Regression in 0.154.0: Astra experimental context management disabled by runtime supports_experimental_context=false

What version of Codex CLI is running?

codex-cli 0.154.0

What subscription do you have?

ChatGPT Pro

Which model were you using?

gpt-6-astra

What platform is your computer?

Linux 4.15.0-76-generic x86_64 x86_64

What issue are you seeing?

Experimental context management works with gpt-6-astra on Codex CLI 0.153.4, but does not activate on 0.154.0 on the same Linux machine and ChatGPT Pro account.

On 0.154.0:

• Bundled catalog: supports_experimental_context = true
• Runtime catalog: supports_experimental_context = false
• <context_window> and <context_window_guidance> are not injected.

On 0.153.4, the same configuration successfully injects both tags.

This looks like a regression related to the new runtime model capability gate in 0.154.0.

What steps can reproduce the bug?

1. Enable experimental context management:

[features]
context_management.experimental_mode = true

2. With Codex CLI 0.154.0, check Astra metadata:

codex debug models

Observed runtime value:

supports_experimental_context = false

while:

codex debug models --bundled

reports:

supports_experimental_context = true

3. Test context-management activation:

codex -m gpt-6-astra \
  -c features.context_management.experimental_mode=true \
  debug prompt-input probe \
  | grep -Eo '<context_window>|<context_window_guidance>' \
  | sort -u

On 0.154.0: no output.

4. Run the same command with Codex CLI 0.153.4 on the same machine.

Observed:

<context_window>
<context_window_guidance>

The only changed variable is the Codex CLI version: 0.153.4 → 0.154.0.

Start by comparing Astra metadata from codex debug models with codex debug models --bundled, then follow the runtime capability check used by debug prompt-input probe. Done means experimental context management activates on Codex CLI 0.154.0 and the probe emits both <context_window> and <context_window_guidance> with the documented configuration.