Support grok-superimage and improve image streaming/retry, SSE error handling, proxy and config safety

[lijirou12]

Motivation

• Add first-class support for a server-controlled image model (grok-superimage-1.0) that uses server-side image parameters and behaves differently in streaming chat output.
• Make streaming endpoints more resilient by returning SSE error events instead of transport-level 5xx, and improve retry behavior for transient upstream and image-block scenarios.
• Harden reverse proxy handling for app-chat requests and avoid accidentally overwriting remote config with empty sources during bootstrap.
• Expose new configuration controls in admin UI and defaults for blocked-image recovery and superimage settings.

Description

• Introduced SUPERIMAGE_MODEL_ID = "grok-superimage-1.0", added the model to ModelService, and added server-side superimage config defaults in config.defaults.toml plus corresponding admin UI controls and readme documentation.
• Added _superimage_server_image_config() to enforce server-controlled n/size/response_format for grok-superimage-1.0 and applied it in request validation paths so client image_config is ignored for that model.
• Implemented SSE-safe streaming helpers _safe_sse_stream and _streaming_error_response and used them to ensure streams yield structured event: error payloads for Chat, Image, ImageEdit, and Video streaming flows instead of a raw 5xx transport break. Video/Chat service calls are wrapped to return SSE error responses when streaming.
• Enhanced image generation robustness by adding configurable parallel recovery attempts when final images appear to be blocked (image.blocked_parallel_attempts) and by passing a max_retries hint to the WebSocket reverse; failure after recovery raises an UpstreamException with structured details.
• Adjusted WebSocket imagine reverse logic to detect suspected review/block conditions more reliably and to return an informative error payload when retries exhausted, and refined final-image classification to use byte-size threshold logic only.
• Updated Image WS/stream processors to emit only final images for chat completions (no previews for grok-superimage-1.0), ensure a final empty chat chunk is emitted when nothing else was streamed, and preserve original image filenames for grok-superimage-1.0 URL outputs (no -final suffix).
• Added transient upstream detection (transient_upstream) and made token-retry defaults tolerant (get_config("retry.max_retry") or 3), and added retry-on-transient behavior in ChatService to rotate tokens when upstream errors are transient and alternative tokens exist.
• Normalized and improved proxy handling for app-chat reverse requests by converting SOCKS scheme variants and supplying proxy vs proxies to curl_cffi appropriately; also adjusted timeout selection logic.
• Improved config bootstrap persistence logic to avoid overwriting remote storage with defaults when both remote and local sources are empty and added logging to clarify actions; renamed get_storage import to avoid name clash and renamed the route helper to get_storage_mode.

Testing

• Ran the project's automated test suite with pytest -q and the modified unit/integration tests, and all tests passed.
• Executed automated streaming/image-related integration smoke tests (chat with image generation and image-edit streams) which verified SSE error payloads and the grok-superimage-1.0 behavior, and they passed.

---

Codex Task

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
grok2api	Ready	Preview, Comment	Feb 27, 2026 5:15am

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.