Run Pre-Release Readiness Reviews Before Production Deploy

Runs a read-only production release readiness review to find missing deploy materials, unsafe ordering, configuration gaps, data migration gaps, and ambiguous production risks before CI/CD or a manual release begins - it never modifies source, configs, migrations, secrets, or deployment files, and never executes migrations, clears caches, uploads assets, triggers CI/CD, deploys, or rotates secrets. Scope selection follows a fixed priority: a given PR reviewed via read-only gh pr view/gh pr diff, an explicit base..head range, a head-only commit compared to the previous reachable release tag, or with no scope at all, the previous release tag against HEAD - falling back to the latest 5 commits with an explicit warning if no usable previous tag exists. Before analyzing anything it reads references/checklist.md (so release domains aren't skipped) and references/report-template.md (for consistent priority/owner/redaction/output conventions), and looks for project-local AGENTS.md/CLAUDE.md guidance to respect project-specific conventions - treating that guidance as interpretive context, never as permission to override its own read-only and no-secret-disclosure rules. Evidence collection uses only safe inspection commands (git status, git diff --stat, git log, git blame, rg, and gh pr diff when available); if a command can't run, the limitation is recorded rather than papered over. The review workflow maps changed code to production requirements across six domains: schema changes to migrations/indexes/seeds/backfills, config reads to env examples/secrets/flags, cache key or TTL changes to invalidation and prewarm work, queue producers/consumers to topic setup/DLQ/idempotency/deploy order, asset references to object storage/CDN/certificates/permissions, and service contract changes to deploy sequence and rollback risk. Owners are inferred from git blame on changed lines or recent git log authors and labeled explicitly as inferred, never including email addresses. Every finding is classified P0, P1, or P2 with module, finding, evidence, inferred owner, risk, and recommended action, sorted highest priority first - clean checklist categories are never listed, only confirmed problems and risks needing confirmation. Secrets are never disclosed in full: only file path, line number, variable name, secret type, and a redacted hint. Uncommitted worktree changes are excluded from the reviewed range by default (always reporting whether the worktree is dirty), unless the user explicitly asks to include them; if dirty/untracked files touch release-relevant areas, that becomes its own P2 confirmation item. The skill distinguishes a real release confirmation item (a diff-linked production risk like an unverifiable new secret or an unclear migration status, classified P1/P2 and forcing conclusion NEEDS_CONFIRMATION) from a neutral "Unable To Verify" limitation (missing deployment credentials with no specific release requirement in the diff, which doesn't change the conclusion by itself) - promoting a limitation to a real finding whenever it blocks confirming a release-critical change. The final conclusion is always exactly one of BLOCKED, NEEDS_CONFIRMATION, or NO_BLOCKER_FOUND, and a service-deployment-order section is included only when the diff touches multiple services, async workers, migrations, queues, cache, or public contracts.

Use it when someone asks for a release audit, pre-release review, go-live review, or deploy readiness check; before publishing a tag, deploying production services, or merging a release branch; when a PR or git range may touch migrations, environment changes, queues, cache, object storage, or service contracts; or when someone asks whether a change is safe to ship and wants a read-only risk report.

Input is a PR URL/number, an explicit git range, a head commit, or (with no scope given) the current HEAD against the previous release tag. Output is a prioritized (P0/P1/P2) risk report with evidence, inferred owner, and recommended action per finding, a fixed conclusion value, and - when relevant - a service deployment order section.

Uses read-only git and gh CLI commands plus ripgrep for evidence collection, and reads project-local AGENTS.md/CLAUDE.md guidance files when present to interpret findings in project-specific context.

Release managers and engineers who want an evidence-linked, read-only risk assessment before shipping - covering migrations, config, cache, queues, assets, and service contracts - without the reviewer touching or executing anything in the repository or infrastructure.