# Integrating agent-skill-bus with Claude Code Add self-improving skill tracking to any Claude Code project in 5 minutes. Every task Claude Code completes gets logged. Over time, the system detects which skills are degrading and surfaces repair suggestions. ## Setup ### 1. Install and initialize ```bash npm install agent-skill-bus npx agent-skill-bus init ``` This creates three directories under `skills/`: ``` skills/ ├── prompt-request-bus/ # Task queue (JSONL) ├── self-improving-skills/ # Skill run logs + health └── knowledge-watcher/ # External change tracking ``` ### 2. Add to CLAUDE.md (or AGENTS.md) Paste this block into your project's `CLAUDE.md` so Claude Code automatically records every task result: ```markdown ## Skill Bus Integration After completing any task, log the result with the CLI: \`\`\`bash npx agent-skill-bus record-run \ --agent claude-code \ --skill \ --task "" \ --result \ --score <0.0-1.0> \`\`\` ### Skill names to use | When you... | Skill name | |------------|------------| | Write new code | `code-generation` | | Fix a bug | `bug-fix` | | Refactor existing code | `code-refactor` | | Write tests | `test-writing` | | Review code | `code-review` | | Write documentation | `documentation` | | Debug an issue | `debugging` | | Create a PR | `pr-management` | ### Scoring guide | Score | Meaning | |-------|---------| | 1.0 | Perfect — no revisions needed | | 0.8 | Good — minor adjustments | | 0.6 | Partial — significant rework needed | | 0.3 | Poor — mostly failed, some salvageable | | 0.0 | Total failure | ### Before starting work, check the queue \`\`\`bash npx agent-skill-bus dispatch \`\`\` Pick the highest-priority task from the output. After finishing it: \`\`\`bash npx agent-skill-bus complete --result "description of what was done" \`\`\` ``` ### 3. Add a health check to your workflow Run this periodically (or add it to a pre-commit hook / CI step): ```bash # Update health report npx agent-skill-bus health # Show skills that need attention npx agent-skill-bus flagged # Detect silent degradation (score drop >15% week-over-week) npx agent-skill-bus drift ``` ## Usage Patterns ### Pattern 1: Record after every task (simplest) In your CLAUDE.md instructions, Claude Code will run `record-run` after each completed task. No code changes needed — it is pure CLI. ```bash # Claude Code runs this after fixing a bug: npx agent-skill-bus record-run \ --agent claude-code \ --skill bug-fix \ --task "Fix null pointer in auth middleware" \ --result success \ --score 0.9 \ --notes "Root cause: missing null check on user.session" ``` ### Pattern 2: Enqueue tasks from issues When you create GitHub issues, also enqueue them in the skill bus: ```bash npx agent-skill-bus enqueue \ --source github-issue-42 \ --priority high \ --agent claude-code \ --task "Implement OAuth2 login flow" \ --files "src/auth/oauth.ts,src/auth/callback.ts" \ --skills "code-generation,test-writing" ``` Claude Code then dispatches from the queue: ```bash # Get next task npx agent-skill-bus dispatch --max 1 # Start it (acquires file locks) npx agent-skill-bus start pr-1710900000-a1b2c3d4 # ... do the work ... # Complete it (releases locks, updates history) npx agent-skill-bus complete pr-1710900000-a1b2c3d4 --result "OAuth2 implemented" ``` ### Pattern 3: DAG pipelines for multi-step tasks Chain tasks with dependencies: ```bash # Step 1: Generate code npx agent-skill-bus enqueue \ --source pipeline \ --agent coder \ --task "Implement feature" \ --dag-id feature-42 # Step 2: Write tests (depends on step 1) npx agent-skill-bus enqueue \ --source pipeline \ --agent tester \ --task "Write tests" \ --depends-on pr-STEP1-ID \ --dag-id feature-42 # Step 3: Code review (depends on step 2) npx agent-skill-bus enqueue \ --source pipeline \ --agent reviewer \ --task "Review implementation" \ --depends-on pr-STEP2-ID \ --dag-id feature-42 ``` ### Pattern 4: Self-healing loop When health checks detect a degrading skill, enqueue a repair task: ```bash # Check for flagged skills FLAGGED=$(npx agent-skill-bus flagged 2>/dev/null) # If any skill is flagged, create a repair task if echo "$FLAGGED" | grep -q '"flagged": true'; then npx agent-skill-bus enqueue \ --source self-improve \ --priority high \ --agent claude-code \ --task "Investigate and fix degrading skill" \ --context "$(echo "$FLAGGED" | head -20)" fi ``` ## Programmatic Usage (ESM) If you prefer to call the API from JavaScript instead of CLI: ```javascript import { SkillMonitor, PromptRequestQueue } from 'agent-skill-bus'; const monitor = new SkillMonitor('./skills/self-improving-skills'); const queue = new PromptRequestQueue('./skills/prompt-request-bus'); // Record a run monitor.recordRun({ agent: 'claude-code', skill: 'bug-fix', task: 'Fix auth timeout', result: 'success', score: 0.95, }); // Check health const health = monitor.analyze(); const flagged = monitor.getFlagged(); // Enqueue a follow-up task if needed if (flagged.length > 0) { queue.enqueue({ source: 'self-improve', priority: 'high', agent: 'claude-code', task: `Repair degrading skill: ${flagged[0].name}`, }); } ``` ## File Structure After Integration ``` your-project/ ├── CLAUDE.md # ← instructions added here ├── skills/ │ ├── prompt-request-bus/ │ │ ├── SKILL.md # Queue skill definition │ │ ├── prompt-request-queue.jsonl # Task queue (auto-managed) │ │ ├── active-locks.jsonl # File locks (auto-managed) │ │ └── prompt-request-history.md # Completed task log │ ├── self-improving-skills/ │ │ ├── SKILL.md # Monitor skill definition │ │ ├── skill-runs.jsonl # Run log (append-only) │ │ ├── skill-health.json # Latest health snapshot │ │ └── skill-improvements.md # Improvement history │ └── knowledge-watcher/ │ ├── SKILL.md # Watcher skill definition │ ├── knowledge-state.json # Source state tracking │ └── knowledge-diffs.jsonl # Detected changes └── ... ``` All data files are JSONL (one JSON object per line) or plain Markdown. They are human-readable, git-friendly, and trivial to parse from any language.