Repository navigation
Background task management for codex - seeking approval to submit PR #6404
Description
Hey team! I've built a feature that lets AI agents run background tasks so they can do multiple things at once, and wanted to get approval before submitting a PR (per the contributing guidelines).
The problem this solves
Right now when the AI runs a command, it blocks until that command finishes. This is a problem for:
- Starting dev servers (blocks the AI from making code changes while server runs)
- Running long builds while continuing to work on other tasks
- Monitoring logs while deploying
- Running tests in background while writing more code
What I built
A background task management system that allows the AI to:
- Start shell commands in the background (e.g.,
npm run devkeeps running while AI does other work) - Check output from background tasks anytime using
bash_outputtool - Kill background tasks when done using
kill_shelltool - View all active background tasks via TUI (Ctrl+B shows list of running tasks)
Implementation details
- Built on Tokio for async process management
- Resource limits: max 20 concurrent tasks, 10k line output buffer per task
- Clean process cleanup when task/session ends
- Full test coverage (12 tests)
- Hex-based shell IDs for easy reference
This is related to #3968 - enables the same workflow where AI can start long-running processes and continue working.
Would you be open to reviewing a PR for this?
Thanks!
Activity
etraut-openai commentedon Nov 8, 2025
Thanks for your interest in codex. This looks like an interesting feature.
We're not accepting new PRs for new features at this time. Each new feature requires the involvement of the codex team to ensure that the design is sound, the feature composes well with existing and upcoming features, the feature works with all of the codex surfaces, and the implementation does not introduce regressions and is maintainable.
We have a long backlog of feature requests. We're generally prioritizing based on community feedback. Feature request #3968 has received some "thumbs up" reactions, but there are many features currently ahead of it in the priority queue.
If you have any thoughts about the design of this feature, we'd appreciate it if you'd add those to the the thread for #3968. We'd welcome any insights that you gained from implementing your proof of concept.