Kb Eval Repo

Skill: kb_eval_repo

Scope

Use this skill only when the user is asking to evaluate an open-source project, usually with a GitHub URL or owner/repo name, and wants a judgment such as:

• 是否对我/团队有用
• 值不值得深入
• 能不能复现
• 环境怎么配
• 风险、坑、适配成本
• 和知识库已有资料是什么关系

Do not use it for normal knowledge-base lookup, paper ingestion, file ingestion, or broad literature review.

Input

Accept the same query-page envelope used by the Java backend:

• protocol = research_kb_agent_task
• taskType = kb_query or kb_eval_repo
• kbTargets[] contains one or two repositories
• payload.question is the user's natural-language request
• Optional: payload.repoUrl, payload.options.writeReview

Only read repositories listed in kbTargets. Do not infer or access any other repository.

Workflow

1. Validate the task JSON.
2. Extract a GitHub repo from payload.repoUrl, payload.question, or owner/repo text.
3. If a GitHub repo is present, fetch public metadata, README, and common dependency/config files.
4. Read relevant Markdown pages from the selected knowledge repositories.
5. Evaluate with five dimensions:
   • relevance to the selected KB context
   • reproducibility and engineering maturity
   • activity and credibility
   • adaptation cost
   • unique value compared with KB materials
6. Give exactly one verdict:
   • 值得深入
   • 选择性参考
   • 暂不建议
7. Return a query-compatible JSON result with:
   • answer
   • citations
   • usedScopes
   • readPages
   • optional generatedPage

If payload.options.writeReview is not false, write an evaluation report to reviews/-开源项目评估.md in the first selected target repository and cite it.

Evidence Rules

• Project facts must come from GitHub metadata/README/dependency files.
• Knowledge-base comparisons must come from pages actually read.
• If no repo URL/name is provided and no related KB project can be found, ask the user to provide a GitHub URL.
• If GitHub cannot be fetched, explain the failure and still use KB evidence if enough exists.
• Do not pretend the project was actually installed or run. State that reproducibility is an initial assessment based on repository materials.

Output

Return only one JSON object. Do not use Markdown fences.

{
  "protocol": "research_kb_agent_result",
  "protocolVersion": "1.0",
  "taskId": "...",
  "taskType": "...",
  "success": true,
  "result": {
    "answer": "...",
    "citations": [],
    "usedScopes": ["personal"],
    "readPages": [],
    "mode": "repo_evaluation",
    "generatedPage": ""
  },
  "errors": []
}

Helper

Prefer the bundled deterministic helper:

python3 scripts/run_task.py --stdin
python3 scripts/run_task.py --task-json