name: autonomous-loop description: Autonomous development loop patterns for iterative self-improvement. Auto-triggers when implementing features autonomously, fixing bugs in a loop, or running until completion. Based on frankbria/ralph-claude-code.

Autonomous Development Loop

Based on Geoffrey Huntley's Ralph technique for Claude Code.

For full autonomous mode, use /ralph. This skill provides the underlying patterns.

Core Principle: Dual-Condition Exit Gate

Never exit prematurely. Exit requires BOTH conditions:

1. Completion Indicators ≥ 2 - Heuristic detection from your work
2. Explicit EXIT_SIGNAL: true - Your conscious declaration

If only ONE is true → KEEP GOING

This innovation (introduced in ralph-claude-code v0.9.9) prevents false exits when completion language appears during productive work.

Completion Indicators

Count how many apply each loop:

Need ≥2 indicators + explicit EXIT_SIGNAL: true to exit

Circuit Breaker Pattern

Track your own progress and halt when stuck:

Halt Thresholds

Progress Detection

Track mentally each loop:

• Files modified (0 = no progress)
• Tests changed (pass/fail delta)
• Tasks completed (fix plan updates)
• Errors encountered (same vs different)

Structured Status Reporting

Every response MUST end with:

## Status Report

STATUS: IN_PROGRESS | COMPLETE | BLOCKED
LOOP: [N]
CIRCUIT: CLOSED | HALF_OPEN | OPEN
EXIT_SIGNAL: false | true

TASKS_COMPLETED: [what you finished]
FILES_MODIFIED: [count and list]
TESTS: [X/Y passing]
ERRORS: [count or "none"]

PROGRESS_INDICATORS:
- [x] Files changed this loop
- [ ] Tests improved
- [ ] Tasks marked complete
- [ ] No repeated errors

NEXT: [next action or "done"]

Exit Signal Checklist

Before setting EXIT_SIGNAL: true, verify ALL:

□ All fix_plan.md items complete (or task done if no plan)
□ All tests passing (or code works if no tests)
□ No execution errors
□ All requirements implemented
□ No meaningful work remaining

If ANY unchecked → EXIT_SIGNAL: false

Work Focus Hierarchy

Anti-Patterns to Avoid

• Premature exit: Tests pass but tasks remain
• Infinite loops: Not detecting repeated failures
• Busy work: Testing without implementing
• Silent failures: Not reporting errors clearly
• Scope creep: Adding unplanned work
• Analysis paralysis: Over-researching instead of implementing

Fix Plan Integration

Maintain fix_plan.md:

## High Priority

- [ ] Critical task

## Medium Priority

- [ ] Supporting task

## Completed

- [x] Done item

Empty fix plan + passing tests = consider exit

Example: Why Dual-Gate Matters

Loop 5:
  Output: "Auth complete! Moving to sessions."
  Indicators: 3 (tests pass, "complete", no errors)
  EXIT_SIGNAL: false (Claude continuing)
  Result: CONTINUE ✅ (respects intent)

Loop 8:
  Output: "All done, tests green, ready for review."
  Indicators: 4 (all signals present)
  EXIT_SIGNAL: true (Claude done)
  Result: EXIT ✅ (both conditions met)

Recovery Protocols

If BLOCKED

1. State: "I am blocked because [reason]"
2. List what you tried (min 2 approaches)
3. Suggest alternatives
4. Set STATUS: BLOCKED, EXIT_SIGNAL: false
5. Wait for input

If Circuit Breaker OPENS

1. State: "Circuit breaker OPEN - halting"
2. Summarize accomplishments
3. Describe stagnation pattern
4. Set CIRCUIT: OPEN, EXIT_SIGNAL: false
5. Recommend next steps

Never infinite loop on a blocker.