prp-loop

> **Arguments:** `$ARGUMENTS` (and `$1`, `$2`, ...) refer to the arguments given when this skill was invoked. Take them from the user's request; if absent, infer them from the conversation. # PRP Loop — autonomous cyclic pipeline Launch the orchestrator that drives `plan → implement → pr → review` and loops `review → fix` until the PR review is clean (or limits are hit). It runs headless `codex exec` once per stage and tracks progress in `.claude/prp-loop.state.json`. ## Run it Start a new loop with the user's request as the feature argument: ```bash uv run .agents/skills/prp-loop/scripts/prp_loop.py "$ARGUMENTS" --cli codex ``` Resume a halted or in-progress loop: ```bash uv run .agents/skills/prp-loop/scripts/prp_loop.py --resume --cli codex ``` 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: ```bash uv run .agents/skills/prp-loop/scripts/prp_loop.py "$ARGUMENTS" --cli codex --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 `codex exec` sessions per iteration and you resume/inspect via the state file rather than watching it live. ## What it does 1. **plan** — `prp-plan` writes `.claude/PRPs/plans/<feature>.plan.md`. 2. **implement** — `prp-implement` executes the plan, looping until all validations pass (bounded by `--max-implement-iterations`), then commits. 3. **pr** — `prp-pr` pushes the branch and opens the PR (once). 4. **review** — `prp-review --agents` reviews the PR and writes a `{clean, blocking}` verdict. 5. **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 from `main`/`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.