codex-cli 0.143.0: gpt-5.5 interactive sessions fail shell tool routing with unsupported call shell_commandshell_command

Summary

In interactive Codex CLI sessions on Windows PowerShell, gpt-5.5 high fails local shell task execution before the command runs, with:

unsupported call: shell_commandshell_command

The same machine, directory, and file set work under gpt-5.4, which strongly suggests a model-specific or tool-routing issue rather than a sandbox/project problem.

Environment

• Codex CLI: 0.143.0
• OS: Windows PowerShell
• Working directory: D:\aaa_Lyl_File\Documents\research
• Model failing: gpt-5.5 high
• Model working: gpt-5.4

Observed symptoms

• SessionStart hook (failed)
• sometimes timed out after 30s
• sometimes hook exited with code 1
• UserPromptSubmit hook (failed)
• sometimes timed out after 10s
• sometimes hook exited with code 1
• sometimes Stop hook (failed) with hook exited with code 1
• a strict read-only acceptance flow stops at step 1 with:

unsupported call: shell_commandshell_command

Important note

A plain Get-Location response is not sufficient proof of real shell execution, because the cwd may be echoed from session context.

Real verification was done later with values that cannot be guessed:

• generate a fresh random GUID
• Test-Path on a specific README
• Get-FileHash -Algorithm SHA256 on that README

That verification succeeded under gpt-5.4.

Reproduction

1. Start interactive Codex CLI on Windows:

``powershell
codex --sandbox workspace-write --disable hooks
``

1. Use gpt-5.5 high
2. Ask it to execute a simple read-only PowerShell command and return the raw result
3. Observe failure before command execution with:

``text
unsupported call: shell_commandshell_command
``

Control / workaround

Run:

codex --model gpt-5.4 --sandbox workspace-write --disable hooks

Then execute a command sequence that cannot be guessed, for example:

• generate a random GUID
• Test-Path on a known file
• compute SHA256 of that file

Observed result: real shell execution succeeds.

Example successful verification output under gpt-5.4:

NONCE=4a779aaf-6dda-492f-9fc3-13fd685cdba6
README_EXISTS=True
README_SHA256=F57F30892359CCA9434580D653780B8EE45AC627A3D3DEFD8C9F3AD81450B7B7

Expected behavior

• gpt-5.5 interactive sessions should route shell tool calls correctly
• disabling hooks should avoid hook-related interference, but should not still produce unsupported tool-call routing errors

Actual behavior

• gpt-5.5 appears to hit a shell tool namespace/routing failure before the PowerShell command actually runs
• gpt-5.4 works in the same environment

Impact

This blocks repository inspection, acceptance checks, and local code-audit work in gpt-5.5 interactive sessions on this setup.

Extra note

--ephemeral is not accepted by the interactive top-level codex command in this version; that appears to be unrelated, but was observed during debugging.