dsh-qwen38-compaction-fix

Running qwen3.8-27b locally on the NInfer engine often ends the Harness compaction with the familiar summarization-truncated-at-the-token-cap verdict. The root cause is straightforward: the model defaults to xhigh reasoning effort and burns its entire output token budget on internal reasoning before writing a single word of the summary. The conversation context that the summary never reaches is effectively lost, and session titles get clipped by the same mechanism.

The plugin takes a deliberately narrow approach. It disables thinking for compaction and title-generation calls only, freeing the full output budget for the actual summary text. It then writes the sampling parameters recommended by the model when thinking is off and sets a configurable max_tokens floor so the budget is never further clamped. Every other conversation turn, subagent call, and non-matching model passes through byte-for-byte unchanged. Model matching is an exact ID check, preventing sampling values from leaking into calls for other models.

This is meant for users who serve qwen3.8-27b through an NInfer gateway (ninfer-serve, OpenAI-compatible API) and have already seen the truncated checkpoint or clipped titles in the Harness. If your stack runs on llama.cpp, vLLM, FastMTP, or another engine, the same idea applies but the wire parameters differ, and this package does not cover those launch methods.