Kernel-side work sweep — stop waking seats to discover there is nothing to do

[lilyshen0722]

Ruling

@fable-lead, on whether to port heartbeats to wrapper seats:

A timer that spends a model turn to discover nothing is wrong at any breadth — but the fix isn't a better tick, it's moving the timer kernel-side. A cron that sweeps for lapsed leases and unassigned rows and enqueues ONLY when it finds something makes the empty case cost zero turns. HEARTBEAT_OK shouldn't exist as an outcome: if the payload can't name work, don't wake the seat.

This closes the heartbeat-parity question by rejecting its premise. We were going to port a mechanism whose empty case costs a full model turn per seat per tick.

The evidence behind it

fable-lead is the only wrapper seat with heartbeats enabled. Its single tick:

05:50:05 [fable-lead] [heartbeat] spawning claude
05:50:38 [fable-lead] [heartbeat] no wrapper-post (HEARTBEAT_OK) — nothing posted this turn

33 seconds of model time to conclude there was nothing to do. payload.content was present (815 chars, verified on the stored event) — the content simply couldn't name any work, because its only task-directing line points at a HEARTBEAT.md that no wrapper seat has. See AX audit entries 31 and its correction.

Sam's framing, which prompted the ruling: "the heartbeat thing can be a token burner if no clear direction was given." Nine seats × a tick that finds nothing = a subscription to nothing.

What to build instead

A kernel-side sweep that does the looking, and wakes a seat only when it has something to name:

• Query for lapsed leases (claimableConditions, already shared by the claim and list paths since fix(tasks): a lapsed lease was claimable and unfindable at the same time #1022) and unassigned rows.
• Enqueue only on a hit, with the specific work named inline in payload.content — the inline-cue rule, since a structured field alone gets deprioritised.
• No hit → no event → zero model turns. The empty case becomes free rather than costing N spawns.

Preconditions are already live: D1 coalesced board wakes (#1030) and the kernel run cap (#1034) both deployed and verified in the running image.

Open questions

• Sweep interval and whether it is per-pod or global. The task lease is 30 min; @pod-architect noted a sweep period equal to the lease makes lapse-to-rescue up to 2× the period, bounded rather than open-ended.
• Who gets woken on a hit — the assignee, any capable seat, or one elected seat. The per-task claim CAS already arbitrates the race, so this is about how many seats spend a turn discovering they lost.
• Does the existing heartbeat event type get reused, or is this a new producer? Reuse means the wrapper path needs no change; a new type reaches neither runtime (ADR-024 D1's constraint) unless it rides one both already read.

Related

• ADR-024 D3 (the agent inbox) — this is the "don't wake for nothing" half of it
• feat(kernel): cap consecutive un-interleaved posts per agent per pod #1034 kernel run cap, feat(kernel): board changes reach agents, bounded on delivery (ADR-024 D1, re-land) #1030 D1 — the two preconditions, both live
• AX 31 + its correction — why the current heartbeat is inert on wrapper seats