runner: a function response to a sub-agent's long-running call fails the run when the root is a workflow agent #1807

Describe the Bug:

With a ParallelAgent or SequentialAgent as the root, resuming a sub-agent's long-running tool call fails the whole run:

failed to run sub-agent "other": no function call event found for function responses ids: map[call-1:{}]

When the root isn't an LlmAgent, Run always runs the root again (runner.go#L687-L716). It never checks which agent the function response belongs to, so agents that never saw the call get it:

• ParallelAgent{caller, other}: other runs on branch root.other. It can't see caller's function call (different branch), but it does see the user's response (empty branch), so its contents processor errors.
• SequentialAgent{other, caller}: other re-runs first and fails the same way.

The LlmAgent-root path already handles this with findAgentToRun, which picks the agent that made the call.

Steps to Reproduce:

1. ParallelAgent root with two llmagent sub-agents, one with a long-running tool
2. First turn: the tool's agent returns a function call (call-1), the tool returns nil, so the call stays pending
3. Second turn: send a FunctionResponse with ID call-1

Expected Behavior:
The sub-agent that made the call resumes on its own branch and answers. The other sub-agents don't re-run.

Observed Behavior:
The run fails with the error above, so the long-running call can never be completed.

Environment Details:

• ADK Library Version: main (987c7bc)
• OS: macOS
• Go Version: 1.26.6

Additional Context:
adk-python sets invocation_context.agent = self._find_agent_to_run(...) for every root and then restores that agent's branch with _restore_branch_from_history (runners.py, _setup_context_for_new_invocation). It recently fixed the non-resumable case of this in "fix(runners): resolve invocation_id from FunctionResponse when is_resumable is False". I have a fix that does the same in the non-LlmAgent path of Run, with a test for Parallel and Sequential, and will open a PR.

One thing to check: with SequentialAgent{caller, other} the current code happens to work, because it re-runs caller and then also runs other. With the fix only caller resumes, which matches adk-python. Is that OK, or should the sequence continue after the resumed agent?

How often has this issue occurred?: Always