Streamable tool calls with Anthropic Models

[ActuallyTaylor]

This PR introduces a streaming transcript object that allows the streaming endpoints to stream tool calls back to the user as the model works.

Features

• Session.transcript automatically updates as content is streamed.
• The Anthropic stream function has been updated with an infinite loop until there are no tool calls.
  • This was mirrored from the OpenAI endpoint.
• (Breaking) Updated Anthropic Thinking structures to match new APIs
  • Changed Thinking.init to take a specified type, budget tokens, and display type instead of a static type.
  • Added two static connivence methods .adaptive(display:) and .enabled(budgetTokens:,display:) to easily create thinking structures configured to match the Anthropic APIs.
• Updated AnthropicLanguageModel tests to reflect the streamable transcripts.

Other Approaches

I first attempted to stream the transcript through a second API named streamTranscript. I decided not to go down this route since it would require duplicated streaming code in the model, as well as the need to make every transcript structure conform to Generable and PartiallyGenerated.

Support

Currently only supported on the Anthropic Provider.

[mattt]

Hi @ActuallyTaylor. Thanks for this, and sorry for not reviewing it sooner.

This and #181 each add a way to observe transcript entries during streaming, for Anthropic and MLX respectively, which is what I asked for in #103. Rather than land two shapes, I'd like to settle on one in #103 and then have both PRs adopt it. I've posted a proposal there; would you take a look?

Separately, the change to Thinking.init is source-breaking. Could you split that into its own PR so it can go in with the next minor?

[ActuallyTaylor]

@mattt I have split this pr into #211 which adds just the thinking and effort changes required to get an anthropic model to think! Also I am happy you are back! No worries about the long review period