# Choosing what to run > Read by Document, Implement, Implement + Review, Quick Task, Review and Finalize before they write, run or offer tests. Never invoked as a skill. ## Three tiers 1. **Per task:** run the tests that cover what the task touched. 2. **Phase end:** run the phase file's `## Phase Testing` lines, plus tests for any other file the phase changed. 3. **Full suite:** only where the plan put it: a `Full suite:` line in a phase file, or Finalize's offer. A project's `AGENTS.md` saying "run `npm test`" (or any whole-suite command) does not put a full run in a phase. ## Finding the tests that cover a file In this order: 1. `AGENTS.md`'s `## Test map` table (`Area`, `Paths`, `Run`, `Shared`; it may sit behind one link to `.agents-docs/AGENTS-test-map.md`), read with `test-map.mjs read` when the skill ships it, plus the "run single test" command. 2. Otherwise match by name, folder and imports of the changed files; `test-match.mjs` does this when the skill ships it. A runner's own related-tests flag may be used when `AGENTS.md` records it. It is never required. ## Quick checks Lint, a type check, a size check: run them at phase end only when the project or `AGENTS.md` shows them to be fast. ## Shared ground A test helper or fixture, build or test config, a widely imported module, a shared type or schema. A test map row marked `Shared: yes` is shared ground by definition. Touching any of these is the moment to offer a wider run: say what was touched and why it reaches wide, and let the person choose. ## Ask when unsure When you cannot tell whether a check is fast, whether a change is shared ground, or which tests cover a file, ask the person. Never guess. ## Record honestly Say which tests ran and which were skipped. Never report "tests pass" for a run that did not happen.