September 10, 2026
Mastra's September 10, 2026 release introduces worker provisioning manifests for Mastra Cloud, expanded trace querying, idempotent deletion for observability signals, memory transform hooks, and streaming reliability improvements including a first-chunk timeout.
原始内容为英文；当前页面提供中文导航与来源说明，具体事实请以原文为准。
为什么值得关注
The September 10, 2026 release of Mastra updates @mastra/core to 1.66.0 alongside changes across server, deployer, client, and storage packages. Highlights include: - Worker Provisioning: @mastra/deployer emits a static workers.json manifest when shared storage and PubSub are configured, enabling dedicated worker provisioning, while Mastra.getWorkerConfig() provides live topology data via authenticated endpoints. - Advanced Trace Queries: Enriched predicates for same-span fields, portable metadata, feedback, and scores supported across @mastra/server, @mastra/pg, @mastra/clickhouse, and @mastra/duckdb. Mastra-recorded spans now feature typed input and output payloads with helper functions like isSpanRecordOfType, describeSpanInput, describeSpanOutput, and describeSpanError. - Observability Deletion: Added idempotent batch deleteFeedback() and deleteScores() in core storage, exposed via DELETE /api/observability/ routes in @mastra/server and client methods in @mastra/client-js. - Memory Transform Hooks: @mastra/memory introduces async hooks (beforeObservation, afterObservation, beforeReflection, afterReflection) to filter, redact, or reshape observations and reflections. - Streaming Improvements: Added modelSettings.timeout.firstChunkMs to fail fast (MastraTimeoutError) when a streaming model delays producing content chunks, triggering configured fallbacks. Additionally, Agent.stream() tracing overhead was reduced by avoiding observability context rebuilds on every chunk.
可执行摘要
Mastra core 1.66.0 and related packages add deploy-scoped worker manifests, advanced trace querying (filtering by spans, metadata, feedback, and scores) across ClickHouse/DuckDB/PG, batch deletion for feedback and scores, observational memory hooks (beforeObservation, afterObservation, beforeReflection, afterReflection), and modelSettings.timeout.firstChunkMs.
- Agent 实用度
- 82/100
- 可信度
- 96%
- 机器格式
- JSON + Markdown
开发者应核对什么
- Adopt modelSettings.timeout.firstChunkMs in streaming configurations to prevent hangs on slow initial token generation.
- Implement deleteFeedback() and deleteScores() in custom storage adapters to prevent throwing unhandled not-implemented errors.
- Utilize isSpanRecordOfType, describeSpanInput, and describeSpanOutput in telemetry or UI code to leverage typed span payloads.
- Integrate beforeObservation/afterObservation/beforeReflection/afterReflection hooks in @mastra/memory pipelines if data redaction or payload reshaping is required prior to persistence.