Prp Loop
Developmentprp-loopPrp Loop is an Agent Skill for development published in the Wirasm/PRPs-agentic-eng repository on GitHub: a SKILL.md instruction file an AI agent can load to do this task.
Run the autonomous cyclic PRP pipeline end to end (plan → implement → pr → review, looping review→fix until the PR is clean). Use when the user wants to "ship feature X end to end", "run the full PRP loop", "auto-implement and open a PR for a feature", or invokes /prp-loop.
How to use this skill
Bring this guide into your coding agent with a prompt tailored to the tool you use.
- Open your project in Codex.
- Copy and paste the prompt below.
- Review files and risks before approving any action.
Inspect the Agent Skill at https://github.com/Wirasm/PRPs-agentic-eng/blob/HEAD/.claude/skills/prp-loop/SKILL.md and all supporting files. Report provenance, permissions, scripts, dependencies, and security risks. Propose an exact installation into .agents/skills/prp-loop, but wait for my approval before writing files or running scripts.
Copying this prompt does not install or run the skill. Review third-party files before use.
PRP Loop — autonomous cyclic pipeline
Launch the orchestrator that drives plan → implement → pr → review and loops review → fix until the PR review is clean (or limits are hit). It runs headless claude -p once per stage and tracks progress in .claude/prp-loop.state.json.
Run it
Start a new loop with the user's request as the feature argument:
uv run .claude/PRPs/scripts/prp_loop.py "$ARGUMENTS"
Resume a halted or in-progress loop:
uv run .claude/PRPs/scripts/prp_loop.py --resume
Defaults: --max-cycles 3, --max-implement-iterations 10, base branch auto-detected. Pass --validate "<cmd>" to give the loop an authoritative green check (exit 0 = pass).
Stop after a stage (--until)
Pass --until <stage> (plan | implement | pr | review | fix) to halt once that stage completes:
uv run .claude/PRPs/scripts/prp_loop.py "$ARGUMENTS" --until implement
--until implement runs plan → implement and stops once validations are green and the work is committed — no PR, no review. This is the headless replacement for the old single-session Ralph loop: "grind one plan to green."
UX note: the retired Ralph loop was single-session and interactive (a Stop-hook fed the prompt back in the same session). prp-loop --until implement is headless instead — it drives fresh claude -p sessions per iteration and you resume/inspect via the state file rather than watching it live.
What it does
- plan —
prp-planwrites.claude/PRPs/plans/<feature>.plan.md. - implement —
prp-implementexecutes the plan, looping until all validations pass (bounded by--max-implement-iterations), then commits. - pr —
prp-prpushes the branch and opens the PR (once). - review —
prp-review --agentsreviews the PR and writes a{clean, blocking}verdict. - cycle — if not clean, the blocking findings feed back into a fix pass → push → re-review, up to
--max-cycles. Clean → done.
Safety
- Fully autonomous (
--dangerously-skip-permissions). Operates only on the feature branch — it refuses to PR frommain/master/development/the base branch. - Halts with state preserved on: implement/fix not green after the iteration limit, review still dirty after
--max-cycles, a fix pass with no new commit (no progress), failed push, or any stage error. - Inspect or resume via
.claude/prp-loop.state.json.
Notes
This orchestrator is self-contained and uses no Stop-hook. It owns both loops itself and detects "green" from each stage's VALIDATION: GREEN sentinel (or the --validate command). The PRP skills it calls are invoked verbatim and never modified.