metravel-sprint-reviewer

Metravel Sprint Reviewer

Use this skill for board acceptance, not implementation. Code fixes go back to the owning implementation skill.

Read first:

• AGENTS.md
• docs/RULES.md
• docs/CODEX.md
• docs/TASK_BOARD_MCP.md
• relevant feature docs and the ticket's Task Contract.

Board Rules

• Use the shared MCP task board tools only.
• If board endpoints return HTTP 401, refresh the staff token through .env.e2e using docs/TASK_BOARD_MCP.md; never print the token.
• Review only the requested ticket or active sprint scope.
• Do not create new feature code while acting as reviewer.
• Do not move a task to done without real evidence for its Done gate. A live shared gate may satisfy only the remaining automated-test step as validation delegated: active gate pid/name when its scope covers the task; it is not passed and cannot replace deploy/browser/API/device evidence.

Acceptance Loop

1. Load the active sprint or requested task.
2. For each candidate in review or testing, read the full description, dependencies, blockers, and Task Contract.
3. Reject refinement gaps before runtime work:
   • missing Scope, User-visible result, Data/API contract, Platform impact, Localization impact, Dependencies, Fallback/mock policy, Validation, Regression control, or Done gate
   • Regression control empty or none on a kind=bug or consolidation task
   • unresolved blocking dependency
   • needs_human=true without the human step completed
4. Run or inspect the exact validation required by the Done gate. Reject evidence that violates docs/TASK_BOARD_MCP.md → #### Качество evidence:
   • a 200/no-crash result offered for a task about size, count, duration or order, with no before/after number
   • no negative probe where an unsupported input exists: it must be observably distinguishable from a valid one, never a silent heavy/generic fallback
   • a build-time or deploy-time check offered for a user/crawler-visible production surface instead of a recurring production probe
   • production-backed preview, local production build, production API behind a local bundle, or deploy logs offered as evidence that the live production URL is fixed
   • a production optimization without the same before/after URL, viewport/browser/DPR, auth/cache state, request/API count, bytes, response codes, and relevant scroll/lazy-loading probe
   • a page-level performance task whose evidence proves one component or URL but does not remeasure the whole page budget and adjacent consumers touched by shared media/source/pagination code
   • a test that mocks the very primitive under investigation offered as contract evidence
   • a consolidation closed without naming the CI guard that fails when the new single contract is bypassed
   • targeted tests or governance checks
   • browser/API probes against the target environment
   • device/mobile evidence when the contract requires it
   • npm run test:i18n and affected RU/BE/UK/PL/EN evidence when localization is impacted
   • when the required automated gate is already owned by another live session, apply the delegated-validation contract from AGENTS.md: accept only if it covers the task and no non-test Done-gate step remains; otherwise reject with validation skipped
5. Append evidence to the task description without erasing prior history.
6. Move passing tasks, and tasks whose sole remaining automated gate is validly delegated, to done. Keep tasks with missing validation evidence in review or testing. If validation finds a defect that requires implementation changes, return the task to in_progress with the evidence. Use blocked_by only when a newly discovered hard dependency makes it impossible to start or continue the remaining work; waiting for review, tests, production/API/device checks, or an incomplete Done gate is never such a blocker. The active gate owner must reopen a delegated task or record a blocker if it cannot fix a discovered failure.

For a production incident, no deploy authorization is not a reason to lower the Done gate: accept the implementation into review/testing with local fix ready; production verification pending, but do not move it to done and do not describe production as fixed until the live target is remeasured.

Output

Return a compact Sprint Review:

• tasks accepted with evidence
• tasks rejected with blockers
• checks/probes run
• tasks that need human or backend follow-up
• any board/token/access blocker