feat: add model fallback verification script (#514)

- Tests model switches with verification prompts
- Validates context window meets 64K minimum
- Checks primary model and fallback chain
- Supports OpenRouter, Anthropic, Nous, Kimi, Ollama
- Exits with error if no viable model found

Closes #514

bin/model-fallback-verify.py

@@ -0,0 +1,443 @@
+#!/usr/bin/env python3
+"""
+Model Fallback Verification Script
+Issue #514: [Robustness] Model fallback verification — test before trusting
+
+Tests model switches with verification prompts, validates context windows,
+and ensures at least one viable model is available before starting loops.
+
+Usage:
+  python3 model-fallback-verify.py              # Run full verification
+  python3 model-fallback-verify.py check <model> # Test specific model
+  python3 model-fallback-verify.py context <model> # Check context window
+  python3 model-fallback-verify.py list          # List available models
+"""
+
+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"))
+CONFIG_FILE = HERMES_HOME / "config.yaml"
+LOG_DIR = HERMES_HOME / "logs"
+LOG_FILE = LOG_DIR / "model-verify.log"
+MIN_CONTEXT_WINDOW = 64 * 1024  # 64K tokens minimum
+
+# Provider endpoints
+PROVIDER_CONFIGS = {
+    "openrouter": {
+        "base_url": "https://openrouter.ai/api/v1",
+        "headers": lambda api_key: {"Authorization": "Bearer " + api_key},
+        "chat_url": "/chat/completions",
+    },
+    "anthropic": {
+        "base_url": "https://api.anthropic.com/v1",
+        "headers": lambda api_key: {"x-api-key": api_key, "anthropic-version": "2023-06-01"},
+        "chat_url": "/messages",
+    },
+    "nous": {
+        "base_url": "https://inference.nousresearch.com/v1",
+        "headers": lambda api_key: {"Authorization": "Bearer " + api_key},
+        "chat_url": "/chat/completions",
+    },
+    "kimi-coding": {
+        "base_url": "https://api.kimi.com/coding/v1",
+        "headers": lambda api_key: {"x-api-key": api_key, "x-api-provider": "kimi-coding"},
+        "chat_url": "/chat/completions",
+    },
+    "custom": {
+        "base_url": None,
+        "headers": lambda api_key: {"Authorization": "Bearer " + api_key},
+        "chat_url": "/chat/completions",
+    },
+}
+
+# Known context windows for common models
+KNOWN_CONTEXT_WINDOWS = {
+    "claude-opus-4-6": 200000,
+    "claude-sonnet-4": 200000,
+    "claude-3.5-sonnet": 200000,
+    "gpt-4o": 128000,
+    "gpt-4": 128000,
+    "gpt-3.5-turbo": 16385,
+    "qwen3:30b": 32768,
+    "qwen2.5:7b": 32768,
+    "hermes4:14b": 32768,
+    "gemma3:1b": 8192,
+    "gemma4": 32768,
+    "phi3:3.8b": 128000,
+    "kimi-k2.5": 128000,
+    "google/gemini-2.5-pro": 1048576,
+    "xiaomi/mimo-v2-pro": 131072,
+    "deepseek/deepseek-r1": 131072,
+    "deepseek/deepseek-chat-v3-0324": 131072,
+}
+
+def log(msg):
+    """Log message to file and optionally to 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 load_config():
+    """Load Hermes config.yaml."""
+    if not CONFIG_FILE.exists():
+        return None
+    
+    with open(CONFIG_FILE) as f:
+        return yaml.safe_load(f)
+
+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 get_ollama_models():
+    """Get list of available Ollama models."""
+    ollama_host = os.environ.get("OLLAMA_HOST", "localhost:11434")
+    try:
+        resp = urllib.request.urlopen("http://" + ollama_host + "/api/tags", timeout=5)
+        data = json.loads(resp.read())
+        return [m["name"] for m in data.get("models", [])]
+    except:
+        return []
+
+def test_model(model, provider, api_key=None, base_url=None):
+    """
+    Test a model with a verification prompt.
+    Returns (success, response, error_message)
+    """
+    if provider == "ollama" or ":" in model:
+        # Local Ollama model
+        ollama_host = os.environ.get("OLLAMA_HOST", "localhost:11434")
+        try:
+            body = json.dumps({
+                "model": model,
+                "prompt": "Say exactly VERIFIED and nothing else.",
+                "stream": False,
+                "options": {"num_predict": 10}
+            }).encode()
+            req = urllib.request.Request(
+                "http://" + ollama_host + "/api/generate",
+                data=body,
+                headers={"Content-Type": "application/json"}
+            )
+            resp = urllib.request.urlopen(req, timeout=30)
+            result = json.loads(resp.read())
+            response_text = result.get("response", "").strip()
+            if "VERIFIED" in response_text.upper():
+                return True, response_text, None
+            return False, response_text, "Unexpected response: " + response_text[:100]
+        except Exception as e:
+            return False, "", "Ollama error: " + str(e)[:200]
+    
+    # Cloud provider
+    config = PROVIDER_CONFIGS.get(provider)
+    if not config:
+        return False, "", "Unknown provider: " + provider
+    
+    url = base_url or config["base_url"]
+    if not url:
+        return False, "", "No base URL for provider: " + provider
+    
+    headers = config["headers"](api_key or "")
+    headers["Content-Type"] = "application/json"
+    
+    try:
+        body = json.dumps({
+            "model": model,
+            "max_tokens": 20,
+            "messages": [{"role": "user", "content": "Say exactly VERIFIED and nothing else."}]
+        }).encode()
+        
+        req = urllib.request.Request(
+            url + config["chat_url"],
+            data=body,
+            headers=headers
+        )
+        resp = urllib.request.urlopen(req, timeout=30)
+        result = json.loads(resp.read())
+        
+        if provider == "anthropic":
+            content = result.get("content", [{}])[0].get("text", "")
+        else:
+            choices = result.get("choices", [{}])
+            content = choices[0].get("message", {}).get("content", "") if choices else ""
+        
+        if "VERIFIED" in content.upper():
+            return True, content, None
+        return False, content, "Unexpected response: " + content[:100]
+    
+    except urllib.error.HTTPError as e:
+        error_body = e.read().decode() if e.fp else str(e)
+        if e.code == 404:
+            return False, "", "Model not found (404): " + error_body[:200]
+        elif e.code == 429:
+            return True, "", "Rate limited but model exists"
+        elif e.code >= 500:
+            return False, "", "Server error (" + str(e.code) + "): " + error_body[:200]
+        else:
+            return False, "", "HTTP " + str(e.code) + ": " + error_body[:200]
+    except Exception as e:
+        return False, "", "Request error: " + str(e)[:200]
+
+def get_context_window(model, provider):
+    """
+    Get the context window size for a model.
+    Returns (window_size, source)
+    """
+    if model in KNOWN_CONTEXT_WINDOWS:
+        return KNOWN_CONTEXT_WINDOWS[model], "known"
+    
+    model_lower = model.lower()
+    if "claude" in model_lower:
+        return 200000, "inferred (claude)"
+    elif "gpt-4" in model_lower:
+        return 128000, "inferred (gpt-4)"
+    elif "gemini" in model_lower:
+        return 1048576, "inferred (gemini)"
+    elif "qwen" in model_lower:
+        return 32768, "inferred (qwen)"
+    elif "gemma" in model_lower:
+        return 8192, "inferred (gemma)"
+    elif "phi" in model_lower:
+        return 128000, "inferred (phi)"
+    
+    return 32768, "default"
+
+def verify_model(model, provider, api_key=None, base_url=None):
+    """
+    Full verification of a model: test prompt + context window.
+    Returns dict with verification results.
+    """
+    result = {
+        "model": model,
+        "provider": provider,
+        "tested": False,
+        "responded": False,
+        "response": "",
+        "error": None,
+        "context_window": 0,
+        "context_source": "unknown",
+        "meets_minimum": False,
+        "viable": False,
+    }
+    
+    success, response, error = test_model(model, provider, api_key, base_url)
+    result["tested"] = True
+    result["responded"] = success
+    result["response"] = response[:200] if response else ""
+    result["error"] = error
+    
+    window, source = get_context_window(model, provider)
+    result["context_window"] = window
+    result["context_source"] = source
+    result["meets_minimum"] = window >= MIN_CONTEXT_WINDOW
+    
+    result["viable"] = success and result["meets_minimum"]
+    
+    return result
+
+def get_fallback_chain(config):
+    """Get the fallback chain from config or defaults."""
+    chain = []
+    
+    model_config = config.get("model", {})
+    if isinstance(model_config, dict):
+        primary = model_config.get("default", "")
+        provider = model_config.get("provider", "")
+        if primary and provider:
+            chain.append({"model": primary, "provider": provider, "role": "primary"})
+    elif model_config:
+        chain.append({"model": str(model_config), "provider": "unknown", "role": "primary"})
+    
+    auxiliary = config.get("auxiliary", {})
+    for aux_name, aux_config in auxiliary.items():
+        if isinstance(aux_config, dict):
+            aux_model = aux_config.get("model", "")
+            aux_provider = aux_config.get("provider", "")
+            if aux_model and aux_provider and aux_provider != "auto":
+                chain.append({"model": aux_model, "provider": aux_provider, "role": "auxiliary:" + aux_name})
+    
+    ollama_models = get_ollama_models()
+    for model in ollama_models[:3]:
+        if not any(c["model"] == model for c in chain):
+            chain.append({"model": model, "provider": "ollama", "role": "local-fallback"})
+    
+    return chain
+
+def run_verification():
+    """Run full model fallback verification."""
+    log("=== Model Fallback Verification ===")
+    
+    config = load_config()
+    if not config:
+        log("ERROR: No config.yaml found")
+        return {"success": False, "error": "No config file"}
+    
+    chain = get_fallback_chain(config)
+    if not chain:
+        log("ERROR: No models configured")
+        return {"success": False, "error": "No models in chain"}
+    
+    results = []
+    viable_models = []
+    
+    for entry in chain:
+        model = entry["model"]
+        provider = entry["provider"]
+        role = entry["role"]
+        
+        api_key = get_provider_api_key(provider) if provider != "ollama" else None
+        
+        base_url = None
+        if provider == "custom":
+            provider_config = config.get("auxiliary", {}).get("vision", {})
+            base_url = provider_config.get("base_url")
+        
+        log("Testing [" + role + "] " + model + " (" + provider + ")...")
+        result = verify_model(model, provider, api_key, base_url)
+        result["role"] = role
+        results.append(result)
+        
+        status = "PASS" if result["viable"] else "FAIL"
+        details = []
+        if not result["responded"]:
+            details.append("no response: " + str(result["error"]))
+        if not result["meets_minimum"]:
+            details.append("context " + str(result["context_window"]) + " < " + str(MIN_CONTEXT_WINDOW))
+        
+        log("  [" + status + "] " + model + " - " + (", ".join(details) if details else "verified"))
+        
+        if result["viable"]:
+            viable_models.append(result)
+    
+    log("=== Results: " + str(len(viable_models)) + "/" + str(len(results)) + " models viable ===")
+    
+    if not viable_models:
+        log("CRITICAL: No viable models found!")
+        for r in results:
+            log("  - " + r["model"] + " (" + r["provider"] + "): responded=" + str(r["responded"]) + ", context=" + str(r["context_window"]))
+        return {"success": False, "results": results, "viable": []}
+    
+    log("Viable models (in priority order):")
+    for i, r in enumerate(viable_models, 1):
+        log("  " + str(i) + ". " + r["model"] + " (" + r["provider"] + ") - context: " + str(r["context_window"]) + " tokens [" + r["role"] + "]")
+    
+    return {
+        "success": True,
+        "results": results,
+        "viable": viable_models,
+        "primary": viable_models[0] if viable_models else None,
+    }
+
+def check_single_model(model):
+    """Check a specific model."""
+    if ":" in model:
+        provider = "ollama"
+    elif "/" in model:
+        provider = "openrouter"
+    else:
+        provider = "unknown"
+    
+    config = load_config() or {}
+    api_key = get_provider_api_key(provider) if provider != "ollama" else None
+    
+    result = verify_model(model, provider, api_key)
+    
+    if result["viable"]:
+        print("PASS: " + model)
+        print("  Context window: " + str(result["context_window"]) + " tokens")
+        print("  Response: " + result["response"][:100])
+    else:
+        print("FAIL: " + model)
+        if result["error"]:
+            print("  Error: " + str(result["error"]))
+        if not result["meets_minimum"]:
+            print("  Context window: " + str(result["context_window"]) + " < " + str(MIN_CONTEXT_WINDOW) + " minimum")
+    
+    return result["viable"]
+
+def check_context_window(model):
+    """Check context window for a model."""
+    if ":" in model:
+        provider = "ollama"
+    elif "/" in model:
+        provider = "openrouter"
+    else:
+        provider = "unknown"
+    
+    window, source = get_context_window(model, provider)
+    meets = window >= MIN_CONTEXT_WINDOW
+    
+    print("Model: " + model)
+    print("Provider: " + provider)
+    print("Context window: " + str(window) + " tokens (" + source + ")")
+    print("Minimum (" + str(MIN_CONTEXT_WINDOW) + "): " + ("PASS" if meets else "FAIL"))
+    
+    return meets
+
+def list_models():
+    """List all available models."""
+    config = load_config() or {}
+    chain = get_fallback_chain(config)
+    
+    print("Configured models:")
+    for entry in chain:
+        print("  " + entry["model"].ljust(30) + " " + entry["provider"].ljust(15) + " [" + entry["role"] + "]")
+    
+    ollama = get_ollama_models()
+    if ollama:
+        print("")
+        print("Ollama models:")
+        for m in ollama:
+            print("  " + m)
+
+def main():
+    if len(sys.argv) < 2:
+        result = run_verification()
+        sys.exit(0 if result["success"] else 1)
+    
+    cmd = sys.argv[1]
+    
+    if cmd == "check" and len(sys.argv) > 2:
+        model = sys.argv[2]
+        success = check_single_model(model)
+        sys.exit(0 if success else 1)
+    
+    elif cmd == "context" and len(sys.argv) > 2:
+        model = sys.argv[2]
+        meets = check_context_window(model)
+        sys.exit(0 if meets else 1)
+    
+    elif cmd == "list":
+        list_models()
+    
+    elif cmd == "test":
+        result = run_verification()
+        sys.exit(0 if result["success"] else 1)
+    
+    else:
+        print("Usage:")
+        print("  model-fallback-verify.py              Run full verification")
+        print("  model-fallback-verify.py check <model> Test specific model")
+        print("  model-fallback-verify.py context <model> Check context window")
+        print("  model-fallback-verify.py list          List available models")
+        sys.exit(1)
+
+if __name__ == "__main__":
+    main()