NewYour coding agent can read the release notes before it upgrades.Set up the MCP server →
npm, #4234 most downloaded on npm
Latest stable
4.0.117
Last release 7 days ago
10 Nov 2026
Ships on a steady schedule
a new release about every 8 days
Most releases are documented
notes for 42 of the last 60 stable releases
Nothing withdrawn
no release was ever pulled
3 years old
743 releases · first in 2024
Drag across the ruler to compare, or tap a release to open it.
0adc679: feat(provider): shared spec v3
7728ac5: The mediaResolution option has been added and is now passed to the Google API.
Your coding agent can read these notes before it upgrades. Set up the MCP server →
10c1322: fix: moved dependency @ai-sdk/test-server to devDependencies
LanguageModelV3@ai-sdk/test-server to devDependenciese833473: chore (provider/google): Add preview modelIds for gemini 2.5 flash and lite
4616b86: chore: update zod peer depenedency version
@ai-sdk/provider-utils@3.1.0-beta.3
Provider-V3ImageModelV30c4822d: feat: EmbeddingModelV3
EmbeddingModelV37dea60e: add promptFeedback outputs
@ai-sdk/test-server@1.0.0-beta.0
78928cb: release: start 5.1 beta
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
f6809e3: fix(workflow): align the prepareStep results of WorkflowAgent with ToolLoopAgent. Returned instructions (or the deprecated system) replace th…
ff89832: fix(workflow): call onAbort instead of onError when a WorkflowAgent call is aborted during a model step with a custom abort reason. generate() still rejects with that reason. stream() now resolves with the steps that finished, as it does for an abort without a reason.
2473a71: fix(workflow): WorkflowAgent uses the last value yielded by an async generator tool as the tool result, like ToolLoopAgent. A generator that throws after yielding produces a tool error.
f43ee39: fix(workflow): log the default step limit warning when WorkflowAgent.generate() stops at the default limit.
7ce7ffc: fix(workflow): log provider warnings in WorkflowAgent
WorkflowAgent.generate() and WorkflowAgent.stream() now log the warnings of every model call through the AI SDK warning logger, like ToolLoopAgent. The AI_SDK_LOG_WARNINGS global controls the output.
f6809e3: fix(workflow): align the prepareStep results of WorkflowAgent with ToolLoopAgent. Returned instructions (or the deprecated system) replace the instructions of the step and carry forward to later steps, returned providerOptions are merged with the agent options, and returned headers are merged over the headers of the call.
9307549: feat(workflow): support experimental_refineToolInput in WorkflowAgent, in the constructor and per call. Refined tool inputs are used for tool execution, step results, callbacks, and the next model call, like in ToolLoopAgent.
bf1dd2a: fix(workflow): WorkflowAgent.stream() keeps the tool call in the step result when the model calls an unknown tool or passes input that does not match the schema. The call is marked invalid and the step has a tool-error part, as in generate() and ToolLoopAgent. The model still receives the error as a tool result.
453d880: fix(workflow): call the onInputAvailable tool hook in stream() when the provider sends a tool call without streaming its input first.
5c0e984: fix(workflow): report the provider name and model ID of the called model in the steps of WorkflowAgent.stream() results, matching ToolLoopAgent and WorkflowAgent.generate()
2559173: feat(workflow): support toolApproval in WorkflowAgent
WorkflowAgent accepts a toolApproval configuration in the constructor and per generate() or stream() call. The configuration takes precedence over a tool's needsApproval setting, and its functions run in workflow context, outside of the durable model steps. Calls that the configuration approves run in the same step, and calls that it denies receive an execution-denied tool result without pausing the workflow.
b9f6c82: feat(workflow): support toolOrder in WorkflowAgent
WorkflowAgent accepts toolOrder in the constructor, per call in generate() and stream(), and from prepareCall and prepareStep. Listed tools are sent first in the listed order, and the remaining tools follow in alphabetical order.
f1de9cb: feat(workflow): support the timeout object in WorkflowAgent
generate() and stream() accept timeout: { totalMs, stepMs, firstChunkMs, chunkMs } in addition to a number. stepMs limits each model call. In stream(), firstChunkMs and chunkMs limit the wait for model output. Invalid timeout values are rejected with an InvalidArgumentError. A timeout now also ends a model step when the provider stream ignores the abort signal.
Updated dependencies [3c09e00]
Updated dependencies [f43ee39]
Updated dependencies [7ce7ffc]
Updated dependencies [f6809e3]
Updated dependencies [2559173]
Updated dependencies [f1de9cb]
b161730: fix(workflow): WorkflowAgent now accepts onError and onAbort in the constructor, like ToolLoopAgent. They are called before the callbacks pas
WorkflowAgent now accepts onError and onAbort in the constructor, like ToolLoopAgent. They are called before the callbacks passed to generate() or stream(). An error thrown by an onError or onAbort callback no longer replaces the error or result of the run.WorkflowAgent no longer receive the system prompt or the assistant message that contains the current tool call in options.messages. The messages passed to onToolExecutionStart and onToolExecutionEnd no longer contain the system prompt either. This matches ToolLoopAgent, generateText, and streamText.model, toolChoice, activeTools, and generation settings returned from prepareStep in WorkflowAgent now apply to the current step only, like in ToolLoopAgent. Before, an override returned for one step stayed in effect for all later steps.WorkflowAgent.stream() now aggregates totalUsage the same way as ToolLoopAgent. Token details such as cached and reasoning tokens are kept, and counts that the provider did not report stay undefined instead of becoming 0.### Patch Changes - Updated dependencies [11323f9] - Updated dependencies [2ebc78b] - Updated dependencies [2ebc78b] - Updated dependencies [cb4fa52]
8549a1d: feat(xai): add safetyIdentifier Responses provider option
Sends safety_identifier with Responses API requests so xAI can attribute a policy violation to one end user, and exposes the echoed value on providerMetadata.xai.safetyIdentifier.
Updated dependencies [49c554d]
Updated dependencies [db18e4f]
Updated dependencies [ce27f15]
Updated dependencies [b753c03]
efd2905: fix(workflow): End the final step's telemetry when streaming stops on a client-side tool, preserving the step span and its model call in the
Updated dependencies [ a1a772f ]
8e40444 : Defer validation schema construction until first use to reduce provider import time.
### Patch Changes - Updated dependencies [6d50947] - @ai-sdk/provider-utils@5.0.81
### Patch Changes - Updated dependencies [f5fc6ba] - @ai-sdk/anthropic@4.0.105
2ebc78b: Move Node.js connection-time DNS validation to the optional guarded-fetch package. It uses an Undici agent to validate every resolved address and pin the connection to those results, independently of global fetch wrappers. It supports explicitly trusted private origins and connection-pool cleanup.
Security behavior change: provider-utils now uses the configured fetch or the runtime's global fetch and no longer depends on Undici. Default URL checks still reject private IP literals, local hostnames, unsafe schemes, and unsafe redirect URLs, but do not block public-looking hostnames that resolve to private addresses or DNS rebinding. Applications fetching untrusted URLs should inject createGuardedFetch() on Node.js or enforce equivalent network egress restrictions.
Allow downloadBlob to accept a fetch transport so applications can apply the same protection to file downloads.
MCP OAuth discovery keeps its strict DNS requirement: because Node.js no longer has a DNS-pinned default transport, untrusted authorization-server, redirect, token, and registration URLs now require an explicit custom fetch with connection-time DNS validation on Node.js too. Pass createGuardedFetch() through transport.fetch. The configured MCP server origin and browser window discovery remain supported.
Updated dependencies [2ebc78b]
Updated dependencies [cb4fa52]
6309382: Add support for baseUriDomains in MCP App CSP metadata.
baseUriDomains in MCP App CSP metadata.Patch Changes ai@7.0.164
85b920f : feat(workflow): remove experimental_transform from WorkflowAgent.stream()
85b920f: feat(workflow): remove experimental_transform from WorkflowAgent.stream()
Stream transforms cannot reliably run across workflow step boundaries, so the option has been removed along with the StreamTextTransform type export.
Updated dependencies [5d18351]
Updated dependencies [8da48db]
Updated dependencies [fd7e8a7]
Updated dependencies [36802d4]
Updated dependencies [933dae3]
Updated dependencies [e0143b0]
Updated dependencies [934100c]
Updated dependencies [c389374]
Updated dependencies [892af4d]
Updated dependencies [892af4d]
Updated dependencies [89aff56]
Updated dependencies [1a1da8d]
Updated dependencies [41465fd]
Updated dependencies [9754014]
Updated dependencies [4bb73b0]
Updated dependencies [2e44144]
4d6d830 : Support a single telemetry operation for durable WorkflowAgent runs, including the prompt, final output, steps, and tools. Events recorded i
WorkflowAgent runs, including the prompt, final output, steps, and tools. Events recorded in the workflow are reported on completion or suspension with their original timestamps. For OpenTelemetry, pass the same OpenTelemetryIdGenerator instance as the tracer provider's idGenerator and sampler and as new OpenTelemetry({ durable }). This explicit bridge carries parent trace context and sampling across steps, reports model spans live with provider requests nested beneath them, and reconstructs the enclosing invoke_agent tree when the run completes. Without the bridge, existing live per-model tracing is preserved. Add typed experimental durable context hooks for telemetry integrations; remove reliance on process-global ID-generator detection.f5fc6ba: Derive known model ID suggestions from each provider's model catalog declarations so catalog updates automatically update autocomplete, while
### Patch Changes - Updated dependencies [6d50947] - @ai-sdk/provider-utils@5.0.81
5c247e9: feat(sandbox): add stat, listDirectory, and deleteFile to Experimental_SandboxSession, and specify that relative paths in every method resolve against the directory run and spawn start in
The just-bash and Vercel Sandbox adapters implement the new methods. Both now resolve a relative workingDirectory for run and spawn against the sandbox's working directory, and the just-bash adapter rejects writes to a directory instead of replacing it with a file.
Updated dependencies [7716e18]
Updated dependencies [8ba3a5c]
Updated dependencies [5c247e9]
Updated dependencies [dcd2f8d]
### Patch Changes - Updated dependencies [2ebc78b] - Updated dependencies [cb4fa52] - @ai-sdk/provider-utils@5.0.80 - @ai-sdk/provider@4.0.38
filterActiveTools from experimental. It is now exported as filterActiveTools; experimental_filterActiveTools remains as a deprecated alias.df0cbe9: feat(google): add gemini-3.7-flash model
gemini-3.7-flash model### Patch Changes - Updated dependencies [6d50947] - @ai-sdk/provider-utils@5.0.81
1e632f7: feat(xai): add Grok 4.6 models
### Patch Changes - Updated dependencies [6d50947] - @ai-sdk/provider-utils@5.0.81
### Patch Changes - Updated dependencies [6d50947] - @ai-sdk/provider-utils@5.0.81
### Patch Changes - Updated dependencies [2ebc78b] - Updated dependencies [cb4fa52] - @ai-sdk/provider-utils@5.0.80 - @ai-sdk/provider@4.0.38
8ee1a11: Centralize QuiverAI language and image model IDs, editing and animation restrictions, capabilities, and provider option metadata in a model c
2ebc78b: Forward the provider's configured fetch transport when downloading file inputs. Add a fetch option to createDownload so core downloads can also use guarded-fetch for opt-in DNS validation on Node.js. The function returned by createDownload can now also be passed as experimental_download to generateText and streamText, and experimental_download now also handles generated file URLs returned by the model, so one hook covers all core downloads. Document that the default URL checks do not resolve DNS and recommend the guarded transport or network egress controls for untrusted URLs.
Keep Azure image and Vertex tool-result downloads separate from provider authentication wrappers so file hosts never receive Azure bearer tokens or Vertex Express Mode API keys.
Pass the call's abort signal to custom download requests, including prompt assets and generated files, so experimental_download: createDownload({ fetch }) preserves cancellation and timeouts.
Updated dependencies [2ebc78b]
Updated dependencies [cb4fa52]
### Patch Changes - Updated dependencies [2ebc78b] - Updated dependencies [cb4fa52] - @ai-sdk/provider-utils@5.0.80 - @ai-sdk/provider@4.0.38
0a5d0e7: Fix toolChoice: 'none' by forwarding it to the Hugging Face Responses API and rejecting tool calls before local execution in generateText, streamText, and WorkflowAgent. Streamed tool parts are withheld so client-side tools cannot act on them, and repairToolCall and experimental_refineToolInput are not invoked for rejected tool calls.
10026d1: Add the interrupted finish reason for provider streams that end without a completion signal. Preserve partial output without introducing an error solely for interruption, and stop tool execution, agent continuation, and automatic UI follow-ups for interrupted responses, including late tool results and approvals. Provider detection and emission will be added separately.
Compatibility notice: Adding interrupted to the public finish-reason types can break type checking after upgrading, even though this is a patch release. Exhaustive switches (for example, using an assertNever fallback) and mappings such as Record<FinishReason, string> must handle the new value. Update any custom validation or application logic that assumes the existing set of finish reasons is exhaustive.
When providers begin emitting interrupted, a stream ending without a completion signal will preserve partial output and report this finish reason without emitting an error solely for the interruption. Applications that need to detect incomplete responses should check finishReason === 'interrupted' as well as their existing error handling.
08335ae: Retain custom content parts (e.g. openai.compaction) in WorkflowAgent.stream() step results and subsequent model requests
Updated dependencies [1727442]
Updated dependencies [0a5d0e7]
Updated dependencies [10026d1]
Updated dependencies [ b089ce8 ]
Nothing published for this version
@ai-sdk/openai-compatible@2.0.69
@ai-sdk/openai-compatible@2.0.68
793220f : Refactor Voyage model ID logic into a model catalog that declares capabilities and provider options, and generate provider capability docume
@ai-sdk/openai-compatible@2.0.69
@ai-sdk/openai-compatible@2.0.68
@ai-sdk/openai-compatible@2.0.69
@ai-sdk/openai-compatible@2.0.68
@ai-sdk/openai-compatible@2.0.69
@ai-sdk/openai-compatible@2.0.68
WorkflowChatTransport retrying indefinitely when successful reconnect responses contain no UI message chunks. Empty streams now count toward maxConsecutiveErrors, while streams that receive chunks continue to advance the resume cursor and reset the error count.7440d89: fix(openai-compatible): preserve unmapped usage fields in usage.raw
7440d89: fix(openai-compatible): preserve unmapped usage fields in usage.raw
usage.raw is specified as usage "in the shape that the provider returns",
and the chat model already parsed the usage object loosely so that extra
top-level fields survived. The nested prompt_tokens_details and
completion_tokens_details objects were still strict, so anything a provider
reported inside them was dropped — which is where providers put their most
distinguishing detail. The existing "should preserve extra usage fields"
fixture was itself losing audio_tokens, image_tokens and text_tokens
this way.
Both nested objects are now parsed loosely, as is the completion model's usage
schema, which was strict throughout despite feeding the same raw field.
Only usage.raw changes. The mapped token counts are unaffected.
### Patch Changes - Updated dependencies [31205a4] - @ai-sdk/provider-utils@4.0.46
Nothing published for this version
Nothing published for this version
Patch Changes Updated dependencies [9184a35] ai@7.0.133 @ai-sdk/provider@4.0.25 @ai-sdk/provider-utils@5.0.57
Updated dependencies [ 0c2a9de ]
Nothing published for this version
Updated dependencies [ 6f6b9c0 ]
Nothing published for this version
d9e04cb : fix(workflow): reuse persisted tool denial results during approval resumption
a471b4c : Document and validate WorkflowAgent generation deadlines, retry boundaries, hook cancellation, and selected result serialization across work
WorkflowAgent.generate() with a shared tool loop, a 20-step default, typed output, and core generation result semantics. Reuse core result and content construction through internal exports while preserving existing Workflow streaming behavior. Transport-independent approval creation remains separate follow-up work.One column per month.