A person's type label loses to the review's on edit #91

Problem

On #88 the review set type:feat and retitled the pull request feat(…). The maintainer's account then removed type:feat, added type:ci and retitled it ci(…) (2026-10-10 23:23:22); 28 seconds later dilux-bot retitled it feat(…) again (23:23:50), from the review's last verdict, in .github/workflows/claude-review.yml (the step that runs gh pr edit … --title "$retitled"). docs/pull-requests.md says a person's own type:* label wins over the review's. Because the title decides which issue Type the gate accepts (a feat must close a Feature), the pull request stayed blocked although its issue was accepted.

Scope

• The step of claude-review.yml that applies the review's type to the title reads the pull request's labels first and keeps the type of a type:* label added by a person (not by a bot), retitling from that one instead.
• A test of that decision (whose type wins), in a script with its --test, run by the scripts job.

Out of scope

• How the review decides the type from the diff.
• Any other workflow.

Acceptance criteria

• C1: Given a pull request whose review set type:feat, when a person replaces it with type:ci and the review runs again, then the title keeps ci(…).
• C2: Given a pull request with only the review's type:feat, when the review runs, then the title becomes feat(…) as today.
• C3: Given the decision as a script, when its self-test runs, then it covers a person's label winning and the review's label applying.

Criticality

Normal

Spec: approach

Move the "whose type wins" decision out of claude-review.yml into a script, scripts/review-type.sh. It takes the review's type, the bot's login, the pull request's issue timeline (JSON) and its current type:* labels. It prints the type of the last type:* label a person (not a bot) added, if that label is still on the pull request, and the review's type otherwise. The workflow calls it in two places: the verdict step (which already has this logic inline, at the human= line) and the "Correct the title's type" step. The second call reads the labels again just before gh pr edit --title, so a verdict that was reused, or a label a person changed after the verdict was read, can no longer put feat(…) back over a person's type:ci. The script has a --test with fixture timelines, and the scripts job runs it.

Spec: files

• scripts/review-type.sh (new: the decision and its --test)
• .github/workflows/claude-review.yml (the verdict step and the "Correct the title's type" step call the script)
• the scripts job's list of self-tests, if it does not pick up scripts/*.sh --test on its own
• docs/pull-requests.md (one line: the title follows a person's type:* label too)

Spec: tests

• C1 → scripts/review-type.sh (--test: review feat, timeline has the bot labelling type:feat and then a person labelling type:ci, current labels type:ci → ci)
• C2 → scripts/review-type.sh (--test: review feat, only the bot's type:feat in the timeline → feat)
• C3 → scripts/review-type.sh (--test covers both cases above, plus a person's label that was later removed → the review's type)

🤖 AI-generated · Claude Opus 5.5 (Anthropic)