Codex CLI 0.143.0 on Amazon Linux 2023 returns `unsupported call: exec_command` for basic shell commands

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

• #31609

Powered by Codex Action

I’m seeing what looks like the same regression on macOS, but with a slightly different rejected tool name.

Environment

• OS: macOS Big Sur 11.7.11
• Architecture: macos-x86_64
• Terminal: Apple Terminal
• Codex CLI: 0.143.0
• Install method: pnpm global
• Codex path: /Users/<USER>/Library/pnpm/bin/codex
• Model/provider failing: gpt-5.5 + openai provider / ChatGPT sign-in route
• Auth mode: ChatGPT sign-in
• Sandbox: workspace-write
• Approval policy: on-request

What happens

Basic shell commands fail before execution.

<details>
<summary>Minimal repro</summary>

touch myfile.txt

Codex responds with something like:

I couldn’t create myfile.txt because the shell tool failed to launch.

The command never runs and no file is created.

</details>

Debug log evidence

I enabled logging and reproduced the issue.

<details>
<summary>Logging command</summary>

RUST_LOG=debug codex -c log_dir=./codex-repro-log -m gpt-5.5

</details>

<details>
<summary>Relevant redacted log excerpt</summary>

Output item item=FunctionCall {
  name: "exec_command",
  namespace: Some("exec_command"),
  arguments: "",
  call_id: "REDACTED"
}

ToolCall: exec_commandexec_command {
  "cmd":"source \"$HOME/.zshrc\"\nprint -r -- \"$PATH\"\ntouch myfile.txt\nls -l myfile.txt",
  "shell":"/usr/local/bin/zsh",
  "login":false,
  "workdir":"/Users/<USER>/Desktop",
  "yield_time_ms":10000,
  "max_output_tokens":2000
}

event.name="codex.tool_result"
tool_name=exec_commandexec_command
call_id=REDACTED
duration_ms=0
success=false
output=unsupported call: exec_commandexec_command
app.version=0.143.0
auth_mode="Chatgpt"
originator=codex-tui
terminal.type=Apple_Terminal/440
model=gpt-5.5
slug=gpt-5.5

ERROR codex_core::tools::router: error=unsupported call: exec_commandexec_command

</details>

So in my case, the model output starts as name = exec_command and namespace = exec_command, but the routed tool call becomes exec_commandexec_command. The local router then rejects it before shell execution.

Behavior matrix

<details>
<summary>Working / failing combinations</summary>

gpt-5.4 + ChatGPT sign-in route = works
gpt-5.5 + openai-responses / API key route = works
gpt-5.5 + ChatGPT sign-in route = fails

</details>

I also tested disabling unified_exec and memories, but the same issue still happened.

Notes

This does not look like a shell, sandbox, approval, or project rules denial, because the command is rejected before execution by the tool router.

It looks like a tool-schema / routing mismatch in Codex CLI 0.143.0, specifically on the gpt-5.5 + ChatGPT sign-in route. The issue here reports unsupported call: exec_command; mine reports unsupported call: exec_commandexec_command, but both seem to be the same class of regression: shell tool calls are being emitted/routed in a form the local CLI does not support.

Update: rolling back from Codex CLI 0.143.0 to 0.142.5 fixes the issue for me.

Working/failing matrix:

0.143.0 + gpt-5.5 + ChatGPT sign-in = fails with unsupported call: exec_commandexec_command
0.142.5 + gpt-5.5 + ChatGPT sign-in = works
0.143.0 + gpt-5.4 + ChatGPT sign-in = works
0.143.0 + gpt-5.5 + openai-responses / API key = works

So this looks even more like a 0.143.0 regression in the ChatGPT-auth shell/tool routing path, rather than a local config/sandbox issue.

I can reproduce what appears to be the same issue on Debian 13 using Codex CLI 0.143.0.

Environment:

• Debian 13
• VS Code Remote SSH
• ChatGPT Go authentication
• bubblewrap 0.11.0 installed
• Trusted Git workspace

In my case the router reports:

ERROR codex_core::tools::router:
error=unsupported call: exec_commandexec_command

This happens both in:

• interactive codex
• codex exec

For example:

codex exec "List the files in the current directory."

The shell tool never executes and Codex responds that the shell tool is unavailable.

Installing bubblewrap did not change the behavior.
Updating Codex via codex update also did not resolve the issue.

Suffering from the same issue -- "the shell tool returned an internal “unsupported call” error." (Triedls and pwd on GPT-5.5)

Also, "rolling back from Codex CLI 0.143.0 to 0.142.5" did not completely solve the problem for me -- it only fixes this issue in non-interactive mode (codex exec), but not otherwise.

----

Update: It works for interactive + non-interactive modes now after restarting the app-server via rm -rf ~/.codex/app-server-daemon/

Same issue on fedora 45 amd64, switching back to 0.142.5 fixed the issue

Same issue on ubuntu26.04

Same issue on MacOS 26.5.2

Same issue. macOS 26.5.2, codex v0.143.0.

<img width="819" height="263" alt="Image" src="https://github.com/user-attachments/assets/ca8d0dae-7c0d-484c-aaf5-145007f3c6df" />

I think 0.144.0 fixed it, been using 0.144.1 all day, no issues..