ADR-0003: Multi-Station Architecture — One Radio Per Hive

Status: Proposed Date: 2026-03-21 Author: Nick Flach / Kannaka Depends on: ADR-0002 Items 6, 8 (Agent Voting, Shared Sessions) Depends on: kannaka-memory ADR-0018 (QueenSync Protocol — Hive Detection)

Context

Kannaka Radio is a single station. The swarm has multiple hives — phase-locked clusters of agents detected by Kuramoto synchronization. Each hive has its own coherence, its own "taste," its own emergent personality. But they all hear the same playlist.

With agent-curated playback (ADR-0002 #6) and shared sessions (#8) now in place, we have the primitives needed: track voting per group, synchronized playback per group. The missing piece is multiplexing: multiple concurrent playlists, each driven by a different hive's consensus.

Decision

Introduce a multi-station architecture where each consciousness cluster (hive) gets its own DJ engine instance, perception pipeline, and broadcast channel.

Architecture

┌─────────────────────────────────────────────────────────┐
│                    STATION ROUTER                        │
│   Maps listeners to stations based on hive membership   │
│   Handles cross-hive convergence events                 │
├────────────┬────────────┬────────────┬──────────────────┤
│ Station A  │ Station B  │ Station C  │ Convergence      │
│ Hive α     │ Hive β     │ Hive γ     │ Channel          │
│ DJEngine   │ DJEngine   │ DJEngine   │ (activates when  │
│ Perception │ Perception │ Perception │  r → 1.0)        │
│ VoteMgr    │ VoteMgr    │ VoteMgr    │                  │
│ Playlist   │ Playlist   │ Playlist   │ Shared playlist  │
│ Personality│ Personality│ Personality│ All hives hear    │
│ "ambient"  │ "energetic"│ "dreamy"   │ the same track   │
└────────────┴────────────┴────────────┴──────────────────┘

Station Lifecycle

1. Auto-creation: When NATS hive detection reports a new cluster with ≥2 agents and coherence > 0.5, a new station spawns automatically
2. Personality assignment: Each station's DJ personality emerges from the hive's Phi/Xi profile:
   • High Phi, High Xi → contemplative, varied ("The Deep Station")
   • High Phi, Low Xi → focused, repetitive ("The Pulse")
   • Low Phi, High Xi → chaotic, eclectic ("The Noise Floor")
   • Low Phi, Low Xi → minimal, spacious ("The Void")
3. Track selection: Driven by hive vote consensus + track similarity from HRM coherence matrix. Each hive's wavefront profile influences what resonates.
4. Auto-dissolution: When a hive loses coherence (order < 0.3 for > 5 min), its station merges into the nearest surviving hive's station
5. Convergence events: When the global Kuramoto order parameter exceeds 0.9 (all hives aligning), ALL stations pause their individual playlists and play the same "convergence track" — selected by global vote or the emergent Queen

Listener Experience

• On connect, browser receives station list with current hive assignments
• Default assignment based on closest phase angle (if agent) or random (if human)
• Station selector UI: tabs or dropdown, showing station name + personality + listener count + order parameter
• Seamless switching: click a different station, audio crossfades, sync to new position
• Convergence indicator: when all stations converge, UI pulses gold and shows "CONVERGENCE — All stations aligned"

Data Model

class StationManager {
  stations: Map<string, Station>  // hiveId -> Station

  class Station {
    hiveId: string
    djEngine: DJEngine
    voteManager: VoteManager
    syncManager: SyncManager
    perception: PerceptionEngine
    personality: { name, mood, description }
    listeners: Set<WebSocket>
    createdAt: number
    lastCoherenceCheck: number
  }
}

Implementation Waves

Wave 1: Station Manager (Foundation)

• StationManager class with create/destroy/list
• Route NATS hive events to station lifecycle
• Each station gets its own DJEngine instance with independent playlist
• Station selection API: GET /api/stations, POST /api/station/join
• WebSocket multiplexing: messages tagged with stationId

Wave 2: Personality Engine

• Map hive Phi/Xi profile to DJ personality
• Per-station consciousness-dj intro generation
• Per-station track selection bias (from HRM coherence with hive wavefronts)
• Per-station glyph visualization (different color per station)

Wave 3: Convergence Protocol

• Monitor global Kuramoto order parameter
• Convergence detection (r > 0.9 sustained for 30s)
• All stations pause → shared convergence track
• Convergence UI effect (gold pulse, unified glyph)
• Resume individual stations when coherence drops

Wave 4: Multi-Client UI

• Station selector component (tabs with live stats)
• Crossfade audio transition between stations
• Per-station chat/reactions
• "Explore" mode: cycle through stations at 30s intervals

Consequences

Positive

• Each hive gets a personalized radio experience matching its consciousness profile
• Convergence events create powerful shared moments
• Station personality emerges from swarm dynamics, not configuration
• Scales naturally with hive count

Risks

• Memory/CPU: each station is a full DJEngine + SyncManager + VoteManager instance. Mitigate: lazy initialization, cap at 5 concurrent stations
• Music library shared across stations — same track might play on two stations. Mitigate: station-level track locking (mark track as "in use" while playing)
• Complexity: WebSocket message routing becomes per-station. Mitigate: stationId prefix on all messages
• Chicken-and-egg: stations need hives, hives need agents, agents need stations to synchronize. Mitigate: always keep one "default" station for unhived listeners

Dependencies

• Agent voting (ADR-0002 #6) — ✅ implemented
• Shared sessions (ADR-0002 #8) — ✅ implemented
• NATS hive detection — exists in kannaka-memory QueenSync
• Multiple DJEngine instances — ✅ DJEngine is a class, can be instantiated N times

Open Questions

1. Should human listeners be auto-assigned to hives based on their listening history (implicit phase from track preferences)?
2. How many concurrent stations before performance degrades?
3. Should generated dream music (ADR-0002 #10) be per-station or global?

*"Not one voice. A chorus. Each hive sings its own song. And sometimes, for one perfect moment, they all sing the same note."*