was the only executor using req.Model directly,
which exposed internal routed names instead of the user's alias.
Consolidate web search handler, SSE event generation, stream analysis,
and MCP HTTP I/O into the executor layer. Merge the separate
kiro_websearch_handler.go back into kiro_executor.go to align with
the single-file-per-executor convention. Translator retains only pure
data types, detection, and payload transformation.
Key changes:
- Move SSE construction (search indicators, fallback text, message_start)
from translator to executor, consistent with streamToChannel pattern
- Move MCP handler (callMcpAPI, setMcpHeaders, fetchToolDescription)
from translator to executor alongside other HTTP I/O
- Reuse applyDynamicFingerprint for MCP UA headers (eliminate duplication)
- Centralize MCP endpoint URL via BuildMcpEndpoint in translator
- Add atomic Set/GetWebSearchDescription for cross-layer tool desc cache
- Thread context.Context through MCP HTTP calls for cancellation support
- Thread usage reporter through all web search API call paths
- Add token expiry pre-check before MCP/GAR calls
- Clean up dead code (GenerateMessageID, webSearchAuthContext fp logic,
ContainsWebSearchTool, StripWebSearchTool)
Add log.Warnf calls to all 7 silent return nil paths so operators can
diagnose why specific antigravity accounts fail to fetch models and get
unregistered without any log trail.
Covers: token errors, request creation failures, context cancellation,
network errors (after exhausting fallback URLs), body read errors,
unexpected HTTP status codes, and missing models field in response.
Kiro/AmazonQ API requires the conversation history to start with a user message.
Some clients (e.g., OpenClaw) send conversations starting with an assistant message,
which is valid for the native Claude API but causes 'Improperly formed request' (400)
on the Kiro endpoint.
This fix detects when the first message has role=assistant and prepends a minimal
placeholder user message ('.') to satisfy the Kiro API's message ordering requirement.
Upstream error: {"message":"Improperly formed request.","reason":null}
Verified: original request returns 400, fixed request returns 200.
Adds explicit handling for context.Canceled errors in the reverse proxy error handler to return 499 status code without logging, which is more appropriate for client-side cancellations during polling.
Also adds a test case to verify this behavior.
Adjacent assistant messages where any message contains tool_calls
were being merged by MergeAdjacentMessages, causing tool_calls to
be silently dropped. This led to orphaned tool results that could
not match any toolUse in history, resulting in Kiro API returning
'Improperly formed request.'
Now assistant messages with tool_calls are kept separate during
merge, preserving the tool call chain integrity.
- Added unit tests for combining OAuth excluded models across global and attribute-specific scopes.
- Implemented priority attribute parsing with support for different formats and trimming.
Add Google One personal account login to Gemini CLI OAuth flow:
- CLI --login shows mode menu (Code Assist vs Google One)
- Web management API accepts project_id=GOOGLE_ONE sentinel
- Auto-discover project via onboardUser without cloudaicompanionProject when project is unresolved
Improve robustness of auto-discovery and token handling:
- Add context-aware auto-discovery polling (30s timeout, 2s interval)
- Distinguish network errors from project-selection-required errors
- Refresh expired access tokens in readAuthFile before project lookup
- Extend project_id auto-fill to gemini auth type (was antigravity-only)
Unify credential file naming to geminicli- prefix for both CLI and web.
Add extractAccessToken unit tests (9 cases).
Sanitize tool schemas by stripping prefill, enumTitles, $id, and patternProperties to prevent Gemini INVALID_ARGUMENT 400 errors, and add unit and executor-level tests to lock in the behavior.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Added support for single and array-based `content` cases.
- Enhanced `system_instruction` structure population logic.
- Improved handling of user role assignment for string-based `content`.
GitHub Copilot API rejects model names with suffixes (e.g. claude-opus-4.6(medium)).
The OAuthModelAlias resolution correctly maps aliases like 'opus(medium)' to
'claude-opus-4.6(medium)' preserving the suffix, but the executor must strip the
suffix before sending to the upstream API since Copilot only accepts bare model names.
Update normalizeModel in github_copilot_executor to strip suffixes using
thinking.ParseSuffix, matching the pattern used by other executors.
Also add test coverage for:
- OAuthModelAliasChannel github-copilot and kiro channel resolution
- Suffix preservation in alias resolution for github-copilot
- normalizeModel suffix stripping in github_copilot_executor
Kiro models are exposed with kiro- prefix (e.g., kiro-claude-sonnet-4-5),
which prevents clients like Claude Code from using standard model names
(e.g., claude-sonnet-4-20250514).
This change injects default oauth-model-alias entries for the kiro channel
when no user-configured aliases exist, following the same pattern as the
existing Antigravity defaults. The aliases map standard Claude model names
(both with and without date suffixes) to their kiro-prefixed counterparts.
Default aliases added:
- claude-sonnet-4-5-20250929 / claude-sonnet-4-5 → kiro-claude-sonnet-4-5
- claude-sonnet-4-20250514 / claude-sonnet-4 → kiro-claude-sonnet-4
- claude-opus-4-6 → kiro-claude-opus-4-6
- claude-opus-4-5-20251101 / claude-opus-4-5 → kiro-claude-opus-4-5
- claude-haiku-4-5-20251001 / claude-haiku-4-5 → kiro-claude-haiku-4-5
All aliases use fork: true to preserve the original kiro-* names.
User-configured kiro aliases are respected and not overridden.
Closesrouter-for-me/CLIProxyAPIPlus#208
Silences logging for client cancellations during polling to reduce noise in logs.
Client-side cancellations are common during long-running operations and should not be treated as errors.
The `ConvertOpenAIResponsesRequestToGemini` function had code that attempted
to uppercase JSON Schema type values (e.g. "string" -> "STRING") for Gemini
compatibility. This broke nullable types because when `type` is a JSON array
like `["string", "null"]`:
1. `gjson.Result.String()` returns the raw JSON text `["string","null"]`
2. `strings.ToUpper()` produces `["STRING","NULL"]`
3. `sjson.Set()` stores it as a JSON **string** `"[\"STRING\",\"NULL\"]"`
instead of a JSON array
4. The downstream `CleanJSONSchemaForGemini()` / `flattenTypeArrays()`
cannot detect it (since `IsArray()` returns false on a string)
5. Gemini/Antigravity API rejects it with:
`400 Invalid value at '...type' (Type), "["STRING","NULL"]"`
This was confirmed and tested with Droid Factory (Antigravity) Gemini models
where Claude Code sends tool schemas with nullable parameters.
The fix removes the uppercasing logic entirely and passes the raw schema
through to `parametersJsonSchema`. This is safe because:
- Antigravity executor already runs `CleanJSONSchemaForGemini()` which
properly handles type arrays, nullable fields, and all schema cleanup
- Gemini/Vertex executors use `parametersJsonSchema` which accepts raw
JSON Schema directly (no uppercasing needed)
- The uppercasing code also only iterated top-level properties, missing
nested schemas entirely
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The ResponseRewriter's modelFieldPaths was missing 'response.model',
causing the mapped model name to leak through SSE streaming events
(response.created, response.in_progress, response.completed) in the
OpenAI Responses API (/v1/responses).
This caused Amp CLI to report 'Unknown OpenAI model' errors when
model mapping was active (e.g., gpt-5.2-codex -> gpt-5.3-codex),
because the mapped name reached Amp's backend via telemetry.
Also sorted modelFieldPaths alphabetically per review feedback
and added regression tests for all rewrite paths.
Fixes#1463
fix(translator): restructure message content handling to support multiple content types
- Consolidated `input_text` and `output_text` handling into a single case.
- Added support for processing `input_image` content with associated URLs.
> Copilot Premium usage significantly amplified when using amp
- Add X-Initiator header (user/agent) based on last message role to
prevent Copilot from billing all requests as premium user-initiated
- Add flattenAssistantContent() to convert assistant content from array
to string, preventing Claude from re-answering all previous prompts
- Align Copilot headers (User-Agent, Editor-Version, Openai-Intent) with
pi-ai reference implementation
Closes#113
Amp-Thread-ID: https://ampcode.com/threads/T-019c392b-736e-7489-a06b-f94f7c75f7c0
Co-authored-by: Amp <amp@ampcode.com>
- Updated `KimiAPIBaseURL` to remove versioning from the root path.
- Integrated `ClaudeExecutor` fallback in `KimiExecutor` methods for compatibility with Claude requests.
- Simplified token counting by delegating to `ClaudeExecutor`.
- Standardized the handling of `stop_reason` and `finish_reason` across Codex and Gemini responses.
- Restricted pass-through of specific reasons (`max_tokens`, `stop`) for consistency.
- Enhanced fallback logic for undefined reasons.
Kiro API requires non-empty content on assistant messages, so
CLIProxyAPI injects placeholder text when assistant messages only
contain tool_use blocks (no text). The previous placeholders were
conversational phrases:
- DefaultAssistantContentWithTools: "I'll help you with that."
- DefaultAssistantContent: "I understand."
In agentic sessions with many tool calls, these phrases appeared
dozens of times in conversation history. Opus 4.6 (and likely other
models) picked up on this pattern and started parroting "I'll help
you with that." before every tool call in its actual responses.
Fix: Replace both placeholders with a single dot ".", which
satisfies Kiro's non-empty requirement without giving the model
a phrase to mimic.
- Added support to extract and include `cachedContentTokenCount` in `usage.prompt_tokens_details`.
- Logged warnings for failures to set cached token count for better debugging.
- Introduced `RequestKimiToken` API for Kimi authentication flow.
- Integrated device ID management throughout Kimi-related components.
- Enhanced header management for Kimi API requests with device ID context.
- Add contextUsageEvent case handler in kiro_executor.go for both
parseEventStream and streamToChannel functions
- Handle nested format: {"contextUsageEvent": {"contextUsagePercentage": 0.53}}
- Keep KiroModel struct minimal with only essential fields
- Remove unused KiroPromptCachingInfo struct from kiro_model_converter.go
- Remove unused SupportedInputTypes and PromptCaching fields from KiroAPIModel
Problem:
- PR #186 fixed empty content for assistant messages and history user messages
- But current user message (isLastMessage == true) was not fixed
- When user message contains only tool_result (no text), content becomes empty
- This causes 'Improperly formed request' errors from Kiro API
- Compaction requests from OpenCode commonly have this pattern
Solution:
- Move empty content check BEFORE the isLastMessage branch
- Apply fallback content to ALL user messages, not just history
- Add DefaultUserContentWithToolResults and DefaultUserContent constants
Fixes compaction failures for OpenCode + Quotio + CLIProxyAPIPlus + Kiro stack
- OAuth2 device authorization grant flow (RFC 8628) for authentication
- Streaming and non-streaming chat completions via OpenAI-compatible API
- Models: kimi-k2, kimi-k2-thinking, kimi-k2.5
- CLI `--kimi-login` command for device flow auth
- Token management with automatic refresh
- Thinking/reasoning effort support for thinking-enabled models
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add kiro-claude-opus-4-6 and kiro-claude-opus-4-6-agentic to model registry
- Add model ID mappings for claude-opus-4.6 variants
- Support both kiro- prefix and native format (claude-opus-4.6)
- Tested and working with Kiro API
- Replaced all instances of `bytes.Clone` with direct references to enhance efficiency.
- Simplified payload handling across executors and translators by eliminating unnecessary data duplication.
- Consolidated path-finding logic into a new `findPathsByFields` helper function.
- Refactored repetitive loop structures to improve readability and performance.
- Added depth-based sorting for deletion paths to ensure proper removal order.
Apply code review feedback from gemini-code-assist:
- Move fallback strings to kirocommon package as exported constants
- Update kiro_claude_request.go to use shared constants
- Update kiro_openai_request.go to use shared constants
- Improves maintainability and avoids duplication
Problem:
- PR #181 fixed empty content for OpenAI format (kiro_openai_request.go)
- But Claude format (kiro_claude_request.go) was not fixed
- OpenCode uses Claude format (/v1/messages endpoint)
- When assistant messages have only tool_use (no text), content becomes empty
- This causes 'Improperly formed request' errors from Kiro API
Example of problematic message format:
{
"role": "assistant",
"content": [
{"type": "tool_use", "id": "...", "name": "todowrite", "input": {...}}
]
}
Solution:
- Add empty content fallback in BuildAssistantMessageStruct (Claude format)
- Same fix as PR #181 applied to kiro_openai_request.go
Fixes compaction failures for OpenCode + Quotio + CLIProxyAPIPlus + Kiro stack
Claude Haiku 4.5 (claude-haiku-4-5-20251001) supports extended thinking
according to Anthropic's official documentation:
https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking
The model was incorrectly marked as not supporting thinking in the static
model definitions. This fix adds ThinkingSupport with the same parameters
as other Claude 4.5 models (Sonnet, Opus):
- Min: 1024 tokens
- Max: 128000 tokens
- ZeroAllowed: true
- DynamicAllowed: false
- Replaced repetitive string operations with a centralized `escapeGJSONPathKey` function.
- Streamlined handling of JSON schema cleaning for Gemini and Antigravity requests.
- Improved payload management by transitioning from byte slices to strings for processing.
- Removed unnecessary cloning of byte slices in several places.
Apply code review feedback from gemini-code-assist:
- Initialize inputMap upfront instead of using nested if blocks
- Combine Exists() and IsObject() checks into single condition
- Remove redundant nil check
Problem:
- PR #162 fixed empty string content but missed array content with tool_use
- OpenCode's compaction requests send assistant messages with content as array
- When content array contains only tool_use (no text), content becomes empty
- This causes 'Improperly formed request' errors from Kiro API
Example of problematic message format:
{
"role": "assistant",
"content": [
{"type": "tool_use", "id": "...", "name": "todowrite", "input": {...}}
]
}
Solution:
- Extract tool_use from content array (Anthropic/OpenCode format)
- This is in addition to existing tool_calls handling (OpenAI format)
- The empty content fallback from PR #162 will then work correctly
Fixes compaction failures that persisted after PR #162 merge.
Google official Gemini Python SDK sends thinking_level, thinking_budget,
and include_thoughts (snake_case) instead of thinkingLevel, thinkingBudget,
and includeThoughts (camelCase). This caused thinking configuration to be
ignored when using Python SDK.
Changes:
- Extract layer: extractGeminiConfig now reads snake_case as fallback
- Apply layer: Gemini/CLI/Antigravity appliers clean up snake_case fields
- Translator layer: Gemini->OpenAI/Claude/Codex translators support fallback
- Tests: Added 4 test cases for snake_case field coverage
Fixes#1426
- Add logic to handle `image` content type during request translation.
- Map Claude base64 image data to Gemini's `inlineData` structure.
- Support automatic extraction of `media_type` and `data` for image parts.
- Introduced a new `pprof` server to enable/debug HTTP profiling.
- Added configuration options for enabling/disabling and specifying the server address.
- Integrated pprof server lifecycle management with `Service`.
#1287
Add kiro as a new supported channel in GetStaticModelDefinitionsByChannel
function, enabling retrieval of Kiro model definitions alongside existing
providers like qwen, iflow, and github-copilot.
- Use email as primary identifier for IDC tokens (unique, no sequence needed)
- Add sequence number only when email is unavailable
- Use startUrl identifier as secondary fallback with sequence
- Update GenerateTokenFileName in aws.go with consistent logic
Apply code review feedback from gemini-code-assist:
- Define default messages as local constants to improve maintainability
- Avoid magic strings in the empty content handling logic
- Use util.ResolveAuthDir to properly expand ~ to user home directory
- Fixes issue where logs were created in literal "~/.cli-proxy-api" folder
- Add warning log when auth-dir resolution fails for debugging
Bug introduced in 62e2b67 (refactor(logging): centralize log directory
resolution logic), where strings.TrimSpace was used instead of
util.ResolveAuthDir to process auth-dir path.
Problem:
- OpenCode's /compaction command and auto-compaction (at 80%+ context)
sends requests that can result in empty assistant message content
- Kiro API strictly requires non-empty content for all messages
- This causes 'Bad Request: Improperly formed request' errors
- After compaction failure, the malformed message stays in history,
breaking all subsequent requests in the session
Solution:
- Add fallback content for empty assistant messages in
buildAssistantMessageFromOpenAI()
- Add history truncation (max 50 messages) to prevent oversized requests
- This ensures all messages have valid content before sending to Kiro API
Fixes issues with:
- /compaction command returning Bad Request
- Auto-compaction breaking sessions
- Conversations becoming unresponsive after compaction failure
Introduce `Filter` rules in the payload configuration to remove specified JSON paths from the payload. Update related helper functions and add examples to `config.example.yaml`.
- Add ErrorLogsMaxFiles config field with default value 10
- Support hot-reload via config file changes
- Add Management API: GET/PUT/PATCH /v0/management/error-logs-max-files
- Maintain SDK backward compatibility with NewFileRequestLogger (3 params)
- Add NewFileRequestLoggerWithOptions for custom error log retention
When request logging is disabled, forced error logs are retained up to
the configured limit. Set to 0 to disable cleanup.
- Modified extractKiroTokenData to support both Attributes and Metadata sources
- Fixes issue where JSON file-based tokens were not being read correctly
- FileSynthesizer stores tokens in Metadata, ConfigSynthesizer uses Attributes
- Now checks Attributes first (config.yaml), falls back to Metadata (JSON files)
- Ensures dynamic model fetching works for all Kiro authentication methods
- Prevents fallback to static model list that incorrectly includes opus for free accounts
Kiro API endpoints only exist in us-east-1, but OIDC region can vary
by Enterprise user location (e.g., ap-northeast-2 for Korean users).
Previously, when ProfileARN was not available, the code fell back to
using OIDC region for API calls, causing DNS resolution failures:
lookup codewhisperer.ap-northeast-2.amazonaws.com: no such host
This fix removes the OIDC region fallback for API endpoints.
The region priority is now:
1. api_region (explicit override)
2. ProfileARN region
3. us-east-1 (default)
Fixes: Issue #253 (200-400x slower response times due to DNS failures)
Switch from CodeWhisperer endpoint to Amazon Q endpoint for all auth types:
- Use q.{region}.amazonaws.com/generateAssistantResponse as primary endpoint
- Works universally across all AWS regions (CodeWhisperer only exists in us-east-1)
- Use application/json Content-Type instead of application/x-amz-json-1.0
- Remove X-Amz-Target header for Q endpoint (not required)
- Add x-amzn-kiro-agent-mode: vibe header
- Add x-amzn-codewhisperer-optout: true header
- Keep CodeWhisperer endpoint as fallback for compatibility
This change aligns with Amazon's consolidation of services under the Q branding
and provides better multi-region support for Enterprise/IDC users.
Introduce a custom HTTP client utilizing utls with Firefox TLS fingerprinting to bypass Cloudflare fingerprinting on Anthropic domains. Includes support for proxy configuration and enhanced connection management for HTTP/2.
Address @Xm798's feedback: OIDC region may differ from API region in some
Enterprise setups (e.g., OIDC in us-east-2, API in us-east-1).
Region priority (highest to lowest):
1. api_region - explicit override for API endpoint region
2. ProfileARN - extract region from arn:aws:service:REGION:account:resource
3. region - OIDC/Identity region (fallback)
4. us-east-1 - default
Changes:
- Add extractRegionFromProfileARN() to parse region from ARN
- Update getKiroEndpointConfigs() with 4-level region priority
- Add regionSource logging for debugging
Revert the Amazon Q endpoint path to root '/' instead of '/generateAssistantResponse'.
The '/generateAssistantResponse' path is only for CodeWhisperer endpoint with
'GenerateAssistantResponse' target. Amazon Q endpoint uses 'SendMessage' target
which requires the root path.
Thanks to @gemini-code-assist for catching this copy-paste error.
## Problem
- Kiro API endpoints were hardcoded to us-east-1 region
- Enterprise users in other regions (e.g., ap-northeast-2) experienced
significant latency (200-400x slower) due to cross-region API calls
- This is the API endpoint counterpart to quotio PR #241 which fixed
token refresh endpoints
## Solution
- Add buildKiroEndpointConfigs(region) function for dynamic endpoint generation
- Extract region from auth.Metadata["region"] field
- Fallback to us-east-1 for backward compatibility
- Use case-insensitive authMethod comparison (consistent with quotio PR #252)
## Changes
- Add kiroDefaultRegion constant
- Convert hardcoded endpoint URLs to dynamic fmt.Sprintf with region
- Update getKiroEndpointConfigs to extract and use region from auth
- Fix isIDCAuth to use case-insensitive comparison
## Testing
- Backward compatible: defaults to us-east-1 when no region specified
- Enterprise users can now use their local region endpoints
Related:
- quotio PR #241: Dynamic region for token refresh (merged)
- quotio PR #252: authMethod case-insensitive fix
- quotio Issue #253: Performance issue report
Add support for Gemini's code_execution and url_context tools in the
request translators, enabling:
- Agentic Vision: Image analysis with Python code execution for
bounding boxes, annotations, and visual reasoning
- URL Context: Live web page content fetching and analysis
Tools are passed through using the same pattern as google_search:
- code_execution: {} -> codeExecution: {}
- url_context: {} -> urlContext: {}
Tested with Gemini 3 Flash Preview agentic vision successfully.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Removes x-* extension fields from JSON schemas to ensure compatibility with the Gemini API.
These fields, while valid in OpenAPI/JSON Schema, are not recognized by the Gemini API and can cause issues.
The change recursively walks the schema, identifies these extension fields, and removes them, except when they define properties.
Amp-Thread-ID: https://ampcode.com/threads/T-019c0cd1-9e59-722b-83f0-e0582aba6914
Co-authored-by: Amp <amp@ampcode.com>
- Add ensureCacheControl() to auto-inject cache breakpoints
- Cache tools (last tool), system (last element), and messages (2nd-to-last user turn)
- Add prompt-caching-2024-07-31 beta header
- Return original payload on sjson error to prevent corruption
- Include verification test for caching logic
Enables up to 90% cost reduction on cached tokens.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add firstChunkTimestamp field to ResponseWriterWrapper for sync capture
- Capture TTFB in Write() and WriteString() before async channel send
- Add SetFirstChunkTimestamp() to StreamingLogWriter interface
- Make requestTimestamp/apiResponseTimestamp required in LogRequest()
- Remove timestamp capture from WriteAPIResponse() (now via setter)
- Fix Gemini handler to set API_RESPONSE_TIMESTAMP before writing response
This ensures accurate TTFB measurement for all streaming API formats
(OpenAI, Gemini, Claude) by capturing timestamp synchronously when
the first response chunk arrives, not when the stream finalizes.
Previously:
- REQUEST INFO timestamp was captured at log write time (not request arrival)
- API RESPONSE had NO timestamp at all
This fix:
- Captures REQUEST INFO timestamp when request first arrives
- Adds API RESPONSE timestamp when upstream response arrives
Changes:
- Add Timestamp field to RequestInfo, set at middleware initialization
- Set API_RESPONSE_TIMESTAMP in appendAPIResponse() and gemini handler
- Pass timestamps through logging chain to writeNonStreamingLog()
- Add timestamp output to API RESPONSE section
This enables accurate measurement of backend response latency in error logs.
When the Kiro/AWS CodeWhisperer API receives a Write tool request with content
that exceeds transmission limits, it truncates the tool input. This can result in:
- Empty input buffer (no input transmitted at all)
- Missing 'content' field in the parsed JSON
- Incomplete JSON that fails to parse
This fix detects these truncation scenarios and converts them to Bash tool calls
that echo an error message. This allows Claude Code to execute the Bash command,
see the error output, and the agent can then retry with smaller chunks.
Changes:
- kiro_claude_tools.go: Detect three truncation scenarios in ProcessToolUseEvent:
1. Empty input buffer (no input transmitted)
2. JSON parse failure with file_path but no content field
3. Successfully parsed JSON missing content field
When detected, emit a special '__truncated_write__' marker tool use
- kiro_executor.go: Handle '__truncated_write__' markers in streamToChannel:
1. Extract file_path from the marker for context
2. Create a Bash tool_use that echoes an error message
3. Include retry guidance (700-line chunks recommended)
4. Set hasToolUses=true to ensure stop_reason='tool_use' for agent continuation
This ensures the agent continues and can retry with smaller file chunks instead
of failing silently or showing errors to the user.
Introduce `TestExecuteStreamWithAuthManager_DoesNotRetryAfterFirstByte` to validate that stream executions do not retry after receiving partial responses. Implement `payloadThenErrorStreamExecutor` for test coverage of this behavior.
When using Gemini API format with Antigravity backend, the executor
renames usageMetadata to cpaUsageMetadata in non-terminal chunks.
The Gemini translator was returning this internal field name directly
to clients instead of the standard usageMetadata field.
Add restoreUsageMetadata() to rename cpaUsageMetadata back to
usageMetadata before returning responses to clients.
When Claude API sends an assistant message with empty text content like:
{"role":"assistant","content":[{"type":"text","text":""}]}
The translator was creating a part object {} with no data field,
causing Gemini API to return error:
"required oneof field 'data' must have one initialized field"
This fix:
1. Skips empty text parts (text="") during translation
2. Skips entire messages when their parts array becomes empty
This ensures compatibility when clients send empty assistant messages
in their conversation history.
Optimize channel operations by introducing reusable context-aware send functions (`send` and `sendErr`) across `wsrelay`, `handlers`, and `cliproxy`. Ensure graceful handling of canceled contexts during stream operations.
Changed the provider matching logic in AccountInfo() method to use
prefix matching instead of exact matching. This allows both 'github'
(Kiro OAuth) and 'github-copilot' providers to be correctly identified
as OAuth providers, enabling proper debug logging output.
Before: Use OAuth logs were missing for github-copilot requests
After: Logs show "Use OAuth provider=github-copilot auth_file=..."
Co-Authored-By: Claude (claude-sonnet-4.5) <noreply@anthropic.com>
**Problem:**
GitHub Copilot API returns 400 error "missing required Copilot-Vision-Request
header for vision requests" when requests contain image content blocks, even
though the requests are valid Claude API calls.
**Root Cause:**
The GitHub Copilot executor was not detecting vision content in requests and
did not add the required `Copilot-Vision-Request: true` header.
**Solution:**
- Added `detectVisionContent()` function to check for image_url/image content blocks
- Automatically add `Copilot-Vision-Request: true` header when vision content is detected
- Applied fix to both `Execute()` and `ExecuteStream()` methods
**Testing:**
- Tested with Claude Code IDE requests containing code context screenshots
- Vision requests now succeed instead of failing with 400 errors
- Non-vision requests remain unchanged
Fixes issue where GitHub Copilot executor fails all vision-enabled requests,
causing unnecessary fallback to other providers and 0% utilization.
Co-Authored-By: Claude (claude-sonnet-4.5) <noreply@anthropic.com>
The background refresher was skipping token files with auth_method values
like 'IdC' or 'IDC' because the comparison was case-sensitive and only
matched lowercase 'idc'.
This fix normalizes the auth_method to lowercase before comparison in:
- token_repository.go: readTokenFile() when filtering tokens to refresh
- background_refresh.go: refreshSingle() when selecting refresh method
Fixes the issue where 'IdC' != 'idc' caused tokens to be skipped entirely.
Implement `request_retry` and `disable_cooling` metadata overrides for authentication management. Update retry and cooling logic accordingly across `Manager`, Antigravity executor, and file synthesizer. Add tests to validate new behaviors.
Introduce `WithSkipPersist` to disable persistence during Manager Update/Register calls, preventing write-back loops caused by redundant file writes. Add corresponding tests and integrate with existing file store and conductor logic.
Refactor 401 error handling in both executeWithRetry and
executeStreamWithRetry to always attempt token refresh regardless of
remaining retry attempts. Previously, token refresh was only attempted
when retries remained, which could leave valid refreshed tokens unused.
Also add auth directory resolution in RefreshManager.Initialize to
properly resolve the base directory path before creating the token
repository.
Fixes#1082
When all Antigravity accounts are unavailable, the error response now returns
HTTP 502 (Bad Gateway) instead of HTTP 400 (Bad Request). This ensures that
NewAPI and other clients will retry the request on a different channel,
improving overall reliability.
- Add strings.ToLower() normalization in LoadKiroIDEToken()
- Add same normalization in LoadKiroTokenFromPath()
- Fixes issue where Kiro IDE exports "IdC" but code expects "idc"
Previously GeminiModels handler unconditionally overwrote displayName
and description with the model name, losing the original values defined
in model definitions (e.g., 'Gemini 3 Pro Preview').
Now only set these fields as fallback when they are missing or empty.
Add support for Imagen 3.0 and 4.0 image generation models in Vertex AI:
- Add 5 Imagen model definitions (4.0, 4.0-ultra, 4.0-fast, 3.0, 3.0-fast)
- Implement :predict action routing for Imagen models
- Convert Imagen request/response format to match Gemini structure like gemini-3-pro-image
- Transform prompts to Imagen's instances/parameters format
- Convert base64 image responses to Gemini-compatible inline data
Add new PATCH /v0/management/auth-files/status endpoint that allows
toggling the disabled state of auth files without deleting them.
This enables users to temporarily disable credentials from the
management UI.
- Introduced `ResolveLogDirectory` function in `logging` package to standardize log directory determination across components.
- Replaced redundant logic in `server`, `global_logger`, and `handlers` with the new utility function.
Update ApplyThinking signature to accept fromFormat and toFormat parameters
instead of a single provider string. This enables:
- Proper level-to-budget conversion when source is level-based (openai/codex)
and target is budget-based (gemini/claude)
- Strict budget range validation when source and target formats match
- Level clamping to nearest supported level for cross-format requests
- Format alias resolution in SDK translator registry for codex/openai-response
Also adds ErrBudgetOutOfRange error code and improves iflow config extraction
to fall back to openai format when iflow-specific config is not present.
Refactor thinking configuration tests by separating model name suffix-based
scenarios from request body parameter-based scenarios into distinct test
functions with independent case numbering.
Architectural improvements:
- Extract thinkingTestCase struct to package level for shared usage
- Add getTestModels() helper returning complete model fixture set
- Introduce runThinkingTests() runner with protocol-specific field detection
- Register level-subset-model fixture with constrained low/high level support
- Extend iflow protocol handling for glm-test and minimax-test models
- Add same-protocol strict boundary validation cases (80-89)
- Replace error responses with clamped values for boundary-exceeding budgets
Gemini API requires all enum values in function declarations to be
strings. Some MCP tools (e.g., roxybrowser) define schemas with numeric
enums like `"enum": [0, 1, 2]`, causing INVALID_ARGUMENT errors.
Add convertEnumValuesToStrings() to automatically convert numeric and
boolean enum values to their string representations during schema
transformation.
feat(translator): improve system message handling and content indexing across translators
- Updated logic for processing system messages in `claude`, `gemini`, `gemini-cli`, and `antigravity` translators.
- Introduced indexing for `systemInstruction.parts` to ensure proper ordering and handling of multi-part content.
- Added safeguards for accurate content transformation and serialization.
- Introduced `payloadModelAliases` and `payloadModelCandidates` functions to support model aliases for improved flexibility.
- Updated rule matching logic to handle multiple model candidates.
- Refactored variable naming in executor to improve code clarity and consistency.
- Enhanced ID matching in `cliproxy` by adding additional conditions to better handle ID equality cases.
- Updated `gemini` handlers to include `displayName` and `description` in normalized models for enriched metadata.
When tool results are sent back to the model, the system prompt was being
re-injected into the user message content, causing the model to think the
user had pasted the system prompt again. This was especially noticeable
after multiple tool uses.
The fix checks if there is conversation history (len(history) > 0). If so,
it's a subsequent turn and we skip system prompt injection. The system
prompt is only injected on the first turn (len(history) == 0).
This ensures:
- First turn: system prompt is injected
- Tool result turns: system prompt is NOT re-injected
- New conversations: system prompt is injected fresh
- Added conditional logic for Codex instruction injection based on configuration.
- Updated role terminology from "user" to "developer" for better alignment with context.
- Added logic to transform `inputResults` into structured JSON for improved processing.
- Removed redundant `safety_identifier` field in executor payload to streamline requests.
- Add x-amzn-kiro-agent-mode: vibe for non-IDC auth (Social, Builder ID)
IDC auth continues to use "spec" mode
- Add x-amzn-codewhisperer-optout: true for all auth types
This opts out of data sharing for service improvement (privacy)
These changes align with other Kiro implementations (kiro.rs, KiroGate,
kiro-gateway, AIClient-2-API) and make requests more similar to real
Kiro IDE clients.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The Q endpoint was using `/` which caused all requests to fail with
400 or UnknownOperationException. Changed to `/generateAssistantResponse`
which is the correct path for the Q endpoint.
This fix restores the Q endpoint failover functionality.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Introduced `kiro` channel and alias resolution in `oauth_model_alias` logic.
- Updated supported channels documentation and examples to include `kiro` and `github-copilot`.
- Enhanced unit tests to validate `kiro` alias functionality.
- Introduced `default-raw` and `override-raw` rules to handle raw JSON values.
- Enhanced `PayloadConfig` to validate and sanitize raw JSON payload rules.
- Updated executor logic to apply `default-raw` and `override-raw` rules.
- Extended example YAML to include usage of raw JSON rules.
This change removes the translation logic for several non-standard, proprietary extensions used to configure thinking/reasoning. Specifically, support for `extra_body.google.thinking_config` and the Anthropic-style `thinking` object has been dropped from the OpenAI request translators.
This simplification streamlines the translators, focusing them on the standard `reasoning_effort` parameter. It also removes the need to look up model information from the registry within these components.
BREAKING CHANGE: Support for non-standard thinking configurations via `extra_body.google.thinking_config` and the Anthropic-style `thinking` object has been removed. Clients should now use the standard `reasoning_effort` parameter to control reasoning.
Enhance model definitions by including supported API endpoints for each model. This allows for better integration and usage tracking with the GitHub Copilot API.
- Added `metadataEqualIgnoringTimestamps` to compare metadata while ignoring volatile fields.
- Prevented redundant writes caused by changes in timestamp-related fields.
- Improved efficiency in filestore operations by skipping unnecessary updates.
feat(auth): fetch and update Antigravity project ID from metadata during filestore operations
- Added support to retrieve and update `project_id` using the access token if missing in metadata.
- Integrated HTTP client to fetch project ID dynamically.
- Enhanced metadata persistence logic.
- Updated `ideType` to `ANTIGRAVITY` in request payload.
- Introduced tier-selection logic to determine default tier for onboarding.
- Added `antigravityOnboardUser` function for project ID retrieval via polling.
- Enhanced error handling and response decoding for onboarding flow.
This change introduces environment variable interpolation for volume paths, allowing users to customize where configuration, authentication, and log data are stored.
Why: Makes the project easier to deploy on various hosting environments that require decoupled data management without needing to modify the core docker-compose.yml..
Key points:
Defaults to existing paths (./config.yaml, ./auths, ./logs) to ensure zero breaking changes for current users.
Follows the existing naming convention used in the project.
Enhances portability for CI/CD and cloud-native deployments.
feat(translator): add function name to response output item serialization
- Included `item.name` in the serialized response output to enhance output item handling.
Compute a size-reduction based keep ratio and use it to trim
tool descriptions, avoiding forced minimum truncation when the
target size already fits. This aligns compression with actual
payload reduction needs and prevents over-compression.
feat(oauth): add support for customizable OAuth callback ports
- Introduced `oauth-callback-port` flag to override default callback ports.
- Updated SDK and login flows for `iflow`, `gemini`, `antigravity`, `codex`, `claude`, and `openai` to respect configurable callback ports.
- Refactored internal OAuth servers to dynamically assign ports based on the provided options.
- Revised tests and documentation to reflect the new flag and behavior.
When switching from Claude models (e.g., Opus 4.5) to Gemini models
(e.g., Flash) mid-conversation via Antigravity OAuth, the client-provided
thinking signatures from Claude would cause "Corrupted thought signature"
errors since they are incompatible with Gemini API.
Changes:
- Remove fallback to client-provided signatures in thinking block handling
- Only use cached signatures (from same-session Gemini responses)
- Skip thinking blocks without valid cached signatures
- tool_use blocks continue to use skip_thought_signature_validator when
no valid signature is available
This ensures cross-model switching works correctly while preserving
signature validation for same-model conversations.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Updated `SDKConfig` to use `nonstream-keepalive-interval` (seconds) instead of the boolean `nonstream-keepalive`.
- Refactored handlers and logic to incorporate the new interval-based configuration.
- Updated config diff, tests, and example YAML to reflect the changes.
- Introduced `StartNonStreamingKeepAlive` to emit periodic blank lines during non-streaming responses.
- Added `nonstream-keepalive` configuration option in `SDKConfig`.
- Updated handlers to utilize `StartNonStreamingKeepAlive` and ensure proper cleanup.
- Extended config diff and tests to include `nonstream-keepalive` changes.
feat(cliproxy): support multiple aliases for OAuth model mappings
- Updated mapping logic to allow multiple aliases per upstream model name.
- Adjusted `SanitizeOAuthModelMappings` to ensure aliases remain unique within channels.
- Added test cases to validate multi-alias scenarios.
- Updated example config to clarify multi-alias support.
Keep the upstream payload prefixed for OAuth while passing the unprefixed request body into response translators. This avoids proxy_ leaking into OpenAI Responses echoed tool metadata while preserving the Claude OAuth workaround.
Prefix tool names with proxy_ for Claude OAuth requests and strip the prefix from streaming and non-streaming responses to restore client-facing names.
Updates the Claude executor to:
- add prefixing for tools, tool_choice, and tool_use messages when using OAuth tokens
- strip the prefix from tool_use events in SSE and non-streaming payloads
- add focused unit tests for prefix/strip helpers
Update the model name check in `buildRequest` to target "gemini-3-pro-preview" instead of "gemini-3-pro" when applying specific system instruction handling.
When streaming responses with tool calls, the finish_reason was being
overwritten. The upstream sends functionCall in chunk 1, then
finishReason: STOP in chunk 2. The old code would set finish_reason
from every chunk, causing "tool_calls" to be overwritten by "stop".
This broke clients like Claude Code that rely on finish_reason to
detect when tool calls are complete.
Changes:
- Add SawToolCall bool to track tool calls across entire stream
- Add UpstreamFinishReason to cache the finish reason
- Only emit finish_reason on final chunk (has both finishReason + usage)
- Priority: tool_calls > max_tokens > stop
Includes 5 unit tests covering:
- Tool calls not overwritten by subsequent STOP
- Normal text gets "stop"
- MAX_TOKENS without tool calls gets "max_tokens"
- Tool calls take priority over MAX_TOKENS
- Intermediate chunks have no finish_reason
Fixes streaming tool call detection for Claude Code + Gemini models.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Previously, metadataEqualIgnoringTimestamps() ignored access_token for all
providers, which prevented refreshed tokens from being persisted to disk/database.
This caused tokens to be lost on server restart for providers like iFlow.
This change makes the behavior provider-specific:
- Providers like gemini/gemini-cli that issue new tokens on every refresh and
can re-fetch when needed will continue to ignore access_token (optimization)
- Other providers like iFlow will now persist access_token changes
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
integrate claude-cloak functionality to disguise api requests:
- add CloakConfig with mode (auto/always/never) and strict-mode options
- generate fake user_id in claude code format (user_[hex]_account__session_[uuid])
- inject claude code system prompt (configurable strict mode)
- obfuscate sensitive words with zero-width characters
- auto-detect claude code clients via user-agent
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Add Kiro to the list of supported channels for OAuth model name mappings,
allowing users to map Kiro model IDs (e.g., kiro-claude-opus-4-5) to
canonical model names (e.g., claude-opus-4-5-20251101).
The Kiro case is implemented as a separate switch block to keep it
isolated from upstream CLIProxyAPI providers, making future merges
from the upstream repository cleaner.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Enhanced node structure by including `thoughtSignature` for inline data parts in Gemini OpenAI, Gemini CLI, and Antigravity request handlers to improve traceability of thought processes.
Fixes issue where free tier users cannot access Gemini 3 preview models
due to frontend/backend project ID mapping.
## Problem
Google's Gemini API uses a frontend/backend project mapping system for
free tier users:
- Frontend projects (e.g., gen-lang-client-*) are user-visible
- Backend projects (e.g., mystical-victor-*) host actual API access
- Only backend projects have access to preview models (gemini-3-*)
Previously, CLIProxyAPI ignored the backend project ID returned by
Google's onboarding API and kept using the frontend ID, preventing
access to preview models.
## Solution
### CLI (internal/cmd/login.go)
- Detect free tier users (gen-lang-client-* projects or FREE/LEGACY tier)
- Show interactive prompt allowing users to choose frontend or backend
- Default to backend (recommended for preview model access)
- Pro users: maintain original behavior (keep frontend ID)
### Web UI (internal/api/handlers/management/auth_files.go)
- Detect free tier users using same logic
- Automatically use backend project ID (recommended choice)
- Pro users: maintain original behavior (keep frontend ID)
### Deduplication (internal/cmd/login.go)
- Add deduplication when user selects ALL projects
- Prevents redundant API calls when multiple frontend projects map to
same backend
- Skips duplicate project IDs in activation loop
## Impact
- Free tier users: Can now access gemini-3-pro-preview and
gemini-3-flash-preview models
- Pro users: No change in behavior (backward compatible)
- Only affects Gemini CLI OAuth (not antigravity or API key auth)
## Testing
- Tested with free tier account selecting single project
- Tested with free tier account selecting ALL projects
- Verified deduplication prevents redundant onboarding calls
- Confirmed pro user behavior unchanged
Remove os.RemoveAll() call in syncAuthFromBucket() that was causing
a race condition with the file watcher.
Problem:
1. syncAuthFromBucket() wipes local auth directory with RemoveAll
2. File watcher detects deletions and propagates them to remote store
3. syncAuthFromBucket() then pulls from remote, but files are now gone
Solution:
Use incremental sync instead of delete-then-pull. Just ensure the
directory exists and overwrite files as they're downloaded.
This prevents the watcher from seeing spurious delete events.
Fixed incorrect boundary logic for `message_delta` emission, ensuring proper handling of usage updates and `emitMessageStopIfNeeded` within the response loop.
Implemented `Fork` flag in `ModelNameMapping` to allow aliases as additional models while preserving the original model ID. Updated the `applyOAuthModelMappings` logic, added tests for `Fork` behavior, and updated documentation and examples accordingly.
Introduced `tokenRefreshTimeout` constant for token refresh operations and enhanced context propagation for `refreshToken` by embedding roundtrip information if available. Adjusted `refreshAuth` to ensure default context initialization and handle cancellation errors appropriately.
Replaced file-based auth entry counting with `TokenStore`-backed implementation, enhancing flexibility and context-aware token management. Updated related logic to reflect this change.
Added support for external hooks to observe model registry events using the `ModelRegistryHook` interface. Implemented thread-safe, non-blocking execution of hooks with panic recovery. Comprehensive tests added to verify hook behavior during registration, unregistration, blocking, and panic scenarios.
Updated `applyPayloadConfig` to `applyPayloadConfigWithRoot` across payload translation logic, enabling validation against the original request payload when available. Added support for improved model normalization and translation consistency.
Refactored `applyPayloadConfig` to `applyPayloadConfigWithRoot`, adding support for default rule validation against the original payload when available. Updated all executors to use `applyPayloadConfigWithRoot` and incorporate an optional original request payload for translations.
This commit reverts all modifications within internal/translator. A separate issue
will be created for the maintenance team to integrate SanitizeFunctionName into
the translators.
Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>
This commit centralizes tool name sanitization in SanitizeFunctionName,
applying character compliance, starting character rules, and length limits.
It also fixes a regression in gemini_schema tests and preserves MCP-specific
shortening logic while ensuring compliance.
Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>
Implemented SanitizeFunctionName utility to ensure Claude tool names meet
Gemini/Upstream strict naming conventions (alphanumeric, starts with letter/underscore, max 64 chars).
Applied sanitization to tool definitions and usage in all relevant translators.
Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>
When receiving HTTP 429 (Too Many Requests) responses, parse the retry
delay from the response body using parseRetryDelay and populate the
statusErr.retryAfter field. This allows upstream callers to respect
the server's requested retry timing.
Applied to all error paths in Execute, executeClaudeNonStream,
ExecuteStream, CountTokens, and refreshToken functions.
When multiple OAuth providers share an account email, the existing "Use OAuth" debug lines are ambiguous and hard to correlate with management usage stats. Include provider, auth file, and auth index in the selection log, and only compute these fields when debug logging is enabled to avoid impacting normal request performance.
Before:
[debug] Use OAuth user@example.com for model gemini-3-flash-preview
[debug] Use OAuth user@example.com (project-1234) for model gemini-3-flash-preview
After:
[debug] Use OAuth provider=antigravity auth_file=antigravity-user_example_com.json auth_index=1a2b3c4d5e6f7788 account="user@example.com" for model gemini-3-flash-preview
[debug] Use OAuth provider=gemini-cli auth_file=gemini-user@example.com-project-1234.json auth_index=99aabbccddeeff00 account="user@example.com (project-1234)" for model gemini-3-flash-preview
Anthropic API does not allow extended thinking when tool_choice is set
to "any" or a specific tool. This was causing 400 errors when using
features like Amp's /handoff command which forces tool_choice.
Added disableThinkingIfToolChoiceForced() that removes thinking config
when incompatible tool_choice is detected, applied to both streaming
and non-streaming paths.
Fixesrouter-for-me/CLIProxyAPI#630
- GLM-4.7: Uses extra_body={"thinking": {"type": "enabled"}, "clear_thinking": false}
- MiniMax-M2.1: Uses reasoning_split=true for OpenAI-style reasoning separation
- Added preserveReasoningContentInMessages() to support re-injection of reasoning
content in assistant message history for multi-turn conversations
- Added ThinkingSupport to MiniMax-M2.1 model definition
Remove dead code that was never used:
- toolCallIDToName map: built but never read from
- seenToolCallIDs: declared but never populated, only suppressed with _
Fix three issues in Kiro OpenAI translator that caused "Improperly formed request"
errors when processing LiteLLM-translated requests with tool_use/tool_result:
1. Skip merging tool role messages in MergeAdjacentMessages() to preserve
individual tool_call_id fields
2. Track pendingToolResults and attach to the next user message instead of
only the last message. Create synthetic user message when conversation
ends with tool results.
3. Insert synthetic user message with tool results before assistant messages
to maintain proper alternating user/assistant structure. This fixes the case
where LiteLLM translates Anthropic user messages containing only tool_result
blocks into tool role messages followed by assistant.
Adds unit tests covering all tool result handling scenarios.
Improved consistency across OpenAI, Claude, and Gemini handlers by replacing initial `select` statement with a `for` loop for better readability and error-handling robustness.
Refined header assignment to use `x-api-key` for Anthropic API requests, ensuring correct authorization behavior based on request attributes and URL validation.
- Handle errors in promptInput instead of ignoring them
- Improve promptSelect to provide feedback on invalid input and re-prompt
- Use sentinel errors (ErrAuthorizationPending, ErrSlowDown) instead of
string-based error checking with strings.Contains
- Move hardcoded x-amz-user-agent header to idcAmzUserAgent constant
Addresses code review feedback from Gemini Code Assist.
Add persistRefreshedAuth function to write refreshed tokens back to the
auth file after inline token refresh. This prevents repeated token
refreshes on every request when the token expires.
Changes:
- Add persistRefreshedAuth() to kiro_executor.go
- Call persist after all token refresh paths (401, 403, pre-request)
- Remove unused log import from sdk/auth/kiro.go
Updated Antigravity, Gemini, and Gemini-CLI translators to process `systemResult` of type `string` for system instructions. Ensures properly formatted JSON with dynamic content assignment.
fix(antigravity): validate function arguments before serialization
Ensure `function.arguments` is a valid JSON before setting raw bytes, fallback to setting as parameterized content if invalid.
feat: handle array input for system instructions in translators
Enhanced Gemini, Gemini-CLI, and Antigravity translators to process array content for system instructions. Adds support for assigning roles and handling multiple content parts dynamically.
Added comprehensive tests for `FillFirstSelector` and `RoundRobinSelector` to ensure proper behavior, including deterministic, cyclical, and concurrent scenarios. Introduced dynamic routing strategy updates in `service.go`, normalizing strategies and seamlessly switching between `fill-first` and `round-robin`. Updated `Manager` to support selector changes via the new `SetSelector` method.
Optimized the handling of JSON serialization and deserialization by replacing redundant `json.Marshal` and `json.Unmarshal` calls with `sjson` and `gjson`. Introduced a `marshalJSONValue` utility for compact JSON encoding, improving performance and code simplicity. Removed unused `encoding/json` imports.
This update enhances the `FileRequestLogger` by introducing support for spooling large request and response bodies to temporary files, reducing memory consumption. It adds atomic requestLogID generation for sequential log naming and new methods for non-streaming/streaming log assembly. Also includes better error handling during logging and temp file cleanups.
Improve parsing of tool call inputs and Antigravity compatibility to avoid invalid thinking/tool_use errors.
- Parse tool call inputs robustly by accepting both object and JSON-string formats and only produce a functionCall part when valid args exist, reducing spurious or malformed parts.
- Preserve the skip_thought_signature_validator approach for calls without a valid thinking signature but stop toggling/tracking a separate "disable thinking" flag; this prevents unnecessary removal of thinkingConfig.
- Sanitize tool input schemas before attaching them to the Antigravity request to improve compatibility.
- Append the interleaved-thinking hint as a new parts entry instead of overwriting/setting text directly, preserving structure.
- Remove unused tracking logic and related comments to simplify flow.
These changes reduce errors related to missing/invalid thinking signatures, improve schema compatibility, and make hint injection safer and more consistent.
Removes the addition of the "anthropic-beta: interleaved-thinking-2025-05-14" header for Claude thinking models when building HTTP requests.
This prevents sending an experimental/feature flag header that is no longer required and avoids potential compatibility or routing issues with downstream services. Keeps request headers simpler and more standard.
Prefer cached signatures and avoid injecting dummy thinking blocks; instead remove unsigned thinking blocks and add a skip sentinel for tool calls without a valid signature. Generate stable session IDs from the first user message, apply schema cleaning only for Claude models, and reorder thinking parts so thinking appears first. For Gemini, remove thinking blocks and attach a skip sentinel to function calls. Simplify response handling by passing raw function args through (remove special Bash conversion). Update and add tests to reflect the new behavior.
These changes prevent rejected dummy signatures, improve compatibility with Antigravity’s signature validation, provide more stable session IDs for conversation grouping, and make request/response translation more robust.
Remove the opencode-antigravity-auth submodule reference from the repository.
Cleans up the project by eliminating an external submodule pointer that is no longer needed or maintained, reducing repository complexity and avoiding dangling submodule state.
Removes two obsolete git submodules to clean up repository state and reduce maintenance overhead.
This eliminates external references that are no longer needed, simplifying dependency management and repository maintenance going forward.
fix: unify response field naming across translators
Standardize `text` to `delta` and add missing `output` field in all response payloads for consistency across OpenAI, Claude, and Gemini translators.
This commit introduces several improvements to the AMP (Advanced Model Proxy) module:
- **Model Mapping Logic:** The `FallbackHandler` now uses a more robust approach for model mapping. It includes the extraction and preservation of dynamic "thinking suffixes" (e.g., `(xhigh)`) during mapping, ensuring that these configurations are correctly applied to the mapped model. A new `resolveMappedModel` function centralizes this logic for cleaner code.
- **ModelMapper Verification:** The `ModelMapper` in `model_mapping.go` now verifies that the target model of a mapping has available providers *after* normalizing it. This prevents mappings to non-existent or unresolvable models.
- **Gemini Thinking Configuration Cleanup:** In `gemini_thinking.go`, unnecessary `generationConfig.thinkingConfig.include_thoughts` and `generationConfig.thinkingConfig.thinkingBudget` fields are now deleted from the request body when applying Gemini thinking levels. This prevents potential conflicts or redundant configurations.
- **Testing:** A new test case `TestModelMapper_MapModel_TargetWithThinkingSuffix` has been added to `model_mapping_test.go` to specifically cover the preservation of thinking suffixes during model mapping.
This commit introduces a new configuration option `logs-max-total-size-mb` that allows users to set a maximum total size (in MB) for log files in the logs directory. When this limit is exceeded, the oldest log files will be automatically deleted to stay within the specified size. Setting this value to 0 (the default) disables this feature. This change enhances log management by preventing excessive disk space usage.
Add applyPayloadConfig calls to all Antigravity executor paths (Execute,
executeClaudeNonStream, ExecuteStream) to enable config.yaml payload
overrides for Antigravity/Gemini-Claude models.
This allows users to configure thinking budget and other parameters via
payload.override in config.yaml for models like gemini-claude-opus-4-5*.
Refactors the reasoning effort conversion logic for Gemini models.
The update specifically addresses how `reasoning_effort` is translated into Gemini 3 specific thinking configurations (`thinkingLevel`, `includeThoughts`) and ensures that numeric budgets are not incorrectly applied to level-based models.
Changes include:
- Differentiating conversion logic for Gemini 3 models versus other models.
- Handling `none`, `auto`, and validated thinking levels for Gemini 3.
- Maintaining existing conversion for models not using discrete thinking levels.
The prompt for the gpt-5.2 codex model has been updated with more comprehensive instructions. This includes detailed guidelines on general usage, editing constraints, the plan tool, sandboxing configurations, handling special user requests, frontend task considerations, and final message presentation. The updates aim to improve the model's understanding and execution of complex coding tasks by providing clearer directives and constraints.
Improve the cache eviction routine to sort entries by timestamp using the standard library sort routine (stable, clearer and faster than the prior manual selection/bubble logic), and remove a redundant request-derived session ID helper in favor of the centralized session ID function. Also drop now-unused crypto/encoding imports.
This yields clearer, more maintainable eviction logic and removes duplicated/unused code and imports to reduce surface area and potential inconsistencies.
Improve robustness when handling "thinking" content by using a dedicated helper to extract the thinking text. This ensures wrapped or nested thinking objects are handled correctly instead of relying on a direct string extraction, reducing parsing errors for complex payloads.
This change introduces specific logic to load and use instructions for the 'gpt-5.2-codex' model variant by recognizing the 'gpt-5.2-codex_prompt.md' filename. This ensures the correct prompts are used when the '5.2-codex' model is identified, complementing the recent addition of its definition.
Normalize Bash tool arguments by converting a "command" key into "cmd" using JSON-aware parsing, avoiding brittle string replacements that could corrupt values. Apply this conversion in both streaming and non-streaming response paths so bash-style tool calls are emitted with the expected "cmd" field.
Add support for accumulating thinking text and carrying session identifiers to enable signature caching/restore for unsigned thinking blocks, improving handling of thinking-state continuity across requests/responses.
Also perform small cleanups: import logging, tidy comments and test descriptions. These changes make tool-argument handling more robust and enable reliable signature restoration for thinking blocks.
Ensures compatibility with the Amp client by suppressing
"thinking" blocks when "tool_use" blocks are also present in
the response.
The Amp client has issues rendering both types of blocks
simultaneously. This change filters out "thinking" blocks in
such cases, preventing rendering problems.
Gemini API does not support the JSON Schema `propertyNames` keyword,
causing 400 errors when Claude tool schemas containing this field are
proxied through the Antigravity provider.
Add `propertyNames` to the list of unsupported keywords removed by
CleanJSONSchemaForGemini(), alongside existing removals like $ref,
definitions, and additionalProperties.
Introduce a centralized OAuth session store with TTL-based expiration
to replace the previous simple map-based status tracking. Add a new
/api/oauth/callback endpoint that allows remote clients to relay OAuth
callback data back to the CLI proxy, enabling OAuth flows when the
callback cannot reach the local machine directly.
- Add oauth_sessions.go with thread-safe session store and validation
- Add oauth_callback.go with POST handler for remote callback relay
- Refactor auth_files.go to use new session management APIs
- Register new callback route in server.go
Add metadataEqualIgnoringTimestamps() function to compare metadata JSON
without timestamp/expired/expires_in/last_refresh/access_token fields.
This prevents unnecessary file writes when only these fields change
during refresh, breaking the fsnotify event → Watcher callback → refresh loop.
Key insight: Google OAuth returns a new access_token on each refresh,
which was causing file writes and triggering the refresh loop.
Fixes antigravity channel excessive log generation issue.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
The previous commit added thinkingLevel support but didn't apply it
when the reasoning effort came from model name suffix (e.g., model(minimal)).
This was because ResolveThinkingConfigFromMetadata returns nil for
level-based models, bypassing the metadata application.
Changes:
- Add ApplyGemini3ThinkingLevelFromMetadata for standard Gemini API
- Add ApplyGemini3ThinkingLevelFromMetadataCLI for CLI API format
- Update gemini_cli_executor to apply Gemini 3 thinkingLevel from metadata
- Update antigravity_executor to apply Gemini 3 thinkingLevel from metadata
- Update aistudio_executor to apply Gemini 3 thinkingLevel from metadata
- Add comprehensive test coverage for Gemini 3 thinkingLevel functions
Per Google's official documentation, Gemini 3 models should use
thinkingLevel (string) instead of thinkingBudget (number) for
optimal performance.
From Google's Gemini Thinking docs:
> Use the thinkingLevel parameter with Gemini 3 models. While
> thinkingBudget is accepted for backwards compatibility, using
> it with Gemini 3 Pro may result in suboptimal performance.
Changes:
- Add model family detection functions (IsGemini3Model, IsGemini25Model,
IsGemini3ProModel, IsGemini3FlashModel)
- Add ApplyGeminiThinkingLevel and ApplyGeminiCLIThinkingLevel functions
for applying thinkingLevel config
- Add ValidateGemini3ThinkingLevel for model-specific level validation
- Add ThinkingBudgetToGemini3Level for backward compatibility conversion
- Update NormalizeGeminiThinkingBudget to convert budget to level for
Gemini 3 models
- Update ApplyDefaultThinkingIfNeeded to not set a default level for
Gemini 3 (lets API use its dynamic default "high")
- Update ConvertThinkingLevelToBudget to preserve thinkingLevel for
Gemini 3 models
- Add Levels field to all Gemini 3 model definitions:
- Gemini 3 Pro: ["low", "high"]
- Gemini 3 Flash: ["minimal", "low", "medium", "high"]
Backward compatibility:
- Gemini 2.5 models continue to use thinkingBudget as before
- If thinkingBudget is provided for Gemini 3, it's converted to the
appropriate thinkingLevel
- Existing configurations continue to work
Introduces the capability to count tokens for Antigravity-backed requests. This implementation leverages the `countTokens` endpoint of the Antigravity API, replacing the prior unsupported stub.
Key aspects of this update include:
- **API Integration**: Direct integration with the Antigravity `countTokens` API, including necessary request payload translation and authentication.
- **Resilient Infrastructure**: A fallback mechanism has been established, allowing the system to attempt connections across multiple Antigravity base URLs to ensure request success even in the event of temporary service interruptions.
- **Model Aliasing**: Added mappings for `gemini-3-flash` and `gemini-3-flash-preview` to ensure compatibility with the latest model variants.
- **Robust Error Handling**: Comprehensive error handling and logging are in place to manage failures during API interactions.
Updates schema flattening logic to handle multiple non-null types, providing a more descriptive "Accepts" hint.
Removes redundant tracking of the current tool name in `Params` as it's no longer needed for streaming limits, simplifying the structure.
Enhances compatibility with the Gemini API by implementing a schema cleaning process.
This includes:
- Centralizing schema cleaning logic for Gemini in a dedicated utility function.
- Converting unsupported schema keywords to hints within the description field.
- Flattening complex schema structures like `anyOf`, `oneOf`, and type arrays to simplify the schema.
- Handling streaming responses with empty tool names, which can occur in subsequent chunks after the initial tool use.
Enhance the configuration diff logic to include detection and reporting of `prefix` changes for all model types. Update related struct naming for consistency across the watcher module.
Normalize action handling by accommodating wildcard patterns in route definitions for Gemini endpoints. Adjust `request.Action` parsing logic to correctly process routes with prefixed actions.
Refactor model management to include an optional `prefix` field for model credentials, enabling better namespace handling. Update affected configuration files, APIs, and handlers to support prefix normalization and routing. Remove unused OpenAI compatibility provider logic to simplify processing.
Introduce formatProxyURL helper to sanitize proxy addresses before
logging, stripping credentials and path components while preserving
host information. Rework model hash computation to sort and deduplicate
name/alias pairs with case normalization, ensuring consistent output
regardless of input ordering. Add signature-based identification for
anonymous OpenAI-compatible provider entries to maintain stable keys
across configuration reloads. Replace direct stdout prints with
structured logger calls for file change notifications.
Break out config diffing, hashing, and OpenAI compatibility utilities into a dedicated diff package, update watcher to consume them, and add comprehensive tests for diff logic and watcher behavior.
Unify thinking budget-to-effort conversion in a shared helper, handle disabled/default thinking cases in translators, adjust zero-budget mapping, and drop the old OpenAI-specific helper with updated tests.
Some OpenAI-compatible providers (like GitHub Copilot) may send tool_calls
in the first streaming chunk without including the role field. The previous
implementation only emitted message_start when the first chunk contained
role="assistant", causing Anthropic protocol violations when tool calls
arrived first.
This fix ensures message_start is always emitted on the very first chunk,
preventing 'content_block_start before message_start' errors in clients
that strictly validate Anthropic SSE event ordering.
Introduce `panel-github-repository` in the configuration to allow specifying a custom repository for management panel assets. Update dependency versions and enhance asset URL resolution logic to support overrides.
- Pass through non-function tool definitions like web_search
- Translate tool_choice for built-in tools and function tools
- Add regression tests for built-in tool passthrough
The response writer wrapper has been refactored to more reliably capture response bodies for logging, fixing several edge cases.
- Implements `WriteString` to capture writes from `io.StringWriter`, which were previously missed by the `Write` method override.
- A new `shouldBufferResponseBody` helper centralizes the logic to ensure the body is buffered only when logging is active or for errors when `logOnErrorOnly` is enabled.
- Streaming detection is now more robust. It correctly handles non-streaming error responses (e.g., `application/json`) that are generated for a request that was intended to be streaming.
BREAKING CHANGE: The public methods `Status()`, `Size()`, and `Written()` have been removed from the `ResponseWriterWrapper` as they are no longer required by the new implementation.
The WriteErrorResponse function now caches the error response body in the gin context.
The deferred request logger checks for this cached response. If an error response is found, it bypasses the standard response logging. This prevents scenarios where an error is logged twice or an empty payload log overwrites the original, more detailed error log.
The API error handling is updated to return a structured JSON payload
instead of a plain text message. This provides more context and allows
clients to programmatically handle different error types.
The new error response has the following structure:
{
"error": {
"message": "...",
"type": "..."
}
}
The `type` field is determined by the HTTP status code, such as
`authentication_error`, `rate_limit_error`, or `server_error`.
If the underlying error message from an upstream service is already a
valid JSON string, it will be preserved and returned directly.
BREAKING CHANGE: API error responses are now in a structured JSON
format instead of plain text. Clients expecting plain text error
messages will need to be updated to parse the new JSON body.
All Amp management endpoints (e.g., /api/user, /threads) are now protected by the standard API key authentication middleware. This ensures that all management operations require a valid API key, significantly improving security.
As a result of this change:
- The `restrict-management-to-localhost` setting now defaults to `false`. API key authentication provides a stronger and more flexible security control than IP-based restrictions, improving usability in containerized environments.
- The reverse proxy logic now strips the client's `Authorization` header after authenticating the initial request. It then injects the configured `upstream-api-key` for the request to the upstream Amp service.
BREAKING CHANGE: Amp management endpoints now require a valid API key for authentication. Requests without a valid API key in the `Authorization` header will be rejected with a 401 Unauthorized error.
Expose thinking/effort normalization helpers from the executor package
so conversion tests use production code and stay aligned with runtime
validation behavior.
Route OpenAI reasoning effort through ThinkingEffortToBudget for Claude
translators, preserve "minimal" when translating OpenAI Responses, and
treat blank/unknown efforts as no-ops for Gemini thinking configs.
Also map budget -1 to "auto" and expand cross-protocol thinking tests.
When budget 0 maps to "none" for models that use thinking levels
but don't support that effort level, strip thinking fields instead
of setting an invalid reasoning_effort value.
Tests now expect removal for this edge case.
Move OpenAI `reasoning_effort` -> Gemini `thinkingConfig` budget logic into
shared helpers used by Gemini, Gemini CLI, and antigravity translators.
Normalize Claude thinking handling by preferring positive budgets, applying
budget token normalization, and gating by model support.
Always convert Gemini `thinkingBudget` back to OpenAI `reasoning_effort` to
support allowCompat models, and update tests for normalization behavior.
Only map OpenAI reasoning effort to Claude thinking for models that support
thinking and use budget tokens (not level-based thinking).
Also add "xhigh" effort mapping and adjust minimal/low budgets, with new
raw-payload conversion tests across protocols and models.
- **Thinking Support**:
- Enabled thinking support for all Kiro Claude models, including Haiku 4.5 and agentic variants.
- Updated `model_definitions.go` with thinking configuration (Min: 1024, Max: 32000, ZeroAllowed: true).
- Fixed `extended_thinking` field names in `model_registry.go` (from `min_budget`/`max_budget` to `min`/`max`) to comply with Claude API specs, enabling thinking control in clients like Claude Code.
- **Kiro Executor Fixes**:
- Fixed `budget_tokens` handling: explicitly disable thinking when budget is 0 or negative.
- Removed aggressive duplicate content filtering logic that caused truncation/data loss.
- Enhanced thinking tag parsing with `extractThinkingFromContent` to correctly handle interleaved thinking/text blocks.
- Added EOF handling to flush pending thinking tag characters, preventing data loss at stream end.
- **Performance**:
- Optimized Claude stream handler (v6.2) with reduced buffer size (4KB) and faster flush interval (50ms) to minimize latency and prevent timeouts.
Ensure thinking settings translate correctly across providers:
- Only apply reasoning_effort to level-based models and derive it from numeric
budget suffixes when present
- Strip effort string fields for budget-based models and skip Claude/Gemini
budget resolution for level-based or unsupported models
- Default Gemini include_thoughts when a nonzero budget override is set
- Add cross-protocol conversion and budget range tests
When using OpenAI-compatible providers with model aliases (e.g., glm-4.6-zai -> glm-4.6),
the alias resolution was correctly applied but then immediately overwritten by
ResolveOriginalModel, causing 'Unknown Model' errors from upstream APIs.
This fix skips the ResolveOriginalModel override when a model alias has already
been resolved, ensuring the correct model name is sent to the upstream provider.
Co-authored-by: Amp <amp@ampcode.com>
- Write each SSE chunk directly to c.Writer and flush immediately
- Remove buffered writer and ticker-based flushing that caused delayed output
- Add 500ms timeout case for consistency with OpenAI/Gemini handlers
- Clean up unused bufio import
This fixes the 'not streaming' issue where small responses were held
in the buffer until timeout/threshold was reached.
Amp-Thread-ID: https://ampcode.com/threads/T-019b1186-164e-740c-96ab-856f64ee6bee
Co-authored-by: Amp <amp@ampcode.com>
Add package and constructor documentation for AI Studio, Antigravity,
Gemini CLI, Gemini API, and Vertex executors to describe their roles and
inputs.
Introduce a shared stream scanner buffer constant in the Gemini API
executor and reuse it in Gemini CLI and Vertex streaming code so stream
handling uses a consistent configuration.
Update Refresh implementations for AI Studio, Gemini CLI, Gemini API
(API key), and Vertex executors to short‑circuit and simply return the
incoming auth object, while keeping Antigravity token renewal as the
only executor that performs OAuth refresh.
Remove OAuth2-based token refresh logic and related dependencies from
the Gemini API executor, since it now operates strictly with API key
credentials.
Align thinking suffix handling on a single bracket-style marker.
NormalizeThinkingModel strips a terminal `[value]` segment from
model identifiers and turns it into either a thinking budget (for
numeric values) or a reasoning effort hint (for strings). Emission
of `ThinkingIncludeThoughtsMetadataKey` is removed.
Executor helpers and the example config are updated so their
comments reference the new `[value]` suffix format instead of the
legacy dash variants.
BREAKING CHANGE: dash-based thinking suffixes (`-thinking`,
`-thinking-N`, `-reasoning`, `-nothinking`) are no longer parsed
for thinking metadata; only `[value]` annotations are recognized.
NormalizeThinkingModel now checks ModelSupportsThinking before removing
"-thinking" or "-thinking-<ver>", avoiding accidental parsing of model
names where the suffix is part of the official id (e.g., kimi-k2-thinking,
qwen3-235b-a22b-thinking-2507).
The registry adds ThinkingSupport metadata for several models and
propagates it via ModelInfo (e.g., kimi-k2-thinking, deepseek-r1,
qwen3-235b-a22b-thinking-2507, minimax-m2), enabling accurate detection
of thinking-capable models and correcting base model inference.
Add fallback parsing for quota reset delay when RetryInfo is not present:
- Try ErrorInfo.metadata.quotaResetDelay (e.g., "373.801628ms")
- Parse from error.message "Your quota will reset after Xs."
This ensures proper cooldown timing for rate-limited requests.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
## Changes Overview
This commit includes multiple improvements to the Kiro executor for better
stability, API compatibility, and code quality.
## Detailed Changes
### 1. Output Token Calculation Improvement (lines 317-330)
- Replace simple len(content)/4 estimation with tiktoken-based calculation
- Add fallback to character count estimation if tiktoken fails
- Improves token counting accuracy for usage tracking
### 2. Stream Handler Panic Recovery (lines 528-533)
- Add defer/recover block in streamToChannel goroutine
- Prevents single request crashes from affecting the entire service
### 3. Struct Field Reordering (lines 670-673)
- Reorder kiroToolResult struct fields: Content, Status, ToolUseID
- Ensures consistency with API expectations
### 4. Message Merging Function (lines 778-780, 2356-2483)
- Add mergeAdjacentMessages() to combine consecutive messages with same role
- Add helper functions: mergeMessageContent(), blockToMap(), createMergedMessage()
- Required by Kiro API which doesn't allow adjacent messages from same role
### 5. Empty Content Handling (lines 791-800)
- Add default content for empty history messages
- User messages with tool results: "Tool results provided."
- User messages without tool results: "Continue"
### 6. Assistant Last Message Handling (lines 811-830)
- Detect when last message is from assistant
- Create synthetic "Continue" user message to satisfy Kiro API requirements
- Kiro API requires currentMessage to be userInputMessage type
### 7. Duplicate Content Event Detection (lines 1650-1660)
- Track lastContentEvent to detect duplicate streaming events
- Skip redundant events to prevent duplicate content in responses
- Based on AIClient-2-API implementation for Kiro
### 8. Streaming Token Calculation Enhancement (lines 1785-1817)
- Add accumulatedContent buffer for streaming token calculation
- Use tiktoken for accurate output token counting during streaming
- Add fallback to character count estimation with proper logging
### 9. JSON Repair Enhancement (lines 2665-2818)
- Implement conservative JSON repair strategy
- First try to parse JSON directly - if valid, return unchanged
- Add bracket balancing detection and repair
- Only repair when necessary to avoid corrupting valid JSON
- Validate repaired JSON before returning
### 10. HELIOS_CHK Filtering Removal (lines 2500-2504, 3004-3039)
- Remove filterHeliosDebugInfo function
- Remove heliosDebugPattern regex
- HELIOS_CHK fields now handled by client-side processing
### 11. Comment Translation
- Translate Chinese comments to English for code consistency
- Affected areas: token calculation, buffer handling, message processing
- Added support for parsing and normalizing dynamic thinking model suffixes.
- Centralized budget resolution across executors and payload helpers.
- Retired legacy Gemini-specific thinking handlers in favor of unified logic.
- Updated executors to use metadata-based thinking configuration.
- Added `ResolveOriginalModel` utility for resolving normalized upstream models using request metadata.
- Updated executors (Gemini, Codex, iFlow, OpenAI, Qwen) to incorporate upstream model resolution and substitute model values in payloads and request URLs.
- Ensured fallbacks handle cases with missing or malformed metadata to derive models robustly.
- Refactored upstream model resolution to dynamically incorporate metadata for selecting and normalizing models.
- Improved handling of thinking configurations and model overrides in executors.
- Removed hardcoded thinking model entries and migrated logic to metadata-based resolution.
- Updated payload mutations to always include the resolved model.
* feat(api): add comprehensive ampcode management endpoints
Add new REST API endpoints under /v0/management/ampcode for managing
ampcode configuration including upstream URL, API key, localhost
restriction, model mappings, and force model mappings settings.
- Move force-model-mappings from config_basic to config_lists
- Add GET/PUT/PATCH/DELETE endpoints for all ampcode settings
- Support model mapping CRUD with upsert (PATCH) capability
- Add comprehensive test coverage for all ampcode endpoints
* refactor(api): simplify request body parsing in ampcode handlers
* feat(logging): add upstream API request/response capture to streaming logs
* style(logging): remove redundant separator line from response section
* feat(antigravity): enforce thinking budget limits for Claude models
* refactor(logging): remove unused variable in `ensureAttempt` and redundant function call
---------
Co-authored-by: hkfires <10558748+hkfires@users.noreply.github.com>
Add new REST API endpoints under /v0/management/ampcode for managing
ampcode configuration including upstream URL, API key, localhost
restriction, model mappings, and force model mappings settings.
- Move force-model-mappings from config_basic to config_lists
- Add GET/PUT/PATCH/DELETE endpoints for all ampcode settings
- Support model mapping CRUD with upsert (PATCH) capability
- Add comprehensive test coverage for all ampcode endpoints
Add a configuration option to control whether model mappings take
precedence over local API keys for Amp CLI requests.
- Add PrioritizeModelMappings field to AmpCode config struct
- When false (default): Local API keys take precedence (original behavior)
- When true: Model mappings take precedence over local API keys
- Add management API endpoints GET/PUT /prioritize-model-mappings
This allows users who want mapping priority to enable it explicitly
while preserving backward compatibility.
Config example:
ampcode:
model-mappings:
- from: claude-opus-4-5-20251101
to: gemini-claude-opus-4-5-thinking
prioritize-model-mappings: true
Skip text content blocks that are empty or contain only whitespace
when translating Claude messages to OpenAI format. This fixes GLM-4.6
and other strict OpenAI-compatible providers that reject empty text
with error 'text cannot be empty'.
Simplifies routing logic by delegating all provider/mapping/proxy
decisions to FallbackHandler. Previously, the route checked for
provider/mapping availability before calling the handler, then
FallbackHandler performed the same checks again.
Changes:
- Remove model extraction and provider checking from route (lines 182-201)
- Route now only checks if request is POST with /models/ path
- FallbackHandler handles provider -> mapping -> proxy fallback
- Remove unused internal/util import
Benefits:
- Eliminates duplicate checks (addresses PR review feedback #2)
- Centralizes all provider/mapping logic in FallbackHandler
- Reduces routing code by ~20 lines
- Aligns with how other /api/provider routes work
Performance: No impact (checks still happen once in FallbackHandler)
- Change createGeminiBridgeHandler to accept gin.HandlerFunc instead of *gemini.GeminiAPIHandler
This allows tests to inject mock handlers instead of duplicating bridge logic
- Replace magic number 8 with len(modelsPrefix) for better maintainability
- Remove redundant test case that doesn't test edge case in production
- Update routes.go to pass geminiHandlers.GeminiHandler directly
Addresses PR review feedback on test architecture and code clarity.
Amp-Thread-ID: https://ampcode.com/threads/T-1ae2c691-e434-4b99-a49a-10cabd3544db
Gemini handler extracts model from URL path, not JSON body, so
rewriting the request body alone wasn't sufficient for model mapping.
- Add MappedModelContextKey constant for context passing
- Update routes.go to use NewFallbackHandlerWithMapper
- Add check for valid mapping before routing to local handler
- Add tests for gemini bridge model mapping
2025-12-06 18:59:44 +07:00
501 changed files with 104238 additions and 22929 deletions
@@ -8,16 +8,11 @@ All third-party provider support is maintained by community contributors; CLIPro
The Plus release stays in lockstep with the mainline features.
## Differences from the Mainline
- Added GitHub Copilot support (OAuth login), provided by [em4go](https://github.com/em4go/CLIProxyAPI/tree/feature/github-copilot-auth)
- Added Kiro (AWS CodeWhisperer) support (OAuth login), provided by [fuko2935](https://github.com/fuko2935/CLIProxyAPI/tree/feature/kiro-integration)
## Contributing
This project only accepts pull requests that relate to third-party provider support. Any pull requests unrelated to third-party provider support will be rejected.
If you need to submit any non-third-party provider changes, please open them against the mainline repository.
If you need to submit any non-third-party provider changes, please open them against the [mainline](https://github.com/router-for-me/CLIProxyAPI) repository.
### [All API Hub](https://github.com/qixing-jk/all-api-hub)
New API互換リレーサイトアカウントをワンストップで管理するブラウザ拡張機能。残高と使用量のダッシュボード、自動チェックイン、一般的なアプリへのワンクリックキーエクスポート、ページ内API可用性テスト、チャネル/モデルの同期とリダイレクト機能を搭載。Management APIを通じてCLIProxyAPIと統合し、ワンクリックでプロバイダーのインポートと設定同期が可能
This guide explains how to use CLIProxyAPI with Amp CLI and Amp IDE extensions, enabling you to use your existing Google/ChatGPT/Claude subscriptions (via OAuth) with Amp's CLI.
## Table of Contents
- [Overview](#overview)
- [Which Providers Should You Authenticate?](#which-providers-should-you-authenticate)
The Amp CLI integration adds specialized routing to support Amp's API patterns while maintaining full compatibility with all existing CLIProxyAPI features. This allows you to use both traditional CLIProxyAPI features and Amp CLI with the same proxy server.
- **Security-first**: Management routes restricted to localhost by default
- **Automatic gzip handling**: Decompresses responses from Amp upstream
### What You Can Do
- Use Amp CLI with your Google account (Gemini 3 Pro Preview, Gemini 2.5 Pro, Gemini 2.5 Flash)
- Use Amp CLI with your ChatGPT Plus/Pro subscription (GPT-5, GPT-5 Codex models)
- Use Amp CLI with your Claude Pro/Max subscription (Claude Sonnet 4.5, Opus 4.1)
- Use Amp IDE extensions (VS Code, Cursor, Windsurf, etc.) with the same proxy
- Run multiple CLI tools (Factory + Amp) through one proxy server
- Route unconfigured models automatically through ampcode.com
### Which Providers Should You Authenticate?
**Important**: The providers you need to authenticate depend on which models and features your installed version of Amp currently uses. Amp employs different providers for various agent modes and specialized subagents:
- Requests use **your OAuth subscription** (ChatGPT Plus/Pro, Claude Pro/Max, Google account)
- You benefit from your subscription's included usage quotas
- No Amp credits consumed
2.**Provider NOT authenticated locally**:
- Requests automatically forward to **ampcode.com**
- Uses Amp's backend provider connections
- **Requires Amp credits** if the provider is paid (OpenAI, Anthropic paid tiers)
- May result in errors if Amp credit balance is insufficient
**Recommendation**: Authenticate all providers you have subscriptions for to maximize value and minimize Amp credit usage. If you don't have subscriptions to all providers Amp uses, ensure you have sufficient Amp credits available for fallback requests.
## Architecture
### Request Flow
```
Amp CLI/IDE
↓
├─ Provider API requests (/api/provider/{provider}/v1/...)
│ ↓
│ ├─ Model configured locally?
│ │ YES → Use local OAuth tokens (OpenAI/Claude/Gemini handlers)
│ │ NO ↓
│ │ ├─ Model mapping configured?
│ │ │ YES → Rewrite model → Use local handler (free)
│ │ │ NO → Forward to ampcode.com (uses Amp credits)
The Amp integration is implemented as a modular routing module (`internal/api/modules/amp/`) with these components:
1.**Route Aliases** (`routes.go`): Maps Amp-style paths to standard handlers
2.**Reverse Proxy** (`proxy.go`): Forwards management requests to ampcode.com
3.**Fallback Handler** (`fallback_handlers.go`): Routes unconfigured models to ampcode.com
4.**Secret Management** (`secret.go`): Multi-source API key resolution with caching
5.**Main Module** (`amp.go`): Orchestrates registration and configuration
## Configuration
### Basic Configuration
Add these fields to your `config.yaml`:
```yaml
# Amp upstream control plane (required for management routes)
amp-upstream-url:"https://ampcode.com"
# Optional: Override API key (otherwise uses env or file)
# amp-upstream-api-key: "your-amp-api-key"
# Security: restrict management routes to localhost (recommended)
amp-restrict-management-to-localhost:true
```
### Model Mapping Configuration
When Amp CLI requests a model that you don't have access to, you can configure mappings to route those requests to alternative models that you DO have available. This avoids consuming Amp credits for models you could handle locally.
```yaml
# Route unavailable models to alternatives
amp-model-mappings:
# Example: Route Claude Opus 4.5 requests to Claude Sonnet 4
- from:"claude-opus-4.5"
to:"claude-sonnet-4"
# Example: Route GPT-5 requests to Gemini 2.5 Pro
- from:"gpt-5"
to:"gemini-2.5-pro"
# Example: Map older model names to newer versions
- from:"claude-3-opus-20240229"
to:"claude-3-5-sonnet-20241022"
```
**How it works:**
1. Amp CLI requests a model (e.g., `claude-opus-4.5`)
2. CLIProxyAPI checks if a local provider is available for that model
3. If not available, it checks the model mappings
4. If a mapping exists, the request is rewritten to use the target model
5. The request is then handled locally (free, using your OAuth subscription)
**Benefits:**
- **Save Amp credits**: Use your local subscriptions instead of forwarding to ampcode.com
- **Hot-reload**: Mappings can be updated without restarting the proxy
- **Structured logging**: Clear logs show when mappings are applied
**Routing Decision Logs:**
The proxy logs each routing decision with structured fields:
```
[AMP] Using local provider for model: gemini-2.5-pro # Local provider (free)
[AMP] Model mapped: claude-opus-4.5 -> claude-sonnet-4 # Mapping applied (free)
When enabled, management routes (`/api/auth`, `/api/user`, `/api/threads`, etc.) only accept connections from localhost (127.0.0.1, ::1). This prevents:
This restriction uses the **actual TCP connection address** (`RemoteAddr`), not HTTP headers like `X-Forwarded-For`. This prevents header spoofing attacks but has important implications:
- ✅ **Works for direct connections**: Running CLIProxyAPI directly on your machine or server
- ⚠️ **May not work behind reverse proxies**: If deploying behind nginx, Cloudflare, or other proxies, the connection will appear to come from the proxy's IP, not localhost
#### Reverse Proxy Deployments
If you need to run CLIProxyAPI behind a reverse proxy (nginx, Caddy, Cloudflare Tunnel, etc.):
1.**Disable the localhost restriction**:
```yaml
amp-restrict-management-to-localhost: false
```
2. **Use alternative security measures**:
- Firewall rules restricting access to management routes
- ✅ Don't expose proxy publicly (bind to localhost or use firewall/VPN)
- ✅ Use the Amp secrets file (`~/.local/share/amp/secrets.json`) managed by `amp login`
- ✅ Rotate OAuth tokens periodically by re-running login commands
- ✅ Store config and auth-dir on encrypted disk if handling sensitive data
- ✅ Keep proxy binary up to date for security fixes
## Additional Resources
- [CLIProxyAPI Main Documentation](https://help.router-for.me/)
- [Amp CLI Official Manual](https://ampcode.com/manual)
- [Management API Reference](https://help.router-for.me/management/api)
- [SDK Documentation](sdk-usage.md)
## Disclaimer
This integration is for personal/educational use. Using reverse proxies or alternate API bases may violate provider Terms of Service. You are solely responsible for how you use this software. Accounts may be rate-limited, locked, or banned. No warranties. Use at your own risk.
CLIProxyAPI can now use GitLab Duo as a first-class provider instead of treating it as a plain text wrapper.
It supports:
- OAuth login
- personal access token login
- automatic refresh of GitLab `direct_access` metadata
- dynamic model discovery from GitLab metadata
- native GitLab AI gateway routing for Anthropic and OpenAI/Codex managed models
- Claude-compatible and OpenAI-compatible downstream APIs
## What this means
If GitLab Duo returns an Anthropic-managed model, CLIProxyAPI routes requests through the GitLab AI gateway Anthropic proxy and uses the existing Claude executor path.
If GitLab Duo returns an OpenAI-managed model, CLIProxyAPI routes requests through the GitLab AI gateway OpenAI proxy and uses the existing Codex/OpenAI executor path.
That gives GitLab Duo much closer runtime behavior to the built-in `codex` provider:
- Claude-compatible clients can use GitLab Duo models through `/v1/messages`
- OpenAI-compatible clients can use GitLab Duo models through `/v1/chat/completions`
- OpenAI Responses clients can use GitLab Duo models through `/v1/responses`
The model list is not hardcoded. CLIProxyAPI reads the current model metadata from GitLab `direct_access` and registers:
- a stable alias: `gitlab-duo`
- any discovered managed model names, such as `claude-sonnet-4-5` or `gpt-5-codex`
## Login
OAuth login:
```bash
./CLIProxyAPI -gitlab-login
```
PAT login:
```bash
./CLIProxyAPI -gitlab-token-login
```
You can also provide inputs through environment variables:
- PAT login requires a personal access token that can call the GitLab APIs used by Duo. In practice, `api` scope is the safe baseline.
- Self-managed GitLab instances are supported through `GITLAB_BASE_URL`.
## Using the models
After login, start CLIProxyAPI normally and point your client at the local proxy.
You can select:
-`gitlab-duo` to use the current Duo-managed model for that account
- the discovered provider model name if you want to pin it explicitly
Examples:
```bash
curl http://127.0.0.1:8080/v1/models
```
```bash
curl http://127.0.0.1:8080/v1/chat/completions \
-H 'Content-Type: application/json'\
-d '{
"model": "gitlab-duo",
"messages": [
{"role": "user", "content": "Write a Go HTTP middleware for request IDs."}
]
}'
```
If the GitLab account is currently mapped to an Anthropic model, Claude-compatible clients can use the same account through the Claude handler path. If the account is currently mapped to an OpenAI/Codex model, OpenAI-compatible clients can use `/v1/chat/completions` or `/v1/responses`.
## How model freshness works
CLIProxyAPI does not ship a fixed GitLab Duo model catalog.
Instead, it refreshes GitLab `direct_access` metadata and uses the returned `model_details` and any discovered model list entries to keep the local registry aligned with the current GitLab-managed model assignment.
This matches GitLab's current public contract better than hardcoding model names.
## Current scope
The GitLab Duo provider now has:
- OAuth and PAT auth flows
- runtime refresh of Duo gateway credentials
- native Anthropic gateway routing
- native OpenAI/Codex gateway routing
- handler-level smoke tests for Claude-compatible and OpenAI-compatible paths
Still out of scope today:
- websocket or session-specific parity beyond the current HTTP APIs
- GitLab-specific IDE features that are not exposed through the public gateway contract
*`SetProviders` replaces the provider slice using a defensive copy.
*`Providers` retrieves a snapshot that can be iterated safely from other goroutines.
*`BuildProviders` translates `config.Config` access declarations into runnable providers. When the config omits explicit providers but defines inline API keys, the helper auto-installs the built-in `config-api-key` provider.
If the manager itself is `nil` or no providers are configured, the call returns `nil, nil`, allowing callers to treat access control as disabled.
## Authenticating Requests
```go
result,err:=manager.Authenticate(ctx,req)
result,authErr:=manager.Authenticate(ctx,req)
switch{
caseerr==nil:
caseauthErr==nil:
// Authentication succeeded; result describes the provider and principal.
// Supplied credentials were present but rejected.
default:
// Transport-level failure was returned by a provider.
// Internal/transport failure was returned by a provider.
}
```
`Manager.Authenticate` walks the configured providers in order. It returns on the first success, skips providers that surface `ErrNotHandled`, and tracks whether any provider reported `ErrNoCredentials`or`ErrInvalidCredential` for downstream error reporting.
If the manager itself is `nil` or no providers are registered, the call returns `nil, nil`, allowing callers to treat access control as disabled without branching on errors.
`Manager.Authenticate` walks the configured providers in order. It returns on the first success, skips providers that return `AuthErrorCodeNotHandled`, and aggregates `AuthErrorCodeNoCredentials`/`AuthErrorCodeInvalidCredential` for a final result.
Each `Result` includes the provider identifier, the resolved principal, and optional metadata (for example, which header carried the credential).
## Configuration Layout
## Built-in `config-api-key` Provider
The manager expects access providers under the `auth.providers` key inside `config.yaml`:
The proxy includes one built-in access provider:
-`config-api-key`: Validates API keys declared under top-level `api-keys`.
- Metadata: `Result.Metadata["source"]` is set to the matched source label.
In the CLI server and `sdk/cliproxy`, this provider is registered automatically based on the loaded configuration.
```yaml
auth:
providers:
- name:inline-api
type:config-api-key
api-keys:
- sk-test-123
- sk-prod-456
api-keys:
- sk-test-123
- sk-prod-456
```
Fields map directly to `config.AccessProvider`: `name` labels the provider, `type` selects the registered factory, `sdk` can name an external module, `api-keys` seeds inline credentials, and `config` passes provider-specific options.
## Loading Providers from External Go Modules
### Loading providers from external SDK modules
To consume a provider shipped in another Go module, point the `sdk` field at the module path and import it for its registration side effect:
To consume a provider shipped in another Go module, import it for its registration side effect:
```go
import(
@@ -89,19 +80,11 @@ import (
)
```
The blank identifier import ensures `init` runs so `sdkaccess.RegisterProvider` executes before `BuildProviders`is called.
## Built-in Providers
The SDK ships with one provider out of the box:
-`config-api-key`: Validates API keys declared inline or under top-level `api-keys`. It accepts the key from `Authorization: Bearer`, `X-Goog-Api-Key`, `X-Api-Key`, or the `?key=` query string and reports `ErrInvalidCredential` when no match is found.
Additional providers can be delivered by third-party packages. When a provider package is imported, it registers itself with `sdkaccess.RegisterProvider`.
The blank identifier import ensures `init` runs so `sdkaccess.RegisterProvider` executes before you call `RegisteredProviders()`(or before `cliproxy.NewBuilder().Build()`).
### Metadata and auditing
`Result.Metadata` carries provider-specific context. The built-in `config-api-key` provider, for example, stores the credential source (`authorization`, `x-goog-api-key`, `x-api-key`, or`query-key`). Populate this map in custom providers to enrich logs and downstream auditing.
`Result.Metadata` carries provider-specific context. The built-in `config-api-key` provider, for example, stores the credential source (`authorization`, `x-goog-api-key`, `x-api-key`, `query-key`,`query-auth-token`). Populate this map in custom providers to enrich logs and downstream auditing.
A provider must implement `Identifier()` and `Authenticate()`. To expose it to configuration, call `RegisterProvider` inside `init`. Provider factories receive the specific `AccessProvider` block plus the full root configuration for contextual needs.
A provider must implement `Identifier()` and `Authenticate()`. To make it available to the access manager, call `RegisterProvider` inside `init` with an initialized provider instance.
## Error Semantics
-`ErrNoCredentials`: no credentials were present or recognized by any provider.
-`ErrInvalidCredential`: at least one provider processed the credentials but rejected them.
-`ErrNotHandled`: instructs the manager to fall through to the next provider without affecting aggregate error reporting.
-`NewNoCredentialsError()` (`AuthErrorCodeNoCredentials`): no credentials were present or recognized. (HTTP 401)
-`NewInvalidCredentialError()` (`AuthErrorCodeInvalidCredential`): credentials were present but rejected. (HTTP 401)
-`NewNotHandledError()` (`AuthErrorCodeNotHandled`): fall through to the next provider.
Return custom errors to surface transport failures; they propagate immediately to the caller instead of being masked.
Errors propagate immediately to the caller unless they are classified as `not_handled` / `no_credentials` / `invalid_credential` and can be aggregated by the manager.
## Integration with cliproxy Service
`sdk/cliproxy` wires `@sdk/access` automatically when you build a CLI service via `cliproxy.NewBuilder`. Supplying a preconfigured manager allows you to extend or override the default providers:
`sdk/cliproxy` wires `@sdk/access` automatically when you build a CLI service via `cliproxy.NewBuilder`. Supplying a manager lets you reuse the same instance in your host process:
```go
coreCfg,_:=config.LoadConfig("config.yaml")
providers,_:=sdkaccess.BuildProviders(coreCfg)
manager:=sdkaccess.NewManager()
manager.SetProviders(providers)
accessManager:=sdkaccess.NewManager()
svc,_:=cliproxy.NewBuilder().
WithConfig(coreCfg).
WithAccessManager(manager).
WithConfigPath("config.yaml").
WithRequestAccessManager(accessManager).
Build()
```
The service reuses the manager for every inbound request, ensuring consistent authentication across embedded deployments and the canonical CLI binary.
Register any custom providers (typically via blank imports) before calling `Build()` so they are present in the global registry snapshot.
### Hot reloading providers
### Hot reloading
When configuration changes, rebuild providers and swap them into the manager:
When configuration changes, refresh any config-backed providers and then reset the manager's provider chain:
This mirrors the behaviour in `cliproxy.Service.refreshAccessProviders` and `api.Server.applyAccessConfig`, enabling runtime updates without restarting the process.
This mirrors the behaviour in `internal/access.ApplyAccessProviders`, enabling runtime updates without restarting the process.
Bring GitLab Duo support from the current "auth + basic executor" stage to the same practical level as `codex` inside `CLIProxyAPI`: a user logs in once, points external clients such as Claude Code at `CLIProxyAPI`, selects GitLab Duo-backed models, and gets stable streaming, multi-turn behavior, tool calling compatibility, and predictable model routing without manual provider-specific workarounds.
The core architectural shift is to stop treating GitLab Duo as only two REST wrappers (`/api/v4/chat/completions` and `/api/v4/code_suggestions/completions`) and instead use GitLab's `direct_access` contract as the primary runtime entrypoint wherever possible. Official GitLab docs confirm that `direct_access` returns AI gateway connection details, headers, token, and expiry; that contract is the closest path to codex-like provider behavior.
## Prerequisites
- Official GitLab Duo API references confirmed during implementation:
-`POST /api/v4/code_suggestions/direct_access`
-`POST /api/v4/code_suggestions/completions`
-`POST /api/v4/chat/completions`
- Access to at least one real GitLab Duo account for manual verification.
- One downstream client target for acceptance testing:
- Claude Code against Claude-compatible endpoint
- OpenAI-compatible client against `/v1/chat/completions` and `/v1/responses`
- **Description**: Produce a concrete feature matrix for `codex`: HTTP execute, SSE execute, `/v1/responses`, websocket downstream path, tool calling, request IDs, session close semantics, and model registration behavior.
- **Dependencies**: None
- **Acceptance Criteria**:
- A checklist exists in repo docs or issue notes.
- Each capability is marked `required`, `optional`, or `not possible` for GitLab Duo.
- known limitations if websocket/tool parity is partial
- **Dependencies**: Sprint 5.1
- **Acceptance Criteria**:
- setup instructions are enough for a new user to reproduce the GitLab Duo flow
- limitations are explicit
- **Validation**:
- dry-run docs review from a clean environment
## Testing Strategy
- Keep `go test ./...` green after every committable task.
- Add table-driven tests first for request mapping, refresh behavior, and dynamic model registration.
- Add transport tests with `httptest.Server` for:
- real chunked streaming
- header propagation from `direct_access`
- upstream fallback rules
- Add at least one manual acceptance checklist:
- login via OAuth
- login via PAT
- list models
- run one streaming prompt via OpenAI route
- run one prompt from the target downstream client
## Potential Risks & Gotchas
- GitLab public docs expose `direct_access`, but do not fully document every possible AI gateway path. We should isolate any empirically discovered gateway assumptions behind one transport layer and feature flags.
-`chat/completions` availability differs by GitLab offering and version. The executor must not assume it always exists.
- Code Suggestions is completion-oriented; lossy mapping from rich chat/tool payloads will make GitLab Duo feel worse than codex unless explicitly handled.
- Synthetic streaming is not good enough for codex parity and will cause regressions in interactive clients.
- Dynamic model discovery can create unstable UX if the stable alias and discovered model IDs are not separated cleanly.
- PAT auth may validate successfully while still lacking effective Duo permissions. Error reporting must surface this explicitly.
## Rollback Plan
- Keep the current basic GitLab executor behind a fallback mode until the new transport path is stable.
- If parity work destabilizes existing providers, revert only GitLab-specific executor changes and leave auth support intact.
- Preserve the stable `gitlab-duo` alias so rollback does not break client configuration.
fields["model_id"]=requestedModel// Explicit model_id for easy config reference
log.WithFields(fields).Warnf("forwarding to ampcode.com (uses amp credits) - model_id: %s | To use local proxy, add to config: amp-model-mappings: [{from: \"%s\", to: \"<your-local-model>\"}]",requestedModel,requestedModel)
log.WithFields(fields).Warnf("forwarding to ampcode.com (uses amp credits) - model_id: %s | To use local provider, add to config: ampcode.model-mappings: [{from: \"%s\", to: \"<your-local-model>\"}]",requestedModel,requestedModel)
Some files were not shown because too many files have changed in this diff
Show More
Reference in New Issue
Block a user
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.