Isolated ChatGPT exec with gpt-5.6-luna fails before first turn across CLI 0.146.0 and 0.147.0

What version of Codex CLI is running?

0.146.0 (also reproduced with temporary 0.147.0 and 0.145.0 binaries; global CLI was not changed)

What subscription do you have?

ChatGPT (device-auth login)

Which model were you using?

gpt-5.6-luna

What platform is your computer?

macOS (Apple Silicon)

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

Codex CLI invoked from a macOS terminal; terminal emulator not relevant to the repro

Codex doctor report

Not attached. The full report contains environment diagnostics; the sanitized connectivity, authentication, and model-visibility results are included below.

What issue are you seeing?

Summary

An isolated safety canary using only synthetic, de-identified input cannot complete on Codex CLI 0.146.0 or 0.147.0 with gpt-5.6-luna under ChatGPT authentication. No real conversations, personal paths, media, Wiki, memory content, or credentials were sent.

Observed result

• 0.146.0 and 0.147.0 both emit thread.started and then a server configuration error event before the first response item.
• The result is identical with and without the local Skill content, so the failure appears before Skill processing and output-schema handling.
• 0.145.0 accepts the relevant command options but exits locally before producing any JSONL event, so it is not a usable fallback.

Preconditions verified

• ChatGPT authentication was re-established and verified.
• Provider HTTP reachability and the Responses WebSocket handshake succeeded.
• gpt-5.6-luna is listed as visible and supported_in_api=true.
• The isolated runtime uses a temporary HOME, CODEX_HOME, empty working directory, and temporary auth copy; the auth copy is removed after exit.

Isolation configuration

The run uses read-only sandboxing, ephemeral execution, no user config or rules, JSON output, no approval, no inherited shell environment, and disables shell/browser/MCP/media/memory/plugin-sharing/multi-agent capabilities.

Requested help

Please confirm why ChatGPT-authenticated Codex CLI 0.146.0/0.147.0 returns a server configuration error immediately after the first JSONL lifecycle event when invoking gpt-5.6-luna under this isolation configuration, and whether the 0.145.0 no-JSONL local exit is a known compatibility limitation.

What steps can reproduce the bug?

1. Start from a clean temporary runtime with only a temporary ChatGPT authentication copy.
2. Run a synthetic, de-identified prompt through isolated codex exec with read-only sandboxing, ephemeral mode, no user config/rules, JSON output, no approval, and no inherited shell environment.
3. Invoke gpt-5.6-luna.
4. Observe thread.started followed by a server configuration error on 0.146.0/0.147.0 before the first response item. Repeating with and without the local Skill produces the same outcome.

I intentionally omit the command line, raw error text, IDs, paths, and diagnostics to avoid exposing environment details; the fixed constraints above are sufficient to reproduce the shape.

What is the expected behavior?

The isolated synthetic canary should produce a normal JSONL response lifecycle and a receipt, or return a documented, actionable compatibility error that distinguishes model availability, authentication, isolation configuration, and service configuration.

Additional information

_No response_