FrameworksOfficial

assistant-ui Generative Frame 0.0.10 adds host confirmation and render budgets, while fixing pre-approval frontend execution

assistant-ui released generative-frame 0.0.10 and assistant-stream 0.3.56 on November 8. Catalog actions marked `confirm: true` now require a host `confirmAction`; rendered specs are capped at 5,000 elements and 64 levels; and the paired stream fixes preserve final AI SDK tool results after approval while preventing a same-named frontend tool from running before server approval completes.

Why it matters

The boundary for declarative generative UI is not only what a model may render. It also includes when generated interactions may trigger host effects, how much rendering work a malformed structure may consume, and which post-approval result enters the visible interface and later model history. This release moves those boundaries into the catalog, renderer, and stream decoder.

A closer look

A catalog can mark a sensitive action `confirm: true`, and the React/Vue SpecRenderer and createSpecToolkit accept a host-owned `confirmAction`. Declining stops the action list, a protected action fails when the callback is missing, and `watch` cannot trigger it. Rendering resolves at most 5,000 elements and 64 levels: a streaming spec retains its last tree within budget, while a completed oversized spec shows an error placeholder. assistant-stream 0.3.56 fills later `tool-output-available` or error parts into the original approved tool call, replaces preliminary model-facing content with the final result, and skips pending frontend tools that carry server approval. The release also adds BCP 47 locale formatting for numbers, currencies, and percentages.

Who this is for

Frontend, platform, and security teams using assistant-ui, Generative Frame, declarative UI catalogs, AI SDK tool approvals, or React/Vue agent interfaces.

Original sources

Revisions & corrections (1)

— One coordinated immutable release group adds host-owned action confirmation, deterministic render budgets, and approval-result ordering to declarative generative UI; the official release and linked commits provide exact dates, behavior, tests, and explicit limits.

Related events

Frameworks

AI SDK 7.0.165 sanitizes request URLs before telemetry and closes a transcription word-timing leak

Vercel released AI SDK 7.0.165 on November 7. The patch moves provider-request URL sanitization inside generation and streaming execution callbacks so telemetry integrations and Node tracing subscribers no longer observe raw paths, query parameters, or endpoint credentials first. It also removes transcription words from telemetry when recordOutputs=false while preserving the full result returned to the caller.

Frameworks

AI SDK Harness 1.0.175 fixes host-tool approval deadlocks and aborts turns after permanent host disconnects

Vercel released @ai-sdk/harness 1.0.175 and workflow-harness 1.0.175 with AI SDK 7.0.164 on November 6. Harness fixes a deadlock when a host tool pauses for approval and aborts a running turn when the host is considered disconnected for good; Workflow Harness receives the same behavior through its dependency update.

Frameworks

AG-UI Kotlin 0.5.0 aligns with 1.0 through interrupts, activities, multimodality, and strict state boundaries

AG-UI released Kotlin core, client, and tools 0.5.0 on November 5. The SDK now targets AG-UI 1.0 with REASONING events, interrupt/resume outcomes, ACTIVITY snapshots and deltas, tool errors, and multimodal user messages, while adding browser Wasm and macOS Arm64 targets and tightening event, input, and RFC 6902 state validation.