fix: escalation tool missing runtime parameter

[saksharthakkar]

Summary

Fixes escalation tool that was broken due to missing tool_call_id when returning Command.

Problem

• Escalation tool required runtime: ToolRuntime parameter to get tool_call_id
• Nothing was injecting ToolRuntime into the tool
• Result: TypeError: escalation_tool_fn() missing 1 required positional argument: 'runtime'

Solution

Use wrapper pattern consistent with other tools:

1. Tool returns dict (like process_tool, context_tool, etc.)

   • escalation_tool_fn returns {"action": ..., "output": ..., "escalation_action": ...}
   • Graph-agnostic, no knowledge of Commands or Messages

2. Wrapper handles graph integration

   • escalation_wrapper receives (tool, call, state) where call["id"] is the tool_call_id
   • Converts result to Command when termination needed
   • Uses ToolWrapperMixin to attach wrapper to tool

3. Simplified tool_node.py

   • Removed ToolRuntime injection logic that was never fully implemented
   • Clean separation: tools return data, wrappers handle graph concerns

Why escalation is different from other tools

• Other tools (process, context, integration) always continue - just return data
• Escalation can END the graph based on user action - needs to return Command
• Wrapper pattern elegantly handles this: tool returns data, wrapper decides Command vs dict

Changes

• escalation_tool.py: Returns dict, wrapper converts to Command
• tool_node.py: Removed unused ToolRuntime injection code
• tool_factory.py: Fixed return type
• test_tool_node.py: Cleaned up imports
• pyproject.toml: Version bump to 0.2.5

Test plan

• pytest tests/agent/tools/test_tool_node.py - all pass
• ruff check . - no errors
• mypy - no new errors

[Chibionos]

Approving with a design comment, @saksharthakkar can't the StructuredTool be improved ? rather than creating a new class ?

Addressed feedback: refactored to wrapper pattern - tool now returns dict (graph-agnostic), wrapper handles Command.