| name | verify-loop |
| description | Prove a change works before calling it done โ build, test, lint, smoke-check, and report pass/fail. Use after any code change, before any commit/PR, and as the self-verify step inside an autonomous loop. Trigger whenever the agent is about to claim a task is complete. |
verify-loop
"Done" is a claim, not a proof. An open loop that writes code with no feedback is a machine for
generating confident mistakes. This skill is the feedback.
Steps
- Run the build (
AGENTS.md ## Project โ Build). If it fails, fix the cause and repeat.
- Run the test suite (
โ Test). For a bug fix, ensure a test reproduces the bug and now passes.
- Run lint/format (
โ Lint).
- Smoke-check the changed behavior directly where feasible.
- Where relevant, diff behavior between
main and the change.
- Report PASS or FAIL. On FAIL, show the minimal failing output (not the whole log) and the
likely cause.
Notes
- The maker shouldn't be the only checker โ for risky changes, hand verification to a separate
reviewer subagent (a stronger model). See
primitives/subagents/.
- This is the
self_verify() every loop in harness/ and loops/ depends on.