fix(py): record subagent end_time at SubagentStop, not at conversation end

[James Won (jwon)]

Problem

When a session spawns multiple subagents in parallel (e.g. the assistant fires several Agent tool calls in one turn), the resulting subagent chain runs are recorded with end_time equal to conversation termination, not actual subagent completion. In trace UIs this manifests as sibling subagent chain spans visually outlasting their parent Agent tool spans — sometimes by many minutes — and all closing at the same instant.

Root cause

subagent_stop_hook (python/langsmith/integrations/claude_agent_sdk/_hooks.py) intentionally defers both .end() and .patch() until clear_active_tool_runs() runs, so that PostToolUse for the parent Agent tool can attach outputs to the subagent run first.

Deferring .patch() is required (outputs aren't known yet at SubagentStop). Deferring .end() is not — it just records end_time. Because both were deferred together, RunTree.end_time was set to datetime.now() inside clear_active_tool_runs(), which only runs once at the end of receive_response().

This deferred-end behavior dates back to #2670, which introduced subagent tracing — .end() and .patch() were paired in the same code path, but only .patch() actually needs to wait.

Fix

Record end_time immediately when SubagentStop fires by calling subagent_run.end() in the hook. .patch() remains deferred until clear_active_tool_runs() so outputs added later by PostToolUse are still flushed. clear_active_tool_runs() no longer re-calls .end() for already-ended subagents (which would overwrite the correct timestamp with now()).

Orphan paths (SubagentStop never fired, or no matching Agent tool) are unchanged.

Testing

• Updated test_subagent_stop_and_post_tool_use_set_outputs — previously asserted end_time is None after SubagentStop / PostToolUse (locking in the deferred behavior). Now asserts end_time is set at SubagentStop and is unchanged by subsequent PostToolUse and clear_active_tool_runs calls.
• Added test_subagent_end_time_recorded_at_stop_not_conversation_end — regression test with synthetic delays between hooks, asserting the recorded end_time matches subagent stop time, not cleanup time.
• make format && make lint && make tests pass locally (Python SDK only).