Lev Self-Improvement

Transform event streams into actionable system improvements. Proactive learning (not just reactive fixes).

Quick Reference

Trigger	Action
"what can lev learn?"	Full analysis cycle
"propose improvements"	Generate proposals from patterns
"why did X fail?"	Root cause analysis (fail-forward)
"self-learn"	Automated improvement cycle

Architecture

events.jsonl → Analyze → Patterns → Proposals → Human Review → Apply
                 ↑                                              │
                 └──────────── feedback loop ──────────────────┘

Core Workflows

1. Event Analysis

lev learn analyze              # Detect patterns in events.jsonl
lev learn analyze --since 24h  # Last 24 hours only

See: references/event-analysis.md

2. Fail-Forward Protocol

When something fails, extract learning:

1. Capture: What happened? (exact error, context)
2. Root Cause: Why? (5 whys, dependencies)
3. Proposal: How to prevent? (skill patch, config change)
4. Confidence: How sure? (low/medium/high)

See: references/fail-forward.md

3. Proposal Workflow

# ~/.config/lev/proposals/{id}.yaml
id: prop-abc123
type: skill-patch | config-change | new-workflow
target: ~/.claude/skills/lev/SKILL.md
confidence: 0.85
description: "Add timeout handling"
diff: |
  + timeout: 30s
status: pending | approved | rejected | applied

See: references/proposals.md

Evolution Targets

Target	Location	When
Skills	~/.claude/skills/*/SKILL.md	Behavior improvements
Config	~/lev/.lev/config.yaml	Settings optimization
Daemons	~/.config/lev/daemons.yaml	Process tuning
Workflows	~/lev/workflows/*.yaml	Pattern codification

Relationship to Other Skills

• skill-evolver (absorbed): Reactive skill fixes
• lev (sibling): Behavior definition (lev-self improves it)
• bd (integration): Track proposals as issues

Storage Schema

~/.config/lev/
├── proposals/           # Pending proposals
│   └── {id}.yaml
├── memory/
│   ├── patterns.jsonl   # Detected patterns
│   ├── proposals.jsonl  # Proposal history
│   └── applied.jsonl    # Applied changes + outcomes
└── patterns.jsonl       # Active patterns for matching

Confidence Thresholds

Confidence	Action
≥90%	Auto-apply (after notification)
70-89%	Propose with recommendation
50-69%	Propose, request review
<50%	Log only, don't propose

See: references/tracking-schema.md

CLI Quick Reference

# Analysis
lev learn analyze             # Full event analysis
lev learn patterns            # Show detected patterns

# Proposals
lev learn propose             # Generate proposals from patterns
lev learn review              # Interactive proposal review
lev learn apply <id>          # Apply approved proposal

# Tracking
lev learn status              # Show pending proposals
lev learn history             # Show applied changes

Human-in-the-Loop

All changes require confirmation:

1. Proposal generated → notification
2. Human reviews diff
3. Approve/reject/modify
4. Applied changes logged with outcome

Rollback:

lev learn rollback <id>       # Revert applied proposal

---

For detailed schemas and protocols, see references/