pm-verify-release-v1
Post-release acceptance verification, grade a shipped release against the spec that promised it. Use when someone says "verify the release", "did the release meet its acceptance scenarios", "post-release verification", "execute the measurement plan", "week-one metrics check", or hands a shipped scope that needs product-side evidence. Every pre-declared acceptance scenario is graded PASS / PARTIAL / GAP / UNVERIFIED against the DEPLOYED build, never the repository, and the pre-declared measurement plan is executed read-only. Returns an acceptance report plus a feedback pack that feeds the next requirements cycle. The owner decides acceptance; this verb prepares the evidence and never marks anything accepted. NOT for tests, CI, deploys, merges or code fixes; NOT for writing a handoff package (pm-requirements-v1); NOT for backlog scoring (pm-portfolio-v1).
naderelewa · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add naderelewa/Product-to-Prod --skill pm-verify-release-v1技能说明
name: pm-verify-release-v1 description: | Post-release acceptance verification, grade a shipped release against the spec that promised it. Use when someone says "verify the release", "did the release meet its acceptance scenarios", "post-release verification", "execute the measurement plan", "week-one metrics check", or hands a shipped scope that needs product-side evidence. Every pre-declared acceptance scenario is graded PASS / PARTIAL / GAP / UNVERIFIED against the DEPLOYED build, never the repository, and the pre-declared measurement plan is executed read-only. Returns an acceptance report plus a feedback pack that feeds the next requirements cycle. The owner decides acceptance; this verb prepares the evidence and never marks anything accepted. NOT for tests, CI, deploys, merges or code fixes; NOT for writing a handoff package (pm-requirements-v1); NOT for backlog scoring (pm-portfolio-v1). argument-hint: " --tier <dev|staging|prod> [--url ] [--scope ] [--stack ]" user-invocable: true
pm-verify-release-v1, deployed build vs spec → acceptance evidence → feedback pack
Input is a shipped scope: the cycle's handoff package directory and the tier it is deployed on. Output is a scenario-by-scenario verdict table with receipts, an acceptance report prepared for the owner, and a feedback pack that makes the next cycle smarter.
Completion is not acceptance. This verb produces evidence. A named human owns the acceptance decision. This skill never marks anything accepted, never flips a gate, and never softens a verdict to make a release look accepted.
Posture: read-only everywhere. Behavioural probes look; analytics probes query; nothing is written to any tier, flag, dashboard or repository. Where an engineering toolchain is configured, its run records are inputs only, never authored, never edited, never re-minted.
Preflight (run first)
PKG_ROOT="${CLAUDE_PLUGIN_ROOT:-${PKG_ROOT:-}}"
[ -n "$PKG_ROOT" ] && [ -r "$PKG_ROOT/.claude-plugin/plugin.json" ] || { echo "STOP: set PKG_ROOT to this package's root directory, the one holding .claude-plugin/plugin.json, then re-run."; exit 2; }
bash "$PKG_ROOT/scripts/preflight.sh" pm-verify-release-v1 [--stack <keys>]
A miss on a required capability blocks the run and prints its exact one-time fix. Conditional keys
are passed once the scenario scope is assembled: analytics-verification (any analytics-touch
scenario in scope) · eng-handoff (the downstream registry, for deployed coordinates) ·
repo-state (change-request state reads). A selected-conditional miss blocks exactly like a required
one, never silently degrade into a partial verification that reads like a full one.
Standalone vs supercharged. Standalone (python3 only): every pre-declared scenario graded
against the deployed build via behavioural probes, plus the report and the feedback pack;
analytics-touch receipts ride NEEDS-CONFIRMATION rows instead of measured values, and say so.
Supercharged: analytics-verification executes the pre-declared measurement plan and compares
receipts against the declared thresholds; eng-handoff resolves deployed coordinates from the
registry instead of from an answer.
Hard rules
- Grade only against what was pre-declared. The spec directory's acceptance scenarios are the only rubric, and the handoff's machine-actionable success-metrics section is the only query set, execute it, never re-derive its intent. A receipt produced from a query nobody pre-declared is INFERRED at best, and must say so. If the spec's scenario is wrong, that is a finding about the spec, not licence to invent a better scenario mid-verification.
- The deployed build, not the repository. Every verdict is stamped
tier + build. Web and API coordinates come from the registry when configured, otherwise from an explicit--url. Mobile builds are resolved live at probe time, build numbers are counters, never pinned in a report drafted earlier. Repository state, diffs and green CI are facts about a repository; by themselves they prove nothing about a deployed tier. - Never touch the building side's records. Run records, gate conclusions and handoff ledgers are read-only inputs.
- Honesty over neatness. UNVERIFIED is a first-class verdict, never force one. A red blocker stays red until its owner reports it unblocked. A feature behind a flag is graded on the tier where the flag is ON, and the production precondition is listed as a follow-up rather than scored as a GAP. Quote fidelity as recorded: measured outranks inspected.
- A missing input is a structured stop, not a guess. Record which input is absent, tag the affected scenarios NEEDS-CONFIRMATION, and proceed with the verifiable remainder only.
- Analytics: aggregates and metadata only. Never user-level exports or profiles. Name the environment explicitly on every read, an unnamed project or property answers from whichever one is default, and that is the single most common source of a confidently wrong number. Budget any quota-limited query: enumerate the calls you need before making the first one. No secret is ever read or echoed; the config holds pointers only.
- Reports are versioned, never overwritten. The owner-edited copy becomes the new canonical base.
- No volume targets. Queue movement is recorded as a measured fact, never as a quota.
- Artifacts land where the run said they land. Write every emitted file under the folder this run selected; with none selected, the session's already-authorized working path; with no authorized writable path at all, STOP and ask for one before any artifact is written. The harness enforces that boundary either way, the ask is what turns a refusal into a decision.
Inputs
| Input | Required | What it is |
|---|---|---|
spec-dir | yes | the handoff package |