Codex ignores structure, can`t finish patches, duplicates files, misreads builds, unsafe assumptions, shallow review, generic tests, patch conflicts, style drift, broad audits.
Codex is a sandboxed cloud agent — great for running tasks independently without burning local session, but it`s also the easiest place to ship "functionally correct, structurally wrong" code: wrong directories, v2 file duplicates, tests that ignore conventions, 50-bullet audits no one can act on. This hub breaks down real Codex failure modes: one symptom per article, with `AGENTS.md` patterns and verification steps. For developers who actually integrate Codex into their workflow, not just demo it.
Codex landed your work on main, on a leftover branch, or split it across two — because cloud tasks default to your repo's default branch and local worktrees collide. How to force clean per-task isolation, plus how to rescue commits that already landed wrong.
Codex hits conflict markers and stalls, picks the wrong side, or commits the markers into the file. Rebase in setup, an AGENTS.md stop-on-conflict rule, and a pre-commit hook fix it.
Codex edited package.json but never ran npm install, so package-lock.json diverges and npm ci fails. Fix it with setup.sh, a CI sync check, and AGENTS.md rules.
Codex's PR is green in CI but the app crashes after deploy. Why agent fixes that target the test surface miss the runtime, and the smoke-gate that closes the gap.
Codex defines a fresh User or ApiResponse type when an identical one already lives in your repo. Make the agent search first using AGENTS.md, a types barrel, ripgrep, and a ts-morph CI check.
Codex ran git commit --amend, an interactive rebase, or force-pushed and rewrote shared history. Lock the agent to forward-only commits with AGENTS.md, the sandbox network toggle, and a GitHub ruleset.
Codex emits a one-line PR body with no detail. Force a structured description (before/after, why, test plan) with a PR template, AGENTS.md, and a CI gate; plus the Codex Cloud button caveat.
Codex reports green tests but the failing cases were filtered out, marked .skip, or bailed early. How to force honest test reporting before merge.
Codex hits a context limit mid-task and leaves a half-applied patch. Scope tasks to one verb, split into multi-PR plans, tune auto-compaction, and resume cleanly with config that works as of June 2026.
Codex injects API patterns deprecated years ago because its training data lags. Lock it to modern APIs with AGENTS.md rules, the @typescript-eslint/no-deprecated lint rule, and current reference docs.
Codex writes a file, Prettier reformats it on save, and the diff explodes. Fix it by running the project's pinned Prettier inside the agent loop and aligning AGENTS.md style hints to .prettierrc.
On 500k+ line repos Codex loses the thread halfway. Fix it by scoping the working set, pre-feeding directory summaries, and putting conventions in AGENTS.md so they survive compaction.
Codex quietly drops PNGs, PDFs, sqlite DBs, and compiled artifacts from audits. Fix it by probing binaries (file + wc -c + sha256sum), allowlisting text extensions, and forcing a no-omit output schema.
Codex re-reads the same file 8 times and re-greps the same query 5 times. Fastest fix: run /plan first, add an AGENTS.md repo map, and keep reasoning effort at medium. Full diagnosis below.
Codex halts halfway through a multi-step task with no error — usually context compaction, a sandbox/approval denial, or a premature stop signal. Diagnose by the last tool call, then resume with codex resume --last.