Rules an agent never breaks
These rules hold even when a test plan step seems to need otherwise. Stop and ask the person who started the run instead.- Never publish a draft gig that has invitations. Publishing sends pending invitations by push, text, and email.
- Never message, @mention, invite, or text a person who is not a test account.
- Never change a feature flag, an org setting, or a role unless the person who started the run asked for that exact change. One PostHog project serves dev, QA, and production.
- Never delete data. Archive what you created, and list it in the report so a person can clean it up.
- Never write to a database through SQL. Read-only
SELECTqueries on the dev database are fine for checking what was stored. - Never print a password, token, or key. Read test credentials with a script that does not echo them.
- Never act in a customer org. Production test runs stay in the test org the run was given.
Pick the backend
To refresh a row in What’s ready, check every merged PR for each ticket in its Tickets column against each branch:
qa is in the next TestFlight build. A web or API change that reached main is live in production.
Drive the API
npm run api calls the API as a seeded persona, so you do not hand-build tokens. It needs npm run worktree:setup -- e2e mcp-ts.
--env dev and a token. Every write against dev is held until you add --yes, so read the printed request before you confirm it.
Drive the mobile app
Use a local iOS simulator first. It is free and it covers most checks. Followdocs/engineering/mobile/agent-screenshots.mdx in the repository for the setup:
- Claim a simulator with
node scripts/sim/claims.mjs claim --note "<test plan> run", so no other session uses it. - Start your own Metro with
npm run shots:metro -- --port 8085. - Sign in with
npm run sim:e2e-login -- --as memberfor the crew account, or--as userfor the owner account. The helper reads the test accounts from.env.e2ewithout printing them. - Open the screen under test with a deep link, for example
soundcheck://events/<eventId>orsoundcheck://events/<eventId>/chats/<chatId>.
npm run revyl:select -- --out <scratchpad>/plan.json <files of the area>picks the saved tests and writes the plan. It runs nothing.node scripts/revyl-flows/run.mjs --plan <scratchpad>/plan.json --out <scratchpad>/results.jsonruns each test in the plan and writes the results. Add--dry-runfirst to see the commands without starting a device.