/ouroboros:ralph
Persistent self-referential loop until verification passes. "The boulder never stops."
Usage
ooo ralph "<your request>"
/ouroboros:ralph "<your request>"Trigger keywords: "ralph", "don't stop", "must complete", "until it works", "keep going"
How It Works
Ralph mode includes parallel execution + automatic verification:
- Execute (parallel where possible)
- Independent tasks run concurrently - Dependency-aware scheduling
- Verify (verifier)
- Check completion - Validate tests pass - Measure drift
- Loop (if failed)
- Analyze failure - Fix issues - Repeat from step 1
Instructions
When the user invokes this skill:
- Parse the request: Extract what needs to be done
- Initialize loop:
- Generate a session_id (UUID) - Track iteration, verification_history in conversation context - No file I/O needed — evolve_step stores all execution data in EventStore
- Enter the loop (non-blocking background execution):
while iteration < max_iterations: # Start evolve_step in background — returns immediately job = await start_evolve_step(lineage_id, seed_content, execute=true) job_id = job.meta["job_id"] cursor = job.meta["cursor"] # Poll for progress (non-blocking, shows intermediate state) # Use timeout_seconds=120 (2-min long-poll) to reduce context consumption # Only report when AC completed count changes (level-based polling) prev_completed = 0 while not terminal: wait_result = await job_wait(job_id, cursor, timeout_seconds=120) cursor = wait_result.meta["cursor"] status = wait_result.meta["status"] # Parse AC Progress from response, report only on level completion current_completed = <parse AC completed from response> if current_completed > prev_completed: # Report progress concisely (one line per poll) print: [Level complete] AC: {current_completed}/{total} | Phase: {phase} prev_completed = current_completed terminal = status in ("completed", "failed", "cancelled") # Fetch final result result = await job_result(job_id) # Parse QA from evolve_step response text # (EvolveStepHandler runs QA internally and appends verdict) verification.passed = (qa_verdict == "pass") verification.score = qa_score # Record in conversation context verification_history.append({"iteration": iteration, "passed": verification.passed, "score": verification.score, "verdict": qa_verdict}) if verification.passed: # SUCCESS break # Failed - analyze and continue iteration += 1 if iteration >= max_iterations: # Max iterations reached breakTool mapping:
- start_evolve_step = ouroboros_start_evolve_step - job_wait = ouroboros_job_wait - job_result = ouroboros_job_result
- On termination, display a next-step:
- Success (QA passed): Next: ooo evaluate for formal 3-stage verification - Max iterations reached: Next: ooo interview to re-examine the problem — or ooo unstuck to try a different approach
- Report progress each iteration:
[Ralph Iteration <i>/<max>] Execution complete. Running QA... QA Verdict: <PASS/REVISE/FAIL> (score: <score>) Differences: - <difference 1> - <difference 2> Suggestions: - <suggestion 1> - <suggestion 2> The boulder never stops. Continuing... - Handle interruption:
- If user says "stop": exit gracefully - If user says "continue": call ouroboros_query_events(aggregate_id=<lineage_id>) to reconstruct iteration history from EventStore
The Boulder Never Stops
This is the key phrase. Ralph does not give up:
- Each failure is data for the next attempt
- Verification drives the loop
- Only complete success or max iterations stops it
Example
User: ooo ralph fix all failing tests
[Ralph Iteration 1/10]
Started background execution (job_abc123)
Polling progress...
Phase: Executing | AC Progress: 1/3
Phase: Executing | AC Progress: 2/3
Phase: Executing | AC Progress: 3/3
Execution complete. Fetching result...
QA Verdict: REVISE (score: 0.65)
Differences:
- 3 tests still failing
- Type errors in src/api.py
Suggestions:
- Fix type annotations in api.py before retrying
The boulder never stops. Continuing...
[Ralph Iteration 2/10]
Executing in parallel...
Fixing remaining issues...
Running QA...
QA Verdict: REVISE (score: 0.85)
Differences:
- 1 test edge case failing
Suggestions:
- Add boundary check in parse_input()
The boulder never stops. Continuing...
[Ralph Iteration 3/10]
Executing in parallel...
Fixing edge case...
Running QA...
QA Verdict: PASS (score: 1.0)
Ralph COMPLETE
==============
Request: Fix all failing tests
Iterations: 3
QA History:
- Iteration 1: REVISE (0.65)
- Iteration 2: REVISE (0.85)
- Iteration 3: PASS (1.0)
All tests passing. Build successful.
Next: `ooo evaluate` for formal 3-stage verification