Plan Review

Run this at the strongest reasoning model available to you. model: inherit takes the session's model, so this is a request rather than a guarantee — if the session is on a small or fast model, escalate before reviewing. A downgraded adversarial review is the failure mode that matters here: it returns generic approval and nothing signals that it was shallow.

Dispatch one fresh reviewer to attack an implementation plan the way an experienced staff engineer would, then fix what it finds before writing code.

This is a gate, and it runs exactly once. Blocking and Worth-fixing findings are fixed before implementation starts. The plan is not sent back for a second review — you verify your own fixes against the plan diff and proceed. One pass, then implement.

That single pass is the whole design. A gate that re-reviews until it is satisfied is unbounded: the previous version of this gate ran to a median of ~5 rounds and a maximum of 9, which cost more than the defects it caught were worth.

When to use

• The user asks to review, sanity-check, grill, or poke holes in a plan
• A plan is written and it is worth a second pair of eyes before implementing