Deep code review

Review the requested state as a connected system, not as isolated files or edited lines. Establish the review mode, intended behavior, reachable flows, invariants, evidence for each conclusion, and what remains unknown.

Treat review as read-only. Do not edit, stage, commit, push, post comments, or resolve threads unless the user separately asks for changes. Durable review-state files are review artifacts, not product edits: use a user-requested tracked location when supplied, otherwise use writable scratch space outside the repository and never stage it. If no such location is writable, ask before adding review state to the repository. If the request includes fixes, finish and report the review before starting a clearly separated repair phase.

0. Calibrate the active reviewer

Before inspecting the repository or running repository commands, read model-and-reasoning-profiles.md. Unless the current request already states both the coordinator model and reasoning level, ask one concise blocking question for the missing values. Ask which model and reasoning level should coordinate the review; state that Terra and Sol coordinators default to risk-routed Luna, Terra, and Sol specialists, while every Luna coordinator is restricted to explicit Luna/max descendants. Ask about the specialist pool only when the user wants to restrict a non-Luna coordinator further. Do not infer either coordinator value from prior tasks, model availability, or a model name without an effort suffix. For Luna at max on a snapshot audit or repository-wide multi-agent review, immediately read luna-max-whole-repository-audit.md and treat it as the controlling orchestration contract.

Record the selected model, reasoning level, mixed-model policy, orchestration profile, concurrency and nesting limits, evidence-slice ceiling, and validation plan. For any Luna coordinator, set the mixed-model policy to prohibited; Luna/max descendants only. Apply the matching profile throughout the review. If an interactive answer is impossible, record unknown, use the strict fallback profile, and disclose the fallback; do not silently choose a model identity.

Model and reasoning calibration changes orchestration, checkpoint frequency, lane size, and validation scheduling. It never weakens the finding contract, read-only boundary, exact-state requirement, or completeness gate.

1. Bind the review to an exact state

First select one mode and record it: