Add grok-superimage support, robust SSE streaming, image recovery and proxy/ retry fixes

[lijirou12]

Motivation

• Introduce a server-controlled "superimage" model and improve image generation reliability when upstreams are blocked or return only previews.
• Make streaming endpoints more robust by wrapping streams to emit SSE-formatted error events instead of transport-level 5xx breaks.
• Improve token retry behavior by treating certain upstream errors as transient and falling back to alternate tokens, and defaulting retry counts when config is missing.
• Fix proxy handling for app-chat reverse requests and avoid HTTP CONNECT mishandling for SOCKS proxies.

Description

• Added a new model entry grok-superimage-1.0 and server-side [superimage] config options (n, size, response_format) and wired them into the Chat API to force server-controlled image parameters for that model (app/services/grok/services/model.py, app/api/v1/chat.py, config.defaults.toml, readme.md).
• Implemented SSE-safe streaming helpers to catch exceptions and return SSE error events (_safe_sse_stream, _streaming_error_response) and wrapped all streaming responses with them so clients receive structured SSE errors (app/api/v1/chat.py).
• Added logic to use server-side superimage parameters when model == grok-superimage-1.0 and ensured streaming for that model only emits final images and preserves original image names for URL outputs (app/services/grok/services/image.py, app/services/grok/services/image_edit.py).
• Added blocked-image recovery: run configurable parallel retry attempts when insufficient finals are produced and fail with a clear UpstreamException if still insufficient (image.blocked_parallel_attempts) (app/services/grok/services/image.py, config.defaults.toml, admin UI updates).
• Tightened final-image classification to require a byte-size threshold and adjusted WS imagine logic to detect suspected blocking/review and raise a blocked error to trigger recovery attempts (app/services/reverse/ws_imagine.py).
• Improved image stream processors to only expose final images for chat completion streams, ensure a final chat completion chunk and always emit the terminal [DONE] marker for chat-formatted streams (app/services/grok/services/image.py, app/services/grok/services/image_edit.py).
• Added transient upstream detection helper and integrated it into token-retry logic to try alternate tokens on transient errors, and defaulted retry.max_retry fallbacks to 3 when unset (app/services/grok/utils/retry.py, app/services/grok/services/chat.py, app/services/grok/services/image.py, app/services/grok/services/image_edit.py).
• Normalized and improved proxy handling for the app-chat reverse path, selecting between proxy and proxies parameters for curl-cffi based on scheme and logging proxy usage, and simplified timeout selection (app/services/reverse/app_chat.py).
• Minor API/admin changes: renamed local import to avoid shadowing (get_storage -> resolve_storage) and exposed storage mode endpoint name change; added admin UI fields and locales for superimage and blocked_parallel_attempts (app/api/v1/admin_api/config.py, app/static/admin/js/config.js).

Testing

• Ran the project's automated test suite with pytest (existing unit/integration tests) and no test failures were observed.
• Performed basic smoke checks of the chat/image generation flows (non-blocking) to validate SSE error formatting and that grok-superimage-1.0 uses server-controlled image parameters during request validation.

---

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 4:47am

[chatgpt-codex-connector]

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