majiayu000 Auto-Improvement Skill
Purpose
Documents the automated feedback loop: metrics collection, lesson capture, and pattern detection across sessions.
Architecture
SessionStart hook Stop hooks (sequential)
│ │
write marker ──→ session-metrics.cjs ──→ lesson-capture.cjs
current-session.json │ │
append JSONL evaluate thresholds
sessions.jsonl │
significant? ──→ prompt knowledge-capture
│
/review-improvements
(on-demand detection)
Data Flow
Storage Tiers
| Tier | Location | Tracked | Retention |
|---|---|---|---|
| Raw metrics | .epost-data/improvements/sessions.jsonl |
No (gitignored) | Auto-rotate at 1000 lines |
| Session marker | .epost-data/improvements/current-session.json |
No | Deleted on Stop |
| Detection report | .epost-data/improvements/latest-report.json |
No | Overwritten per run |
| Significant learnings | docs/ |
Yes (git tracked) | Permanent |
JSONL Schema
Each line in sessions.jsonl is a JSON object:
{
"sessionId": "string",
"timestamp": "ISO-8601",
"duration_ms": "number|null",
"branch": "string",
"git": { "filesChanged": 0, "insertions": 0, "deletions": 0 },
"tasks": { "total": 0, "completed": 0, "failed": 0 },
"errors": { "count": 0, "types": ["build|test|lint|runtime"] },
"rework": { "fixIterations": 0, "verificationFailures": 0 },
"skills": { "discovered": [], "loaded": [], "unused": [] },
"knowledge": { "retrieved": 0, "captured": 0, "staleHits": 0 },
"routing": { "intent": "string", "command": "string", "platform": "string" }
}
Full schema: packages/core/assets/improvements-schema.json
Significance Thresholds
Lesson capture triggers automatically when:
| Trigger | Threshold | Capture Type |
|---|---|---|
| Errors fixed | errors.count > 0 |
FINDING |
| Rework detected | fixIterations >= 2 |
PATTERN |
| Verification failure | verificationFailures >= 1 |
CONV |
| New skill first-seen | Not in previous 5 sessions | NOTE |
Detection Engine
Analyzes accumulated metrics from .epost-data/improvements/sessions.jsonl:
| Detection | Method | Recommendation |
|---|---|---|
| Repeat errors | Same error type 3+ times in 7 days | Create FINDING or fix root cause |
| Skill gaps | Platform detected but no platform-skill loaded | Check skill wiring |
| Stale knowledge | docs/ entries >90 days with no references |
Review or archive |
| Rework patterns | Avg fixIterations >2 over recent sessions | Review verification process |
| Unused skills | Loaded but unreferenced 10+ sessions | Optimize or remove |
Run via command: /review --improvements
Hooks
| Hook | Event | Purpose |
|---|---|---|
session-init.cjs |
SessionStart | Writes current-session.json marker |
session-metrics.cjs |
Stop | Collects metrics, appends to JSONL |
lesson-capture.cjs |
Stop (prompt) | Evaluates significance, prompts capture |
Related Skills
data-store—.epost-data/convention for persistent agent dataknowledge-capture— Post-task capture workflow (triggered by lesson-capture hook)knowledge-retrieval— Search existing knowledge before actingknowledge-retrieval— Knowledge system structure and conventionsverification-before-completion— Verification failures tracked for rework detectiondebug— Error metrics auto-captured by session-metrics hookcode-review— Convention violations auto-detected across sessions