david-li0406/meta-skill-evloving/jikime-workflow-loop

Ralph Loop Workflow

Intelligent iterative code improvement with LSP/AST-grep feedback.

Quick Reference (30 seconds)

Ralph Loop enables autonomous code improvement through:

• Evidence-Based Iteration: Every loop collects LSP/AST-grep diagnostics
• Progressive Improvement: Track improvement rate across iterations
• Smart Termination: Auto-stop on stagnation, max iterations, or completion

Start a Loop

jikime hooks start-loop --task "Fix all TypeScript errors" --max-iterations 10

Completion Markers

Output one of these when done:

• <jikime:done />
• <jikime:complete />
• <jikime>DONE</jikime>

Execution Protocol

Phase 1: Initial Analysis

When loop starts:

1. Assess current state (error/warning counts)
2. Identify scope (files, directories)
3. Prioritize issues (errors before warnings)

Phase 2: Iterative Execution

Each iteration:

1. Focus on highest severity issues first
2. Make targeted fixes (one problem at a time)
3. Verify changes don't introduce new issues
4. Let LSP/AST-grep validate automatically

Phase 3: Completion

Loop completes when:

• All conditions satisfied (e.g., zero errors)
• Completion marker output (<jikime:done />)
• Max iterations reached
• Stagnation detected (no improvement in N iterations)

Feedback Integration

The loop collects diagnostics automatically via hooks:

• PostToolUse (LSP): Error/warning counts per file
• PostToolUse (AST-grep): Security issue detection
• Stop Hook: Evaluates completion, calculates progress

Example Feedback Output

Ralph Loop: CONTINUE | Iteration: 3/10 | Current: 5 error(s), 12 warning(s)
Progress: 45% improvement | Next: Fix 5 remaining error(s)

Completion Criteria Options

Safety Guidelines

Avoid Infinite Loops

1. Always set --max-iterations (default: 10)
2. Don't ignore stagnation warnings
3. Use completion markers when truly done

Best Practices

1. Start with specific, measurable goals
2. Fix one issue type at a time
3. Verify each fix before proceeding
4. Output <jikime:done /> when complete

Command Reference

Start Loop

jikime hooks start-loop \
  --task "Description" \
  --max-iterations 10 \
  --zero-errors \
  --tests-pass

Cancel Loop

jikime hooks cancel-loop

Check Status

The Stop hook automatically reports status after each Claude response.

Integration with DDD

Ralph Loop integrates with Domain-Driven Development:

1. ANALYZE: Loop collects diagnostic data
2. PRESERVE: Each iteration validates existing behavior
3. IMPROVE: Progressive fixes with measurable progress

Use manager-ddd for behavior-preserving iterations.