Playwright Fix Verifier

Takes defect IDs, digs up each one's minimal repro, re-runs it against the current build, and updates the ledger with an evidence-backed verdict. Closes the loop that otherwise dead-ends in "the fix is deployed, probably."

Relationship to the other skills

• bug-triage produces the input this skill needs: a defect record with minimal repro steps and original evidence. If a defect ID has no repro on file, route through triage first — improvising a repro from a one-line symptom description verifies your guess, not the defect.
• Repros that depend on forced API failures reuse network-assertion techniques (mocking a 500, a distinctive GraphQL error) — a backend that has been fixed won't fail on demand, so the mock is the repro.
• report-generator consumes the updated ledger; verdicts from this skill should land where that skill will find them.
• After a verdict, optionally transition the corresponding Jira ticket via jira-connector (dry-run first, as ever). A REGRESSED verdict is a new ticket, not a comment on the closed one — hand it to bug-creator, which links it back to the original.

Core principles (and why)

Re-run the recorded repro, not your memory of the symptom. The bug bundle's "Minimal repro steps" section is the contract — same preconditions, same tenant class, same data shape, same forced-state setup. A verification that reproduces something like the original conditions produces a verdict about a different bug.

Pin the build before anything else. Read the app's version identifiers first and record them next to the verdict. Staging builds can change several times a day here; a verdict without a build number expires silently and can't be compared with the last one. If the build is identical to the one where the defect was confirmed broken, say so — re-verifying the same build mostly measures flakiness, and the user may want to wait for the deploy instead.

Match or exceed the original rigor. If the defect was intermittent or the previous verification ran the repro five times (forced and natural paths), a single clean pass now is not comparable evidence. Read what the last verification actually did and do at least that. One trial is only sufficient for defects that were deterministic in both directions.