anthropic client + interleaved thinking support

[mikasenghaas]

Description

Based on #774. Do not merge before.

This PR adds support for Anthropic clients and interleaved thinking

• Accepts AsyncOpenAI | AsyncAnthropic in public-facing environment methods
• Internally we wrap those native clients with vf.Client which maps between a normalized message, response, and tool types and the client-native types
• New environment setter Environment.set_interleaved_thinking which passes the thinking flag to the client (e.g. OAI clients do a best-effort parsing of reasoning content)

Tested Features

• Single-turn + OAI completions

uv run vf-eval continuation-quality -n1 -r1 -d -v 

• Single-turn + Anthropic completions (didn't test bc didn't find a Anthropic model with completions compatiblity)

uv run vf-eval continuation-quality -n1 -r1 -d -v -m haiku

• Single-turn + OAI chat completions

uv run vf-eval gsm8k -n1 -r1 -d -v 

• Single-turn + Anthropic completions

uv run vf-eval gsm8k -n1 -r1 -d -v -m haiku

• Multi-turn + OAI chat completions

uv run vf-eval alphabet-sort -n1 -r1 -d -v

• Multi-turn + Anthropic messages

uv run vf-eval alphabet-sort -n1 -r1 -d -v -m haiku

• Multi-turn + Tool + OAI chat completion

uv run vf-eval wiki-search -n1 -r1 -d -v

• Multi-turn + Tool + Anthropic responses

uv run vf-eval wiki-search -n1 -r1 -d -v -m haiku

• Multi-turn + OAI chat completions tokens (PRIME-RL vLLM server)

uv run vf-eval alphabet-sort -n1 -r1 -v -d -m Qwen/Qwen3-4B-Instruct-2507 -b http://localhost:8000/v1

• Multi-turn + OAI chat completions tokens (PRIME-RL vLLM server)

uv run vf-eval alphabet-sort -n1 -r1 -v -d -m Qwen/Qwen3-4B-Instruct-2507 -b http://localhost:8000/v1 -x '{"interleaved_rollouts": true}'

• Multi-turn + Tool + Interleaved Thinking + OAI chat completions

uv run vf-eval math-python -n1 -r1 -v -d -m deepseek-reasoner

Client Interface

The vf.Client wraps a native client and asks it to implement normalizers to enable a standard way to obtain a Response. In particular, the following four methods have to be implemented

1. Convert the vf.Messages prompt to a client-native prompts (to_native_prompt)
2. Get a client-native response (get_native_response)
3. (Optionally) Raise and handle exceptions (e.g. overlong prompt) for client-native exception (raise_from_native_response)
4. Convert the client-native response to a vf.Response (from_native_response)

In principle, we could continue to use the OAI-types as our normalizing type system (this would be less breaking), however there are some features from other clients that are difficult/ impossible to bake into the OAI types, and so it seems natural to define a custom type which supports the union of all features.

So far, we implement the following clients:

• OpenAICompletionClient
• OpenAIChatCompletionClient
• OpenAIChatCompletionTokenClient
• AnthropicCompletionClient
• AnthropicMessagesClient

This list should be easily extensible to more client and API types in the future, such as OAI responses.

Backwards compatibility

• The public-facing environment API is fully backwards compatible (client types are extended, oai_tools is converted to tool_defs wherever possible
• We use a CustomBaseModel which allows dict-like access as to not break environments which return messages (e.g. BFCL v3)

Brealing:

• oai_tools is deprecated. Environments which oai_tools in setup_state break silently becaus the environment doesn't read from oai_tools anymore.

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Documentation update
• Test improvement

Testing

• All existing tests pass when running uv run pytest locally.
• New tests have been added to cover the changes

Checklist

• My code follows the style guidelines of this project as outlined in AGENTS.md
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• Any dependent changes have been merged and published

Additional Notes

[mikasenghaas]

merged in #897