diff --git a/AGENTS.md b/AGENTS.md
index 810cf0d..51130f7 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -79,6 +79,8 @@ src/
 │   └── recovery.rs     # Error recovery and user guidance
 ├── streaming/          # HLS pipeline
 │   ├── hls_generator.rs    # FFmpeg subprocess per camera (HLS muxer)
+│   ├── supervisor.rs       # Polls FFmpeg every 2s, respawns with exponential backoff
+│   │                       # (1s→30s), reports Streaming/Restarting/Failed into Dashboard
 │   ├── hls_uploader.rs     # Watches HLS dir, drives playlist updates + motion event channel
 │   ├── segment_uploader.rs # Posts each .ts to POST /push-segment with retry/backoff
 │   ├── motion_detector.rs  # Parallel FFmpeg scene-change scorer
@@ -100,7 +102,7 @@ src/
 2. Detect cameras (`camera::detect_cameras()`)
 3. Register with Command Center (`api_client.register()`)
 4. Detect hardware encoder once (NVENC/QSV/AMF), persist to DB
-5. Create HLS generator per camera (FFmpeg subprocess)
+5. Spawn an FFmpeg **supervisor** per camera (`streaming/supervisor.rs`) that owns the `HlsGenerator`, polls the child every 2s, respawns it with exponential backoff (1s → 2s → 4s → … capped at 30s) when it dies, and trips the camera into `Failed` state if the backoff window sees 5+ crashes in 60s. Each transition (`Starting` / `Streaming` / `Restarting { reason }` / `Failed { reason }`) is pushed into the `Dashboard` so the heartbeat / WS messages carry the real pipeline state (with `last_error`) rather than the old hardcoded `"streaming"`.
 6. Spawn HLS uploader tasks (segment push + playlist update + codec detection)
 7. Spawn motion detector per camera (second FFmpeg probe for scene-change scoring)
 8. Launch local HTTP server (port 8080) + WebSocket client
@@ -182,8 +184,8 @@ All outbound calls use `ApiClient` in `src/api/client.rs`:
 
 | Method | Path | Header | Body | When |
 |--------|------|--------|------|------|
-| POST | `/api/nodes/register` | `X-API-Key` | `RegisterRequest` JSON | Startup |
-| POST | `/api/nodes/heartbeat` | `X-API-Key` | `HeartbeatRequest` JSON | Every `heartbeat_interval` s (fallback path; WS heartbeat is primary) |
+| POST | `/api/nodes/register` | `X-Node-API-Key` | `RegisterRequest` JSON | Startup |
+| POST | `/api/nodes/heartbeat` | `X-Node-API-Key` | `HeartbeatRequest` JSON (includes per-camera `CameraStatus { camera_id, status, last_error }` with real pipeline state) | Every `heartbeat_interval` s (fallback path; WS heartbeat is primary) |
 | POST | `/api/cameras/{id}/codec` | `X-Node-API-Key` | `{video_codec, audio_codec}` JSON | After first segment or codec change |
 | POST | `/api/cameras/{id}/push-segment?filename=…` | `X-Node-API-Key` | raw `.ts` bytes (`video/mp2t`) | Every segment |
 | POST | `/api/cameras/{id}/playlist` | `X-Node-API-Key` | playlist text (`text/plain`) | Every playlist rewrite |
diff --git a/README.md b/README.md
index f1d791d..a34ec35 100644
--- a/README.md
+++ b/README.md
@@ -244,6 +244,8 @@ docker run -d \
 
 **Hardware encoding:** At startup, CloudNode probes for a hardware encoder (NVENC, QSV, AMF) and caches the result in the database. Falls back to `libx264` if none is found.
 
+**FFmpeg supervisor:** Each camera's FFmpeg pipeline is wrapped in a supervisor (`streaming/supervisor.rs`) that polls the child every 2 seconds. If FFmpeg exits (disk full, V4L2 disconnect, segment-writer failure, etc.) the supervisor respawns it with exponential backoff — 1s → 2s → 4s → … capped at 30s. A pipeline that crashes 5+ times inside a 60-second window is flagged `Failed` and stops retrying, so a permanently broken camera can't spin forever. Real pipeline state (`starting` / `streaming` / `restarting` / `failed`) is reported on every heartbeat (HTTP and WebSocket) together with the human-readable failure reason, which is what the Command Center dashboard surfaces — it replaces the legacy hardcoded `"streaming"` status that used to make every node look healthy regardless of what FFmpeg was doing.
+
 ---
 
 ## API Endpoints
@@ -317,6 +319,7 @@ src/
 ├── setup/                # Interactive TUI setup wizard (crossterm + inquire)
 ├── streaming/            # HLS pipeline
 │   ├── hls_generator.rs   # FFmpeg subprocess per camera (HLS muxer)
+│   ├── supervisor.rs      # Polls FFmpeg, respawns with exponential backoff, reports real pipeline state
 │   ├── hls_uploader.rs    # Watches HLS dir, hands segments to SegmentUploader, updates playlist, drives motion events
 │   ├── segment_uploader.rs# Posts each .ts to POST /push-segment with retry
 │   ├── motion_detector.rs # Parallel FFmpeg scene-change scorer
diff --git a/config.example.yaml b/config.example.yaml
index 2693da8..56842ad 100644
--- a/config.example.yaml
+++ b/config.example.yaml
@@ -27,18 +27,50 @@ cameras:
 streaming:
   # Target frames per second
   fps: 30
-  
+
   # JPEG quality (1-100)
   jpeg_quality: 85
 
+  # Video encoder override. Leave empty for auto-detect (NVENC/QSV/AMF/libx264).
+  # Examples: "h264_nvenc", "h264_qsv", "h264_amf", "libx264"
+  encoder: ""
+
+  # HLS muxer settings
+  hls:
+    # Enable HLS streaming (must be true for cloud streaming)
+    enabled: true
+
+    # Segment duration in seconds
+    segment_duration: 1
+
+    # Number of segments kept in the rolling playlist
+    playlist_size: 15
+
+    # Video bitrate (e.g. "2500k")
+    bitrate: "2500k"
+
+# Motion detection (second FFmpeg process per camera runs a scene-change filter)
+motion:
+  # Enable motion detection
+  enabled: true
+
+  # Scene-change threshold (0.0 = identical frames, 1.0 = totally different).
+  # Lower values are more sensitive — 0.02 is a good default for typical indoor
+  # lighting. Raise if you get noise from compression/autoexposure flicker.
+  threshold: 0.02
+
+  # Minimum seconds between reported motion events per camera (cooldown).
+  cooldown_secs: 30
+
 # Recording settings
 recording:
   # Enable local recording
   enabled: true
-  
+
   # Recording format: "mp4" or "mkv"
   format: "mp4"
-  
+
+
 storage:
   # Base path for recordings and snapshots
   path: "./data"