trend-harvester

Trend Harvester — Full Pipeline

Architecture Overview

[수집]   trend_harvester.py     (no_agent, 6h)
  │
  ▼
[평가]   kanban worker          (LLM, daily)
  │       ┌──────────────┐
  │       │ apply/discuss │
  │       │ defer/discard │
  │       │ + 비교 (기존과)│
  ▼       └──────────────┘
[적용]   kanban worker          (LLM, per-item)
  │       ┌─────────────────────┐
  │       │ skill patch/create  │
  │       │ neuron update       │
  │       │ config change       │
  ▼       └─────────────────────┘
[관찰]   trend_usage_watch.py   (no_agent, daily)
  │
  ▼
[폐기]   kanban worker          (LLM, weekly)
          ┌─────────────────────┐
          │ 검토 → retire       │
          │ 또는 keep           │
          └─────────────────────┘

Pipeline Stages

1. 수집 (Collection)

• Cron: trend-collect (no_agent, 0 */6 * * *)
• Script: trend_harvester.py
• Output: collected/ → analyzed/{keep,review,graveyard}/
• LLM 필요?: ❌ No

2. 평가 (Evaluation)

• Trigger cron: trend-evaluate-trigger (LLM, daily 10:00 KST)
• Worker: Kanban worker (profile: default, trend-harvester skill loaded)
• Input: analyzed/keep/*.json (아직 evaluated/에 없는 항목)
• Process:
  1. Read keep JSON file
  2. Search existing skills/config/neurons for comparison
  3. Evaluate using 5-axis filter + comparison logic
  4. Decide: APPLY / DISCUSS / DEFER / DISCARD
  5. Write result to evaluated/YYYY-MM-DD-hash.json
  6. If APPLY: copy to pending/, create child kanban task trend-apply-<name>
  7. If DISCUSS: create kanban task for human review
• Output: Workers create child tasks for each APPLY item
• LLM 필요?: ✅ Yes

3. 적용 (Application)

• Trigger: Created as child task by evaluation worker
• Worker: Kanban worker
• Input: pending/<name>.json
• Process:
  1. Decide tier: Tier 1-2 (auto-apply) or Tier 3-4 (human approval)
  2. Auto-apply: create skill (Tier 1), patch existing skill (Tier 2)
  3. Human-approval: kanban task with draft + options
  4. Move to applied/<name>.json with provenance metadata
• LLM 필요?: ✅ Yes