Codex ignoring rules at the bottom of your AGENTS.md? It stops loading instructions at 32 KiB

#What changed

Nothing new shipped; this is a documented limit that is easy to miss. When Codex starts a run it builds one instruction chain: your global ~/.codex/AGENTS.md (or AGENTS.override.md), then one AGENTS.md per directory from the repo root down to where you launched it. It joins them root-first and stops adding files once the combined size reaches project_doc_max_bytes, which is 32 KiB by default. Whatever comes after the cut, usually the most specific folder-level rules, never reaches the model. Tip spotted via @CodexUpdate on X (https://x.com/CodexUpdate/status/2109316872843985119), checked against OpenAI's AGENTS.md guide.

#Who is affected

Anyone with a long global AGENTS.md or a big repo-root file, especially monorepos where the nested files closest to the code are the ones that get dropped.

#What to do now

1. Check which files actually loaded: from the folder you work in, run codex --ask-for-approval never "List the instruction sources you loaded."
2. Trim the global and root files, and move team- or service-specific rules into AGENTS.md files inside those folders.
3. If you really need more, raise the cap in ~/.codex/config.toml, for example project_doc_max_bytes = 65536, then start a new Codex session.

#What is not confirmed

OpenAI's docs describe the cap but not exactly how a file that crosses the limit is cut, so treat anything near the boundary as possibly missing. A higher cap also means more of every prompt goes to instructions.