fix(models): allow slash-containing model IDs with unknown prefixes

[OiPunk]

Summary

Fixes #2492 by making MultiProvider treat unknown prefixes as plain model names instead of hard-failing with Unknown prefix.

This keeps explicit provider prefixes (openai/, litellm/, and prefixes configured in provider_map) working as before, while allowing model IDs that naturally contain slashes (for example openrouter/openai/gpt-5) to be sent to the default OpenAI provider unchanged.

What changed

• Added _is_known_prefix() in MultiProvider.
• Updated get_model() routing logic:
  • Known prefix: keep prefix-based routing.
  • Unknown prefix: do not treat it as provider prefix; route as plain model name.
• Added new tests in tests/models/test_multi_provider.py covering:
  • unknown prefix passthrough
  • known fallback prefix routing
  • provider_map routing
  • helper map mutations
  • fallback cache behavior
  • unknown fallback error path
  • dynamic litellm fallback import path

Validation

• uv run --with ruff ruff check src/agents/models/multi_provider.py tests/models/test_multi_provider.py
• uv run mypy src/agents/models/multi_provider.py tests/models/test_multi_provider.py
• uv run --with pytest pytest -q tests/models/test_multi_provider.py
• uv run --with coverage --with pytest coverage run -m pytest -q tests/models/test_multi_provider.py
• uv run --with coverage coverage report -m src/agents/models/multi_provider.py

Coverage result for changed source file:

• src/agents/models/multi_provider.py: 100%

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 9b438acb60

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Preserve fallback routing for custom prefixes

Clearing prefix for every non-hardcoded/non-provider_map prefix means get_model("custom/..." ) can no longer reach _get_fallback_provider, so subclasses that extend MultiProvider by overriding _create_fallback_provider lose custom prefix routing. Before this change, those prefixes flowed into fallback creation; now they are silently sent to the OpenAI provider with the unmodified model string, which is a behavior regression for existing custom MultiProvider extensions.

Useful? React with 👍 / 👎.

[seratch]

These changes may actually work, but we generally recommend using either LiteLLM or model object for non-OpenAI models. Making the parser complex for further flexibility (the litellm was the special case because we generally recommend using litellm for non-OAI models) may not be a change we would like to have.

[OiPunk]

Thanks for the clear guidance, this makes sense.

I agree we should keep the parser behavior conservative and align with the current recommendation (LiteLLM or explicit model objects for non-OpenAI providers). I can close this PR to avoid adding parser complexity unless you prefer a much narrower variant.