- Introduce `sanitizeOpenAIResponsesReasoningEncryptedContentKeepForeign` to retain unknown-format reasoning signatures
- Update Meta executor to preserve foreign reasoning encrypted payloads for third-party models like Muse
Closes: #6450
- Pass configured proxy URL to the npm registry updater client
- Validate fetched versions against strict semver and minimum server version floor (`1.0.13`)
- Update default fallback client version to `1.0.46`
- Manage background updater lifecycle and context via service lifecycle initialization
- Track dropped attachments and reject emptied user turns during OpenAI Responses to Interactions translation
- Register checked request translator for OpenAI Responses to Interactions format conversion
- Fetch latest @xai-official/grok client version from npm registry on startup
- Periodically refresh Grok CLI version every 3 hours without multiple retries
- Update hardcoded fallback version to 1.0.50 if npm registry is unreachable
- Dynamically inject active version into x-grok-client-version and User-Agent headers
- Relocate caller system prompts into conversation messages when using OAuth to prevent triggering Anthropic's third-party classifier
- Apply the same relocation policy to token counting requests
Closes: #6432
An attachment a translator could not represent was dropped without a word, and
a turn whose only content was that attachment went upstream as an empty
messages, parts, or input array. The model then answered as if nothing was
attached, or the upstream rejected a request that never named the attachment.
A base64 document now reaches OpenAI as a file part. A file the target cannot
receive, such as container_upload, an OpenAI file_id, or input_audio sent to
Claude, is reported by name so the route answers 400 instead of forwarding an
empty turn. Text beside it still travels.
One rule, ErrIfNothingLeft, decides when to refuse, and RegisterCheckedRequest
carries the refusal on the request envelope. Every executor that translates
checks it before calling upstream. The exported converters keep their []byte
signature, so no caller or sibling test had to change.
- Reject speech-only models (such as `grok-tts` and `grok-voice-tts-1.0`) on non-speech entry protocols to prevent invalid chat upstream routing
- Mark upstream 404 errors for xAI speech as request-scoped unless indicating model unavailability, preventing unnecessary credential rotation and model cooldowns
- Align xAI speech payload finalizer protocol with `openai` for consistent payload override rules
- Buffer and publish observed upstream usage when apply patch transformation or translation encounters an error in Meta and xAI executors
- Add stream usage buffer helpers to report failure metrics with captured upstream token usage
Closes: #6427
- Update cloaking check to retain caller system instructions at the top level when a conversation begins with multiple consecutive user messages followed by assistant turns
Closes: #6426
- Update terminal user run check in Claude cloaking logic so conversations ending with a single user turn keep caller system instructions at the top level instead of relocating them into mid-conversation messages
Closes: #6423
- Detect terminal stream chunks and avoid recording client cancellations as execution failures once completion events have been delivered
Closes: #6421
- Emit carrier thinking blocks for thought signatures accompanying visible text or tool calls in both streaming and non-streaming responses
- Ensure thinking content and visible text are properly segregated without misclassifying regular output as reasoning
Closes: #6409
- Map Gemini safety, recitation, and malformed function call finish reasons to Claude `refusal` stop reason across streaming and non-streaming responses
- Ensure `content_block` and `message_delta` events are properly emitted on partless stream completions and `[DONE]` signals
Closes: #6408
- Pre-allocate `oauthToolAliases` in Claude and Kimi executor constructors so request copies created via `ForAPIKey` reuse the same store
- Replace struct-level mutex with package-level mutex for lazy store allocation to prevent copying mutexes by value
Closes: #6401
- Retain unknown-format `encrypted_content` during reasoning signature sanitization when compatibility mode is enabled
- Continue stripping empty signatures and known foreign provider signatures in compat mode
Closes: #6399
- Ensure schema nodes defining `properties` explicitly set `type` to `object`
- Prefer `object` type when resolving multi-type unions on nodes declaring `properties`
Closes: #6394
- Support base64 `document` content blocks alongside `image` blocks when converting Claude requests to Gemini inline data
- Avoid creating empty contents turns when all message parts are filtered or unsupported
Closes: #5960
- Record reasoning effort from finalized upstream payload after cache helper processing across execution, streaming, compaction, and image requests
Closes: #6388
- Track pending completion state in OpenAI-to-Responses translator when item output is complete and a finish reason is received
- Synthesize terminal `[DONE]` event on clean EOF when response completion is pending instead of failing the stream
Closes: #6381
- Honor `prompt_cache_options.mode = "explicit"` to suppress automatic `cache_control` insertion while preserving client-defined cache breakpoints
- Strip `prompt_cache_options` from request payloads before forwarding upstream and during token counting
Closes: #6380
Replace Codex-specific response model APIs with shared model observation and remove redundant substitution warning throttling. Consolidate repeated Antigravity context-hash rejection messages into a single request-level log while preserving rejection details across index rebuilds.
- Track terminal chunks including finish reasons, stop messages, and completion events during streaming
- Commit replay settlements and publish observed usage instead of reporting stream failure when the client disconnects after a terminal chunk
Closes: #6376
- Parse and map `url_citation` annotations to OpenAI message and delta annotations for both stream and non-stream responses
- Track emitted text rune offsets to calculate citation start and end indices accurately
- Deduplicate citations across streaming annotation events and message completion updates
Closes: #6361
- Buffer and assemble multi-line or split JSON chunks until a valid JSON payload is formed
- Intercept upstream backend error objects during streaming and surface them as terminal errors with appropriate status codes
Closes: #6357
- Track terminal connection errors on websocket session when upstream read loops fail or receive unexpected binary payloads
- Return recorded terminal errors immediately during session activation to prevent callers from blocking indefinitely after upstream disconnects
- Reset terminal error state when configuring new connections
Closes: #6354
- Add bounded in-memory store to retain Claude OAuth tool alias mappings keyed by message ID
- Restore alias mappings for thread continuation requests omitting tool declarations, returning a request-scoped not found error when state is missing
- Record tool aliases on execution and streaming response completion
Closes: #6346
- Treat Claude missing thread state 404 errors as client errors to bypass credential rotation and cooldown
- Inject `thread_not_found` error code into Claude error response details to signal client-side replay
- Support extracting upstream response bodies across error formatting and classification helpers
Closes: #6344
- Add `claudeMidConversationSystemMessagesAtEnd` to detect consecutive user messages ending the message array
- Append forwarded system blocks to top-level `system` payload instead of inserting mid-conversation message turns when trailing user turns are present
Closes: #6336
- Add `FlushResponse` to prefer error-aware flushing and unwrap wrapped response writers
- Bypass Gin's void `Flush` while maintaining header accounting via `WriteHeaderNow` to keep transport errors observable
Closes: #6332
- Add `CatalogSources` configuration to allow overriding general, Codex, and Devin catalog sources via URLs or absolute local file paths
- Centralize model catalog fetch, validation, and scheduled background updates in `catalogUpdater`
- Support dynamic catalog source updates and hot reloading during service runtime
- Update `--local-model` flag to use embedded models as defaults while respecting explicit overrides
Closes: #6319Closes: #5024
- Strip provider namespaces and prefix before evaluating model family
- Parse major and minor version numbers dynamically while ignoring snapshot dates
- Correctly identify unsupported assistant prefill models for Fable, Opus 5+, and Sonnet 4.6+
Closes: #5483
- Add `PayloadFinalizer` to apply configuration rules at the final serialization barrier
- Support decoding, modifying, and re-encoding Devin protobuf business fields
- Add multipart form-data payload normalization to JSON for media payload rules
Closes: #5067