Check whether a fresh Codex session receives a project instruction
At a glance
Agent Relay Check uses a temporary random marker to test whether a fresh Codex session can return guidance from a chosen AGENTS.md path. This workflow is listed under agent configuration, Codex, AGENTS.md, diagnostics.
- Agents / tools
- Agent Relay Check, Codex CLI, Node.js, AGENTS.md
SourceBuilding a tool for these workflows? See weekly rates.
What they did
1) On macOS or Linux, use Node.js 22+, an authenticated Codex CLI 0.163.0 or newer, and access to gpt-6.1-sol/MAX; extract Agent Relay Check and create a fresh challenge with `node cli.mjs prepare --out <tmp>/before --keys project_probe`. 2) Append the generated instruction block to the AGENTS.md file being checked, keeping the challenge and output outside the project, then start a fresh session and run `node cli.mjs run --project <project> --challenge <tmp>/before/challenge.json --out <tmp>/result-before --live`. 3) Read the local report for RECEIVED, NOT_OBSERVED, or INCONCLUSIVE, then remove only the temporary instruction block; repeat with a fresh challenge after a later change if you want a before/after comparison.
How it went
In the author's macOS example on version 0.1.0, a root and child marker were both RECEIVED; adding an empty child AGENTS.override.md left the root RECEIVED and child NOT_OBSERVED; removing the override restored both. These were the author's own runs, not an independent user trial.
Worth knowing
A live run uses one call from the existing Codex quota. RECEIVED proves only that the exact marker was returned, while NOT_OBSERVED does not prove non-delivery; the check does not certify rule compliance or security. For assembled input alone, `codex debug prompt-input` is the simpler inspection path.
