Codex Desktop injects codex_app.automation_update with a large oneOf+$ref
schema. On xAI free/build Responses, that request accepts HTTP but never
emits SSE, so clients hang until cancel.
- Simplify automation_update (and similar large oneOf+$ref schemas) to a
permissive object schema before upstream xAI
- Map subscription free-usage-exhausted 429s to a 24h RetryAfter for
account rotation/cooldown
- Add unit tests for both behaviors
- Added logic to convert OpenAI `max_tokens` and `max_completion_tokens` to `generationConfig.maxOutputTokens`.
- Implemented unit tests to validate proper mapping behavior, including precedence of `max_tokens` over `max_completion_tokens`.
Closes: #4108
- Introduced `AttachCacheControl` and `AttachMessageCacheControl` utilities for consistent cache control injection in content parts and messages.
- Updated Claude translator modules to preserve and apply cache control metadata for responses, messages, tools, and parts.
- Added unit tests to validate proper cache control behavior across multiple scenarios.
Closes: #4146
- Replaced hardcoded model allowlist with dynamic model registry metadata lookup.
- Enhanced `xaiSupportsReasoningEffort` to utilize metadata for reasoning capability validation.
- Updated `sanitizeXAIResponsesBody` to log reasoning effort stripping for unsupported models.
- Added unit tests to verify registry-based reasoning effort handling across models, including Grok 4.5.
Closes: #4147
- Added `tryRefreshAfterUnauthorized` to refresh OAuth credentials on 401 errors during requests.
- Implemented `refreshLocks` to prevent concurrent refreshes for the same auth ID.
- Updated auth/state handling to reset unauthorized model states and resume operations after a successful refresh.
- Enhanced refresh logic with error handling, synchronization, and state updates.
Closes: #4087
- Registered Grok 4.5 with advanced agentic software and workflow task handling.
- Increased context length and max completion tokens for enhanced performance.
- Updated `CredentialFileName` to include trimmed `hashAccountID` for team-scoped plans.
- Added `isTeamScopedPlan` helper to handle plan-specific logic.
- Ensured filenames are unique across emails for multiple teams.
Closes: #4075
- Consolidated thinking content block handling for improved clarity and efficiency.
- Unified `generationConfig` handling, including normalization of `thinkingConfig` settings.
- Added `applyOpenAIThinkingCompatibilityToAntigravity` for consistent reasoning and thought configuration compatibility.
Closes: #4067
- Introduced `StreamUsageBuffer` to encapsulate stream usage observation and publishing logic.
- Updated OpenAI-compatible executors and helpers to use `StreamUsageBuffer` for consistent stream usage tracking.
- Added unit tests to validate `StreamUsageBuffer` behavior, ensuring proper handling of usage details.
Closes: #4053
- Added `mapCodexWebsocketReadError` to handle `CloseMessageTooBig` errors with proper status and error code mapping.
- Updated error propagation and logging in Codex WebSocket executor.
- Introduced corresponding unit test to verify `message_too_big` error mapping in streamed responses.
Closes: #4017
- Added handling of input/output modalities metadata for Codex and compatibility clients.
- Updated model registry and configurations to support "text" and "image" modalities.
- Introduced new tests to validate input/output modalities registration and processing.
Closes: #3976
- Added `applyResponsesToolResultContent` to process and structure tool response content.
- Introduced conversion logic for text, image, and file parts in tool responses.
- Updated tests to validate correct handling of data URL images in tool response outputs.
Closes: #4116
- Verified proper chunking and streaming of SSE events in `ExecuteStream`.
- Refactored executor to ensure complete SSE event delivery without partial fragments.
Closes: #4121
- Added fallback and suspension handling for `invalid_grant` errors in auth conductor.
- Introduced helper methods to detect `invalid_grant` error patterns.
- Updated tests to validate `invalid_grant` fallback and suspension scenarios in execution and stream flows.
Closes: #4120
- Introduced API handlers and executor logic for Google Interactions
- Added request and response transformations for OpenAI and Claude Interactions.
- Integrated Gemini API with Interactions support.
- Updated tests to validate Interactions request parsing and error handling.
- Refactored translator logic for Interactions data flows.
shouldRetryAfterError approves a retry using the pre-jitter wait, so the jitter added in waitForCooldown could push the actual sleep up to 2s past the configured max-retry-interval ceiling. Clamp the jitter range to the remaining headroom below maxWait so the documented maximum stays a hard bound; waits at the ceiling simply sleep unjittered.
Previously every 429 incremented the backoff level unconditionally, so N concurrent in-flight failures from a single quota event inflated the level by N at once (12+ failures jump straight to the 30 minute plateau even when the upstream limit clears in seconds). Failures that land while a previous quota window is still open now reuse that window instead of escalating, so the ladder advances at most once per window. Provider supplied retry hints still take effect unchanged.
waitForCooldown also slept until the exact recovery deadline, waking every waiting request in lockstep against the single recovered credential and re-escalating it in one burst. Cooldown waits now carry a small random jitter (up to wait/4, capped at 2s) to spread the wakeups.
Start the normal server when template api-keys are detected, but keep proxy API endpoints disabled until the keys are updated.
Show the warning page on / and /management.html, add a button to open the management panel via /management.html?safe-mode=configure, and remove the old warning-only server path.
Fixes#4063
Enhance reasoning content retrieval by falling back to 'reasoning' if 'reasoning_content' is absent or empty.
In some selfhosted openai-like server serving deepseek, the response.delta may contains reasoning rather than reasoning_content. So the client like codex does not recognize the thinking process.
For downstream websocket with CPA-mediated HTTP/SSE upstream (non-passthrough),
always merge the full conversation transcript instead of sending incremental
previous_response_id turns. Release pinned websocket auths after timeout and
gateway failures, and after WS bootstrap failover pin the successful SSE
credential so the next turn keeps merged context instead of retrying a
suspended websocket auth.
Fixes#4048
Force-mapping rewrites streaming OpenAI Responses SSE through StreamRewriter before the /v1/responses websocket forwarder. Antigravity/Gemini and Codex emit frames without reliable trailing newlines, so buffered chunks glued as ...}event:..., ...}data:..., or event:/data: without separators. The rewriter dropped pending tails and downstream WS synthesized 408 without response.completed even when upstream returned HTTP 200.
- safeReplaceGlued for }event: and }data: when data JSON is complete\n- Finish flush with glue normalization and line-wise fallback\n- Newline between pending event: and next data: line (Codex scanner lines)\n- Regression tests: Antigravity sim, Codex data lines, force-map WS forward
Use the aidev_client parenthetical suffix for default Antigravity UA and
route loadCodeAssist through the short UA. Keep onboardUser on the long UA
with google-api-nodejs-client.
Recognize model-visible output_text even when message.role is user, accept
assistant string content, and preserve raw merged text without trimming.
Always merge the adjacent visible item into native Gemini reasoning turns so
trailing output is not dropped by prefill stripping.
Build Gemini-native model contents for reasoning items when the routed model
uses the Gemini signature provider: thought summary on a thought part and the
replay signature on the adjacent visible part. Merge following assistant
visible text when it is not trailing prefill, and keep reasoning history when
stripping trailing model-authored prefill.
Extend Responses translator tests for signature compatibility and Antigravity
wrapper coverage; clarify stale replay signature expectations in executor tests.
Wraient