Zum Inhalt

sprint-review

Schließt einen aktiven Sprint gemäß Sprint-Spec: validiert das Deploy-Artefakt und protokolliert den Value-Delivery-Audit-Trail.

Close an active sprint per the project sprint spec, validating the deployable artefact and recording the value-delivery audit trail. Invoke when the user asks to close a sprint, review a sprint, finish a sprint, ship a sprint, or wrap a sprint. Also handles equivalent German-language requests. Promotes active → review, validates artifact_ref per the release-artifact spec's per-project-type rules, confirms the named verifies_sprint_value acceptance criterion is checked, optionally chains into release-notes-curate and release-publish-trigger (operator-opt-in, recorded verbatim in ## Review notes), then promotes review → closed. Falls back to review → cancelled with a one-paragraph rationale when artefact validation fails unrecoverably. Supports resume on re-invocation per spec/claude/resumable-work/.

Anwenden wenn

  • you want to close, finish, or ship an active sprint
  • you want to validate the sprint's artifact_ref against the release-artifact spec
  • you want the optional chain into release-notes-curate and release-publish-trigger

Nicht anwenden wenn

  • You want to drive day-to-day mechanics of the sprint → sprint-execute
  • You want to plan or open a new sprint → sprint-plan

Siehe auch

Referenziert von

Sprint Review

Closes an active sprint per spec/project/sprint/<canonical_language>.md. This skill is the canonical write authority for active → review, review → closed, and the cancellation paths from any stage. It delegates artefact validation to the rules declared by spec/project/release-artifact/<canonical_language>.md and may chain (with explicit operator opt-in) into release-notes-curate and release-publish-trigger per spec/project/release-skill-layer/<canonical_language>.md.

Why this is a skill, not an agent

  • Externally-visible mutations gate on user confirmation — promoting a sprint to closed is irreversible bookkeeping, the artefact-validation results need to be reviewed by the user before the promotion fires, and the release-skill-layer chain is operator-opt-in by spec; an agent's fire-and-forget shape would lose those gates.
  • Orchestrator that chains other skills — this skill conditionally dispatches release-notes-curate and release-publish-trigger mid-flow, and falls through to a cancellation path on validation failure; the skill-orchestrates pattern (per skill-vs-agent) defaults the orchestrator to skill form.
  • Output flows back into the main conversation — the validation transcript, the ## Review notes audit block, and the chosen lifecycle outcome (closed vs cancelled) all belong in the user's working context, not behind an agent's structured-report boundary.
  • Counter-dimension considered: a narrower agent could sharpen the per-kind artefact verification step (git rev-parse, gh release view, docker manifest inspect, etc.), but the load-bearing dimension is the multi-step orchestration with operator-opt-in branches — skill wins.

User-language policy

Detect the user's language and respond in it. Sprint markdown content (including ## Review notes) stays in the project's primary language; frontmatter keys, lifecycle tokens, verification command outputs (exit codes, key output lines), and any gh / git / docker invocations remain English so portfolio automation reads them deterministically.

Preconditions

Before mutating the sprint or dispatching any chained skill, confirm:

  • Current working directory is inside a git repository with project/sprints/ populated.
  • The user has named the sprint to close (or there's exactly one active sprint—use that one and confirm with the user before mutating).
  • The sprint's status is active. Refuse on planned (operate via sprint-execute first), review (mid-flow recovery; ask the user how to proceed), closed, or cancelled (already terminal).
  • Every feature listed in the sprint's features frontmatter is done (read each feature file and check status). On any non-done feature, stop and report the offending features; this skill doesn't drive feature transitions—sprint-execute does.
  • The sprint's features frontmatter list is non-empty. An empty list makes the value-delivery contract unsatisfiable per spec/project/sprint/ §Lifecycle; refuse the active → review transition with a verbatim error.
  • The sprint's last_commit frontmatter field is non-null per spec/project/sprint/ §Acceptance Criteria. If it's null, hand back—sprint-execute is the canonical writer of that field.

Operations

The skill follows a strict order. The artefact-validation block (step 3) MUST complete before the release-skill-layer dispatch (step 5); a failure in step 3 short-circuits to step 7 (cancellation), never to step 5. The order is fixed because publishing a release whose artefact ancestry is broken is the exact failure mode release-artifact exists to prevent.

1. Promote active → review

  1. Confirm every precondition above. On any failure, stop and report; don't write.
  2. Set status: review on the sprint's frontmatter. Don't set ended yet—the field is reserved for closed or cancelled.
  3. Surface the sprint summary to the user: number, slug, value_statement, the feature list with each feature's verifies_sprint_value field, and the current artifact_ref (which MUST be non-null per spec/project/sprint/ §Artefact contract; if null, refuse the transition and hand back).

2. Detect the project type

Use the same signals as release-skill-layer and github-issue-templates-apply per spec/project/release-artifact/ §Project-type detection:

  • .claude-plugin/plugin.json → Claude plugin
  • pyproject.toml shape → Python application or library
  • package.json → Node / TypeScript
  • declared CLI entry → CLI tool
  • custom_components/<name>/manifest.json → HACS integration
  • mkdocs.yml (or equivalent) → documentation-only

A per-repo override at .github/release-skill-layer.yml (artifact_kinds: [...]) takes precedence; MUST NOT introduce any other override file.

3. Validate artifact_ref per the release-artifact spec

Parse the sprint's artifact_ref (string or list of strings) per the detected project type and the optional artifact_kinds override. For every component, run the per-kind verification command listed in spec/project/release-artifact/ §Validation at sprint closure:

  • Git tag → git rev-parse <tag>
  • GitHub release → gh release view <tag>
  • Container image → docker manifest inspect <image>
  • PyPI distribution → pip index versions <dist>
  • npm package → npm view <pkg>@<version>
  • Doc-site deploy URL → HEAD request
  • Claude plugin → git rev-parse <plugin-version-tag> plus a marketplace-resolution probe (read .claude-plugin/marketplace.json at HEAD and confirm the plugin version is listed)
  • HACS integration → git rev-parse <tag> plus a manifest.json version-field equality check at the artefact's commit

Then confirm: the resolved artefact's commit (when applicable) MUST equal the sprint's last_commit frontmatter field or be reachable from it. A missing or null last_commit blocks closure; surface the failure verbatim.

A bare commit SHA as artifact_ref is rejected unless the project explicitly opts in via .github/release-skill-layer.yml.

Record the verification transcript (commands run, exit codes, key output lines) for later persistence in ## Review notes. On any failure, do not proceed to step 4 or 5. Surface the failed check verbatim, ask the user whether the failure is recoverable (re-cut the artefact and rerun) or unrecoverable (cancel the sprint), and route accordingly: recoverable → stop; unrecoverable → step 7.

4. Confirm verifies_sprint_value

Walk the sprint's feature list and confirm:

  • Exactly one feature carries a non-null verifies_sprint_value field per spec/project/sprint/ §Value-delivery contract.
  • The named acceptance-<n> identifier exists on that feature.
  • The named acceptance-criterion bullet is checked (- [x] **acceptance-<n>** …).

On any violation (zero or multiple verifies_sprint_value declarations, missing identifier, or unchecked bullet), refuse closure and surface the offence verbatim. Sprint closure with a missing or unchecked verifier is a hard violation per spec/project/sprint/ §Acceptance Criteria.

5. Operator-opt-in chain into release machinery (optional)

This step runs only when the project type publishes via release-publish.yml (typical for Claude plugin, Python library, Node library, CLI tool, HACS integration—the project types named by spec/project/release-artifact/ §Dispatch boundary to release machinery).

Ask the user explicitly whether to chain. The chain points are fixed:

  • release-notes-curate — augments the open release-drafter draft body with project-context-aware sections; idempotent on re-runs per spec/project/release-skill-layer/ §Skill A.
  • release-publish-trigger — validates every gate from spec/project/release-automation/ §Pre-publish verification and dispatches release-publish.yml via gh workflow run.

Outcomes to record verbatim in ## Review notes:

  • "Chained: release-notes-curate (run URL or summary), release-publish-trigger (workflow run URL)" — when the user opts in.
  • "Skipped: release-notes-curate / release-publish-trigger; reason: " — when the user declines, naming who or what dispatches publication later (operator manually, automated post-merge job, or a follow-up sprint) per spec/project/sprint/ §Dispatch into release machinery.

The recording is mandatory; a sprint that closes without a recorded operator decision on the chain fails validation per spec/project/sprint/ §Acceptance Criteria.

6. Promote review → closed

  1. Set status: closed and ended: <today's ISO date> on the sprint's frontmatter.
  2. Populate ## Review notes with: the verification transcript from step 3 (commands and outcomes), the verifying feature's full location (features/<slug>.md plus the acceptance-<n> ID) per spec/project/sprint/ §Value-delivery contract, the operator's release-skill-layer chain decision from step 5, and any cross-references to out-of-band artefacts published during the sprint (per spec/project/release-artifact/ §Out-of-band artefacts).
  3. If the sprint's roadmap items each have every feature done, advance their status to done per spec/project/roadmap/ §Lifecycle. A roadmap item whose pending features fall outside the sprint stays active until a successor sprint closes.
  4. Surface unconsumed blog-trigger deferrals. Per spec/project/blog-author-trigger/ §Deferral artefact (SHOULD: the source consumer's sprint-review surfaces unconsumed deferrals at sprint close so they don't accumulate silently), scan project/blog-triggers/*.yml for entries with status: deferred. List each by feature-slug plus its deferral_reason and deferred_at, so the operator can revisit them (re-invoke blog-author-trigger to author / update, or leave deferred). This is a read-only surfacing step; sprint-review never consumes or mutates the deferral artefact (only a later trigger run sets status: consumed). When the directory is absent or holds no deferred entry, note "no unconsumed blog-trigger deferrals" and move on.
  5. Surface the closed sprint summary to the user: path, ended date, the artefact reference, the verifying feature, the chain decision.

7. Cancellation path: review → cancelled (or earlier)

Triggered when step 3 fails unrecoverably (the underlying release pipeline is broken and re-cutting the artefact isn't feasible at this time) or when the user explicitly cancels at any earlier stage.

  1. Set status: cancelled and ended: <today's ISO date>.
  2. Populate ## Review notes with a one-paragraph rationale that MUST name (a) the lifecycle stage at which cancellation occurred (planned, active, or review), and (b) why recovery wasn't feasible at this time. Missing rationale or stage citation is a hard validation failure per spec/project/sprint/ §Acceptance Criteria.
  3. Re-target every roadmap item still pointing at the cancelled sprint (target_sprint) per spec/project/roadmap/ §Sprint and feature linkage: clear to null or point at a planned successor. Roadmap items MUST NOT advance to done from a cancelled sprint even when every feature is individually done; the items remain active until a successor sprint reaches closed.
  4. Surface the cancelled sprint summary to the user, naming the rationale and the re-targeting outcome.

Gotchas

  • artifact_ref validation is per-project-type. The skill reads the project type from project/portfolio.yml (or, when absent, from heuristic detection) and applies the matching artifact-ref shape from the release-artifact spec: a Claude plugin sprint expects a published-release tag, a Python library expects a PyPI version, a documentation-only project expects a deployed-docs commit SHA, and so on. A wrong project-type detection produces a misleading "artifact_ref invalid" refusal. The skill surfaces the detected type before validating so the operator can override.
  • active → review is reversible (a failed artifact validation can route back to active); review → closed and review → cancelled are NOT. The skill warns explicitly before either terminal transition and requires verbatim operator confirmation.
  • The verifies_sprint_value acceptance criterion is the load-bearing closure gate. When the named criterion isn't checked on the carrying feature, the skill refuses closed and routes to cancelled (with a clear rationale paragraph) — operators sometimes try to argue the criterion "was achieved differently"; the spec is strict that the named criterion is the audit trail, not a related observation.
  • Optional chaining into release-notes-curate and release-publish-trigger is operator-opt-in. Defaulting to chain leads to surprise releases; the skill stays explicit about each chain hop and records the operator's opt-in verbatim in ## Review notes.
  • Cancelled sprints leave a value-delivery gap that the next sprint must explain. When the skill cancels a sprint, the unfinished features need re-targeting (typically to the next sprint or back to the roadmap queue). The skill surfaces the affected feature IDs and asks where each one lands; nothing is silently re-targeted.
  • Unconsumed blog-trigger deferrals are surfaced read-only, not consumed. At review → closed the skill lists project/blog-triggers/*.yml entries with status: deferred (per spec/project/blog-author-trigger/ §Deferral artefact) so they stay visible, but it never edits them: the deferred → consumed transition is owned by a later blog-author-trigger run, and cancelled is operator-set. The surfacing is a SHOULD reminder, not a closure gate — a sprint still closes when deferrals remain.

Examples

  • Read examples/01-clean-close-claude-plugin.md when closing a sprint cleanly for a Claude plugin project with all features done.
  • Read examples/02-artifact-validation-fails-cancel.md when artifact validation fails and the sprint must be cancelled instead of closed.
  • Read examples/03-chain-into-release-skill-layer.md when the user opts in to chaining into release-notes-curate and release-publish-trigger after sprint closure.

Resumability

Per spec/claude/resumable-work/, this skill is resumable: true. State is persisted to .resume/sprint-review/<run-id>.yml after every successful user-approval gate and after each named phase boundary. On re-invocation, scan that directory for files with status: in_progress whose inputs: snapshot matches the current invocation; if one matches, prompt the operator with Resume run <run_id> from phase <phase> (last checkpoint <last_checkpoint_at>)? [resume / start-new / discard]. The state-file envelope (schema_version, run_id, inputs, phase, decisions[], status, ...) and the fail-closed semantics on schema or YAML errors are load-bearing in the spec; don't duplicate those rules here.

Hard rules

  • Never call gh release edit --draft=false, gh api -X PATCH /repos/.../releases/<id> draft=false, or any other path that flips the draft state outside release-publish.yml. The rule comes from spec/project/release-automation/, spec/project/release-skill-layer/, and spec/project/release-artifact/ §Dispatch boundary to release machinery and is non-negotiable here too. The only acceptable publish path is dispatching release-publish-trigger.
  • Never transition active → closed without passing through review. The value-coverage check lives in review per spec/project/sprint/ §Lifecycle.
  • Never close a sprint while any feature in features is non-done, or while the features frontmatter list is empty, or while last_commit is null, or while artifact_ref is null. Each is a hard refusal point.
  • Never close a sprint with zero or more than one verifies_sprint_value declarations across its features, or with the named acceptance-criterion bullet unchecked.
  • Never accept a bare commit SHA as artifact_ref unless the project explicitly opts in via .github/release-skill-layer.yml.
  • Never chain into release-notes-curate or release-publish-trigger without the user's explicit opt-in, and never close the sprint without recording the operator's chain decision (chained or skipped) verbatim in ## Review notes.
  • Never advance a roadmap item to status: done from a cancelled sprint, even when every feature is individually done. Re-target the pending features to a successor sprint instead.
  • Never edit any sprint other than the one being reviewed in this run. sprint-plan and sprint-execute own the other sprints' lifecycles.
  • When spec/project/sprint/, spec/project/release-artifact/, spec/project/release-skill-layer/, or spec/project/release-automation/ disagrees with this skill, the spec wins. Propose updating this skill rather than silently diverging.

Multi-model testing

Examples and operations in this skill are verified on Claude Sonnet 4.6 as the default model; spot-checked on Haiku 4.5 for cost-sensitive runs; Opus 4.7 is appropriate for high-stakes audits that require deeper reasoning. The skill body has no model-specific assumptions beyond standard tool-call semantics.