Certify Production Runtime Health with Evidence-Backed Verdicts

Production Runtime Certification produces an evidence-backed release verdict that keeps code quality, CI success, deployment success and live runtime health separate. It prevents a green pipeline or a successful deploy event from being treated as proof that a service is usable in its target environment.

Before testing, the parties agree a certification contract: the exact source revision, repository and target environment, the deployment identifier and expected artifact or image digest, the health and readiness endpoints with expected success semantics, the migration mechanism and expected schema version, the critical routes or user journeys, and the authorized, non-destructive probes and credential boundaries. If the target, authorization or success criteria are ambiguous, it stops and asks. It never probes an unrelated environment, exposes credentials in output or modifies production data to obtain a pass.

Every gate is recorded as PASS, FAIL, BLOCKED (could not be tested because access, tooling, configuration or a dependency is unavailable) or NOT APPLICABLE with a written reason. BLOCKED is not PASS. The overall result is GREEN only when every required gate passes, RED when any required gate fails, and BLOCKED when a required gate is blocked with no failure.

The workflow has eight steps:

• Pin the subject: capture an immutable source revision and runtime artifact. A branch name, local working tree or "latest deployment" label is not immutable evidence.
• Verify source and CI: run the current validation and test commands on the pinned revision and inspect the required CI jobs for the same revision, including conclusion and timestamp.
• Verify deployment identity: the environment reports the expected deployment, revision or artifact digest.
• Verify database state through the supported migration status or a read-only schema-version query. Do not apply, roll back, repair or stamp migrations without separate authorization and a recovery plan.
• Probe liveness and readiness separately, recording timestamp, target, status code, bounded response summary and latency, and redacting secrets.
• Exercise the smallest authorized smoke suite on critical routes, preferring synthetic records and read-only probes with least-privilege test identities.
• Check operational signals in the deployment window, such as crash loops, unhandled exceptions, dependency failures and elevated error rates. Missing log access is BLOCKED and not proof of health.
• Issue the verdict with the evidence matrix and unresolved risks, keeping observations separate from inference.

Use it before declaring a release, environment or migration production-ready, after a deployment when CI is green but runtime health is unknown or failing, in release audits that need a concise PASS, FAIL or BLOCKED decision, and when handing a deployed service to operations or a client. Do not use it as a substitute for feature acceptance, security review, load testing, disaster-recovery exercises or regulatory approval. A certification is a point-in-time result for one revision and environment, and it expires when the artifact, configuration, schema, dependencies or environment changes.

Input is the agreed contract above. Output is a report listing environment, source revision, deployment or artifact identifier, observed time, one line per gate, the overall verdict, evidence and unresolved risks.

SOURCE VALIDATION: PASS | FAIL | BLOCKED | NOT APPLICABLE
CI: PASS | FAIL | BLOCKED | NOT APPLICABLE
DEPLOYMENT IDENTITY: PASS | FAIL | BLOCKED | NOT APPLICABLE

The skill includes an example where CI is green but liveness and readiness return HTTP 503, giving an overall RED verdict, and an example of a fully supported GREEN verdict.

It points to @verification-before-completion for checking that completion claims have fresh evidence, and to @deployment-procedures for executing the deployment before returning to certify it. It works with Claude, Cursor, Codex and Gemini.

DevOps and release engineers, and teams handing deployed services to operations or clients, who need a reproducible go or no-go decision based on fresh runtime evidence.