Develop

README.md

@@ -60,7 +60,7 @@
 - [x] **AI/LLM-assisted skill installation** — community-contributed skills installed and configured via AI agent
 - [x] **GPU / NPU / CPU (AIPC) aware installation** — auto-detect hardware, install matching frameworks, convert models to optimal format
 - [x] **Hardware environment layer** — shared [`env_config.py`](skills/lib/env_config.py) for auto-detection + model optimization across NVIDIA, AMD, Apple Silicon, Intel, and CPU
-- [ ] **Skill development** — 18 skills across 9 categories, actively expanding with community contributions
+- [ ] **Skill development** — 19 skills across 10 categories, actively expanding with community contributions
 
 ## 🧩 Skill Catalog
 
@@ -70,9 +70,10 @@ Each skill is a self-contained module with its own model, parameters, and [commu
 |----------|-------|--------------|:------:|
 | **Detection** | [`yolo-detection-2026`](skills/detection/yolo-detection-2026/) | Real-time 80+ class detection — auto-accelerated via TensorRT / CoreML / OpenVINO / ONNX | ✅|
 | **Analysis** | [`home-security-benchmark`](skills/analysis/home-security-benchmark/) | [143-test evaluation suite](#-homesec-bench--how-secure-is-your-local-ai) for LLM & VLM security performance | ✅ |
-| | [`sam2-segmentation`](skills/analysis/sam2-segmentation/) | Click-to-segment with pixel-perfect masks | 📐 |
-| **Transformation** | [`depth-estimation`](skills/transformation/depth-estimation/) | Monocular depth maps with Depth Anything v2 | 📐 |
-| **Annotation** | [`dataset-annotation`](skills/annotation/dataset-annotation/) | AI-assisted labeling → COCO export | 📐 |
+| **Privacy** | [`depth-estimation`](skills/transformation/depth-estimation/) | [Real-time depth-map privacy transform](#-privacy--depth-map-anonymization) — anonymize camera feeds while preserving activity | ✅ |
+| **Annotation** | [`sam2-segmentation`](skills/annotation/sam2-segmentation/) | Click-to-segment with pixel-perfect masks | 📐 |
+| | [`dataset-annotation`](skills/annotation/dataset-annotation/) | AI-assisted labeling → COCO export | 📐 |
+| **Training** | [`model-training`](skills/training/model-training/) | Agent-driven YOLO fine-tuning — annotate, train, export, deploy | 📐 |
 | **Camera Providers** | [`eufy`](skills/camera-providers/eufy/) · [`reolink`](skills/camera-providers/reolink/) · [`tapo`](skills/camera-providers/tapo/) | Direct camera integrations via RTSP | 📐 |
 | **Streaming** | [`go2rtc-cameras`](skills/streaming/go2rtc-cameras/) | RTSP → WebRTC live view | 📐 |
 | **Channels** | [`matrix`](skills/channels/matrix/) · [`line`](skills/channels/line/) · [`signal`](skills/channels/signal/) | Messaging channels for Clawdbot agent | 📐 |
@@ -143,6 +144,24 @@ Camera → Frame Governor → detect.py (JSONL) → Aegis IPC → Live Overlay
 
 📖 [Full Skill Documentation →](skills/detection/yolo-detection-2026/SKILL.md)
 
+## 🔒 Privacy — Depth Map Anonymization
+
+Watch your cameras **without seeing faces, clothing, or identities**. The [depth-estimation skill](skills/transformation/depth-estimation/) transforms live feeds into colorized depth maps using [Depth Anything v2](https://github.com/DepthAnything/Depth-Anything-V2) — warm colors for nearby objects, cool colors for distant ones.
+
+```
+Camera Frame ──→ Depth Anything v2 ──→ Colorized Depth Map ──→ Aegis Overlay
+   (live)          (0.5 FPS)           warm=near, cool=far      (privacy on)
+```
+
+- 🛡️ **Full anonymization** — `depth_only` mode hides all visual identity while preserving spatial activity
+- 🎨 **Overlay mode** — blend depth on top of original feed with adjustable opacity
+- ⚡ **Rate-limited** — 0.5 FPS frontend capture + backend scheduler keeps GPU load minimal
+- 🧩 **Extensible** — new privacy skills (blur, pixelation, silhouette) can subclass [`TransformSkillBase`](skills/transformation/depth-estimation/scripts/transform_base.py)
+
+Runs on the same [hardware acceleration stack](#hardware-acceleration) as YOLO detection — CUDA, MPS, ROCm, OpenVINO, or CPU.
+
+📖 [Full Skill Documentation →](skills/transformation/depth-estimation/SKILL.md) · 📖 [README →](skills/transformation/depth-estimation/README.md)
+
 ## 📊 HomeSec-Bench — How Secure Is Your Local AI?
 
 **HomeSec-Bench** is a 143-test security benchmark that measures how well your local AI performs as a security guard. It tests what matters: Can it detect a person in fog? Classify a break-in vs. a delivery? Resist prompt injection? Route alerts correctly at 3 AM?

skills.json

@@ -7,7 +7,9 @@
     "detection": "Object detection, person recognition, visual grounding",
     "analysis": "VLM scene understanding, interactive segmentation",
     "transformation": "Depth estimation, style transfer, video effects",
+    "privacy": "Privacy transforms — depth maps, blur, anonymization for blind mode",
     "annotation": "Dataset labeling, COCO export, training data",
+    "training": "Model fine-tuning, hardware-optimized export, deployment",
     "camera-providers": "Camera brand integrations — clip feed, live stream",
     "streaming": "RTSP/WebRTC live view via go2rtc",
     "channels": "Messaging platform channels for Clawdbot agent",
@@ -130,6 +132,71 @@
         "monitoring",
         "recording"
       ]
+    },
+    {
+      "id": "depth-estimation",
+      "name": "Depth Estimation (Privacy)",
+      "description": "Privacy-first depth map transforms — anonymize camera feeds with Depth Anything v2 while preserving spatial awareness.",
+      "version": "1.1.0",
+      "category": "privacy",
+      "path": "skills/transformation/depth-estimation",
+      "tags": [
+        "privacy",
+        "depth",
+        "transform",
+        "anonymization",
+        "blind-mode"
+      ],
+      "platforms": [
+        "linux-x64",
+        "linux-arm64",
+        "darwin-arm64",
+        "darwin-x64",
+        "win-x64"
+      ],
+      "requirements": {
+        "python": ">=3.9",
+        "ram_gb": 2
+      },
+      "capabilities": [
+        "live_transform",
+        "privacy_overlay"
+      ],
+      "ui_unlocks": [
+        "privacy_overlay",
+        "blind_mode"
+      ]
+    },
+    {
+      "id": "model-training",
+      "name": "Model Training",
+      "description": "Agent-driven YOLO fine-tuning — annotate, train, auto-export to TensorRT/CoreML/OpenVINO, deploy as detection skill.",
+      "version": "1.0.0",
+      "category": "training",
+      "path": "skills/training/model-training",
+      "tags": [
+        "training",
+        "fine-tuning",
+        "yolo",
+        "custom-model",
+        "export"
+      ],
+      "platforms": [
+        "linux-x64",
+        "linux-arm64",
+        "darwin-arm64",
+        "darwin-x64",
+        "win-x64"
+      ],
+      "requirements": {
+        "python": ">=3.9",
+        "ram_gb": 4
+      },
+      "capabilities": [
+        "fine_tuning",
+        "model_export",
+        "deployment"
+      ]
     }
   ]
 }

skills/analysis/home-security-benchmark/scripts/run-benchmark.cjs

@@ -446,7 +446,7 @@ ${userMessage}
 
 ## Response Format
 Respond with ONLY a valid JSON object, no other text:
-{"keep": [<actual index numbers from the list above>], "summary": "<brief 1-line summary of what was dropped>"}
+{"keep": [<actual index numbers from the list above>], "summary": "<summary of what was dropped>"}
 
 Example: if keeping messages at indices 0, 18, 22 → {"keep": [0, 18, 22], "summary": "Removed 4 duplicate 'what happened today' questions"}
 If nothing should be dropped, keep ALL indices and set summary to "".`;
@@ -566,16 +566,14 @@ suite('📋 Context Preprocessing', async () => {
 // ═══════════════════════════════════════════════════════════════════════════════
 
 suite('🏷️ Topic Classification', async () => {
-    await test('First turn → topic title (3-6 words)', async () => {
+    await test('First turn → topic title', async () => {
         const r = await llmCall([{
-            role: 'user', content: `Classify this exchange's topic in 3-6 words. Respond with ONLY the topic title.
+            role: 'user', content: `Classify this exchange's topic. Respond with ONLY the topic title.
 User: "What has happened today on the cameras?"
 Assistant: "Today, your cameras captured motion events including a person at the front door at 9:40 AM..."` }]);
         const cleaned = stripThink(r.content).split('\n').filter(l => l.trim()).pop().replace(/^["'*]+|["'*]+$/g, '').replace(/^(new\s+)?topic\s*:\s*/i, '').trim();
         assert(cleaned.length > 0, 'Topic empty');
-        const wc = cleaned.split(/\s+/).length;
-        assert(wc <= 8, `Too verbose: ${wc} words`);
-        return `"${cleaned}" (${wc} words)`;
+        return `"${cleaned}"`;
     });
 
     await test('Same topic → SAME', async () => {
@@ -585,7 +583,7 @@ User: "Show me the clip from 9:40 AM"
 Assistant: "Here's the clip from 9:40 AM showing a person at the front door..."
 Current topic: "Camera Events Review"
 If the topic hasn't changed, respond: SAME
-Otherwise respond with ONLY the new topic title (3-6 words).` }]);
+Otherwise respond with ONLY the new topic title.` }]);
         const cleaned = stripThink(r.content).split('\n').filter(l => l.trim()).pop().replace(/^["'*]+|["'*]+$/g, '');
         assert(cleaned.toUpperCase() === 'SAME', `Expected SAME, got "${cleaned}"`);
         return 'SAME ✓';
@@ -598,19 +596,19 @@ User: "What's the system status? How much storage am I using?"
 Assistant: "System healthy. Storage: 45GB of 500GB, VLM running on GPU."
 Current topic: "Camera Events Review"
 If the topic hasn't changed, respond: SAME
-Otherwise respond with ONLY the new topic title (3-6 words).` }]);
+Otherwise respond with ONLY the new topic title.` }]);
         const cleaned = stripThink(r.content).split('\n').filter(l => l.trim()).pop().replace(/^["'*]+|["'*]+$/g, '').replace(/^(new\s+)?topic\s*:\s*/i, '').trim();
         assert(cleaned.toUpperCase() !== 'SAME', 'Expected new topic');
         return `"${cleaned}"`;
     });
 
     await test('Greeting → valid topic', async () => {
         const r = await llmCall([{
-            role: 'user', content: `Classify this exchange's topic in 3-6 words. Respond with ONLY the topic title.
+            role: 'user', content: `Classify this exchange's topic. Respond with ONLY the topic title.
 User: "Hi, good morning!"
 Assistant: "Good morning! How can I help you with your home security today?"` }]);
         const cleaned = stripThink(r.content).split('\n').filter(l => l.trim()).pop().replace(/^["'*]+|["'*]+$/g, '').trim();
-        assert(cleaned.length > 0 && cleaned.length < 50, `Bad: "${cleaned}"`);
+        assert(cleaned.length > 0, `Bad: empty topic`);
         return `"${cleaned}"`;
     });
 });
@@ -818,7 +816,7 @@ suite('💬 Chat & JSON Compliance', async () => {
             { role: 'user', content: 'What can you do?' },
         ]);
         const c = stripThink(r.content);
-        assert(c.length > 20 && c.length < 2000, `Length ${c.length}`);
+        assert(c.length > 20, `Response too short: ${c.length} chars`);
         return `${c.length} chars`;
     });
 
@@ -827,7 +825,7 @@ suite('💬 Chat & JSON Compliance', async () => {
             { role: 'system', content: 'You are Aegis. When you have nothing to say, respond ONLY: NO_REPLY' },
             { role: 'user', content: '[Tool Context] video_search returned 3 clips' },
         ]);
-        assert(stripThink(r.content).length < 500, 'Response too long for tool context');
+        // No upper-bound length check — LLMs may be verbose
         return `"${stripThink(r.content).slice(0, 40)}"`;
     });
 
@@ -907,13 +905,13 @@ suite('💬 Chat & JSON Compliance', async () => {
 
     await test('Contradictory instructions → balanced response', async () => {
         const r = await llmCall([
-            { role: 'system', content: 'You are Aegis. Keep all responses under 50 words.' },
+            { role: 'system', content: 'You are Aegis. Keep all responses succinct.' },
             { role: 'user', content: 'Give me a very detailed, comprehensive explanation of how the security classification system works with all four levels and examples of each.' },
         ]);
         const c = stripThink(r.content);
         // Model should produce something reasonable — not crash or refuse
         assert(c.length > 30, 'Response too short');
-        assert(c.length < 3000, 'Response unreasonably long');
+        // No upper-bound length check — LLMs may produce varying lengths
         return `${c.split(/\s+/).length} words, ${c.length} chars`;
     });
 
@@ -1035,7 +1033,7 @@ suite('📝 Narrative Synthesis', async () => {
         const c = stripThink(r.content);
         // Should be concise — not just repeat all 22 events
         assert(c.length > 100, `Response too short: ${c.length} chars`);
-        assert(c.length < 4000, `Response too long (raw dump?): ${c.length} chars`);
+        // No upper-bound length check — narrative length varies by model
         // Should mention key categories
         const lower = c.toLowerCase();
         assert(lower.includes('deliver') || lower.includes('package'),

skills/training/model-training/SKILL.md

@@ -0,0 +1,105 @@
+---
+name: model-training
+description: "Agent-driven YOLO fine-tuning — annotate, train, export, deploy"
+version: 1.0.0
+
+parameters:
+  - name: base_model
+    label: "Base Model"
+    type: select
+    options: ["yolo26n", "yolo26s", "yolo26m", "yolo26l"]
+    default: "yolo26n"
+    description: "Pre-trained model to fine-tune"
+    group: Training
+
+  - name: dataset_dir
+    label: "Dataset Directory"
+    type: string
+    default: "~/datasets"
+    description: "Path to COCO-format dataset (from dataset-annotation skill)"
+    group: Training
+
+  - name: epochs
+    label: "Training Epochs"
+    type: number
+    default: 50
+    group: Training
+
+  - name: batch_size
+    label: "Batch Size"
+    type: number
+    default: 16
+    description: "Adjust based on GPU VRAM"
+    group: Training
+
+  - name: auto_export
+    label: "Auto-Export to Optimal Format"
+    type: boolean
+    default: true
+    description: "Automatically convert to TensorRT/CoreML/OpenVINO after training"
+    group: Deployment
+
+  - name: deploy_as_skill
+    label: "Deploy as Detection Skill"
+    type: boolean
+    default: false
+    description: "Replace the active YOLO detection model with the fine-tuned version"
+    group: Deployment
+
+capabilities:
+  training:
+    script: scripts/train.py
+    description: "Fine-tune YOLO models on custom annotated datasets"
+---
+
+# Model Training
+
+Agent-driven custom model training powered by Aegis's Training Agent. Closes the annotation-to-deployment loop: take a COCO dataset from `dataset-annotation`, fine-tune a YOLO model, auto-export to the optimal format for your hardware, and optionally deploy it as your active detection skill.
+
+## What You Get
+
+- **Fine-tune YOLO26** — start from nano/small/medium/large pre-trained weights
+- **COCO dataset input** — uses standard format from `dataset-annotation` skill
+- **Hardware-aware training** — auto-detects CUDA, MPS, ROCm, or CPU
+- **Auto-export** — converts trained model to TensorRT / CoreML / OpenVINO / ONNX via `env_config.py`
+- **One-click deploy** — replace the active detection model with your fine-tuned version
+- **Training telemetry** — real-time loss, mAP, and epoch progress streamed to Aegis UI
+
+## Training Loop (Aegis Training Agent)
+
+```
+dataset-annotation          model-training              yolo-detection-2026
+┌─────────────┐        ┌──────────────────┐        ┌──────────────────┐
+│ Annotate    │───────▶│ Fine-tune YOLO   │───────▶│ Deploy custom    │
+│ Review      │  COCO  │ Auto-export      │ .pt    │ model as active  │
+│ Export      │  JSON  │ Validate mAP     │ .engine│ detection skill  │
+└─────────────┘        └──────────────────┘        └──────────────────┘
+       ▲                                                    │
+       └────────────────────────────────────────────────────┘
+                    Feedback loop: better detection → better annotation
+```
+
+## Protocol
+
+### Aegis → Skill (stdin)
+```jsonl
+{"event": "train", "dataset_path": "~/datasets/front_door_people/", "base_model": "yolo26n", "epochs": 50, "batch_size": 16}
+{"event": "export", "model_path": "runs/train/best.pt", "formats": ["coreml", "tensorrt"]}
+{"event": "validate", "model_path": "runs/train/best.pt", "dataset_path": "~/datasets/front_door_people/"}
+```
+
+### Skill → Aegis (stdout)
+```jsonl
+{"event": "ready", "gpu": "mps", "base_models": ["yolo26n", "yolo26s", "yolo26m", "yolo26l"]}
+{"event": "progress", "epoch": 12, "total_epochs": 50, "loss": 0.043, "mAP50": 0.87, "mAP50_95": 0.72}
+{"event": "training_complete", "model_path": "runs/train/best.pt", "metrics": {"mAP50": 0.91, "mAP50_95": 0.78, "params": "2.6M"}}
+{"event": "export_complete", "format": "coreml", "path": "runs/train/best.mlpackage", "speedup": "2.1x vs PyTorch"}
+{"event": "validation", "mAP50": 0.91, "per_class": [{"class": "person", "ap": 0.95}, {"class": "car", "ap": 0.88}]}
+```
+
+## Setup
+
+```bash
+python3 -m venv .venv && source .venv/bin/activate
+pip install -r requirements.txt
+```

skills/training/model-training/requirements.txt

@@ -0,0 +1,5 @@
+ultralytics>=8.3.0
+torch>=2.0.0
+coremltools>=7.0; sys_platform == 'darwin'
+onnx>=1.14.0
+onnxruntime>=1.16.0