Repository navigation
notify_observers drains a process-global task set, so one loop's on_loop_end blocks on another loop's passive hooks #2
Description
agent_core/loop_types.py keeps fire-and-forget observer tasks in a module-global set:
_background_tasks: set[asyncio.Task[None]] = set()
...
if method == "on_loop_end":
await drain_background_observers() # drains every loop's tasksdrain_background_observers() gathers all pending tasks, not the ones this loop created. With two agent loops in one process — the ordinary sub-agent case — loop A's on_loop_end waits on loop B's still-running passive hooks.
Observed: a passive on_turn_end sleeping 3s in one logical loop made an unrelated loop's on_loop_end return 3.0s late. A passive hook awaiting something that is never signalled turns this into an indefinite hang of an unrelated loop's shutdown.
Fix direction: a per-loop task registry, or scope the drain to the tasks this loop scheduled. Note the drain exists for a real reason (a caller reading the event sink right after run() would otherwise race pending writes), so it must stay — just be scoped.
Provenance: pre-existing in ApodexHarness/miroharness/core/loop_types.py, ported verbatim by #1 — not introduced by the refactor. Fixing it here is the source-of-truth fix; ApodexHarness carries the same code and needs the same change until it consumes AgentCore.
Activity
fix: harden runtime observer contracts
zhanghanduo commentedon Sep 3, 2026
Fixed on
mainby 63dba6f ("fix: harden runtime observer contracts"). The commit's trailer was written with a literal\n\nbeforeRefs #2, #3, #4, #5, #6., so GitHub never linked it — closing by hand.The drain is now loop-scoped:
_background_tasks_by_owner: dict[asyncio.Task, set[asyncio.Task]]keys passive hooks by the owning agent-loop task (agent_core/loop_types.py:485),notify_observersregisters each task underasyncio.current_task()(:574), anddrain_background_observersgathers only that owner's pending tasks (:583). The drain itself is preserved, as the issue required — one loop'son_loop_endno longer waits on another loop's hooks.Downstream:
miroharness/core/loop_types.py:653still carries the process-global_background_tasksset. Tracking that backport separately.zhanghanduo commentedon Sep 3, 2026
Correction to the "Downstream" note above.
That note is accurate for MiroHarness
main, which still carries the standalone 790-line copy — but it is already stale for the integration branch. Onrefactor/agent-core,miroharness/core/loop_types.pyis a 75-line compatibility facade that re-exportsagent_core.loop_types(done by1bedcc52 refactor: consume shared runtime contracts), and that branch pins:git merge-base --is-ancestor 63dba6f a9b5272confirms the pin already contains this fix.So no backport is needed. The downstream copy is not being patched in parallel; it is being deleted by the extraction. MiroHarness
mainkeeps the old code only untilrefactor/agent-corelands, and this fix arrives there through the pin rather than through a second edit.The only product-specific symbol that stays behind in the facade is
wall_deadline_remaining_s, which bindsmiroharness.core.execution_contextand is registered as an allowed delta inscripts/contract/manifest.py.