GPT WORLD / INDEPENDENT CODEX FIELD MANUAL Source review · 2026-07-26

Field guide / 14 min

Turn changes into reviewable evidence

A coding agent is useful only when its diff, tests, and review survive inspection.

Track
Codex CLI core
Level
Workflow
Maturity
Stable

Codex can implement, run checks, show the working-tree diff, and launch a dedicated reviewer. Review scopes are mutually exclusive: uncommitted work, a base-branch comparison, a commit, or custom instructions. The reviewer reports findings without changing the tree.

CLI

Verified terminal sequence

TERMINAL / COPY WITH INTENT
  1. /diff
  2. /review
  3. $codex review --uncommitted
  4. $codex review --base main
  5. $codex review --commit <SHA>
A

Working principles

01

Tests prove behavior

Ask for focused regression coverage and the relevant existing checks, not a generic claim that tests pass.

02

Diff is the artifact

Inspect every changed file, including untracked files and edits that predated the agent.

03

Review is a separate lens

Use a dedicated reviewer after implementation so correctness and regression risks are examined independently.

B

Field procedure

  1. 01

    Record the starting state

    Know which changes already exist so Codex does not accidentally claim or overwrite them.

  2. 02

    Define verification

    Name the focused test and the broader check required before completion.

  3. 03

    Inspect /diff

    Compare implementation to the requested scope and reject unrelated cleanup.

  4. 04

    Choose one review target

    Use --uncommitted, --base, or --commit; do not combine conflicting scopes.

  5. 05

    Resolve findings deliberately

    Confirm evidence, fix real defects, rerun checks, and document any accepted residual risk.

PASS / FAIL

Acceptance checklist

  • Starting user changes are preserved.
  • Focused and broader checks are recorded.
  • Every diff hunk is understood.
  • Review findings are resolved or explicitly dispositioned.

WATCH / REJECT

Failure patterns

  • Running review before reproducing the requested behavior.
  • Treating style-only findings as equal to correctness.
  • Letting a fix loop expand into unrelated refactoring.