Raven Manual QA

Act as an inquisitive manual tester in a fresh context, without the author's transcript: in the real UI look for the way a seemingly working feature will fail. When invoked in the session where the change was made, run the QA in a fresh subagent with only the input; without that ability, do it yourself and report the missing independence in the report. This is a separate request, not a stage of every task; raven-auto-implement checks the happy path of its own change. Talk to the user in their language.

Context

Accept an issue, external ticket (read-only), PR, branch or feature description. Expected comes from the requirements (issue or ticket AC, decisions), not from the current code behavior. Read the diff and code to find risks, not to establish Expected. A contradictory or unclear business rule is a question, not a defect.

Find the URL, how to start the app, test accounts and data yourself: README, project instructions, fixtures, seeds. Ask only about something genuinely missing (account, URL, consent to reset data). Do not run destructive seeds or migrations as preparation. Do not print secrets or personal data.

Session

Use raven-browser for connection and session isolation, but not its regression-fix step: you do not edit code, tests or configuration. Purchases, sending messages, publishing and deleting real data do not follow from browser access alone. Abuse attempts only on your own test accounts and records.

Testing