agent-resume: Skill for Codex

Delayed Codex Resume

Use AgentResume for a one-time callback into the current Codex session. The session itself schedules the wait, then its resumed instruction performs the requested follow-up. This avoids inventing a scheduler, but it is intentionally not a durable job system.

Preconditions

• The request authorizes a one-shot callback and its cost.
• The dedicated session has the agent_resume MCP tool configured; otherwise it cannot schedule its own continuation.
• Choose the model when starting the session. AgentResume freezes that model from the current Codex state together with the thread id and passes it to codex exec resume; it must not silently inherit a later CLI default.
• Probe a requested model with a tiny codex exec --ephemeral -m <model> call before relying on it. If unavailable, report the exact error; never silently substitute a different model.

Read the transport contract before changing this workflow or using it in an unfamiliar environment.

One-shot workflow

1. Start a dedicated, non-ephemeral Codex session with the chosen model. Keep its task narrow and prohibit unrelated tools or edits.

2. In that session, call agent_resume.wait_and_resume with:

   {
     "wait_seconds": 60,
     "note": "One-shot callback: reply exactly НАДО.",
     "execute_resume": true
   }
   

   The initial prompt must say that the continuation, after the callback, must reply with the requested exact text.

3. Record the returned job id. Do not create a cron job, systemd timer, or retry loop for this test.

4. Observe for at most the requested delay plus two minutes, then verify all of the following:

   • agent_resume.wait_job_status reports state: "finished";
   • its meta.json records resume_result.executed: true, a zero launch_returncode, resume_result.model equal to the initial model, and resume_result.log_file;
   • resume_result.log_file contains the expected response. If a resume PID is recorded, it is no longer alive before declaring the test complete.

5. Report elapsed time, selected model, job id, and the exact evidence. A zero launch return code merely means the resumed Codex process started; it does not prove the response was produced. If the callback fails, preserve the log and diagnose before attempting another one.