Mastra Release - 2025-08-12

Agents

Agent documentation to include the onChunk callback and addresses a user query from issue #6480. #6627
A bug where the scorers endpoint returned agent names instead of IDs in the agentIds field, and also removes stray console logs and corrects typos in related tests. #6683
The ability to update the agent model. #6687
System memory message is always available during agent calls, including for new threads. #6709

A test suite to verify that Qdrant's metadata filtering for thread_id and resource_id works correctly in the Memory system, with all tests passing and no implementation changes required. #6607
A metadata filtering test suite for the Chroma vector store to verify compatibility with the Memory system, confirming that all shared tests pass without requiring code changes. #6645
TypeScript build errors correctly cause the build to fail by propagating tsc's exit code in the types-builder script. #6659
A broken documentation deployment by importing a missing component, resolving a build failure. #6684

The CLI's tool bundling logic to exclude test files by adjusting the glob pattern and function signatures, ensuring test files are not bundled and extending type compatibility. #6647
The playground CLI to correctly use user-specified host and port values instead of always defaulting to the standard settings. #6624
The temperature parameter to 1 for the gpt-5 model in the playground. #6663
Development environment detection by replacing the spoofable 'referer' header check with a secure 'MASTRA_DEV' environment variable, preventing authentication and authorization bypass in production. #6680

Cloudflare Worker compatibility by deferring PostgresStore initialization and adding Babel plugins to ensure proper Mastra instance scoping and initialization within Cloudflare Workers. #6086
An early warning in the deployer for invalid configuration by introducing a new Babel check in analyzeBundle, improving error detection before bundling fails. #6682
A set model handler to both the server deployer and the JavaScript client SDK. #6690

The docs-chatbot template into a monorepo containing separate agent and mcp-server applications. #6685
A new coding agent template that utilizes e2b sandboxes. #6649

The schema compatibility logic to avoid relying on the full language model. #6641
The MastraModelInput feature, though specific details of the changes are not provided in the description. #6642
The codebase to support v4 model input by introducing MastraInputStream, removing extraneous code, sharing ChunkType, and upgrading to aisdk v4 input. #6658

Schema compatibility issues with certain model providers by ensuring memory tools use makeCoreTool, resolving errors with OpenAI reasoning models and addressing issue #6548. #6628
The working memory tool to use the user-provided schema directly, improving LLM compatibility and tool invocation clarity. #6640

The vnext agent network stream to support accepting an array of messages. #6617
[TIER-2] The A2A implementation to spec version 0.3.0 across multiple packages, renames key methods, adopts official a2a-js SDK types, and fixes a bug with in-memory task storage, without adding new functionality. #6441

A bug by asynchronously capturing telemetry data on stream completion via an intercepted onFinish callback, ensuring immediate stream returns and preserving existing behaviors for both agent.stream() and agent.streamVNext(). #6265

[IMPORTANT] A new 'semantic-markdown' chunking strategy to the RAG package, enabling more context-aware and size-controlled splitting of markdown documents by intelligently merging related header sections, along with documentation, refactoring, performance improvements, and tests. #6595

[TIER-2] Support for filtering scorer scores by source in storage, fixes libSQL parsing and scorer implementation issues for storage providers, and updates tests for compatible storage providers. #6670

A bug where tool calls would incorrectly appear after messages when refreshing the agent chat page. #6637
That tool arguments are correctly preserved in tool invocations when tool-result messages arrive separately or when messages are restored from the database, preventing loss of argument data in these scenarios. #6655
Tool name formatting by replacing invalid characters, prefixing names that start with invalid characters, and truncating long names to ensure compatibility with LLM provider requirements such as Gemini and OpenAI. #6612
Tool input validation that returns structured error objects instead of throwing exceptions, enabling agents and MCP servers to handle and correct validation failures gracefully. #6466
A handler that returns the list of providers for which the user has keys in their environment. #6705