Add grok-superimage support, safer SSE streaming, image recovery & proxy/ retry improvements

[lijirou12]

Motivation

• Support a new server-controlled image model grok-superimage-1.0 with unified server-side image parameters and safer streaming semantics for chat image outputs.
• Make streaming endpoints resilient to upstream errors by emitting SSE error events instead of causing transport 5xx breaks.
• Improve image generation robustness for cases where upstream may block or review images by adding recovery attempts and better final-image classification.
• Improve retry logic and reverse proxy handling to better handle transient upstream failures and SOCKS proxy behavior.

Description

• Added grok-superimage-1.0 ModelInfo and a server-side config section [superimage] in config.defaults.toml and admin UI support (app/static/admin/js/config.js) to control n, size, and response_format, and added SUPERIMAGE_MODEL_ID and _superimage_server_image_config() in app/api/v1/chat.py so the server overrides client image_config for this model.
• Implemented safe SSE streaming helpers in app/api/v1/chat.py: _safe_sse_stream() converts exceptions into SSE error events and yields a final data: [DONE], and _streaming_error_response() to return a one-shot SSE error response for sync exceptions during stream setup; integrated these into image generation/edit and general chat/video streaming return paths.
• Image generation improvements in app/services/grok/services/image.py: added configurable parallel recovery attempts controlled by image.blocked_parallel_attempts, run extra parallel _fetch_batch attempts when finals are insufficient, and raise UpstreamException when still insufficient after recovery.
• Image stream processors changes in image.py and image_edit.py: only expose final images for chat-format streams, preserve original imagine image name for grok-superimage-1.0 (no -final suffix), ensure a final empty chat chunk and data: [DONE] are emitted if no intermediate chat chunk was emitted, and avoid exposing preview partials for chat streams.
• WebSocket imagine reverse changes in app/services/reverse/ws_imagine.py: classify final images by byte-size threshold (final_min_bytes) rather than filename extension, log and raise a _BlockedError on suspected blocked/review conditions to trigger recovery behavior.
• Retry and token picking logic in app/services/grok/utils/retry.py: added transient_upstream() that detects transient upstream errors/timeouts; used in app/services/grok/services/chat.py to attempt switching tokens on transient upstream errors when alternatives exist.
• Reverse app-chat changes in app/services/reverse/app_chat.py: normalize proxy URLs (convert socks5:// -> socks5h://, socks4:// -> socks4a://), prefer proxy param for SOCKS proxies while using proxies for HTTP/HTTPS, and adjust timeout selection logic.
• Minor API/admin changes: renamed import get_storage -> resolve_storage and router function get_storage -> get_storage_mode to avoid name collision in app/api/v1/admin_api/config.py.
• Documentation updates in readme.md to list the new model and behavior notes, and added new default config keys in config.defaults.toml and localized admin UI strings in app/static/admin/js/config.js.

Testing

• Ran the test suite with pytest -q and all tests passed.
• Ran static checks with flake8 and no new lint errors were reported.
• Performed local manual smoke checks of POST /v1/chat/completions for an image model stream and confirmed SSE error payloads and grok-superimage-1.0 behavior (server-side n/size/response_format) were emitted as expected.

---

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 4:30am

[chatgpt-codex-connector]

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