test: add tests for describe_image, describe_image_sharegpt, generate_training_pairs

Addresses review item #5 (test coverage gaps):
- describe_image: oversize skip, valid JSON parse, truncated JSON repair, fallback
- describe_image_sharegpt: oversize skip, natural language return
- generate_training_pairs: JSONL format, ShareGPT format, dry run

bin/glitch_patterns.py

@@ -31,6 +31,14 @@ class GlitchCategory(Enum):
     WATER_REFLECTION = "water_reflection"
     SKYBOX_SEAM = "skybox_seam"
 
+    # Three.js-specific categories (ref: timmy-config#543)
+    SHADER_FAILURE = "shader_failure"
+    TEXTURE_PLACEHOLDER = "texture_placeholder"
+    UV_MAPPING_ERROR = "uv_mapping_error"
+    FRUSTUM_CULLING = "frustum_culling"
+    SHADOW_MAP_ARTIFACT = "shadow_map_artifact"
+    BLOOM_OVERFLOW = "bloom_overflow"
+
 
 @dataclass
 class GlitchPattern:
@@ -241,6 +249,123 @@ MATRIX_GLITCH_PATTERNS: list[GlitchPattern] = [
         ],
         confidence_threshold=0.45,
     ),
+
+    # --- Three.js-Specific Glitch Patterns (ref: timmy-config#543) ---
+    GlitchPattern(
+        category=GlitchCategory.SHADER_FAILURE,
+        name="Shader Compilation Failure",
+        description="Three.js shader failed to compile, rendering the material as solid black. "
+                    "Common when custom ShaderMaterial has syntax errors or missing uniforms.",
+        severity=GlitchSeverity.CRITICAL,
+        detection_prompts=[
+            "Look for objects or surfaces rendered as pure black (#000000) that should have visible textures or materials.",
+            "Identify geometry that appears completely dark while surrounding objects are normally lit.",
+            "Check for objects where the material seems to 'absorb all light' — flat black with no shading gradient.",
+        ],
+        visual_indicators=[
+            "solid black object with no shading",
+            "geometry rendered as silhouette",
+            "material appears to absorb light entirely",
+            "black patch inconsistent with scene lighting",
+        ],
+        confidence_threshold=0.7,
+    ),
+    GlitchPattern(
+        category=GlitchCategory.TEXTURE_PLACEHOLDER,
+        name="Three.js Texture Not Loaded",
+        description="Three.js failed to load the texture asset, rendering a 1x1 white pixel "
+                    "stretched across the entire surface. Distinguished from missing-texture by "
+                    "the uniform white/grey appearance rather than magenta.",
+        severity=GlitchSeverity.CRITICAL,
+        detection_prompts=[
+            "Look for surfaces that are uniformly white or light grey with no texture detail, even on large geometry.",
+            "Identify objects where the texture appears as a single solid color stretched across complex UVs.",
+            "Check for surfaces that look 'blank' or 'unloaded' — flat white/grey where detail should exist.",
+        ],
+        visual_indicators=[
+            "uniform white or light grey surface",
+            "no texture detail on large geometry",
+            "stretched single-color appearance",
+            "1x1 pixel placeholder stretched to fill UV space",
+        ],
+        confidence_threshold=0.65,
+    ),
+    GlitchPattern(
+        category=GlitchCategory.UV_MAPPING_ERROR,
+        name="BufferGeometry UV Mapping Error",
+        description="Three.js BufferGeometry has incorrect UV coordinates, causing textures to "
+                    "appear stretched, compressed, or mapped to the wrong faces.",
+        severity=GlitchSeverity.HIGH,
+        detection_prompts=[
+            "Look for textures that appear dramatically stretched in one direction on specific faces.",
+            "Identify surfaces where the texture pattern is distorted but other nearby surfaces look correct.",
+            "Check for faces where the texture seems 'smeared' or mapped with incorrect aspect ratio.",
+        ],
+        visual_indicators=[
+            "texture stretching on specific faces",
+            "distorted pattern on geometry",
+            "smeared texture appearance",
+            "aspect ratio mismatch between texture and surface",
+        ],
+        confidence_threshold=0.6,
+    ),
+    GlitchPattern(
+        category=GlitchCategory.FRUSTUM_CULLING,
+        name="Frustum Culling Artifact",
+        description="Three.js frustum culling incorrectly marks objects as outside the camera "
+                    "frustum, causing them to pop in/out of existence at screen edges.",
+        severity=GlitchSeverity.MEDIUM,
+        detection_prompts=[
+            "Look for objects that are partially visible at the edge of the frame — half-rendered or cut off unnaturally.",
+            "Identify geometry that seems to 'pop' into existence as the view angle changes.",
+            "Check screen edges for objects that appear suddenly rather than smoothly entering the viewport.",
+        ],
+        visual_indicators=[
+            "half-visible object at screen edge",
+            "object popping into frame",
+            "abrupt appearance of geometry",
+            "bounding box visible but mesh missing",
+        ],
+        confidence_threshold=0.55,
+    ),
+    GlitchPattern(
+        category=GlitchCategory.SHADOW_MAP_ARTIFACT,
+        name="Shadow Map Resolution Artifact",
+        description="Three.js shadow map has insufficient resolution, causing pixelated, "
+                    "blocky shadows with visible texel edges instead of smooth shadow gradients.",
+        severity=GlitchSeverity.MEDIUM,
+        detection_prompts=[
+            "Look for shadows with visible blocky or pixelated edges instead of smooth gradients.",
+            "Identify shadow maps where individual texels (texture pixels) are clearly visible.",
+            "Check for shadows that appear as jagged stair-stepped patterns rather than soft edges.",
+        ],
+        visual_indicators=[
+            "blocky shadow edges",
+            "visible texel grid in shadows",
+            "stair-stepped shadow boundary",
+            "pixelated shadow gradient",
+        ],
+        confidence_threshold=0.55,
+    ),
+    GlitchPattern(
+        category=GlitchCategory.BLOOM_OVERFLOW,
+        name="Post-Processing Bloom Overflow",
+        description="Three.js UnrealBloomPass or similar post-processing bloom effect is too "
+                    "intense, causing bright areas to bleed glow into surrounding geometry.",
+        severity=GlitchSeverity.LOW,
+        detection_prompts=[
+            "Look for bright areas that have an unusually large, soft glow bleeding into adjacent surfaces.",
+            "Identify scenes where light sources appear to have a 'halo' that extends beyond physical plausibility.",
+            "Check for bright objects whose glow color bleeds onto nearby unrelated geometry.",
+        ],
+        visual_indicators=[
+            "excessive glow bleeding from bright surfaces",
+            "halo around light sources",
+            "bloom color tinting adjacent geometry",
+            "glow bleeding beyond object boundaries",
+        ],
+        confidence_threshold=0.5,
+    ),
 ]
 
 
@@ -289,6 +414,23 @@ def build_vision_prompt(patterns: list[GlitchPattern] | None = None) -> str:
     )
 
 
+
+# Three.js-specific category set for filtering (ref: timmy-config#543)
+THREEJS_CATEGORIES = {
+    GlitchCategory.SHADER_FAILURE,
+    GlitchCategory.TEXTURE_PLACEHOLDER,
+    GlitchCategory.UV_MAPPING_ERROR,
+    GlitchCategory.FRUSTUM_CULLING,
+    GlitchCategory.SHADOW_MAP_ARTIFACT,
+    GlitchCategory.BLOOM_OVERFLOW,
+}
+
+
+def get_threejs_patterns() -> list[GlitchPattern]:
+    """Return only Three.js-specific glitch patterns."""
+    return [p for p in MATRIX_GLITCH_PATTERNS if p.category in THREEJS_CATEGORIES]
+
+
 if __name__ == "__main__":
     import json
     print(f"Loaded {len(MATRIX_GLITCH_PATTERNS)} glitch patterns:\n")

bin/matrix_glitch_detector.py

@@ -9,7 +9,7 @@ Usage:
     python matrix_glitch_detector.py <url> [--angles 4] [--output report.json]
     python matrix_glitch_detector.py --demo  # Run with synthetic test data
 
-Ref: timmy-config#491
+Ref: timmy-config#491, timmy-config#543
 """
 
 import argparse
@@ -33,6 +33,7 @@ from glitch_patterns import (
     MATRIX_GLITCH_PATTERNS,
     build_vision_prompt,
     get_patterns_by_severity,
+    get_threejs_patterns,
 )
 
 
@@ -345,14 +346,17 @@ def _parse_vision_response(
 
 def _infer_severity(category: str, confidence: float) -> str:
     """Infer severity from category and confidence when not provided."""
-    critical_cats = {"missing_textures", "clipping"}
-    high_cats = {"floating_assets", "broken_normals"}
+    critical_cats = {"missing_textures", "clipping", "shader_failure", "texture_placeholder"}
+    high_cats = {"floating_assets", "broken_normals", "uv_mapping_error"}
+    medium_cats = {"frustum_culling", "shadow_map_artifact"}
 
     cat_lower = category.lower()
     if any(c in cat_lower for c in critical_cats):
         return "critical" if confidence > 0.7 else "high"
     if any(c in cat_lower for c in high_cats):
         return "high" if confidence > 0.7 else "medium"
+    if any(c in cat_lower for c in medium_cats):
+        return "medium" if confidence > 0.6 else "low"
     return "medium" if confidence > 0.6 else "low"
 
 
@@ -389,9 +393,9 @@ def build_report(
             ),
         },
         metadata={
-            "detector_version": "0.1.0",
+            "detector_version": "0.2.0",
             "pattern_count": len(MATRIX_GLITCH_PATTERNS),
-            "reference": "timmy-config#491",
+            "reference": "timmy-config#491, timmy-config#543",
         },
     )
 
@@ -460,6 +464,30 @@ def run_demo(output_path: Optional[Path] = None) -> ScanResult:
             screenshot_index=3,
             screenshot_angle="left",
         ),
+        DetectedGlitch(
+            id=str(uuid.uuid4())[:8],
+            category="shader_failure",
+            name="Black Material on Portal Frame",
+            description="Portal frame rendered as solid black — shader compilation failed (missing uniform u_time)",
+            severity="critical",
+            confidence=0.91,
+            location_x=45.0,
+            location_y=30.0,
+            screenshot_index=0,
+            screenshot_angle="front",
+        ),
+        DetectedGlitch(
+            id=str(uuid.uuid4())[:8],
+            category="shadow_map_artifact",
+            name="Pixelated Character Shadow",
+            description="Character shadow shows visible texel grid — shadow map resolution too low (512x512)",
+            severity="medium",
+            confidence=0.78,
+            location_x=52.0,
+            location_y=75.0,
+            screenshot_index=1,
+            screenshot_angle="right",
+        ),
     ]
 
     print(f"[*] Detected {len(demo_glitches)} glitches")
@@ -496,6 +524,11 @@ Examples:
         help="Minimum severity to include in report",
     )
     parser.add_argument("--verbose", "-v", action="store_true", help="Verbose output")
+    parser.add_argument(
+        "--threejs",
+        action="store_true",
+        help="Focus on Three.js-specific glitch patterns only (shader, texture, UV, culling, shadow, bloom)",
+    )
 
     args = parser.parse_args()
 
@@ -525,9 +558,13 @@ Examples:
     screenshots = capture_screenshots(args.url, angles, screenshots_dir)
     print(f"[*] Captured {len(screenshots)} screenshots")
 
-    # Filter patterns by severity
+    # Filter patterns by severity and type
     min_sev = GlitchSeverity(args.min_severity)
     patterns = get_patterns_by_severity(min_sev)
+    if args.threejs:
+        threejs_patterns = get_threejs_patterns()
+        patterns = [p for p in patterns if p in threejs_patterns]
+        print(f"[*] Three.js-focused mode: {len(patterns)} patterns")
 
     # Analyze with vision AI
     print(f"[*] Analyzing with vision AI ({len(patterns)} patterns)...")

bin/preflight-provider-check.py

@@ -0,0 +1,271 @@
+#!/usr/bin/env python3
+"""
+Pre-Flight Provider Check Script
+Issue #508: [Robustness] Credential drain detection — provider health checks
+
+Pre-flight check before session launch: verifies provider credentials and balance.
+
+Usage:
+  python3 preflight-provider-check.py              # Check all providers
+  python3 preflight-provider-check.py --launch      # Check and return exit code
+  python3 preflight-provider-check.py --balance     # Check OpenRouter balance
+"""
+
+import os, sys, json, yaml, urllib.request
+from datetime import datetime, timezone
+from pathlib import Path
+
+# Configuration
+HERMES_HOME = Path(os.environ.get("HERMES_HOME", Path.home() / ".hermes"))
+LOG_DIR = Path.home() / ".local" / "timmy" / "fleet-health"
+LOG_FILE = LOG_DIR / "preflight-check.log"
+
+def log(msg):
+    """Log message to file and optionally console."""
+    timestamp = datetime.now(timezone.utc).strftime("%Y-%m-%d %H:%M:%S")
+    log_entry = "[" + timestamp + "] " + msg
+    
+    LOG_DIR.mkdir(parents=True, exist_ok=True)
+    with open(LOG_FILE, "a") as f:
+        f.write(log_entry + "\n")
+    
+    if "--quiet" not in sys.argv:
+        print(log_entry)
+
+def get_provider_api_key(provider):
+    """Get API key for a provider from .env or environment."""
+    env_file = HERMES_HOME / ".env"
+    if env_file.exists():
+        with open(env_file) as f:
+            for line in f:
+                line = line.strip()
+                if line.startswith(provider.upper() + "_API_KEY="):
+                    return line.split("=", 1)[1].strip().strip("'\"")
+    
+    return os.environ.get(provider.upper() + "_API_KEY")
+
+def check_openrouter_balance(api_key):
+    """Check OpenRouter balance via /api/v1/auth/key."""
+    if not api_key:
+        return False, "No API key", 0
+    
+    try:
+        req = urllib.request.Request(
+            "https://openrouter.ai/api/v1/auth/key",
+            headers={"Authorization": "Bearer " + api_key}
+        )
+        resp = urllib.request.urlopen(req, timeout=10)
+        data = json.loads(resp.read())
+        
+        # Check for credits
+        credits = data.get("data", {}).get("limit", 0)
+        usage = data.get("data", {}).get("usage", 0)
+        remaining = credits - usage if credits else None
+        
+        if remaining is not None and remaining <= 0:
+            return False, "No credits remaining", 0
+        elif remaining is not None:
+            return True, "Credits available", remaining
+        else:
+            return True, "Unlimited or unknown balance", None
+    
+    except urllib.error.HTTPError as e:
+        if e.code == 401:
+            return False, "Invalid API key", 0
+        else:
+            return False, "HTTP " + str(e.code), 0
+    except Exception as e:
+        return False, str(e)[:100], 0
+
+def check_nous_key(api_key):
+    """Check Nous API key with minimal test call."""
+    if not api_key:
+        return False, "No API key"
+    
+    try:
+        req = urllib.request.Request(
+            "https://inference.nousresearch.com/v1/models",
+            headers={"Authorization": "Bearer " + api_key}
+        )
+        resp = urllib.request.urlopen(req, timeout=10)
+        
+        if resp.status == 200:
+            return True, "Valid key"
+        else:
+            return False, "HTTP " + str(resp.status)
+    
+    except urllib.error.HTTPError as e:
+        if e.code == 401:
+            return False, "Invalid API key"
+        elif e.code == 403:
+            return False, "Forbidden"
+        else:
+            return False, "HTTP " + str(e.code)
+    except Exception as e:
+        return False, str(e)[:100]
+
+def check_anthropic_key(api_key):
+    """Check Anthropic API key with minimal test call."""
+    if not api_key:
+        return False, "No API key"
+    
+    try:
+        req = urllib.request.Request(
+            "https://api.anthropic.com/v1/models",
+            headers={
+                "x-api-key": api_key,
+                "anthropic-version": "2023-06-01"
+            }
+        )
+        resp = urllib.request.urlopen(req, timeout=10)
+        
+        if resp.status == 200:
+            return True, "Valid key"
+        else:
+            return False, "HTTP " + str(resp.status)
+    
+    except urllib.error.HTTPError as e:
+        if e.code == 401:
+            return False, "Invalid API key"
+        elif e.code == 403:
+            return False, "Forbidden"
+        else:
+            return False, "HTTP " + str(e.code)
+    except Exception as e:
+        return False, str(e)[:100]
+
+def check_ollama():
+    """Check if Ollama is running."""
+    try:
+        req = urllib.request.Request("http://localhost:11434/api/tags")
+        resp = urllib.request.urlopen(req, timeout=5)
+        
+        if resp.status == 200:
+            data = json.loads(resp.read())
+            models = data.get("models", [])
+            return True, str(len(models)) + " models loaded"
+        else:
+            return False, "HTTP " + str(resp.status)
+    
+    except Exception as e:
+        return False, str(e)[:100]
+
+def get_configured_provider():
+    """Get the configured provider from global config."""
+    config_file = HERMES_HOME / "config.yaml"
+    if not config_file.exists():
+        return None
+    
+    try:
+        with open(config_file) as f:
+            config = yaml.safe_load(f)
+        
+        model_config = config.get("model", {})
+        if isinstance(model_config, dict):
+            return model_config.get("provider")
+    except:
+        pass
+    
+    return None
+
+def run_preflight_check():
+    """Run pre-flight check on all providers."""
+    log("=== Pre-Flight Provider Check ===")
+    
+    results = {}
+    
+    # Check OpenRouter
+    or_key = get_provider_api_key("openrouter")
+    or_ok, or_msg, or_balance = check_openrouter_balance(or_key)
+    results["openrouter"] = {"healthy": or_ok, "message": or_msg, "balance": or_balance}
+    
+    # Check Nous
+    nous_key = get_provider_api_key("nous")
+    nous_ok, nous_msg = check_nous_key(nous_key)
+    results["nous"] = {"healthy": nous_ok, "message": nous_msg}
+    
+    # Check Anthropic
+    anthropic_key = get_provider_api_key("anthropic")
+    anthropic_ok, anthropic_msg = check_anthropic_key(anthropic_key)
+    results["anthropic"] = {"healthy": anthropic_ok, "message": anthropic_msg}
+    
+    # Check Ollama
+    ollama_ok, ollama_msg = check_ollama()
+    results["ollama"] = {"healthy": ollama_ok, "message": ollama_msg}
+    
+    # Get configured provider
+    configured = get_configured_provider()
+    
+    # Summary
+    healthy_count = sum(1 for r in results.values() if r["healthy"])
+    total_count = len(results)
+    
+    log("Results: " + str(healthy_count) + "/" + str(total_count) + " providers healthy")
+    
+    for provider, result in results.items():
+        status = "HEALTHY" if result["healthy"] else "UNHEALTHY"
+        extra = ""
+        if provider == "openrouter" and result.get("balance") is not None:
+            extra = " (balance: " + str(result["balance"]) + ")"
+        
+        log("  " + provider + ": " + status + " - " + result["message"] + extra)
+    
+    if configured:
+        log("Configured provider: " + configured)
+        if configured in results and not results[configured]["healthy"]:
+            log("WARNING: Configured provider " + configured + " is UNHEALTHY!")
+    
+    return results, configured
+
+def check_launch_readiness():
+    """Check if we're ready to launch sessions."""
+    results, configured = run_preflight_check()
+    
+    # Check if configured provider is healthy
+    if configured and configured in results:
+        if not results[configured]["healthy"]:
+            log("LAUNCH BLOCKED: Configured provider " + configured + " is unhealthy")
+            return False, configured + " is unhealthy"
+    
+    # Check if at least one provider is healthy
+    healthy_providers = [p for p, r in results.items() if r["healthy"]]
+    if not healthy_providers:
+        log("LAUNCH BLOCKED: No healthy providers available")
+        return False, "No healthy providers"
+    
+    log("LAUNCH READY: " + str(len(healthy_providers)) + " healthy providers available")
+    return True, "Ready"
+
+def show_balance():
+    """Show OpenRouter balance."""
+    api_key = get_provider_api_key("openrouter")
+    if not api_key:
+        print("No OpenRouter API key found")
+        return
+    
+    ok, msg, balance = check_openrouter_balance(api_key)
+    
+    if ok:
+        if balance is not None:
+            print("OpenRouter balance: " + str(balance) + " credits")
+        else:
+            print("OpenRouter: " + msg)
+    else:
+        print("OpenRouter: " + msg)
+
+def main():
+    if "--balance" in sys.argv:
+        show_balance()
+    elif "--launch" in sys.argv:
+        ready, message = check_launch_readiness()
+        if ready:
+            print("READY")
+            sys.exit(0)
+        else:
+            print("BLOCKED: " + message)
+            sys.exit(1)
+    else:
+        run_preflight_check()
+
+if __name__ == "__main__":
+    main()

bin/provider-health-monitor.py

@@ -0,0 +1,411 @@
+#!/usr/bin/env python3
+"""
+Provider Health Monitor Script
+Issue #509: [Robustness] Provider-aware profile config — auto-switch on failure
+
+Monitors provider health and automatically switches profiles to working providers.
+
+Usage:
+  python3 provider-health-monitor.py              # Run once
+  python3 provider-health-monitor.py --daemon     # Run continuously
+  python3 provider-health-monitor.py --status     # Show provider health
+"""
+
+import os, sys, json, yaml, urllib.request, time
+from datetime import datetime, timezone
+from pathlib import Path
+
+# Configuration
+HERMES_HOME = Path(os.environ.get("HERMES_HOME", Path.home() / ".hermes"))
+PROFILES_DIR = HERMES_HOME / "profiles"
+LOG_DIR = Path.home() / ".local" / "timmy" / "fleet-health"
+STATE_FILE = LOG_DIR / "tmux-state.json"
+LOG_FILE = LOG_DIR / "provider-health.log"
+
+# Provider test endpoints
+PROVIDER_TESTS = {
+    "openrouter": {
+        "url": "https://openrouter.ai/api/v1/models",
+        "method": "GET",
+        "headers": lambda api_key: {"Authorization": "Bearer " + api_key},
+        "timeout": 10
+    },
+    "anthropic": {
+        "url": "https://api.anthropic.com/v1/models",
+        "method": "GET",
+        "headers": lambda api_key: {"x-api-key": api_key, "anthropic-version": "2023-06-01"},
+        "timeout": 10
+    },
+    "nous": {
+        "url": "https://inference.nousresearch.com/v1/models",
+        "method": "GET",
+        "headers": lambda api_key: {"Authorization": "Bearer " + api_key},
+        "timeout": 10
+    },
+    "kimi-coding": {
+        "url": "https://api.kimi.com/coding/v1/models",
+        "method": "GET",
+        "headers": lambda api_key: {"x-api-key": api_key, "x-api-provider": "kimi-coding"},
+        "timeout": 10
+    },
+    "ollama": {
+        "url": "http://localhost:11434/api/tags",
+        "method": "GET",
+        "headers": lambda api_key: {},
+        "timeout": 5
+    }
+}
+
+def log(msg):
+    """Log message to file and optionally console."""
+    timestamp = datetime.now(timezone.utc).strftime("%Y-%m-%d %H:%M:%S")
+    log_entry = "[" + timestamp + "] " + msg
+    
+    LOG_DIR.mkdir(parents=True, exist_ok=True)
+    with open(LOG_FILE, "a") as f:
+        f.write(log_entry + "\n")
+    
+    if "--quiet" not in sys.argv:
+        print(log_entry)
+
+def get_provider_api_key(provider):
+    """Get API key for a provider from .env or environment."""
+    env_file = HERMES_HOME / ".env"
+    if env_file.exists():
+        with open(env_file) as f:
+            for line in f:
+                line = line.strip()
+                if line.startswith(provider.upper() + "_API_KEY="):
+                    return line.split("=", 1)[1].strip().strip("'\"")
+    
+    return os.environ.get(provider.upper() + "_API_KEY")
+
+def test_provider(provider, api_key=None):
+    """Test if a provider is healthy."""
+    config = PROVIDER_TESTS.get(provider)
+    if not config:
+        return False, "Unknown provider: " + provider
+    
+    headers = config["headers"](api_key or "")
+    
+    try:
+        req = urllib.request.Request(
+            config["url"],
+            headers=headers,
+            method=config["method"]
+        )
+        resp = urllib.request.urlopen(req, timeout=config["timeout"])
+        
+        if resp.status == 200:
+            return True, "Healthy"
+        else:
+            return False, "HTTP " + str(resp.status)
+    
+    except urllib.error.HTTPError as e:
+        if e.code == 401:
+            return False, "Unauthorized (401)"
+        elif e.code == 403:
+            return False, "Forbidden (403)"
+        elif e.code == 429:
+            return True, "Rate limited but accessible"
+        else:
+            return False, "HTTP " + str(e.code)
+    except Exception as e:
+        return False, str(e)[:100]
+
+def get_all_providers():
+    """Get all providers from profiles and global config."""
+    providers = set()
+    
+    # Global config
+    global_config = HERMES_HOME / "config.yaml"
+    if global_config.exists():
+        try:
+            with open(global_config) as f:
+                config = yaml.safe_load(f)
+            
+            # Primary model provider
+            model_config = config.get("model", {})
+            if isinstance(model_config, dict):
+                provider = model_config.get("provider", "")
+                if provider:
+                    providers.add(provider)
+            
+            # Auxiliary providers
+            auxiliary = config.get("auxiliary", {})
+            for aux_config in auxiliary.values():
+                if isinstance(aux_config, dict):
+                    provider = aux_config.get("provider", "")
+                    if provider and provider != "auto":
+                        providers.add(provider)
+        except:
+            pass
+    
+    # Profile configs
+    if PROFILES_DIR.exists():
+        for profile_dir in PROFILES_DIR.iterdir():
+            if profile_dir.is_dir():
+                config_file = profile_dir / "config.yaml"
+                if config_file.exists():
+                    try:
+                        with open(config_file) as f:
+                            config = yaml.safe_load(f)
+                        
+                        model_config = config.get("model", {})
+                        if isinstance(model_config, dict):
+                            provider = model_config.get("provider", "")
+                            if provider:
+                                providers.add(provider)
+                        
+                        auxiliary = config.get("auxiliary", {})
+                        for aux_config in auxiliary.values():
+                            if isinstance(aux_config, dict):
+                                provider = aux_config.get("provider", "")
+                                if provider and provider != "auto":
+                                    providers.add(provider)
+                    except:
+                        pass
+    
+    # Add common providers even if not configured
+    providers.update(["openrouter", "nous", "ollama"])
+    
+    return list(providers)
+
+def build_health_map():
+    """Build a health map of all providers."""
+    providers = get_all_providers()
+    health_map = {}
+    
+    log("Testing " + str(len(providers)) + " providers...")
+    
+    for provider in providers:
+        api_key = get_provider_api_key(provider)
+        healthy, message = test_provider(provider, api_key)
+        
+        health_map[provider] = {
+            "healthy": healthy,
+            "message": message,
+            "last_test": datetime.now(timezone.utc).isoformat(),
+            "api_key_present": bool(api_key)
+        }
+        
+        status = "HEALTHY" if healthy else "UNHEALTHY"
+        log("  " + provider + ": " + status + " - " + message)
+    
+    return health_map
+
+def get_fallback_providers(health_map):
+    """Get list of healthy providers in priority order."""
+    # Priority order: nous, openrouter, ollama, others
+    priority_order = ["nous", "openrouter", "ollama", "anthropic", "kimi-coding"]
+    
+    healthy = []
+    for provider in priority_order:
+        if provider in health_map and health_map[provider]["healthy"]:
+            healthy.append(provider)
+    
+    # Add any other healthy providers not in priority list
+    for provider, info in health_map.items():
+        if info["healthy"] and provider not in healthy:
+            healthy.append(provider)
+    
+    return healthy
+
+def update_profile_config(profile_name, new_provider):
+    """Update a profile's config to use a new provider."""
+    config_file = PROFILES_DIR / profile_name / "config.yaml"
+    
+    if not config_file.exists():
+        return False, "Config file not found"
+    
+    try:
+        with open(config_file) as f:
+            config = yaml.safe_load(f)
+        
+        # Update model provider
+        if "model" not in config:
+            config["model"] = {}
+        
+        old_provider = config["model"].get("provider", "unknown")
+        config["model"]["provider"] = new_provider
+        
+        # Update auxiliary providers if they were using the old provider
+        auxiliary = config.get("auxiliary", {})
+        for aux_name, aux_config in auxiliary.items():
+            if isinstance(aux_config, dict) and aux_config.get("provider") == old_provider:
+                aux_config["provider"] = new_provider
+        
+        # Write back
+        with open(config_file, "w") as f:
+            yaml.dump(config, f, default_flow_style=False)
+        
+        log("Updated " + profile_name + ": " + old_provider + " -> " + new_provider)
+        return True, "Updated"
+    
+    except Exception as e:
+        return False, str(e)
+
+def check_profiles(health_map):
+    """Check all profiles and update unhealthy providers."""
+    if not PROFILES_DIR.exists():
+        return
+    
+    fallback_providers = get_fallback_providers(health_map)
+    if not fallback_providers:
+        log("CRITICAL: No healthy providers available!")
+        return
+    
+    updated_profiles = []
+    
+    for profile_dir in PROFILES_DIR.iterdir():
+        if not profile_dir.is_dir():
+            continue
+        
+        profile_name = profile_dir.name
+        config_file = profile_dir / "config.yaml"
+        
+        if not config_file.exists():
+            continue
+        
+        try:
+            with open(config_file) as f:
+                config = yaml.safe_load(f)
+            
+            model_config = config.get("model", {})
+            if not isinstance(model_config, dict):
+                continue
+            
+            current_provider = model_config.get("provider", "")
+            if not current_provider:
+                continue
+            
+            # Check if current provider is healthy
+            if current_provider in health_map and health_map[current_provider]["healthy"]:
+                continue  # Provider is healthy, no action needed
+            
+            # Find best fallback
+            best_fallback = None
+            for provider in fallback_providers:
+                if provider != current_provider:
+                    best_fallback = provider
+                    break
+            
+            if not best_fallback:
+                log("No fallback for " + profile_name + " (current: " + current_provider + ")")
+                continue
+            
+            # Update profile
+            success, message = update_profile_config(profile_name, best_fallback)
+            if success:
+                updated_profiles.append({
+                    "profile": profile_name,
+                    "old_provider": current_provider,
+                    "new_provider": best_fallback
+                })
+        
+        except Exception as e:
+            log("Error processing " + profile_name + ": " + str(e))
+    
+    return updated_profiles
+
+def load_state():
+    """Load state from tmux-state.json."""
+    if STATE_FILE.exists():
+        try:
+            with open(STATE_FILE) as f:
+                return json.load(f)
+        except:
+            pass
+    return {}
+
+def save_state(state):
+    """Save state to tmux-state.json."""
+    LOG_DIR.mkdir(parents=True, exist_ok=True)
+    
+    with open(STATE_FILE, "w") as f:
+        json.dump(state, f, indent=2)
+
+def run_once():
+    """Run provider health check once."""
+    log("=== Provider Health Check ===")
+    
+    state = load_state()
+    
+    # Build health map
+    health_map = build_health_map()
+    
+    # Check profiles and update if needed
+    updated_profiles = check_profiles(health_map)
+    
+    # Update state
+    state["provider_health"] = health_map
+    state["last_provider_check"] = datetime.now(timezone.utc).isoformat()
+    
+    if updated_profiles:
+        state["last_profile_updates"] = updated_profiles
+    
+    save_state(state)
+    
+    # Summary
+    healthy_count = sum(1 for p in health_map.values() if p["healthy"])
+    total_count = len(health_map)
+    
+    log("Health: " + str(healthy_count) + "/" + str(total_count) + " providers healthy")
+    
+    if updated_profiles:
+        log("Updated " + str(len(updated_profiles)) + " profiles:")
+        for update in updated_profiles:
+            log("  " + update["profile"] + ": " + update["old_provider"] + " -> " + update["new_provider"])
+
+def show_status():
+    """Show provider health status."""
+    state = load_state()
+    health_map = state.get("provider_health", {})
+    
+    if not health_map:
+        print("No provider health data available. Run without --status first.")
+        return
+    
+    print("Provider Health (last updated: " + str(state.get("last_provider_check", "unknown")) + ")")
+    print("=" * 80)
+    
+    for provider, info in sorted(health_map.items()):
+        status = "HEALTHY" if info["healthy"] else "UNHEALTHY"
+        message = info.get("message", "")
+        api_key = "yes" if info.get("api_key_present") else "no"
+        
+        print(provider.ljust(20) + " " + status.ljust(10) + " API key: " + api_key + " - " + message)
+    
+    # Show recent updates
+    updates = state.get("last_profile_updates", [])
+    if updates:
+        print()
+        print("Recent Profile Updates:")
+        for update in updates:
+            print("  " + update["profile"] + ": " + update["old_provider"] + " -> " + update["new_provider"])
+
+def daemon_mode():
+    """Run continuously."""
+    log("Starting provider health daemon (check every 300s)")
+    
+    while True:
+        try:
+            run_once()
+            time.sleep(300)  # Check every 5 minutes
+        except KeyboardInterrupt:
+            log("Daemon stopped by user")
+            break
+        except Exception as e:
+            log("Error: " + str(e))
+            time.sleep(60)
+
+def main():
+    if "--status" in sys.argv:
+        show_status()
+    elif "--daemon" in sys.argv:
+        daemon_mode()
+    else:
+        run_once()
+
+if __name__ == "__main__":
+    main()

cron/jobs.json

@@ -196,7 +196,37 @@
       "paused_reason": null,
       "skills": [],
       "skill": null
+    },
+    {
+      "id": "tmux-supervisor-513",
+      "name": "Autonomous Cron Supervisor",
+      "prompt": "Load the tmux-supervisor skill and execute the monitoring protocol.\n\nCheck both `dev` and `timmy` tmux sessions for idle panes. Only send Telegram notifications on actionable events (idle, overflow, failure). Be silent when all agents are working.\n\nSteps:\n1. List all tmux sessions (skip 'Alexander')\n2. For each session, list windows and panes\n3. Capture each pane and classify state (idle vs active)\n4. For idle panes: read context, craft context-aware prompt\n5. Send /queue prompts to idle panes\n6. Verify prompts landed\n7. Only notify via Telegram if:\n   - A pane was prompted (idle detected)\n   - A pane shows context overflow (>80%)\n   - A pane is stuck or crashed\n8. If all panes are active: respond with [SILENT]",
+      "schedule": {
+        "kind": "interval",
+        "minutes": 7,
+        "display": "every 7m"
+      },
+      "schedule_display": "every 7m",
+      "repeat": {
+        "times": null,
+        "completed": 0
+      },
+      "enabled": true,
+      "created_at": "2026-04-15T03:00:00.000000+00:00",
+      "next_run_at": null,
+      "last_run_at": null,
+      "last_status": null,
+      "last_error": null,
+      "deliver": "telegram",
+      "origin": null,
+      "state": "scheduled",
+      "paused_at": null,
+      "paused_reason": null,
+      "skills": [
+        "tmux-supervisor"
+      ],
+      "skill": "tmux-supervisor"
     }
   ],
   "updated_at": "2026-04-13T02:00:00+00:00"
-}
+}

docs/visual-evidence-689.md

@@ -0,0 +1,74 @@
+# Visual Evidence — Gemma 4 Multimodal Scene Description Generator
+
+## Test Image: Coffee Beans (Macro Photo)
+
+### Gemma 4 Vision Analysis (via Ollama)
+
+**Model:** gemma4:latest (8B, Q4_K_M)
+**Input:** sample_photo.jpg (46KB JPEG)
+
+**Structured Output (JSONL):**
+```json
+{
+  "mood": "dark",
+  "colors": ["dark brown", "espresso", "black"],
+  "composition": "close-up",
+  "camera": "static",
+  "lighting": "soft",
+  "description": "An extreme close-up shot captures a dense pile of roasted coffee beans. The beans are a uniform, deep dark brown and appear slightly oily, filling the entire frame. The focus emphasizes the rich texture and individual shapes of the beans."
+}
+```
+
+### Hermes Vision Analysis (Cross-Validation)
+
+**Scene ID:** COFFEE_MACRO_001
+**Mood:** Warm, aromatic, and comforting
+**Dominant Colors:** Deep umber, burnt sienna, espresso black, mahogany
+**Composition:** Full-frame fill, centrally weighted
+**Camera:** High-angle, close-up (Macro)
+**Lighting:** Soft, diffused top-lighting
+
+## Test Image: Abstract Geometric Composition
+
+### Gemma 4 Vision Analysis
+
+**Input:** scene1.jpg (10KB, PIL-generated)
+
+**Structured Output (JSONL):**
+```json
+{
+  "mood": "energetic",
+  "colors": ["deep blue", "yellow", "coral"],
+  "composition": "wide-shot",
+  "camera": "static",
+  "lighting": "artificial",
+  "description": "This is an abstract graphic composition set against a solid, deep blue background. A bright yellow square is placed in the upper left quadrant, while a large, solid coral-colored circle occupies the lower right quadrant. The geometric shapes create a high-contrast, minimalist visual balance."
+}
+```
+
+## Verification Summary
+
+| Test | Status | Details |
+|------|--------|---------|
+| Model detection | ✅ PASS | `gemma4:latest` auto-detected |
+| Image scanning | ✅ PASS | 2 images found recursively |
+| Vision analysis | ✅ PASS | Both images described accurately |
+| JSON parsing | ✅ PASS | Structured output with all fields |
+| Training format | ✅ PASS | JSONL with source, model, timestamp |
+| ShareGPT format | ⚠️ PARTIAL | Works but needs retry on rate limit |
+
+## Running the Generator
+
+```bash
+# Check model availability
+python scripts/generate_scene_descriptions.py --check-model
+
+# Generate scene descriptions from assets
+python scripts/generate_scene_descriptions.py --input ./assets --output training-data/scene-descriptions-auto.jsonl
+
+# Limit to 10 files with specific model
+python scripts/generate_scene_descriptions.py --input ./assets --model gemma4:latest --limit 10
+
+# ShareGPT format for training pipeline
+python scripts/generate_scene_descriptions.py --input ./assets --format sharegpt
+```

scripts/generate_scene_descriptions.py

@@ -0,0 +1,409 @@
+#!/usr/bin/env python3
+"""
+Auto-generate scene descriptions from image/video assets.
+
+Scans a directory for media files, generates scene descriptions using
+a local vision model (Ollama), and outputs training pairs in JSONL format.
+
+Supports Gemma 4 multimodal vision via Ollama. Falls back gracefully when
+models are unavailable.
+
+Usage:
+    python scripts/generate_scene_descriptions.py --input ./assets --output training-data/scene-descriptions-auto.jsonl
+    python scripts/generate_scene_descriptions.py --input ./assets --model gemma4:latest --limit 50
+    python scripts/generate_scene_descriptions.py --input ./assets --format sharegpt
+    python scripts/generate_scene_descriptions.py --dry-run  # List files without generating
+    python scripts/generate_scene_descriptions.py --input ./assets --check-model  # Verify model availability
+
+Ref: timmy-config#689
+"""
+
+import argparse
+import base64
+import json
+import os
+import re
+import subprocess
+import sys
+import time
+import urllib.request
+from datetime import datetime, timezone
+from pathlib import Path
+from typing import Optional
+
+# Supported media extensions
+IMAGE_EXTS = {".jpg", ".jpeg", ".png", ".gif", ".webp", ".bmp"}
+VIDEO_EXTS = {".mp4", ".webm", ".mov", ".avi", ".mkv"}
+ALL_EXTS = IMAGE_EXTS | VIDEO_EXTS
+
+# File size limit (50MB) — prevents unbounded memory usage on large images
+MAX_FILE_SIZE = 50 * 1024 * 1024
+
+# Vision models in preference order (best first)
+VISION_MODELS = [
+    "gemma4:latest",       # Gemma 4 — multimodal vision (8B, Q4_K_M)
+    "gemma3:12b",          # Gemma 3 — fallback vision
+    "llava:latest",        # LLaVA — generic vision
+    "llava-phi3:latest",   # LLaVA-Phi3 — lightweight vision
+]
+
+# Vision model prompt template (structured JSON output)
+SCENE_PROMPT = """Describe this image for a visual scene database. Output ONLY valid JSON (no markdown, no explanation):
+{
+  "mood": "one of: calm, energetic, dark, warm, cool, chaotic, serene, tense, joyful, melancholic",
+  "colors": ["dominant color 1", "dominant color 2", "dominant color 3"],
+  "composition": "one of: close-up, wide-shot, medium-shot, low-angle, high-angle, bird-eye, profile, over-shoulder",
+  "camera": "one of: static, slow-pan, tracking, handheld, crane, dolly, steady, locked-off",
+  "lighting": "one of: natural, artificial, mixed, dramatic, soft, harsh, backlit",
+  "description": "2-3 sentence visual description of the scene"
+}
+
+Be specific. Describe what you see, not what you imagine."""
+
+# ShareGPT format prompt (for training pipeline integration)
+SHAREGPT_SCENE_PROMPT = """Analyze this image and describe the visual scene. Include mood, dominant colors, composition, camera angle, lighting, and a vivid 2-3 sentence description."""
+
+
+def check_model_available(model: str, ollama_url: str = "http://localhost:11434") -> bool:
+    """Check if a model is available in Ollama."""
+    try:
+        req = urllib.request.Request(f"{ollama_url}/api/tags")
+        resp = urllib.request.urlopen(req, timeout=10)
+        data = json.loads(resp.read())
+        available = [m["name"] for m in data.get("models", [])]
+        return model in available
+    except Exception:
+        return False
+
+
+def auto_detect_model(ollama_url: str = "http://localhost:11434") -> Optional[str]:
+    """Auto-detect the best available vision model."""
+    for model in VISION_MODELS:
+        if check_model_available(model, ollama_url):
+            print(f"Auto-detected vision model: {model}", file=sys.stderr)
+            return model
+    return None
+
+
+def scan_media(input_dir: str) -> list[Path]:
+    """Scan directory for media files recursively."""
+    media_files = []
+    input_path = Path(input_dir)
+    if not input_path.exists():
+        print(f"Error: {input_dir} does not exist", file=sys.stderr)
+        return media_files
+
+    for ext in sorted(ALL_EXTS):
+        media_files.extend(input_path.rglob(f"*{ext}"))
+        media_files.extend(input_path.rglob(f"*{ext.upper()}"))
+
+    return sorted(set(media_files))
+
+
+def extract_video_frame(video_path: Path, output_path: Path) -> bool:
+    """Extract a representative frame from a video using ffmpeg."""
+    try:
+        result = subprocess.run(
+            # FIX #3: Seek 2s in before grabbing frame — avoids black/title frames
+            ["ffmpeg", "-ss", "2", "-i", str(video_path), "-vframes", "1",
+             "-q:v", "2", str(output_path), "-y"],
+            capture_output=True, timeout=30,
+        )
+        if result.returncode != 0 and result.stderr:
+            print(f"  ffmpeg stderr: {result.stderr.decode(errors='replace')[:200]}", file=sys.stderr)
+        return output_path.exists() and output_path.stat().st_size > 0
+    except FileNotFoundError:
+        print("  ffmpeg not found — skipping video frame extraction", file=sys.stderr)
+        return False
+    except Exception as e:
+        print(f"  ffmpeg error: {e}", file=sys.stderr)
+        return False
+
+
+def describe_image(
+    image_path: Path,
+    model: str = "gemma4:latest",
+    ollama_url: str = "http://localhost:11434",
+    max_retries: int = 2,
+) -> Optional[dict]:
+    """Generate scene description using Ollama vision model with retry."""
+    # FIX #1: Check file size before reading into memory
+    if image_path.stat().st_size > MAX_FILE_SIZE:
+        print(f"  Skipping {image_path.name}: exceeds {MAX_FILE_SIZE // (1024*1024)}MB limit", file=sys.stderr)
+        return None
+
+    for attempt in range(max_retries + 1):
+        try:
+            with open(image_path, "rb") as f:
+                image_b64 = base64.b64encode(f.read()).decode()
+
+            req = urllib.request.Request(
+                f"{ollama_url}/api/generate",
+                data=json.dumps({
+                    "model": model,
+                    "prompt": SCENE_PROMPT,
+                    "images": [image_b64],
+                    "stream": False,
+                    "options": {"temperature": 0.3, "num_predict": 1024}
+                }).encode(),
+                headers={"Content-Type": "application/json"},
+            )
+            resp = urllib.request.urlopen(req, timeout=120)
+            data = json.loads(resp.read())
+            response_text = data.get("response", "")
+
+            # Parse JSON from response (handle both complete and truncated JSON)
+            json_match = re.search(r"\{[\s\S]*\}", response_text)
+            if not json_match:
+                # Try to find opening brace for truncated JSON
+                brace_match = re.search(r"\{", response_text)
+                if brace_match:
+                    json_match = brace_match
+
+            if json_match:
+                raw_json = json_match.group() if hasattr(json_match, 'group') else response_text[json_match.start():]
+                # Try strict parse first
+                try:
+                    parsed = json.loads(raw_json)
+                    required = ["mood", "colors", "composition", "camera", "description"]
+                    if all(k in parsed for k in required) and parsed.get("description"):
+                        return parsed
+                except json.JSONDecodeError:
+                    # Attempt repair: extract fields from truncated JSON
+                    repaired = {}
+                    for field in ["mood", "colors", "composition", "camera", "lighting", "description"]:
+                        pat = rf'"\s*{field}"\s*:\s*"([^"]*)"'
+                        m = re.search(pat, response_text)
+                        if m:
+                            repaired[field] = m.group(1)
+                        elif field == "colors":
+                            colors_match = re.search(r'"colors"\s*:\s*\[([^\]]*)\]', response_text)
+                            if colors_match:
+                                repaired[field] = [c.strip().strip('"') for c in colors_match.group(1).split(",") if c.strip()]
+                            else:
+                                repaired[field] = []
+                        else:
+                            repaired[field] = "unknown"
+                    if repaired.get("description") or repaired.get("mood") != "unknown":
+                        return repaired
+
+            # Final fallback: natural language response
+            clean = re.sub(r"[*_`#]", "", response_text).strip()
+            clean = re.sub(r"\n{3,}", "\n\n", clean)
+            return {
+                "description": clean[:500] if clean else response_text[:500],
+                "mood": "unknown",
+                "colors": [],
+                "composition": "unknown",
+                "camera": "unknown",
+                "lighting": "unknown"
+            }
+
+        except (urllib.error.URLError, TimeoutError) as e:
+            if attempt < max_retries:
+                wait = 2 ** attempt
+                print(f"  Retry {attempt + 1}/{max_retries} after {wait}s: {e}", file=sys.stderr)
+                time.sleep(wait)
+            else:
+                print(f"  Error describing {image_path.name}: {e}", file=sys.stderr)
+                return None
+        except Exception as e:
+            print(f"  Error describing {image_path.name}: {e}", file=sys.stderr)
+            return None
+
+
+def describe_image_sharegpt(
+    image_path: Path,
+    model: str = "gemma4:latest",
+    ollama_url: str = "http://localhost:11434",
+    max_retries: int = 2,
+) -> Optional[str]:
+    """Generate scene description in natural language for ShareGPT format."""
+    # FIX #1: Check file size before reading into memory
+    if image_path.stat().st_size > MAX_FILE_SIZE:
+        print(f"  Skipping {image_path.name}: exceeds {MAX_FILE_SIZE // (1024*1024)}MB limit", file=sys.stderr)
+        return None
+
+    for attempt in range(max_retries + 1):
+        try:
+            with open(image_path, "rb") as f:
+                image_b64 = base64.b64encode(f.read()).decode()
+
+            req = urllib.request.Request(
+                f"{ollama_url}/api/generate",
+                data=json.dumps({
+                    "model": model,
+                    "prompt": SHAREGPT_SCENE_PROMPT,
+                    "images": [image_b64],
+                    "stream": False,
+                    "options": {"temperature": 0.5, "num_predict": 256}
+                }).encode(),
+                headers={"Content-Type": "application/json"},
+            )
+            resp = urllib.request.urlopen(req, timeout=120)
+            data = json.loads(resp.read())
+            return data.get("response", "").strip()
+
+        except (urllib.error.URLError, TimeoutError) as e:
+            if attempt < max_retries:
+                time.sleep(2 ** attempt)
+            else:
+                return None
+        except Exception:
+            return None
+
+
+def generate_training_pairs(
+    media_files: list[Path],
+    model: str,
+    ollama_url: str,
+    limit: int = 0,
+    dry_run: bool = False,
+    output_format: str = "jsonl",
+) -> list[dict]:
+    """Generate training pairs from media files."""
+    pairs = []
+    files = media_files[:limit] if limit > 0 else media_files
+
+    print(f"Processing {len(files)} files with model {model}...", file=sys.stderr)
+
+    for i, media_path in enumerate(files):
+        print(f"  [{i + 1}/{len(files)}] {media_path.name}...", file=sys.stderr, end=" ", flush=True)
+
+        if dry_run:
+            print("(dry run)", file=sys.stderr)
+            pairs.append({"source": str(media_path), "status": "dry-run"})
+            continue
+
+        is_video = media_path.suffix.lower() in VIDEO_EXTS
+        work_path = media_path
+
+        if is_video:
+            frame_path = media_path.with_suffix(".frame.jpg")
+            if extract_video_frame(media_path, frame_path):
+                work_path = frame_path
+            else:
+                print("SKIP (frame extraction failed)", file=sys.stderr)
+                continue
+
+        try:
+            if output_format == "sharegpt":
+                # ShareGPT format for training pipeline
+                description = describe_image_sharegpt(work_path, model, ollama_url)
+                if description:
+                    pair = {
+                        "conversations": [
+                            {"from": "human", "value": f"<image>\n{SHAREGPT_SCENE_PROMPT}"},
+                            {"from": "gpt", "value": description}
+                        ],
+                        "source": str(media_path),
+                        "media_type": "video" if is_video else "image",
+                        "model": model,
+                        "generated_at": datetime.now(timezone.utc).isoformat(),
+                    }
+                    pairs.append(pair)
+                    print("OK", file=sys.stderr)
+                else:
+                    print("FAIL", file=sys.stderr)
+            else:
+                # Structured JSONL format
+                description = describe_image(work_path, model, ollama_url)
+                if description:
+                    pair = {
+                        "source": str(media_path),
+                        "media_type": "video" if is_video else "image",
+                        "description": description,
+                        "model": model,
+                        "generated_at": datetime.now(timezone.utc).isoformat(),
+                    }
+                    pairs.append(pair)
+                    print("OK", file=sys.stderr)
+                else:
+                    print("FAIL", file=sys.stderr)
+        finally:
+            # FIX #6: Cleanup temp frame in try/finally — survives crashes
+            if is_video and work_path != media_path:
+                try:
+                    work_path.unlink()
+                except Exception:
+                    pass
+
+        # Small delay between files (reduced from 0.5s — Ollama is local)
+        time.sleep(0.1)
+
+    return pairs
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Auto-generate scene descriptions from media assets using vision AI"
+    )
+    parser.add_argument("--input", "-i", default="", help="Input directory with media files")
+    parser.add_argument("--output", "-o", default="training-data/scene-descriptions-auto.jsonl")
+    parser.add_argument("--model", "-m", default=None, help="Ollama model name (auto-detects best available if omitted)")
+    parser.add_argument("--ollama-url", default="http://localhost:11434")
+    parser.add_argument("--limit", "-l", type=int, default=0, help="Max files to process (0=all)")
+    parser.add_argument("--dry-run", action="store_true", help="List files without generating")
+    parser.add_argument("--check-model", action="store_true", help="Check model availability and exit")
+    parser.add_argument("--format", choices=["jsonl", "sharegpt"], default="jsonl",
+                        help="Output format: jsonl (structured) or sharegpt (training pipeline)")
+    args = parser.parse_args()
+
+    # Model detection
+    if args.check_model:
+        if args.model:
+            available = check_model_available(args.model, args.ollama_url)
+            print(f"Model '{args.model}': {'✅ available' if available else '❌ not found'}")
+        else:
+            model = auto_detect_model(args.ollama_url)
+            if model:
+                print(f"✅ Best available: {model}")
+            else:
+                print("❌ No vision models found in Ollama — install one with: ollama pull gemma4:latest")
+        sys.exit(0)
+
+    # Auto-detect model if not specified
+    model = args.model
+    if not model:
+        model = auto_detect_model(args.ollama_url)
+        if not model:
+            # Fall back to best default even if not installed — let Ollama handle the error
+            model = "gemma4:latest"
+            print(f"Warning: No vision models detected. Falling back to {model}", file=sys.stderr)
+
+    # Validate input
+    if not args.input:
+        print("Error: --input is required (unless using --check-model)", file=sys.stderr)
+        sys.exit(1)
+
+    # Scan and process
+    media_files = scan_media(args.input)
+    print(f"Found {len(media_files)} media files", file=sys.stderr)
+
+    if not media_files:
+        print("No media files found.", file=sys.stderr)
+        sys.exit(1)
+
+    pairs = generate_training_pairs(
+        media_files, model, args.ollama_url,
+        args.limit, args.dry_run, args.format
+    )
+
+    # Write output
+    output_path = Path(args.output)
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+    with open(output_path, "w") as f:
+        for pair in pairs:
+            f.write(json.dumps(pair, ensure_ascii=False) + "\n")
+
+    print(f"\nWrote {len(pairs)} pairs to {output_path}", file=sys.stderr)
+
+    # Summary
+    success = len([p for p in pairs if "description" in p or "conversations" in p])
+    failed = len(pairs) - success
+    if failed > 0:
+        print(f"  ⚠️  {failed} files failed", file=sys.stderr)
+
+
+if __name__ == "__main__":
+    main()

tests/test_glitch_detector.py

@@ -19,9 +19,11 @@ from glitch_patterns import (
     GlitchPattern,
     GlitchSeverity,
     MATRIX_GLITCH_PATTERNS,
+    THREEJS_CATEGORIES,
     build_vision_prompt,
     get_pattern_by_category,
     get_patterns_by_severity,
+    get_threejs_patterns,
 )
 
 from matrix_glitch_detector import (
@@ -40,7 +42,7 @@ class TestGlitchPatterns(unittest.TestCase):
 
     def test_pattern_count(self):
         """Verify we have a reasonable number of defined patterns."""
-        self.assertGreaterEqual(len(MATRIX_GLITCH_PATTERNS), 8)
+        self.assertGreaterEqual(len(MATRIX_GLITCH_PATTERNS), 14)  # 10 generic + 6 Three.js
 
     def test_all_patterns_have_required_fields(self):
         """Every pattern must have category, name, description, severity, prompts."""
@@ -88,6 +90,9 @@ class TestGlitchPatterns(unittest.TestCase):
         self.assertIn("Floating Object", prompt)
         self.assertIn("Z-Fighting", prompt)
         self.assertIn("Missing", prompt)
+        # Three.js patterns should be included
+        self.assertIn("Shader Compilation Failure", prompt)
+        self.assertIn("Bloom Overflow", prompt)
 
     def test_build_vision_prompt_subset(self):
         """Vision prompt with subset should only include specified patterns."""
@@ -248,7 +253,7 @@ class TestGlitchDetector(unittest.TestCase):
 
         try:
             report = run_demo(output_path)
-            self.assertEqual(len(report.glitches), 4)
+            self.assertEqual(len(report.glitches), 6)  # 4 original + 2 Three.js
             self.assertGreater(report.summary["total_glitches"], 0)
             self.assertTrue(output_path.exists())
 
@@ -260,6 +265,93 @@ class TestGlitchDetector(unittest.TestCase):
             output_path.unlink(missing_ok=True)
 
 
+class TestThreeJsPatterns(unittest.TestCase):
+    """Tests for Three.js-specific glitch patterns (timmy-config#543)."""
+
+    def test_get_threejs_patterns_returns_only_threejs(self):
+        """get_threejs_patterns() should return only Three.js categories."""
+        patterns = get_threejs_patterns()
+        self.assertEqual(len(patterns), 6)
+        for p in patterns:
+            self.assertIn(p.category, THREEJS_CATEGORIES)
+
+    def test_threejs_patterns_have_required_fields(self):
+        """All Three.js patterns must have valid fields."""
+        for p in get_threejs_patterns():
+            self.assertIsInstance(p.category, GlitchCategory)
+            self.assertTrue(p.name)
+            self.assertTrue(p.description)
+            self.assertIsInstance(p.severity, GlitchSeverity)
+            self.assertGreater(len(p.detection_prompts), 0)
+            self.assertGreater(len(p.visual_indicators), 0)
+
+    def test_shader_failure_is_critical(self):
+        """Shader compilation failure should be CRITICAL severity."""
+        p = get_pattern_by_category(GlitchCategory.SHADER_FAILURE)
+        self.assertIsNotNone(p)
+        self.assertEqual(p.severity, GlitchSeverity.CRITICAL)
+
+    def test_texture_placeholder_is_critical(self):
+        """Texture placeholder (1x1 white) should be CRITICAL severity."""
+        p = get_pattern_by_category(GlitchCategory.TEXTURE_PLACEHOLDER)
+        self.assertIsNotNone(p)
+        self.assertEqual(p.severity, GlitchSeverity.CRITICAL)
+
+    def test_infer_severity_shader_failure(self):
+        """Shader failure should infer critical/high."""
+        self.assertEqual(_infer_severity("shader_failure", 0.8), "critical")
+        self.assertEqual(_infer_severity("shader_failure", 0.5), "high")
+
+    def test_infer_severity_texture_placeholder(self):
+        """Texture placeholder should infer critical/high."""
+        self.assertEqual(_infer_severity("texture_placeholder", 0.8), "critical")
+        self.assertEqual(_infer_severity("texture_placeholder", 0.5), "high")
+
+    def test_infer_severity_uv_mapping(self):
+        """UV mapping error should infer high/medium."""
+        self.assertEqual(_infer_severity("uv_mapping_error", 0.8), "high")
+        self.assertEqual(_infer_severity("uv_mapping_error", 0.5), "medium")
+
+    def test_infer_severity_frustum_culling(self):
+        """Frustum culling should infer medium/low."""
+        self.assertEqual(_infer_severity("frustum_culling", 0.7), "medium")
+        self.assertEqual(_infer_severity("frustum_culling", 0.4), "low")
+
+    def test_infer_severity_shadow_map(self):
+        """Shadow map artifact should infer medium/low."""
+        self.assertEqual(_infer_severity("shadow_map_artifact", 0.7), "medium")
+        self.assertEqual(_infer_severity("shadow_map_artifact", 0.4), "low")
+
+    def test_infer_severity_bloom_overflow(self):
+        """Bloom overflow should infer medium/low (default path)."""
+        self.assertEqual(_infer_severity("bloom_overflow", 0.7), "medium")
+        self.assertEqual(_infer_severity("bloom_overflow", 0.4), "low")
+
+    def test_threejs_patterns_in_vision_prompt(self):
+        """Three.js patterns should appear in the composite vision prompt."""
+        prompt = build_vision_prompt()
+        self.assertIn("shader_failure", prompt)
+        self.assertIn("texture_placeholder", prompt)
+        self.assertIn("uv_mapping_error", prompt)
+        self.assertIn("frustum_culling", prompt)
+        self.assertIn("shadow_map_artifact", prompt)
+        self.assertIn("bloom_overflow", prompt)
+
+    def test_threejs_subset_prompt(self):
+        """Building prompt from Three.js-only patterns should work."""
+        threejs = get_threejs_patterns()
+        prompt = build_vision_prompt(threejs)
+        self.assertIn("Shader Compilation Failure", prompt)
+        self.assertNotIn("Floating Object", prompt)  # generic, not Three.js
+
+    def test_report_metadata_version(self):
+        """Report metadata should reference both issues."""
+        report = run_demo()
+        self.assertEqual(report.metadata["detector_version"], "0.2.0")
+        self.assertIn("543", report.metadata["reference"])
+
+
+
 class TestIntegration(unittest.TestCase):
     """Integration-level tests."""
 
@@ -276,6 +368,13 @@ class TestIntegration(unittest.TestCase):
         expected = {"floating_assets", "z_fighting", "missing_textures", "clipping", "broken_normals"}
         self.assertTrue(expected.issubset(category_values))
 
+    def test_patterns_cover_threejs_themes(self):
+        """Patterns should cover Three.js-specific glitch themes (#543)."""
+        category_values = {p.category.value for p in MATRIX_GLITCH_PATTERNS}
+        threejs_expected = {"shader_failure", "texture_placeholder", "uv_mapping_error",
+                           "frustum_culling", "shadow_map_artifact", "bloom_overflow"}
+        self.assertTrue(threejs_expected.issubset(category_values))
+
 
 if __name__ == "__main__":
     unittest.main()

tests/test_scene_descriptions.py

@@ -0,0 +1,333 @@
+#!/usr/bin/env python3
+"""
+Tests for generate_scene_descriptions.py
+
+Tests the scene description generation pipeline including:
+- Media file scanning
+- Model detection
+- JSON parsing from vision responses
+- Output format validation
+
+Ref: timmy-config#689
+"""
+
+import json
+import tempfile
+import unittest
+from pathlib import Path
+from unittest.mock import MagicMock, patch
+
+# Add scripts to path for import
+import sys
+sys.path.insert(0, str(Path(__file__).resolve().parent.parent / "scripts"))
+
+from generate_scene_descriptions import (
+    IMAGE_EXTS,
+    VIDEO_EXTS,
+    ALL_EXTS,
+    VISION_MODELS,
+    auto_detect_model,
+    check_model_available,
+    scan_media,
+    extract_video_frame,
+)
+
+
+class TestMediaScanning(unittest.TestCase):
+    """Test media file scanning."""
+
+    def test_scan_empty_directory(self):
+        with tempfile.TemporaryDirectory() as tmpdir:
+            result = scan_media(tmpdir)
+            self.assertEqual(result, [])
+
+    def test_scan_nonexistent_directory(self):
+        result = scan_media("/nonexistent/path/that/does/not/exist")
+        self.assertEqual(result, [])
+
+    def test_scan_with_images(self):
+        with tempfile.TemporaryDirectory() as tmpdir:
+            # Create test files
+            for ext in [".jpg", ".png", ".webp"]:
+                (Path(tmpdir) / f"test{ext}").touch()
+
+            result = scan_media(tmpdir)
+            self.assertEqual(len(result), 3)
+
+    def test_scan_recursive(self):
+        with tempfile.TemporaryDirectory() as tmpdir:
+            subdir = Path(tmpdir) / "sub" / "dir"
+            subdir.mkdir(parents=True)
+            (subdir / "deep.jpg").touch()
+            (Path(tmpdir) / "top.png").touch()
+
+            result = scan_media(tmpdir)
+            self.assertEqual(len(result), 2)
+
+    def test_scan_ignores_unsupported(self):
+        with tempfile.TemporaryDirectory() as tmpdir:
+            (Path(tmpdir) / "image.jpg").touch()
+            (Path(tmpdir) / "document.pdf").touch()
+            (Path(tmpdir) / "script.py").touch()
+
+            result = scan_media(tmpdir)
+            self.assertEqual(len(result), 1)
+
+    def test_scan_sorted_output(self):
+        with tempfile.TemporaryDirectory() as tmpdir:
+            for name in ["z.jpg", "a.png", "m.webp"]:
+                (Path(tmpdir) / name).touch()
+
+            result = scan_media(tmpdir)
+            names = [p.name for p in result]
+            self.assertEqual(names, sorted(names))
+
+
+class TestModelDetection(unittest.TestCase):
+    """Test model availability detection."""
+
+    @patch('generate_scene_descriptions.urllib.request.urlopen')
+    def test_check_model_available(self, mock_urlopen):
+        mock_resp = MagicMock()
+        mock_resp.read.return_value = json.dumps({
+            "models": [{"name": "gemma4:latest"}]
+        }).encode()
+        mock_urlopen.return_value.__enter__ = MagicMock(return_value=mock_resp)
+        mock_urlopen.return_value.__exit__ = MagicMock(return_value=False)
+        mock_urlopen.return_value = mock_resp
+
+        result = check_model_available("gemma4:latest")
+        self.assertTrue(result)
+
+    @patch('generate_scene_descriptions.urllib.request.urlopen')
+    def test_check_model_not_available(self, mock_urlopen):
+        mock_resp = MagicMock()
+        mock_resp.read.return_value = json.dumps({
+            "models": [{"name": "llama2:7b"}]
+        }).encode()
+        mock_urlopen.return_value = mock_resp
+
+        result = check_model_available("gemma4:latest")
+        self.assertFalse(result)
+
+    @patch('generate_scene_descriptions.check_model_available')
+    def test_auto_detect_prefers_gemma4(self, mock_check):
+        def side_effect(model, url):
+            return model == "gemma4:latest"
+        mock_check.side_effect = side_effect
+
+        result = auto_detect_model()
+        self.assertEqual(result, "gemma4:latest")
+
+    @patch('generate_scene_descriptions.check_model_available')
+    def test_auto_detect_falls_back(self, mock_check):
+        def side_effect(model, url):
+            return model == "llava:latest"
+        mock_check.side_effect = side_effect
+
+        result = auto_detect_model()
+        self.assertEqual(result, "llava:latest")
+
+    @patch('generate_scene_descriptions.check_model_available')
+    def test_auto_detect_returns_none_when_no_models(self, mock_check):
+        mock_check.return_value = False
+        result = auto_detect_model()
+        self.assertIsNone(result)
+
+
+class TestConstants(unittest.TestCase):
+    """Test constant definitions."""
+
+    def test_image_extensions(self):
+        self.assertIn(".jpg", IMAGE_EXTS)
+        self.assertIn(".png", IMAGE_EXTS)
+        self.assertIn(".webp", IMAGE_EXTS)
+
+    def test_video_extensions(self):
+        self.assertIn(".mp4", VIDEO_EXTS)
+        self.assertIn(".webm", VIDEO_EXTS)
+
+    def test_all_extensions_union(self):
+        self.assertEqual(ALL_EXTS, IMAGE_EXTS | VIDEO_EXTS)
+
+    def test_vision_models_ordered(self):
+        self.assertEqual(VISION_MODELS[0], "gemma4:latest")
+        self.assertIn("llava:latest", VISION_MODELS)
+
+
+class TestVideoFrameExtraction(unittest.TestCase):
+    """Test video frame extraction."""
+
+    def test_extract_nonexistent_video(self):
+        result = extract_video_frame(Path("/nonexistent.mp4"), Path("/tmp/frame.jpg"))
+        self.assertFalse(result)
+
+
+class TestDescribeImage(unittest.TestCase):
+    """Test describe_image() with mocked Ollama responses."""
+
+    def test_skips_oversized_file(self):
+        """Files exceeding MAX_FILE_SIZE should be skipped without API call."""
+        import generate_scene_descriptions
+        with tempfile.NamedTemporaryFile(suffix=".jpg", delete=False) as f:
+            f.write(b"\x00" * (51 * 1024 * 1024))
+            f.flush()
+            result = generate_scene_descriptions.describe_image(Path(f.name))
+        Path(f.name).unlink()
+        self.assertIsNone(result)
+
+    @patch('generate_scene_descriptions.urllib.request.urlopen')
+    def test_parses_valid_json_response(self, mock_urlopen):
+        """Valid JSON response should be parsed and returned."""
+        import generate_scene_descriptions
+        resp_data = {
+            "response": '{"mood": "calm", "colors": ["blue", "white"], "composition": "wide-shot", "camera": "static", "lighting": "natural", "description": "A serene ocean scene."}'
+        }
+        mock_resp = MagicMock()
+        mock_resp.read.return_value = json.dumps(resp_data).encode()
+        mock_urlopen.return_value = mock_resp
+
+        with tempfile.NamedTemporaryFile(suffix=".jpg", delete=False) as f:
+            f.write(b"\xff\xd8\xff\xe0" + b"\x00" * 1000)
+            f.flush()
+            result = generate_scene_descriptions.describe_image(Path(f.name))
+        Path(f.name).unlink()
+
+        self.assertIsNotNone(result)
+        self.assertEqual(result["mood"], "calm")
+        self.assertIn("lighting", result)
+
+    @patch('generate_scene_descriptions.urllib.request.urlopen')
+    def test_repair_truncated_json(self, mock_urlopen):
+        """Truncated JSON should be repaired with regex extraction."""
+        import generate_scene_descriptions
+        resp_data = {
+            "response": '{"mood": "dark", "colors": ["red"], "composition": "close-up", "camera": "handheld", "lighting": "dramatic", "description": "A shadowy figure in a dimly lit alley'
+        }
+        mock_resp = MagicMock()
+        mock_resp.read.return_value = json.dumps(resp_data).encode()
+        mock_urlopen.return_value = mock_resp
+
+        with tempfile.NamedTemporaryFile(suffix=".jpg", delete=False) as f:
+            f.write(b"\xff\xd8\xff\xe0" + b"\x00" * 1000)
+            f.flush()
+            result = generate_scene_descriptions.describe_image(Path(f.name))
+        Path(f.name).unlink()
+
+        self.assertIsNotNone(result)
+        self.assertEqual(result["mood"], "dark")
+        self.assertEqual(result["lighting"], "dramatic")
+
+    @patch('generate_scene_descriptions.urllib.request.urlopen')
+    def test_fallback_on_invalid_json(self, mock_urlopen):
+        """Completely invalid JSON response should still return a fallback."""
+        import generate_scene_descriptions
+        resp_data = {"response": "This is just plain text describing a beautiful sunset over mountains."}
+        mock_resp = MagicMock()
+        mock_resp.read.return_value = json.dumps(resp_data).encode()
+        mock_urlopen.return_value = mock_resp
+
+        with tempfile.NamedTemporaryFile(suffix=".jpg", delete=False) as f:
+            f.write(b"\xff\xd8\xff\xe0" + b"\x00" * 1000)
+            f.flush()
+            result = generate_scene_descriptions.describe_image(Path(f.name))
+        Path(f.name).unlink()
+
+        self.assertIsNotNone(result)
+        self.assertIn("description", result)
+        self.assertIn("lighting", result)
+
+
+class TestDescribeImageSharegpt(unittest.TestCase):
+    """Test describe_image_sharegpt() with mocked Ollama responses."""
+
+    def test_skips_oversized_file(self):
+        """Files exceeding MAX_FILE_SIZE should be skipped."""
+        import generate_scene_descriptions
+        with tempfile.NamedTemporaryFile(suffix=".jpg", delete=False) as f:
+            f.write(b"\x00" * (51 * 1024 * 1024))
+            f.flush()
+            result = generate_scene_descriptions.describe_image_sharegpt(Path(f.name))
+        Path(f.name).unlink()
+        self.assertIsNone(result)
+
+    @patch('generate_scene_descriptions.urllib.request.urlopen')
+    def test_returns_natural_language(self, mock_urlopen):
+        """Should return the raw response text."""
+        import generate_scene_descriptions
+        resp_data = {"response": "A warm sunset over rolling hills with golden light."}
+        mock_resp = MagicMock()
+        mock_resp.read.return_value = json.dumps(resp_data).encode()
+        mock_urlopen.return_value = mock_resp
+
+        with tempfile.NamedTemporaryFile(suffix=".jpg", delete=False) as f:
+            f.write(b"\xff\xd8\xff\xe0" + b"\x00" * 1000)
+            f.flush()
+            result = generate_scene_descriptions.describe_image_sharegpt(Path(f.name))
+        Path(f.name).unlink()
+
+        self.assertIsNotNone(result)
+        self.assertIn("sunset", result)
+
+
+class TestGenerateTrainingPairs(unittest.TestCase):
+    """Test generate_training_pairs() orchestration."""
+
+    @patch('generate_scene_descriptions.describe_image')
+    def test_jsonl_output_format(self, mock_describe):
+        """JSONL format should produce structured description objects."""
+        import generate_scene_descriptions
+        mock_describe.return_value = {"mood": "calm", "description": "Test"}
+
+        with tempfile.NamedTemporaryFile(suffix=".jpg", delete=False) as f:
+            f.write(b"\x00" * 1000)
+            f.flush()
+            pairs = generate_scene_descriptions.generate_training_pairs(
+                [Path(f.name)], "test-model", "http://localhost:11434",
+                output_format="jsonl"
+            )
+        Path(f.name).unlink()
+
+        self.assertEqual(len(pairs), 1)
+        self.assertIn("description", pairs[0])
+        self.assertIn("generated_at", pairs[0])
+
+    @patch('generate_scene_descriptions.describe_image_sharegpt')
+    def test_sharegpt_output_format(self, mock_describe):
+        """ShareGPT format should produce conversation objects."""
+        import generate_scene_descriptions
+        mock_describe.return_value = "A description of the scene."
+
+        with tempfile.NamedTemporaryFile(suffix=".jpg", delete=False) as f:
+            f.write(b"\x00" * 1000)
+            f.flush()
+            pairs = generate_scene_descriptions.generate_training_pairs(
+                [Path(f.name)], "test-model", "http://localhost:11434",
+                output_format="sharegpt"
+            )
+        Path(f.name).unlink()
+
+        self.assertEqual(len(pairs), 1)
+        self.assertIn("conversations", pairs[0])
+        self.assertEqual(len(pairs[0]["conversations"]), 2)
+
+    @patch('generate_scene_descriptions.describe_image')
+    def test_dry_run_skips_api_calls(self, mock_describe):
+        """Dry run should not call describe_image."""
+        import generate_scene_descriptions
+        with tempfile.NamedTemporaryFile(suffix=".jpg", delete=False) as f:
+            f.write(b"\x00" * 1000)
+            f.flush()
+            pairs = generate_scene_descriptions.generate_training_pairs(
+                [Path(f.name)], "test-model", "http://localhost:11434",
+                dry_run=True
+            )
+        Path(f.name).unlink()
+
+        mock_describe.assert_not_called()
+        self.assertEqual(len(pairs), 1)
+        self.assertEqual(pairs[0]["status"], "dry-run")
+
+
+if __name__ == "__main__":
+    unittest.main()