drscrewdriver/dsh-thinking-levels
deepseek-harness dsh thinging levels adjustment reasoning intensity adjustment
dsh-thinking-levels gives per-round reasoning-effort control for DeepSeek Harness. Pick Auto in the session model selector and the plugin schedules low/high/max from the recent tool-call history before submitting the API effort — cheap tool rounds stay cheap while heavy work never starves; or fix a wire level (off/on/minimal/low/medium/high/xhigh/max) manually. It hooks the agent/request waterfall that dsh re-resolves for every step (registered with prepend so the session model-selection assembly cannot overwrite it) and injects the level into the next model request. Wire facts are verified against the official docs and dsh's llm-deepseek adapter: low maps 1:1 on deepseek-v4-flash/v4-pro while medium/xhigh collapse onto high; auto is a plugin mask never sent to the API. A settings card exposes the auto scheduler boundaries and an llm-pi-ai custom-provider capability table with per-model short-circuit takeover and apply-to-all presets.
Install
dsh plugin --profile web add dsh-thinking-levels -wnpm dsh-thinking-levels 0.6.0 verified 2026-09-02 (repository field → github.com/drscrewdriver/dsh-thinking-levels; README EN with zh/ja/ko editions). Install per README: dsh plugin --profile web add dsh-thinking-levels -w — the web profile is a pnpm workspace root so -w is required; GitHub alternative dsh plugin --profile web add https://github.com/drscrewdriver/dsh-thinking-levels.git -w or a local-path add. Then restart dsh web (a running instance does not hot-load new bundle layers). The dsh runtime uses pnpm 11: a fresh package may hit the minimum-release-age rule — add the version to minimumReleaseAgeExclude in ~/.dsh/profiles/web/pnpm-workspace.yaml to lift the cooling period.
Compatibility
DSH web profile (any profile works); pnpm 11 runtime; official DSH exposes zh/en locales — ja/ko dictionaries exist but only work on a DSH fork that registers those locale IDs.
Details
- Repo: drscrewdriver/dsh-thinking-levels
- Category: Other
- Stars: 1
- Version: npm dsh-thinking-levels 0.6.0
- Last push: 2026-09-20
- First seen: 2026-08-18
Recent updates
Auto per-step effort scheduling from tool history; 8 fixed levels incl. off/on toggle-only handling; llm-pi-ai custom-model capability editor; settings card; verified wire mapping.
FAQ
- What is the Auto level?
- Auto is a mask: the plugin schedules low/high/max per step from the recent tool-call history and resolves it to a concrete wire level before submission — it is never sent to the API as-is.
- Which wire values does the DeepSeek adapter accept?
- off, low, high and max; medium and xhigh collapse onto high on the official adapter, and UNSUPPORTED_REASONING_EFFORT is rejected otherwise.
- Does it work with custom llm-pi-ai models?
- Yes — a settings card maps each thinking level to the exact value your hand-declared model expects, with per-model short-circuit takeover.
Alternatives
Dbi-Eshuh/dsh-thinking-status-customizer · LiangYin233/dsh-provider-model-configurator