review-loop

Ask the agent to do the task using review-loop, optionally setting a quality gate and loop limits. The agent completes the work, spawns a separate critic subagent that scores it 1-10 with feedback, revises, and repeats until the gate is met or max loops are reached.

Use a reviewer subagent with fresh context that receives only the work product, not the worker's reasoning. This avoids anchoring bias and produces honest 1-10 scores with specific, actionable feedback on issues like missing validation or error handling.

Yes, Mode B reviews existing work. Point the agent at the file or module, and it reads the code, spawns the reviewer subagent, fixes flagged issues, and re-reviews until the quality gate is met.

The reviewer must be equal to or more capable than the worker. Use a balanced model for standard tasks and an advanced reasoning model for complex work like security or distributed systems, since a weaker reviewer cannot catch subtle logic errors.

Skip it for simple fixes, one-liner changes, exploratory spike work, and tasks where tests are the real quality gate. It also requires clear acceptance criteria; define those first or the loop may not converge.

If the score stalls for two consecutive loops with the same feedback, the worker model is likely too weak to act on the critique. Escalate to a more capable model, lower the gate temporarily, or split the task into smaller pieces.