dsh-quality-review

dsh-quality-review

English | 中文

AI response quality review for DeepSeek Harness (DSH): audits each finished assistant turn with an independent reviewer model and steers the agent to fix unreasonable output (up to 2 review rounds).

How it works

On agent/turn-stopping — the moment a turn is about to close — the plugin takes the assistant's latest visible reply, audits it with a reviewer model, and when the verdict is fail it steers the agent with a concrete fix request so the turn stays open and the model revises. Each turn allows at most maxRounds steered follow-ups (default 2), which is the hard loop guard.

Reviewed aspects (each can be toggled):

• Factual accuracy — obvious factual errors and fabricated information
• Completeness — any part of the user's question left unanswered
• Logical consistency — contradictory reasoning and unsupportable conclusions
• Instruction following — violations of explicit user format/language/constraint requirements

Install

dsh plugin --profile web add dsh-quality-review

Or add it to your profile's package.json dependencies and list it under dsh.profile.bundles, then restart dsh web.

Configuration

Every field has a safe default, so the bundle row can be inserted with no config at all. Override individual keys from your profile's cordis.patch.yml:

- id: quality-review
  config:
    enabled: true            # master switch; the plugin stays loaded but inert when false
    reviewer:
      provider: ''           # reviewer provider id; empty reuses the agent's own provider
      model: ''              # reviewer model id; empty reuses the agent's own model
    maxRounds: 2             # max steered fix rounds per turn (loop guard, max 5)
    aspects:
      factualAccuracy: true
      completeness: true
      logicalConsistency: true
      instructionFollowing: true
    minReplyChars: 200       # skip replies shorter than this
    sop:
      enabled: true          # SOP reference-standard switch (default on)
      dir: ''                # SOP folder; empty uses DSH_HOME/quality-review/sop
    exemptPatterns:          # extra static keywords: skip review on match (optional)
      - 代码 review

Common-task SOP reference standards (sop)

For routine SOP-style tasks, reviewing against generic dimensions is noisy. The plugin reads a SOP folder and treats each file name (extension stripped) as a match keyword for the related task; when the user prompt matches any name (case-insensitive substring), it loads the content of every file in the folder — a SOP may span several files — and injects it into the review prompt as the task's quality standard, adding a "是否符合 SOP 规范" dimension the reviewer checks the answer against. Non-matching turns still use the generic four dimensions.

Default folder: DSH_HOME/quality-review/sop. Drop files in and fill them with the task's standard (what a good output must contain / avoid); no config or restart needed — the folder is re-scanned every turn. exemptPatterns is separate: a static keyword list that skips review on match.

Development

npm install
npm run build    # tsc — compiles src/ to lib/

The published package ships the compiled lib/ output; src/ holds the TypeScript sources.

License

MIT