dsh-loop-doctor

中文 | English

A self-diagnosing plugin family for DeepSeek Harness (dsh): deterministic detour detection over the harness's own SessionEvent log, human-reviewable optimization suggestions with replay verification, and accepted-fix sediment that grows over time — the harness reading its own log to tell the operator where it wasted work.

Observation before advice, advice before writes, writes behind human review.

The four packages

The self-optimization loop

session log ──▶ detector ──▶ signal ──▶ advisor ──▶ suggestion + replay
                                    │                        │
                                    │                        ▼
                                    │                 apply: ledger (human)
                                    │                        │
                                    ▼                        ▼
                       tool: loop_doctor_diagnose      accepted-fix sediment
                                                             │
                                    advisor reload ◀──────────┘
                                    (next match injects the validated fix)

Safety baseline

• Rules before LLM. Detection is deterministic; no model call is on the detection path.
• Confidence ceiling 0.9. A rule engine never claims certainty; expert-labeled patterns are marked unproven.
• Replay-verified ≠ simulated. Only behavior-modeling simulations measured on the exact log earn replay-verified; policy-cap arithmetic carries simulated and never drives auto-apply.
• Writes behind review. The family's only writes are its own ledger and sediment files; auto-apply is off by default and gated to replay-verified, ceiling-confidence fixes.
• Two-layer honesty. The keyless snapshot proves the pipeline in a real loop; the emergent proof (a real model naturally detouring) needs a keyed run and is documented as a recipe, never claimed as verified.

Get started

From this repository

pnpm install
pnpm run typecheck   # tsc over the whole family
pnpm test            # vitest — 140 tests across the four packages

The end-to-end keyless proof lives at examples/headless-agent/loop-doctor.cordis.snapshot.yml (three real detours, self-diagnosis twice, auto-apply + reload observed in-snapshot). It runs inside a full deepseek-harness checkout with the branch applied; see the Agent Note for the replay-fixture constraints.

Install from npm

pnpm add @hubert-hwk/dsh-loop-doctor-detector \
         @hubert-hwk/dsh-loop-doctor-advisor \
         @hubert-hwk/dsh-loop-doctor-apply \
         @hubert-hwk/dsh-loop-doctor-tool

Inside DeepSeek Harness

The four packages are designed to be mounted as plugins of a dsh host: wire the installed packages into your dsh plugin configuration. To run them from a source checkout of the harness instead, copy packages/loop-doctor/ into the harness workspace and rename the packages back to the upstream @hubert-hwk/dsh-loop-doctor-* names (or use the upstream feat/loop-doctor branch directly).

Relationship to upstream

This family was authored as an upstream contribution to deepseek-ai/deepseek-harness (branch feat/loop-doctor, based on the official master). The upstream project currently does not accept external pull requests (see their CONTRIBUTING.md), so this repository exists as the public, community-visible home of the work:

• Designed to upstream: layout and gates mirror the harness monorepo so the branch can be re-based onto a future upstream master when contributions reopen. The packages are published under the community scope @hubert-hwk/ because the @deepseek-ai/ npm scope is reserved for the official organization; inside the harness they carry the upstream names @hubert-hwk/dsh-loop-doctor-*.
• Agent Note included: .agents/notes/implemented/feature/2026-08-19-loop-doctor-self-diagnosis.md.

Community

• Discuss / upvote on the DeepSeek Harness GitHub Discussions (category: Show Your Plugins!).
• This repository carries the dsh-plugin topic so the harness community can discover it.

License

MIT