verify-run

name	verify-run
description	Post-run verification of an autonomous implementation run (implement-flow / implement-auto / implement). Runs the run-verifier agent to catch the failure modes the evidence gate can't see — staged tails, dead code, vacuous receipts, collapsed commits — and, with --fix, remediates them. Use after an unattended run, before trusting its "done".

Independently verify what an autonomous run produced — before trusting its result or opening the PR: $ARGUMENTS

Detection lives in the run-verifier agent (read-only; structured verdict). This skill is the entry point: it runs that agent, presents the verdict, and — only with --fix — remediates the mechanical findings. Run it from the same worktree the run used.

$ARGUMENTS may name the worktree/branch to check (default: the current tree) and/or --fix.

Why this exists: these runs close tasks on executed evidence, but "a test executed and passed" is neither necessary nor sufficient for "the feature works" — correct work can be left uncommitted (false negatives) and green-but-broken work reported as done (false positives). The agent catches both.

Run

1. Spawn the run-verifier agent with a one-line brief naming the worktree/branch to verify (default: the current tree) and whether the run reported all tasks closed. It resolves the run's commits, executes its checks in that worktree, and returns { clean, findings, learnings_path }.

2. Present the verdict:

   • clean: true → one line: verified · <N> commits on <branch> · no blocking findings. Surface learnings_path if set (it may be out-of-tree and invisible in the diff). Stop.
   • findings → list them most-severe first, each with the exact file/symbol/commit and the fix. A block (staged-tail, vacuous-receipt, dead-code) means the run's "done" does not hold; a warn (commit-boundary) is cosmetic history.

3. --fix only (never otherwise — default is advisory): remediate the mechanical findings yourself — finish a staged tail's commit (repo's own conventions, one commit, explicit-path staging), delete confirmed dead code, or re-split a collapsed commit — then re-spawn run-verifier to confirm.

Notes

• General across Go / JS-TS / Elixir; the agent carries the per-language recipes. A repo-specific cross-boundary invariant (e.g. an event's SNS→SQS filter wiring, unexecutable locally) does not belong here — encode it as a hermetic test in that repo so it runs in CI for everyone.
• Same agent, three callers. implement-flow's Finalize spawns run-verifier as its integration gate, and implement / implement-auto can run it post-run — so the /verify-run command and the automated gates verify identically.
• Defense in depth. The checks overlap the workflow's own receipts on purpose. As the evidence contract absorbs one (worktree-pinned receipts, per-file staging, reachability in the audit), the agent's matching finding goes quiet — the signal it's safe to retire it.