Pine/deepseek-harness
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases 1 Wiki Activity

feat(tasks): wake an idle owner when a background task completes

Browse Source 
Completion notices went through agent.inject(), which never reserves a
driver, so a task settling after its turn closed left the notice parked
until unrelated input woke the agent — while the same prompt told the
model not to poll for it.

An unreported completion now picks its lane from the owner's state: a
busy owner is injected as before, an idle owner is woken with
followup(). This adopts the delivery rule the subagent continuation
manager already ships. maxConsecutiveWakes bounds the self-exciting
chain and is reset by user-authored input; completionDelivery: quiet
restores the old lane for deterministic transcripts.

Teardown cancellation now claims the terminal report the way kill()
already does, so an owner being destroyed is never woken, and settle()
announces completion last so a reporter that opens a turn synchronously
sees a committed record.
This commit is contained in:
Yichen Jiang
2026-08-11 19:36:49 +08:00
parent 1d12ae62e7
commit d78b796a51
25 changed files with 531 additions and 96 deletions
Download Patch File Download Diff File Expand all files Collapse all files
packages/tasks/tasks-local/README.md
+1 -1
View File
@@ -10,7 +10,7 @@ Tasks belong to their owner and backend, not the producer tool fiber, so produce
Service disposal closes listeners, cancels all live tasks, awaits their records, and detaches effects from surviving owner scopes. If teardown cancellation throws, the service force-fails the record and warns that work may be orphaned instead of deadlocking. A cancellation that returns but never settles `done` remains indistinguishable from a slow stop and can stall teardown.
Settlement is first-wins: the earliest terminal outcome — producer settlement, a rejected `done` contained as `failed`, or a teardown force-failure — records once, notifies listeners once with per-listener containment, and releases waiters. Pending waits mark the task reported before listeners run so completion reporters do not duplicate notices.
Settlement is first-wins: the earliest terminal outcome — producer settlement, a rejected `done` contained as `failed`, or a teardown force-failure — records once, releases waiters, and notifies listeners once with per-listener containment. Pending waits mark the task reported before listeners run so completion reporters do not duplicate notices, and a teardown cancel marks it for the same reason: nothing will read a notice addressed to an owner being destroyed. Completion is the last thing a settlement announces, after the record is committed and the visible-set change is published, because a reporter may open a model turn synchronously and every other observer must already have seen the settled record.
Controllers and listeners are layered by the scope that registered them, in the tools-registry shape: a registration files into its registering context's scope, and a read unions the global layer with the owner's scope chain. One process-wide registry therefore answers per-owner questions per owner — `start()` refuses `background tasks unavailable: no task controller serves this agent (load @deepseek-ai/dsh-tool-tasks in its composition)` for an owner whose own composition attaches none, however many other compositions attach theirs, and a settlement reaches only the listeners its owner's composition registered.