[BlueX888]

Problem

A function tool awaits an AgentTask. While it waits, update_agent() replaces the agent. When the task then completes, it unwinds on the tool task itself and the superseded-activity branch awaits old_activity.aclose() — whose tool-executor drain waits for that same non-cancellable tool task. The task awaits itself: the tool never returns, session.aclose() hangs in _wait_for_inactive, no close event is emitted, and teardown can hit RecursionError.

Reproduced on 049aea8; the new regression tests fail there (timeout in _wait_for_inactive with fut_done=True, inactive_ev=False, tool task non-cancellable; the tool-never-returns and late-unwind cases also time out).

Two adjacent patterns came out of review and are covered by this revision:

• a tool that awaits a second AgentTask after the first was superseded: the detached close holds the activity's lock that the second task's handoff needs, so task, tool and close all hang;
• a task that unwinds after session.aclose() returned: the close path had already walked past the superseded activity, which then leaked (tools/MCP toolsets never closed).

Fix

• _close_detached_activity() closes a superseded activity from a detached task instead of awaiting it inside the tool task. During a session close, the close path keeps ownership; _will_close_activity() checks whether the teardown walk actually reaches the activity and the detached close is used when it does not.
• A closed or detaching activity is never resumed again, and a second inline AgentTask on it is refused with ToolError — the same contract a tool already sees when the session is closing.
• The session keyterm detector records its owning activity: a stale detached close no longer disables detection for the activity that now owns it.
• drain()/aclose() are idempotent on an already closed activity.

Tests

tests/test_agent_task_close_race.py gains:

• test_aclose_after_update_agent_while_tool_awaits_agent_task — the original close-time hang
• test_agent_task_completion_after_update_agent_returns_to_tool — the tool that never returned after the task completed
• test_second_agent_task_in_tool_after_update_agent_does_not_deadlock — the two-task cycle
• test_late_agent_task_unwind_after_aclose_closes_superseded_activity — the late-unwind leak
• test_detached_close_keeps_keyterm_detection_on_replacement — the keyterm ownership

Receipts on 049aea8: the new close-race tests fail/hang for the reasons above. With this branch: tests/test_agent_task_close_race.py 14 passed (55+ reruns), agent-session + task suites 260 passed, full unit suite 6035 passed with the same 10 pre-existing failures / 10 environment errors as 049aea8; ruff check / ruff format --check clean. Merge-tree is clean against #8147, #7956, #7930, #7940, #7957 and current main.

Residual (theoretical): nested AgentTask chains racing update_agent() could still miss a close; no reachable case was constructed.

🤖 Generated with Claude Code