How to Wire GPT-5 Codex into a Pull Request Bot

GPT-5 Codex works well as a pull request bot: it reads the diff, comments on real problems, and can push a lint fix to the branch. The design that survives contact with a real team is narrow — review only, or fix only defined failures, never “improve the code.”

What should the bot actually do?

Two jobs are safe and useful. First, review incoming diffs and post line comments for bugs, missing tests, and risky changes. Second, when CI fails on formatting or lint, push a fix commit. Anything broader — refactoring, architecture suggestions — generates noise and gets muted, so skip it.

How do you hook it up?

The flow is straightforward:

1. Trigger on pull_request events and fetch the diff via the GitHub API.
2. Send the diff plus a short repo summary to the Codex model.
3. Post returned findings as review comments, each anchored to a file and line.
4. For lint failures, run the agent with write access on a bot branch and push.

Keep the prompt scoped: “flag correctness bugs and missing tests; ignore style.” Named exclusions reduce false positives more than any model upgrade.

How do you keep review quality high?

• Cap the diff size reviewed — skip PRs over a few thousand lines or split them.
• Require the model to cite the exact line it is commenting on.
• Suppress comments on files it already flagged in the last commit.
• Let a human approve any auto-pushed fix before merge.

Bots that comment once per real issue get kept. Bots that chatter get disabled within a week.

What does it cost per PR?

Reviewing a 300-line diff costs roughly 10k-30k tokens — a few cents. An auto-fix run that iterates with tests is larger, maybe 50k-100k tokens, still under a dollar. On a busy repo with 50 PRs a week, expect single-digit dollars. Caching the repo summary across reviews trims it further.

How does it compare to a full coding agent?

A PR bot is a narrow slice of what droid or similar agents do interactively. The bot’s value is that it runs without a human, on every PR, at low cost. The trade-off is judgment: it cannot ask clarifying questions, so its comments must be conservative. Bias toward silence when unsure.

Quick Summary

• Use GPT-5 Codex as a PR reviewer or lint-fixer, not a general refactorer.
• Fetch the diff via webhook, post anchored line comments.
• Cap diff size and require line citations to cut false positives.
• Costs a few cents per review, under a dollar per auto-fix run.
• Bias to conservative comments; humans approve all auto-pushes.

Last updated September 2026.