dsh-plugin-reasoning-efforts

dsh's chat UI shows a reasoning-effort selector only when a model advertises the reasoning.efforts capability. Official DeepSeek channels include it out of the box, but third-party relay models declared via llm-pi-ai generally do not, so the selector simply disappears and there is no corresponding field in the Web settings page to work around it.

This plugin watches the llm-pi-ai settings section and, following per-family rules, writes the missing capability declarations through dsh's regular settings service. Changes land transparently in ~/.dsh/settings.yaml. The merge is idempotent: it only fills in absent fields and never overwrites existing ones, so re-running it is safe and no feedback loop can form. After installing and restarting dsh web, every model on your relay routes gains a working effort selector with zero manual configuration, and any model added later is picked up automatically.

By default every model gets six levels (Off / Low / Medium / High / Xhigh / Max) with High as the starting point. DeepSeek and Reasoner families keep their native dialect where Off truly disables thinking via thinking:{type:"disabled"}; all other models use the generic reasoning_effort parameter. Mixed relay routes are safe because the thinking format is written per model, preventing cross-contamination. Effort selection is session-scoped, so each conversation window can use a different level independently.

Built for dsh 0.1.x users who route through third-party relays or aggregators and want the same one-click effort-level experience the official channels provide—without hand-editing YAML for every model in every route.