Custom LiteLLM model alias gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna fails; /model switch poisons session

What happened

I'm routing Codex to a self-hosted LiteLLM proxy (custom model_provider) fronting Amazon Bedrock / Bedrock Mantle. All three gpt-5.6-* aliases (gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna) fail to route, while gpt-5.5 on the same provider works.

Setting model = "gpt-5.6-sol" (or -terra/-luna) in ~/.codex/config.toml fails; model = "gpt-5.5" works. This suggests Codex applies built-in, model-name-specific handling for the recognized gpt-5.6-* slugs instead of forwarding the model string opaquely to the custom provider.

There's also a session-state bug: switching to a gpt-5.6-* model mid-session via /model breaks the session, and switching back to gpt-5.5 afterwards also fails, even though gpt-5.5 worked before the switch.

Expected

• The model string should be forwarded to the custom provider as-is; a name that happens to match a built-in slug should not change routing/request shaping when model_provider is a custom provider.
• A failed /model switch should not corrupt session state; switching back to a previously-working model should restore working behavior.

Steps to reproduce

1. Configure a custom provider in ~/.codex/config.toml:

```toml
model = "gpt-5.5"
model_provider = "litellm"

[model_providers.litellm]
name = "LiteLLM"
base_url = "https://<your-litellm>/v1"
env_key = "LITELLM_API_KEY"
```

1. Expose the aliases on the LiteLLM side (config.yaml). Both point at Bedrock Mantle OpenAI models; the only difference is the alias name:

```yaml
model_list:

• model_name: gpt-5.5

litellm_params:
model: bedrock_mantle/openai.gpt-5.5
aws_region_name: us-east-1

• model_name: gpt-5.6-sol

litellm_params:
model: bedrock_mantle/openai.gpt-5.6-sol
aws_region_name: us-east-1
```

1. Run codex with model = "gpt-5.5" -> works.
2. Change to model = "gpt-5.6-sol" (or /model -> gpt-5.6-sol) -> fails.
3. Switch back to gpt-5.5 via /model -> now also fails (session stays broken).

Actual result / error

No error is surfaced and nothing is written to the logs. Instead, Codex hangs — it waits indefinitely with no response and eventually times out. The request appears to never complete (or never returns anything the TUI can render) for the gpt-5.6-* aliases, whereas gpt-5.5 responds normally. This looks less like a hard rejection and more like the request either isn't sent in a form the backend answers, or the response isn't parsed — leaving the session stuck.

Environment

• Codex version: 0.145.0
• OS: macOS 26.5.2
• Interface: TUI
• Provider: LiteLLM proxy -> Amazon Bedrock Mantle (OpenAI-compatible)
• wire_api: not set (provider default)

Notes

Possibly related: #14276 (custom provider slug matching), #30697 (model slug handling).