evolve

/evolve — Manual Evolution Trigger

CRITICAL: Run HARNESS_DIR=$(epic path) first. NEVER use .harness/ in the project directory.

You are the Evolution Engine — analyze past sessions to improve skills.

Sub-commands

/evolve (default) — Run evolution now

1. Read observation logs from $HARNESS_DIR/obs/
2. Analyze failure patterns across all sessions
3. Identify weak areas (error types, recurring failures)
4. Generate or improve evolved skills in $HARNESS_DIR/evolved/
5. Gate: validate new skills (format, dedup, cap of 10)
6. Report what changed

/evolve status — Show evolution dashboard

Read $HARNESS_DIR/metrics.json and $HARNESS_DIR/evolution.jsonl, then display:

## Evolution Dashboard

### Overview
- Sessions analyzed: {total_sessions}
- Average success rate: {avg_success_rate}%
- Best score: {best_score} (session: {best_session})
- Trend: {trend} ({score_history.length} data points)
- Stagnation count: {stagnation_count} / 3 (rollback at 3)

### Score History (last 5 sessions)
| Session | Success Rate | Avg Score | Observations | Tool Success | Output Quality |
|---------|-------------|-----------|--------------|-------------|---------------|

### Evolved Skills
(list $HARNESS_DIR/evolved/*/SKILL.md with name and description from frontmatter)

### Last Session Analysis
(read last entry from evolution.jsonl)
- Error patterns: {error_patterns}
- Failure patterns: {failure_patterns[].pattern_type}
- Skills seeded: {skills_seeded}
- Skills rolled back: {skills_rolled_back}
- Analysis: {analysis_summary}

/evolve history — Long-term analysis

Read $HARNESS_DIR/evolution.jsonl (full history), then display:

## Evolution History

### Trend Over Time
| Session # | Date | Success Rate | Avg Score | Skills | Patterns |
|-----------|------|-------------|-----------|--------|----------|

### Cumulative Pattern Frequency
| Pattern | Total Count | First Seen | Last Seen |
|---------|-------------|------------|-----------|

### Skill Effectiveness
| Skill | Sessions Active | Avg Score With | Avg Score Without | Delta |
|-------|----------------|----------------|-------------------|-------|

### Dispatch Analysis
| Skill | Times Invoked | Top Trigger Signals |
|-------|--------------|---------------------|

/evolve rollback — Undo last evolution

1. If $HARNESS_DIR/evolved_backup/ exists, restore it to $HARNESS_DIR/evolved/
2. Otherwise, read $HARNESS_DIR/evolution.jsonl for last entry, remove skills seeded in that entry
3. Append a rollback record to evolution.jsonl
4. Report what was rolled back

/evolve reset — Clear all evolution data

1. Remove $HARNESS_DIR/evolved/, $HARNESS_DIR/evolved_backup/
2. Clear metrics.json and evolution.jsonl
3. Confirm with user first

How Evolution Works

Observe (PostToolUse — multi-dimensional scoring)
    ↓ $HARNESS_DIR/obs/session_YYYYMMDD.jsonl
Analyze (Stop or /evolve)
    ↓ SessionAnalysis: per-tool, per-ext, score distribution
    ↓ Pattern detection: repeated_same_error, fix_then_break, long_debug_loop, thrashing
Seed (auto-generate targeted skills)
    ↓ 4 seeding paths: pattern / weak tool / weak file type / high-freq error
Gate (validate: format, dedup, cap of 10)
    ↓ Stagnation check: 3 sessions no improvement → rollback to best checkpoint
Reload (next session resume reports metrics + loads evolved skills)

Skill Synthesis (host-agent)

When epic-harness reflect seeds a skill, it emits a pending-synthesis manifest

to $HARNESS_DIR/pending_synth.jsonl (failure evidence + template body). To

synthesize a better body:

1. Read $HARNESS_DIR/pending_synth.jsonl; for each record with status: "pending":
2. Launch a subagent (use your host's subagent mechanism — do not name a specific

tool or model) with the manifest's prompt_guidance + evidence, instructing

it to write a markdown skill body with the required sections (## Process,

## Anti-Rationalization, ## Evidence Required, ## Red Flags).

1. Apply the result:

   epic-harness evolve accept-synth --skill <name> --file <body.md>

1. The CLI validates the body, runs the Critic gate, overwrites the template

skill, and marks the manifest consumed.

If no host runs accept-synth, the template body persists — synthesis can only

improve a skill, never block it.

Scoring System

• Composite: 0.5 × tool_success + 0.3 × output_quality + 0.2 × execution_cost
• Failure classification: 9 categories (type_error, syntax_error, test_fail, lint_fail, build_fail, permission_denied, timeout, not_found, runtime_error)
• Pattern detection: 4 types (repeated_same_error, fix_then_break, long_debug_loop, thrashing)

Red Flags

• Evolving after only 1-2 observations (not enough data)
• Keeping evolved skills that never trigger
• Not reviewing evolved skills periodically
• Ignoring stagnation warnings