Add asynchronous tool support with agent pause/resume

[weeco]

Summary

This PR adds support for asynchronous tools that require external completion (e.g., user input, long-running deployments). When an asynchronous tool is executed, the agent pauses with FinishReasonInputRequired and can be resumed later via Runner.Resume() with the final tool results.

Key Changes

• Tool Interface Enhancement: Added IsAsynchronous() method to the Tool interface to declare whether a tool requires external completion

• Agent Pause/Resume Logic:

  • Modified LLMAgent.executeSingleTurn() to detect asynchronous tools and pause execution with FinishReasonInputRequired
  • Tracks tool call IDs requiring external input and includes them in InvocationEndEvent.InputRequiredToolIDs
  • Emits StatusStageInputRequired event when pausing for external input

• Runner Resume Support:

  • Added Runner.Resume() method to continue paused invocations by providing tool results
  • Extracted shared agent execution logic into runAgent() helper used by both Run() and Resume()
  • Tool results are added as a user message with ToolResponse parts before resuming

• Tool Registry Enhancement: Registry now appends a note to asynchronous tool descriptions instructing the LLM not to re-invoke them after a pending status

• MCP Elicitation Support: Added ElicitationHandler and WithElicitationHandler() to support MCP server requests for user input during tool execution

• Built-in Tool Updates:

  • RequireInputTool now declares itself as asynchronous
  • All existing tools implement IsAsynchronous() returning false

• Comprehensive Tests: Added TestRun_AsyncToolPausesExecution and TestRun_MixedSyncAndAsyncTools to verify pause/resume behavior with mixed tool types

Implementation Details

• Asynchronous tool detection happens during concurrent tool execution in executeTools()
• The executeToolsResult struct now tracks both response parts and IDs of tools requiring external input
• Session persistence occurs after each assistant message and on exit, ensuring consistency across pause/resume cycles
• The LLM receives full context: original tool call, initial pending result, and final result provided via Resume()

https://claude.ai/code/session_01WFwbKjni1JS6Rf17VNySb1

[blacksmith-sh]

Found 1 test failure on Blacksmith runners:

Failure

Test	View Logs
github.com/redpanda-data/ai-sdk-go/tool/TestRegistry_WebfetchToolWithLLM_Integration	View Logs

[weeco]

This was a test using the new ultraplan feature. Still have to review and check if I'd be happy with that API, I only planned and let it go create the PR without loooking into it