exec_command reports running before command dispatch while awaiting hook approval

What issue are you seeing?

When an exec_command tool call is held by a blocking PermissionRequest approval hook, Codex returns this model-visible tool state:

Script running with cell ID 45

The hook command is running, the requested shell command has not started.

That state is indistinguishable from a command that was dispatched and is now hung or legitimately long-running. The agent therefore diagnoses the wrong layer: it polls the cell, searches the process table, terminates the cell, or retries the command. A retry can create another long running hook.

What steps can reproduce the bug?

1. Use Codex with a hook that runs for longer time durations.
2. Run an exec_command that is expected to finish quickly.
3. The LLM thinks the shell command / tool call itself hung and may attempt to debug it although it didn't even run.

Environment:

• Codex CLI 0.154.0, confirmed current by codex doctor
• Model: gpt-5.6-sol
• Platform: Darwin 25.5.0 arm64 arm
• Terminal: VS Code 1.137.0
• unified_exec and hooks enabled

What is the expected behavior?

The tool protocol should expose the pre-dispatch phase explicitly, for example:

{
  "status": "waiting_for_hook",
  "hook_command": "...",
  "command_started": false,
  "cell_id": "45"
}

Only report running after the requested command has actually been dispatched.

At minimum, the model-visible text should say that execution is waiting for hooks to complete and that the shell command has not started.

Additional information

Impact:

• Agents falsely diagnose healthy shell commands or repository hooks as hung.
• Agents terminate cells that haven't really executed.
• Agents retry and can create duplicate long running hooks.
• Users receive incorrect explanations about what is blocking execution.

Related but not duplicate: #39835 requests racing a PermissionRequest hook with the terminal approval prompt. This report concerns the exec_command lifecycle state exposed to the model while a hook command is pending.

The repository was searched across open and closed issues for hook waiting, command running, hanging command, and pre-dispatch lifecycle reports. No issue describing this protocol-state ambiguity was found.

Start at the exec_command lifecycle and the PermissionRequest approval hook, focusing on how the model-visible running state is emitted before dispatch. Trace the state transition from hook approval through command start. Done means hook-blocked calls clearly report that the shell command has not started, while running is reported only after dispatch.