FrameworksOfficial

AI SDK 7.0.157 stabilizes tool filtering, while tool visibility still is not execution authority

Released October 29, AI SDK 7.0.157 graduates `filterActiveTools` to a stable export and makes `LanguageModelStreamPart` a stable type while retaining deprecated aliases. It also fixes Node API warnings caused by telemetry runtime detection in the Next.js Edge Runtime.

Event 2026-10-29Source published 2026-10-29Indexed 2026-10-30Verified 2026-10-29

Why it matters

Generative interfaces and agents often connect large sets of search, data, action and rendering tools. Narrowing the model-visible set per step can remove irrelevant schemas, reduce selection ambiguity and preserve type inference. Treating 'not shown to the model' as 'not allowed to execute', however, turns a prompt optimization into a false authorization boundary.

A closer look

Stable `filterActiveTools` takes a full tool registry and an activeTools list and returns a typed subset. `experimental_filterActiveTools` remains as a deprecated alias for gradual migration. Stable `LanguageModelStreamPart` unifies model-stream event typing, with `Experimental_LanguageModelStreamPart` retained until v8; the custom-loop function `experimental_streamLanguageModelCall` itself remains experimental. The Edge fix changes `process.release` detection to optional `globalThis.process` access, and the maintainer verified that warnings disappeared in a Next.js end-to-end example. An updated gateway dependency also preserves caller cancellations instead of fabricating retryable HTTP 500 errors, but that is a dependency change rather than an AI core API change.

THE LIMITS OF THE EVIDENCE

Evidence is the maintainer release, commits and open design issue; the behavior was not independently run here. Graduation does not announce new runtime authorization enforcement, dynamic tool registration or cache semantics. The still-open repository issue #18021 says request-level `activeTools` changes can invalidate prompt caches; inactive tools remain in the full runtime registry, prior reports show they may still execute, and error paths can expose the full registry. The issue proposes separating registry, model availability, history changes and execution enforcement and had asked to keep the helper experimental until that design was resolved, creating tension with the 7.0.157 graduation. Teams must continue to enforce permissions, approvals and argument validation in executors and business logic rather than treating `filterActiveTools` as a security sandbox. The release provides no token, cache-hit, tool-selection or end-to-end latency benchmark. Deprecated aliases remain, so existing code does not require an immediate rewrite.

Who this is for

TypeScript teams using AI SDK for multi-tool agents, generative UI, MCP catalogs, per-step tool selection, Edge deployment and authorization

Original sources

Revisions & corrections (2)

2026-10-30 — Publish the verified 7.0.157 release. The open design issue is a documented limitation and counterexample, not a date or identity conflict.

2026-10-30 — Publish a verified framework boundary change and preserve the official open issue showing that model-visible tool filtering is not an authorization control.

AI SDK 7.0.157 stabilizes tool filtering, while tool visibility still is not execution authority

Why it matters

A closer look

Who this is for

Original sources

Related events

assistant-ui's October 28 release tightens prompt review, tool-result ownership and configuration writes

AI SDK 7.0.155 removes runtime checks already guaranteed by static types

MCP Apps 2.0.4 trims approval-card sprawl and repairs example state and dependencies