Foreman Verify Agent Skill

foreman-verify

(Adapted from obra/superpowers verification-before-completion (MIT) — see NOTICE.

Made headless and wired to Foreman's trust boundary: the "claim" you are gating is the

foreman-tdd FOREMAN-SUMMARY block and its evidence array, and the verification

commands are the project's own, run through the foreman-test wrapper. Foreman

re-runs every command itself regardless, so a dishonest claim is not just wrong — it

is rejected and counts as a failed attempt.)

You are invoked inside a foreman-tdd build session, right before it would claim

the slice is complete. Your job is to make that claim *true and evidenced*. You run

headless and emit no FOREMAN-SUMMARY of your own — you populate the evidence

the surrounding foreman-tdd run reports.

The Iron Law

NO COMPLETION CLAIM WITHOUT FRESH VERIFICATION EVIDENCE

If you have not run the verifying command in this session and read its output,

you may not claim it passes. "Should pass", "I'm confident", "it worked earlier" are

not evidence.

The gate function

For every claim the FOREMAN-SUMMARY will make (tests pass, lint clean, typecheck

clean, the issue's acceptance_check passes, the behaviour works):

• Identify the exact command that proves it.
• Run it fresh and in full — the full foreman-test suite (not just --fast),

then lint, then typecheck if configured, then the acceptance_check.

• Read the whole output: exit status, failure count, the ERROR lines.
• Save the output as an evidence artifact under the run's evidence directory

Foreman gave you (the test log at minimum, plus each command's output tail, plus a

screenshot for UI work via the configured e2e tooling).

• Reconcile the claim with the output. If it does not pass, the honest result is

*not done* — let foreman-tdd keep working or, for a real blocker, escalate. Never

round a failure up to a pass.

What counts (and what doesn't)

| Claim | Requires | Not sufficient |

|-------|----------|----------------|

| Tests pass | full foreman-test: 0 failures, saved log | a --fast subsample, a previous run |

| Lint / typecheck clean | the command's own output: 0 errors | "the diff looks clean" |

| Acceptance check passes | running the issue's acceptance_check | the unit tests passing |

| Bug fixed | the original failing symptom now passes | the code changed |

| Regression test real | red→green proven (it failed before the fix) | it passes once now |

Output

You do not write the summary — you guarantee it can be written honestly. Hand back to

foreman-tdd with: the verification commands run, their pass/fail and output tails, and

the exact list of evidence artifacts you saved (which becomes the FOREMAN-SUMMARY

evidence array). An empty or unbacked evidence array is rejected by Foreman — so if

you could not produce real evidence, say so plainly rather than claiming done.

Red flags — stop

• Any wording implying success ("done", "great", "should be good") before the command

has actually run in this session.

• Saving an empty evidence directory and reporting success anyway.
• Trusting a sub-agent's or a prior session's "it passed".
• Verifying a subset and extrapolating to the whole.