Skill: kb_literature_review

Scope

Use this skill when the user wants a cross-document synthesis from the selected personal/team knowledge bases, not a single fact lookup.

Typical triggers:

• 写一篇/一份关于 X 的综述
• 梳理知识库里关于 X 的资料
• 对比几种方法/路线
• 总结主流方法、优缺点、研究缺口
• 基于知识库内容做专项报告

Do not use it for:

• simple lookup like "有没有 X"
• one document's conclusion
• open-source repository usefulness evaluation
• ingestion or source scanning

Input

Accept the Java backend query-page envelope:

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

Only read repositories listed in kbTargets.

Workflow

1. Validate the task JSON.
2. Determine the review topic from payload.topic or payload.question.
3. Read repository Markdown trees and collect candidate pages, excluding source_files/.
4. Score candidates by topic terms, title, path, headings, and content.
5. Read a focused set of relevant pages. Prefer concept/review/summary pages before low-level archive pages.
6. If fewer than two relevant pages exist, say the knowledge base is insufficient and do not fabricate a review.
7. Generate a Chinese review with:
   • scope
   • topic overview
   • method/category map
   • comparison of strengths and limitations
   • contradictions or disagreements
   • research gaps
   • implications for the selected KB context
   • source list
8. If payload.options.writeReview is not false, write the review to reviews/<topic>-专项综述.md in the first selected target repository and cite it.
9. Return one query-compatible JSON result.

Evidence Rules

• Every knowledge-base fact must come from pages actually read.
• Do not fill missing areas with general knowledge unless explicitly marked as non-KB background.
• Citations must be repository-relative .md paths and must not point to source_files/.
• The answer body should be readable as a concise executive summary; detailed source list belongs in citations[].

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", "team"],
    "readPages": [],
    "mode": "literature_review",
    "generatedPage": ""
  },
  "errors": []
}

Helper

Prefer the bundled deterministic helper:

python3 scripts/run_task.py --stdin
python3 scripts/run_task.py --task-json <path>