pydantic-ai 2.54.0

v2.54.0 (2026-10-02)

on Python PyPI

What's Changed

⚠️ Compatibility Notes

• Handle draft-7 list-form items in JsonSchemaTransformer, TestModel and Mistral streamed output, and make TestModel accept boolean subschemas and cap prefixItems at maxItems by @dsfaccini in #9560
• Fix ImageGeneration under FallbackModel and dataclasses.replace, and compare Embedder by identity by @dsfaccini in #9477
• Refuse a second durable execution engine on an agent before either binds by @DouweM in #9391
• Make wrap_* hooks enclose complete stage lifecycles by @adtyavrdhn in #7053
• Re-raise model errors from Temporal model activities with their original type in workflow code by @DouweM in #9580
• End the run when a BackgroundTools tool raises an unexpected exception, as a sequential tool does by @DouweM in #9434
• Drop the SubAgents agent_folders default-change warning: since v2.52.0 (#9023) the agents folder is not loaded by default, so pass agent_folders='agents' to keep loading it by @DouweM in #9424

🚀 Features

• Tint clai2 file diffs from the selected /theme palette by @mpfaffenberger in #9648
• Make clai2 /update always install the newest build and restart into it, resuming the session by @mpfaffenberger in #9659
• Disable RepoContext inventory by default by @DouweM in #9382
• Give the CodeMode missing-return-schema warning its own CodeModeReturnSchemaWarning category by @DouweM in #9423
• Add cancel_and_resume example: cancel and resume a streaming run by @DouweM in #7012
• Reconnect dropped OpenAILiveModel sessions by forking a stored session or replaying the history by @DouweM in #9401
• Default Pydantic AI-created HTTP clients to the OpenAI SDK's connection pool limits, and make create_async_httpx2_client() public by @DouweM in #9600
• Use gpt-live-transcribe for OpenAI realtime's default input_transcription_model='auto' and list gpt-transcribe by @DouweM in #9402
• Support WebSearchTool on OpenAI GPT-Live through its delegated backend by @DouweM in #9049
• Add native to ExaSearch and YouSearch to prefer the model's native web search, and web_search_tool() for WebSearch(local=...) by @DouweM in #9386
• Add a double-Esc conversation rewind menu to clai2 by @mpfaffenberger in #9688
• Allow deleting unused models from the clai2 /model menu by @mpfaffenberger in #9696

🐛 Bug Fixes

• Fail closed in the shell realpath fallback when readlink cannot read a symlink by @dsfaccini in #9613
• Stop importing the undeclared sniffio package in pydantic-evals and pydantic-ai-harness by @pydanty in #9620
• Fall back from ImageGenerationTool to the local tool on Gemini text models, and keep tool output when a native tool falls back locally on Gemini 2.5 by @dsfaccini in #9551
• Nest worktrees under their repository in the clai2 resume sidebar by @mpfaffenberger in #9667
• Remove the clai2 resume-directory confirmation by @mpfaffenberger in #9682
• Report clai2 startup plugin failures through observability by @mpfaffenberger in #9681
• Show file diffs in compact clai2 tool output by @mpfaffenberger in #9687
• Treat a null finish_reason on a SnowflakeModel response like an empty one by @DouweM in #9590
• Refuse a LocalFileStore root another user owns, and log ToolOutputLimits spill and summarize fallbacks by @DouweM in #9020
• State when the ConversationSearch default-scope warning will be removed by @DouweM in #9383
• Warn when read_only=True removes every hosted MCP tool by @DouweM in #9387
• Pin the request that started the current run, not the first prompt of the conversation, in the TrajectoryJudge window by @DouweM in #9425
• Identify the search_tools tool by its tool_kind instead of its name by @DouweM in #9603
• Find the tool search tool in CodeMode by its tool_kind, so a prefixed search tool still gets its notes by @DouweM in #9604
• Log LLMReminder generation failures instead of silently falling back to GoalReanchor text by @DouweM in #9024
• Include function tools run after the final result in run_stream() usage by @DouweM in #9694
• Build the same ModelResponse from streamed and complete GoogleModel responses, and add Gemini 3 web search sources by @DouweM in #9594
• Recover run state with RunCancelled.from_cancellation() when a DBOS workflow is cancelled by @DouweM in #9394
• Make custom setup_logfire callbacks replay-safe and add replay_safe to LogfirePlugin by @DouweM in #9698
• Raise ModelAPIError for undecodable 200 response bodies from Anthropic, Bedrock, Groq, Hugging Face, Google, and OpenAI embeddings and images by @DouweM in #9350
• Keep the run state recoverable when Ctrl-C interrupts run_sync() by @DouweM in #9605
• Clamp OpenAI realtime barge-in truncation at the profile's audio_output_sample_rate by @DouweM in #9399
• Load a capability once when load_capability is called twice for it in one response by @DouweM in #9397
• Keep a first-position ToolAvailabilityDeltaPart announcement out of the standing system prompt by @DouweM in #9413
• Commit an ACP turn up to its usage limit instead of rolling it back by @DouweM in #9426
• Preserve dependency extras in clai2 archive updates by @mpfaffenberger in #9690
• Preserve text and reasoning part boundaries in OpenAIChatModel streams by @pydanty in #8729
• Propagate a native cancel that lands during a command timeout's stop by @dsfaccini in #9675
• Return the live token from PKCESignIn.token() when another clai2 session saved one without a refresh token by @dsfaccini in #9670
• Keep BubblewrapWorkspace commands from running code on the SSH host through shell startup files or the SSHWorkspaceBackend stop script's PATH (VERIA-212) by @mpfaffenberger in #9673

Full Changelog: v2.53.0...v2.54.0