What "never approves" means for an AI code review tool

Updated: 2026-10-02.

Code Terrier publishes COMMENT reviews. It never publishes APPROVE or REQUEST_CHANGES, and never merges, pushes, or commits code. The people working on the repository retain authority to change and merge it.

Read evidence, report findings

The Pi Durable review agent uses codemode to read and search a confined immutable checkout. It has no shell, network, filesystem mutation, or GitHub publication capability. Local terminal reviews receive the uploaded diff instead of a complete checkout.

The Worker owns policy, verification, and publication. Separating those responsibilities limits what a model tool call can do. It does not eliminate incorrect findings or prove the security of every change.

Verification has limits

When an independent checker is available, findings are evaluated in a separate context before publication. If none is available, that limitation is recorded rather than disguising the original finding as independently checked. A different context need not mean a different model family.

A clean review means no findings were reported within the evidence and execution limits. It is not approval, a guarantee, or a replacement for tests and human judgment. Infrastructure failures and missing terminal reports cannot certify a clean result.

Follow-up stays under human control

Trigger modes determine when reviews run. Auto mode rechecks eligible later pushes; on-creation and manual modes constrain automatic work. Replies do not start a conversation-model loop. On a later review, Adopted and Skipped findings are suppressed and resolved; Escalated findings remain visible while duplicate publication is suppressed.

Suggestions are optional patches to inspect. Security, authentication, and payment findings ask questions instead of inviting blind patch application. Install Code Terrier if a reviewer that reports evidence and leaves the merge decision to people fits your workflow.