| name | loop-review-agents |
| description | Use only when the user explicitly asks to run loop-review-agents, invokes /loop-review-agents, or requests repeated independent review rounds over changes, a commit, a plan, a PRD, or another review target. |
| disable-model-invocation | true |
Loop Review Agents
Run repeated three-agent review rounds and apply only consensus findings with clear value. This is an opt-in workflow; do not use it for ordinary reviews unless the user invokes it.
Review Target
Infer the target from the current context or command arguments:
- current uncommitted changes,
- the last commit or branch diff,
- an accepted implementation plan,
- a PRD or design document,
- another explicit target named by the user.
If the target is genuinely ambiguous, ask one concise clarification before starting.
Round Workflow
- Launch three independent reviewers against the same target. Do not share reviewers' same-round outputs with each other.
- Ask each reviewer to focus on high-signal issues: correctness, contracts, security/privacy, accessibility, data loss, tests/build, release/publishing risk, and user-visible regressions.
- Merge findings into a coordinator ledger.
- Apply feedback only when it meets one of these thresholds:
- at least two reviewers strongly agree on a high-value finding,
- all three reviewers agree on a medium-value finding.
- Do not apply taste-only style feedback, unsupported speculation, or findings already rejected in a prior round unless new evidence appears.
- After applying accepted feedback, run another review round against the updated state.
Value Levels
- High value: likely bug, contract break, security/privacy issue, accessibility issue, data loss, failing test/build, release/publishing risk, or user-visible regression.
- Medium value: maintainability, test coverage, documentation, or API clarity issue with concrete impact and agreement across all reviewers.
- Low value: style, preference, naming taste, speculative refactor, or change without demonstrated impact.
Treat reviewers as strongly agreeing when they identify the same underlying issue, assign similar severity, cite evidence, and point toward compatible fixes.
Coordinator Ledger
Maintain this ledger during the loop:
- Round: review pass number and target reviewed.
- Applied: finding, consensus evidence, files changed, and verification performed.
- Rejected or Deferred: finding, reviewers who raised it, and rationale.
- Already Handled: repeated findings from prior rounds and where they were addressed.
- Human Review Pointers: disputed or variable areas the user should inspect if the loop does not converge.
Use the ledger to avoid bouncing between stylistic alternatives across rounds.
Stop Conditions
- Stop when a round produces no qualifying consensus findings.
- If more than 5 rounds are needed, narrow the remaining review focus and explicitly look for taste or variability churn.
- If 10 rounds complete without convergence, stop. Ask the user to review and report the specific disputed areas, accepted changes, rejected findings, and remaining uncertainty.
Final Report
Report:
- whether the final state is acceptable as-is,
- rounds completed,
- applied consensus fixes,
- findings rejected or deferred and why,
- verification run or not run,
- human-review pointers if convergence failed.