dsh-quality-review
DeepSeek Harness 的 AI 响应质量审查：使用独立的审查模型审核每个已完成的助手回合，并引导智能体修正不合理的输出（最多 2 轮审查）。
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:CAI-MH/dsh-quality-review#cc6aff67f5305b41f613f8a84674ea1815ba057bQuality Review 以 dsh-quality-review 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。
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).
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):
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.
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
npm install
npm run build # tsc — compiles src/ to lib/
The published package ships the compiled lib/ output; src/ holds the TypeScript sources.