dsh-finish-reason-patch

When you route GPT-family models, gateways, or OpenAI-compatible proxies through DSH, some providers deliver the full streaming output yet never emit a terminal finish_reason. The underlying pi-ai adapter in DSH interprets this as a transport error and aborts a perfectly complete model turn. dsh-finish-reason-patch applies a precise tolerance patch at exactly that layer.

The plugin hooks the llm/stream waterfall that every model call passes through via LlmRuntime.streamWithRegistration—agent turns, session-title requests, subagent calls, and prepared invocations. When the terminal finish is precisely the missing-finish_reason error, it is rewritten to a normal stop (or to tool-calls if tool-call deltas were already delivered), letting the turn complete gracefully. In edge cases where a gateway streams only usage metadata or nothing at all, the empty response is accepted as a finished turn instead of entering a retry loop. All other error types—transport resets, rate limits, HTTP failures—pass through untouched, so the existing retry policy still applies to genuinely broken responses.

This plugin is for developers who rely on DSH to call GPT-family models or OpenAI-compatible endpoints through gateways and occasionally see turns abort without a clear cause. It has zero external dependencies, composes automatically at boot, and leaves no trace in any shipped DSH file when removed.