KanoNoUta/dsh-captain

@kanonouta/dsh-captain

English | 中文

Captain is a dual-face DeepSeek Harness plugin that exposes one synthetic captain model route. GPT is the planner and independent reviewer, while DeepSeek workers make the incremental repository changes.

Install into DeepSeek Harness

Captain ships its bundle manifest and built Host/Client entry points, so a Web profile can install it directly from GitHub:

dsh plugin --profile web add github:KanoNoUta/dsh-captain

For source development inside a DeepSeek Harness checkout:

cd F:\path\to\deepseek-harness
git submodule add https://github.com/KanoNoUta/dsh-captain.git packages/extensions/captain
git apply packages/extensions/captain/patches/deepseek-harness-integration.patch
pnpm install
pnpm exec tsc -b packages/extensions/captain/tsconfig.host.json --pretty false
pnpm --filter @kanonouta/dsh-captain run bundle

The optional source integration patch adds both TypeScript compiler faces and keeps official DeepSeek and OpenCode DeepSeek routes distinguishable in the native model selector. Normal dsh plugin add installation uses cordis.patch.yml and does not require that patch.

Composition

The Host registers captain through the normal LLM adapter directory, so the existing model selector renders a route such as GPT-5.6 Terra -> DeepSeek V4 Flash beside ordinary providers. The route id is captain:<planner-model>-><worker-model> and the model entry advertises text and image input plus Balanced, High Quality, and Ultra reasoning efforts.

Captain resolves each nested provider/model route before dispatch. A policy effort is downgraded to the strongest effort that route advertises, while routes with no selectable reasoning effort receive no effort field.

Captain calls the configured provider routes through the existing OpenAI-compatible LLM adapters. It does not implement OAuth or a second credential store; configure relay providers in the normal LLM settings and reference their provider and model ids from the Captain settings namespace.

The planner returns a JSON dependency DAG. Workers with disjoint file ownership run concurrently under the configured token budget and adaptive parallel limit. A worker response that contains only unexecuted DSML tool-call markup receives one correction attempt; a second such response fails the task. Worker execution failures stop before review and repair. A reviewer receives the acceptance criteria, worker reports, and the current incremental Git diff, and receives one strict JSON correction attempt after a malformed response. A failed code review selects only tasks named by findings for repair; an unscoped finding rechecks the full plan. A passing review advances the in-memory checkpoint used by the next run.