trend-harvester

name	trend-harvester
title	trend-harvester
type	skill
space	outcome
tags	["outcome"]
created	"2026-05-20T00:00:00.000Z"
updated	"2026-06-14T00:00:00.000Z"
description	AI Trend Collection & Filtering — collect GitHub trending repos, score via 5-axis philosophy filter, evaluate, apply, and retire.
links	["[[@action/skills/SKILL-INDEX]]","[[@memory/growth/trend-harvester/index]]"]

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

4. 관찰 (Usage Watch)

• Cron: trend-usage-watch (no_agent, 0 6 * * *)
• Script: trend_usage_watch.py
• Process:
  1. Read applied/ items
  2. Check each against:
     • Skill directories: matching skill exists?
     • Config files: referenced?
     • Neuron/rules files: referenced?
  3. Classify: active / stale (30d+ no reference) / recent_unreferenced
  4. Write .usage_report.json
• Output: Summary to Discord, stale candidates flagged
• LLM 필요?: ❌ No

5. 폐기 (Retirement)

• Trigger cron: trend-retire-trigger (LLM, weekly Monday 10:00)
• Worker: Kanban worker
• Input: .usage_report.json (stale candidates)
• Process:
  1. Read stale items from usage report
  2. For each: confirm no hidden dependencies
  3. Archive: move to archived/ with retirement note
  4. If still needed: reset last_seen to "exempt"
• Output: Archived items + summary
• LLM 필요?: ✅ Yes

판정 기준 (Evaluation Decision Matrix)

Decision	Condition	Action
APPLY	score ≥ 6.0 AND no_model_dependency ≥ 0.7 AND (no existing OR clearly better)	→ pending/ + child task
UPGRADE	existing skill found AND new item is strictly better	→ pending/ (replace target)
DISCUSS	needs human judgment (architectural, Tier 3-4)	→ kanban task (human)
DEFER	promising but not now (high effort, low urgency)	→ evaluated/ with defer date
DISCARD	score < 4.0 OR model-dependent OR worse than existing	→ graveyard/

Pipeline State

Directory	Purpose
collected/	Raw GitHub/RSS scrape output
analyzed/keep/	Passes 5-axis filter (score ≥ 6.0)
analyzed/review/	Borderline (4.0 ≤ score < 6.0)
analyzed/graveyard/	Rejected (score < 4.0)
evaluated/	Keep items that have been LLM-evaluated
pending/	Approved for application, waiting for worker
applied/	Successfully applied
archived/	Previously applied but retired

Cron Jobs Summary

Job	Schedule	Type	Model	LLM?
trend-collect	0 */6 * * *	no_agent	—	❌
trend-evaluate-trigger	0 10 * * *	LLM agent	opencode-go/deepseek-v4-flash	✅ (light)
trend-usage-watch	0 6 * * *	no_agent	—	❌
trend-retire-trigger	0 10 * * 1	LLM agent	opencode-go/deepseek-v4-flash	✅ (light)

Troubleshooting

Stale PID Lock Recovery

Same as before — pkill -f trend_harvester + rm -f .harvester.lock

LLM Timeout (Evaluate Trigger)

Uses fast model (deepseek-v4-flash). If still timing out, check opencode-go API status.

Kanban Worker Not Picking Up Tasks

Check kanban dispatcher is running (launchctl list | grep cron-runner)

Integration with {{AGENT_NAME}}

• P4-cortex/growth/trend-harvester/ — Storage
• P4-cortex/knowledge/harvester_sync_state.json — Memory sync state
• P4-cortex/growth/trend-harvester/.eval_state.json — Evaluation state
• P4-cortex/growth/trend-harvester/.usage_report.json — Usage watch report
• kanban board — Worker tasks for evaluation/application/retirement

Related

• [[@action/skills/SKILL-INDEX]]
• [[@memory/growth/trend-harvester/index]]
• [[taste-review]]