Goblin1024/ralph

Ralph Skill

Purpose

Guaranteed task completion through persistent iteration and mandatory verification.

Activation

Trigger: $ralph <task> or $ralph "continue"

Core Principle

The task is not complete until verification proves it.

Workflow

Phase 1: Context Loading

1. Check for approved plan in .omk/plans/
2. Load context snapshot from .omk/context/
3. Review current state
4. Identify what's already done vs. remaining

Phase 2: Execution Loop

while not complete:
  1. Assess progress
  2. Delegate parallel tasks using Agent tool
  3. Run long operations in background
  4. Verify results
  5. Update TODO list

Phase 3: Verification (Mandatory)

Before claiming completion, you MUST submit evidence via MCP tools:

1. Submit Test Evidence: After running npm test, call:

   omk_submit_evidence({skill:"ralph", step:"tests_passed", phase:"verifying", evidenceType:"command_output", command:"npm test", output:"<test output>", exitCode:0})
   

2. Submit Build Evidence: After running npm run build, call:

   omk_submit_evidence({skill:"ralph", step:"build_passed", phase:"verifying", evidenceType:"command_output", command:"npm run build", output:"<build output>", exitCode:0})
   

3. Submit Lint Evidence: After running npm run lint, call:

   omk_submit_evidence({skill:"ralph", step:"lint_clean", phase:"verifying", evidenceType:"command_output", command:"npm run lint", output:"<lint output>", exitCode:0})
   

4. Submit Type Evidence: After running tsc --noEmit, call:

   omk_submit_evidence({skill:"ralph", step:"types_clean", phase:"verifying", evidenceType:"command_output", command:"tsc --noEmit", output:"<tsc output>", exitCode:0})
   

5. Check Required Evidence: Before advancing phase, verify all evidence is present:

   omk_list_required_evidence({skill:"ralph", phase:"verifying"})
   

6. TODO Check:

   • Zero pending items
   • Zero in-progress items

7. Architect Review:

   • Delegate to architect subagent
   • After approval, submit:

   omk_submit_evidence({skill:"ralph", step:"architect_approved", phase:"completing", evidenceType:"review_signature", reviewerAgent:"architect", reviewResult:"approved"})
   

Phase 4: Completion

1. Run final verification and submit all evidence
2. Submit diff evidence:

   omk_submit_evidence({skill:"ralph", step:"diff_recorded", phase:"completing", evidenceType:"diff_record", filesModified:["src/foo.ts"], linesAdded:10, linesRemoved:2})
   

3. Call omk_assert_phase({skill:"ralph", phase:"completed"}) to confirm transition is allowed
4. Update state to "completed"
5. Run $cancel to clean up
6. Report success with evidence references

Execution Policy

• Parallel First: Delegate independent tasks simultaneously
• Background for Long Ops: Use run_in_background: true for builds/tests
• Evidence Required: No "should work" - prove it works
• No Scope Reduction: Complete the full task, not a subset
• Fresh Verification: Re-run checks, don't trust old results

State Management

• On start: Create ralph-state.json with phase "executing", iteration: 1
• Each iteration: Update iteration count, progress
• On verification: Phase "verifying"
• On fix: Phase "fixing"
• On complete: Phase "complete", active: false

Progress Tracking

Use TODO list to track:

[ ] Task 1
[ ] Task 2
[x] Task 3 (completed)

Subagent Delegation

Use Kimi's Agent tool:

Agent(description="Implement auth module", prompt="...")
Agent(description="Write tests", prompt="...", run_in_background=true)
Agent(description="Review code", prompt="...")

Examples

Basic Execution:

$ralph "implement user login"
→ Load approved plan
→ Delegate implementation to coder subagent
→ Delegate tests to test-engineer subagent
→ Collect results
→ Run verification
→ Report completion

Resume After Interruption:

$ralph "continue"
→ Read current state
→ Identify incomplete tasks
→ Continue from where we left off
→ Complete and verify

With PRD:

$ralph --prd "build the authentication system"
→ Check for prd-authentication.md
→ Follow the approved plan
→ Verify each milestone
→ Complete when all user stories pass

Stop Conditions

• Task is complete with verification
• User says "$cancel"
• Fundamental blocker (external service down, missing credentials)
• Same error recurs 3+ times (escalate to user)

Completion Checklist

• [ ] All requirements met
• [ ] Tests pass (fresh run)
• [ ] Build succeeds (fresh run)
• [ ] No type errors
• [ ] No lint errors
• [ ] TODO list empty
• [ ] Architect verification passed
• [ ] State cleaned up

Notes

• Ralph NEVER gives up until complete or explicitly cancelled
• Verification is mandatory, not optional
• Use parallel execution for speed
• Keep user informed of progress
• Escalate blockers, don't guess