Self-Evolve — Continuous Improvement Engine

The system that makes the system better. Every error is a lesson. Every feedback is a signal. Every week, the toolkit gets smarter.

Self-evolve closes the loop between usage and improvement. Instead of static skills that decay, this engine ensures your solo-cto-agent installation learns from every session.

---

Architecture

bin/self-evolve.js               ← Public module API
bin/self-evolve/
├── self-evolve-orchestrator.js  ← Master orchestrator (full cycle management)
├── error-collector.js           ← Error pattern tracking → error-patterns.md
├── quality-analyzer.js          ← Post-task quality checks → quality-log.md
├── skill-improver.js            ← 3-strike auto-patch → skill-changelog.md
├── feedback-collector.js        ← L1 user feedback (1-5) → feedback-log.md
├── weekly-report.js             ← Weekly synthesis → reports/weekly-*.md
├── skill-scout.js               ← Skill discovery & compatibility check
└── external-trends.js           ← L3 npm/GitHub/API trends → reports/trends-*.md

Data Files

All markdown-based. No database required. Lives in your project directory.

{project}/
├── error-patterns.md      ← Error pattern registry (auto-appended)
├── quality-log.md         ← Quality check records (auto-appended)
├── skill-changelog.md     ← Skill modification history
├── feedback-log.md        ← User satisfaction scores (L1)
├── reports/
│   ├── weekly-*.md        ← Weekly synthesis reports
│   └── trends-*.md        ← External trends reports (L3)
└── memory/
    ├── index.md           ← Episode search index
    ├── episodes/          ← Session episodes (14-day retention)
    └── knowledge/         ← Auto-generated knowledge articles (permanent)

---

CLI Commands

Initialize

solo-cto-agent self-evolve init          # Create all data files & directories
solo-cto-agent self-evolve status        # System health check

Error Tracking

# Record an error
solo-cto-agent self-evolve error \
  --skill review --category build \
  --symptom "TypeScript strict mode failure" \
  --cause "Missing type annotation" \
  --fix "Added explicit return type" \
  --severity medium

# View top repeated errors
solo-cto-agent self-evolve errors --top 10

Quality Analysis

# Record quality check results
solo-cto-agent self-evolve quality \
  --type code --skill review \
  --checks "build:pass,typescript:pass,imports:warn"

# View quality trends + feedback summary
solo-cto-agent self-evolve summary

User Feedback (L1)

# Record satisfaction score (1-5)
solo-cto-agent self-evolve feedback \
  --skill review --score 4 \
  --task "PR code review" \
  --reason "Caught the N+1 query"

Skill Improvement

# Check for pending improvement triggers
solo-cto-agent self-evolve improve

# Apply all pending improvements
solo-cto-agent self-evolve improve --apply

Reports

# Weekly synthesis report
solo-cto-agent self-evolve report [--weeks 2]

# External trends (npm outdated + GitHub trending)
solo-cto-agent self-evolve trends --npm-dir /path/to/project

Skill Scouting

# List installed skills
solo-cto-agent self-evolve scout --installed

---

How It Works

1. Error Pattern Collection

Every time an error occurs during a task, it gets recorded in error-patterns.md with a structured format:

### [2025-04-15] Error ID: E-1
- **skill**: review
- **category**: build
- **severity**: high
- **symptom**: ReferenceError: COLORS is not defined
- **root cause**: Sub-module extraction missed constant delegation
- **fix**: Added const COLORS = reviewParser.COLORS
- **repeat count**: 1
- **needs skill improvement**: no

Duplicate detection: If the same skill + similar symptom (>60% word overlap) is recorded again, the existing entry's repeat count increments instead of creating a duplicate.

Circuit breaker: When an error pattern hits 3 repeats, it triggers the skill improvement engine.

2. Quality Analysis

After each significant task, quality checks run against type-specific checklists:

Type	Checks
code	build-success, typescript-strict, no-console-log, import-accuracy, deployable, error-handling
design	anti-ai-slop, 8px-grid, color-tokens, mobile-responsive, accessibility, consistent-spacing
document	fact-tags, source-citations, no-exaggeration, actionable-next-steps, correct-format
analysis	fact-tags, source-citations, data-backed, risk-first, actionable-next-steps

Results scored as pass, warn, or fail. Same type/skill hitting warn/fail 3 times triggers skill improvement.

3. Skill Auto-Improvement (3-Strike Rule)

When triggers fire (from errors or quality), the engine:

1. Scans error-patterns + quality-log + feedback-log for the specific skill
2. Generates a prevention rule or checklist reinforcement
3. Appends it to the skill's SKILL.md under "Auto-Improvements" section
4. Records the change in skill-changelog.md
5. Reports to the user (Level 2: do first, report after)

Protected areas: Core philosophy sections, routing tables, and cross-skill dependencies are never auto-modified.

4. L1/L2 Feedback Loop

L1: After significant tasks, collect a 1-5 satisfaction score.

• 5 = Perfect, use as-is
• 4 = Good, minor tweaks
• 3 = Okay, room for improvement
• 2 = Poor, major rework needed
• 1 = Unusable

L2: When the same skill gets ≤2 score 3 times, the engine auto-diagnoses and patches the skill based on accumulated feedback reasons.

5. Weekly Report

Generated on Monday mornings (or on demand). Synthesizes:

• Project activity overview
• Skill changes made
• Quality pass/warn/fail trends
• User feedback averages
• Top repeated errors
• Skill scouting findings
• Next week priorities

6. External Trends (L3)

Scans three sources for updates relevant to your stack:

1. npm outdated — Dependency update check (major vs minor)
2. GitHub trending — Repos matching your tech stack keywords
3. API changelogs — Breaking changes, new features, model updates

7. Skill Scouting

Discovers new skills and evaluates them against your setup:

• Conflict detection: Name overlap, description overlap with installed skills
• Compatibility check: Filters by your tech stack keywords
• Action classification: auto-install (system improvements) vs recommend (new capabilities)
• Protection rules: Never installs 2+ skills at once, never overwrites custom rules

---

Module API

const selfEvolve = require('solo-cto-agent/bin/self-evolve');

// Error tracking
selfEvolve.collectError(projectDir, { skill, category, symptom, cause, fix, severity });
selfEvolve.getTopErrors(projectDir, 10);
selfEvolve.getImprovementTriggers(projectDir, 3);

// Quality
selfEvolve.analyzeQuality(projectDir, { type, skill, checks, notes });
selfEvolve.getQualityTrend(projectDir, 20);

// Feedback
selfEvolve.recordSatisfaction(projectDir, { skill, score, task, reason });
selfEvolve.getLowScorePatterns(projectDir, 3);
selfEvolve.getFeedbackSummary(projectDir);

// Improvement
selfEvolve.checkTriggers(projectDir);
selfEvolve.applyImprovement(skillsDir, trigger, projectDir);

// Reports
selfEvolve.generateWeeklyReport(projectDir, { weeks: 1 });
selfEvolve.generateTrendsReport(projectDir, { npmOutdated, trending });

// Scouting
selfEvolve.getInstalledSkills(skillsDir);
selfEvolve.evaluateSkill(newSkill, skillsDir);

// Orchestrator
selfEvolve.runPostTask(projectDir, taskInfo);
selfEvolve.runSessionEnd(projectDir, skillsDir);
selfEvolve.getStatus(projectDir);
selfEvolve.initializeDataFiles(projectDir);

---

Integration with Existing Commands

Self-evolve hooks into the existing solo-cto-agent workflow:

• After review: Quality check runs automatically (if enabled)
• After knowledge: Knowledge articles feed into the weekly report
• feedback command: Existing accept/reject feeds into personalization; self-evolve adds 1-5 scoring
• doctor command: Now also checks self-evolve health
• watch mode: Error patterns update on each iteration
• Session end: Improvement triggers check automatically