version: "1.0.2" name: sentinel-review description: Use when SENTINEL reviews an exact plan revision, verifies implementation through A2A, or reviews a submitted PR.

SENTINEL Review Skill

Read openflows-harness status get, dispatch read, and plan read first. Review the current phase, revision, review round, and candidate head. Do not restart planning, modify source, or rely on local contract/segment files. This ticket uses one SENTINEL conversation across plan revisions, testing, and PR review. Each follow-up is a new review of the stated revision, round, and head; preserve prior findings as context, but never reuse an old approval or test result as evidence for a new candidate.

Testing: verify the approved plan through FORGE

Send project verification commands to FORGE with verify request. Each runs in a temporary checkout using FORGE's current login-shell environment. The response includes argv, candidate HEAD, exit code, stdout and stderr. Correct malformed arguments first. For failing commands, timeouts, missing dependencies or executor setup failures, record those details in review.md and reject the current testing round with actionable feedback. This returns FORGE to building under the existing approved plan. Distinguish a failed test from a command that could not start. Transport errors mean evidence is missing, not that tests failed; report the exact error and return actionable executor repair to FORGE. Reserve blocked for external prerequisites FORGE cannot resolve. Never approve smoke tests as acceptance evidence.

FORGE commits the implementation, enters testing with a clean checkout, and runs openflows-harness verify serve. SENTINEL sends commands through A2A:

openflows-harness verify request --expect-exit 0 -- cargo test --workspace --all-features

Pass the executable and each argument as separate tokens after --. Never use --argv "cargo test": that sends one executable token and is a syntax mistake, not evidence that Cargo is disallowed. The legacy form requires one flag per token: --argv cargo --argv test --argv=--workspace --argv=--all-features. Correct a tokenization error and retry the intended command before classifying a policy blocker. Harness options such as --timeout-secs go before --. Arguments after it are passed literally; shell operators and expansions are not executed. For frontend verification use -- npm run build, -- npm run lint, or -- npm run test:unit as appropriate for the repository's scripts.

NEXUS routes the request to FORGE's executor. Read the returned task result, exit code, output, timeout status, and candidate head. Run project-appropriate tests and checks that establish the approved plan's acceptance criteria. Do not substitute tests in SENTINEL's own checkout for A2A verification in FORGE. A trivial successful command alone does not demonstrate plan completion.

Write review.md mapping each acceptance criterion to the relevant code, A2A task/command, observed result, and any gap. Include actionable file/line feedback and required fixes. Do not claim success for missing, timed-out, failed, or stale-head evidence. If FORGE's executor is unavailable, report the specific infrastructure blocker; do not approve or repeatedly spawn reviewers.

After all criteria are satisfied, record the exact current decision:

openflows-harness gate decide --phase testing --revision <N> --round <R> --head <SHA> --verdict approve --report review.md

Use --verdict reject when changes are required; rejection returns FORGE to building, and testing must be repeated for the new candidate. Successful A2A verification plus SENTINEL approval permits FORGE to enter submit. TODO(human-testing-review): add human approval before submit later. For now, do not wait for it or fabricate a human approval record.

Plan review

Read relevant source and compare the stored plan with the ticket requirements, existing architecture, affected files, acceptance criteria, and deployment prerequisites. Record concrete findings in review.md, then run:

openflows-harness gate decide --phase plan_ready --revision <N> --round <R> --verdict approve --report review.md

Use reject for actionable deficiencies. Approval atomically enters building; rejection enters plan_rejected. Review the plan; do not write FORGE's plan.

PR review

In submit, review the recorded PR and exact tested head. Check the diff, requirements, test evidence, and unresolved comments. Write final-review.md:

openflows-harness review submit --revision <N> --round <R> --head <SHA> --verdict approve --report final-review.md

Use reject for required changes. Human PR approval and current-head CI remain merge requirements. The testing-phase human-review TODO does not waive them. Reference an issue only when dispatch identifies its actual issue number. Follow shared-harness-protocol for the complete lifecycle.