dsh-tool-turbo
DSH 主机插件：通过针对简单工具任务自动降低 reasoning_effort，减少工具调用延迟，并提供按工具划分的耗时遥测。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:Electricitysheep/dsh-tool-turbo#d29e7d9d73032516e37797ea69592ce18e0fe05fTool Turbo 以 dsh-tool-turbo 发布，当前版本为 0.1.0。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
Cut tool-call latency in DeepSeek Harness (dsh) by auto-adjusting reasoning_effort per tool round.
中文文档 · English
In a multi-step tool chain, the model re-thinks before every tool call — and that thinking dominates the wall-clock time (a 50-step agent task can spend minutes in reasoning between tools). dsh-tool-turbo watches the recent tool calls of a step and injects the lowest sensible reasoning effort into the next model request, then lifts it again the moment the work gets heavy.
DeepSeek's API exposes reasoning_effort in three steps (low / high / max, shipped 2026-08-13). dsh re-resolves the request config for every step through an agent/request waterfall (see packages/core/agent-loop/src/agent.ts — "plugins propose the next request config"). dsh-tool-turbo plugs into that waterfall:
tool/call records from the session.write, read, grep, glob, bash, fs_*, …) with small payloads → low; mixed/heavy work → high; very heavy payloads → max (opt-in).agent/request config for the next model call of that step.Long tool chains keep the cheap rounds cheap, and never starve the hard rounds of reasoning.
# 1. clone + build the plugin
git clone https://github.com/Electricitysheep/dsh-tool-turbo.git
cd dsh-tool-turbo && npm install
# 2. register into your dsh profile (web shown; any profile works)
# ~/.dsh/profiles/web/package.json dependencies:
# "dsh-tool-turbo": "link:<absolute path to dsh-tool-turbo>"
# ~/.dsh/profiles/web/cordis.patch.yml:
# - insert:
# - id: tool-turbo
# name: dsh-tool-turbo
cd ~/.dsh/profiles/web && pnpm install
# 3. restart dsh web
dsh web
[tool-turbo] agent/request: baseline=high calls=[] => reasoningEffort=high
[tool-turbo] agent/request: baseline=high calls=[{"name":"write",…}] => reasoningEffort=low
decideEffort): fresh prompt keeps the baseline, simple-tool chains downgrade to low, downgrades respect the user toggle, heavy payloads upgrade to max (opt-in), mixed tools lift to high.tsc --noEmit clean.| Recent tool calls | Decision |
|---|---|
| none (fresh prompt) | keep user's selected effort |
| ≥75% simple tools, small args, downgrade allowed | low |
| mixed / heavy tools | high (when upgrades allowed) |
| very heavy payloads, upgrade allowed | max |
| otherwise | keep user's selected effort |
Toggles (settings namespace planned): allowDowngrade (default on), allowUpgrade (default off — keep max conservative), baseline (default high).
headless/tui)MIT