§1 — Scope and sources
This audit compares the current Codex instruction and skill-loading contract with the workstation conventions built for Claude Code. Sources were the official Codex manual, the user-global agent file, repository and nested agent files in Ambix and Nova, the installed user skill directories, and live Codex skill discovery in this session.
§2 — Findings
| Surface | Observed contract | Gap |
|---|---|---|
| User instructions | Codex looks for ~/.codex/AGENTS.md. | The canonical workstation file is ~/.agents/AGENTS.md; the Codex entrypoint is absent. |
| Project instructions | Codex builds one root-to-launch-directory chain. | Nested files below the launch directory are not dynamically added when later editing those paths. |
| Context budget | Project instruction bytes are capped by configuration. | Ambix root plus nested guidance exceeds the default 32 KiB budget. |
| Skills | Metadata is discovered from user and repository .agents/skills; bodies load when triggered. | Useful Claude-only skills are not visible, and two HPC copies are stale against current H200 policy. |
| Nova | Main and develop checkouts carry different root instructions. | Stale branch guidance includes git actions prohibited by the current global policy. |
§3 — Required evidence
The implementation must include black-box Codex launches from repository root and nested directories, captured instruction-chain manifests, byte-limit tests, skill metadata/body activation tests, and negative fixtures proving stale or out-of-scope files are not silently treated as loaded. Evidence must record Codex version and configuration.