Codex Long-time Orchestrator: a local host for long-running plan, execute, wait, wake, and resume loops

[mayii2001]

I built an open-source project around Codex CLI for long-running engineering loops:

https://github.com/mayii2001/codex-long-time-orchestrator

The problem I wanted to solve is that many real tasks do not fit into a single Codex call. In practice, work often looks more like this:

• plan a change
• execute code edits
• start a smoke test or long-running job
• wait for minutes or longer
• wake the agent again to inspect logs
• fix issues
• resume execution from saved state

This project splits those responsibilities between Codex and a persistent local host process.

What it does

• browser-based main agent chat for planning and progress inspection
• freezeable execution plans
• background task execution
• periodic wake-ups for long-running tasks
• per-project run history
• task process inspection
• resume from interrupted runs
• completed-task skipping on rerun
• compressed checkpoints for long sessions
• repeated execution on the same task can reuse its own resumed Codex session

Current design

• the main agent keeps a durable planner session
• repeated execution on the same task can resume its own session as well
• the orchestrator owns waiting, wake timing, state persistence, and UI
• run state is stored locally under .orchestrator/runs/<run-id>/

Validation

One recent manual validation scenario was a repeated 10-second wake-up loop. I treated the run as successful once checkIteration >= 3, which confirmed that the orchestrator could wake the same task repeatedly instead of surviving only a single wait.

Current limits

This is still an active engineering project, not a fully validated production system. The main flow is usable, but I would still describe the project as experimental.

If people here are building similar long-horizon Codex workflows, I would be interested in how you handle wake-up and resume loops.