Feature request: per-tool `eager_input_streaming` so `assistant.tool_call_delta` fires early for Anthropic models

What

The recently added assistant.tool_call_delta provides access to incremental tool call deltas. For recent OpenAI models (e.g. GPT 5.5), it works as expected to send deltas smoothly as they are generated. However, for Anthropic models (e.g. Claude Opus 4.8), the deltas arrive as a burst because Claude buffers them. Anthropic provided the eager_input_streaming flag for the expected behavior: https://platform.claude.com/docs/en/agents-and-tools/tool-use/fine-grained-tool-streaming#how-to-use-fine-grained-tool-streaming. But the SDK hasn't exposed this option yet.

Proposed Usage Example

defineTool("my_tool", {
    description,
    parameters,
    handler,
    eagerInputStreaming: true, // or a different name
});

Why

Responsiveness: for tools capable of handling partial arguments, this improves responsiveness when the whole response might take many seconds - sometimes minutes - to generate.

Commencez par suivre le flux de assistant.tool_call_delta et l’API de définition par outil utilisée par l’exemple defineTool proposé, puis examinez où les options de modèle Anthropic sont envoyées. Le travail est considéré comme terminé lorsqu’une option eager input streaming par outil est exposée et fait parvenir les tool-call deltas progressivement pour les modèles Anthropic, sans modifier le comportement existant.