Release Readiness
Judge the exact artifact set that will ship, then put the promotion decision to the user. This skill never deploys or publishes.
When to use
- Integrated work is approaching an initial or later release promotion.
- Skip internal work with no release surface. A library or package release is not skipped merely because it has no running production service.
Step 1: Pin down what is being released
Open assets/release-candidate.md before this step's first action; each step
fills its sections. Read references/gate-details.md before filling the
expected rows: each of its numbered rows is an expected row or an approved
omission.
- Fill the
Immutable release-input descriptor. Write no secret values. - Write every attempt and its evidence in the
External attempt and evidence ledger(the ledger: add to it, never rewrite it), never in the descriptor. - Freeze the artifact set: one finished, successful build per required member, from the recorded source. Record the set's identity (row 10).
- Only that set is tested and promoted. A later rebuild is a different set, however equivalent it looks; reject it.
Step 2: Judge it
Work every Readiness rows entry until it is evidenced, not applicable
with an approved disposition (the row's owner accepts the reason it does not
apply, and that acceptance is recorded under the descriptor's approver rule),
or blocking. In that table's Status column, pass is evidenced,
fail is blocking, and a pending or inconclusive row is neither and
holds the verdict at inconclusive until it resolves, unless a fail row
already makes it not ready.
- Install, start, or load each member through the paths a real consumer uses. Passing checks on the source tree are not evidence about the artifact.
- REQUIRED SUB-SKILL: invoke
verification-before-completion. Run every check that protects this promotion over its expected inventory, then compare that inventory with what actually ran and account for every difference (row 11). - Exercise the cutover and the recovery, including the claimed RPO and RTO,
and observe what happens; do not infer it (rows 3, 13). Drill on an
isolated representative copy; shared or production state needs exact
direct authority, else block. Fill
Rollout and recovery. - Check the target: configuration, secrets, capacity, dependencies, and the order of the steps (row 6). Verify that each is present and has the expected shape; never print a value.
- Tie the rollout control to the assurance, migration, or release policy
that owns it (rows 5, 12). If there is no such policy, the row is blocking;
for a bounded feature,
rollout-controlsdrafts one for its owner to accept. Never invent a policy or lower one. - If the promotion is an expansion or a full release, attach the evidence observed in the prior stage.
- Account for the consumers (rows 7, 8, 15). If the release is a package, install the packed artifact into a clean, representative consumer before publishing anything.
Step 3: Decide
Record a decision (its disposition) for every deviation in
Deviations(row 14). Lowering a check so the frozen set passes is a new assurance decision, not a release fix.Write the
Verdictwith every identity it is bound to.Recompute every identity immediately before any decision. If any has changed, reopen the decision.
Present the verdict and the choice, then end the turn:
Release candidate {{artifact-set identity}} → {{target}} Verdict: {{ready for the target | ready with accepted deviations | not ready | inconclusive}} — {{n}} gates evidenced, {{n}} deviations owned, rollback {{state}} 1. Promote this exact set 2. Hold 3. Cancel Recommendation: {{1 only when ready or ready with accepted deviations, otherwise 2}} — {{one sentence}}.Promote nothing until the user names one of the three; record the answer in
Decision. On 1, the promotion runs as its own authorized action, outside this skill, never on the strength of the verdict. On 2 or 3, stop.
Gotchas
- Step 2.6's prior-stage evidence has to be captured while that stage is live. A rollout that expands on its own schedule — a percentage ramp, a timed cohort — has already moved by the time anyone looks, and the expansion is then approved against evidence nobody observed.
- A row recorded as not applicable counts the same as an evidenced one in a bare tally of checks, which is why Step 3.4 presents "gates evidenced" and "deviations owned" as separate numbers. Collapsing them into one hides how much of the release nobody checked.
Hard stops
- The release tests ran on the source tree, or on a different or incomplete artifact set.
- The release-input, artifact-set, terminal-evidence, or verdict identity is missing, stale, incomplete, or changed.
- Any required cell, evidence, or change is missing, duplicated, skipped or rejected with only a generic reason, unresolved, unowned, out of order, or beyond its bound.
- A decommission is partial, failed, awaiting validation, failed validation, or unresolved.
- The rollback names only the source, not a restorable artifact together with the data and config state to restore.
- The fixer, builder, or deployer accepts an unapproved deviation on their own.
- “Ready” is treated as authority to deploy or publish.
Common mistakes
- Calling a bare feature flag a rollout plan, or calling a recovery promise with no stated bound a rollback.