AI SDK 7.0.163 honors telemetry content controls and preserves Google streamed tool order and signature history
Vercel released AI SDK 7.0.163 on November 5. `recordInputs` and `recordOutputs` now govern content delivered to integrations and the `ai:telemetry` channel, filtering prompts, tool inputs, response text, and raw span results while retaining application results such as usage. The same release preserves the order of text, complete function calls, and thought signatures in streamed Google responses and across replayed history.
Why it matters
Generative UI often sends stream events, tool calls, and telemetry into both visible timelines and observability systems. If disabling content recording affects only one tracing path, sensitive prompts can leak through another; if tool order or signature history changes, the UI timeline and the next model turn can become inconsistent.
A closer look
New dispatcher tests clear instructions, messages, state, questions, and tool arguments when inputs are disabled, and clear content, text, and response messages while retaining token usage when outputs are disabled. The application event object remains unchanged, and raw span results are withheld whenever either content-recording flag is off. Google fixes add tests for streamed tool order, empty-text metadata, and Anthropic/Bedrock history conversion, preserving standalone thought signatures while omitting foreign-provider empty signature text.
Who this is for
Platform, security, and generative-UI teams using AI SDK streamed tools, Google/Vertex reasoning history, OpenTelemetry, or custom telemetry integrations.
Original sources
- Vercel AI SDK official release and linked commits ↗github.com
- Telemetry recording flags fix ↗github.com
- Google streamed tool ordering and signature history fix ↗github.com
Revisions & corrections (1)
— The official release fixes two evidence-complete integrity boundaries for streamed generative UI: telemetry content controls and provider-order/history preservation.