Major Changes

• 986c6fd: feat(ai): change type of experimental_context from unknown to generic

• b0c2869: chore(ai): remove deprecated media type part from ToolResultOutput

• f7d4f01: feat(provider): add support for reasoning-file type for files that are part of reasoning

• 776b617: feat(provider): adding new 'custom' content type

• ef992f8: Remove CommonJS exports from all packages. All packages are now ESM-only ("type": "module"). Consumers using require() must switch to ESM import syntax.

• 493295c: Remove the deprecated ToolCallOptions export.

  Use ToolExecutionOptions instead.

• c29a26f: feat(provider): add support for provider references and uploading files as supported per provider

• 3887c70: feat(provider): add new top-level reasoning parameter to spec and support it in generateText and streamText

• 61753c3: ### @ai-sdk/openai: remove redundant name argument from openai.tools.customTool()

  openai.tools.customTool() no longer accepts a name field. the tool name is now derived from the sdk tool key (the object key in the tools object).

  migration: remove the name property from customTool() calls. the object key is now used as the tool name sent to the openai api.

  before:

  tools: {
    write_sql: openai.tools.customTool({
      name: 'write_sql',
      description: '...',
    }),
  }
  

  after:

  tools: {
    write_sql: openai.tools.customTool({
      description: '...',
    }),
  }
  

  @ai-sdk/provider-utils: createToolNameMapping() no longer accepts the resolveProviderToolName parameter

  before: tool name can be set dynamically

  const toolNameMapping = createToolNameMapping({
    tools,
    providerToolNames: {
      "openai.code_interpreter": "code_interpreter",
      "openai.file_search": "file_search",
      "openai.image_generation": "image_generation",
      "openai.local_shell": "local_shell",
      "openai.shell": "shell",
      "openai.web_search": "web_search",
      "openai.web_search_preview": "web_search_preview",
      "openai.mcp": "mcp",
      "openai.apply_patch": "apply_patch",
    },
    resolveProviderToolName: (tool) =>
      tool.id === "openai.custom"
        ? (tool.args as { name?: string }).name
        : undefined,
  });
  

  after: tool name is static based on tools keys

  const toolNameMapping = createToolNameMapping({
    tools,
    providerToolNames: {
      'openai.code_interpreter': 'code_interpreter',
      'openai.file_search': 'file_search',
      'openai.image_generation': 'image_generation',
      'openai.local_shell': 'local_shell',
      'openai.shell': 'shell',
      'openai.web_search': 'web_search',
      'openai.web_search_preview': 'web_search_preview',
      'openai.mcp': 'mcp',
      'openai.apply_patch': 'apply_patch',
    }
  });
  

• 7e26e81: chore: rename experimental_context to context

• 8359612: Start v7 pre-release

• 5463d0d: feat(provider): align tool result output content file part types with top-level message file part types

Patch Changes

• 2427d88: feat(ai): change Tool.sensitiveContext to telemetry.includeToolsContext and make it opt-in

• 785fe16: feat: distinguish provider-defined and provider-executed tools

• ee798eb: chore(provider-utils): rename Experimental_Sandbox to Experimental_SandboxSession

• 531251e: fix(security): validate redirect targets in download functions to prevent SSRF bypass

  Both downloadBlob and download now validate the final URL after following HTTP redirects, preventing attackers from bypassing SSRF protections via open redirects to internal/private addresses.

• 67df0a0: feat: add sensitiveContext property to Tool

• 105f95b: Ensure the default empty tool input schema includes type: "object" for OpenAI-compatible providers that require object schemas.

• eea8d98: refactoring: rename tool execution events

• d848405: feat: add optional abortSignal parameters to sandbox command execution

• 46d1149: chore(provider-utils,google): fix grammar errors in error and warning messages

• 1f509d4: fix(ai): force template check on 'kind' param

• ca446f8: feat: flexible tool descriptions

• 3ae1786: fix: better context type inference

• a7de9c9: fix: make sandbox experimental

• 9f0e36c: trigger release for all packages after provenance setup

• befb78c: refactoring: remove real-time delays in unit tests

• f634bac: feat(mcp): add new McpProviderMetadata type

• 2e17091: fix(types): move shared tool set utility types into provider-utils

  Moved ToolSet, InferToolSetContext, and UnionToIntersection into @ai-sdk/provider-utils and updated ai internals to import them directly from there. This keeps the shared tool typing utilities colocated with the core tool type definitions.

• ca39020: Add an optional workingDirectory parameter to sandbox command execution.

• 0458559: fix: deprecate needsApproval on Tool

• 5852c0a: refactoring(provider-utils): add controller as property to StreamingToolCallTracker

• 2e98477: fix: retain stack traces on async errors

• add1126: refactoring: executeTool uses tool as parameter

• aeda373: fix: only send provider credentials to same-origin response-supplied URLs

  Several provider clients followed a URL taken from the provider's API response (a polling/status URL or a final media URL such as polling_url, urls.get, result_url, result.sample, or video.uri) and reused the authenticated headers — or appended ?key=<API_KEY> — on that request. Because the host of the response-supplied URL was never validated, the long-lived API key was sent to whatever host the response named (a CDN in the benign case, or an attacker-chosen host if the provider response was tampered with), allowing credential exfiltration.

  A new isSameOrigin helper is added to @ai-sdk/provider-utils, and the affected fetches in @ai-sdk/black-forest-labs, @ai-sdk/fireworks, @ai-sdk/replicate, @ai-sdk/gladia, @ai-sdk/fal, and @ai-sdk/google now attach credentials only when the followed URL is same-origin with the provider's configured API origin. Requests to a foreign origin are made without the credential.

• 350ea38: refactoring: introduce Arrayable type

• 7fc6bd6: Raise minimum supported Node.js version to 22. Supported versions: 22, 24, and 26.

• f807e45: Extract shared StreamingToolCallTracker class into @ai-sdk/provider-utils to deduplicate streaming tool call handling across OpenAI-compatible providers. Also adds missing generateId() fallback for toolCallId in Alibaba's doGenerate path and ensures all providers finalize unfinished tool calls during stream flush.

• 08d2129: feat(mcp): propagate the server name through dynamic tool parts

• 0c4c275: trigger initial canary release

• 6fd51c0: fix(provider): preserve error type prefix in getErrorMessage

• 69254e0: feat(ai): add toolMetadata for tool specific metdata

• 6c93e36: feat(provider-utils): add spawnCommand method to Experimental_Sandbox to allow for detached command execution

• 9bd6512: feat(provider): change file part data property to be tagged with a type and remove the image part type

• 258c093: chore: ensure consistent import handling and avoid import duplicates or cycles

• 375fdd7: fix: harden download URL SSRF guard against hostname and redirect bypasses

  validateDownloadUrl and the file download helpers (downloadBlob, download) could be bypassed in several ways when handling untrusted URLs:

  • A fully-qualified hostname with a trailing dot (e.g. localhost., myhost.local.) skipped the localhost/.local blocklist.
  • IPv6 addresses that embed an IPv4 address in their last 32 bits — IPv4-compatible (::127.0.0.1), IPv4-translated (::ffff:0:127.0.0.1), and NAT64 (64:ff9b::127.0.0.1, including the 64:ff9b:1::/48 local-use prefix) — were not decoded and checked against the private IPv4 ranges.
  • Redirects were validated only after fetch had already followed them, so the request to a redirect target (e.g. an internal/metadata address) had already been issued before the check ran.
  • Several reserved/internal address ranges were not blocked: CGNAT (100.64.0.0/10, used by some cloud providers for internal traffic), benchmarking (198.18.0.0/15), IETF protocol assignments (192.0.0.0/24), the reserved 240.0.0.0/4 block (including the 255.255.255.255 broadcast address), and IPv6 site-local (fec0::/10) and multicast (ff00::/8).

  The validator now strips trailing dots before the hostname checks and fully expands IPv6 addresses to detect embedded private IPv4 targets. The download helpers now follow redirects manually (redirect: 'manual'), re-validating each hop before requesting it, so an unsafe redirect target is never fetched. When a redirect cannot be inspected because the runtime returns an opaque response, the helpers fail closed (reject the redirect) on the server; only in a real browser — where SSRF is not reachable (fetch is constrained by CORS and cannot reach a server's internal network or cloud-metadata endpoints) — is the redirect followed natively so legitimate redirected downloads keep working.

• b6783da: refactoring: restructure Tool types

• 3015fc3: feat: sandbox shell execution abstraction

• b8396f0: trigger initial beta release

• daf6637: feat(provider-utils): add env option to spawn and run methods of Experimental_SandboxSession

• a6617c5: feat(provider-utils): add readFile and writeFile plus convenience wrappers to Experimental_Sandbox abstraction

• 28dfa06: fix: support tools with optional context

• 083947b: feat(ai): separate toolsContext from context

• bae5e2b: fix(security): re-validate tool approvals from client message history before execution

  The approval-replay path in generateText/streamText (and WorkflowAgent.stream) reconstructed approved tool calls from the client-supplied messages array and executed them without re-validating input against the tool's schema or re-applying the approval policy. A client could forge an assistant message with a pre-approved tool-call part and have the server execute a tool with attacker-chosen arguments.

  The replay path now validates HMAC signature (when experimental_toolApprovalSecret is configured), re-validates tool-call input against the tool's input schema, and re-resolves the approval policy before execution.

• f617ac2: feat(provider-utils): narrow tool() return type to ExecutableTool<...> when execute is provided

• 90e2d8a: chore: fix unused vars not being flagged by our lint tooling

• b4507d5: fix(provider-utils): cancel response body on download rejection to prevent socket leak

  When a download was rejected early — because the Content-Length header exceeded the size limit, the response status was not ok, or a redirect resolved to a blocked URL — the fetch response body was left unconsumed and uncancelled. With WHATWG Fetch/undici this leaves the underlying TCP socket open instead of returning it to the connection pool, allowing an attacker-controlled origin to exhaust file descriptors and cause a denial of service. The body is now cancelled on all early-rejection paths in readResponseWithSizeLimit, download, and downloadBlob, and fetchWithValidatedRedirects cancels each redirect hop's body before following or rejecting the next hop.

• e93fa91: rename Sandbox.executeCommand to Sandbox.runCommand

• fc92055: feat(ai): automatic tool approval

• b3976a2: Add workflow serialization support to all provider models.

  @ai-sdk/provider-utils: New serializeModel() helper that extracts only serializable properties from a model instance, filtering out functions and objects containing functions. Third-party provider authors can use this to add workflow support to their own models.

  All providers: headers is now optional in provider config types. This is non-breaking — existing code that passes headers continues to work. Custom provider implementations that construct model configs manually can now omit headers, which is useful when models are deserialized from a workflow step boundary where auth is provided separately.

  All provider model classes now include WORKFLOW_SERIALIZE and WORKFLOW_DESERIALIZE static methods, enabling them to cross workflow step boundaries without serialization errors.

• ff5eba1: feat: roll image-* tool output types into their equivalent file-* types