Electricitysheep/dsh-tool-turbo
Per-round reasoning_effort optimizer for DeepSeek Harness (dsh): auto-downgrades tool-call reasoning for simpl
dsh-tool-turbo cuts tool-call latency in DeepSeek Harness by auto-adjusting reasoning_effort per tool round. In a multi-step tool chain, the model re-thinks before every call — and that thinking dominates wall-clock time. dsh-tool-turbo watches the recent tool calls of a step, determines the complexity, and injects the lowest sensible reasoning_effort into the next model request: simple deterministic tools (write, read, grep, bash, fs_*) get low; mixed/heavy work gets high; very heavy payloads get max (opt-in). Long chains keep cheap rounds cheap without starving hard rounds of reasoning.
Install
git clone https://github.com/Electricitysheep/dsh-tool-turbo.git && cd dsh-tool-turbo && npm installNo npm package — clone, build, then register as a link dependency. After npm install, add to ~/.dsh/profiles/web/package.json: "dsh-tool-turbo": "link:<absolute-path-to-dsh-tool-turbo>" under dependencies. Add to ~/.dsh/profiles/web/cordis.patch.yml: - insert:\n - id: tool-turbo\n name: 'dsh-tool-turbo'. Restart dsh web. The plugin uses the agent/request waterfall available in DeepSeek Harness 0.1.x (reasoning_effort feature added 2026-08-13).
Compatibility
DeepSeek Harness 0.1.x with reasoning_effort API support (three steps: low / high / max; added 2026-08-13). Hooks the agent/request waterfall. No npm package — local link install required.
Details
- Repo: Electricitysheep/dsh-tool-turbo
- Category: Other
- Stars: 5
- Version: GitHub source link install (README-documented 2026-08-24; no npm package). Requires DeepSeek API reasoning_effort support (introduced 2026-08-13).
- Last push: 2026-08-17
- First seen: 2026-08-13
Recent updates
Current: agent/request waterfall hook; whitelist of simple deterministic tools → reasoning_effort low; mixed/heavy → high; very heavy → max (opt-in); reasoning_effort API: low/high/max (DeepSeek 2026-08-13). Local link install (no npm).
FAQ
- How does it decide which reasoning level to apply?
- It watches the recent tool/call records from the current step. Simple, deterministic tools (write, read, grep, glob, bash, fs_*) with small payloads → reasoning_effort low. Mixed or heavier work → high. Very heavy payloads → max (opt-in). The decision is deterministic, not model-based.
- Can this hurt reasoning quality on hard tool calls?
- No — the decision only lowers effort for known-simple tools. Hard or ambiguous work defaults to high, and very heavy payloads can be opted into max. Hard calls are never downgraded.
- Does this require DeepSeek's reasoning_effort API?
- Yes — it depends on the reasoning_effort parameter (low/high/max) introduced in the DeepSeek API on 2026-08-13. Runs with DeepSeek Harness 0.1.x releases that expose this through the agent/request waterfall.
Alternatives
omdsh-dev/dsh-tool-stat · omdsh-dev/dsh-tool-calculator · LaoYueHanNi/dsh-token-usage