Scheduler Decisions
Why delegated tasks use the scheduler’s durable jobs, conversation scope, and cancellation rules.
Why is a delegated task a scheduler job?
Because Link Scheduler already keeps jobs durably, fires them through
an agent's way in, counts them against its caps and carries them
through a move. A subagent, a message to a teammate and the steps after
them are each a job, and one that ended is kept in the store contract's
ended place with its outcome
(Where the Jobs Are Kept).
A separate task service would need its own store, caps and moves, and a
team needs no other.
Who says what a task waits on?
The scheduler, outside the model. A task's coordinator is the
conversation the call came from, never a field the model writes, and
after takes only that conversation's own delegations. So a model
can't attach its work to another conversation's, wait on what it never
started, or slip past the hop, daily and subagent caps a team's members
already share.
What does cancelling a task cancel?
Its children and what waits on it, and in turn theirs, under the jobs' lock. A step whose input failed or was cancelled is cancelled rather than run on what never came, so a coordinator never pays for work whose plan already broke. Without a retained host, a running subagent finishes its turn and its answer is discarded. A retained subagent already admitted by its host is left running: direct cancellation refuses it, and a cascade leaves it alone.
Why does a recurring schedule ask first?
A recurring schedule acts again and again with no one watching, so
cron__create asks you unless it says recurring = false; a one-off
reminder doesn't. The owner sees every conversation's schedules with
lnk agent task list --all.
How does an owner let recurring schedules through?
With the agent's spec, [tool.scheduler] recurring = "allow". An
agent with no one to answer, in the background or on a box, would have
every recurring schedule refused, and --allow can't help: an argument
that asks (--ask-if) asks even of a call --allow names. The setting
takes the --ask-if away for that agent alone, and leaves ask and
every other ask as they are.
Why is the recurring setting the spec's?
The spec is a control file the agent never writes, so a model can't let
its own schedules through, and its default, ask, keeps the reason
above. A refusal names the setting,
carried to the sandbox as the hint of what lets the call through
(--ask-if-hint), so the sandbox stays unaware of the scheduler.
Why is there a shortest interval?
An every-minute cron would spend a day's messages in under two hours,
so a recurring schedule repeats no more often than the spec's
interval, five minutes in Link's. A subagent's work is outside the
day's budget, which is for what schedules send.
What happens across a daylight saving change?
As Vixie cron does: a time skipped forward runs once, at the change, and a time that repeats runs once, so no firing is lost or doubled.