AI SDK 7.0.162 restores server tool context on approval continuation and adds one durable agent trace
Vercel's AI SDK 7.0.162 fixes approval continuation reconstruction so server-defined toolMetadata, dynamic, and title are preserved for policy evaluation. The same release adds a single telemetry operation for durable WorkflowAgent runs, collecting prompts, final output, steps, and tools and reconstructing the invoke_agent tree after completion or suspension.
Why it matters
If approval continuation changes the tool context seen by policy, the same sensitive action can receive different decisions before and after resume. Durable workflows are also difficult to reconstruct when evidence is split across live model spans and pause boundaries.
A closer look
The approval fix rebuilds toolMetadata, dynamic, and title from the current server tool definition, preventing history from overriding those fields. The official commit reproduces the issue with a local MockLanguageModelV4 end-to-end script: continuation fields were null before the fix and matched the initial call after it; focused and full test suites passed. Durable telemetry creates one operation for the WorkflowAgent run, including original timestamps, prompt, final output, steps, and tools, with parent-trace and sampling bridging. Live nested model and provider spans remain, while an invoke_agent tree can be reconstructed on completion or suspension.
Who this is for
Platform and application teams using AI SDK tool approval, durable workflows, nested agents, and OpenTelemetry.
Original sources
- Vercel AI SDK official release and linked commits ↗github.com
- Preserve server tool metadata in approval continuations ↗github.com
- Durable WorkflowAgent telemetry ↗github.com
Revisions & corrections (1)
— Publish the evidence-complete official AI SDK 7.0.162 approval-context and durable-telemetry event.