--- name: relay-run-and-review description: Write a Relay Test from a description, run it and read the verdict, check a code change against an App's Tests, inspect a failure, or export evidence. --- # Describe, run, read the verdict 1. Call `relay_get_guide` for `describe` (for a failure, `debug`). Call `relay_health`, then `relay_list_apps` and `relay_list_tests` so you reuse an existing Test instead of writing a duplicate. 2. New coverage: call `relay_create_test` with one sentence (or one step per line) and the `url` or `app`. It saves the Test and returns the exact `relay_run_test` call. Steps written from words need a model key. 3. Call `relay_run_test`. It waits and returns one verdict: passed, failed, blocked or cancelled, with the failing step's expected vs. saw and a screenshot. Pass `targetId` from `relay_list_devices` when several targets are ready. If it returns `running`, call `relay_get_verdict` later. A risk report needs a deliberate repeat with transport `confirm: true`. 4. After changing code, call `relay_check_change` with the App and, when known, the changed `areas` or `testIds`. It runs the relevant ready Tests and returns their verdicts. This is a quick signal, not a merge decision; gated merge checks use the Proof flow (`relay-proof` skill, full profile). 5. For a failure, call `relay_inspect_failure` for evidence and repair options. Propose a repair only when asked; the original attempt is kept. Export a Run with `relay_export_evidence` for a reviewer. 6. For explicitly requested data values, use `relay_repeat_test` for one pilot, then `relay_continue_repeat` with explicit confirmation. Finish with the App/Test/Run IDs, the verdict, and the failing step if any. A human makes screenshot review decisions; agents summarize evidence.