Deep Research 🔬

Multi-round web research with question decomposition, cross-verification, and structured reports with numbered citations.

Architecture

Three-step agent-assisted pipeline. No external API keys needed — uses OpenClaw built-in web_search and web_fetch tools.

┌─────────┐     ┌──────────┐     ┌──────────┐
│  plan    │ ──▶ │ analyze  │ ──▶ │ report   │
│ (脚本)   │     │ (脚本)    │     │ (脚本)    │
└────┬─────┘     └────┬─────┘     └────┬─────┘
     │ search_        │ fetch_         │
     │ commands       │ commands       │ report
     ▼                ▼                ▼ skeleton
  Agent 执行       Agent 执行       Agent 填充
  web_search      web_fetch        分析内容

Script path: scripts/research.py (relative to this skill directory)

Workflow

Step 1: Plan — 生成搜索指令

python3 scripts/research.py plan "topic" --depth standard

Output JSON contains search_commands — a list of web_search tool calls. Agent executes each one and collects results into a JSON array:

[
  {"query": "...", "results": [{"title": "...", "url": "...", "snippet": "..."}, ...]},
  ...
]

Save to a temp file (e.g., /tmp/search-results.json).

Step 2: Analyze — 去重排序 + 生成 fetch 指令

python3 scripts/research.py analyze /tmp/search-results.json --top 8

Output JSON contains:

• selected_sources: Deduplicated, tier-sorted source list
• fetch_commands: web_fetch tool calls for top URLs
• summary: Human-readable source overview

Agent executes each web_fetch and collects results into a JSON array:

[
  {"url": "...", "text": "extracted content...", "length": 12345},
  ...
]

Save to /tmp/fetch-results.json.

Step 3: Report — 生成报告骨架

python3 scripts/research.py report \
  --topic "topic" \
  --search /tmp/search-analysis.json \
  --fetch /tmp/fetch-results.json \
  --depth standard --save

Output: Markdown report with {FILL: ...} placeholders. Agent reads the Source Content section and fills in analysis.

Depth Levels

Auto-detect from query complexity, or user specifies.

Quick ⚡ (≤30s)

• --depth quick → 1 sub-question, 2 fetches
• Agent may skip the script entirely for trivial queries
• Output: IM message only (≤2000 chars)

Standard 🔍 (2-3 min)

• --depth standard → 4-5 sub-questions, 8 fetches
• Full 3-step pipeline
• Output: IM summary + full report saved to research/

Deep 🔬 (5-10 min)

• --depth deep → 7 sub-questions (includes contrarian), 15 fetches
• Spawn sub-agents for parallel search:

  sessions_spawn:
    mode: run
    task: |
      Execute these web_search calls: {subset of search_commands}
      Return results as JSON array.
  

• Main agent runs analyze + report after collecting all results
• Output: IM summary + comprehensive report (2000-5000 words)

User overrides: "快速搜一下" → Quick, "详细研究" → Standard, "深度调研" → Deep

Source Authority Tiers

Script auto-classifies:

• Tier 1 🟢 Official docs, .gov/.edu, arxiv, RFCs
• Tier 2 🟡 Major tech blogs, Stack Overflow, vendor blogs
• Tier 3 🟠 Personal blogs, Medium, forum posts
• Tier 4 🔴 AI-generated, marketing landing pages

Cross-Verification (Standard & Deep)

After report skeleton is generated, agent verifies:

• Source diversity: ≥3 independent sources?
• Recency: Prefer last 12 months; flag outdated
• Conflicts: When sources disagree, present both with tier labels
• If gaps found → run additional web_search + web_fetch → append

Report Output

IM Summary

🔬 Research: {topic}

{2-4 paragraph summary}

**Key takeaways:**
- Finding 1 [1][2]
- Finding 2 [3]

📊 Confidence: {🟢/🟡/🔴}
📄 Full report: research/{slug}-{date}.md

Full Report

Saved to research/{topic-slug}-{YYYYMMDD}.md with:

• Executive Summary
• Key Findings (per sub-topic)
• Conflicting Information
• Confidence Assessment table
• Numbered Sources with tier labels
• Research Log

Progress Feedback

• Quick: No progress messages
• Standard: One mid-point: "🔍 已搜索 N 轮，找到 M 来源，正在抓取..."
• Deep: Per sub-agent updates

Edge Cases

• Too broad: Ask user to narrow; suggest 3-4 angles
• No results: Try EN↔CN keywords, then report honestly
• Rate limits (429): Wait 5s + retry; fallback to web_fetch on known URLs
• Large topics: Summarize each round to ≤500 chars before next

Language

• Match user's language for report
• Auto-add cross-language search (EN topic → add 1 CN query; CN → add 1 EN)
• Chinese: 「」直角引号、——全破折号、：全角冒号