fix: scope single-turn node inputs to workflow branch

[he-yufeng]

Summary

Fixes #5989.

prepare_llm_agent_input() appends the single-turn node input directly to
session.events, so it bypasses NodeRunner._enrich_event(). In a dynamic
workflow sub-branch, that left the private node input as a branchless user event.
Branchless user events are visible to sibling branches, which can make another
worker pick the wrong "current turn" boundary and slice out its own function
call.

This stamps the single-turn input event with the current invocation branch when
one exists, while preserving the existing isolation-scope stamping.

To verify

• PYTHONPATH=src python -m pytest tests\unittests\workflow\test_llm_agent_as_node.py -q -k "single_turn_input_event_inherits_branch_and_scope"
• PYTHONPATH=src python -m pytest tests\unittests\workflow\test_llm_agent_as_node.py -q
• python -m py_compile src\google\adk\workflow\_llm_agent_wrapper.py tests\unittests\workflow\test_llm_agent_as_node.py
• python -m pyink --check src\google\adk\workflow\_llm_agent_wrapper.py tests\unittests\workflow\test_llm_agent_as_node.py
• python -m isort --check-only src\google\adk\workflow\_llm_agent_wrapper.py tests\unittests\workflow\test_llm_agent_as_node.py
• git diff --check

[adk-bot]

Response from ADK Triaging Agent

Hello @he-yufeng, thank you for creating this PR!

While reviewing your PR, we noticed it is mostly following the contribution guidelines. However, to help speed up the review process, could you please provide:

• A summary of the passed pytest results.
• Logs or screenshots after the fix is applied (or the output showing the issue resolved) to help reviewers better understand the fix.

You can find more details in our contribution guidelines.

Thank you!

[adk-bot]

🔍 ADK Pull Request Analysis: PR #6008

Title: fix: scope single-turn node inputs to workflow branch
Author: @he-yufeng
Status: OPEN
Impact: 27 additions, 2 deletions across 2 files

Executive Summary

1. Core Objective: Prevents the single-turn agent node-input user-role event from being treated as globally visible across concurrent execution branches, avoiding subsequent "turn-boundary" truncation errors.
2. Justification & Value: [Justified Fix] - A highly valuable fix addressing a severe intermittent concurrency bug (ValueError: No function call event found for function responses ids) when executing parallel tool-using agents as dynamic workflows.
3. Alignment with Principles: [Pass] - Follows all modern python conventions (annotations, standard types), keeps boundaries clean without breaking public API compatibility, and provides clean test coverage using the standard Arrange-Act-Assert pattern.
4. Recommendation: Approve

Detailed Findings & Analysis

1. Objectives & Impact ("What does it do?")

• Context & Background: Resolves linked Issue #5989. When executing concurrent tool-using LlmAgents via a dynamic sub-branch orchestration flow (use_sub_branch=True), the orchestrator concurrently executes node iterations. Because the input user event for single-turn mode was previously appended without a branch property, it was globally matched as belonging to every branch. During multi-round tool loops, a sibling's input event could interleave into another worker's timeline, acting as an erroneous "latest turn boundary" and slicing out that worker's own function_call event context, raising a ValueError.
• Implementation Mechanism:
  In _llm_agent_wrapper.py, during the assembly of the user_event, the branch is now dynamically stamped from the active invocation context if it exists:

  branch = getattr(ctx._invocation_context, 'branch', None)
  if branch:
    user_event.branch = branch

  This matches the behavior of NodeRunner._enrich_event() and ensures that the local agent input is correctly isolated.
• Affected Surface: It safely restricts scope purely to the private/internal _llm_agent_wrapper.py logic, ensuring zero changes to any public-facing API signatures.

2. Justification & Value ("Is it a valid and useful change?")

• Workspace Verification:
  • Investigating _llm_agent_wrapper.py verifies that prior to the PR, only containment (via isolation_scope) was stamped, while the branch remained completely absent on the user-role event.
  • Investigating _is_event_belongs_to_branch() confirms that not event.branch forces any event to belong to all branches:

    if not invocation_branch or not event.branch:
      return True

    This confirms the issue description: a branch-less user event erroneously scopes to concurrent sibling nodes.
• Value Assessment: Highly critical. Without this, multi-agent frameworks using parallel tasks under the same session would consistently crash under heavy concurrency and staggered tool loops.
• Alternative Approaches: The developer chose the most elegant design—applying the existing branch context to the event rather than patching the branch-matching consumer, keeping the "blast radius" completely localized.
• Scope & Depth: [Systematic Fix] & [Root Cause] — Rather than applying a bandaid workaround at the parser level, this fixes the root cause: the node-input event was lacking the identifier for the branch it was spawned in.

3. Principle & Style Alignment Checklist ("Does it follow rules?")

• Public API & Visibility Boundaries:
  • Status: Pass
  • Analysis: No breaking changes. The modified function prepare_llm_agent_input remains fully internal to the private framework workflow orchestration.
• Code Quality, Typing & Conventions:
  • Status: Pass
  • Analysis: Complies with from __future__ import annotations. Uses standard getattr properties for defensive retrieval, leaving the core implementation clean, typed, and style-compliant. No imports from __init__.py.
• Robustness & Edge Cases:
  • Status: Pass
  • Analysis: Defensive checks using getattr(..., None) handle the edge case where the context is run outside of an active workflow sub-branch, maintaining backward compatibility seamlessly.
• Test Integrity & Quality:
  • Status: Pass
  • Analysis: The new test test_single_turn_input_event_inherits_branch_and_scope is beautifully integrated into test_llm_agent_as_node.py. It avoids mocks and uses real ADK invocation wrappers, cleanly formatted using the Arrange-Act-Assert pattern.

---

Summary of Work

• Retrieved and analyzed PR fix: scope single-turn node inputs to workflow branch #6008 detail and diff.
• Verified CLA signature successfully (status SUCCESS).
• Fetched and parsed Issue Concurrent tool-using LlmAgents run via ctx.run_node crash with "No function call event found for function responses ids" (single_turn appends branchless/scopeless user event to shared session.events) #5989 context to map root-cause analysis precisely against the baseline codebase.
• Inspected files _llm_agent_wrapper.py, test_llm_agent_as_node.py, and contents.py using view_file to confirm event belonging logic.
• Drafted and presented the structured analysis report indicating approve and highlighting the underlying root cause alignment.