dsh-model-fix

The opencode aggregate endpoints for muse-spark-1.2 and gpt-5.6-luna carry a subtle stream-termination defect: content flows through perfectly, yet neither finish_reason nor the [DONE] sentinel is ever emitted before the stream closes. Because DSH's pi-ai SDK mandates a finish_reason at stream end, every turn surfaces a Stream ended without finish_reason error (mapped to TRANSPORT). That error sits in the default retry list, so the agent-level retry loop re-executes the entire step up to maxRetries times, burning extra tokens on every single turn. No proxy swap or reconnection tweak helps — the root cause is on the model endpoint side, and direct vs. proxied responses are identical.

dsh-model-fix inserts a precise stream-termination patch into DSH's llm/stream waterfall. It fires only when the stream has already produced content (text-delta, reasoning-delta, or tool-call-delta), the closing error code is TRANSPORT or STREAM_CLOSED, and the message explicitly points to a missing terminal event — in that case the defective ending is rewritten to a clean stop. Real transport failures (e.g., SocketError: other side closed) are left untouched and still fail as expected; empty zero-content responses are not swallowed either. Two configuration knobs in cordis.patch.yml — a modelPattern regex and an optional providers whitelist — let you scope the fix to exact model IDs and routing keys so no other models are affected.

This plugin is aimed at developers who route muse-spark or gpt-5.6-luna through opencode-go in DSH and hit the recurring content-fine-but-turn-fails symptom. After installing, just restart the target profile and the spurious TRANSPORT errors and wasted retries go away. No data-training opt-in on the opencode platform is required, and no proxy is needed.