github mastra-ai/mastra @mastra/core@0.24.6
2025-11-27
on GitHub

7 hours ago

Highlights

Stream nested execution context from Workflows and Networks to your UI

Agent responses now stream live through workflows and networks, with complete execution metadata flowing to your UI.

In workflows, pipe agent streams directly through steps: 

const planActivities = createStep({
  execute: async ({ mastra, writer }) => {
    const agent = mastra?.getAgent('weatherAgent');
    const response = await agent.stream('Plan activities');
    await response.fullStream.pipeTo(writer);
    
    return { activities: await response.text };
  }
});

In networks, each step now tracks properly—unique IDs, iteration counts, task info, and agent handoffs all flow through with correct sequencing. No more duplicated steps or missing metadata.

Both surface text chunks, tool calls, and results as they happen, so users see progress in real time instead of waiting for the full response.

AI-SDK voice models are now supported

CompositeVoice now accepts AI SDK voice models directly—use OpenAI for transcription, ElevenLabs for speech, or any combination you want. 

import { CompositeVoice } from "@mastra/core/voice";
import { openai } from "@ai-sdk/openai";
import { elevenlabs } from "@ai-sdk/elevenlabs";

const voice = new CompositeVoice({
  input: openai.transcription('whisper-1'),
  output: elevenlabs.speech('eleven_turbo_v2'),
});

const audio = await voice.speak("Hello from AI SDK!");
const transcript = await voice.listen(audio);

Works with OpenAI, ElevenLabs, Groq, Deepgram, LMNT, Hume, and more. AI SDK models are automatically wrapped, so you can swap providers without changing your code.

Changelog

@mastra/ai-sdk

  • Support streaming agent text chunks from workflow-step-output

    Adds support for streaming text and tool call chunks from agents running inside workflows via the workflow-step-output event. When you pipe an agent's stream into a workflow step's writer, the text chunks, tool calls, and other streaming events are automatically included in the workflow stream and converted to UI messages.

Features:

  • Added includeTextStreamParts option to WorkflowStreamToAISDKTransformer (defaults to true)
  • Added isMastraTextStreamChunk type guard to identify Mastra chunks with text streaming data
  • Support for streaming text chunks: text-start, text-delta, text-end
  • Support for streaming tool calls: tool-call, tool-result
  • Comprehensive test coverage in transformers.test.ts
  • Updated documentation for workflow streaming and workflowRoute()

Example: 

const planActivities = createStep({
  execute: async ({ mastra, writer }) => {
    const agent = mastra?.getAgent('weatherAgent');
    const response = await agent.stream('Plan activities');
    await response.fullStream.pipeTo(writer);
    
    return { activities: await response.text };
  }
});

When served via workflowRoute(), the UI receives incremental text updates as the agent generates its response, providing a smooth streaming experience. (#10568)

  • Fix chat route to use agent ID instead of agent name for resolution. The /chat/:agentId endpoint now correctly resolves agents by their ID property (e.g., weather-agent) instead of requiring the camelCase variable name (e.g., weatherAgent). This fixes issue #10469 where URLs like /chat/weather-agent would return 404 errors. (#10565)

  • Fixes propagation of custom data chunks from nested workflows in branches to the root stream when using toAISdkV5Stream with {from: 'workflow'}.

    Previously, when a nested workflow within a branch used writer.custom() to write data-* chunks, those chunks were wrapped in workflow-step-output events and not extracted, causing them to be dropped from the root stream.

    Changes:

    • Added handling for workflow-step-output chunks in transformWorkflow() to extract and propagate data-* chunks

  • When a workflow-step-output chunk contains a data-* chunk in its payload.output, the transformer now extracts it and returns it directly to the root stream

  • Added comprehensive test coverage for nested workflows with branches and custom data propagation

    This ensures that custom data chunks written via writer.custom() in nested workflows (especially those within branches) are properly propagated to the root stream, allowing consumers to receive progress updates, metrics, and other custom data from nested workflow steps. (#10447)

  • Fix network data step formatting in AI SDK stream transformation

    Previously, network execution steps were not being tracked correctly in the AI SDK stream transformation. Steps were being duplicated rather than updated, and critical metadata like step IDs, iterations, and task information was missing or incorrectly structured.

    Changes:

    • Enhanced step tracking in AgentNetworkToAISDKTransformer to properly maintain step state throughout execution lifecycle

  • Steps are now identified by unique IDs and updated in place rather than creating duplicates

  • Added proper iteration and task metadata to each step in the network execution flow

  • Fixed agent, workflow, and tool execution events to correctly populate step data

  • Updated network stream event types to include networkId, workflowId, and consistent runId tracking

  • Added test coverage for network custom data chunks with comprehensive validation

    This ensures the AI SDK correctly represents the full execution flow of agent networks with accurate step sequencing and metadata. (#10432)

  • [0.x] Make workflowRoute includeTextStreamParts option default to false (#10574)

  • Add support for tool-call-approval and tool-call-suspended events in chatRoute (#10205)

  • Backports the messageMetadata and onError support from PR #10313 to the 0.x branch, adding these features to toAISdkFormat function.

    • Added messageMetadata parameter to toAISdkFormat options
    • Function receives the current stream part and returns metadata to attach to start and finish chunks
    • Metadata is included in start and finish chunks when provided

  • Added onError parameter to toAISdkFormat options

    • Allows custom error handling during stream conversion
    • Falls back to safeParseErrorObject utility when not provided

  • Added safeParseErrorObject utility function for error parsing

  • Updated AgentStreamToAISDKTransformer to accept and use messageMetadata and onError

  • Updated JSDoc documentation with parameter descriptions and usage examples

  • Added comprehensive test suite for messageMetadata functionality (6 test cases)

  • Fixed existing test file to use toAISdkFormat instead of removed toAISdkV5Stream

    • All existing tests pass (14 tests across 3 test files)

  • New tests verify:

    • messageMetadata is called with correct part structure

    • Metadata is included in start and finish chunks

    • Proper handling when messageMetadata is not provided or returns null/undefined

    • Function is called for each relevant part in the stream

      • Uses UIMessageStreamOptions<UIMessage>['messageMetadata'] and UIMessageStreamOptions<UIMessage>['onError'] types from AI SDK v5 for full type compatibility

      • Backport of: #10313 (#10314)

  • Fixed workflow routes to properly receive request context from middleware. This aligns the behavior of workflowRoute with chatRoute, ensuring that context set in middleware is consistently forwarded to workflows.

    When both middleware and request body provide a request context, the middleware value now takes precedence, and a warning is emitted to help identify potential conflicts.

    See #10427 (#10427)

@mastra/astra

  • deleteVectors, deleteFilter when upserting, updateVector filter (#10244) (#10244)

@mastra/auth-clerk

  • remove organization requirement from default authorization (#10551)

@mastra/chroma

  • deleteVectors, deleteFilter when upserting, updateVector filter (#10244) (#10244)

@mastra/clickhouse

  • fix: ensure score responses match saved payloads for Mastra Stores. (#10570)
  • Fix message sorting in getMessagesPaginated when using semantic recall (include parameter). Messages are now always sorted by createdAt after combining paginated and included messages, ensuring correct chronological ordering of conversation history. All stores now consistently use MessageList for deduplication followed by explicit sorting. (#10573)

@mastra/cloudflare

  • fix: ensure score responses match saved payloads for Mastra Stores. (#10570)
  • Fix message sorting in listMessages when using semantic recall (include parameter). Messages are now always sorted by createdAt instead of storage order, ensuring correct chronological ordering of conversation history. (#10545)

@mastra/cloudflare-d1

  • fix: ensure score responses match saved payloads for Mastra Stores. (#10570)
  • Fix message sorting in getMessagesPaginated when using semantic recall (include parameter). Messages are now always sorted by createdAt after combining paginated and included messages, ensuring correct chronological ordering of conversation history. All stores now consistently use MessageList for deduplication followed by explicit sorting. (#10573)

@mastra/core

  • Fix base64 encoded images with threads - issue #10480

    Fixed "Invalid URL" error when using base64 encoded images (without data: prefix) in agent calls with threads and resources. Raw base64 strings are now automatically converted to proper data URIs before being processed.

    Changes:

  • Updated attachments-to-parts.ts to detect and convert raw base64 strings to data URIs

  • Fixed MessageList image processing to handle raw base64 in two locations:

    • Image part conversion in aiV4CoreMessageToV1PromptMessage
    • File part to experimental_attachments conversion in mastraDBMessageToAIV4UIMessage

  • Added comprehensive tests for base64 images, data URIs, and HTTP URLs with threads

    Breaking Change: None - this is a bug fix that maintains backward compatibility while adding support for raw base64 strings. (#10483)

  • SimpleAuth and improved CloudAuth (#10569)

  • Fixed OpenAI schema compatibility when using agent.generate() or agent.stream() with structuredOutput.

    Changes

    • Automatic transformation: Zod schemas are now automatically transformed for OpenAI strict mode compatibility when using OpenAI models (including reasoning models like o1, o3, o4)

  • Optional field handling: .optional() fields are converted to .nullable() with a transform that converts null → undefined, preserving optional semantics while satisfying OpenAI's strict mode requirements

  • Preserves nullable fields: Intentionally .nullable() fields remain unchanged

  • Deep transformation: Handles .optional() fields at any nesting level (objects, arrays, unions, etc.)

  • JSON Schema objects: Not transformed, only Zod schemas

    Example

const agent = new Agent({
  name: 'data-extractor',
  model: { provider: 'openai', modelId: 'gpt-4o' },
  instructions: 'Extract user information',
});

	const schema = z.object({
  name: z.string(),
  age: z.number().optional(),
  deletedAt: z.date().nullable(),
});

	// Schema is automatically transformed for OpenAI compatibility
const result = await agent.generate('Extract: John, deleted yesterday', {
  structuredOutput: { schema },
});

	// Result: { name: 'John', age: undefined, deletedAt: null }

(#10454)

  • deleteVectors, deleteFilter when upserting, updateVector filter (#10244) (#10244)

  • Fix generateTitle model type to accept AI SDK LanguageModelV2

    Updated the generateTitle.model config option to accept MastraModelConfig instead of MastraLanguageModel. This allows users to pass raw AI SDK LanguageModelV2 models (e.g., anthropic.languageModel('claude-3-5-haiku-20241022')) directly without type errors.

    Previously, passing a standard LanguageModelV2 would fail because MastraLanguageModelV2 has different doGenerate/doStream return types. Now MastraModelConfig is used consistently across:

  • memory/types.ts - generateTitle.model config

  • agent.ts - genTitle, generateTitleFromUserMessage, resolveTitleGenerationConfig

  • agent-legacy.ts - AgentLegacyCapabilities interface (#10567)

  • Fix message metadata not persisting when using simple message format. Previously, custom metadata passed in messages (e.g., {role: 'user', content: 'text', metadata: {userId: '123'}}) was not being saved to the database. This occurred because the CoreMessage conversion path didn't preserve metadata fields.

    Now metadata is properly preserved for all message input formats:

  • Simple CoreMessage format: {role, content, metadata}

  • Full UIMessage format: {role, content, parts, metadata}

  • AI SDK v5 ModelMessage format with metadata

    Fixes #8556 (#10488)

  • feat: Composite auth implementation (#10359)

  • Fix requireApproval property being ignored for tools passed via toolsets, clientTools, and memoryTools parameters. The requireApproval flag now correctly propagates through all tool conversion paths, ensuring tools requiring approval will properly request user approval before execution. (#10562)

  • Fix Azure Foundry rate limit handling for -1 values (#10409)

  • Fix model headers not being passed through gateway system

    Previously, custom headers specified in MastraModelConfig were not being passed through the gateway system to model providers. This affected:

  • OpenRouter (preventing activity tracking with HTTP-Referer and X-Title)

  • Custom providers using custom URLs (headers not passed to createOpenAICompatible)

  • Custom gateway implementations (headers not available in resolveLanguageModel)

    Now headers are correctly passed through the entire gateway system:

  • Base MastraModelGateway interface updated to accept headers

  • ModelRouterLanguageModel passes headers from config to all gateways

  • OpenRouter receives headers for activity tracking

  • Custom URL providers receive headers via createOpenAICompatible

  • Custom gateways can access headers in their resolveLanguageModel implementation

    Example usage: 

// Works with OpenRouter
const agent = new Agent({
  name: 'my-agent',
  instructions: 'You are a helpful assistant.',
  model: {
    id: 'openrouter/anthropic/claude-3-5-sonnet',
    headers: {
      'HTTP-Referer': 'https://myapp.com',
      'X-Title': 'My Application',
    },
  },
});

	// Also works with custom providers
const customAgent = new Agent({
  name: 'custom-agent',
  instructions: 'You are a helpful assistant.',
  model: {
    id: 'custom-provider/model',
    url: 'https://api.custom.com/v1',
    apiKey: 'key',
    headers: {
      'X-Custom-Header': 'custom-value',
    },
  },
});

Fixes #9760
(#10564)

  • fix(agent): persist messages before tool suspension

    Fixes issues where thread and messages were not saved before suspension when tools require approval or call suspend() during execution. This caused conversation history to be lost if users refreshed during tool approval or suspension.

    Backend changes (@mastra/core):

  • Add assistant messages to messageList immediately after LLM execution

  • Flush messages synchronously before suspension to persist state

  • Create thread if it doesn't exist before flushing

  • Add metadata helpers to persist and remove tool approval state

  • Pass saveQueueManager and memory context through workflow for immediate persistence

    Frontend changes (@mastra/react):

  • Extract runId from pending approvals to enable resumption after refresh

  • Convert pendingToolApprovals (DB format) to requireApprovalMetadata (runtime format)

  • Handle both dynamic-tool and tool-{NAME} part types for approval state

  • Change runId from hardcoded agentId to unique uuid()

    UI changes (@mastra/playground-ui):

  • Handle tool calls awaiting approval in message initialization

  • Convert approval metadata format when loading initial messages

    Fixes #9745, #9906 (#10369)

  • Fix race condition in parallel tool stream writes

    Introduces a write queue to ToolStream to serialize access to the underlying stream, preventing writer locked errors (#10463)

  • Remove unneeded console warning when flushing messages and no threadId or saveQueueManager is found. (#10369)

  • Fixes GPT-5 reasoning which was failing on subsequent tool calls with the error: 

Item 'fc_xxx' of type 'function_call' was provided without its required 'reasoning' item: 'rs_xxx'

(#10489)

  • Add optional includeRawChunks parameter to agent execution options,
    allowing users to include raw chunks in stream output where supported
    by the model provider. (#10456)

  • When mastra dev runs, multiple processes can write to provider-registry.json concurrently (auto-refresh, syncGateways, syncGlobalCacheToLocal). This causes file corruption where the end of the JSON appears twice, making it unparseable.

    The fix uses atomic writes via the write-to-temp-then-rename pattern. Instead of: 

fs.writeFileSync(filePath, content, 'utf-8');
We now do:

const tempPath = `${filePath}.${process.pid}.${Date.now()}.${randomSuffix}.tmp`;
fs.writeFileSync(tempPath, content, 'utf-8');
fs.renameSync(tempPath, filePath); // atomic on POSIX

fs.rename() is atomic on POSIX systems when both paths are on the same filesystem, so concurrent writes will each complete fully rather than interleaving. (#10529)

  • Ensures that data chunks written via writer.custom() always bubble up directly to the top-level stream, even when nested in sub-agents. This allows tools to emit custom progress updates, metrics, and other data that can be consumed at any level of the agent hierarchy.

    • Added bubbling logic in sub-agent execution: When sub-agents execute, data chunks (chunks with type starting with data-) are detected and written via writer.custom() instead of writer.write(), ensuring they bubble up directly without being wrapped in tool-output chunks.

    • Added comprehensive tests:

    • Test for writer.custom() with direct tool execution

    • Test for writer.custom() with sub-agent tools (nested execution)

    • Test for mixed usage of writer.write() and writer.custom() in the same tool

      When a sub-agent's tool uses writer.custom() to write data chunks, those chunks appear in the sub-agent's stream. The parent agent's execution logic now detects these chunks and uses writer.custom() to bubble them up directly, preserving their structure and making them accessible at the top level.

      This ensures that:

      • Data chunks from tools always appear directly in the stream (not wrapped)

  • Data chunks bubble up correctly through nested agent hierarchies

  • Regular chunks continue to be wrapped in tool-output as expected (#10309)

  • Adds ability to create custom MastraModelGateway's that can be added to the Mastra class instance under the gateways property. Giving you typescript autocompletion in any model picker string. 

import { MastraModelGateway, type ProviderConfig } from '@mastra/core/llm';
import { createOpenAICompatible } from '@ai-sdk/openai-compatible';
import type { LanguageModelV2 } from '@ai-sdk/provider';

	class MyCustomGateway extends MastraModelGateway {
  readonly id = 'custom';
  readonly name = 'My Custom Gateway';

	  async fetchProviders(): Promise<Record<string, ProviderConfig>> {
    return {
      'my-provider': {
        name: 'My Provider',
        models: ['model-1', 'model-2'],
        apiKeyEnvVar: 'MY_API_KEY',
        gateway: this.id,
      },
    };
  }

	  buildUrl(modelId: string, envVars?: Record<string, string>): string {
    return 'https://api.my-provider.com/v1';
  }

	  async getApiKey(modelId: string): Promise<string> {
    const apiKey = process.env.MY_API_KEY;
    if (!apiKey) throw new Error('MY_API_KEY not set');
    return apiKey;
  }

	  async resolveLanguageModel({
    modelId,
    providerId,
    apiKey,
  }: {
    modelId: string;
    providerId: string;
    apiKey: string;
  }): Promise<LanguageModelV2> {
    const baseURL = this.buildUrl(`${providerId}/${modelId}`);
    return createOpenAICompatible({
      name: providerId,
      apiKey,
      baseURL,
    }).chatModel(modelId);
  }
}

	new Mastra({
  gateways: {
    myGateway: new MyCustomGateway(),
  },
});

(#10535)

  • Support AI SDK voice models

    Mastra now supports AI SDK's transcription and speech models directly in CompositeVoice, enabling seamless integration with a wide range of voice providers through the AI SDK ecosystem. This allows you to use models from OpenAI, ElevenLabs, Groq, Deepgram, LMNT, Hume, and many more for both speech-to-text (transcription) and text-to-speech capabilities.

    AI SDK models are automatically wrapped when passed to CompositeVoice, so you can mix and match AI SDK models with existing Mastra voice providers for maximum flexibility.

    Usage Example

import { CompositeVoice } from "@mastra/core/voice";
import { openai } from "@ai-sdk/openai";
import { elevenlabs } from "@ai-sdk/elevenlabs";

	// Use AI SDK models directly with CompositeVoice
const voice = new CompositeVoice({
  input: openai.transcription('whisper-1'),      // AI SDK transcription model
  output: elevenlabs.speech('eleven_turbo_v2'),  // AI SDK speech model
});

	// Convert text to speech
const audioStream = await voice.speak("Hello from AI SDK!");

	// Convert speech to text
const transcript = await voice.listen(audioStream);
console.log(transcript);

Fixes #9947 (#10558)

  • Fix network data step formatting in AI SDK stream transformation

    Previously, network execution steps were not being tracked correctly in the AI SDK stream transformation. Steps were being duplicated rather than updated, and critical metadata like step IDs, iterations, and task information was missing or incorrectly structured.

    Changes:

    • Enhanced step tracking in AgentNetworkToAISDKTransformer to properly maintain step state throughout execution lifecycle

  • Steps are now identified by unique IDs and updated in place rather than creating duplicates

  • Added proper iteration and task metadata to each step in the network execution flow

  • Fixed agent, workflow, and tool execution events to correctly populate step data

  • Updated network stream event types to include networkId, workflowId, and consistent runId tracking

  • Added test coverage for network custom data chunks with comprehensive validation

    This ensures the AI SDK correctly represents the full execution flow of agent networks with accurate step sequencing and metadata. (#10432)

  • Fix generating provider-registry.json (#10535)

  • Fix message-list conversion issues when persisting messages before tool suspension: filter internal metadata fields (__originalContent) from UI messages, keep reasoning field empty for consistent cache keys during message deduplication, and only include providerMetadata on parts when defined. (#10552)

  • Fix agent.generate() to use model's doGenerate method instead of doStream

    When calling agent.generate(), the model's doGenerate method is now correctly invoked instead of always using doStream. This aligns the non-streaming generation path with the intended behavior where providers can implement optimized non-streaming responses. (#10572)

@mastra/couchbase

  • deleteVectors, deleteFilter when upserting, updateVector filter (#10244) (#10244)

@mastra/deployer

  • Rename "Playground" to "Studio" (#10443)
  • Fixed a bug where imports that were not used in the main entry point were tree-shaken during analysis, causing bundling errors. Tree-shaking now only runs during the bundling step. (#10470)

@mastra/deployer-cloud

  • SimpleAuth and improved CloudAuth (#10569)
  • Do not initialize local storage when using mastra cloud storage instead (#10495)

@mastra/dynamodb

  • fix: ensure score responses match saved payloads for Mastra Stores. (#10570)
  • Fix message sorting in getMessagesPaginated when using semantic recall (include parameter). Messages are now always sorted by createdAt after combining paginated and included messages, ensuring correct chronological ordering of conversation history. All stores now consistently use MessageList for deduplication followed by explicit sorting. (#10573)

@mastra/inngest

  • Emit workflow-step-result and workflow-step-finish when step fails in inngest workflow (#10555)

@mastra/lance

  • deleteVectors, deleteFilter when upserting, updateVector filter (#10244) (#10244)
  • fix: ensure score responses match saved payloads for Mastra Stores. (#10570)

@mastra/libsql

  • deleteVectors, deleteFilter when upserting, updateVector filter (#10244) (#10244)
  • Fix message sorting in getMessagesPaginated when using semantic recall (include parameter). Messages are now always sorted by createdAt after combining paginated and included messages, ensuring correct chronological ordering of conversation history. All stores now consistently use MessageList for deduplication followed by explicit sorting. (#10573)

@mastra/mcp

  • Fix MCP client to return structuredContent directly when tools define an outputSchema, ensuring output validation works correctly instead of failing with "expected X, received undefined" errors. (#10442)

@mastra/mcp-docs-server

  • Ensure changelog truncation includes at least 2 versions before cutting off (#10496)

@mastra/mongodb

  • deleteVectors, deleteFilter when upserting, updateVector filter (#10244) (#10244)
  • fix: ensure score responses match saved payloads for Mastra Stores. (#10570)
  • Fix message sorting in getMessagesPaginated when using semantic recall (include parameter). Messages are now always sorted by createdAt after combining paginated and included messages, ensuring correct chronological ordering of conversation history. All stores now consistently use MessageList for deduplication followed by explicit sorting. (#10573)

@mastra/mssql

  • fix: ensure score responses match saved payloads for Mastra Stores. (#10570)
  • Fix message sorting in getMessagesPaginated when using semantic recall (include parameter). Messages are now always sorted by createdAt after combining paginated and included messages, ensuring correct chronological ordering of conversation history. All stores now consistently use MessageList for deduplication followed by explicit sorting. (#10573)

@mastra/opensearch

  • deleteVectors, deleteFilter when upserting, updateVector filter (#10244) (#10244)

@mastra/pg

  • deleteVectors, deleteFilter when upserting, updateVector filter (#10244) (#10244)
  • fix: ensure score responses match saved payloads for Mastra Stores. (#10570)
  • Fix message sorting in getMessagesPaginated when using semantic recall (include parameter). Messages are now always sorted by createdAt after combining paginated and included messages, ensuring correct chronological ordering of conversation history. All stores now consistently use MessageList for deduplication followed by explicit sorting. (#10573)

@mastra/playground-ui

  • fix(agent): persist messages before tool suspension

    Fixes issues where thread and messages were not saved before suspension when tools require approval or call suspend() during execution. This caused conversation history to be lost if users refreshed during tool approval or suspension.

    Backend changes (@mastra/core):

  • Add assistant messages to messageList immediately after LLM execution

  • Flush messages synchronously before suspension to persist state

  • Create thread if it doesn't exist before flushing

  • Add metadata helpers to persist and remove tool approval state

  • Pass saveQueueManager and memory context through workflow for immediate persistence

    Frontend changes (@mastra/react):

  • Extract runId from pending approvals to enable resumption after refresh

  • Convert pendingToolApprovals (DB format) to requireApprovalMetadata (runtime format)

  • Handle both dynamic-tool and tool-{NAME} part types for approval state

  • Change runId from hardcoded agentId to unique uuid()

    UI changes (@mastra/playground-ui):

  • Handle tool calls awaiting approval in message initialization

  • Convert approval metadata format when loading initial messages

    Fixes #9745, #9906 (#10369)

@mastra/rag

  • Add metadata filtering support a GraphRag (#10352) (#10403)

@mastra/react

  • Configurable resourceId in react useChat (#10561)

  • fix(agent): persist messages before tool suspension

    Fixes issues where thread and messages were not saved before suspension when tools require approval or call suspend() during execution. This caused conversation history to be lost if users refreshed during tool approval or suspension.

    Backend changes (@mastra/core):

  • Add assistant messages to messageList immediately after LLM execution

  • Flush messages synchronously before suspension to persist state

  • Create thread if it doesn't exist before flushing

  • Add metadata helpers to persist and remove tool approval state

  • Pass saveQueueManager and memory context through workflow for immediate persistence

    Frontend changes (@mastra/react):

  • Extract runId from pending approvals to enable resumption after refresh

  • Convert pendingToolApprovals (DB format) to requireApprovalMetadata (runtime format)

  • Handle both dynamic-tool and tool-{NAME} part types for approval state

  • Change runId from hardcoded agentId to unique uuid()

    UI changes (@mastra/playground-ui):

  • Handle tool calls awaiting approval in message initialization

  • Convert approval metadata format when loading initial messages

    Fixes #9745, #9906 (#10369)

@mastra/schema-compat

  • Fixed OpenAI schema compatibility when using agent.generate() or agent.stream() with structuredOutput.

    Changes

    • Automatic transformation: Zod schemas are now automatically transformed for OpenAI strict mode compatibility when using OpenAI models (including reasoning models like o1, o3, o4)

  • Optional field handling: .optional() fields are converted to .nullable() with a transform that converts null → undefined, preserving optional semantics while satisfying OpenAI's strict mode requirements

  • Preserves nullable fields: Intentionally .nullable() fields remain unchanged

  • Deep transformation: Handles .optional() fields at any nesting level (objects, arrays, unions, etc.)

  • JSON Schema objects: Not transformed, only Zod schemas

    Example

const agent = new Agent({
  name: 'data-extractor',
  model: { provider: 'openai', modelId: 'gpt-4o' },
  instructions: 'Extract user information',
});

	const schema = z.object({
  name: z.string(),
  age: z.number().optional(),
  deletedAt: z.date().nullable(),
});

	// Schema is automatically transformed for OpenAI compatibility
const result = await agent.generate('Extract: John, deleted yesterday', {
  structuredOutput: { schema },
});

	// Result: { name: 'John', age: undefined, deletedAt: null }

(#10454)

@mastra/upstash

  • Fix message sorting in listMessages when using semantic recall (include parameter). Messages are now always sorted by createdAt instead of storage order, ensuring correct chronological ordering of conversation history. (#10545)

@mastra/voice-deepgram

  • feat(voice-deepgram): add speaker diarization support for STT (#10536)

mastra

  • Rename "Playground" to "Studio" (#10443)
Check out latest releases or
releases around mastra-ai/mastra @mastra/core@0.24.6

Don't miss a new mastra release

NewReleases is sending notifications on new releases.

Get notifications