TensorZero

[!WARNING] Breaking Changes

• parallel_tool_calls now defaults to null (i.e. follows the model provider's default behavior) instead of false (disabled). OpenAI and other providers default to true, so this behavior was counterintuitive.

[!WARNING] Deprecating Functionality

The following features are getting deprecated. Both versions will work for now, but we encourage you to migrate as soon as possible. A future release will remove the deprecated approach.

• When using the OpenAI-compatible inference endpoint, TensorZero-specific arguments (e.g. episode_id, variant_name) should be provided in the request's body using the tensorzero:: prefix. Previously, these arguments were provided as headers (without a prefix).

Bug Fixes

• ClickHouse introduced a bug (https://github.com/ClickHouse/ClickHouse/issues/77848) in their 2025.2 release that affected the inference detail page in the TensorZero UI. This release introduces a workaround for the upstream bug.

Notable Changes

• Add embedded gateway support for openai-python (using tensorzero.patch_openai_client)
• Add recipe notebook for supervised fine-tuning on demonstrations for Together AI
• Support signature field for thought content blocks for Anthropic Extended Thinking
• Support multiple system / developer messages in OpenAI-compatible inference endpoint (via concatenation)

& multiple under-the-hood improvements