## Background Tool execution approval (e.g. for human-in-the-loop) requires complex logic for checking and tool execution. ## Summary - introduce `needsApproval` flag on `Tool` - introduce `ToolApprovalRequest` part on assistant model message - introduce `ToolApprovalResponse` part on tool model message - introduce `ToolApprovalRequestOutput` - add `execution-denied` tool output to language model v3 prompt specification - add tool approval logic to `generateText` and `streamText` - extract `executeToolCall` function - add `UIMessage` and UI message chunk support for tool approvals - add `addToolApprovalResponse` function to `Chat` - add `ChatAddToolApproveResponseFunction` type - update conversion and validation function for UI messages, model messages, and language model emssages Missing tool approvals are not handled in any special way, because some models may support tool calls without results. If the used model requires tool results for tool calls, the corresponding API will throw errors. ## Manual Verification - [x] run `examples/ai-core/src/generate-text/openai-tool-approval.ts` - [x] with positive approval responses - [x] with negative approval responses - [x] run `examples/ai-core/src/stream-text/openai-tool-approval.ts` - [x] with positive approval responses - [x] with negative approval responses - [x] run UI test `next-openai` http://localhost:3000/test-tool-approval - [x] approved with follow up message - [x] denied with follow up message - [x] multiple approvals in one message ## Future Work * discovered OpenAI responses bug: need to ensure tool result order matches tool call order * add tool execution approvals for dynamic tools * allow functions as values for the `Tool.needsApproval` property * needsApproval can return metadata that becomes part of the approval request * approval can have options * approval is available during tool execution (incl. options) * access initial tool execution step result similar to steps * different id generator overrides for different ids * UI sync when approvals happen server side (e.g. via email) * infinite client/server loop when there are errors * ai sdk stream versioning * user facing documentation @nicoalbanese ## Related Issues Closes vercel#8718 Closes vercel#8700

Tool Execution Approval (human-in-the-loop) #8718

Description

Metadata

Metadata

Assignees

Labels

Type

Fields

Projects

Milestone

Relationships

Development

Issue actions