| name | verify |
| description | Prove a change actually works before it ships. Use whenever a change is done, after implementing a feature/fix, before opening a PR, or when the user says verify/check this works/is this actually working. Tests passing is not sufficient; this skill requires real entry-point evidence. |
Verify
Core principle: tests passing ≠ feature working. Verification means executing the changed behaviour through the same boundary a user, service, or downstream caller uses.
Process
1. Restate the contract
Briefly state what the change is supposed to do from the user's perspective. If you cannot state it, ask or mark NOT VERIFIED.
2. Run the existing gates
Run the repo's relevant gates, escalating from cheapest to broadest:
- format/lint/typecheck/build
- focused tests for touched area
- full test suite when practical before PR/merge
If a gate was already failing before the change, record that evidence and do not pretend it passed.
3. Drive the real entry point
Exercise the actual surface:
- CLI → run the installed/local binary or package command with real args.
- API/service → start the service, call the endpoint, inspect body and side effects.
- UI → run the app, interact with the changed path, capture screenshots if useful.
- Workflow/worker/pipeline → run a small real input and inspect output artifacts/rows/state.
- Library → write and run a throwaway consumer script.
Capture command + representative output. Screenshots or logs count only when tied to the command/path that produced them.
4. Edge/failure sweep
Try the highest-value applicable cases, not an infinite checklist:
- empty/zero/nil input
- boundary/off-by-one
- duplicate/repeated invocation/idempotency
- malformed dependency response or timeout
- unicode/long strings if strings cross a boundary
- permissions/auth for new access paths
- backwards compatibility for old config/data/callers
5. Reliability sweep
Check that new failure modes are diagnosable:
- errors carry context and are not swallowed
- external calls have timeout/retry/cancellation where appropriate
- logs are useful but contain no secrets/PII
- resources are cleaned up
6. Verdict
End with exactly one:
- VERIFIED — evidence proves the contract through the real entry point.
- NOT VERIFIED — something failed, could not run, or only tests/builds ran.
Partial verification is NOT VERIFIED with notes.
Gotchas
- Do not verify by re-reading code.
- Do not count executor claims or CI summaries as evidence unless you retrieved the underlying output.
- Any code change after VERIFIED invalidates the verdict; re-run the relevant verification.
