_get_current_turn_contents picks wrong anchor when single_turn agents run in parallel #7488

#5961 was closed as stale after a repro was requested but never provided. This is that repro.

The bug

When two single_turn sub-agents run on the same session in parallel, one ends up with the other's output in its context window. The reverse scan in _get_current_turn_contents starts at len(events) - 1 — the tail of the shared session list. If agent A finishes first and appends, agent B's scan starts at A's event, skips it on a branch mismatch, and anchors on the parent event directly above. _get_contents is then called with events[anchor:], which includes A's output. ADK wraps it in a safety block, but the content is still there.

Confirmed on v2.11.0. PR #6047 (v2.11.0) fixed include_contents field-override — a different code path. This scan is untouched.

To reproduce

pip install google-adk
python3 parallel_agent_context_contamination.py

No API key. No model. No LiteLLM.

Output (see poc_output.txt):

[FAIL] Unconditional tail-anchored reverse scan found:
  >>>  41|   for i in range(len(events) - 1, -1, -1):

[FAIL] sub_agent_b's context contains sub_agent_a's output!

Scan log at the moment sub_agent_b calls the function while sub_agent_a has already appended:

SCAN   sub_agent_a   root.a   no
SCAN   parent        root     YES   ← anchor — but events[anchor:] includes sub_agent_a's event
       ─── BUILD phase ───
BUILD  parent        root     YES
BUILD  sub_agent_a   root.a   no    ← safety-wrapped into sub_agent_b's context

Fix

Snapshot len(session.events) when each sub-agent's InvocationContext is created and use that as the scan start index. Scanning from there instead of len(events)-1 keeps the anchor search bounded to events that existed before the sibling started appending.

adk_parallel_context_bug.zip