ascorbic/macrodata/distill

find ~/.claude/projects -name "*.jsonl" -mtime -1 -type f 2>/dev/null

Task(subagent_type="general-purpose", prompt=`
Read the conversation at {path}.
 
Filter to actual conversation content:
- Include: human messages, assistant text responses
- Exclude: tool calls, tool results, system messages, thinking blocks
 
Extract and return as JSON:
{
  "distilled_actions": [
    {
      "summary": "Fixed auth bug in src/auth.ts where token refresh was racing",
      "files": ["src/auth.ts"],
      "outcome": "Added mutex lock around refresh"
    }
  ],
  "facts": [
    {
      "topic": "project-name",
      "content": "Uses JWT tokens with 15min expiry"
    },
    {
      "topic": "person-name",
      "content": "Prefers explicit error handling over try/catch"
    }
  ],
  "decisions": [
    "Chose Redis over in-memory cache for session storage because of multi-instance deployment"
  ]
}
 
Focus on:
- What was accomplished (not just discussed)
- Decisions made and their rationale
- New information about projects, people, or preferences
- File paths and specific technical details that should survive compression
 
Return ONLY the JSON, no explanation.
`)

For each action in all results:
  log_journal(topic="distilled", content=action.summary + " Files: " + action.files.join(", "))

log_journal(topic="distill-summary", content="Processed N conversations. Extracted X actions, Y facts.")

{
  "distilled_actions": [
    {
      "summary": "Added /distill skill to macrodata plugin",
      "files": ["plugins/macrodata/skills/distill/SKILL.md"],
      "outcome": "Skill extracts facts from conversations via sub-agents"
    }
  ],
  "facts": [
    {
      "topic": "macrodata",
      "content": "Distillation separates narrative context from retained facts for better compression"
    }
  ],
  "decisions": [
    "Coordinator updates state directly to prevent race conditions from parallel sub-agents"
  ]
}