v1.0.3 (2025-09-10)

What's Changed

• Include thinking parts in subsequent model requests to improve performance and cache hit rates by @DouweM in #2823
• Add on_complete callback to AG-UI functions to get access to AgentRunResult by @ChuckJonas in #2429
• Support ModelSettings.seed in GoogleModel by @DouweM in #2842
• Add with agent.sequential_tool_calls(): contextmanager and use it in DBOSAgent by @DouweM in #2856
• Ensure ModelResponse fields are set from actual model response when streaming by @DouweM in #2848
• Send AG-UI thinking start and end events by @DouweM in #2855
• Support models that return output tool args as {"response': "<JSON string>"} by @shaheerzaman in #2836
• Support NativeOutput with FunctionModel by @DouweM in #2843
• Raise error when WebSearchTool is used with OpenAIChatModel and unsupported model by @safina57 in #2824

New Contributors

• @safina57 made their first contribution in #2824
• @shaheerzaman made their first contribution in #2836

Full Changelog: v1.0.2...v1.0.3

v1.0.2 (2025-09-08)

What's Changed

• Add support for durable execution with DBOS by @qianl15 in #2638
• Support sequential tool calling by @strawgate in #2718
• Add GoogleModelSettings.google_cached_content to pass cached_content by @Hojland in #2832
• Add ModelResponse.finish_reason and set provider_response_id while streaming by @fatelei in #2590
• Add support for gen_ai.response.id by @Kludex in #2831
• Only send tool choice to Bedrock Converse API for Anthropic and Nova models by @DouweM in #2819
• Handle errors in cost calculation in InstrumentedModel by @alexmojaki in #2834

New Contributors

• @Hojland made their first contribution in #2832
• @qianl15 made their first contribution in #2638
• @fatelei made their first contribution in #2590

Full Changelog: v1.0.1...v1.0.2

v1.0.1 (2025-09-05)

Breaking Changes

The following breaking change was slated to go into v1.0.0 but accidentally left out. Because it's a security issue, an uncommonly used feature, and few people will have updated to v1 yet within 24 hours, we decided it was justified to make an exception to the no-breaking-changes policy to get it out ASAP.

• Remove Python evaluator from pydantic_evals for security reasons by @dmontagu in #2808

Other Changes

• Remove eval-type-backport dependency by @Viicos in #2814

Full Changelog: v1.0.0...v1.0.1

v1.0.0 (2025-09-04)

What's Changed

• Drop support for Python 3.9 by @Kludex in #2725
• Deprecate OpenAIModelProfile.openai_supports_sampling_settings by @Kludex in #2730
• Add support for human-in-the-loop tool call approval by @DouweM in #2581
• Add tool_calls_limit to UsageLimits and tool_calls to RunUsage by @tradeqvest in #2633
• Add LiteLLM provider for OpenAI API compatible models by @mochow13 in #2606
• Add identifier field to FileUrl and subclasses by @kyuam32 in #2636
• Support NativeOutput with Groq by @DouweM in #2772
• Add docstring_format, require_parameter_descriptions, schema_generator to FunctionToolset by @g-eoj in #2601
• Gracefully handle errors in evals by @dmontagu in #2295
• Include logfire with pydantic-ai package by @Kludex in #2683
• Let almost all types used in docs examples be imported directly from pydantic_ai by @DouweM in #2736
• Bump temporalio to 1.17.0 by @DouweM in #2811
• Default InstrumentationSettings version to 2 by @alexmojaki in #2726
• Remove cases and averages from eval span by @DouweM in #2715
• Make many more dataclasses kw-only by @dmontagu in #2738
• Don't emit empty AG-UI thinking message events by @DouweM in #2754
• Update mcp package version by @BrokenDuck in #2741
• Raise error if MCP server __aexit__ is called when _running_count is already 0 by @federicociner in #2696
• Fix error when streaming from Gemini includes only executable_code or code_execution_result by @binaryCrossEntropy in #2719
• Close original response when retrying HTTP request by @DouweM in #2753
• Agent.__aenter__ returns Self, use default instrumentation for MCP sampling model by @alexmojaki in #2765
• Fix Anthropic streaming usage counting by @DouweM in #2771
• Create separate ThinkingParts for separate OpenAI Responses reasoning summary parts by @DouweM in #2775
• Handle Groq tool_use_failed errors by getting model to retry by @DouweM in #2774
• Raise error when trying to use Google built-in tools with user/output tools by @DouweM in #2777
• Move mcp-run-python to its own repo by @samuelcolvin in #2776
• Fix Azure OpenAI streaming when async content filter is enabled by @frednijsvrt in #2763
• Don't emit AG-UI text message content events with empty text part deltas by @celeritatem in #2779
• Handle streaming thinking signature deltas from Bedrock Converse API by @DouweM in #2785
• Don't require MCPServerStreamableHTTP and MCPServerSSE url to be a keyword argument by @franciscovilchezv in #2758
• Add operation.cost span attribute to model request spans, rename ModelResponse.price() to .cost() by @alexmojaki in #2767
• Ensure that old ModelResponses stored in a DB can still be deserialized by @DouweM in #2792
• Type ModelRequest.parts and ModelResponse.parts as Sequence by @moritzwilksch in #2798
• Always run event_stream_handler inside Temporal activity by @DouweM in #2806
• Document that various functions need to be async to be used with Temporal by @DouweM in #2809

New Contributors

• @BrokenDuck made their first contribution in #2741
• @federicociner made their first contribution in #2696
• @g-eoj made their first contribution in #2601
• @binaryCrossEntropy made their first contribution in #2719
• @Trollgeir made their first contribution in #2729
• @kyuam32 made their first contribution in #2636
• @richhuth made their first contribution in #2680
• @frednijsvrt made their first contribution in #2763
• @celeritatem made their first contribution in #2779
• @mochow13 made their first contribution in #2606
• @franciscovilchezv made their first contribution in #2758
• @moritzwilksch made their first contribution in #2798

Full Changelog: v0.8.1...v1.0.0

v1.0.0b1 (2025-08-30)

What's Changed

• Drop support for Python 3.9 by @Kludex in #2725
• Add support for human-in-the-loop tool call approval by @DouweM in #2581
• Deprecate OpenAIModelProfile.openai_supports_sampling_settings by @Kludex in #2730
• Gracefully handle errors in evals by @dmontagu in #2295
• Include logfire with pydantic-ai package by @Kludex in #2683
• Remove errors when passing Retrying instead of RetryConfig to TenacityTransport by @DouweM in #2717
• Default InstrumentationSettings version to 2 by @alexmojaki in #2726
• Remove cases and averages from eval span by @DouweM in #2715
• Make many more dataclasses kw-only by @dmontagu in #2738

Full Changelog: v0.8.1...v1.0.0b1

v0.8.1 (2025-08-29)

What's Changed

• Add gen_ai.system_instructions attribute to agent run spans by @alexmojaki in #2699
• Bump temporalio to 1.16.0 by @DouweM in #2703
• Rename StreamedRunResult methods to be consistent with AgentStream by @DouweM in #2692
• Deprecate specifying a model name without a provider prefix, and the vertexai provider name by @DouweM in #2711
• Rename ModelResponse.provider_request_id to provider_response_id by @DouweM in #2710
• docs: Add message_history parameter documentation for CLI methods by @ryx2 in #2695

Full Changelog: v0.8.0...v0.8.1

v0.8.0 (2025-08-26)

What's Changed

• Add elicitation callback support to MCP servers by @yamanahlawat in #2373
• Add message_history parameter to agent.to_cli() by @ryx2 in #2674
• Properly deserialize complex tool arguments with Temporal by @DouweM in #2686
• Fix serialization / deserialization of FileUrl.media_type by @mwildehahn in #2677
• Handle missing token details in vLLM/OpenAI-compatible APIs by @DouweM in #2669
• Make AgentStreamEvent union of ModelResponseStreamEvent and HandleResponseEvent by @DouweM in #2689

Full Changelog: v0.7.6...v0.8.0

v0.7.6 (2025-08-26)

What's Changed

• Replace all_messages_events with pydantic_ai.all_messages for InstrumentationSettings(version=2) by @alexmojaki in #2658
• Fix inability to call response.raise_for_status in AsyncTenacityTransport by @dmontagu in #2668
• Deprecate OpenAIModel in favor of OpenAIChatModel by @Kludex in #2676
• anthropic: drop new lines on empty system prompt by @Kludex in #2678
• fix(bedrock): skip SystemPromptPart with empty content by @essamgouda97 in #2672
• BREAKING CHANGE: Fix tenacity implementation for improved retry behavior by @dmontagu in #2670
• Add Cerebras provider by @Kludex in #2643

New Contributors

• @essamgouda97 made their first contribution in #2672

Full Changelog: v0.7.5...v0.7.6

v0.7.5 (2025-08-25)

What's Changed

• Handle 'STOP' finish_reason in GeminiStreamedResponse by @ArneZsng in #2631
• Add price() method to ModelResponse by @Kludex in #2584
• Include thoughts tokens in output_tokens for Google models by @alexmojaki in #2634
• Add span_id and trace_id to EvaluationReport by @Kludex in #2627
• Allow proper type on AnthropicProvider when using Bedrock by @akoshel in #2490
• Use new OpenTelemetry GenAI chat span attribute conventions by @alexmojaki in #2349
• Ensure content is always set for assistant tool call messages for OpenAI. by @vimota in #2641

New Contributors

• @ArneZsng made their first contribution in #2631
• @akoshel made their first contribution in #2490
• @vimota made their first contribution in #2641

Full Changelog: v0.7.4...v0.7.5

v0.7.4 (2025-08-20)

What's Changed

• Fix bug with google model safety handling by @dmontagu in #2066
• Add takes_ctx arg to Tool.from_schema by @dedeswim in #2615
• feat: support Google's url_context builtin tool by @vinnymeller in #2604
• Add missing UrlContextTool into all by @Kludex in #2617
• Drop assertion on Google streaming by @Kludex in #2618

New Contributors

• @dedeswim made their first contribution in #2615
• @vinnymeller made their first contribution in #2604
• @ivo-1 made their first contribution in #2610

Full Changelog: v0.7.3...v0.7.4