pydantic-ai 2.57.0

v2.57.0 (2026-10-15)

on Python PyPI

What's Changed

⚠️ Compatibility Notes

• Attach namespaced origin and tool_call_id metadata to BackgroundTools completions by @michaeldunn9 in #10196
• Take GPT-Live realtime history, usage, and reply waits from the session core by @DouweM in #10083
• Honour disable-model-invocation and user-invocable in Skills, and skip stray skill library paths by @DouweM in #10378
• Take Gemini Live realtime history, usage, and reply waits from the session core by @DouweM in #10082

🚀 Features

• Add MetaProvider for Meta Muse models, with meta: using the Responses API by @mpfaffenberger in #10156
• Add Harness ApprovalReviewer to let a model approve low-risk tool calls before a person sees them by @mpfaffenberger in #10143
• Price requests at their reported service tier by @adtyavrdhn in #10244
• Add dataset_name to EvaluationReport by @jirikuncar in #10307
• Forward MCP Apps tool results to AG-UI and Vercel AI frontends and add MCPToolset.read_resource_contents by @adtyavrdhn in #10179
• Relay MCP Apps tool calls with MCPToolset.call_tool_for_app by @adtyavrdhn in #10190

🐛 Bug Fixes

• Fix token-based compaction retaining too little recent history by @AndrewTilson in #9784
• Record and stream one part per realtime output item when its transcript arrives late by @pydanty in #10097
• Reference multimodal URL choices in the ToolReturnContent discriminator mapping by @Kludex in #10239
• Stop OpenAIResponsesModel sending unavailable tool_choice names in allowed_tools by @dsfaccini in #10232
• Send Gemini tool-call thought signatures (extra_content) back from OpenAIChatModel by @dsfaccini in #10235
• Give a tool's prepare its own retry count under .prefixed(), .renamed() and PrefixTools by @dsfaccini in #10219
• Support Gemini agentic media processing in GoogleModel by @miguel-flowstate in #8028
• Stop granting native tool_search to gpt-5.4-nano by @dsfaccini in #10272
• Omit unset optional fields from Vercel AI UIMessage JSON so the AI SDK accepts it by @adtyavrdhn in #10175
• Keep Planning tool loops in the prompt cache, and attach a leading CachePoint to the previous user message on OpenRouterModel by @mpfaffenberger in #10257
• Release concurrency slots after streamed request cleanup by @dsfaccini in #10225
• Backfill an empty thinking field on current-turn tool calls in OpenAIChatModel's default 'auto' mode by @dsfaccini in #10251
• Keep available_tool_names and active_capability_ids on run context copies inside Temporal activities and Render child tasks by @dsfaccini in #10280
• Reconnect an xAI realtime session past max_duration when a reconnect policy is set by @DouweM in #10056
• Await ProcessEventStream cleanup after early exit and cancellation by @dsfaccini in #10341
• Count the request of a GPT-Live backend response that ends while the session closes by @DouweM in #10084
• Warn when FallbackModel response handlers are skipped on a streamed request by @DouweM in #10066
• Close the HTTP client of a model a run builds from a model name, and reuse one per name inside async with agent: by @DouweM in #10067
• Make waits between continuation polls durable DBOS.sleep steps under DBOSDurability and DBOSAgent by @DouweM in #10383
• Charge the partial response of an interrupted stream in SpendLimits by @DouweM in #10069
• Stop WarnNearLimits from deleting user prompts that quote its [WarnNearLimits] marker by @DouweM in #10391

New Contributors

• @AndrewTilson made their first contribution in #9784
• @miguel-flowstate made their first contribution in #8028
• @jirikuncar made their first contribution in #10307
• @michaeldunn9 made their first contribution in #10196

Full Changelog: v2.56.0...v2.57.0