codex: MPS model left unset assumes canonical OpenAI names — Azure OpenAI deployment names produce an unusable config

[NeilMazumdar]

Summary

When a Model Provider Service is configured for Codex, write_tool_config deliberately leaves model unset so Codex sends its own canonical name. That assumption holds for providers exposing OpenAI-canonical ids, but Azure OpenAI MPS targets are deployment names, which are routinely not canonical (gpt-5.6-luna, my-gpt5-prod, …). The result is that ucode configure --agents codex against an Azure OpenAI MPS configures Codex onto a model the service has no target for.

This is independent of #323 — it still bites after that 400 is fixed.

Where

ucode/agents/codex.py, write_tool_config:

# With a Model Provider Service the gateway routes by header and Codex sends
# its own canonical model name (e.g. `gpt-5`) — leave `model` unset so no
# Databricks endpoint id is pinned. ...
chosen_model = None if provider else (model or default_model(state))

The comment states the assumption plainly; it just isn't true for EXTERNAL_MODEL_PROVIDER_TYPE_AZURE_OPENAI, where config.targets[].model is whatever the Azure deployment is called.

Impact

A developer runs ucode configure --agents codex, gets a "configured" success, and the first request fails because the model name Codex picked isn't a target on the service. Nothing in the configure output hints that the model is the problem.

Workaround

Author a managed config with an explicit default_model — the launch path pins that, so it overrides Codex's own choice:

{"agent": "CODING_AGENT_CODEX",
 "model_config": {"codex": {"model_provider_service": "<catalog>.<schema>.<service>",
                            "default_model": "<azure-deployment-name>"}}}

Suggested fix

The service's targets are already discoverable via GET /api/2.1/unity-catalog/model-provider-services/<full_name> → config.targets[].model. Options, roughly in order of preference:

1. When a provider is set, read its targets and pin model when the target list doesn't contain the name Codex would otherwise send (or when there is exactly one target).
2. Failing that, validate at configure time and fail loudly — "service <name> has no target matching <model>; targets are: …" — rather than writing a config that cannot work.
3. At minimum, document that default_model is effectively required for Azure OpenAI MPS.