Levels

Every review runs the same way today: rules sort the change, a planner groups it, the experts each group needs review it, and what they find is verified. The cheaper layers below are planned.

Who decides how closely to look

Nobody picks a level. The models decide in three places:

• The planner marks the groups where a mistake costs most as core. A core group gets the correctness, skeptic, security and design experts (The plan).
• Each expert reads everything the change does to its files and reports every problem it finds.
• Each expert grades its findings. error and warning findings are verified; note findings are folded into the nitpicks (Review feedback).

Budget

A review has a budget in tokens and money, and each command the verifier runs has a timeout. There is no limit on total time. Tokens are new input, output and cache writes. Tokens an agent reads back from the provider's cache are recorded apart and do not count against the budget; they grow with every turn an agent takes, so a review's bill is higher than its token count.

• Go checks the budget before starting each agent, so running agents still finish.
• The planner and the experts may spend three quarters; the rest is kept for verification.
• At most eight experts run at once (Running locally).
• When the budget runs out the review starts no more agents and returns what it verified, with result partial.

Planned layers

The plan is to start an agent only when a cheaper check has found something for it:

Layer	Would run	Model cost
L0	The CI check, triage and file types	none
L1	lint, typecheck, secret_scan, dependency_audit, openapi_diff, complexity, coverage_diff	none
L2	The planner, the readability expert, turning tool output into comments	small
L3	The other experts, and reproduction	most

Only L0 exists. Routing by tool output comes with the first L1 tool, together with a measurement of what it loses against the full panel (Evaluation).