fix: the bug reproduced from #720 by dilux-bot[bot] · Pull Request #723 · DiluxOne/.github

dilux-bot · 2026-11-11T21:30:44Z

📝 What changes

The review labels a pull request type:docs (and the "Correct the title's type" step writes docs: into its title) from the model's reading alone, while scripts/conventions.sh rejects a docs title as soon as one changed file is outside scripts/surface.py --docs-paths <kind>. A diff with a doc and a test is therefore labelled docs by one and refused by the other. The fix bounds the review's type by the same paths, in the one place both workflow steps already call, scripts/review-type.sh.

scripts/review-type.sh gains an optional trailing --diff KIND BASE HEAD: when the review's TYPE is docs and git diff --no-renames --name-only BASE...HEAD has a file that matches none of the globs python3 surface.py --docs-paths KIND prints (matched with the same shell case as conventions.sh), the review's type becomes chore. A type:* label a person set still wins, as today (it is theirs; the conventions check then tells them). If the docs paths cannot be read, the script fails, like an unreadable timeline. Without --diff it behaves exactly as now, so older callers keep working.
.github/workflows/claude-review.yml: the Verdict step and the "Correct the title's type" step pass --diff "$KIND" "$BASE" "$SHA" (KIND from steps.policy.outputs.kind, BASE the pull request's base SHA, the checkout already has full history); the review prompt's line on type says docs only when every file is documentation.
Self-tests in review-type.sh --test for: docs with a doc and a test → chore; docs with docs only → docs; a kind's own docs path (wordpress-plugin readme.txt) → docs; a non-docs type untouched; a person's type:docs label still stands; unreadable kind fails.
The reproduction test is added to the default branch exactly as written.

Option to decide: what the review's docs becomes when the diff has more than documentation.

(a) chore (recommended): closes any accepted issue, never claims new behaviour, and the reviewer cannot be second-guessed into feat/fix by a script.
(b) test when every other file is a test, else chore: closer to the Conventional Commits meaning, but needs a notion of "test path" per kind that the packs do not have yet.

Nothing widens a permission, secret, lane or limit.

💡 Why

Closes #721. The autopilot works it here in the open: one commit per criterion, each box ticked once its commit is pushed.

🎯 Acceptance criteria

C1 Given the reproduction test tests/repro/test_issue_720.py exactly as written, when it runs on the fix, then it passes, where it failed on b25a3cf.
Test: tests/repro/test_issue_720.py

🧭 Decisions

2026-11-11: The review's docs on a diff with files outside the docs paths becomes chore (the Spec's recommended option a): it closes any accepted issue and needs no per-kind notion of a test path.

🙌 Credits

🧪 How I tested it

The reproduction test, run in the sandbox at HEAD (repro-run.sh run-ref HEAD tests/repro/test_issue_720.py): outcome=passed. It failed on b25a3cf.
bash scripts/review-type.sh --test: every check passes, including the new ones for fix: a docs pull request that carries its test never passes conventions #720. They cover a doc plus a test becoming chore, docs only staying docs, wordpress-plugin's readme.txt counting as docs (and not counting without that kind), a non-docs type left as it was, a person's type:docs label still winning, an unreadable kind failing, and --diff without HEAD being refused. shellcheck is clean.
Every self-test the pull-request workflow's scripts job runs, plus pytest on tests/. Five self-tests fail here, and all five fail the same way on the base: plugin-i18n and design-lint need gettext and stylelint, which this runner lacks, and review-brief, autopilot.py and runner-vm/dx-runner-slot fail because of this runner's environment. The one pytest failure is in autopilot.py, which this change does not touch. actionlint is not installed here, so claude-review.yml was only checked to parse as YAML.
This repository has no make pre-pr.

The reproduction test tests/repro/test_issue_720.py at f35229f, the pull request's head, is the one that failed on the default branch (its SHA-256 is the sealed PRD's). The agent ran it in the sandbox (a clean copy of its commit; no network, no secrets) before finishing, and gate G7 runs it again the workflow way it was written, at the head, before it merges.