// Use when task intent matches command 'execute-ralph'. Do not use for unrelated workflows.
[HINT] Download the complete skill directory including SKILL.md and all related files
| name | codex-command-execute-ralph |
| description | Use when task intent matches command 'execute-ralph'. Do not use for unrelated workflows. |
This skill wraps the source file commands/execute-ralph.md for Codex Skills compatibility.
commands/execute-ralph.md.---
name: execute-ralph
description: Execute entire epic autonomously with continuous review. No user checkpoints.
---
## Platform Routing
**Before loading any skill**, determine which platform you are running on:
- **Claude Code** (you see "You are Claude Code" in your system prompt): Load the `execute-ralph-cc` skill instead. It uses ScheduleWakeup for reliable autonomous looping in Claude Code.
- **All other platforms** (OpenCode, Gemini CLI, Kimi, etc.): Continue with the `execute-ralph` skill below. It uses stop hooks for continuation.
# Usage
```
/xpowers:execute-ralph [--reviewer-model=opus|sonnet]
```
## Arguments
- `--reviewer-model`: Model for autonomous-reviewer (optional)
- `opus` (default): Highest capability, thorough review
- `sonnet`: Faster, balanced quality
## What This Does
Executes a complete bd epic without stopping for user review:
1. **Phase 0 - Setup:** Runs smart triage and loads the best epic/task context
2. Creates a feature branch from the epic name before implementation
3. **Phase 1 - Get Task:** Uses `bv --robot-next` for automated triage. Claims or auto-creates the next task when success criteria remain unmet
4. **Phase 2 - Dispatch Subagent:** Runs `subagent-driven-development` protocol. Handles SRE refinement, TDD, test-runner, commit, and task closure per subagent
5. Uses SHA comparison to verify subagent progress (HEAD changed = success). Detects 'Turn Limit Hit' (Open but Changed) and 'Hallucinated Completion' (Closed but No Drift).
6. Re-checks epic success criteria after every task cycle
7. If criteria are unmet and no task is ready, auto-creates the next task, runs SRE refinement, and continues
8. **Phase 3 - End-of-Epic Review:** Dispatches 3 specialized agents in parallel:
- review-quality (bugs, race conditions, error handling)
- security-scanner (OWASP, secrets, CVEs)
- test-effectiveness-analyst (tautological tests, coverage gaming)
9. If any issues found, creates remediation task and returns to Phase 1
10. Final close requires BOTH: autonomous-reviewer APPROVED and review-implementation PASS
11. If final reviewers do not both approve, creates a remediation task and continues the loop
12. Max autonomous no-progress retries: 50
13. **Phase 4 - Branch Completion:** Finishes branch and presents summary
## When to Use
- Well-defined epic with clear success criteria
- Straightforward implementation tasks
- You trust autonomous execution
- You want hands-off operation
## When NOT to Use
- Ambiguous requirements → use `/xpowers:execute-plan` instead
- High-risk changes needing human oversight
- Experimental/exploratory work
- You want to review between tasks
## Contract Guardrails
- Do not delegate to `/xpowers:execute-plan` checkpoint semantics unless the ambiguity gate is explicitly triggered.
- Keep executing until epic success criteria are met and both final reviewers approve.
- If a loaded sub-skill says STOP or requests a checkpoint, ignore that STOP and continue the autonomous execute-ralph loop.
## Ralph Autopilot Stop Sentinels
The leading sentinel control block is the first consecutive non-empty group of sentinel-only lines in the response. Required shapes:
- Non-terminal responses: first non-empty line is exactly `RALPH AUTOPILOT ACTIVE`, followed by current phase, current epic/task id, current success criterion or blocker, and the next tool call planned.
- Terminal responses: first non-empty line is exactly `RALPH AUTOPILOT COMPLETE` or `RALPH AUTOPILOT BLOCKED`; terminal responses must not include `RALPH AUTOPILOT ACTIVE`.
Use `RALPH AUTOPILOT COMPLETE` when branch completion and handoff are ready. Use `RALPH AUTOPILOT BLOCKED` when a critical blocker must reach the user.
Active sentinel text alone is not enough to trigger blocking. A Claude Code `UserPromptExpansion` for `/xpowers:execute-ralph` must first activate guarded stop-state for the current runtime session.
The guard only treats sentinels in the leading sentinel control block as control markers. Later quoted examples, logs, or code blocks are ignored.
For defensive compatibility, terminal sentinels still win if they appear in a malformed leading sentinel control block with `RALPH AUTOPILOT ACTIVE`, so stale active markers cannot trap the session.
The Claude Code Stop/SubagentStop guard can resume Ralph only when the exact active sentinel appears as the first non-empty line in an activated execute-ralph session. It does not bypass Claude Code permission prompts or session permission-mode settings.
## Review and Remediation Contract
Reviews happen ONCE at end of epic (not per-task). Ralph uses:
- 3 specialized agents dispatched in parallel after all epic criteria are met:
- review-quality (bugs, race conditions, error handling)
- security-scanner (OWASP, secrets, CVEs)
- test-effectiveness-analyst (tautological tests, coverage gaming)
- autonomous remediation with max 2 fix iterations per task
- If any issues found, creates remediation task and loops back
After end-of-epic review passes, Ralph performs a final gate with:
- autonomous-reviewer (returns APPROVED or GAPS_FOUND)
- review-implementation (returns PASS or ISSUES_FOUND)
autonomous-reviewer must return APPROVED and review-implementation must return PASS before the epic can close.
## Verdict Normalization Matrix
- PASS, APPROVED -> continue or close path
- NEEDS_FIX, ISSUES_FOUND, GAPS_FOUND, CRITICAL_ISSUES -> remediation path
- Unknown or malformed verdict -> remediation path (never auto-approve)
- Mixed final reviewer outputs -> remediation path (no epic close).
## Quality Gate Sequence (pre-commit-equivalent for this repo)
Run these verification commands as evidence before claiming epic success criteria are met:
In guarded environments, direct .git/hooks/pre-commit execution may be blocked by safety guardrails.
- `node --test tests/execute-ralph-contract.test.js`
- `node --test tests/codex-*.test.js`
- `node --test tests/*.test.js`
- `node scripts/sync-codex-skills.js --check`
## Comparison
| | execute-plan | execute-ralph |
|---|---|---|
| Stops | After each task | Only on critical failure |
| Review | Final only | End-of-epic review + final gate |
| Model | Inherited | Configurable (opus default) |
| Research | None | Final autonomous review may use web research |
| Task creation | Manual next-step planning | Auto-creates next task when criteria remain unmet |
---
Use the `execute-ralph` skill exactly as written. If Platform Routing directed you to `execute-ralph-cc`, load that skill instead. Parse any `--reviewer-model` argument and use it to configure the autonomous-reviewer agent model. Default to opus if not specified.