Codex 5.4 (high) says it will do something and stops, requires my intervention

Resolved 💬 9 comments Opened Mar 12, 2026 by fullofcaffeine Closed Mar 29, 2026

💡 Likely answer: A maintainer (github-actions[bot], contributor) responded on this thread — see the highlighted reply below.

What version of Codex CLI is running?

codex-cli 0.114.0

What subscription do you have?

Pro

Which model were you using?

gpt-5.4-high

What platform is your computer?

Darwin 24.4.0 arm64 arm

What terminal emulator and version are you using (if applicable)?

iTerm2

What issue are you seeing?

I've been having similar issues lately where codex will just stop when it should keep going. In this case, it said it'd do something, but then stopped. I had to manually type and submit "continue".

This is related (and might be a duplicate of) this one: https://github.com/openai/codex/issues/14303: I noticed that Codex would say it was watching a process (e.g test run, compiler running) but in reality it wasn't really in the agentic loop, it just stopped. I actually had to babysit the process and nudge the agent. It seems to be related to the "background agent" feature? Before it'd just start in the foreground and block the agent, but it'd not break the agentic flow. I can see the appeal of "async" background processes, but the agent should be notified when they finish no? Is that how it works now?

What steps can reproduce the bug?

Uploaded thread: 019cb669-5422-7ca0-b7bb-0229699c49bf

What is the expected behavior?

_No response_

Additional information

_No response_

View original on GitHub ↗

9 Comments

github-actions[bot] contributor · 4 months ago

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

#14303
#14314
#14341

Powered by Codex Action

nilswx · 4 months ago

I use subagents a lot and no longer have to wait or _"continue"_ anywhere since adding this to my AGENTS.md.

## Autonomy Policy
If you're working towards goals, do NOT end your turn. This allows for continuous autonomous work.

The user will interrupt you when required, but they will mostly provide steering messages.

Do not pester the user by ending your turn after a unit of work, as that requires them to keep nudging you to keep working.

You MUST continue working autonomously towards any known objectives until the user interrupts you. Do NOT end your turn until there is absolutely nothing left to do.

They're always going! Caffeinated goblins. Also works for dumber models on OpenRouter, etc.

fullofcaffeine · 4 months ago

@nilswx Oh, interesting! Will try it out. Thanks!

EDIT:

Caffeinated goblins

My type of goblins!

Taeu · 4 months ago

Still seeing this on codex-cli 0.114.0 on macOS, and the painful part is that permissive execution settings do not solve it. Even with effectively YOLO / full-access style runs, the agent still falls into handoff behavior like "next step is..." or requires a manual continue nudge instead of just executing the next known step.

Prompt / AGENTS workarounds help, but they are not reliable enough for long-run harnesses. I think Codex needs an explicit opt-in autonomy mode:

do not ask whether to continue while known work remains
if the workflow gate or task file already defines the next step, execute it in the same turn
stop only for real blockers, destructive approval boundaries, or completion
preserve objective, in_progress, and next_action_1 across compaction/resume

This feels closely related to both #4357 and #5957: from the user's side the issue is still "I have to babysit the session even when I intentionally enabled a low-friction run mode."

fullofcaffeine · 4 months ago

morozow · 4 months ago

I fixed the issue discussed here: https://github.com/openai/codex/issues/14731. <s>I'm waiting for permission to create the PR and merge it, unless this entire</s>. I created PR, fix is already in this commit https://github.com/openai/codex/commit/c2845daf318e70a08ccb579c99a7b45fc8c423dc and a patch release is needed.

Update: PR is closed due to restricted contribution policy

ariccio · 4 months ago

Ya, very frustrating

uehlbran · 3 months ago

Yeah, I'm still seeing this non-sense as well. Ask codex to do something and it acknowledges your request and doesn't do anything or half asses it lol.

etraut-openai contributor · 3 months ago

Thanks for the bug report. We'll share the feedback with the team responsible for training our codex models. We're tracking this feedback and similar reports so we can improve future models.

@fullofcaffeine, I wasn't able to look at your logs because we don't have any records of the thread ID you posted. Perhaps the /feedback upload failed.