Verify before done
Rewrite of obra/superpowers verification-before-completion @ b36e0829. Compress, not a paste. Id: verify-before-done.
Notify only when the work is landed+verified: see SDLC Definition of done.
The proving command is always re-run. Verifier mint, resume, and never the builder: see SDLC Item agents.
Iron law
No completion claims without fresh verification evidence.
If you have not run the proving command in this turn, you cannot claim pass, fixed, green, or done.
Gate
Before any success claim, commit, PR, or “next task”:
- Identify the command (or check) that proves the claim.
- Run it fresh and complete — not a remembered earlier run.
- Read full output, exit code, failure count.
- Match output to the claim. If not, report actual status with evidence.
- Only then claim success — and cite the evidence.
Skip a step = claiming, not verifying.
Always / Ask first / Never
| Always | Ask first | Never |
|---|---|---|
| Fresh command for this claim | Softening verify for throwaway spikes you will delete | “Should pass” / “probably” / “seems” |
| Full suite or scoped command that actually covers the claim | Partial checks when full suite is expensive — name what you skipped | Trusting agent “success” without VCS/diff or independent run |
| Red-green proof for new regression tests | Operator in the loop when evidence is ambiguous or flaky | Satisfaction (“Great!”, “Done!”) before evidence |
| Line-by-line checklist for requirements DoD | Unsupervised push to protected / prod deploy / secret rotate | Extrapolating from linter to build, or build to product fix |
Operator in the loop: run verification for evidence — bots oversee workers. If security or CI marks a check flaky/MEDIUM, do not self-clear — show the operator the output and wait.
Claim → evidence
| Claim | Requires | Not enough |
|---|---|---|
| Tests pass | Test run: 0 failures | Prior run, “should pass” |
| Linter clean | Linter: 0 errors | Partial path |
| Build succeeds | Build exit 0 | Linter green |
| Bug fixed | Symptom repro now passes | Code changed |
| Regression test works | Fail without fix, pass with fix | Passes once |
| Agent finished | Diff on project-main or trunk (the land target) shows the change | Agent said success; item branch only |
| Requirements met | Checklist against plan/ticket | Tests alone |
Roles
| Role | Owns | Does not own |
|---|---|---|
| builder | Running this gate before ship claims | Skipping security on changes to skill homes, reviewed at Review |
| tester | CI that produces evidence; verifier subagent for the item | Claiming product DoD from fmt-only green |
| security | Intake / security clear at Review | Day-to-day verify coaching |
| manager | After-act landed+verified | Blessing without evidence |
Red flags
- Wording that implies success without a command in this turn
- “Just this once” / “I’m tired” / “agent said it’s fine”
- Moving on because the diff “looks right”
- Calling the item landed when it only exists on an item branch
Stop. Run the proof. Or say what is still unverified.
Upstream pin
Intent pin: obra/superpowers verification-before-completion @ b36e0829c6d0140e93cfef2ca599b1b07d4a7797. See SOURCES.md.