Codex Agent Rules

OpenAI's coding agent, running in the terminal.

Overview

Codex packs 122 rules into its root file, covering Rust style, module size, reviews, snapshot tests, and app-server APIs. Several rules limit the size or scope of a change, not just the code it touches.

Context budgeting Hard prohibitions Verification by change type Good and bad pairs Generated file guards Behavioral framing

Documents

├── 
└── codex-rs
    └── tui
        ├── src
        │   └── bottom_pane
        │       └── 
        └──

Techniques in this file

01 / Context budgeting

Limits on the model context

Model history is incremental to preserve prompt caching. Every injected item has a hard size cap, anything over 1,000 tokens needs review, and fragments use a named trait.

973. No unbounded items - everything injected in the model context must have a bounded size and a hard cap.

A section arguing against the largest crate

The file says codex-core grew because adding code was easier than moving it out. New work should use an existing crate or a new workspace member, and reviewers are told to push back on further growth.

Source excerpt starting at line 76.
76To that end: **resist adding code to codex-core**!

A line budget on the diff

Non-mechanical changes stop at 800 lines and complex logic at 500. Work over either limit returns as a staged proposal based on the diff, dependencies, and call sites, starting with the smallest coherent change.

127Unless the change is mechanical the total number of changed lines should not exceed 800 lines.

Module size limits with the affected files listed

Modules target 500 lines excluding tests, and past roughly 800 new functionality goes in a new module. The rule names the specific files it applies to most, on the stated grounds that they already attract unrelated changes.

Explain protected sandbox variables

Two environment variables are off limits because existing tests use them to skip cases the agent cannot run in its sandbox. That rationale keeps the checks from looking like dead code.

9  - You operate in a sandbox where `CODEX_SANDBOX_NETWORK_DISABLED=1` will be set whenever you use the `shell` tool. Any existing code that uses `CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR` was authored with this fact in mind. It is often used to early exit out of tests that the author knew you would not be able to run given your sandbox limitations.

06 / Verification by change type

Test scope that escalates, with a permission gate

Run the changed project's tests first. Common, core, or protocol changes require the full suite after user approval, and slow Rust commands must not be killed by PID.

62- When running Rust commands (e.g. `just fix` or `just test`) be patient with the command and never try to kill them using the PID. Rust lock can make the execution slow, this is expected.

Ideas for your repo

Put a line budget on changes, and ask for a staging plan when one exceeds it.
Name the modules that are already too big. An agent cannot infer which files attract unrelated changes.
Record constraints code cannot express where agents will see them before deleting related checks.
Gate expensive test runs behind asking, and state which changes make them mandatory.