Patches I sent upstream.

On a shared auth profile, one model hitting a rate limit took every other model down with it. Scoped cooldowns per-model, replaced the 1m→1h exponential escalation with a capped 30s→5m ladder, added an actual error message.

LiveSessionModelSwitchError couldn't tell apart a user typing /model, an automatic 429 fallback, and a heartbeat override. All three looked identical, so legitimate fallbacks got misclassified. Added the disambiguation.

A session row can exist before its transcript has events. A reset at that moment claims sequence zero, no header ever gets written, and the next turn dies with a cryptic migration error. Fix: write the header first.

A transient catalog fingerprint mismatch permanently wedged a worker — it cached one preparation promise for its entire lifetime, so once that promise rejected, every subsequent request replayed the same cached failure.

Reasoning models emit hundreds of empty streaming chunks before real content. The null check passed for empty strings, nothing accumulated, and users watched two minutes of silence. Eight lines changed.

Five MCP tools for GUI automation with zero local GPU — screen parsing and action reasoning run entirely on HuggingFace Inference Endpoints. Configurable via env vars for any OmniParser-compatible parser and OpenAI-compatible vision model. Addressed several open feature requests.

Proposed a fix so opencode can speak Copilot Business and Enterprise auth — bearer exchange, dynamic endpoint resolution, and VS Code identity headers.

Right now --max_train_steps gets silently overridden. Proposed a warning so the user actually knows.