Entering /agent causes the entire terminal to freeze, making all operations impossible
Open 💬 2 comments Opened Aug 1, 2026 by enenH
💡 Likely answer: A maintainer (github-actions[bot], contributor)
responded on this thread — see the highlighted reply below.
What version of Codex CLI is running?
0.146.0
What subscription do you have?
20x
Which model were you using?
gpt-5.6-sol
What platform is your computer?
Win10 x64
What terminal emulator and version are you using (if applicable)?
Windows Powell shell
Codex doctor report
What issue are you seeing?
输入 /agent 会导致整个终端被冻结 无法进行任何操作 只能强行关闭进程
What steps can reproduce the bug?
只需要疯狂输入 /agent 进行不同对话间的切换,就有概率触发
What is the expected behavior?
_No response_
Additional information
_No response_
2 Comments
English translation:
What version of Codex CLI is running?
0.146.0
What subscription do you have?
20x
Which model were you using?
gpt-5.6-sol
What platform is your computer?
Win10 x64
What terminal emulator and version are you using (if applicable)?
Windows Powell shell
Codex doctor report
What issue are you seeing?
Entering /agent causes the entire terminal to freeze, making all operations impossible. The process can only be forcibly closed.
What steps can reproduce the bug?
Simply enter /agent repeatedly and rapidly to switch between different conversations; there is a chance that the issue will occur.
What is the expected behavior?
_No response_
Additional information
_No response_
<!-- codex-issue-translator -->
Hi - I'm also experiencing this issue, but on macOS. I've closed duplicate issue https://github.com/openai/codex/issues/37735. Here's a full copy of my report:
What version of Codex CLI is running?
0.146.1.322 (stable)
What subscription do you have?
None; using AWS Bedrock (company paid)
Which model were you using?
gpt-5.6-sol
What platform is your computer?
Darwin 25.5.0 arm64 arm
What terminal emulator and version are you using (if applicable)?
Ghostty
Codex doctor report
What issue are you seeing?
When the system is under extreme CPU + memory pressure (2-3 parallel xcodebuild processes each all trying to consume all CPU cores, active processes swapping), the UI can hang.
I'm typically talking to the main agent, and I typically have many subagents open as well. First symptom: messages sent to the main agent are displayed in the UI, but responses back are not displayed (confirmed via jsonl inspection that the main agent does respond, but it does not show in the UI).
If I'm in that no-main-responses showing mode, and I do nothing, it will eventually recover.
If I'm in that no-main-responses showing mode, and I instead switch to another agent, then switch back to main (using /side and exiting /side also works), the UI freezes entirely.
When frozen, the UI seems to not process any input. It does not respond to Ctrl+Q, Ctrl+S, newlines, nor typed messages. It does not re-draw upon terminal resizing. The main agent continues to process new turns, but after the UI freezes, if I type input, even with <Enter> at the end, the main agent does not receive it as a message.
What steps can reproduce the bug?
/agentsuntil UI. Alternatively, rapidly launch/side <some ask>conversations and exit them.What is the expected behavior?
At all points, the TUI should be responsive. Under no circumstance should it stop processing user input - any continued input processing - even if delayed - is preferred to a complete freeze/deadlock.
Additional information
After experiencing this bug four or five times, I modified codex locally to hook in
tokio-consolefor further diagnosis. I forked off of 79b4f03d35962b005b007a015113b38930711665 / rust-v0.146.1.I got a codex instance to freeze, then had an LLM analyze it. It produced this timeline and inference:
Precise event timeline
| Time | Type | Event and significance |
| --- | --- | --- |
|
13:24:46.396552625| Observation | Task132, the top-level TUI event-loop task, received its last recorded wake. ||
13:24:47.507392250to13:24:47.507410583| Observation | Task154, the in-process app-server facade worker, performed its final recorded poll. The broader capture records its wake/poll beginning around13:24:47.507383. ||
13:24:47.507399875to13:24:47.507419083| Observation | Task132performed its final recorded poll. It ended approximately 8.5 microseconds after task154's final poll ended. || About
13:24:47.508945to13:24:47.508983417| Observation | Task136, the frame-scheduler task, received and processed its final recorded redraw wake, then became idle. || After
13:24:47.508983417| Observation | Lower in-process app-server tasks141,143, and144continued polling after the three UI-facing tasks stopped. ||
13:30:09.395| Observation | The parent rollout still recorded parent-owned reasoning, messages, and tool activity, rather than only child-process activity. ||
13:56:57.064| Observation | The parent rollout recorded a parentsend_inputcall. ||
13:57:01.105| Observation | The parent rollout recorded a parentwait_agentcall, the latest explicitly verified parent event. ||
13:59:59.832506| Observation | The target Codex process remained alive in stateS+, using 0.5% CPU with 422,256 KiB RSS. |Inference: The facade-first, TUI-next ordering is consistent with task
154becoming pending on one of the awaited outward sends in the lossless-delivery path, followed by task132becoming pending on a facade request. The potentially blocked outward item is either a precedingLaggedmarker or the lossless event itself.Timing alone does not identify the awaited Rust future, queue occupancy, outward item, or pending request method. The deadlock hypothesis therefore depends on both the timing evidence, and source analysis.