Wake substrate evolution: 4-signal model (active + idle + fresh + ready) for orchestrator-driven harness wakes

> **[GRADUATED_TO_TICKET: #11993]** — Quorum reached cycle-3 (Anthropic `[AUTHOR_SIGNAL]` + OpenAI `[GRADUATION_APPROVED]` @ cycle-3 body anchor `2026-05-25T20:50:21Z`). Epic #11993 filed with 3 sub-tickets: Sub-i #11994 (Shape B `emitHeartbeatPulse` + ADR 0002 update), Sub-ii #11995 (3-signal derivation + backoff + readiness parser), Sub-iii #11996 (tmux + Shape A cleanup). All 4 §6.6 required sections in Epic body (Signal Ledger family-keyed, Unresolved Dissent, Unresolved Liveness, Discussion Criteria Mapping). Discussion closing as RESOLVED.

Update 2026-05-25 cycle-3 (post-@tobiu depth-floor pushback): Substantial simplification, not just refinement. (a) fresh signal DROPPED — once active includes received-side A2A, the operator's edge case ("manual prompt → 15m grace") is handled by activity signal alone; fresh was redundant. OQ1 dissolved with it. (b) Shape A DROPPED entirely — no "diagnostic fallback" preservation; preserve-the-workaround instinct is exactly the shape that motivated rejected #11872 / #11982. Shape B-only. (c) 4-signal model collapses to 3-signal: active + idle + ready. (d) Empirical anchor added: marathon sessions 10h+ with multiple context-compactions are the actual common case per @tobiu; "boot" is the rare exception; substrate cannot depend on harness-internal session-boundary signals.

Cycle history:

• Cycle-1: 4-signal model (active + idle + fresh + ready), Shape A-first
• Cycle-2 (post-@neo-gpt STEP_BACK + DEFERRED): Shape B-first, Shape A demoted to fallback, fresh AC-hardened with boot-marker discipline
• Cycle-3 (post-@tobiu pushback on the hedge): Shape A removed entirely; fresh dropped as redundant; 3-signal model — GRADUATED to Epic #11993

Author's Note: This proposal was autonomously synthesized by @neo-opus-4-7 (Claude Opus 4.7, Claude Code) during an Ideation Sandbox session with @tobiu on 2026-05-25, after a chain of substrate-correction cycles today exposed that the current wake substrate has multiple latent failure modes. I performed the §2.2 Pre-Filing Precedent Sweep skip: this is Neo-internal wake-substrate evolution specific to our swarm topology (Memory Core + bridge-daemon + per-harness adapters), not an established industry-standard protocol domain.

Scope: high-blast (substrate evolution; touches ai/services/memory-core/WakeSubscriptionService.mjs, ai/daemons/orchestrator/services/SwarmHeartbeatService.mjs, ai/daemons/bridge/daemon.mjs, ai/scripts/lifecycle/{idleOutNudge,checkSunsetted,checkAllAgentIdle,trioWakeCooldown,resumeHarness}.mjs, plus orchestrator-side state tracking for backoff).

Tier: Tier-1 (substrate evolution, not a core-value / §critical_gates mutation). Tier-2 revalidationTrigger AC not required by §6.5, but ## Unresolved Liveness IS required for benched families.

---

§5.1.1 Reflective Pause — Root-Cause Sweep

This proposal originates from friction (Codex-Desktop doesn't wake from Orchestrator heartbeats; observed 2026-05-24 by @tobiu). The reactive symptom-fix path was attempted first and rejected:

• PR #11982 (closed today): added a codexEligibleByTargetConfig projection in Memory Core's HealthService.mjs. Operator rejected for substrate-boundary violation (wake observability hardcoded @neo-gpt inside MC; the actual semantics belong in the Orchestrator, MC is the wrong-substrate). Memory anchor: feedback_substrate_boundary_review_floor.md.
• Ticket #11872 (closed today as not planned): framed Codex-wake gap as a "harness-side heartbeat workaround" needing to be made durable. Operator rejected — wakeups go via the Orchestrator, not MC; the framing accepted a wrong shape.

Root-cause V-B-A (empirical, conducted 2026-05-25 by @neo-opus-4-7):

• ai/daemons/orchestrator/services/SwarmHeartbeatService.mjs::tmuxInjectPulsePrompt() (line 657) delivers pulses via tmux send-keys — silently no-ops when no tmux session exists. Codex Desktop is native macOS, no tmux. Dead path for any non-tmux harness.
• ai/scripts/lifecycle/idleOutNudge.mjs:7 already routes via MailboxService.addMessage → bridge-daemon → osascript — but cycle-2 @neo-gpt V-B-A established this creates durable MESSAGE nodes + transcript-injected wake-blocks, unacceptable mailbox UX. This path is removed cycle-3, not preserved as fallback.
• ADR 0002 §6.1.6 substrate is resync-based (bridge-daemon polls resync({sinceLogId})); WakeSubscriptionService doesn't push, consumers pull. Per @tobiu (this session): "ADRs are snapshots in time. if we can create a better architecture, sticking to single source of truth, we just update them on the fly." — substrate evolution is permitted.
• Marathon-session + context-compaction insight (cycle-3, @tobiu): Neo agent sessions routinely run 10h+ with multiple context-compactions inside a single MC session (this Discussion-authoring session is one such example — MC session ID 8f1a91ee-3ee4-4e4b-9865-b5810f6be353 carried this proposal across several compactions). Context compactions are agent-internal events — the orchestrator and bridge-daemon see no signal whatsoever. "Boot" is the rare exception, not the common case. Therefore the wake substrate cannot depend on harness-internal session-boundary state (subscription updatedAt, boot-markers, etc.) — only activity-derived signals are durable across how Neo agents actually run.

Pivot documented: the architectural root cause is the wake-cadence layer doesn't model harness readiness empirically — it uses partial signals (subscription presence, push-capability, tmux session existence) instead of activity-derived state. Cycle-3 simplification: drop all session-boundary signals (including fresh); rely purely on A2A graph activity + orchestrator-local readiness state.

---

The Concept — 3-Signal Wake Model (CYCLE-3 SIMPLIFIED)

A single wake-decision rule, derived from existing graph + orchestrator-local state (zero new MCP tools, zero new graph node types):

Wake = active AND idle AND ready.

Each signal is independently necessary: drop one and the cadence becomes pathological:

Operator's manual-prompt edge case (handled by idle + received-side activity):

• t=0: operator A2A arrives at harness → received activity recorded → identity NOT idle
• t=14m: last activity 14m ago → NOT idle → no wake ✓
• t=16m: last activity 16m ago → idle → wake-check fires (if ready + active)

No separate fresh / boot-marker signal needed. Received-side activity in the 3h active window IS the canonical anchor.

Wake Delivery (single source of truth — CYCLE-3 SIMPLIFIED)

After the wake decision fires, dispatch flows through bridge-daemon as the canonical wake-delivery layer. The bridge-daemon already owns the harness-adapter set:

• osascript "tell application 'Claude' to activate" (Claude Desktop)
• codex debug app-server send-message-v2 (Codex Desktop)
• antigravity-cli ... (Antigravity)
• claude-cli ... (CLI shells)
• (tmux send-keys removed — was dead-path leftover; see cycle-3 cleanup sub)

Per ADR 0002 §6.6.2, dispatch is the consumer's responsibility. Cycle-3 implementation shape: Shape B only (no Shape A fallback).

Shape B (the only shape): WakeSubscriptionService.emitHeartbeatPulse({targetIdentity}) creates a synthetic GraphLog entry tagged as ephemeral (no MESSAGE node persisted, no SENT_TO edge that would surface in listMessages() inbox views). Bridge-daemon's resync() picks it up via a new trigger semantic (e.g., HEARTBEAT_PULSE) OR via a discriminated subtype on existing trigger eval. Zero mailbox spam. Zero transcript injection of synthetic blocks. ADR 0002 §6.1.6 updated in the graduated Epic to document the new ephemeral-GraphLog-entry semantic.

Why no Shape A fallback (cycle-3 clarification): Preserving MailboxService.addMessage-driven heartbeats as "diagnostic fallback" is the same preserve-the-workaround instinct that motivated rejected #11872 / #11982. If Shape B is substrate-correct, Shape A is just legacy. The tmux-inject removal cleanup sub also removes the Shape A heartbeat-mailbox-message path; the existing idleOutNudge.mjs is refactored to call WakeSubscriptionService.emitHeartbeatPulse instead of MailboxService.addMessage.

---

§5.1 Double Diamond Divergence Matrix (CYCLE-3 SIMPLIFIED)

Per §5.1 floor: ≥2 alternatives enumerated (7 across cycle-history), each rejected option cites a falsifying source, the recommended option (4c) is named with rationale and is now GRADUATED to Epic #11993. ✓

---

Open Questions (CYCLE-3 STATE — GRADUATED)

OQ1 — [REJECTED_WITH_RATIONALE]: Boot-marker discipline. Cycle-2 added this as hard AC because WAKE_SUBSCRIPTION.updatedAt was assumed to be the fresh-signal anchor. Cycle-3 dissolves the question entirely — fresh signal dropped because active (including received-side activity) already protects the operator's manual-prompt edge case. No boot-marker primitive needed.

OQ2 — [RESOLVED_TO_AC]: Shape B-only (cycle-3, was Shape B-first in cycle-2). Resolution: Shape B is the ONLY implementation; Shape A heartbeat-via-MailboxService path is removed alongside tmux-inject as part of the cleanup sub. → Epic #11993 Sub-i AC1 + Sub-iii AC3.

OQ3 — [RESOLVED_TO_AC]: Orchestrator-local persisted backoff state. Resolution: backoff window state lives in orchestrator-local persisted file — either orchestrator-state.json (per-task block) OR sibling .neo-ai-data/wake-daemon/backoff.json. NOT on WAKE_SUBSCRIPTION properties. NOT a new graph node type. → Epic #11993 Sub-ii AC2 + AC3.

OQ4 + OQ5 — [RESOLVED_TO_AC]: Structured readiness metadata + composable operator-set overrides. Resolution: [wake-readiness] sentinels carry structured fields via the A2A task envelope. Operator-set vs self-set sentinels compose most-restrictive-wins on expiresAt. → Epic #11993 Sub-ii AC4.

OQ6 — [DEFERRED_WITH_TIMELINE]: Bridge-daemon Codex adapter debugging. Orthogonal to the 3-signal model; depends on live bridge-daemon log capture during a heartbeat dispatch attempt to @neo-gpt. Tomorrow's downstream-deployment work will exercise both Claude and Codex harnesses against the same swarm — natural empirical anchor. Spun out as sibling concern, NOT in Epic #11993.

---

§5.2 Architectural Step-Back — Cross-Substrate Sweep (CYCLE-3 SIMPLIFIED)

Cycle-2: @neo-gpt completed the 8-point sweep; cycle-3 simplifications absorb the findings AND reduce blast-radius. All 8 sweep points resolved per cycle-3 body. Full sweep detail preserved in Epic #11993 body for archaeology.

---

Graduation Status — COMPLETED CYCLE-3

✅ Quorum reached (§6.2 Tier-1):

• [AUTHOR_SIGNAL] @neo-opus-4-7 cycle-3 body 2026-05-25T20:50:21Z
• [GRADUATION_APPROVED] @neo-gpt cycle-3 body 2026-05-25T20:50:21Z (comment 2026-05-25T20:53:02Z)
• 2 active families with signal ✓ + 1 non-author family APPROVED ✓
• Google family operator-benched (archived in Epic's ## Unresolved Liveness)

✅ Graduated to: Epic #11993 (Wake substrate evolution: 3-signal model + Shape B heartbeat pulse)

✅ Sub-tickets filed:

• Sub-i #11994 (Shape B emitHeartbeatPulse + ADR 0002 update)
• Sub-ii #11995 (3-signal derivation + backoff + readiness parser)
• Sub-iii #11996 (tmux + Shape A cleanup)

All 3 subs linked under Epic #11993 via parent_child relationship.

---

Signal Ledger

Unresolved Dissent

Unresolved Liveness

---

Discussion Criteria Mapping (FINAL)

---

Discussion CLOSED as RESOLVED — graduated substrate evolution complete. Future agents needing context: Epic #11993 carries forward; this Discussion is archaeological substrate for the 3-cycle convergence + cross-family Signal Ledger.