feat: add ToolRuntime injection for interruptible tools

[saksharthakkar]

Summary

Enables escalation tools (and future interruptible tools) to access runtime.tool_call_id for building proper ToolMessage responses.

Problem

Escalation tools fail with:

TypeError: escalation_tool_fn() missing 1 required positional argument: 'runtime'

The escalation tool needs tool_call_id when returning a Command with ToolMessage:

async def escalation_tool_fn(runtime: ToolRuntime, **kwargs):
    # ... interrupt and resume logic ...
    if outcome == "end":
        return Command(
            update={"messages": [
                ToolMessage(
                    content="...",
                    tool_call_id=runtime.tool_call_id,  # ← needs this
                )
            ]},
            goto=TERMINATE,
        )

Without tool_call_id, LangGraph cannot match the response to the original LLM tool call → broken state.

Solution

• Detect tools with runtime: ToolRuntime parameter via type hint inspection
• Inject ToolRuntime with tool_call_id, state, and config when detected
• Skip injection for regular tools (no signature change = no breakage)

Related

• Companion to UiPath/uipath-agents-python#65 (trace context preservation for interruptible tools)

Test plan

• Unit tests for runtime detection (True for tools with runtime param, False otherwise)
• Unit tests for async/sync tool execution with runtime injection
• Unit tests for Command return with correct tool_call_id
• All 357 existing tests pass

🤖 Generated with Claude Code

[andreitava-uip]

I'm really not a fan of introducing such hacks to "solve" circular imports.
If we have circular imports we should just restructure our modules in order to avoid them.

[andreitava-uip]

Personally I would take an entirely different approach here.
I'd say tools themselves should never return messages or commands. A tool should only ever return its output schema.
Keep them completely graph-agnostic.
In the custom tool node design, the idea is for the wrappers to return commands if needed.

If anything, escalation tool should be refactored to no longer be coupled to the graph, which is something that should be in the works already.