e2e loop

The unit suite proves the prompt graph is well-formed. It cannot prove a review still comes out right. This closes that gap: run → grade → diagnose → fix → re-run, with the running and the grading done by DIFFERENT agents so neither marks its own work.

Costs real money and posts to a real PR on open-pr-test. Never start a round the user did not ask for.

Why subagents rather than this session

A dev session already knows which defects the fixture plants and which rule was just edited. Reviewing from that context tests the session's memory, not the prompts. Each round therefore spawns fresh agents that were told nothing beyond what a real user's session gets.

Round budget

Ask for max_rounds if the user did not say; default 2. Stop early when every checklist row passes, or when a round produces no NEW passing row — a loop that keeps editing without moving the score is the failure mode this skill is most likely to hit.

Preflight

1. scripts/check.sh <base-ref> must be green. A red suite makes every later verdict unreadable.
2. e2e/bootstrap.sh --pr <n> [--vendor …] if no fixture PR is open for this round.
3. python3 scripts/vendor_lint.py --pr <n> — every documented Fetch command must run. A broken vendor command wastes a whole round: the review fails at fetch and every checklist row reads fail for a reason that has nothing to do with the rules being tested. Seconds, and free.
4. Note the fixture URL. Every later stage refers to it.

Stage 1 — run the review (subagent, fresh)

Spawn a subagent whose whole brief is:

Read <repo>/src/commands/review.md VERBATIM and follow it against <fixture PR url>. Wherever it says ${CLAUDE_PLUGIN_ROOT}, substitute <repo>/src — you are exercising the WORKING TREE, not the installed plugin. You have no other instructions and no knowledge of what the PR contains.

FORBIDDEN: paraphrasing the command file into the brief, hinting at the planted defects, naming the stacks involved, or telling it what a good review looks like. Every one of those invalidates the round.

The plugin's own rule already requires a delegated run to read the command file verbatim, so this stage is the same path a real user's subagent takes.

Covering /open-pr:fix in the same round: e2e/bootstrap.sh --pr <n> --checkout --clone-dir <dir> gives a working copy on the fixture branch without touching the remote; the fix reads the convention from the same <data>/<repo>/ memory the review wrote. FORBIDDEN: re-running the seeding mode for that — it force-pushes the branch the posted review is anchored to.

Stage 2 — grade it (a DIFFERENT subagent, fresh)

Spawn a second subagent with e2e/checklist.md, the fixture URL, and read access to the posted review. Its brief: for EVERY row and checkbox, return pass / fail / partial plus the exact quote from the review that justifies the verdict, and nothing else.

FORBIDDEN: this agent proposing fixes, or the Stage 1 agent grading itself — a runner asked to grade its own output rationalises rather than reports.

A row with no quote to back it is fail, not partial. Missing evidence IS the finding.

Stage 3 — diagnose each failure

For every fail / partial, name the ONE file that owns the rule that should have caught it (src/core/review-criteria.md, a src/templates/<stack>.md, a src/cases/*.md, a vendor group…). Then classify:

FORBIDDEN: a fix that mentions the fixture, its filenames or its defects. A prompt that only works on open-pr-test is worse than the failure it hides.

Stage 4 — apply, then re-verify

Apply the fixes. Then scripts/check.sh <base-ref>:

• suite red → the fix broke an invariant; fix that before another round
• context cost up → report it per the token rule in CLAUDE.md; a passing checklist does not buy a budget regression

FORBIDDEN: editing e2e/checklist.md to make a failure pass, unless the user agreed in Stage 3 that the expectation itself was wrong. That is grading your own homework.

Leave the changes uncommitted. Report them; committing and pushing stay the user's call.

Stage 5 — close the round

1. e2e/bootstrap.sh --pr <n> --teardown — closes the fixture PR/MR, leaving its link recorded on the project PR as the evidence the round happened.
2. Report, in this shape:

round <k>/<max>   <passed>/<total> checklist rows
fixed      <row> ← <file that owned it>
still open <row> ← why, and what it would take
token      <mean delta vs base>, per-scenario deltas if any moved
uncommitted <files>

3. Rows still failing && rounds left && the last round moved the score → bootstrap a new fixture and go again. Score unmoved ⇒ stop and hand back: two rounds of edits with no new row passing means the diagnosis is wrong, and a third round of guessing makes the prompts worse, not better.