Summary

Fixes a regression where subagent narration leaks into the parent agent's
final response (e.g. a streamed answer that begins with "I'll look up the
weather...", "I'll search for...", "Now let me..." before the real answer).

The leak appears after upgrading langgraph to >= 1.2 and is not caused
by the LLM client choice (ChatBedrockConverse vs ChatAnthropicBedrock) —
both stream identically. The trigger is a change in the shape of langgraph's
tasks-mode stream metadata that silently broke subagent namespace
correlation in the AG-UI encoder.

Root cause

AGUIStreamEncoder (via LangGraphStreamHelper._handle_tasks_chunk) builds a
{namespace -> tool_call_id} mapping from tasks-mode chunks so that subagent
events can be attributed to the task tool call that spawned them. Clients use
this attribution (emitted as NAMESPACE_CONTEXT events) to tell subagent output
apart from the parent agent's own response.

langgraph changed the shape of the tasks chunk input field between versions:

# langgraph < 1.2
input = {
    "__type": "tool_call_with_context",
    "tool_call": {"name": "task", "id": "toolu_...", "type": "tool_call", ...},
    "state": {...},
}

# langgraph >= 1.2
input = [
    {"name": "task", "id": "toolu_...", "type": "tool_call", ...}
]

_handle_tasks_chunk only read the dict shape (input.get("tool_call")). On
langgraph >= 1.2, input is a list, so the lookup returned {}, the mapping
stayed empty, and correlate_namespace took its fallback branch:

# Unknown namespace — treat as parent agent
return ()

Every subagent namespace was therefore classified as the parent. As a result:

1. No NAMESPACE_CONTEXT events were emitted (the namespace never "changed"
   from the parent's empty tuple).
2. Consumers that suppress subagent text based on namespace had nothing to key
   on, so subagent narration was rendered as the parent's answer.

This also produced a flood of [sse:correlate] Unknown namespace ... warnings
in the logs, which is the visible fingerprint of the bug.

Fix

Parse both input shapes via a new _extract_task_tool_calls helper and
iterate the resulting tool calls. The dict (tool_call_with_context) and list
forms both normalize to a list of tool-call dicts, so the mapping populates
correctly regardless of langgraph version. No protocol or consumer changes are
required — restoring the mapping makes the existing NAMESPACE_CONTEXT
attribution work again for all clients.

How it was verified

Reproduced end-to-end against a parent agent invoking parallel subagents,
streaming through the real AGUIStreamEncoder, comparing langgraph 1.0.x vs
1.2.x:

Tests

Adds TestTasksChunkNamespaceMapping covering:

• legacy dict shape populates the mapping
• list shape populates the mapping
• end-to-end correlation after a list-shape chunk
• the uncorrelated-fallback failure mode (guard)
• non-task tool calls are ignored
• result chunks without a tool call are ignored

All encoder/namespace tests pass; ruff clean.

Type of Change

• Bugfix