zhubaohi/dsh-qwen38-compaction-fix

Stops a local qwen3.8 gateway on NInfer from returning truncated compaction checkpoints by turning thinking off for those calls.

Fixes one failure precisely. A local qwen3.8-27b on NInfer thinks at xhigh, spends the entire output token budget on reasoning before reaching a conclusion, and the compaction checkpoint comes back truncated — so DSH condenses the conversation into the truncated text and everything the summary never reached is lost from context. The plugin disables thinking for compaction calls only, leaving normal turns untouched, and applies to those calls the sampling parameters the model's own guidance recommends when thinking is off. It has a secondary fix the README documents as a consequence of the same cause: session title generation runs on the same route with a tiny output budget (64 tokens), so xhigh thinking truncates titles too, and the plugin writes the configured reasoning_effort wire value into title request bodies as well. Reasoning effort is configurable, with a preference order of your setting, then off, then low, and a model that offers none of them keeps its own default with a single warning.

Models & Providers ★ 3
View on GitHub ↗

Install

dsh plugin --profile web add dsh-qwen38-ninfer-compaction-fix

The npm line from the README's Install section. Nothing else is required for the common case, but the configuration section matters: the plugin only acts when the outgoing request's model field matches an id in its models allowlist, which defaults to exactly qwen3.8-27b, case sensitive.

Compatibility

Scope is NInfer only, and the README says so in a blockquote: it is written for gateways served by the NInfer engine (ninfer-serve, an OpenAI-compatible API), and the wire fields it rewrites are engine-specific. The match rule is the part that decides whether it does anything at all: it is an exact id comparison, not a family or substring match, because the sampling parameters it writes into the body are specific to that model. If your gateway serves the model under a different id — a custom name, a suffixed variant, different casing — add that id to models. If nothing matches, the plugin silently does nothing: every request passes through byte for byte with no warning on a miss.

Details

Alternatives

gausszhou/dsh-opencode-session-id · mubaid/dsh-opencode-freeaccess · drscrewdriver/dsh-llm-openai-completions

More plugins in Models & Providers

Browse more in Models & Providers

Guides for Models & Providers plugins