Research Flywheel

Research Flywheel

Mine conversations for topics worth researching, dispatch deep research, store findings, and produce digests. Closes the loop between what Chris talks about and what gets properly researched.

Trigger

Use when the user says any of:

• "research flywheel"
• "mine conversations for research topics"
• "what should I research"
• "research digest"
• "flywheel status"
• "flywheel run"

Commands

research-flywheel mine — Find research candidates

Scans conversation history for research-worthy topics.

research-flywheel mine [--since "2026-03-17"] [--agents all|researcher|claw|builder]

Flow:

1. Run configurable mine patterns (from references/mine-patterns.md) via lcm_grep
2. Extract topic mentions with source references (conversation ID + message context)
3. Deduplicate against already-completed research in state
4. Score candidates: frequency × recency × Chris-initiated bonus
5. Store as pending candidates in state file
6. Print ranked list

Patterns: See references/mine-patterns.md. Configurable, not hardcoded regex. Default patterns look for:

• Questions about topics ("how does X work", "what's the best Y")
• Mentions of tools/products/technologies in context of evaluation
• "research this", "look into", "investigate" directives
• Topics mentioned across multiple conversations (high frequency signal)

research-flywheel research "<topic>" — Dispatch deep research

Wraps the existing deep-research skill to research a specific topic.

research-flywheel research "edge AI for snow sports" --breadth 6 --depth 3

Flow:

1. Look up topic in pending candidates (or add if new)
2. Delegate to ~/.openclaw/skills/deep-research/scripts/deep-research.mjs with the topic
3. Monitor progress via state file updates
4. On completion: store report path in state, update candidate status to "completed"
5. Optionally write vault note via obsidian-cli

research-flywheel status — Show research state

research-flywheel status [--topic "snow"]

Prints:

• Completed research (topic, report path, date)
• Pending candidates (ranked by score)
• Gaps (topics mentioned but never researched)

research-flywheel digest — Produce summary

research-flywheel digest [--since "2026-03-17"] [--output discord|vault|stdout]

Flow:

1. Read all completed reports since the date
2. Generate a concise digest (key findings per topic, 2-3 sentences each)
3. Deliver to specified output channel
4. Discord delivery uses the message tool
5. Vault delivery uses obsidian-cli create

research-flywheel run — Full cycle

research-flywheel run [--dry-run] [--auto] [--top N]

Flow:

1. Mine for candidates
2. Filter to top N (default 3)
3. Research each (unless --dry-run)
4. Generate digest
5. If --auto: skip all confirmations
6. If --dry-run: print what would happen, exit

State File

~/.openclaw/agents/researcher/research/flywheel/state.json:

{
  "candidates": [
    {
      "topic": "edge AI for snow sports",
      "source": "conv:27/msg:8251",
      "priority": "high",
      "score": 8.5,
      "status": "pending",
      "tags": ["snowboard", "edge-ai"],
      "discoveredAt": "2026-03-18T10:00:00+10:00"
    }
  ],
  "completed": [
    {
      "topic": "snowboard apps competitive analysis",
      "reportPath": "research/reports/snowboard-apps-competitive-v3.md",
      "completedAt": "2026-03-18T14:06:00+10:00",
      "vaultNote": "snowboard-apps-competitive",
      "digest": "ShredApp leads with AI coaching, Snocru has best social features..."
    }
  ],
  "config": {
    "defaultBreadth": 4,
    "defaultDepth": 2,
    "topN": 3,
    "notifyChannel": "1483345530363973768",
    "dryRun": false
  }
}

Integration Points

• deep-research skill — research delegates to ~/.openclaw/skills/deep-research/scripts/deep-research.mjs
• LCM tools — mine uses lcm_grep + lcm_describe for conversation mining
• Obsidian vault — digest and research can write notes via obsidian-cli
• Discord — digest delivery via message tool
• OpenClaw cron — schedulable for automated daily runs (mine + digest cycle)

Cron Integration

# Daily flywheel: mine topics overnight, research top 1, post morning digest
# 0 7 * * * research-flywheel run --auto --top 1

Edge Cases

• No LCM data: If lcm_grep returns nothing, report "No conversation data available for mining" and exit cleanly
• Empty state file: Initialize with defaults on first run
• Duplicate topics: Dedupe by normalized topic string (lowercase, stripped)
• Researcher agent: Primary runner is the Researcher agent, but any agent can dispatch via sessions_send

Example Session

User: "flywheel status"
→ Shows 2 completed, 5 pending candidates

User: "flywheel run --dry-run"
→ Would mine 3 new candidates, research "GLM-5 turbo benchmarks", post digest

User: "flywheel run"
→ Mines, asks "Research top 3 topics?", confirms, runs deep-research on each, posts digest