Add model_name, agent_name, and response_id to RequestUsage for better tracking

[ihower]

Resolved: #2100

This PR adds three useful fields to RequestUsage: model_name, agent_name, and response_id.

These fields make it easier to:

• analyze usage across different models
• track usage by agent
• correlate each usage record with the exact response ID for inspection

Note

model_name is taken from the Model#model attribute. All real subclasses of Model (such as OpenAIResponsesModel, OpenAIChatCompletionsModel, and LitellmModel) do initialize a model attribute.

However, the parent Model class does not define this field, which leads to type-checking errors. So this PR adds the attribute to the base class. To avoid breaking custom Model implementations (for example, FakeModel in tests that do not set a model name), the default value is an blank string.

[seratch]

in fact, it seems available models have self.model but this could be potentially a breaking change for some custom models. it will require a minor version upgrade and release note annoucements

[seratch]

Can we get rid of this property this time? When a developer uses a ModelProvider, this property could be absent, and it sounds surprisingly inconsistent.

[seratch]

these properties are in the request usage data, so i don't think these data should be passed this way. at least, we should avoid passing these in order to keep flexibility of future enhancement (meaning having additional-properties argument, which can accept more properties in the future)

[ihower]

@seratch Thanks for the code review.

I’ve updated the PR so that all of these fields go into a single metadata dict. This should keep future enhancements flexible.

Regarding model.model, I’m not fully sure about the breaking-change risk. Since the base Model class now provides a default value "", any custom model that inherits from Model will have the attribute and won’t raise errors.

Are you concerned about custom models that don’t inherit from Model at all (basically duck typing)?

If we want to avoid touching the Model class to fully eliminate this risk, I can switch to a runtime check when calling add(), something like:

if hasattr(model, "model") and model.model:
    metadata["model_name"] = model.model

I can update the PR if you prefer this approach. Just let me know. Thanks!

[github-actions]

This PR is stale because it has been open for 10 days with no activity.

[seratch]

There is no rush but once this project starts working on 0.7 series changes, resolving the conflicts would be appreciated

[ihower]

@seratch Thanks! I’ve resolved the conflicts and rebased onto the latest main.

[ihower]

Rebased again on latest main to resolve new conflicts.

[seratch]

Can we get rid of this property this time? When a developer uses a ModelProvider, this property could be absent, and it sounds surprisingly inconsistent.

[seratch]

As mentioned above, I'd like to avoid having model_name this time. Also, I've added endpoint: str | None = None (responses.create, responses.compact, or None) to TS SDK for supporting responses.compact API: openai/openai-agents-js#785 We'll add the same to the Usage data class in Python SDK. So, for these SDKs' consistency plus better developer experience, I'd like to avoid dict[str, Any] here and simply having agent_name: str | None and response_id: str: None to request usage data.

[ihower]

I’ve removed the metadata dict and now put the fields directly on Usage.

I think model_name is the most important part here. The original issue was about tracking model cost.
Since the base Model provides a default value of "", this should not be a breaking change.
If a developer uses a custom provider and doesn’t set a model name, it should be easy to add once when they notice. This is also consistent with the other built-in Model implementations in the SDK, which all provide a model name.

Thank you for reconsidering.