Repository navigation
Releases
Releases · pydantic/pydantic-ai
v2.69.0 (2026-11-16)
What's Changed
⚠️ Compatibility Notes
- Fold
RequestInputintoHandleDeferredToolCalls, whose handler can take a resume token by @DouweM in #13285 - Support Gemini 4 Hydrogen (
gemini-4-hydrogen) onGoogleModelby @dsfaccini in #13299 - Generalize response-scoped tool call IDs into the
tool_call_ids_are_response_scopedprofile flag by @DouweM in #13128 - Add
service_tiersupport toGroqModelby @dsfaccini in #13521 - Detect loops in streamed thinking, text and tool call arguments in
DetectLoopsby @mpfaffenberger in #12222 - Make
Advisoracceptnative=andlocal=like the otherNativeOrLocalToolcapabilities, deprecatingmodeby @adtyavrdhn in #13149 - Treat nested CodeMode
ToolFailedcalls as failed results by @pydanty in #13076 - Keep settled sub-agent task histories in their saved files instead of in every
DelegationTasksowner's memory by @DouweM in #13221 - Raise
ModelQuotaExceededErrorfor Anthropic's tier spend cap and user-set spend limits by @dsfaccini in #13269 - Send structured tool return items next to a file to the model as JSON, not as escaped JSON strings by @dsfaccini in #13271
- Raise a gave-up transport status or connection error from Google embeddings, image generation and
GoogleModel.count_tokensasModelAPIErrorby @dsfaccini in #13312 - Raise a Gemini API error from
GoogleImageGenerationModel.generateas its error category, likeModelRateLimitErrorby @dsfaccini in #13335 - Raise an error status a retrying transport gave up on during a streamed Gemini long decode as
ModelHTTPErrorby @dsfaccini in #13356 - Raise a gave-up transport status from OpenAI and Cohere embeddings, OpenAI image generation and
GoogleModelbatch methods asModelHTTPErrorby @dsfaccini in #13327 - Cache opaque Prefect run IDs across retries by @dsfaccini in #11896
- Reject a
TextOutputfunction that takes a model, dataclass orTypedDictwhen the agent is created by @dsfaccini in #13422 - Keep the body of an error status that
HTTPX2TenacityTransportandAsyncHTTPX2TenacityTransportgave up on by @dsfaccini in #13411 - Raise OpenAI API and connection errors from
OpenAIImageGenerationModel.generateas their error category, likeModelRateLimitErrorby @dsfaccini in #13433 - Raise realtime connection errors with an HTTP status as their error category, like
ModelRateLimitErrorby @dsfaccini in #13436 - Keep Bedrock GPT-5.6 and GPT-6 encrypted reasoning out of Claude-on-Bedrock requests by @pydanty in #12943
- Raise
ModelContextWindowExceededErrorfor Mistral3059prompt-too-long errors, not for other3051invalid-argument errors by @pydanty in #12882 - Raise a status a retrying transport gave up on as the error category its body names, like
ModelQuotaExceededErrorby @dsfaccini in #13485 - Stop counting OpenRouter's
is_byokflag as a usage detail by @pydanty in #13025 - Map Gemini
IMAGE_RECITATIONandTOO_MANY_TOOL_CALLSfinish reasons tocontent_filteranderrorby @pydanty in #10918 - Make
TestModelintegers satisfy mixed inclusive and exclusive bounds, and fractional bounds by @pydanty in #13298 - Add
minimax_model_profileand use it for MiniMax models on Bedrock,OpenAICompatibleProvider,LiteLLMProviderandTogetherProviderby @pydanty in #12693 - Map Mistral Large 4
thinkinglevels to itslow/medium/highreasoning_effortvalues by @pydanty in #12876 - Apply
FunctionToolset'sstrict,sequential,defer_loadingandinclude_return_schemadefaults toToolinstances by @pydanty in #12572 - Make
TestModelstreamed usage matchrun()for tool calls and text by @pydanty in #11877 - Run the token count
count_tokens_before_requestasks for in its own durable unit by @dsfaccini in #13460 - Skip a failing
TieredCompactiontier on model errors, and let tiers omit their trigger by @mpfaffenberger in #13660 - Rename harness
subagents,filesystemandspendpackages to their capability names and align store,ACP*and*Errorclass names by @mpfaffenberger in #13401 - Rename the
DockerSandboxharness capability toDockerContainer, keepingDockerSandboxas a deprecated alias by @adtyavrdhn in #13245 - Rename
BubblewrapSandboxtoBubblewrappedWorkspaceand clarify its docs by @mpfaffenberger in #13179
🚀 Features
- Trim
Coder's default instructions and addCoder(autonomy=...)by @mpfaffenberger in #12542 - Fold
RequestInputintoHandleDeferredToolCalls, whose handler can take a resume token by @DouweM in #13285 - Support Gemini 4 Hydrogen (
gemini-4-hydrogen) onGoogleModelby @dsfaccini in #13299 - Add
is_retryto find retries in message history and list the v2.60.0RetryPromptPartchange as breaking by @DouweM in #11404 - Move benchmark-only
Coderinstructions into opt-inBenchmarkInstructionsby @mpfaffenberger in #13395 - Generalize response-scoped tool call IDs into the
tool_call_ids_are_response_scopedprofile flag by @DouweM in #13128 - Add peer messaging with
Peersinpydantic-ai-harnessby @dsfaccini in #11671 - Add an
anthropic_inference_geosetting for Anthropic data residency by @rscholz98 in #12769 - Add opt-in recursive node registration with
GraphBuilder.add_reachable()by @dsfaccini in #12706 - Add
WindowsWorkspace, a PowerShell workspace for Windows hosts by @mpfaffenberger in #13475 - Add
service_tiersupport toGroqModelby @dsfaccini in #13521 - Add
KubernetesSandboxfor kubernetes-sigs/agent-sandbox pods by @adtyavrdhn in #10612 - Add
E2BSandbox(network=), passing E2B's network options through to a new sandbox by @adtyavrdhn in #13587 - Add
BoxdSandbox(secrets=), naming the boxd secrets a new machine gets by @adtyavrdhn in #13589 - Add
pin(source=)so harness compaction re-injects tool- and model-written pins as framed data, not the user's words by @mpfaffenberger in #13615 - Let
Advisor.max_usestake a callable so a consultation budget can span several runs by @mpfaffenberger in #13617 - Add
DynamicWorkflowToolset.create_executorso subclasses can customize how sub-agent calls are scheduled by @mpfaffenberger in #13623 - Detect loops in streamed thinking, text and tool call arguments in
DetectLoopsby @mpfaffenberger in #12222 - Add
openai_safety_identifiermodel setting by @pydanty in #12258 - Add
DaytonaSandbox(secrets=), mounting Daytona secrets in a new sandbox by @adtyavrdhn in #13590 - Add
OpenSandbox(network_policy=)andOpenSandboxBackend(credential_proxy=), passing OpenSandbox's egress options through by @adtyavrdhn in #13591 - Wait a model error's
retry_afterdurably before DBOS retries a model request, via a sharedmodel_error_retry_afterhelper by @DouweM in #13270 - Add
CapabilityCreation.load_active(workspace), which refuses a storedirectorybelow'operator'authority before loading it by @mpfaffenberger in #13641 - Report
pydantic-ai-harnessin MongoDB handshake driver metadata by @pydanty in http...
CLAI2 bleeding
Sdists rebuilt from main for /update on the CLAI bleeding channel. Not a release.
v2.68.0 (2026-11-11)
What's Changed
⚠️ Compatibility Notes
- Note text that looks like an application frame in user and tool content instead of escaping it by @DouweM in #11813
- Add
OutputTokenLimitExceeded, and fall back to the next model inModelFallbackwhen a response is cut off atmax_tokensby @DouweM in #12407 - Ask for approvals and answers inside the run with
RequestInput, putAskUseron the deferred pipeline, and mark deferred calls withkindandanswered_byby @DouweM in #12757 - Add
DelegationTasks.discover()and stop one damaged saved task from blockingDelegationTasks.opened()by @mpfaffenberger in #10140 - Require workspace backends to leave detached processes running, and check it in
WorkspaceBackendSuiteby @mpfaffenberger in #13070 - Run
PromptInjectionDefendersemantic classification as durable operations so it works on Temporal by @mpfaffenberger in #12424 - Give sub-agent task-control events the
Eventsuffix and past-participle names, withagent_nameandparent_task_idfields, keeping the old names as deprecated aliases by @mpfaffenberger in #13074 - Spell out abbreviated identifiers in
pydantic_ai_harness, renamingnum_resultstomax_resultsandmax_duration_secstomax_duration_secondsby @mpfaffenberger in #13075 - Rename harness
CapabilityEventfields to*_count,*_name,duration_msandprevious_item, keeping the old names as deprecated aliases by @mpfaffenberger in #13068 - Rename harness
CapabilityEventoutliers to<Subject><PastParticiple>Eventand addSTEP_PERSISTENCE_EVENTSby @mpfaffenberger in #13065 - Add the
RefineOutputharness capability by @mpfaffenberger in #13120
🚀 Features
- Load and save UI adapter conversations through a
ConversationStorewithconversation_storeby @DouweM in #12326 - Use the AI Gateway's streamed cost for
GoogleModelandBedrockConverseModelstreams by @ddanielcruzz in #13168 - Support running
SubAgentsdelegations to durable delegates as Temporal and DBOS child workflows by @mpfaffenberger in #12516 - Add the
RefineOutputharness capability by @mpfaffenberger in #13120 - Add a per-command
cwdtoWorkspace.runby @mpfaffenberger in #12642 - Require workspace backends to leave detached processes running, and check it in
WorkspaceBackendSuiteby @mpfaffenberger in #13070 - Add
DelegationTasks.discover()and stop one damaged saved task from blockingDelegationTasks.opened()by @mpfaffenberger in #10140 - Bound the wait for the detached process to start in the
WorkspaceBackendSuitereattach rule by @mpfaffenberger in #13081 - Ask for approvals and answers inside the run with
RequestInput, putAskUseron the deferred pipeline, and mark deferred calls withkindandanswered_byby @DouweM in #12757 - Add
OutputTokenLimitExceeded, and fall back to the next model inModelFallbackwhen a response is cut off atmax_tokensby @DouweM in #12407 - Allow a
CancellationTokenfired from a workflow signal to cancel a Temporal durable run by @DouweM in #7210 - Note text that looks like an application frame in user and tool content instead of escaping it by @DouweM in #11813
- Add a
GoogleDecisionsModelbackend for the Gemini Decisions API by @dsfaccini in #13181
🐛 Bug Fixes
- Resolve
vendor/modelids through one case-insensitive family map onOpenAICompatibleProvider,LiteLLMProviderandTogetherProviderby @DouweM in #12743 - Map unified
thinkingtoreasoning_effortfor Z.AI GLM and DeepSeek V3.2 on Bedrock by @EdAbati in #11491 - Honor
Permissionsremembered approvals inside durable execution by @mpfaffenberger in #12410 - Make
AnthropicModel.count_tokenssend the request's betas and raise its all-deferred toolsUserErrorby @dsfaccini in #13098 - Drop the account-age claim from
AnthropicStaleThinkingBlockWarning, since Claude Fable 5.2 and Mythos 5.2 bind thinking blocks on every account by @dsfaccini in #13155 - Run
PromptInjectionDefendersemantic classification as durable operations so it works on Temporal by @mpfaffenberger in #12424 - Carry
ModelResponse.canonical_systemthrough the Vercel AI and AG-UI event streams by @DouweM in #13134 - Lower a
max_tokensabove a Claude model's maximum output to it instead of sending it by @DouweM in #12427 - Report when
TieredCompactioncan't reachtarget_tokensbecause its tiers keep more than the target by @DouweM in #12650 - Summarize the tail messages
SummarizingCompaction(keep_user_messages=True)copies displace instead of dropping them by @DouweM in #12652 - Complete a Gemini Live turn that ends silently after a tool result by @DouweM in #13126
📦 Dependencies
- Raise the
genai-pricesfloor to 0.2.3 so Anthropic fast mode responses are priced at fast mode rates by @dsfaccini in #13104
New Contributors
Full Changelog: v2.67.0...v2.68.0
v2.67.0 (2026-11-10)
What's Changed
⚠️ Compatibility Notes
- Support
SubAgents(dynamic_agents=...)under durable execution by @mpfaffenberger in #12418 - Strip every default-ignorable character before
ArtifactScannerphrase matching and flag hidden variation selectors by @mpfaffenberger in #11475 - Load one instruction file per directory and nested files by default in
RepoContext, addwalk_up, and haveCoderbuild on them by @DouweM in #11691 - Cap inlined nested instructions at 32,000 characters per conversation by default with
RepoContext(nested_max_chars=...)by @DouweM in #11699 - Send documents of any media type
mimetypesknows, with the format taken from it by @arjun-krishna1 in #12365 - Move provider-specific native-tool fields to
provider_settingsby @adtyavrdhn in #7807 - Add
framing_announcementsetting and only announce the<system>voice tag once a conversation needs it by @DouweM in #12661 - Surface nested
RepoContextinstruction files when the model writes or edits in a directory by @DouweM in #11701 - Record
InputGuardrail,OutputGuardrailandToolGuardrailverdicts as durable operations by @mpfaffenberger in #12428 - Refuse a
CapabilityCreationrun whendirectoryis below'operator'authority by @DouweM in #12558 - Stop continuing an OpenAI response without a
history_digestthroughopenai_previous_response_idby @DouweM in #12575 - Replay Prefect tasks on flow retry when the message history holds binary content by @sangkyoonnam in #12188
- Reject
CodeModetools named after Python builtins, which the sandbox called instead of the tool by @adtyavrdhn in #12698 - Break chat-template special tokens in content on open-model routes, and restore them in model output, with
ModelProfile.may_tokenize_special_tokens_in_contentby @DouweM in #12556 - Use the engine's codec and
serialization_failurefor durable capability results by @maksimzayats in #11057 - Send
FileSearchTool.max_num_resultsto OpenAI Responses by @adtyavrdhn in #12728 - Raise
UserErrorwhen validation requests approval for an external tool, keep single-model tool arguments through durable re-validation, and document concurrent DBOS runs and shared MCPhttp_clients by @dsfaccini in #12958 - Record a
StepPersistencetool call that ends inModelRetryorToolFailedasfailedinstead of leaving itstartedby @dsfaccini in #12638 - Raise
UserErrorfromStreamedRunResult.resultwhen arun_stream()run stops while finishing by @dsfaccini in #13087 - Run
CodeModeremote sandboxes over gRPC with pydantic-monty 1.5.0 by @dmontagu in #12673 - Replace
httpxwithhttpx2inpydantic-ai-harnessby @Kludex in #13026 - Curate
KnownModelNameand the provider model-name literals to the models a user would pick by @dsfaccini in #12817
🚀 Features
- Add
dequeueandset_enqueue_priorityto change a message queued withenqueuebefore delivery by @adtyavrdhn in #12345 - Support
SubAgents(dynamic_agents=...)under durable execution by @mpfaffenberger in #12418 - Run tool calls as DBOS child workflows via
child_workflowtool metadata inDBOSDurabilityby @mpfaffenberger in #12445 - Checkpoint
TrajectoryJudgeevaluations as a durable operation when the judge has anidby @mpfaffenberger in #12414 - Identify Azure Durable Task workers with
pydantic-aiin their gRPC user agent by @tamirdresher in #12493 - Add
clienttoModalSandboxandModalSandboxBackendby @adtyavrdhn in #12501 - Strip every default-ignorable character before
ArtifactScannerphrase matching and flag hidden variation selectors by @mpfaffenberger in #11475 - Add
clienttoE2BSandboxandE2BSandboxBackendby @adtyavrdhn in #12500 - Add per-path workspace authority with
Workspace.authority_ofby @DouweM in #12473 - Add
openai_include_unmapped_response_datato keep Chat Completions response data Pydantic AI doesn't map by @DouweM in #12034 - Load one instruction file per directory and nested files by default in
RepoContext, addwalk_up, and haveCoderbuild on them by @DouweM in #11691 - Cap inlined nested instructions at 32,000 characters per conversation by default with
RepoContext(nested_max_chars=...)by @DouweM in #11699 - Support OpenAI
fastandultrafastservice tiers inopenai_service_tierby @adtyavrdhn in #12527 - Send documents of any media type
mimetypesknows, with the format taken from it by @arjun-krishna1 in #12365 - Run DBOS child-workflow tools in parallel when a workspace is attached by @mpfaffenberger in #12522
- Add the
KeenableSearchcapability for keyless web research by @ilya-bogin-keenable in #9079 - Add
gateway/bedrock:global.moonshotai.kimi-k3andgateway/bedrock:global.moonshotai.kimi-k3.1toKnownModelNameby @dsfaccini in #12869 - Add provider-specific settings to native tools by @adtyavrdhn in #7633
- Support Anthropic
web_search_20260318andweb_fetch_20260318viaprovider_settingsby @adtyavrdhn in #12610 - Move provider-specific native-tool fields to
provider_settingsby @adtyavrdhn in #7807 - Support
SubAgents(task_control=...)under Temporal and DBOS durability by @mpfaffenberger in #12521 - Support OpenAI's
gpt-image-2.6-flareandgpt-image-2.6-sunburstimage models by @dsfaccini in #12820 - Add Claude Fable 5.2 (
claude-fable-5-2) and Claude Mythos 5.2 (claude-mythos-5-2) support by @dsfaccini in #12970 - List Claude Fable 5.2 on Bedrock (
global.anthropic.claude-fable-5-2) and send per-message effort to it on OpenRouter by @dsfaccini in #13037 - Support Gemini 4 Carbon (
gemini-4-carbon) onGoogleModelby @dsfaccini in #13057 - Add
framing_announcementsetting and only announce the<system>voice tag once a conversation needs it by @DouweM in #12661 - Surface nested
RepoContextinstruction files when the model writes or edits in a directory by @DouweM in #11701
🐛 Bug Fixes
- Record
InputGuardrail,OutputGuardrailandToolGuardrailverdicts as durable operations by @mpfaffenberger in #12428 - Stop commands under
crun1.14.1 inDockerSandboxand buildAsyncDaytonaoff the event loop by @adtyavrdhn in #12509 - Record
Advisor,SmartSearchandConversationSearchtool I/O as durable operations by @mpfaffenberger in #12415 - Include rejected
FallbackModelattempts in the usageBrowserUsereports to browser-use by @mpfaffenberger in #11879 - Publish
BoxdSandboxfields in the agent spec schema by @adtyavrdhn in #12518 - Publish
DaytonaSandboxfields in the agent spec schema by @adtyavrdhn in #12517 - Publish
RenderSandboxfields in the agent spec schema by @adtyavrdhn in #12520 - Publish
SpritesSandboxfields in the agent spec schema by @adtyavrdhn in #12519 - State a loaded capability's instructions where the load is when a UI client sends its history back by @DouweM in #11775
- Send Gemini tool returns that contain a
$refkey as JSON text, which Gemini otherwise rejects by @DouweM in #12429 - Make Temporal wait a model error's
retry_afterbefore retrying the model request activity by @DouweM in #12106 - Raise right away when the token limit cuts off
NativeOutputorPromptedOutputtext that fails validation, instead of retrying by @DouweM in https://github.com/pydantic/p...
v2.66.0 (2026-11-04)
What's Changed
🛡️ Security
- GHSA-m5vj-8mcw-82qx (high):
web_fetch_tooland theWebFetchcapability's local fallback re-indented a JSON response on the event loop before applying the returned-content limit, so a deeply nested document within the download size limit could expand to hundreds of times its size, using large amounts of memory and delaying other work in the same process. You are affected if an attacker can choose or influence the URLs the tool fetches. Provider-native web fetching is not affected. Fixed in v2.66.0 and v1.107.10 (pydantic-ai,pydantic-ai-slim). - GHSA-mfx5-pc32-32m8 (high): The
private_keypattern in Pydantic AI Harness's secret redaction backtracked exponentially on crafted text, so a few hundred characters that begin like a private key could occupy the event loop for minutes or longer. You are affected if you useredact_secretsorsecret_data()with theprivate_keypattern, which is included by default, on text an attacker can influence. Fixed inpydantic-ai-harness0.66.0. - GHSA-5fj9-5pjw-vpwc (medium): The cloud-metadata blocklist that URL downloads check did not include Tencent Cloud's instance metadata address
169.254.0.23, so someone who can supply a URL the application downloads could have it fetch the instance's temporary role credentials. You are affected if you run on Tencent Cloud and allow local URLs, throughforce_download='allow-local'on aFileUrlorweb_fetch_tool(allow_local_urls=True). Fixed in v2.66.0 and v1.107.10 (pydantic-ai,pydantic-ai-slim). - GHSA-jp94-mmr3-2pvv (medium): With OpenTelemetry instrumentation enabled for embedding models, embedding spans recorded the
extra_headersandextra_bodyembedding settings, so credentials passed through them could be read by anyone with access to the exported traces. You are affected if you instrument embedding models and pass secrets through those settings. Fixed in v2.66.0 and v1.107.10 (pydantic-ai,pydantic-ai-slim). Restrict access to historical traces and rotate credentials they exposed. - GHSA-7494-925c-rrxv (medium): For models whose profile requires tool schemas without
$defsand$ref, a small tool schema whose definitions reference one another repeatedly expanded exponentially when inlined on the event loop, which can take minutes of CPU time and gigabytes of memory. You are affected if you serve such a model through the AG-UI adapter, where clients submit their own frontend tool schemas. Fixed in v2.66.0 and v1.107.10 (pydantic-ai,pydantic-ai-slim). - GHSA-484x-x9m4-62cj (medium): Decision models (
DecisionModelsubclasses such asOpenAIDecisionsModel, andTypeSafeModel) resolved every$refin a tool's argument schema again at each place it appears, so a small crafted schema could take exponential time and memory on the event loop. You are affected if you use a decision model from v2.46.0 onward with tools whose schemas an attacker can supply, most directly through an AG-UI endpoint. Fixed in v2.66.0 (pydantic-ai,pydantic-ai-slim). - GHSA-r879-rxgj-3g75 (medium): The Vercel AI and AG-UI adapters took time that grows with the square of the number of parts to load a submitted message, on the event loop, so a single request could delay other work in the same process. You are affected if you pass client-submitted conversations to
VercelAIAdapterorAGUIAdapter. Fixed in v2.66.0 and v1.107.10 (pydantic-ai,pydantic-ai-slim). - GHSA-4cj4-hgwf-4hc4 (medium): When message history was serialized and loaded again, a tool result containing a JSON object in the serialized shape of one of Pydantic AI's file types was rebuilt as that file. A third party who controls data a tool returns could use this to make the application download a URL on its local network, or make the model provider read an
s3://object or uploaded file with the application's credentials. You are affected if tool results can contain data third parties control, that history is stored and reloaded, held on the server by a UI adapter, or carried by Temporal, DBOS or Prefect durable execution, and your model or its provider fetches or resolves such files; the advisory lists which do. Fixed in v2.66.0 and v1.107.10 (pydantic-ai,pydantic-ai-slim). - GHSA-5ppx-9jq6-6pmw (low):
allowed_domainson the local web-fetch tool, andPlaywrightBrowseregress policies whose allowlist bounds every request kind, looked a refused hostname up in DNS before rejecting it, so data encoded in the hostname could reach that domain's nameserver. Domain lists are now checked before resolution. You are affected only if you rely on such an allowlist while the model processes content untrusted parties can influence. Fixed in v2.66.0 and v1.107.10 (pydantic-ai,pydantic-ai-slim), andpydantic-ai-harness0.66.0. - GHSA-mgmv-r532-pg36 (medium): An
MCPcapability configured with a server URL and no explicitidderived its identifier from the URL's hostname and last path segment, and the AG-UI and Vercel AI adapters sent that identifier to the frontend with MCP Apps tool results, so a credential in that segment could reach users of the chat frontend. This affects v2.57.0 up to but not including v2.60.0, and was fixed in v2.60.0; it is disclosed with this batch. Rotate MCP server credentials that were carried in the URL path.
⚠️ Compatibility Notes
- Call Anthropic's
20260209web tools directly on the Anthropic API when citations are requested by @dsfaccini in #12278 - Apply
FunctionToolset(requires_approval=True)toToolinstances passed to the toolset by @dsfaccini in #12378 - Re-resolve realtime session instructions before each model turn and deliver their changes as appends by @DouweM in #11930
- Recognize
SummarizingCompaction(keep_user_messages=True)copies only by the marker on their text by @DouweM in #12316
🚀 Features
- Add
OpenSandboxfor self-hosted agent workspaces by @adtyavrdhn in #11841 - Add Azure Durable Task durability to
pydantic_ai_harnessby @DouweM in #11332 - Add
Pixeltabletools and memory storage to Harness by @pierrebrunelle in #9369 - Add
RenderSandboxto run agent workspaces in Render sandboxes by @ojusave in #10238 - Add
include_instructionsto theMCPcapability so server instructions can be left out of the agent's instructions by @DouweM in #12269 - Re-resolve realtime session instructions before each model turn and deliver their changes as appends by @DouweM in #11930
- Report xAI Grok Voice speech timing as
segmenttimestamps withoutput_speech_timestampsby @DouweM in #11782 - Add a
retriesfield toSummarizingCompactionby @gouyu-hou in #12315 - Add
Workspace.authorityandAuthorityWorkspacefor workspaces whose files the operator does not control by @adtyavrdhn in #11836
🐛 Bug Fixes
- Stop
ModalSandboxfailing withAttributeErroron Modal 1.6, which removedFilesystemExecutionErrorby @adtyavrdhn in #12247 - Call Anthropic's
20260209web tools directly on the Anthropic API when citations are requested by @dsfaccini in #12278 - Add streaming support to
CohereModelby @Julien-ser in #6100 - Release run dependencies after exceptions by @dsfaccini in #11255
- Preserve extra keys in
tool_searchandload_capabilitymessage history by @pydanty in #7933 - Inline
$defsin tool schemas for GLM-5.3 on Together, Nebius, Fireworks, Vercel, Ollama, Hugging Face, Cerebras, Mistral and OpenAI-compatible gateways by @pydanty in #12138 - Keep
RepoContextfrom following symlinks out of the working directory or to files that aren't instruction files by @DouweM in #12044 - Reject approvals for tool calls a server-held history isn't waiting on under
manage_message_history='server'by @DouweM in #12313 - Apply
FunctionToolset(requires_approval=True)toToolinstances passed to the toolset by @dsfaccini in #12378 - Recognize
SummarizingCompaction(keep_user_messages=True)copies only by the marker on their text by @DouweM in https://github.com/pydantic/pydant...
v1.107.10 (2026-11-04)
What's Changed
🛡️ Security
- GHSA-m5vj-8mcw-82qx (high):
web_fetch_tooland theWebFetchcapability's local fallback re-indented a JSON response on the event loop before applying the returned-content limit, so a deeply nested document within the download size limit could expand to hundreds of times its size, using large amounts of memory and delaying other work in the same process. You are affected if an attacker can choose or influence the URLs the tool fetches. Provider-native web fetching is not affected. Fixed in v1.107.10 and v2.66.0 (pydantic-ai,pydantic-ai-slim). - GHSA-5fj9-5pjw-vpwc (medium): The cloud-metadata blocklist that URL downloads check did not include Tencent Cloud's instance metadata address
169.254.0.23, so someone who can supply a URL the application downloads could have it fetch the instance's temporary role credentials. You are affected if you run on Tencent Cloud and allow local URLs, throughforce_download='allow-local'on aFileUrlorweb_fetch_tool(allow_local_urls=True). Fixed in v1.107.10 and v2.66.0 (pydantic-ai,pydantic-ai-slim). - GHSA-jp94-mmr3-2pvv (medium): With OpenTelemetry instrumentation enabled for embedding models, embedding spans recorded the
extra_headersandextra_bodyembedding settings, so credentials passed through them could be read by anyone with access to the exported traces. You are affected if you instrument embedding models and pass secrets through those settings. Fixed in v1.107.10 and v2.66.0 (pydantic-ai,pydantic-ai-slim). Restrict access to historical traces and rotate credentials they exposed. - GHSA-7494-925c-rrxv (medium): For models whose profile requires tool schemas without
$defsand$ref, a small tool schema whose definitions reference one another repeatedly expanded exponentially when inlined on the event loop, which can take minutes of CPU time and gigabytes of memory. You are affected if you serve such a model through the AG-UI adapter, where clients submit their own frontend tool schemas. Fixed in v1.107.10 and v2.66.0 (pydantic-ai,pydantic-ai-slim). - GHSA-r879-rxgj-3g75 (medium): The Vercel AI and AG-UI adapters took time that grows with the square of the number of parts to load a submitted message, on the event loop, so a single request could delay other work in the same process. You are affected if you pass client-submitted conversations to
VercelAIAdapterorAGUIAdapter. Fixed in v1.107.10 and v2.66.0 (pydantic-ai,pydantic-ai-slim). - GHSA-4cj4-hgwf-4hc4 (medium): When message history was serialized and loaded again, a tool result containing a JSON object in the serialized shape of one of Pydantic AI's file types was rebuilt as that file. A third party who controls data a tool returns could use this to make the application download a URL on its local network, or make the model provider read an
s3://object or uploaded file with the application's credentials. You are affected if tool results can contain data third parties control, that history is stored and reloaded or carried by Temporal, DBOS or Prefect durable execution, and your model or its provider fetches or resolves such files; the advisory lists which do. Fixed in v1.107.10 and v2.66.0 (pydantic-ai,pydantic-ai-slim). - GHSA-5ppx-9jq6-6pmw (low):
allowed_domainson the local web-fetch tool looked a refused hostname up in DNS before rejecting it, so data encoded in the hostname could reach that domain's nameserver. Domain lists are now checked before resolution. You are affected only if you rely on such an allowlist while the model processes content untrusted parties can influence. Fixed in v1.107.10 and v2.66.0 (pydantic-ai,pydantic-ai-slim).
Full Changelog: v1.107.9...v1.107.10
v2.65.0 (2026-11-02)
What's Changed
⚠️ Compatibility Notes
- Protect the
gh-awcanonical session capture from agent tools by @dsfaccini in #10606 - Add multimodal notes to
AskUserAnswerby @dsfaccini in #11664 - Count
BrowserUsemodel calls toward the host run's usage andUsageLimitsby @mpfaffenberger in #11860 - Name the V2 replacement when code imports a name V2 removed, like
OpenAIModelorMCPServerStdioby @DouweM in #12035 - Fall back to the next model in
ModelFallbackwhen a model refuses, and letContentFilterErrorinfallback_onmatch refusals by @DouweM in #12039 - Raise
ModelQuotaExceededErrorinstead ofModelRateLimitErrorfor Z.ai balance and plan-limit 429s by @DouweM in #12042 - Support the
'12h'prompt cache retention on the OpenAI Responses API for GPT-5.6 and later by @DouweM in #12046 - Send Google a server deadline of at least 20 minutes instead of the client's read timeout, and warn about httpx's 5 s default by @DouweM in #12057
- Stop Temporal from retrying model errors that can't succeed on retry by default, via
non_retryable_model_errorsby @DouweM in #12107
🚀 Features
- Add WebSocket transport to OpenAI Responses by @dsfaccini in #10444
- Add
ConversationStore, which keeps a conversation as a log of messages stored once, andConversation.revisionto catch conflicting saves by @DouweM in #10658 - Add multimodal notes to
AskUserAnswerby @dsfaccini in #11664 - Add
anthropic_web_tools_allowed_callersto setallowed_callerson Anthropic's web tools by @aweis89 in #10489 - Support the
'12h'prompt cache retention on the OpenAI Responses API for GPT-5.6 and later by @DouweM in #12046 - Send mid-conversation system prompts natively on OpenRouter OpenAI routes and the Claude models that support them by @DouweM in #12056
- Name the V2 replacement when code imports a name V2 removed, like
OpenAIModelorMCPServerStdioby @DouweM in #12035 - Fall back to the next model in
ModelFallbackwhen a model refuses, and letContentFilterErrorinfallback_onmatch refusals by @DouweM in #12039
🐛 Bug Fixes
- Raise
ModelContextWindowExceededErrorandModelQuotaExceededErrorfor Bedrock errors onAsyncAnthropicBedrockby @DouweM in #12040 - Raise
ContentFilterErrorinstead ofTypeErrorwhen Azure filters a streamedOpenAIChatModelprompt by @DouweM in #12041 - Re-surface nested
RepoContextinstruction files after compaction, and stop repeating notes still in history by @DouweM in #11704 - Include the refusal text in
ContentFilterErrorwhen the provider also reports a finish reason by @DouweM in #12036 - Raise
ModelQuotaExceededErrorinstead ofModelRateLimitErrorfor Z.ai balance and plan-limit 429s by @DouweM in #12042 - Protect the
gh-awcanonical session capture from agent tools by @dsfaccini in #10606 - Close
PlaywrightBrowser's Chromium and driver when ananyioscope cancels a run by @dsfaccini in #12047 - Ask
Coderto match the surrounding code's style instead of applying SOLID by @mpfaffenberger in #12120 - Return a
Codertool's error to the model once its retries run out instead of ending the run by @mpfaffenberger in #12121 - Keep
*argsand**kwargsdocstring descriptions on tool parameters by @Aryapatel-007 in #12126 - Read tool parameter descriptions from
Keyword ArgsandOther Parametersdocstring sections by @roydonsequeira in #10035 - Let a retired
aws_lambdaagent loop finish its own cleanup at the grace deadline by @dsfaccini in #12140 - Treat
auth=''as unset alongsideclientinGitHub,Notion,Ordinal,PostHog,Pylon, andSlackby @mpfaffenberger in #11852 - Count
BrowserUsemodel calls toward the host run's usage andUsageLimitsby @mpfaffenberger in #11860 - Let
FileSystemtools emit their events underPrefixToolsby @mpfaffenberger in #11867 - Forward
cache='1h'to Claude Haiku 5.5 on Bedrock instead of snapping it to'5m'by @DouweM in #12045 - Mark
<context>elements in user and tool content as not from the application by @DouweM in #12050 - Stop Temporal from retrying subclasses of non-retryable errors, like
ContentFilterErroror aUserErrorsubclass by @DouweM in #12105 - Stop Temporal from retrying model errors that can't succeed on retry by default, via
non_retryable_model_errorsby @DouweM in #12107 - Stop sending a history container on Anthropic requests without code execution by @dsfaccini in #11764
- Send Google a server deadline of at least 20 minutes instead of the client's read timeout, and warn about httpx's 5 s default by @DouweM in #12057
- Place the OpenRouter Gemini message cache breakpoint before the final message, where it no longer doubles the billed prompt by @DouweM in #12048
📦 Dependencies
- Remove the Daytona
close()workaround and requiredaytona>=0.226.1by @mislavivanda in #11977
New Contributors
- @Aryapatel-007 made their first contribution in #12126
- @roydonsequeira made their first contribution in #10035
- @mislavivanda made their first contribution in #11977
Full Changelog: v2.64.0...v2.65.0
v2.64.0 (2026-10-30)
What's Changed
⚠️ Compatibility Notes
- Raise an error status a retrying transport gave up on as
ModelHTTPError, so fallback moves on by @DouweM in #11934 - Send stored response, item, conversation and container IDs back only to the endpoint that created them, via
Model.accepts_server_stateby @DouweM in #11774 - Make
json-repairan optional dependency, needed only byCoderandRepairToolArgumentsby @mpfaffenberger in #11768 - Turn off strict tool schemas for
Coder's tools by default, withCoder(strict=True)to opt in by @mpfaffenberger in #11743 - Surface every nested
RepoContextinstruction file between a traversed path and the working directory, without repeating the root file by @DouweM in #11687 - Add
output_speech_timestampsto report word timing of realtime speech, on Azure AI Voice Live by @DouweM in #11685 - Make the
ModelFallbackcapability constructible from agent specs by @DouweM in #11679 - Add a
first_event_timeoutmodel setting so a stream that never starts can fall back by @DouweM in #11582 - Dispatch
OnlineEvaluationevaluators once per durable run instead of on every replay by @DouweM in #11577 - Show the harness
Planningplan in the plan tools' results instead of a per-request reminder copy by @DouweM in #11521 - Add agent artifact security scanning for
MemoryandSkillswithArtifactScannerby @mpfaffenberger in #11470 - Fix
AnthropicModel.count_tokensrejecting all-deferred tool requests and undercounting server tools on the Anthropic API by @dsfaccini in #11343 - Accept plain functions as evaluators in
pydantic_evalsby @dmontagu in #10292 - Remove unresolvable
RealtimeEventfrompydantic_ai.agent.__all__by @adtyavrdhn in #10189 - Add
RunContext.load_capabilityto load on-demand capabilities from code by @adtyavrdhn in #10090 - Add a
ModelFallbackcapability, a per-attemptprepare_model_requesthook, andRetryModelRequestby @DouweM in #9348
🚀 Features
- Let
@agent.system_prompt(dynamic=True)append its changes withon_change='append'to keep the cached prompt by @DouweM in #11445 - Pair
DetectLoopstool results with the response that made the calls by @mpfaffenberger in #11551 - Add
BoxdSandbox: a workspace that runsCoder,Shell,FileSystemand every other harness capability in a boxd machine by @MichielMAnalytics in #11367 - Add a
ModelFallbackcapability, a per-attemptprepare_model_requesthook, andRetryModelRequestby @DouweM in #9348 - Add Harness
Reminder(scope='conversation')for reminders that stay in the message history by @DouweM in #11642 - Add
output_speech_timestampsto report word timing of realtime speech, on Azure AI Voice Live by @DouweM in #11685 - Cache the schema of
ModelRequestContextdurable capability operations across agent construction by @DouweM in #11776 - Make the
ModelFallbackcapability constructible from agent specs by @DouweM in #11679 - Report word timestamps of ElevenLabs Agents speech with
output_speech_timestampsby @DouweM in #11784 - Record CI provenance on Pydantic Evals experiment spans with OpenTelemetry
vcs.*andcicd.*attributes by @dmontagu in #10294 - Accept plain functions as evaluators in
pydantic_evalsby @dmontagu in #10292 - Add
logfire.scorer.score_typeandlogfire.scorer.run_idto online evaluation result events by @dmontagu in #10293 - Add the OpenTelemetry
test.suite.nameandtest.case.nameattributes to Pydantic Evals spans by @dmontagu in #11538 - Turn off strict tool schemas for
Coder's tools by default, withCoder(strict=True)to opt in by @mpfaffenberger in #11743 - Make
json-repairan optional dependency, needed only byCoderandRepairToolArgumentsby @mpfaffenberger in #11768 - Teach
Coderto read what the task names, check stated requirements, and delegate only when asked by @mpfaffenberger in #11811 - Let
AskUseranswerers receiveRunContextby @dsfaccini in #11662 - Add agent artifact security scanning for
MemoryandSkillswithArtifactScannerby @mpfaffenberger in #11470 - Add
ModelResponse.model_idrecording the model the agent sent each request to by @DouweM in #11807 - Save function evaluators to
Datasetfiles under the function's name by @dmontagu in #10425 - State how nested instruction files relate in
RepoContextinstructions whennested_traversalis on by @DouweM in #11698 - Add a
first_event_timeoutmodel setting so a stream that never starts can fall back by @DouweM in #11582 - Add
stickytoModelFallbackto keep a conversation on the fallback model that answered by @DouweM in #11330 - Add
RunContext.load_capabilityto load on-demand capabilities from code by @adtyavrdhn in #10090
🐛 Bug Fixes
- Preserve nested
SpanTreeattributes during JSON round-trips and mark omitted Google embedding and ElevenLabs usage as unmeasured to avoid false zero automatic costs by @dsfaccini in #11912 - Fall back from Native Output on Bedrock when the installed
botocorepredatesoutputConfigby @DouweM in #11694 - Close the
IPythonKernelclient's zmq context on shutdown with jupyter-client before 8.7 by @DouweM in #11716 - Deliver queued realtime messages after a mid-reply reconnect on OpenAI, Azure and xAI by @DouweM in #11678
- Keep the Gemini Live event stream alive when a resumed session lost a tool call without an id by @DouweM in #11680
- Search a mixed
CodeModeTool Search corpus withsearch_tools, so a native deferred tool no longer declares the sandboxed ones by @DouweM in #11638 - Release a closed realtime session's run state by reference counting by @TejasS1233 in #11240
- Show the harness
Planningplan in the plan tools' results instead of a per-request reminder copy by @DouweM in #11521 - Dispatch
OnlineEvaluationevaluators once per durable run instead of on every replay by @DouweM in #11577 - Fix
AnthropicModel.count_tokensrejecting all-deferred tool requests and undercounting server tools on the Anthropic API by @dsfaccini in #11343 - Stop advertising
CodeExecutionToolon the BedrockAnthropicModelclients by @DouweM in #11790 - Turn on the unified
cachesetting for GPT-5.6 and later with the Azure OpenAI v1 API by @DouweM in #11749 - Keep the usage and response ID a primed stream received when the consumer leaves it without iterating by @DouweM in #11783
- Keep
Shell's job files in a private temp directory when the workspace isn't writable by @mpfaffenberger in #11809 - Inline
$refs underpropertyNamesandcontainsfor OpenRouter Google models and customprefer_inlined_defstransformers by @jukebox in #10859 - Strip regex patterns Bedrock rejects from strict JSON schemas in
BedrockJsonSchemaTransformerby @adtyavrdhn in #11831 - Remove unresolvable
RealtimeEventfrompydantic_ai.agent.__all__by @adtyavrdhn in #10189 - Keep finished
BackgroundToolsresults in the history when a run pauses for deferred tools by @wangrat in #11752 - Let
FileSystemedit_fileedit non-UTF-8 text files thatread_filereads, honoringexpected_hashby @mpfaffenberger in #11851 - Make
SubAgentschildren that share usage run under the parent'susage_limitsby @mpfaffenberger in #11853 - Publish
AgentSpecschema fields for `Adviso...
v2.63.0 (2026-10-27)
What's Changed
⚠️ Compatibility Notes
- Default
CodeModetodynamic_catalog=True, so a revealed tool is appended instead of rewriting the cached prompt by @DouweM in #11358 - Pass the Gemini Live reconnect replay to
GoogleRealtimeConnection'sdialas a keyword-onlyreplayargument by @DouweM in #11515 - Give Harness
LLMRemindera defaultinterval=10instead of generating on every request by @DouweM in #11641 - Replay signed and encrypted history data to providers with the same
canonical_systemby @DouweM in #11504 - Let a deferred
CombinedCapabilityload as one capability, with members keeping their global ordering by @adtyavrdhn in #10321 - End a GPT-Live reply that delegated at Live's next pause instead of mid-sentence when the backend answers by @DouweM in #11250
- Say what the
<system>tag means and escape it in user and tool text on models without mid-conversation system messages by @DouweM in #11506
🚀 Features
- Add
task_timeouttoDataset.evaluateto bound each task attempt by @dmontagu in #10291 - Support
ag-ui-protocol1.0 by @adtyavrdhn in #8627 - Default
CodeModetodynamic_catalog=True, so a revealed tool is appended instead of rewriting the cached prompt by @DouweM in #11358 - Pass the Gemini Live reconnect replay to
GoogleRealtimeConnection'sdialas a keyword-onlyreplayargument by @DouweM in #11515 - Add OpenAI
gpt-6.1-terramodel support by @dsfaccini in #11393 - Add an IPython Code Mode (
IPythonCodeMode) alongside the Monty-basedCodeModeby @mpfaffenberger in #11570 - Give Harness
LLMRemindera defaultinterval=10instead of generating on every request by @DouweM in #11641 - Add a
DetectLoopscapability to interrupt repeated-action loops by @mpfaffenberger in #11471 - Add
manage_message_historyto UI adapters so a server-held history isn't sent to the model twice by @DouweM in #11431 - Keep strict schemas within Anthropic's and OpenAI's complexity limits, falling back to non-strict output where no mode was chosen by @DouweM in #11464
- Add a public
pydantic_ai.framingmodule withwrap,neutralize_tagsandregister_tag, so extensions can frame their own tags by @DouweM in #10957 - Replay signed and encrypted history data to providers with the same
canonical_systemby @DouweM in #11504 - Add
PostgresSpendStoreso harnessSpendLimitsbudgets can be shared through PostgreSQL by @DouweM in #11527
🐛 Bug Fixes
- Let a deferred
CombinedCapabilityload as one capability, with members keeping their global ordering by @adtyavrdhn in #10321 - List the modules Monty 1.2-1.4 added in the
CodeModerun_codedescription by @mpfaffenberger in #11459 - Require
mistralai>=3.1.0and drop the legacyhttpximport soMistralModelimports on fresh installs by @adtyavrdhn in #11549 - Keep the rate-limit category of Azure
too_many_requestsResponses failures onopenai>=3.32.0by @dsfaccini in #11532 - Run
ToolOutputLimitsspill writes in adurable_operationso replays do not write them again by @mpfaffenberger in #11460 - Record the instructions a
before_model_requesthook sent when it rewrote the history with a copy of the request by @DouweM in #11501 - State a deferred capability's instructions on the request where a
before_model_requesthook loaded it by @DouweM in #11516 - Escape framing tags consistently in
BackgroundTools,RepoContext,TrajectoryJudgeandPlanningframes by @DouweM in #11530 - Record a
thinkingchange after an unansweredModelSettingsDeltaPartinstead of replacing it, so UI streams and hook-rebuilt requests keep it by @DouweM in #11514 - Make typed turns on Gemini
gemini-robotics-errealtime models see the image sent just before them by @DouweM in #11484 - End a GPT-Live WebRTC sideband session cleanly when the browser hangs up by @DouweM in #11435
- End a GPT-Live reply that delegated at Live's next pause instead of mid-sentence when the backend answers by @DouweM in #11250
- Withdraw a pending batch request when the run doesn't end with its
PendingModelResponseby @DouweM in #11517 - Send per-message effort changes through
AsyncAnthropicBedrockMantleon Claude Opus 5.5 by @DouweM in #11486 - Enable thinking-block binding recovery on
AsyncAnthropicBedrockMantleby @DouweM in #11487 - Stop offering
AdvisorToolonAsyncAnthropicBedrockMantle, which rejects it by @DouweM in #11488 - Keep the stored
CodeModeREPL state whenrepl_storeload fails, instead of saving a fresh REPL over it by @mpfaffenberger in #11483 - Fix
RuntimeError: coroutine ignored GeneratorExitand an escapingKeyboardInterruptonAgent.iter()teardown by @pydanty in #10258 - Return a
wrap_runoron_run_errorrecovery result fromrun_stream()instead of raisingAgentRunErrorby @DouweM in #11571 - Say what the
<system>tag means and escape it in user and tool text on models without mid-conversation system messages by @DouweM in #11506 - Escape framing tags consistently in the
SummarizingCompactionsummarizer transcript by @DouweM in #11583 - Enforce
allowed_patternson the real path whensearch_filestargets a single symlinked file by @mpfaffenberger in #11639 - Normalize bare
Nonetype arguments inDatasetby @pydanty in #11254 - Stop offering
WebSearchToolandWebFetchToolonAsyncAnthropicBedrockMantle, which rejects them by @DouweM in #11489 - Don't commit an automatic
SummarizingCompactionsummary that would make the history larger by @DouweM in #11649 - Clear the caption in
stream_transcripts(delta=True)when a transcript is corrected to empty by @DouweM in #11667 - Evaluate runs with every
OnlineEvaluationcapability on an agent, not just the outermost by @DouweM in #11663 - Omit realtime span
model_settingswheninclude_content=Falseby @DouweM in #11668 - Call
AsyncAnthropicBedrockMantlethe Bedrock Messages API, not Claude Platform on AWS, in comments and docstrings by @DouweM in #11482 - Record the provider item ID on the model's realtime
SpeechPart.idby @DouweM in #11675
Full Changelog: v2.62.0...v2.63.0
v2.62.0 (2026-10-26)
What's Changed
⚠️ Compatibility Notes
- Change thinking effort mid-conversation via OpenAI
configuration_updateand Anthropic per-messageoutput_config, recorded asModelSettingsDeltaPartby @DouweM in #11111 - Add
ModelRequestRejectedError,ModelAuthenticationErrorandModelNotFoundErrormodel error categories by @DouweM in #11081 - State loaded deferred capabilities' instructions in the system voice, append their changes, and show them to
before_model_requestby @DouweM in #11103 - Prefix hosted MCP capabilities' tool names with the service by default, configurable with
tool_prefixby @DouweM in #10376 - Store the
SummarizingCompactionsummary as aCompactionPartinstead of a system prompt, and send provider-lessCompactionParts to every model by @DouweM in #10573 - Deliver realtime
SystemPromptParts on each provider's operator channel withsession.send()by @DouweM in #11379 - Preserve run identity when replaying Prefect capability operations by @dsfaccini in #11196
- Append changed MCP server instructions with
on_change='append', and give an id-less URLMCPToolsetits hostname asidby @DouweM in #10791 - Record Harness
Memorysnapshots in history once per changed run, so requests stay append-only by @DouweM in #10903 - Tell the realtime model a tool call failed on every provider, so it answers instead of going silent by @DouweM in #11086
- Reserve DBOS handler step positions during tool
ctx.emit()by @dsfaccini in #11157 - Record GPT-Live speech as one
SpeechPartinstead of splitting it at each delegated tool call by @DouweM in #11187 - Deliver
ctx.emit()events into the realtime session event stream by @DouweM in #11369 - Make realtime shared settings warn, raise or ignore like standard
ModelSettings: forcedtool_choiceraises on Gemini Live,parallel_tool_calls=Falsewarns by @DouweM in #11066 - Keep the prompt cache when changing thinking effort mid-conversation on OpenRouter, via its
configuration_updatesystem message by @DouweM in #11426 - Start
GraphBuilderbroadcast branches in.to()order instead of reversed, which can break replay of durable workflows paused across the upgrade by @DouweM in #11408 - Cache
on_change='append'instruction blocks behind a second cache point after the static instructions by @DouweM in #11351 - Build Harness
SystemReminderson request-scoped system prompts and deprecatecache_ttlby @DouweM in #9712 - Add
CodeMode(monty_sandbox=...), which takes a remote Monty URL or a function that builds the pool, and deprecatemonty_sandbox_urlby @adtyavrdhn in #11141
🚀 Features
- Change thinking effort mid-conversation via OpenAI
configuration_updateand Anthropic per-messageoutput_config, recorded asModelSettingsDeltaPartby @DouweM in #11111 - Add
ModelRequestRejectedError,ModelAuthenticationErrorandModelNotFoundErrormodel error categories by @DouweM in #11081 - Export
conversation_id_from_run_idfrompydantic_ai.durable_execfor third-party durability engines by @DouweM in #11338 - Run tool calls as Temporal child workflows via
child_workflowtool metadata inTemporalDurabilityby @YHallouard in #7329 - Restrict
tool_choicesubsets withallowed_toolson OpenAI Chat Completions to preserve the prompt cache by @DouweM in #10396 - State loaded deferred capabilities' instructions in the system voice, append their changes, and show them to
before_model_requestby @DouweM in #11103 - Prefix hosted MCP capabilities' tool names with the service by default, configurable with
tool_prefixby @DouweM in #10376 - Store the
SummarizingCompactionsummary as aCompactionPartinstead of a system prompt, and send provider-lessCompactionParts to every model by @DouweM in #10573 - Add
tool_prefixto theMCPcapability by @DouweM in #11364 - Let
SubAgentsrun delegations as tasks the model can background, steer, answer and cancel, behindtask_controlby @mpfaffenberger in #10893 - Add
Coder(task_control=)to turn onSubAgentstask control by @mpfaffenberger in #11007 - Add
StructuredContentuser content, sent as its value to decision models by @DouweM in #10529 - Use the Pydantic AI Gateway's reported cost for
RequestUsage.costby @adtyavrdhn in #7337 - Let the model request a harness compaction at a natural break in the task with
CompactOnRequestby @mpfaffenberger in #11422 - Deliver realtime
SystemPromptParts on each provider's operator channel withsession.send()by @DouweM in #11379 - Add
AnthropicModel.compact_messages()for on-demand native compaction by @DouweM in #11432 - Add provider batch API support with a
BatchModelRequestscapability and aget_wrapper_modelcapability hook by @DouweM in #10672
🐛 Bug Fixes
- Break
GraphRuniterator andToolManagercontext reference cycles by @dsfaccini in #11158 - Preserve run identity when replaying Prefect capability operations by @dsfaccini in #11196
- Stop
Agent.run_streamname inference from keepingdepsalive until garbage collection by @dsfaccini in #11219 - Preserve Cohere response IDs in
ModelResponse.provider_response_idby @dsfaccini in #11223 - Recognize
deepseek-flashinDeepSeekProviderby @dsfaccini in #11249 - Close the whole
AgentStreamevent stream when a consumer leavesnode.stream()early by @adtyavrdhn in #11009 - Close inner event streams so a raising listener cancels in-flight tool calls by @Kludex in #8822
- Fix
thinking=Falseon OpenRouter'santhropic/claude-sonnet-5.5, which can't turn reasoning off by @dsfaccini in #11261 - Append changed MCP server instructions with
on_change='append', and give an id-less URLMCPToolsetits hostname asidby @DouweM in #10791 - Keep a system prompt after the tool result it followed in AG-UI
dump_messages()by @DouweM in #11188 - Cancel a suspended background job only once when a streamed continuation request fails by @DouweM in #11090
- Keep billed usage when a run stops before a continued response completes by @DouweM in #11068
- Classify Azure
too_many_requestsstream errors forFallbackModelby @pydanty in #11253 - Make
google_proactive_audiowork on the default Gemini provider with a separatev1alphaLive client by @DouweM in #11067 - Keep an empty response in place when AG-UI
load_messages()reloads a dumped history by @DouweM in #11192 - Keep
openai_previous_response_idchaining when a rewrite-mode instruction changes beside anon_change='append'block by @DouweM in #11331 - Record Harness
Memorysnapshots in history once per changed run, so requests stay append-only by @DouweM in #10903 - Tell the realtime model a tool call failed on every provider, so it answers instead of going silent by @DouweM in #11086
- Send the lowest effort for
thinking=Falseon OpenRouter routes that make reasoning mandatory by @DouweM in #11347 - Reserve DBOS handler step positions during tool
ctx.emit()by @dsfaccini in #11157 - Record GPT-Live speech as one
SpeechPartinstead of splitting it at each delegated tool call by @DouweM in #11187 - Keep history before a
CompactionPartwith an empty summary instead of treating it as a compaction boundary by @DouweM in #11368 - Replay local history into a fresh Gemini Live session when a reconnect has no resu...