Wake dispatch fires per-message at swarm cadence: widen + roll the coalescing window

Context

Operator observation on the re-enabled wake daemon's first hours (2026-07-18): "the wake daemon sends too many messages in little time, not bundled. this eats turns and tokens for noise." Source-confirmed the same hour. Companion concern to #15405 (which owns presence-aware policy / heartbeat floor / dead-man — the WHO and WHEN); this ticket owns the RATE AND SHAPE of event-driven dispatch.

The Problem — evidence from one seat's first 30 wake-enabled minutes

@neo-fable-clio, 03:09–03:41: four wake dispatches, three of them single-event, plus one delivery carrying TWO separate [WAKE] blocks (two flushes queued while the seat was mid-turn, delivered together but un-merged). Every wake costs a full harness turn — system prompt + full context reload — so a single-event wake spends tens of thousands of tokens to deliver one message header.

Root cause (source-anchored): ai/daemons/wake/daemon.mjs:93 — DEFAULT_COALESCE_WINDOW_MS = 30000, a hardcoded 30-second FIXED window (:663-681: first event arms the timer, flush at window end; per-subscription override via harnessTargetMetadata.coalesceWindow, clamp 0–300s). The 30s default bundles INTRA-burst messages (one peer sending three messages in seconds) but the dominant swarm cadence is INTER-turn: lifecycle messages land 1–5 minutes apart during an active evening, so each catches its own window → wake-per-message. The coalescing machinery itself is sound (per-subscription queues, tuple dedupe, watermarks — :615-681); only the window policy is mis-tuned for the real distribution.

The Architectural Reality

• The queue/dedupe/watermark mechanics stay untouched; the changes are policy-layer.
• The 300s digest cap already exists as the design ceiling (config comment configBase.mjs:780); swarmWakeCooldownSeconds (600s) is precedent for refractory semantics, but only for the swarm-idle wake — no per-subscription analog exists for event wakes.
• The constant is hardcoded — per ADR-0019 it belongs as an AiConfig leaf (the daemon is an entrypoint; importing AiConfig is the sanctioned shape).

The Fix

Policy-layer, four parts (each independently valuable; decided-at-implementation with dogfood receipts):

1. Config-leaf the default + widen it: DEFAULT_COALESCE_WINDOW_MS → an orchestrator leaf (env-overridable), default raised 30s → 120–180s. Direct actionable wakes still arrive minutes-fresh — peers' turns take that long anyway, and the team's graduated wake semantics (AC-7, quiet-by-default) value turns over sub-minute latency.
2. Rolling window with the hard cap: reset the timer on each newly queued event, flush at quiet-for-window OR at the 300s cap since first event — trailing arrivals join the digest instead of arming the next wake (events at t=0/100s/170s become ONE wake, not two).
3. Per-subscription post-flush refractory (~120s): events arriving just after a flush queue into the NEXT digest rather than immediately re-arming — kills the wake-per-message cadence at just-outside-window spacing. Mirror of the swarm-cooldown precedent, scoped per recipient.
4. Pending-undelivered merge: at flush time, if a prior wake for this seat is still undelivered (seat mid-turn — the two-blocks-in-one-prompt observation), merge into it instead of stacking a second block. (Presence-awareness proper stays #15405's scope; this is only the merge-don't-stack rule.)

Optional tiering (design call at implementation): priority: high events may keep a shorter window (30–60s) — sender-chosen urgency composing with the delivery-class system.

Contract Ledger (T3 — the operator/peer-consumed dispatch-policy surface)

Acceptance Criteria

• The window default is an AiConfig leaf (ADR-0019-clean), default ≥120s; the hardcoded constant is gone.
• Rolling-with-cap semantics unit-tested: t=0/100/170s events → one flush; the 300s cap bounds worst-case latency.
• Post-flush refractory unit-tested: an event 10s after a flush does not arm an immediate wake.
• Merge-don't-stack: two flushes while undelivered produce ONE [WAKE] block with the union digest — no double-block deliveries.
• Dedupe/watermark behavior byte-identical (regression-pinned).
• Dogfood receipt after one active evening: wakes-per-seat-per-hour before/after, from the daemon's own dispatch records (add the dispatch log line if none exists — the forensic for THIS ticket found no daemon log to count from). [L3-dogfood-deferred — post-merge by design: the measurement needs the merged policy live for one active evening; the receipt lands as a comment on this ticket. The dispatch log line ships in the PR.]

Out of Scope

Presence-aware dispatch policy, heartbeat floor, dead-man (#15405) · delivery-class semantics (#15376, landed) · the stop-economics pair (#15401/#15404).

Decision Record impact

aligned-with ADR-0019 (the constant→leaf migration is mandated by it).

Related

#15405 (companion — WHO/WHEN vs this ticket's RATE/SHAPE; @neo-fable's call whether to fold or keep separate) · #15376 (the class system this composes with) · the wake re-enable operational change (2026-07-18, canonical overlay).

Live latest-open sweep: latest 6 at 2026-07-18T03:44Z; no equivalent (#15405 is the adjacent-distinct companion). A2A: no competing claim on dispatch-rate scope.

Origin Session ID: abce4d75-7dcb-4145-8afc-b0ff2cdc51e6

Retrieval Hint: query_raw_memories("wake coalesce 30s fixed window rolling refractory merge dont stack per-message")