qa-heallisted

# qa-heal The official `playwright-test-healer` is useful and will be used. It also cannot be trusted unsupervised, and this is not a suspicion — it is what its own agent definition instructs it to do: > "Code Remediation: … Fixing assertions and **expected values**" > "You will continue this process **until the test runs successfully** without any failures or errors." > "If the error persists … mark this test as **test.fixme()** so that it is skipped." > "Do not ask user questions … do the most reasonable thing possible to **pass the test**." An agent told to make tests pass, given permission to edit expected values, will eventually edit an expected value. This stage exists to catch that moment. ## Procedure 1. **Commit the generated specs first.** That commit is the baseline. Without it the guard has nothing to compare against and will pass everything. 2. Run the healer. 3. Adjudicate: ```bash node dist/tools/heal-guard.js --base "$BASELINE" $(git diff --name-only "$BASELINE" -- '*.spec.ts') ``` 4. **Exit 0** — the healing was locator/wait work. Keep it. 5. **Non-zero** — revert those files and convert each rejected edit into a finding. Do not re-run the healer on the same failure with different wording. The guard's verdict is not a puzzle to route around. ## The distinction that matters A test fails for one of two reasons: - **the test is wrong** — bad locator, bad wait, bad setup → *heal it* - **the app is wrong, or the spec is wrong** — the asserted va