finish-phase

Run the full gate battery with npm run check:phase, which executes lint, type checks, the test suite under coverage, mutation testing over the diff, and selective e2e scenarios. Every gate must pass; a red gate is re-run alone, never by repeating the whole chain.

This workflow requires 85% for bot source code and 80% for tooling scripts, run as two separate Stryker families so tooling mutants cannot hide inside a high aggregate score. A dropped score means tests assert too little — strengthen them, never lower the bar.

Tag only when the phase changed what a player or operator gets, because a tag restarts the bot and pays the full release check. Tooling, documents, and process changes go to main and ride the next tag.

npm install on one platform can prune optional packages another platform requires, producing a lock file that passes locally but fails on the runner. Run npm ci --dry-run before committing and diff the lock file for unexplained removals.

No. The whole-diff review is always delegated to a phase-reviewer subagent because the author still holds the reasoning that made the code look right. Stop editing before launching it, since the reviewer reads live files rather than a snapshot.

Do not load it during stage 1 framing, when a list of changes is only being recognized as a phase. It applies once code is written and gates are due — wrapping up a phase, cutting a release, or checking releasability.