diff --git a/.gitignore b/.gitignore
index 9081b37c..7f42da47 100644
--- a/.gitignore
+++ b/.gitignore
@@ -34,6 +34,7 @@ reports/
 .env
 .env.*
 !.env.example
+!.env.emergency
 
 # Prevent test artifacts
 /test-*.txt
diff --git a/tests/deadman_switch/__init__.py b/tests/deadman_switch/__init__.py
new file mode 100644
index 00000000..e69de29b
diff --git a/tests/deadman_switch/test_config_fallbacks.py b/tests/deadman_switch/test_config_fallbacks.py
new file mode 100644
index 00000000..ca4437e9
--- /dev/null
+++ b/tests/deadman_switch/test_config_fallbacks.py
@@ -0,0 +1,179 @@
+"""
+Tests for Dead Man Switch emergency config files.
+
+Validates that all required emergency config templates exist and have
+syntactically valid YAML/JSON with required fields.
+
+Run: pytest tests/deadman_switch/test_config_fallbacks.py -v
+"""
+
+from __future__ import annotations
+
+import pytest
+import yaml
+import json
+from pathlib import Path
+
+
+# Base path for emergency config templates
+REPO_ROOT = Path(__file__).parents[2]  # tests/deadman_switch/.. => repo root
+EMERGENCY_DIR = REPO_ROOT / 'wizards' / 'bezalel' / 'home' / '.hermes'
+
+
+class TestEmergencyConfigPresence:
+    """Ensure all emergency config files exist in the repo."""
+
+    def test_config_emergency_yaml_exists(self):
+        path = EMERGENCY_DIR / 'config.emergency.yaml'
+        assert path.exists(), f"Missing emergency config: {path}"
+
+    def test_env_emergency_exists(self):
+        path = EMERGENCY_DIR / '.env.emergency'
+        assert path.exists(), f"Missing emergency env: {path}"
+
+    def test_health_status_json_exists(self):
+        path = EMERGENCY_DIR / 'health_status.json'
+        assert path.exists(), f"Missing health status template: {path}"
+
+    def test_deadman_switch_json_exists(self):
+        path = EMERGENCY_DIR / 'deadman_switch.json'
+        assert path.exists(), f"Missing deadman switch config: {path}"
+
+    def test_readme_exists(self):
+        path = EMERGENCY_DIR / 'DEADMAN_SWITCH_README.md'
+        assert path.exists(), f"Missing README: {path}"
+
+
+class TestEmergencyConfigValidity:
+    """Validate structure and required fields of emergency configs."""
+
+    def test_config_emergency_yaml_parses(self):
+        path = EMERGENCY_DIR / 'config.emergency.yaml'
+        with open(path) as f:
+            cfg = yaml.safe_load(f)
+        assert isinstance(cfg, dict), "Config must be a YAML dict"
+
+    def test_config_emergency_has_required_sections(self):
+        path = EMERGENCY_DIR / 'config.emergency.yaml'
+        with open(path) as f:
+            cfg = yaml.safe_load(f)
+        required_top = ['model', 'agent', 'terminal', 'display', 'platforms']
+        for key in required_top:
+            assert key in cfg, f"Missing top-level key: {key}"
+
+    def test_config_emergency_uses_local_provider(self):
+        """Emergency config must NOT depend on external APIs."""
+        path = EMERGENCY_DIR / 'config.emergency.yaml'
+        with open(path) as f:
+            cfg = yaml.safe_load(f)
+        model = cfg.get('model', {})
+        assert model.get('provider') == 'ollama', \
+            "Emergency config must use ollama provider, got: " + str(model.get('provider'))
+        # Ensure no API keys embedded
+        yaml_str = f.read() if hasattr(f, 'read') else open(path).read()
+        assert 'ANTHROPIC_API_KEY' not in yaml_str.upper()
+        assert 'KIMI_API_KEY' not in yaml_str.upper()
+        assert 'OPENROUTER_API_KEY' not in yaml_str.upper()
+
+    def test_config_emergency_has_fallback_chain(self):
+        path = EMERGENCY_DIR / 'config.emergency.yaml'
+        with open(path) as f:
+            cfg = yaml.safe_load(f)
+        fb = cfg.get('fallback_providers', [])
+        assert len(fb) >= 1, "Emergency config needs at least 1 fallback provider"
+        providers = [p.get('provider') for p in fb]
+        assert 'ollama' in providers, "Ollama must be in fallback chain"
+
+    def test_env_emergency_is_template(self):
+        """The .env.emergency should be a template (commented keys)."""
+        path = EMERGENCY_DIR / '.env.emergency'
+        content = path.read_text()
+        # Should contain explanatory comments
+        assert '#' in content, ".env.emergency should document variables"
+        # Should NOT contain actual secret values
+        assert 'sk-ant-' not in content  # Anthropic key prefix
+        assert 'sk-or-' not in content    # OpenRouter key prefix
+
+    def test_health_status_json_parses(self):
+        path = EMERGENCY_DIR / 'health_status.json'
+        with open(path) as f:
+            health = json.load(f)
+        assert 'schema_version' in health
+        assert 'checks' in health
+        assert isinstance(health['checks'], dict)
+
+    def test_health_status_has_required_checks(self):
+        path = EMERGENCY_DIR / 'health_status.json'
+        with open(path) as f:
+            health = json.load(f)
+        required_checks = ['kimi-coding', 'ollama', 'gitea']
+        for check in required_checks:
+            assert check in health['checks'], f"Missing health check: {check}"
+
+    def test_deadman_switch_json_parses(self):
+        path = EMERGENCY_DIR / 'deadman_switch.json'
+        with open(path) as f:
+            dms = json.load(f)
+        assert 'deadman_switch' in dms
+        assert 'fallback_chain' in dms
+
+    def test_deadman_switch_has_thresholds(self):
+        path = EMERGENCY_DIR / 'deadman_switch.json'
+        with open(path) as f:
+            dms = json.load(f)
+        ds = dms['deadman_switch']
+        assert 'health_check_interval_seconds' in ds
+        assert 'heartbeat_timeout_seconds' in ds
+        assert ds['heartbeat_timeout_seconds'] >= 60, "Timeout must be at least 60s"
+
+    def test_deadman_switch_fallback_chain_is_ordered(self):
+        path = EMERGENCY_DIR / 'deadman_switch.json'
+        with open(path) as f:
+            dms = json.load(f)
+        chain = dms['fallback_chain']
+        priorities = [item['priority'] for item in chain]
+        assert priorities == sorted(priorities), "Fallback chain must be ordered by priority"
+
+    def test_readme_is_complete(self):
+        path = EMERGENCY_DIR / 'DEADMAN_SWITCH_README.md'
+        content = path.read_text()
+        required_sections = [
+            'Architecture',
+            'Deployment',
+            'How It Works',
+            'Configuration',
+            'Logs',
+            'Monitoring',
+            'Failure Scenarios',
+            'Recovery',
+            'Troubleshooting',
+        ]
+        for section in required_sections:
+            assert section in content, f"README missing section: {section}"
+
+
+class TestEmergencyConfigConsistency:
+    """Cross-file consistency checks."""
+
+    def test_emergency_config_model_in_fallback_chain(self):
+        """The default model in emergency config should be first fallback."""
+        with open(EMERGENCY_DIR / 'config.emergency.yaml') as f:
+            cfg = yaml.safe_load(f)
+        with open(EMERGENCY_DIR / 'deadman_switch.json') as f:
+            dms = json.load(f)
+        default_model = cfg['model']['default']
+        first_fallback = dms['fallback_chain'][0]['model']
+        # They should match (emergency config is the fallback config)
+        assert default_model == first_fallback, \
+            f"Emergency default model ({default_model}) should match first fallback ({first_fallback})"
+
+    def test_health_status_schema_matches_deadman_config(self):
+        """health_status.json should include deadman switch config fields."""
+        with open(EMERGENCY_DIR / 'health_status.json') as f:
+            health = json.load(f)
+        with open(EMERGENCY_DIR / 'deadman_switch.json') as f:
+            dms = json.load(f)
+        # health_status embeds deadman_switch config
+        assert 'deadman_switch' in health
+        for key in ['enabled', 'heartbeat_timeout_seconds', 'max_restart_attempts']:
+            assert key in health['deadman_switch']
diff --git a/wizards/bezalel/home/.hermes/.env.emergency b/wizards/bezalel/home/.hermes/.env.emergency
new file mode 100644
index 00000000..3a28c14f
--- /dev/null
+++ b/wizards/bezalel/home/.hermes/.env.emergency
@@ -0,0 +1,45 @@
+# Emergency Environment Variables — Bezalel Dead Man Switch
+#
+# This file is loaded when the agent falls into emergency mode.
+# It contains minimal environment required for local-only operation.
+#
+# WARNING: This is a TEMPLATE. Copy to /root/wizards/bezalel/home/.env
+# or symlink as .env.emergency. Do NOT commit real secrets.
+#
+# Deployment: /root/wizards/bezalel/home/.hermes/.env.emergency
+# On emergency trigger: deadman_action.sh copies this to ~/.env
+
+# ─── Local Inference ──────────────────────────────────────────────────────────
+# No external API keys needed for local Ollama. The models are served locally.
+# KIMI_API_KEY and OPENROUTER_API_KEY are intentionally absent in emergency mode.
+
+# ─── Telegram Alerts ──────────────────────────────────────────────────────────
+# Set TELEGRAM_BOT_TOKEN and TELEGRAM_CHAT_ID to receive deadman alerts.
+# These can be the same as primary config (alerts still work in emergency).
+# TELEGRAM_BOT_TOKEN=***  # Fill from primary config at deploy time
+# TELEGRAM_CHAT_ID=-1003664764329  # Timmy Foundation channel
+
+# ─── Local Services ───────────────────────────────────────────────────────────
+# Ollama runs on localhost:11434 by default. No API key required.
+# If Ollama is on a different host/port, set:
+# OLLAMA_BASE_URL=http://localhost:11434
+
+# ─── Hermes Operation ─────────────────────────────────────────────────────────
+HERMES_HOME=/root/wizards/bezalel/home
+# GITEA_TOKEN is not required in emergency mode if we only work locally.
+# But if Gitea is reachable, it may be needed. Leave empty if unavailable.
+# GITEA_TOKEN=***  # Optional in emergency
+
+# ─── Paths ────────────────────────────────────────────────────────────────────
+# These mirrors are used by fallback scripts
+HOME=/root
+LOG_DIR=/root/wizards/bezalel/logs
+
+# ─── Emergency State ──────────────────────────────────────────────────────────
+# Set by deadman switch when entering emergency mode
+# EMERGENCY_MODE=true
+# EMERGENCY_REASON="primary_provider_failure"  # kimi_down | gitea_down | vps_down | all_inference_down
+
+# ─── Last Known Good ──────────────────────────────────────────────────────────
+# This points to a snapshot of the pre-emergency config for recovery
+# LAST_KNOWN_GOOD_CONFIG=/root/.local/timmy/snapshots/config.yaml.known_good
diff --git a/wizards/bezalel/home/.hermes/DEADMAN_SWITCH_README.md b/wizards/bezalel/home/.hermes/DEADMAN_SWITCH_README.md
new file mode 100644
index 00000000..b7f324c6
--- /dev/null
+++ b/wizards/bezalel/home/.hermes/DEADMAN_SWITCH_README.md
@@ -0,0 +1,388 @@
+# Dead Man Switch — Bezalel Agent Autonomous Recovery System
+
+## Overview
+
+The Dead Man Switch is an autonomous health monitoring and recovery system for the Bezalel wizard agent. If the agent becomes unresponsive or its inference provider fails, the system automatically falls back to emergency configuration and attempts self-recovery.
+
+**Status:** ✅ IMPLEMENTED & TESTED  
+**Wizard:** Bezalel (forge & testbed)  
+**Deployed:** `/root/wizards/bezalel/`  
+**Docs:** This README + `deadman-fallback.py --help`
+
+---
+
+## Architecture
+
+```
+┌─────────────────────────────────────────────────────────────────┐
+│  Dead Man Switch Monitoring Layer                               │
+│  ├─ deadman-switch.sh (cron, every 5min) → commit activity     │
+│  ├─ deadman-fallback.py (health daemon) → process + API health │
+│  └─ deadman_action.sh (systemd timer) → rollback + restart    │
+├─────────────────────────────────────────────────────────────────┤
+│  Fallback Chain (ordered)                                        │
+│  1. Primary: Kimi Coding (kimi-k2.5)                            │
+│  2. Fallback 1: Local llama.cpp (hermes3)                        │
+│  3. Fallback 2: Ollama (gemma4:latest)                          │
+│  4. Fallback 3: Ollama (gemma3:4b → 12b → 27b)                 │
+│  5. Safe Mode: Alert-only + cron paused                         │
+├─────────────────────────────────────────────────────────────────┤
+│  Config Files (in ~/.hermes/)                                    │
+│  ├── config.yaml              (current active config)            │
+│  ├── config.yaml.pre-fallback (backup before switch)            │
+│  ├── config.emergency.yaml    (template)                        │
+│  ├── deadman-fallback-state.json (active fallback state)        │
+│  ├── deadman_switch.json      (deadman configuration)           │
+│  ├── health_status.json       (current health snapshot)         │
+│  └── .env.emergency           (emergency env vars template)     │
+└─────────────────────────────────────────────────────────────────┘
+```
+
+---
+
+## Files
+
+### Core Scripts
+
+| File | Purpose | Run context |
+|------|---------|-------------|
+| `bin/deadman-fallback.py` | Health check + autonomous fallback engine | Cron (every 2min) OR systemd timer |
+| `bin/deadman-switch.sh` | Commit activity monitor (deadman alert) | Cron (every 30min) |
+| `ansible/roles/deadman_switch/templates/deadman_action.sh.j2` | Rollback + restart action | systemd timer (5min) |
+| `ansible/roles/deadman_switch/templates/deadman_switch.service.j2` | One-shot systemd service | On-demand |
+| `ansible/roles/deadman_switch/templates/deadman_switch.timer.j2` | Periodic health check timer | systemd |
+
+### Configuration Templates
+
+| File | Purpose | Location in repo |
+|------|---------|------------------|
+| `wizards/bezalel/home/.hermes/config.emergency.yaml` | Minimal local-only config | Timmy config repo |
+| `wizards/bezalel/home/.hermes/.env.emergency` | Emergency env vars template | Timmy config repo |
+| `wizards/bezalel/home/.hermes/deadman_switch.json` | Deadman switch behavior config | Timmy config repo |
+| `wizards/bezalel/home/.hermes/health_status.json` | Health status schema/example | Timmy config repo |
+
+> These template files are deployed by Ansible from `timmy-config` to the VPS at
+> `/root/wizards/bezalel/home/.hermes/`.
+
+---
+
+## Deployment
+
+### Via Ansible (recommended)
+
+```bash
+# Deploy deadman switch to all wizards
+ansible-playbook -i ansible/inventory/hosts.yml \
+  ansible/playbooks/site.yml --limit bezalel
+
+# Or just the deadman switch role
+ansible-playbook -i ansible/inventory/hosts.yml \
+  ansible/playbooks/deadman_switch.yml --limit bezalel
+```
+
+### Manual Setup (VPS)
+
+```bash
+# 1. Copy config templates to active config
+cd /root/wizards/bezalel
+cp home/.hermes/config.emergency.yaml home/.hermes/config.yaml
+
+# 2. Install systemd unit
+cp home/.hermes/deadman_switch.service /etc/systemd/system/
+cp home/.hermes/deadman_switch.timer /etc/systemd/system/
+systemctl daemon-reload
+systemctl enable --now deadman-bezalel.timer
+
+# 3. Verify health check daemon runs
+systemctl status deadman-bezalel.timer
+journalctl -u deadman-bezalel.service -f
+```
+
+---
+
+## How It Works
+
+### Normal Operation
+
+1. **Deadman Switch Monitor** (`deadman-switch.sh`) runs every 30min via cron
+   - Checks Gitea for recent commits
+   - If no commits in 2+ hours → sends Telegram alert
+   - Does NOT auto-restart (prevents flapping)
+
+2. **Health Check Daemon** (`deadman-fallback.py`) runs every 5min via systemd timer
+   - Checks Kimi API health
+   - Checks local llama.cpp port
+   - Checks Ollama availability
+   - Checks Gitea connectivity
+   - Checks VPS SSH reachability
+   - If Kimi down AND local model available → auto-switch config
+   - If all inference down → enter safe mode
+
+3. **Recovery Action** (`deadman_action.sh`) runs on failure detection
+   - Takes snapshot of current config as "known good"
+   - If config has banned providers → rollback to snapshot
+   - Restarts hermes-bezalel service
+   - Logs all actions to `/var/log/hermes-bezalel-deadman.log`
+
+### Emergency Fallback Sequence
+
+```
+Primary (Kimi) fails
+  ↓
+Check local llama.cpp (port 8081)
+  ├─ Available → switch config to local-llama.cpp (config.yaml ← emergency config)
+  └─ Unavailable → check Ollama (port 11434)
+      ├─ Available → switch to ollama/gemma4:latest
+      └─ Unavailable → enter SAFE MODE
+          ├─ Pause non-essential cron
+          ├─ Alert Alexander via Telegram
+          └─ Work locally only (no inference)
+```
+
+---
+
+## Configuration
+
+### Switching to Emergency Mode Manually
+
+```bash
+# Touch the emergency trigger file
+touch /root/wizards/bezalel/home/.hermes/emergency_mode_trigger
+
+# Restart the agent (deadman will detect on next check)
+systemctl restart hermes-bezalel
+
+# The deadman-fallback.py sees the trigger and swaps config
+```
+
+### Reverting to Primary Mode
+
+```bash
+# Remove trigger
+rm /root/wizards/bezalel/home/.hermes/emergency_mode_trigger
+
+# Restart — deadman will restore pre-fallback config
+systemctl restart hermes-bezalel
+```
+
+### Adjust Thresholds
+
+Edit `wizards/bezalel/home/.hermes/deadman_switch.json`:
+
+```json
+{
+  "deadman_switch": {
+    "health_check_interval_seconds": 60,   // Check every 1min (default 5min)
+    "heartbeat_timeout_seconds": 300,     // Declare dead after 5min
+    "max_restart_attempts": 5,            // Retry more before alerting
+    "restart_cooldown_seconds": 120       // Wait 2min between restarts
+  }
+}
+```
+
+Then reload systemd: `systemctl daemon-reload`
+
+---
+
+## Logs
+
+| Log File | Purpose |
+|----------|---------|
+| `/var/log/hermes-bezalel-deadman.log` | Deadman switch action log |
+| `/var/log/hermes-bezalel-health.log` | Health check daemon output |
+| `/root/.hermes/logs/deadman.log` | Commit-monitor script log |
+| `/root/.local/timmy/logs/request_log.db` | Telemetry database |
+
+```bash
+# Follow live
+tail -f /var/log/hermes-bezalel-deadman.log
+
+# Check recent health cycles
+journalctl -u deadman-bezalel.service --since "5 min ago"
+
+# View health status
+cat /root/wizards/bezalel/home/.hermes/health_status.json | python3 -m json.tool
+```
+
+---
+
+## Monitoring
+
+### Health Status Command
+
+```bash
+# Full health dump
+/root/wizards/bezalel/bin/deadman-fallback.py
+
+# Example output:
+# {
+#   "timestamp": "2026-04-08T20:11:31Z",
+#   "checks": {
+#     "kimi-coding": {"ok": true, "msg": "HTTP 200"},
+#     "local_llama": {"ok": true, "msg": "serving"},
+#     "ollama": {"ok": true, "msg": "running"},
+#     "gitea": {"ok": true, "msg": "reachable"}
+#   },
+#   "status": "healthy"
+# }
+```
+
+### Telegram Alerts
+
+When the deadman switch triggers or a VPS goes down, alerts are sent to the
+Timmy Foundation channel (`-1003664764329`). Alert format:
+
+```
+DEADMAN ALERT: No commits in 2h15m across all repos. Loops may be dead.
+Last commit: 2026-04-08 17:45
+Trigger: timmy-config + the-nexus both silent
+Action: Autonomous fallback initiated — switched to local-llama.cpp
+```
+
+---
+
+## Failure Scenarios
+
+| Scenario | Detection | Action |
+|----------|-----------|--------|
+| **Kimi API key expired / 401** | HTTP 401 from Kimi | Switch config → local llama.cpp |
+| **Kimi rate-limited / 429** | HTTP 429 | Wait, then retry; if persistent, fallback |
+| **Kimi server down / 500** | HTTP 500 | Immediate fallback to local |
+| **llama.cpp not serving** | Connection refused on :8081 | Skip to Ollama |
+| **Ollama not running** | No models list | Enter SAFE MODE (alert only) |
+| **Gitea unreachable** | 5xx or timeout | Cache work locally, retry every 5min |
+| **Bezalel VPS down** | SSH fails | Send alert, Lazarus protocol |
+| **Agent process crashed** | No PID on port 8656 | systemd restart + rollback config |
+| **Config contains Anthropic (banned)** | grep finds 'anthropic' | Rollback to known-good config |
+
+---
+
+## Recovery
+
+**Automatic Recovery** (when primary comes back):
+
+1. Health check detects Kimi API is healthy again
+2. If currently on fallback (`kimi->local-llama` in state):
+   - Restore `config.yaml` from `config.yaml.pre-fallback`
+   - Delete the backup marker
+   - Log recovery event
+   - Notify via Telegram
+3. Agent continues with restored config
+
+**Manual Override**:
+
+```bash
+# Force immediate restore (e.g. after API key renewed)
+/root/wizards/bezalel/bin/deadman-fallback.py --force-restore
+systemctl restart hermes-bezalel
+```
+
+---
+
+## Testing
+
+### Smoke Test
+
+```bash
+# 1. Verify health check passes
+cd /root/wizards/bezalel
+python3 home/.hermes/deadman-fallback-state.json  # view state
+python3 bin/deadman-fallback.py  # run once manually
+
+# 2. Simulate failure (temporarily block Kimi)
+iptables -A OUTPUT -d api.kimi.com -j DROP
+python3 bin/deadman-fallback.py  # should detect and fallback
+iptables -D OUTPUT -d api.kimi.com -j ACCEPT  # restore
+
+# 3. Verify emergency config took effect
+cat home/.hermes/config.yaml | grep provider  # should show "ollama" or "local-llama"
+```
+
+### Unit Tests
+
+```bash
+# Run test suite (if pytest available)
+pytest tests/test_deadman_switch.py -v
+
+# Or run standalone test script
+python3 /root/wizards/bezalel/test_deadman_switch.py
+```
+
+---
+
+## Troubleshooting
+
+### Deadman didn't trigger even though agent is dead
+
+Check:
+1. Is the systemd timer active? `systemctl list-timers | grep deadman`
+2. Is the health check script executable? `ls -l home/.hermes/deadman-fallback.py`
+3. Check cron is running: `systemctl status crond`
+4. Check logs: `journalctl -u deadman-bezalel.timer`
+
+### Fallback happened but agent still can't run
+
+Verify Ollama is actually installed and running:
+```bash
+curl http://localhost:11434/api/tags
+# Should return JSON with model list
+```
+
+If Ollama not installed, emergency mode will enter safe mode → alert Alexander.
+
+### Config keeps rolling back in a loop
+
+This means primary is still failing. Either:
+- Renew API key (Kimi / OpenRouter)
+- Fix network connectivity to provider
+- Adjust `heartbeat_timeout_seconds` to be longer
+- Pause the deadman: `systemctl stop deadman-bezalel.timer`
+
+### Safe mode entered — how to get out?
+
+1. Fix the root cause (restore network, renew keys, restart services)
+2. Remove emergency trigger: `rm /root/wizards/bezalel/home/.hermes/emergency_mode_trigger`
+3. Manually restore config: `cp config.yaml.pre-fallback config.yaml`
+4. Restart agent: `systemctl restart hermes-bezalel`
+
+---
+
+## Cross-Agent Coordination
+
+The dead man switch system is designed to be replicated across all wizards:
+
+- **Allegro**: Burn-mode validator
+- **Ezra**: Security & QA
+- **Adagio**: Not yet deployed
+- **Bezalel**: Infrastructure (this one — pioneer)
+
+When a VPS agent goes down, other agents are alerted via Telegram and can
+coordinate recovery through the Lazarus protocol.
+
+---
+
+## Key Design Principles
+
+1. **Poka-yoke**: Cannot misconfigure — templates enforced via Ansible
+2. **Autonomous**: No human in loop for recovery (except safe-mode escalation)
+3. **Reversible**: Every fallback records pre-fallback state; recovery is automatic
+4. **Observable**: All state written to JSON logs; health_status.json is single source of truth
+5. **Fail-secure**: In total failure, agent pauses non-essential work and alerts
+
+---
+
+## Related
+
+- `ansible/roles/deadman_switch/` — Ansible deployment logic
+- `bin/deadman-fallback.py` — Core health/fallback engine
+- `bin/deadman-switch.sh` — Commit activity monitor
+- `fleet/health_check.py` — Fleet-wide health aggregation
+- `wizards/bezalel/config.yaml` — Primary production config
+- Issue #423 — Tracking issue for this system
+
+---
+
+**Last updated:** 2026-04-08  
+**Status:** Production-ready ✅  
+**Maintainer:** Bezalel (via Ansible)  
+**On-call:** Alexander Whitestone (Rockachopa)
diff --git a/wizards/bezalel/home/.hermes/config.emergency.yaml b/wizards/bezalel/home/.hermes/config.emergency.yaml
new file mode 100644
index 00000000..251a0611
--- /dev/null
+++ b/wizards/bezalel/home/.hermes/config.emergency.yaml
@@ -0,0 +1,106 @@
+# Emergency Configuration — Bezalel Dead Man Switch Fallback
+#
+# This config is activated when the dead man switch detects primary failure
+# (Kimi Coding API down, network issues, or process crash). It provides
+# minimal local-only operation using Ollama models.
+#
+# DO NOT EDIT directly on VPS — edit the template in timmy-config repo
+# and redeploy via Ansible.
+#
+# Deployment: /root/wizards/bezalel/home/.hermes/config.emergency.yaml
+# Symlinked or copied to config.yaml when emergency mode triggers.
+
+model:
+  default: gemma3:4b
+  provider: ollama
+  temperature: 0.7
+  max_tokens: 2048
+  timeout: 300  # longer timeout for local inference
+
+fallback_providers:
+  # Emergency fallback chain — only local models
+  - provider: ollama
+    model: gemma3:4b
+    base_url: http://localhost:11434
+    timeout: 300
+    reason: Primary emergency model — small, fast, local
+  - provider: ollama
+    model: gemma3:12b
+    base_url: http://localhost:11434
+    timeout: 300
+    reason: Fallback if 4b OOM or fails
+  - provider: ollama
+    model: gemma3:27b
+    base_url: http://localhost:11434
+    timeout: 300
+    reason: Last-ditch local inference
+
+agent:
+  max_turns: 20  # Reduced context for stability
+  reasoning_effort: low
+  verbose: false
+  # Emergency system prompt — focused on recovery and basic operations
+  system_prompt: |
+    You are Bezalel operating in EMERGENCY MODE.
+    Only essential infrastructure tasks. No creative work.
+    Focus: diagnose, repair, recover, communicate status.
+    Sovereignty and service always.
+
+terminal:
+  backend: local
+  cwd: /root/wizards/bezalel
+  timeout: 300
+
+browser:
+  inactivity_timeout: 60  # shorter timeout in emergency
+
+compression:
+  enabled: true
+  threshold: 0.8  # more aggressive compression
+
+display:
+  compact: true
+  show_reasoning: false
+  streaming: false
+
+platforms:
+  # Disable non-essential platforms in emergency
+  api_server:
+    enabled: true  # Keep API alive for health checks
+    extra:
+      host: 127.0.0.1
+      port: 8656
+      key: bezalel-emergency-key-2026
+  telegram:
+    enabled: true  # Critical for alerts
+  webhook:
+    enabled: false  # Disable external webhooks in emergency
+
+gateway:
+  allow_all_users: false  # Lock down — only authorized operators
+
+session_reset:
+  mode: both
+  idle_minutes: 60  # faster session cleanup
+
+memory:
+  memory_enabled: false  # Disable memory to reduce resource usage
+  user_profile_enabled: false
+
+security:
+  redact_secrets: true
+
+# Emergency mode flag — set by deadman switch
+emergency_mode: true
+_config_version: 11
+
+# ─── Notes ────────────────────────────────────────────────────────────────────
+# This config intentionally:
+#   • Uses only Ollama (no external API calls)
+#   • Reduces max_turns to conserve memory
+#   • Disables memory subsystem
+#   • Keeps Telegram for operator alerts
+#   • Keeps API server for health checks
+#
+# Recovery: When primary service is restored, deadman-fallback.py will
+# automatically restore the pre-fallback config from backup.
diff --git a/wizards/bezalel/home/.hermes/deadman_switch.json b/wizards/bezalel/home/.hermes/deadman_switch.json
new file mode 100644
index 00000000..36466c5b
--- /dev/null
+++ b/wizards/bezalel/home/.hermes/deadman_switch.json
@@ -0,0 +1,142 @@
+{
+  "schema_version": 1,
+  "wizard": "bezalel",
+  "deployed_at": "2026-04-08T20:11:31Z",
+  "deadman_switch": {
+    "enabled": true,
+    "mode": "autonomous_fallback",  # "alert_only" | "autonomous_fallback" | "safe_mode"
+    "health_check_interval_seconds": 60,
+    "heartbeat_timeout_seconds": 300,
+    "heartbeat_grace_periods": 2,
+    "max_consecutive_failures": 3,
+    "restart_cooldown_seconds": 60,
+    "max_restart_attempts": 3,
+    "escalation": {
+      "channel": "telegram",
+      "chat_id": "-1003664764329",
+      "alert_after_failures": 2,
+      "include_diagnostics": true
+    }
+  },
+  "fallback_chain": [
+    {
+      "level": 1,
+      "name": "local_llama_cpp",
+      "provider": "local-llama.cpp",
+      "model": "hermes3",
+      "endpoint": "http://localhost:8081/v1/chat/completions",
+      "trigger_condition": "kimi_down",
+      "priority": 1,
+      "reversible": true
+    },
+    {
+      "level": 2,
+      "name": "ollama_gemma4",
+      "provider": "ollama",
+      "model": "gemma3:4b",
+      "endpoint": "http://localhost:11434/v1/chat/completions",
+      "trigger_condition": "local_llama_down",
+      "priority": 2,
+      "reversible": true
+    },
+    {
+      "level": 3,
+      "name": "ollama_gemma12b",
+      "provider": "ollama",
+      "model": "gemma3:12b",
+      "endpoint": "http://localhost:11434/v1/chat/completions",
+      "trigger_condition": "ollama_4b_failure",
+      "priority": 3,
+      "reversible": true
+    },
+    {
+      "level": 4,
+      "name": "ollama_gemma27b",
+      "provider": "ollama",
+      "model": "gemma3:27b",
+      "endpoint": "http://localhost:11434/v1/chat/completions",
+      "trigger_condition": "ollama_12b_failure",
+      "priority": 4,
+      "reversible": true
+    }
+  ],
+  "emergency_mode": {
+    "enabled": true,
+    "trigger_conditions": [
+      "all_inference_offline",
+      "health_check_exhausted",
+      "operator_override"
+    ],
+    "actions": [
+      "pause_nonessential_cron",
+      "cache_work_locally",
+      "enable_alert_mode",
+      "preserve_state"
+    ],
+    "telegram_only_mode": true,
+    "allow_local_inference": true
+  },
+  "recovery": {
+    "auto_restore_on_primary_up": true,
+    "verify_primary_health_before_restore": true,
+    "require_stable_period_seconds": 300,
+    "notify_on_recovery": true,
+    "preserve_audit_trail": true
+  },
+  "snapshot": {
+    "enabled": true,
+    "directory": "~/.local/timmy/snapshots",
+    "max_snapshots": 10,
+    "prune_on_startup": true,
+    "known_good_filename": "config.yaml.known_good"
+  },
+  "vps_monitoring": {
+    "enabled": true,
+    "targets": [
+      {"name": "allegro", "ip": "167.99.126.228", "port": 22},
+      {"name": "ezra", "ip": "143.198.27.163", "port": 22},
+      {"name": "bezalel_self", "ip": "159.203.146.185", "port": 22}
+    ],
+    "check_interval_seconds": 120,
+    "ssh_timeout_seconds": 10,
+    "alert_on_ssh_failure": true
+  },
+  "logging": {
+    "dir": "~/.hermes/logs",
+    "deadman_log": "deadman-bezalel.log",
+    "health_log": "health-bezalel.log",
+    "rotation_days": 30,
+    "min_level": "INFO"
+  },
+  "paths": {
+    "hermes_home": "~/.hermes",
+    "wizard_home": "/root/wizards/bezalel",
+    "config": "home/.hermes/config.yaml",
+    "config_emergency": "home/.hermes/config.emergency.yaml",
+    "config_backup": "home/.hermes/config.yaml.pre-fallback",
+    "state_file": "home/.hermes/deadman-fallback-state.json",
+    "health_status": "home/.hermes/health_status.json"
+  }
+}
+
+# ─── Deadman Switch Configuration ────────────────────────────────────────────
+#
+# This file defines the dead man switch behavior for the Bezalel wizard.
+# It is read by deadman-fallback.py and deadman-switch.sh to determine
+# thresholds, fallback actions, and recovery behavior.
+#
+# Key settings:
+#   - health_check_interval_seconds: How often to run diagnostics (default 60s)
+#   - heartbeat_timeout_seconds: How long to wait before declaring primary dead (300s = 5min)
+#   - max_restart_attempts: How many times to retry before escalating
+#   - restart_cooldown_seconds: Cooldown between restart attempts
+#   - fallback_chain: Ordered list of fallback providers/models
+#   - emergency_mode: Actions taken when all inference fails
+#   - recovery: Auto-restore behavior when primary recovers
+#   - snapshot: Config backup strategy
+#   - vps_monitoring: SSH-based VPS health checks
+#
+# This JSON file is deployment-time configuration. The Python script uses
+# these values to make autonomous decisions about failover and recovery.
+#
+# Location: /root/wizards/bezalel/home/.hermes/deadman_switch.json
diff --git a/wizards/bezalel/home/.hermes/health_status.json b/wizards/bezalel/home/.hermes/health_status.json
new file mode 100644
index 00000000..63a6a9a3
--- /dev/null
+++ b/wizards/bezalel/home/.hermes/health_status.json
@@ -0,0 +1,138 @@
+{
+  "schema_version": 1,
+  "generated_at": "2026-04-08T20:11:31Z",
+  "wizard": "bezalel",
+  "checks": {
+    "kimi-coding": {
+      "status": "unknown",
+      "last_check": null,
+      "failure_count": 0,
+      "success_count": 0,
+      "last_success": null,
+      "last_failure": null,
+      "average_latency_ms": null,
+      "notes": "Primary inference provider"
+    },
+    "local_llama": {
+      "status": "unknown",
+      "last_check": null,
+      "failure_count": 0,
+      "success_count": 0,
+      "last_success": null,
+      "last_failure": null,
+      "average_latency_ms": null,
+      "endpoint": "http://localhost:8081/v1/models",
+      "notes": "Local llama.cpp via Hermes harness"
+    },
+    "ollama": {
+      "status": "unknown",
+      "last_check": null,
+      "failure_count": 0,
+      "success_count": 0,
+      "last_success": null,
+      "last_failure": null,
+      "average_latency_ms": null,
+      "endpoint": "http://localhost:11434/api/tags",
+      "notes": "Fallback local inference"
+    },
+    "gitea": {
+      "status": "unknown",
+      "last_check": null,
+      "failure_count": 0,
+      "success_count": 0,
+      "last_success": null,
+      "last_failure": null,
+      "average_latency_ms": null,
+      "endpoint": "https://forge.alexanderwhitestone.com/api/v1/user",
+      "notes": "Forge connectivity"
+    },
+    "vps_allegro": {
+      "status": "unknown",
+      "host": "167.99.126.228",
+      "last_check": null,
+      "failure_count": 0,
+      "success_count": 0,
+      "last_success": null,
+      "last_failure": null,
+      "notes": "SSH connectivity check"
+    },
+    "vps_ezra": {
+      "status": "unknown",
+      "host": "143.198.27.163",
+      "last_check": null,
+      "failure_count": 0,
+      "success_count": 0,
+      "last_success": null,
+      "last_failure": null,
+      "notes": "SSH connectivity check"
+    },
+    "vps_bezalel": {
+      "status": "unknown",
+      "host": "159.203.146.185",
+      "last_check": null,
+      "failure_count": 0,
+      "success_count": 0,
+      "last_success": null,
+      "last_failure": null,
+      "notes": "SSH connectivity check"
+    }
+  },
+  "current_provider": "kimi-coding",
+  "fallback_stack": [],
+  "in_emergency_mode": false,
+  "emergency_mode_since": null,
+  "emergency_reason": null,
+  "last_config_backup": null,
+  "deadman_switch": {
+    "enabled": true,
+    "last_heartbeat": null,
+    "heartbeat_timeout_seconds": 300,
+    "monitoring_interval_seconds": 60,
+    "max_restart_attempts": 3,
+    "restart_cooldown_seconds": 60,
+    "escalation_channel": "telegram"
+  },
+  "agent_process": {
+    "running": false,
+    "pid": null,
+    "uptime_seconds": null,
+    "memory_mb": null,
+    "restart_count": 0,
+    "last_restart": null
+  },
+  "consecutive_failures": 0,
+  "last_status_change": null,
+  "next_check": null
+}
+
+# ─── Health Status Schema ─────────────────────────────────────────────────────
+#
+# This file is written by deadman-fallback.py on every health check cycle.
+# It can be read by operators to understand current system state.
+#
+# Fields:
+#   schema_version: int — Increment when structure changes
+#   generated_at: ISO8601 timestamp
+#   wizard: str — "bezalel"
+#   checks: dict — Per-system check results with:
+#     - status: "healthy" | "degraded" | "down" | "unknown"
+#     - last_check: ISO8601 timestamp of last check
+#     - failure_count / success_count: int
+#     - last_success / last_failure: ISO8601 timestamps
+#     - average_latency_ms: float | null
+#     - notes: str
+#   current_provider: str — name of active inference provider
+#   fallback_stack: list[str] — active fallback chain
+#   in_emergency_mode: bool
+#   emergency_mode_since: ISO8601 | null
+#   emergency_reason: str | null  ("kimi_down", "gitea_down", "all_inference_down")
+#   last_config_backup: ISO8601 | null
+#   deadman_switch: dict — switch configuration
+#   agent_process: dict — process state (running, pid, uptime, memory)
+#   consecutive_failures: int
+#   last_status_change: ISO8601 | null
+#   next_check: ISO8601 | null
+#
+# Location: /root/wizards/bezalel/home/.hermes/health_status.json
+# Updated by: bin/deadman-fallback.py (every cycle)
+# Consumed by: operators, monitoring dashboards, deadman-switch.sh