Add grok-superimage, robust image streaming/retries, SSE error handling, proxy and config safety

[lijirou12]

Motivation

• Improve reliability of image generation when upstream returns previews or is intermittently blocked and provide a server-side grok-superimage-1.0 that uses centralized image parameters.
• Make streaming endpoints return well-formed SSE error payloads instead of transport-level 5xx breaks for better client behavior.
• Harden token retry behavior and transient upstream handling to reduce failed requests due to transient network/upstream issues.
• Add small operational safety and UX improvements such as proxy normalization, config bootstrap protection, and admin UI/config defaults for new features.

Description

• Introduced grok-superimage-1.0 in ModelService and server-side image parameters via _superimage_server_image_config used when that model is requested, and added superimage config defaults in config.defaults.toml and UI fields in app/static/admin/js/config.js and numeric handling.
• Implemented robust image-generation recovery flow: WebSocket client class ImagineWebSocketReverse now classifies preview/medium/final by byte-size thresholds and can perform parallel recovery attempts when blocked, and ImageGenerationService will run extra parallel recovery attempts using distinct tokens and fail with a clear UpstreamException if no final images are produced.
• Modified streaming behavior across chat/video/image paths to wrap upstream streams with _safe_sse_stream that encodes exceptions as SSE event: error JSON payloads and added _streaming_error_response to return one-shot SSE errors for streaming requests; also ensured chat-style image streams always emit a final empty chat chunk and data: [DONE].
• Added transient upstream detection in app/services/grok/utils/retry.py and integrated it into ChatService to retry with different tokens on transient errors, and added safer fallbacks when get_config("retry.max_retry") may be unset.
• Hardened config bootstrap in app/core/config.py to avoid overwriting remote storage with defaults when the remote returned None and no local seed existed, and renamed get_storage import to avoid collision in admin API and exposed storage detection.
• Normalized proxy handling in app/services/reverse/app_chat.py including SOCKS scheme normalization and using proxy vs proxies based on scheme, and adjusted timeout selection logic.
• Several image stream/processors updated to only expose final images for chat-format streams, preserve original imagine image names for grok-superimage-1.0 URL outputs, and emit correct SSE finalization; introduced logging changes and limits for parallel attempts and retries.
• Documentation update in readme.md listing grok-superimage-1.0 and behavior notes for streaming and blocked-image recovery.

Testing

• Ran the project's unit and integration test suite and basic API smoke tests locally including POST /v1/chat/completions flows for text, image, and streaming scenarios, and observed no regressions.
• Exercised image generation paths with simulated WS responses to validate final/preview classification, parallel recovery attempts, and SSE error emission, and confirmed expected SSE error events and data: [DONE] termination for streaming clients.
• Verified admin UI renders new superimage fields and blocked_parallel_attempts numeric input in the configuration page and that default config values load from config.defaults.toml.

---

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 6:09am

[chatgpt-codex-connector]

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