dsh-model-fix

The opencode aggregate endpoints for muse-spark-1.2 and gpt-5.6-luna have a subtle bug: the stream delivers all content correctly but never sends a finish_reason or the [DONE] sentinel before closing. DSH's pi-ai SDK strictly requires that terminal marker, so every turn collapses into a Stream ended without finish_reason error (mapped to TRANSPORT). Because TRANSPORT is on the default retry list, the agent-level retry loop replays the whole step up to maxRetries times, silently burning extra tokens. Swapping proxies or re-connecting does not help — the defect lives in the model endpoint itself.

dsh-model-fix patches the llm/stream waterfall with a narrowly scoped fix. It intervenes only when a stream has already produced real content (text-delta, reasoning-delta, or tool-call-delta), the closing error is TRANSPORT or STREAM_CLOSED, and the message explicitly says the terminal event is missing. In that case the defective ending is rewritten to a clean stop. Genuine transport failures such as SocketError remain untouched and still surface as errors; empty zero-content responses are not masked either. Two config fields — modelPattern (a regex over model IDs) and providers (an optional routing-key whitelist) — let you limit the patch to exactly the models and routes you need.

If you route muse-spark or gpt-5.6-luna through opencode-go in DSH and keep hitting content-fine-but-turn-fails errors with wasted retries, this plugin is for you. Install it, restart the target profile, and the spurious TRANSPORT errors disappear. No data-training opt-in is required and no proxy is needed.