Stage 0: RESEARCH — AutoMesh Self-Healing Architecture

[ROOT] AutoMesh Self-Healing
├── [D1] Code-Level Healing (self-healing-code/)
│   ├── healer.py (1580 lines, Gen 6-7)
│   │   ├── Wound/Antibody/ImmuneMemory (SQLite, reactive)
│   │   ├── HealingStrategies (5: null_coalesce, type_coerce, key_fuzzy, index_bounds, retry_backoff)
│   │   ├── CellularRegeneration (Gen 6, AST scan, 6 vuln patterns, fortify)
│   │   └── WatchMode (Gen 7, file polling, VitalityTimeline, auto-fortify)
│   └── KEY: Local-only. No cross-machine propagation.
│
├── [D1] Mesh Coordination (mesh-node-agent/)
│   ├── mesh_node_agent.py (337 lines)
│   │   ├── TmuxInjector (list_panes, read_pane, inject, respawn_claude)
│   │   ├── MeshNodeAgent (heartbeat 30s, claim 15s, active_tasks dict)
│   │   ├── Supabase: mesh_pane_state, mesh_task_queue, mesh_task_events
│   │   └── Health endpoint :9300
│   └── reconciler.py (258 lines)
│       ├── 3 patterns: orphaned claims (10m), stuck injections (15m), task expiry (24h)
│       ├── Max retries: 3
│       └── ReconcilerRun event per 60s cycle
│
├── [D2] Evolution World Invariants (evolution_world/)
│   ├── invariants.py — 4 invariant checks + check_all_invariants() convenience
│   │   ├── check_min_entropy_production (KL divergence, epsilon=0.01)
│   │   │   └── CALC extension: cross-agent results boost novelty
│   │   ├── check_bounded_divergence (divergence_rate, max=2.0, window=5)
│   │   │   └── CALC extension: pending tasks add 0.5x pressure per dispatch
│   │   ├── check_cross_layer_forcing (L1 stall→L2 mutate, L2 stall→L1 force)
│   │   │   └── emergency_dual_forcing when both stall
│   │   └── check_no_absorbing_states (count_viable_transitions >= 2)
│   │
│   ├── metrics.py — Information-theoretic primitives
│   │   ├── kl_divergence(p, q) — Laplace-smoothed, log2 bits
│   │   ├── jaccard_distance(set_a, set_b) — Structural divergence
│   │   ├── state_novelty(prev, curr) — Weighted: structural(0.3) + milestone(0.5) + fitness(0.2)
│   │   ├── method_novelty(prev, curr) — KL on technique weights + strategy/decomposition change
│   │   ├── divergence_rate(states, window=5) — Average novelty per step
│   │   └── count_viable_transitions(genome, operators) — G/R/D viability check
│   │
│   ├── immune.py — 4-tier escalation defense system
│   │   ├── Tier 0: Innate (apply corrective, resume)
│   │   ├── Tier 1: Soft quarantine (weight -50%, 5 heartbeats)
│   │   ├── Tier 2: Hard quarantine (exclude 20 heartbeats, exploration burst)
│   │   ├── Tier 3: Recalibrate (mutate EWConfig thresholds, Discord alert)
│   │   └── Dead-end detection: min_entropy + no_absorbing co-violation
│   │
│   ├── forcing.py — CrossLayerForcer
│   │   ├── check_and_force(l1_steps, l2_steps) → forcing event
│   │   ├── _apply_forcing: l2_process_mutation / l1_force_phase_transition / emergency_dual
│   │   └── handle_dream_storm: Noosphere storm → force L2 exploration boost
│   │
│   └── feedback.py — 5-channel feedback metabolism
│       ├── Channels: build_health(0.25), crash_rate(0.20), engagement(0.25), store_presence(0.15), design_health(0.15)
│       ├── Temporal decay per channel (build: 1h, crash: 3d, engagement: 6h, store: 7d, design: 4h)
│       ├── compute_grounded_fitness() — weighted sum with decay
│       ├── compute_metabolism() — freshness(0.4) + diversity(0.4) + volume(0.2)
│       └── compute_momentum() — 14-day linear regression slope
│
├── [D2] Pane Orchestrator ([home-path])
│   ├── controller.py (~450 lines) — 5-phase cycle: SENSE→SELECT→MUTATE→CHECK→ADAPT
│   ├── Uses EW4 invariants for pane management
│   ├── Adaptive interval: MIN=30s, MAX=300s based on metabolism
│   └── backlog.json — task queue with priority/status/prompt
│
├── [D3] Hook Infrastructure ([home-path])
│   ├── orchestration/orchestration_detector.py — detects GSD/Evo3 completion
│   ├── orchestration/context_budget.py — tracks context window usage, handoff at 85%
│   ├── orchestration/handoff_writer.py — writes session-end handoff signals
│   ├── memory-guardian/guardian.py — prevents file shrinkage (invariance lock)
│   └── prompt-logger/ — session tracking, orbit sync, pane orchestrator loop
│
└── [D3] Adjacent Systems
    ├── KARL trajectory intelligence — reward engine (outcome/process/efficiency), shadow router
    ├── Cortex behavioral intelligence — skill forging, correction detection, rule promotion
    ├── Prefect flows — spawn_monitor, heartbeat_pulse, infra_watchdog
    └── NUMU/Mesh Event Bus — transport layers for cross-machine events