| name | check |
| description | Goal-loop quality gate — run QA/tests and fix until green, with real output as proof |
| argument-hint | [qa|test|all] |
| allowed-tools | Bash(make *), Bash(docker run:*), Bash(docker compose:*), Read, Edit, Grep, Glob |
Drive the project to a verified-green state. Parse $ARGUMENTS: qa, test, or all (default all).
This is a goal-loop: the completion condition is a passing run with its real output shown — not an assertion of success.
Step 1 — Pick the target
qa → make qatest → make testall (default) → make qa, then make test
On a dev machine make qa OOM-kills at the Rector leg (exit 137) — this is the 768M cap, not your diff. When it does, do not conclude the branch is broken: run the QA legs individually via the one-off containers in .claude/local-qa.md and paste their per-leg output as the evidence. make test likewise needs the full stack; the Functional/PHPUnit and Playwright recipes are there too.
Step 2 — Run and inspect
Run the target. If it passes, go to Step 4.
If it fails, read the real error output. Identify the failing tool (PHP-CS-Fixer, Rector, PHPStan, ESLint, Prettier, TS, PHPUnit, Playwright) and the precise cause.
Step 3 — Fix and retry (max 3 runs total)
Apply the smallest fix that makes the check pass — lint/format/type errors, broken assertions, missing typegen. Do not change feature behavior to force a pass: if a test reveals a real regression, stop and report it instead of editing the test to match.
Re-run the target. Repeat up to 2 more times (3 runs total including Step 2). After 3 failed runs, stop and list the remaining errors verbatim — do not loop further.
Step 4 — Report with proof
Declare green only with the real command output pasted as evidence. If all was requested, both make qa and make test must be green.
If you stopped at the attempt cap or on a suspected real regression, say so explicitly and show the outstanding errors.
