Shift note: PR review with an agent and a checklist

A pull request is a proposed change to a shared codebase. Reviewing it means checking whether the change solves the intended problem, fits the surrounding code, and avoids unwanted effects. A coding agent can help by reading files, tracing behavior, and running permitted checks. A written checklist gives that work a clear scope and gives the human reviewer a record to inspect.

The useful result is a set of findings supported by code references, reproducible examples, or test output. The reviewer still decides whether the change is ready to merge.

Start the review with the change request

Before asking an agent to inspect code, explain what the pull request should accomplish. Include the original bug report or feature requirement, the branch being reviewed, and any constraints that matter. For example, an update to password reset behavior might need to preserve the existing response format used by mobile clients.

Ask the agent to read relevant repository instructions and compare the proposed branch with its intended base. It should inspect nearby callers and tests when needed. A changed function can look correct while breaking an assumption elsewhere in the application.

For a first review, keep the assignment narrow. Ask for correctness and regression risks, with style comments only when the repository explicitly requires them.

Write checklist questions that require evidence

A useful checklist turns broad concerns into questions the agent can investigate. Keep a shared version in the repository, then add questions specific to the pull request.

• Does the implementation satisfy the stated requirement, including the reported failure case?
• Do existing callers still receive the response fields, types, and errors they depend on?
• Are missing values, empty inputs, boundary values, and failed dependencies handled as intended?
• Does the change preserve the required authentication, authorization, and handling of sensitive data?
• Do relevant tests cover the changed behavior, and which checks were actually run?
• Could the change affect configuration, stored data, deployment order, or rollback?

Each answer should point to supporting evidence or state what remains unknown. A claim that tests passed should include the command and result. If the environment prevented a check, the report should say so.

Give the agent a review assignment it can follow

A starting instruction is: “Review this pull request against the attached checklist. Explain each suspected defect with a file location, the conditions that trigger it, and its likely effect. Separate confirmed findings from questions. Report the checks you ran and any limits on your review. Do not edit files during this pass.”

Suppose a pull request changes a checkout calculation. A finding might explain that a missing discount value reaches an arithmetic operation and produces an invalid total. The reviewer can then inspect that path and request a focused test. A general suggestion to “improve validation” provides much less help.

Verify findings before requesting changes

Read each finding against the actual code. Agents can misunderstand framework behavior, overlook a guard in another file, or flag a path that cannot occur. Ask for the triggering input or a minimal reproduction when the explanation is incomplete.

Prioritize findings by their effect on users and operations. Incorrect permissions or lost data deserve attention before naming preferences. When a finding is valid, request a focused fix and a suitable regression test. After the author updates the pull request, check the revised behavior and confirm that the original issue is resolved.

Choose access that fits the review workflow

If your team uses several coding tools, Relay365 is an option for managing model access. It is a pay-as-you-go relay desk for multiple model families and works with Claude Code, Codex, Gemini CLI, and OpenAI SDK clients. Its API base is https://api.api365relay.com; OpenAI clients add /v1.

For example, you could use GPT-6 Astra for a checklist review, subject to your tool configuration and repository requirements. The same evidence standards should apply whichever model you choose. Create a Relay365 account if that access setup fits your workflow.

Close the review with a short handoff: confirmed findings, completed checks, unresolved questions, and the human reviewer’s decision. That record helps the next person understand what was examined and what still needs attention.