| name | dev-verify |
| description | Internal /dev VERIFY beat — independent machine verification over authenticated plan identity and TaskList. |
| user-invocable | false |
| disable-model-invocation | true |
| hooks | {"PreToolUse":[{"matcher":"Write|Edit|MultiEdit|NotebookEdit","hooks":[{"type":"command","command":"bun ${CLAUDE_PLUGIN_ROOT}/hooks/orchestrator-mutation-guard.ts --workflow dev"}]},{"matcher":"Bash","hooks":[{"type":"command","command":"bun ${CLAUDE_PLUGIN_ROOT}/hooks/orchestrator-mutation-guard.ts --workflow dev"}]},{"matcher":"Agent","hooks":[{"type":"command","command":"bun ${CLAUDE_PLUGIN_ROOT}/hooks/approved-artifact-gate.ts --workflow dev"}]}]} |
Dev verification — the verify beat
!bun ${CLAUDE_SKILL_DIR}/../../scripts/load-constraints.ts dev-verify
Review the current receipt-selected generated {planFile, planHash} after the test-gap audit returns
COVERED. Never use REVIEW_STATE.md, VALIDATION.md, LEARNINGS.md, a fixed plan, or a visible
review ledger. TaskList is the live finding/retry record; the result below is the review account.
Independent review loop
- Resolve and rehash the approved dev receipt. Read its requirements, architecture, task criteria,
evidence plan, and review surfaces.
- Dispatch fresh read-only reviewer(s) with the exact plan identity, changed-code scope, fresh test
evidence, and no implementation reasoning. Reviewer and doer must be different identities.
- Require evidence for each reported issue, reject pre-existing or speculative findings, and capture
actionable findings as one current-plan TaskList item each (
item_kind: review-finding) with
requirement/criterion references and retry dependencies.
- Send fixes only through
dev-implement; resume the reviewer after fixes and require a full fresh
re-review. The controller reconciles findings; it does not patch or silently suppress them.
- At the bounded retry limit, return
ESCALATE with open TaskList IDs rather than fabricating an
approval.
Optional Codex second pass
After a primary PASS, Codex is optional. Record the user's launch, join, retry, decline, unavailability,
or completed verdict in the current review TaskList item and the returned result. A launch is not a
verdict: requested must be joined before approval. High-confidence actionable Codex findings create
normal TaskList findings and re-enter the same repair/re-review loop. Do not create
codex-second-pass-*.json or any planning ledger.
Return contract
Dev review result
- Plan: {planFile, planHash}
- Reviewers: [{identity, scope, freshEvidence}]
- Findings: [TaskList IDs with requirement/criterion links]
- Codex: {status: completed | declined | unavailable | retrying, verdict?, reviewTaskId}
- Verdict: APPROVED | CHANGES_REQUIRED | ESCALATE | BLOCKED
- Required next action: [TaskList IDs or dev-accept]
APPROVED means every reviewer has fresh evidence, no actionable current-plan finding remains, and any
chosen Codex pass has returned or is explicitly declined/unavailable. Then continue to dev-accept.