dsh-task-effort

DeepSeek reasoning effort comes in three fixed tiers-off, high, max-and DSH previously locked one value globally. You end up paying high-tier tokens for a one-line question while a two-hour refactor never gets the max depth it deserves. dsh-task-effort moves the decision from a single global setting to every individual message, so each call gets exactly the reasoning power it needs.

The plugin hooks into the agent request waterfall and classifies the latest user message before each model call. Explicit tags like [effort=max] take top priority; keywords such as refactor, architecture, and debug escalate to max; pure chit-chat (politeness prefixes stripped) drops to off; messages exceeding the length threshold go to max; short signals with no cue inherit the previous turn. Steering or appending messages mid-turn triggers a re-evaluation, and the cache resets each round for a fresh assessment.

Production-oriented extras make it more than a keyword switch. costAware caps effort during Beijing-time weekday peaks (09-12 / 14-18), throttling max down to peakCap (default high) to cut inference spend. escalateOnError forces the remaining steps of a turn to max after a tool-call error, reducing blind retries. Subagents are classified independently based on their own delegation prompts, decoupled from the parent conversation. If you manually pick a tier in the GUI model selector, the plugin detects the deviation, respects your choice, and pauses auto-judgment until a new session or restart.

If you routinely mix one-liner questions with cross-file refactors on DeepSeek inside DSH and want reasoning effort to follow task weight automatically-while keeping costs in check-this plugin is built for you. It drops in with zero workflow changes, and the [effort=...​] tag is always there as a safety net when heuristics miss.