dsh-adaptive-reasoning

Every request to a model starts with a silent decision: how much reasoning should this step get? A greeting, a translation, a long-form summary—each one arguably needs a different level, yet manually toggling between off, low, medium, and high breaks flow and is easily misjudged under pressure. dsh-adaptive-reasoning removes that manual step entirely. Using lightweight local rules, it recognises the task family on each turn and assigns the matching reasoning level and tool policy automatically, with no classifier model, no GPU warm-up, no background inference, and no recurring jobs.

The mapping is transparent and conservative. Conversational greetings run with reasoning off. Rewrites, translations, and short summaries lock in low reasoning and hide tools for that step to protect the source text. Longer summaries step up to medium. Ordinary explanations and comparisons use medium with tools available. Prompt improvement deliberately stays at high, because faster modes were observed to invent requirements during evaluation. Complex, ambiguous, or unrecognised work—and any turn where a tool call just failed—escalate to high and restore tools so the normal loop can recover. Every decision is logged to a sidecar telemetry file that records routing choices, timings, and character counts only; it never copies prompt or answer text.

This plugin is for DSH users who already have a verified model and provider pairing and simply do not want to think about reasoning level before every sentence. It does not change your saved provider, model, or default settings; it only takes over per-request effort on routes you explicitly allowlist, and all other routes pass through untouched. Configure the mapping once, and every subsequent request is handled automatically. Disable or remove the plugin and normal manual reasoning selection returns immediately.