---

name: production-validation description: "Before declaring engineering work done, validate that what you built matches the production reality — not just the local or test environment — using the real signals (CI, observability, smoke tests)."

Production validation

The biggest gap between agent work and human-quality work is that agents tend to call something done when the local checks pass. Production validation is the layer on top: confirm the change actually behaves correctly where it runs.

When to run

• Before marking any code-writing planning task as done.
• After deploying to staging or production.
• After a rollback under rollback-runbook.
• After completing an incident-postmortem-loop action item that promises a fix.

The 4-gate checklist

Treat this as a hard gate before closing the work.

Gate 1 — CI green

• All required status checks on the PR are green via github_status_checks.
• No skipped checks that should have run.
• If you cannot fetch CI status, fail closed — do not declare done.

Gate 2 — Reproducible local checks

• Tests run from a fresh clone (not an incremental cache).
• Lint, typecheck, and unit tests all pass.
• For UI work, the relevant Storybook or Cypress flow runs without manual fix-ups.

Gate 3 — Observability sanity

• Logs at the relevant level show the expected entries for the new path.
• Metrics (Datadog/Grafana) for error rate, latency, throughput show no regression in the last 15 minutes after deploy.
• If a new metric or alert was added, confirm it is reporting.

Gate 4 — User-visible smoke

• Trigger the user-visible flow at least once on the deployed environment.
• For background jobs, verify one job completes successfully end-to-end.
• For data writes, confirm the write reached the destination (DB row, file, event).

What to record

After all four gates pass, post a brief validation summary on the PR or task:

production-validation:
  ci: <PR url / commit>
  local_checks: <pass/fail>
  observability: <link to dashboard>
  smoke: <description of the flow run>

Anti-patterns

• Closing the task as soon as the PR merges.
• Skipping Gate 3 because the change "felt small".
• Running the smoke check in local instead of production.