loreguard-cli 0.16.0__tar.gz → 0.20.3__tar.gz

{loreguard_cli-0.16.0 → loreguard_cli-0.20.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: loreguard-cli
-Version: 0.16.0
+Version: 0.20.3
 Summary: Local inference client for Loreguard NPCs
 Project-URL: Homepage, https://loreguard.com
 Project-URL: Documentation, https://github.com/beyond-logic-labs/loreguard-cli#readme
@@ -29,7 +29,7 @@ Requires-Dist: rich>=13.0.0
 Requires-Dist: textual>=0.47.0
 Requires-Dist: tf-keras>=2.16.0
 Requires-Dist: torch>=2.0.0
-Requires-Dist: transformers>=5.0.0
+Requires-Dist: transformers<5,>=4.36.0
 Requires-Dist: uvicorn>=0.27.0
 Requires-Dist: websockets>=12.0
 Provides-Extra: build

{loreguard_cli-0.16.0 → loreguard_cli-0.20.3}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "loreguard-cli"
-version = "0.16.0"
+version = "0.20.3"
 description = "Local inference client for Loreguard NPCs"
 readme = "README.md"
 license = "MIT"
@@ -28,7 +28,7 @@ dependencies = [
     "aiofiles>=24.1.0",
     "rich>=13.0.0",
     "textual>=0.47.0",
-    "transformers>=5.0.0",
+    "transformers>=4.36.0,<5",
     "torch>=2.0.0",
     "fastapi>=0.109.0",
     "uvicorn>=0.27.0",

{loreguard_cli-0.16.0 → loreguard_cli-0.20.3}/src/__main__.py

@@ -28,6 +28,29 @@ def main():
         print(json.dumps(status, indent=2))
         sys.exit(0 if status.get("running") else 1)
 
+    # Handle 'download-llama-server' command - for bundle tool delegation (ADR-0027)
+    if args and args[0] == "download-llama-server":
+        import asyncio
+        from pathlib import Path
+        from .llama_server import download_llama_server
+
+        output_dir = None
+        for i, a in enumerate(args):
+            if a == "--output-dir" and i + 1 < len(args):
+                output_dir = Path(args[i + 1])
+
+        if not output_dir:
+            print("Usage: loreguard download-llama-server --output-dir <path>", file=sys.stderr)
+            sys.exit(1)
+
+        output_dir.mkdir(parents=True, exist_ok=True)
+
+        def on_progress(msg, progress=None):
+            print(f"       {msg}")
+
+        asyncio.run(download_llama_server(progress_callback=on_progress, target_dir=output_dir))
+        sys.exit(0)
+
     # Filter out help flags - these should show CLI help
     if any(a in ('-h', '--help') for a in args):
         from .cli import main as cli_main

{loreguard_cli-0.16.0 → loreguard_cli-0.20.3}/src/cli.py

@@ -11,7 +11,8 @@ Environment variables (alternative to args):
     LOREGUARD_MODEL     Path to model file
     LOREGUARD_MODEL_ID  Model ID to download (if not using custom model)
     LOREGUARD_PORT      Local llama-server port (default: 8080)
-    LOREGUARD_BACKEND   Backend URL (default: wss://api.loreguard.com/workers)
+    LOREGUARD_BACKEND   Backend WebSocket URL (default: wss://console.loreguard.com/workers)
+    LOREGUARD_API       API base URL (default: https://console.loreguard.com)
     LOREGUARD_WORKER_ID Worker ID (default: hostname)
 """
 
@@ -26,6 +27,8 @@ from datetime import datetime
 from pathlib import Path
 from typing import Optional
 
+from .config import DEFAULT_API_URL, DEFAULT_BACKEND_URL
+
 # Setup logging
 logging.basicConfig(
     level=logging.INFO,
@@ -44,7 +47,7 @@ class LoreguardCLI:
         model_path: Optional[Path] = None,
         model_id: Optional[str] = None,
         port: int = 8080,
-        backend_url: str = "wss://api.loreguard.com/workers",
+        backend_url: str = DEFAULT_BACKEND_URL,
         worker_id: Optional[str] = None,
         model_family: str = "llama3",
     ):
@@ -454,9 +457,14 @@ Available model IDs:
     )
     parser.add_argument(
         "--backend",
-        default=os.getenv("LOREGUARD_BACKEND", "wss://api.loreguard.com/workers"),
+        default=os.getenv("LOREGUARD_BACKEND", DEFAULT_BACKEND_URL),
         help="Backend WebSocket URL",
     )
+    parser.add_argument(
+        "--api-url",
+        default=os.getenv("LOREGUARD_API", DEFAULT_API_URL),
+        help=f"API base URL (default: {DEFAULT_API_URL})",
+    )
     parser.add_argument(
         "-v", "--verbose",
         action="store_true",

{loreguard_cli-0.16.0 → loreguard_cli-0.20.3}/src/config.py

@@ -50,6 +50,7 @@ class LoreguardConfig:
     context_size: int = 16384  # llama-server context window size (configurable per game)
     max_speech_tokens: int = 50  # Max tokens for NPC speech output (Pass 4). Default: 50 (~40 words)
     model_family: str = "auto"  # Model family profile (auto, llama3, qwen3, gemma, chatml)
+    dialogue_act_enabled: bool = False  # Dialogue act classifier for filler selection
 
     def save(self) -> None:
         """Save configuration to disk."""
@@ -73,6 +74,7 @@ class LoreguardConfig:
                     context_size=data.get("context_size", 16384),
                     max_speech_tokens=data.get("max_speech_tokens", 50),
                     model_family=data.get("model_family", "auto"),
+                    dialogue_act_enabled=data.get("dialogue_act_enabled", False),
                 )
             except (json.JSONDecodeError, KeyError):
                 pass
@@ -121,6 +123,14 @@ class LoreguardConfig:
 # Environment Variable Configuration
 # =============================================================================
 
+DEFAULT_API_URL = "https://console.loreguard.com"
+DEFAULT_BACKEND_URL = "wss://console.loreguard.com/workers"
+
+
+def get_api_url() -> str:
+    """Get the Loreguard API base URL (configurable via LOREGUARD_API env var)."""
+    return os.getenv("LOREGUARD_API", DEFAULT_API_URL)
+
 
 @lru_cache(maxsize=1)
 def load_config() -> dict:
@@ -133,12 +143,13 @@ def load_config() -> dict:
     return {
         # Server settings
         "LLM_ENDPOINT": os.getenv("LLM_ENDPOINT", "http://localhost:8080"),
-        "BACKEND_URL": os.getenv("LOREGUARD_BACKEND", "wss://api.loreguard.com/workers"),
+        "BACKEND_URL": os.getenv("LOREGUARD_BACKEND", DEFAULT_BACKEND_URL),
+        "API_URL": os.getenv("LOREGUARD_API", DEFAULT_API_URL),
         "HOST": os.getenv("HOST", "127.0.0.1"),
         "PORT": os.getenv("PORT", "8081"),
 
         # Worker authentication (required for backend connection)
-        # Get API token from loreguard.com dashboard
+        # Get API token from console.loreguard.com
         "WORKER_ID": os.getenv("LOREGUARD_WORKER_ID", os.getenv("WORKER_ID", "")),
         # LOREGUARD_TOKEN is preferred, WORKER_TOKEN kept for backwards compatibility
         "LOREGUARD_TOKEN": os.getenv("LOREGUARD_TOKEN", os.getenv("WORKER_TOKEN", "")),
@@ -232,20 +243,21 @@ def get_models_dir() -> Optional[Path]:
 
 
 def resolve_model_path(model_name: str, subdir: str = "") -> str:
-    """Resolve a model path, preferring pre-shipped models over HF downloads.
+    """Resolve a model path, preferring local models over HF downloads.
 
     Resolution order:
     1. LOREGUARD_MODELS_DIR/<subdir>  (explicit override)
-    2. Bundle models dir using manifest.txt  (HF name → manifest key → local dir)
-    3. Bundle models dir using HF name → org--model convention  (fallback)
-    4. Original HF model name  (download from HuggingFace)
+    2. Application Support models dir/<subdir>  (standard install location)
+    3. Bundle models dir using manifest.txt  (HF name → manifest key → local dir)
+    4. Bundle models dir using HF name → org--model convention  (fallback)
+    5. Download from HuggingFace to Application Support models dir
 
     Args:
         model_name: HuggingFace model name (e.g., 'vectara/hallucination_evaluation_model')
         subdir: Subdirectory within MODELS_DIR to check (e.g., 'hhem', 'deberta')
 
     Returns:
-        Local path if pre-shipped model found, otherwise the original HF model name.
+        Local path to the model directory.
     """
     # 1. Explicit LOREGUARD_MODELS_DIR/<subdir>
     explicit_dir = get_config_value("MODELS_DIR")
@@ -254,7 +266,14 @@ def resolve_model_path(model_name: str, subdir: str = "") -> str:
         if local_path.exists() and any(local_path.iterdir()):
             return str(local_path)
 
-    # 2 & 3. Bundle directory resolution
+    # 2. Application Support models dir/<subdir>
+    app_models = get_data_dir() / "models"
+    if subdir:
+        local_path = app_models / subdir
+        if local_path.exists() and any(local_path.iterdir()):
+            return str(local_path)
+
+    # 3 & 4. Bundle directory resolution
     bundle_dir = get_bundle_dir()
     if bundle_dir:
         bundle_models = bundle_dir / "models"
@@ -275,9 +294,37 @@ def resolve_model_path(model_name: str, subdir: str = "") -> str:
         if local_path.exists() and any(local_path.iterdir()):
             return str(local_path)
 
+    # 5. Download from HuggingFace to Application Support models dir
+    if subdir:
+        return _download_hf_model(model_name, app_models / subdir)
+
     return model_name
 
 
+def _download_hf_model(model_name: str, target_dir: Path) -> str:
+    """Download a HuggingFace model to the loreguard models directory.
+
+    Returns:
+        Path to the downloaded model directory.
+    """
+    import logging
+    logger = logging.getLogger(__name__)
+    try:
+        from huggingface_hub import snapshot_download
+        target_dir.mkdir(parents=True, exist_ok=True)
+        logger.info(f"Downloading {model_name} to {target_dir}")
+        snapshot_download(
+            model_name,
+            local_dir=str(target_dir),
+            local_dir_use_symlinks=False,
+        )
+        logger.info(f"Downloaded {model_name} to {target_dir}")
+        return str(target_dir)
+    except Exception as e:
+        logger.warning(f"Failed to download {model_name}: {e}")
+        return model_name
+
+
 def get_config_value(key: str, default: Optional[str] = None) -> Optional[str]:
     """Get a single configuration value."""
     config = load_config()

{loreguard_cli-0.16.0 → loreguard_cli-0.20.3}/src/http_server.py

@@ -367,7 +367,7 @@ class EmbeddedHTTPServer:
 
             # Derive HTTP base URL from WebSocket URL
             # ws://localhost:8090/workers → http://localhost:8090
-            # wss://api.loreguard.com/workers → https://api.loreguard.com
+            # wss://console.loreguard.com/workers → https://console.loreguard.com
             backend_ws = server.tunnel.backend_url
             if backend_ws.startswith("wss://"):
                 base_url = "https://" + backend_ws[6:].split("/")[0]
@@ -535,15 +535,15 @@ class EmbeddedHTTPServer:
                     content={"error": "Missing 'model' field"},
                 )
 
-            # Security: prevent path traversal
-            if "/" in model_name or "\\" in model_name or ".." in model_name:
+            # Security: resolve and verify path stays inside models_dir
+            model_path = (server.models_dir / model_name).resolve()
+            if model_path.parent != server.models_dir.resolve():
                 return JSONResponse(
                     status_code=400,
                     content={"error": "Invalid model name"},
                 )
 
-            model_path = server.models_dir / model_name
-            if not model_path.exists() or not model_path.suffix == ".gguf":
+            if not model_path.exists() or model_path.suffix != ".gguf":
                 return JSONResponse(
                     status_code=404,
                     content={"error": f"Model '{model_name}' not found"},
@@ -553,6 +553,9 @@ class EmbeddedHTTPServer:
             if hasattr(server.llama_process, "model_path") and server.llama_process.model_path.name == model_name:
                 return {"status": "already_active", "model": model_name}
 
+            # Save original model_path for rollback on failure
+            original_model_path = server.llama_process.model_path
+
             try:
                 # Stop current llama-server
                 server.llama_process.stop()
@@ -564,21 +567,35 @@ class EmbeddedHTTPServer:
                 # Wait for health check (llama-server takes a few seconds to load model)
                 import httpx
                 llama_url = f"http://127.0.0.1:{server.llama_process.port}/health"
-                for attempt in range(60):  # 60 attempts × 0.5s = 30s timeout
-                    await asyncio.sleep(0.5)
-                    try:
-                        async with httpx.AsyncClient(timeout=2.0) as client:
+                async with httpx.AsyncClient(timeout=2.0) as client:
+                    for attempt in range(60):  # 60 attempts × 0.5s = 30s timeout
+                        await asyncio.sleep(0.5)
+                        try:
                             resp = await client.get(llama_url)
                             if resp.status_code == 200:
+                                # Persist selection so it survives restarts
+                                try:
+                                    from .config import LoreguardConfig
+                                    cfg = LoreguardConfig.load()
+                                    cfg.set_model_path(model_path)
+                                    cfg.save()
+                                except Exception:
+                                    pass  # Best-effort persistence
                                 return {"status": "ok", "model": model_name}
-                    except Exception:
-                        continue
+                        except Exception:
+                            continue
 
                 return JSONResponse(
                     status_code=500,
                     content={"error": "Model loaded but health check timed out after 30s"},
                 )
             except Exception as e:
+                # Rollback: restore original model path and try to restart
+                server.llama_process.model_path = original_model_path
+                try:
+                    server.llama_process.start()
+                except Exception:
+                    pass  # Best-effort rollback
                 return JSONResponse(
                     status_code=500,
                     content={"error": f"Failed to reload model: {e}"},

{loreguard_cli-0.16.0 → loreguard_cli-0.20.3}/src/llama_server.py

@@ -30,7 +30,7 @@ def _get_templates_dir() -> Path:
     return Path(__file__).parent.parent / "templates"
 
 
-LLAMA_VERSION = "b7789"  # Must match loreguard-engine bundle version
+LLAMA_VERSION = "b8467"  # Must match loreguard-engine bundle version
 
 # Download URLs for each platform
 BINARIES = {
@@ -265,18 +265,21 @@ def make_executable(path: Path) -> None:
 
 async def download_llama_server(
     progress_callback: Optional[Callable[[str, DownloadProgress | None], None]] = None,
+    target_dir: Optional[Path] = None,
 ) -> Path:
     """Download and install llama-server for the current platform.
 
     Args:
         progress_callback: Called with (status_message, progress_or_none)
+        target_dir: If provided, install into this directory instead of the default.
+                    Used by the bundle tool to pre-ship llama-server.
 
     Returns:
         Path to the installed llama-server binary
     """
     plat = get_platform()
     config = BINARIES[plat]
-    bin_dir = get_bin_dir()
+    bin_dir = target_dir or get_bin_dir()
 
     def notify(msg: str, progress: DownloadProgress | None = None):
         if progress_callback:
@@ -355,12 +358,12 @@ async def download_llama_server(
                 make_executable(lib)
 
         # Write version marker file for future version checks
-        version_file = get_version_file_path()
+        version_file = bin_dir / ".llama_version" if target_dir else get_version_file_path()
         version_file.write_text(LLAMA_VERSION)
 
         notify(f"llama-server {LLAMA_VERSION} installed successfully!")
 
-    return get_llama_server_path()
+    return bin_dir / config["binary_name"]
 
 
 class LlamaServerProcess:

{loreguard_cli-0.16.0 → loreguard_cli-0.20.3}/src/llm.py

@@ -61,7 +61,9 @@ class LLMRequest:
     stop: list[str] = field(default_factory=lambda: DEFAULT_STOP_SEQUENCES.copy())
 
     # Thinking mode control (for Qwen3)
-    disable_thinking: bool = False
+    # Defaults to True: thinking wastes tokens and breaks pipelines.
+    # Only enable explicitly when extended reasoning is desired.
+    disable_thinking: bool = True
 
     # If true, error if content is empty instead of falling back to reasoning_content
     require_content: bool = False
@@ -257,9 +259,10 @@ class LLMProxy:
             payload["id_slot"] = 0
             logger.info("KV cache: cache_prompt=true, id_slot=0 (verify -np 1 on server)")
 
-        # Disable thinking mode if requested (for Qwen3)
+        # Disable thinking mode (for Qwen3/3.5).
+        # Must use chat_template_kwargs — top-level enable_thinking is ignored by llama.cpp b8467+.
         if req.disable_thinking:
-            payload["enable_thinking"] = False
+            payload.setdefault("chat_template_kwargs", {})["enable_thinking"] = False
 
         # Note: JSON mode is not compatible with streaming in llama.cpp
         # If force_json is requested, fall back to non-streaming
@@ -573,9 +576,10 @@ class LLMProxy:
             payload["id_slot"] = 0
             logger.info("KV cache: cache_prompt=true, id_slot=0 (verify -np 1 on server)")
 
-        # Disable thinking mode if requested (for Qwen3)
+        # Disable thinking mode (for Qwen3/3.5).
+        # Must use chat_template_kwargs — top-level enable_thinking is ignored by llama.cpp b8467+.
         if req.disable_thinking:
-            payload["enable_thinking"] = False
+            payload.setdefault("chat_template_kwargs", {})["enable_thinking"] = False
 
         # Force JSON output if requested
         if req.force_json:

{loreguard_cli-0.16.0 → loreguard_cli-0.20.3}/src/main.py

@@ -31,7 +31,7 @@ from rich.console import Console
 
 from .tunnel import BackendTunnel
 from .llm import LLMProxy
-from .config import get_config_value, resolve_model_path
+from .config import get_config_value, resolve_model_path, DEFAULT_BACKEND_URL
 from .nli import NLIService, is_nli_model_available
 from .intent_classifier import IntentClassifier, is_intent_model_available
 from .dialogue_act_classifier import (
@@ -118,7 +118,7 @@ async def startup():
         console.print("[yellow]Intent classifier disabled (set LOREGUARD_INTENT_ENABLED=true to enable)[/yellow]")
 
     # Initialize dialogue act classifier (optional, for filler selection)
-    enable_dialogue_act = os.getenv("LOREGUARD_DIALOGUE_ACT_ENABLED", "true").lower() == "true"
+    enable_dialogue_act = os.getenv("LOREGUARD_DIALOGUE_ACT_ENABLED", "false").lower() == "true"
     if enable_dialogue_act:
         console.print("[cyan]Initializing dialogue act classifier...[/cyan]")
         if is_dialogue_act_model_available():
@@ -170,7 +170,7 @@ async def startup():
             chunk_detector = None
 
     # Connect to remote backend
-    backend_url = get_config_value("BACKEND_URL", "wss://api.lorekeeper.ai/workers")
+    backend_url = get_config_value("BACKEND_URL", DEFAULT_BACKEND_URL)
     worker_id = get_config_value("WORKER_ID", "")
     worker_token = get_config_value("WORKER_TOKEN", "")
     model_id = get_config_value("MODEL_ID", "default")

{loreguard_cli-0.16.0 → loreguard_cli-0.20.3}/src/models_registry.py

@@ -92,6 +92,18 @@ SUPPORTED_MODELS: list[ModelInfo] = [
         hardware="32GB RAM • 20GB VRAM",
         recommended=False,
     ),
+    ModelInfo(
+        id="qwen3.5-9b-q4km",
+        name="Qwen 3.5 9B Q4_K_M",
+        filename="Qwen3.5-9B-Q4_K_M.gguf",
+        size_gb=5.2,
+        size_bytes=5_627_044_256,
+        context_length=32768,
+        url="https://huggingface.co/unsloth/Qwen3.5-9B-GGUF/resolve/main/Qwen3.5-9B-Q4_K_M.gguf",
+        description="Strong general model. 32K context, good reasoning.",
+        hardware="12GB RAM • 8GB VRAM",
+        recommended=False,
+    ),
 ]
 
 

{loreguard_cli-0.16.0 → loreguard_cli-0.20.3}/src/nli.py

@@ -328,6 +328,26 @@ class NLIService:
 
             return results
 
+    def _resolve_model_dir(self) -> Optional[str]:
+        """Resolve the actual directory containing model files.
+
+        For local paths, returns as-is. For HuggingFace repo IDs (e.g.
+        'vectara/hallucination_evaluation_model'), resolves the cache snapshot dir.
+        """
+        if os.path.isdir(self._model_path):
+            return self._model_path
+        try:
+            from huggingface_hub import snapshot_download
+            return snapshot_download(self._model_path, local_files_only=True)
+        except Exception:
+            # Cache miss — trigger a download so the snapshot exists
+            try:
+                from huggingface_hub import snapshot_download
+                return snapshot_download(self._model_path)
+            except Exception as e:
+                logger.warning(f"Could not resolve model dir for {self._model_path}: {e}")
+                return None
+
     def _patch_hhem_model_files(self):
         """Patch vendored HHEM files for transformers 5.x compatibility.
 
@@ -336,8 +356,31 @@ class NLIService:
         2. Is stricter about model_type matching between config.json and config class
         Since trust_remote_code loads the .py files directly, we patch before loading.
         """
+        model_dir = self._resolve_model_dir()
+        if not model_dir:
+            logger.warning("Cannot resolve model directory for patching")
+            return
+
+        # Also patch the modules cache (transformers copies .py files there)
+        modules_dirs = [model_dir]
+        modules_cache = os.path.join(
+            os.path.expanduser("~"), ".cache", "huggingface", "modules",
+            "transformers_modules",
+        )
+        if os.path.isdir(modules_cache):
+            for root, dirs, files in os.walk(modules_cache):
+                if "modeling_hhem_v2.py" in files:
+                    modules_dirs.append(root)
+
+        for patch_dir in modules_dirs:
+            self._patch_dir(patch_dir)
+
+    def _patch_dir(self, patch_dir: str):
+        """Apply HHEM patches to a single directory."""
         # Patch 1: modeling_hhem_v2.py — add missing class attributes
-        model_file = os.path.join(self._model_path, "modeling_hhem_v2.py")
+        model_file = os.path.join(patch_dir, "modeling_hhem_v2.py")
+        if not os.path.exists(model_file):
+            return
         if os.path.exists(model_file):
             try:
                 content = open(model_file, "r").read()
@@ -354,13 +397,18 @@ class NLIService:
                     if patched != content:
                         with open(model_file, "w") as f:
                             f.write(patched)
-                        logger.info("Patched modeling_hhem_v2.py for transformers 5.x")
+                        # Clear __pycache__ so patched file is reloaded
+                        pycache = os.path.join(patch_dir, "__pycache__")
+                        if os.path.isdir(pycache):
+                            import shutil
+                            shutil.rmtree(pycache)
+                        logger.info(f"Patched {model_file} for transformers 5.x")
             except Exception as e:
                 logger.warning(f"Could not patch modeling_hhem_v2.py: {e}")
 
         # Patch 2: config.json — fix model_type mismatch
         # config.json has "HHEMv2Config" but the config class defines model_type = "HHEMv2"
-        config_file = os.path.join(self._model_path, "config.json")
+        config_file = os.path.join(patch_dir, "config.json")
         if os.path.exists(config_file):
             try:
                 content = open(config_file, "r").read()
@@ -375,24 +423,40 @@ class NLIService:
             except Exception as e:
                 logger.warning(f"Could not patch config.json: {e}")
 
-        # Patch 3: configuration_hhem_v2.py — use local flan-t5-base instead of HuggingFace
-        # The HHEM model downloads google/flan-t5-base config+tokenizer at init.
-        # If we've bundled those files locally, rewrite the foundation path.
-        config_py = os.path.join(self._model_path, "configuration_hhem_v2.py")
-        local_foundation = os.path.join(self._model_path, "flan-t5-base")
+        # Patch 3: configuration_hhem_v2.py, point the foundation at the bundled
+        # flan-t5-base. Resolve it RELATIVE to the config file at runtime (via
+        # __file__) so the path is portable and never a baked absolute machine path.
+        config_py = os.path.join(patch_dir, "configuration_hhem_v2.py")
+        local_foundation = os.path.join(patch_dir, "flan-t5-base")
         if os.path.exists(config_py) and os.path.isdir(local_foundation):
             try:
+                import re
+
                 content = open(config_py, "r").read()
-                if '"google/flan-t5-base"' in content:
-                    # Use absolute path to the bundled flan-t5-base files
-                    abs_path = os.path.abspath(local_foundation)
-                    patched = content.replace(
-                        '"google/flan-t5-base"',
-                        f'"{abs_path}"',
+                resolver = (
+                    'os.path.join(os.path.dirname(os.path.abspath(__file__)), '
+                    '"flan-t5-base")'
+                )
+                patched = content
+                if "import os" not in patched:
+                    patched = patched.replace(
+                        "from transformers import PretrainedConfig",
+                        "from transformers import PretrainedConfig\nimport os",
+                        1,
                     )
+                # Replace the class-level foundation assignment, whatever it is now
+                # (the HuggingFace default OR a previously-baked absolute path), with
+                # the self-resolving expression. count=1 leaves any __init__ default.
+                patched = re.sub(
+                    r'foundation\s*=\s*"[^"]*"',
+                    "foundation = " + resolver,
+                    patched,
+                    count=1,
+                )
+                if patched != content:
                     with open(config_py, "w") as f:
                         f.write(patched)
-                    logger.info(f"Patched foundation to local: {abs_path}")
+                    logger.info("Patched foundation to self-resolving relative path")
             except Exception as e:
                 logger.warning(f"Could not patch configuration_hhem_v2.py: {e}")
 

{loreguard_cli-0.16.0 → loreguard_cli-0.20.3}/src/npc_chat.py

@@ -15,6 +15,7 @@ Rate Limits (when using Player JWT):
 
 import asyncio
 import logging
+import os
 from dataclasses import dataclass
 from typing import Optional
 
@@ -35,8 +36,9 @@ from .term_ui import (
 # Configure module logger
 logger = logging.getLogger(__name__)
 
-# Loreguard API base URL
-LOREGUARD_API_URL = "https://api.loreguard.com"
+# Loreguard API base URL (configurable via LOREGUARD_API env var)
+from .config import get_api_url
+LOREGUARD_API_URL = get_api_url()
 
 
 @dataclass
@@ -407,7 +409,7 @@ class NPCChat:
                 raise Exception("Invalid API token - please check your authentication")
             elif e.response.status_code == 404:
                 logger.warning("No characters found for this account")
-                raise Exception("No characters found - register NPCs at loreguard.com first")
+                raise Exception("No characters found - register NPCs at console.loreguard.com first")
             logger.error("HTTP error fetching characters: %d", e.response.status_code)
             raise
         except httpx.RequestError as e:
@@ -428,7 +430,7 @@ class NPCChat:
             return None
 
         if not characters:
-            print_error("No NPCs registered. Create NPCs at loreguard.com first.")
+            print_error("No NPCs registered. Create NPCs at console.loreguard.com first.")
             return None
 
         items = [
@@ -642,7 +644,7 @@ async def run_npc_chat(
     Args:
         api_token: Loreguard API token for authentication (for server-side use)
         player_jwt: Player JWT from Steam exchange (for game clients)
-        base_url: Loreguard API base URL (default: https://api.loreguard.com)
+        base_url: Loreguard API base URL (default: https://console.loreguard.com)
         config: Optional client configuration for timeouts
         verbose: If True, show pipeline pass updates via WebSocket
         tunnel: BackendTunnel instance for receiving pass_update messages (required for verbose)

{loreguard_cli-0.16.0 → loreguard_cli-0.20.3}/src/steam.py

@@ -12,7 +12,7 @@ Usage:
         read_timeout=15.0,
         max_retries=3
     )
-    steam_auth = SteamAuth(api_url="https://api.loreguard.com", config=config)
+    steam_auth = SteamAuth(api_url="https://console.loreguard.com", config=config)
 
     # Exchange Steam ticket for Player JWT
     result = await steam_auth.exchange_ticket(
@@ -40,8 +40,9 @@ import httpx
 # Configure module logger
 logger = logging.getLogger(__name__)
 
-# Default Loreguard API URL
-LOREGUARD_API_URL = "https://api.loreguard.com"
+# Default Loreguard API URL (configurable via LOREGUARD_API env var)
+from .config import get_api_url
+LOREGUARD_API_URL = get_api_url()
 
 # Validation patterns
 STEAM_APP_ID_PATTERN = re.compile(r"^\d{1,10}$")

{loreguard_cli-0.16.0 → loreguard_cli-0.20.3}/src/tui/modals/auth_menu.py

@@ -12,6 +12,7 @@ from textual.widgets import Input, Static, ListView, ListItem, Label
 from rich.text import Text
 
 from ..styles import PURPLE, CYAN, PINK, FG, FG_DIM, GREEN, RED
+from ...config import get_api_url
 from ..widgets.banner import get_gradient_color
 
 if TYPE_CHECKING:
@@ -227,7 +228,7 @@ class AuthMenuModal(ModalScreen[tuple | None]):
 
         # Update status
         status = self.query_one("#status-line", Static)
-        status.update(Text("Get your token at loreguard.com/dashboard", style=FG_DIM))
+        status.update(Text("Get your token at console.loreguard.com", style=FG_DIM))
 
     def _switch_to_menu(self) -> None:
         """Switch back to menu mode."""
@@ -270,7 +271,7 @@ class AuthMenuModal(ModalScreen[tuple | None]):
         try:
             async with httpx.AsyncClient(timeout=10.0) as client:
                 response = await client.get(
-                    "https://api.loreguard.com/api/auth/me",
+                    f"{get_api_url()}/api/auth/me",
                     headers={"Authorization": f"Bearer {token}"},
                 )
                 if response.status_code == 200:

{loreguard_cli-0.16.0 → loreguard_cli-0.20.3}/src/tui/modals/token_input.py

@@ -59,7 +59,7 @@ class TokenInputModal(ModalScreen[str | None]):
         """Compose the modal layout."""
         with Vertical():
             yield Static("Enter API Token", classes="modal-title")
-            yield Static("Get your token at loreguard.com/dashboard", classes="modal-hint")
+            yield Static("Get your token at console.loreguard.com", classes="modal-hint")
             yield Input(placeholder="Paste your token here...", password=True, id="token-input")
             yield Static("enter submit • esc cancel", classes="modal-footer")
 

{loreguard_cli-0.16.0 → loreguard_cli-0.20.3}/src/tui/modals/unified_palette.py

@@ -284,7 +284,7 @@ class UnifiedPaletteModal(ModalScreen[tuple[str, Any] | None]):
         # Add models
         if self._show_models:
             from ...llama_server import get_models_dir
-            from ...models_registry import SUPPORTED_MODELS
+            from ...models_registry import SUPPORTED_MODELS, ModelInfo
             from ...hf_discovery import discover_models
             from ..widgets.hardware_info import detect_hardware
 
@@ -302,6 +302,33 @@ class UnifiedPaletteModal(ModalScreen[tuple[str, Any] | None]):
             except Exception:
                 all_models = list(SUPPORTED_MODELS)
 
+            # Ensure all static registry models are included
+            known_filenames = {m.filename for m in all_models}
+            for model in SUPPORTED_MODELS:
+                if model.filename not in known_filenames:
+                    all_models.append(model)
+                    known_filenames.add(model.filename)
+
+            # Scan local models dir for GGUF files not in registry/discovery
+            if self._models_dir and self._models_dir.exists():
+                for gguf_file in self._models_dir.glob("*.gguf"):
+                    if gguf_file.name not in known_filenames:
+                        size_bytes = gguf_file.stat().st_size
+                        size_gb = size_bytes / (1024 ** 3)
+                        stem = gguf_file.stem
+                        all_models.append(ModelInfo(
+                            id=f"local-{stem.lower()}",
+                            name=stem.replace("-", " ").replace("_", " "),
+                            filename=gguf_file.name,
+                            size_gb=round(size_gb, 1),
+                            size_bytes=size_bytes,
+                            context_length=8192,
+                            url="",
+                            description="Local model",
+                            hardware="",
+                        ))
+                        known_filenames.add(gguf_file.name)
+
             # Sort: most recent first, then by size descending
             def model_sort_key(m):
                 # Primary: sort by recency (days_ago), None goes last

{loreguard_cli-0.16.0 → loreguard_cli-0.20.3}/src/tui/screens/auth.py

@@ -16,6 +16,7 @@ from ..widgets.banner import LoreguardBanner
 from ..widgets.hardware_info import HardwareInfo
 from ..widgets.footer import LoreguardFooter
 from ..styles import CYAN, PINK, GREEN, RED, FG_DIM
+from ...config import get_api_url
 
 if TYPE_CHECKING:
     from ..app import LoreguardApp
@@ -115,7 +116,7 @@ class AuthScreen(Screen):
         try:
             async with httpx.AsyncClient(timeout=10.0) as client:
                 response = await client.get(
-                    "https://api.loreguard.com/api/auth/me",
+                    f"{get_api_url()}/api/auth/me",
                     headers={"Authorization": f"Bearer {token}"},
                 )
                 if response.status_code == 200:

{loreguard_cli-0.16.0 → loreguard_cli-0.20.3}/src/tui/screens/main.py

@@ -16,7 +16,7 @@ from ..widgets.hardware_info import HardwareInfo
 from ..widgets.server_monitor import ServerMonitor
 from ..widgets.npc_chat import NPCChat
 from ..widgets.footer import LoreguardFooter
-from ...config import LoreguardConfig
+from ...config import LoreguardConfig, get_api_url, DEFAULT_BACKEND_URL
 
 if TYPE_CHECKING:
     from ..app import LoreguardApp
@@ -450,7 +450,7 @@ class MainScreen(Screen):
                     self._log(f"Still loading... ({elapsed}s)")
 
                 # Check if process died
-                if app._llama_process.process and app._llama_process.process.poll() is not None:
+                if app._llama_process and app._llama_process.process and app._llama_process.process.poll() is not None:
                     self._log("llama-server process died", "error")
                     break
 
@@ -548,7 +548,7 @@ class MainScreen(Screen):
 
             # Load dialogue act classifier (filler selection) - run in thread pool
             dialogue_act_classifier = None
-            enable_dialogue_act = os.getenv("LOREGUARD_DIALOGUE_ACT_ENABLED", "true").lower() == "true"
+            enable_dialogue_act = os.getenv("LOREGUARD_DIALOGUE_ACT_ENABLED", "false").lower() == "true"
             if not enable_dialogue_act:
                 self._log("Dialogue act classifier disabled via LOREGUARD_DIALOGUE_ACT_ENABLED")
             else:
@@ -649,7 +649,7 @@ class MainScreen(Screen):
                     self._update_connection_status("connecting")
 
             app._tunnel = BackendTunnel(
-                backend_url="wss://api.loreguard.com/workers",
+                backend_url=os.getenv("LOREGUARD_BACKEND", DEFAULT_BACKEND_URL),
                 llm_proxy=llm_proxy,
                 worker_id=app.worker_id,
                 worker_token=app.api_token,
@@ -851,7 +851,7 @@ class MainScreen(Screen):
         try:
             async with httpx.AsyncClient(timeout=10.0) as client:
                 response = await client.get(
-                    "https://api.loreguard.com/api/characters",
+                    f"{get_api_url()}/api/characters",
                     headers={"Authorization": f"Bearer {app.api_token}"},
                 )
 
@@ -862,7 +862,7 @@ class MainScreen(Screen):
                     npcs = [c for c in characters if c.get("type") != "world"]
 
                     if not npcs:
-                        self._update_status("No NPCs registered. Create NPCs at loreguard.com first.")
+                        self._update_status("No NPCs registered. Create NPCs at console.loreguard.com first.")
                         return
 
                     # Create NPC items

{loreguard_cli-0.16.0 → loreguard_cli-0.20.3}/src/tui/screens/running.py

@@ -17,6 +17,7 @@ from ..widgets.banner import LoreguardBanner
 from ..widgets.hardware_info import HardwareInfo
 from ..widgets.footer import LoreguardFooter
 from ..styles import CYAN, GREEN, YELLOW, RED, FG_DIM
+from ...config import DEFAULT_BACKEND_URL
 
 if TYPE_CHECKING:
     from ..app import LoreguardApp
@@ -256,7 +257,7 @@ class RunningScreen(Screen):
 
                 # Load Dialogue Act Classifier
                 dialogue_act_classifier = None
-                enable_dialogue_act = os.getenv("LOREGUARD_DIALOGUE_ACT_ENABLED", "true").lower() == "true"
+                enable_dialogue_act = os.getenv("LOREGUARD_DIALOGUE_ACT_ENABLED", "false").lower() == "true"
                 if not enable_dialogue_act:
                     self._update_status("dialogue_act", "Dialogue Act", "Disabled", "info")
                     self._log("Dialogue act classifier disabled via LOREGUARD_DIALOGUE_ACT_ENABLED", "info")
@@ -322,7 +323,7 @@ class RunningScreen(Screen):
                 model_id = app.model_path.stem
 
                 self._tunnel = BackendTunnel(
-                    backend_url="wss://api.loreguard.com/workers",
+                    backend_url=os.getenv("LOREGUARD_BACKEND", DEFAULT_BACKEND_URL),
                     llm_proxy=llm_proxy,
                     worker_id=app.worker_id,
                     worker_token=app.api_token,

{loreguard_cli-0.16.0 → loreguard_cli-0.20.3}/src/tui/widgets/npc_chat.py

@@ -6,6 +6,7 @@ Uses the local proxy for NPC conversations with token streaming:
 
 import json
 import logging
+import os
 from typing import TYPE_CHECKING
 
 import httpx
@@ -23,8 +24,9 @@ from ...runtime import RuntimeInfo
 if TYPE_CHECKING:
     from ..app import LoreguardApp
 
-# Fallback to cloud API if local proxy unavailable
-LOREGUARD_API_URL = "https://api.loreguard.com"
+# Fallback to cloud API if local proxy unavailable (configurable via LOREGUARD_API env var)
+from ...config import get_api_url
+LOREGUARD_API_URL = get_api_url()
 
 
 def get_local_proxy_url() -> str | None:

{loreguard_cli-0.16.0 → loreguard_cli-0.20.3}/src/tunnel.py

@@ -81,7 +81,8 @@ class BackendTunnel:
         self.registered = False
         self.backend_version = ""  # Populated from worker_ack
         self._reconnect_delay = 1  # Start with 1 second
-        self._max_reconnect_delay = 60  # Max 60 seconds
+        self._max_reconnect_delay = 3  # Cap at 3s until first successful connection
+        self._has_connected = False  # Tracks if we've ever connected successfully
         self._running = True
         self._shutdown_requested = False
         self._heartbeat_task: asyncio.Task | None = None
@@ -122,7 +123,7 @@ class BackendTunnel:
         """Establish and maintain connection to backend with auto-reconnect."""
         if not self.worker_id or not self.worker_token:
             self._log("Error: Worker ID and API token are required", "error")
-            self._log("Get an API token from loreguard.com dashboard", "warn")
+            self._log("Get an API token from console.loreguard.com", "warn")
             return
 
         last_error = ""
@@ -210,6 +211,11 @@ class BackendTunnel:
         connection_start = time.time()
         self._log("Connected to backend!", "success")
 
+        # After first successful connection, use longer backoff for reconnections
+        if not self._has_connected:
+            self._has_connected = True
+            self._max_reconnect_delay = 60
+
         # Register as worker
         success, error_reason = await self._register_worker()
         if not success:

{loreguard_cli-0.16.0 → loreguard_cli-0.20.3}/src/wizard.py

@@ -54,6 +54,8 @@ from rich.box import ROUNDED
 from rich.align import Align
 from rich.layout import Layout
 
+from .config import get_api_url, DEFAULT_BACKEND_URL
+
 # Logger instance
 log = logging.getLogger("loreguard")
 
@@ -1044,7 +1046,7 @@ async def step_authentication(app: Optional[TUIApp] = None) -> tuple[Optional[st
     try:
         async with httpx.AsyncClient(timeout=10.0) as client:
             response = await client.get(
-                "https://api.loreguard.com/api/auth/me",
+                f"{get_api_url()}/api/auth/me",
                 headers={"Authorization": f"Bearer {token}"},
             )
             if response.status_code == 200:
@@ -1510,7 +1512,7 @@ async def step_start(
 
             model_id = _resolve_backend_model_id(model_path.stem)
             tunnel = BackendTunnel(
-                backend_url="wss://api.loreguard.com/workers",
+                backend_url=os.getenv("LOREGUARD_BACKEND", DEFAULT_BACKEND_URL),
                 llm_proxy=llm_proxy,
                 worker_id=worker_id,
                 worker_token=token,

{loreguard_cli-0.16.0 → loreguard_cli-0.20.3}/uv.lock

@@ -476,7 +476,7 @@ wheels = [
 
 [[package]]
 name = "huggingface-hub"
-version = "0.36.0"
+version = "0.36.2"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "filelock" },
@@ -488,9 +488,9 @@ dependencies = [
     { name = "tqdm" },
     { name = "typing-extensions" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/98/63/4910c5fa9128fdadf6a9c5ac138e8b1b6cee4ca44bf7915bbfbce4e355ee/huggingface_hub-0.36.0.tar.gz", hash = "sha256:47b3f0e2539c39bf5cde015d63b72ec49baff67b6931c3d97f3f84532e2b8d25", size = 463358, upload-time = "2025-10-23T12:12:01.413Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/7c/b7/8cb61d2eece5fb05a83271da168186721c450eb74e3c31f7ef3169fa475b/huggingface_hub-0.36.2.tar.gz", hash = "sha256:1934304d2fb224f8afa3b87007d58501acfda9215b334eed53072dd5e815ff7a", size = 649782, upload-time = "2026-02-06T09:24:13.098Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/cb/bd/1a875e0d592d447cbc02805fd3fe0f497714d6a2583f59d14fa9ebad96eb/huggingface_hub-0.36.0-py3-none-any.whl", hash = "sha256:7bcc9ad17d5b3f07b57c78e79d527102d08313caa278a641993acddcb894548d", size = 566094, upload-time = "2025-10-23T12:11:59.557Z" },
+    { url = "https://files.pythonhosted.org/packages/a8/af/48ac8483240de756d2438c380746e7130d1c6f75802ef22f3c6d49982787/huggingface_hub-0.36.2-py3-none-any.whl", hash = "sha256:48f0c8eac16145dfce371e9d2d7772854a4f591bcb56c9cf548accf531d54270", size = 566395, upload-time = "2026-02-06T09:24:11.133Z" },
 ]
 
 [[package]]
@@ -600,7 +600,7 @@ wheels = [
 
 [[package]]
 name = "loreguard-cli"
-version = "0.14.5"
+version = "0.16.0"
 source = { editable = "." }
 dependencies = [
     { name = "aiofiles" },
@@ -649,7 +649,7 @@ requires-dist = [
     { name = "textual", specifier = ">=0.47.0" },
     { name = "tf-keras", specifier = ">=2.16.0" },
     { name = "torch", specifier = ">=2.0.0" },
-    { name = "transformers", specifier = ">=4.36.0" },
+    { name = "transformers", specifier = ">=4.36.0,<5" },
     { name = "uvicorn", specifier = ">=0.27.0" },
     { name = "websockets", specifier = ">=12.0" },
 ]
@@ -2226,7 +2226,7 @@ wheels = [
 
 [[package]]
 name = "transformers"
-version = "4.57.3"
+version = "4.57.6"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "filelock" },
@@ -2241,9 +2241,9 @@ dependencies = [
     { name = "tokenizers" },
     { name = "tqdm" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/dd/70/d42a739e8dfde3d92bb2fff5819cbf331fe9657323221e79415cd5eb65ee/transformers-4.57.3.tar.gz", hash = "sha256:df4945029aaddd7c09eec5cad851f30662f8bd1746721b34cc031d70c65afebc", size = 10139680, upload-time = "2025-11-25T15:51:30.139Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/c4/35/67252acc1b929dc88b6602e8c4a982e64f31e733b804c14bc24b47da35e6/transformers-4.57.6.tar.gz", hash = "sha256:55e44126ece9dc0a291521b7e5492b572e6ef2766338a610b9ab5afbb70689d3", size = 10134912, upload-time = "2026-01-16T10:38:39.284Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/6a/6b/2f416568b3c4c91c96e5a365d164f8a4a4a88030aa8ab4644181fdadce97/transformers-4.57.3-py3-none-any.whl", hash = "sha256:c77d353a4851b1880191603d36acb313411d3577f6e2897814f333841f7003f4", size = 11993463, upload-time = "2025-11-25T15:51:26.493Z" },
+    { url = "https://files.pythonhosted.org/packages/03/b8/e484ef633af3887baeeb4b6ad12743363af7cce68ae51e938e00aaa0529d/transformers-4.57.6-py3-none-any.whl", hash = "sha256:4c9e9de11333ddfe5114bc872c9f370509198acf0b87a832a0ab9458e2bd0550", size = 11993498, upload-time = "2026-01-16T10:38:31.289Z" },
 ]
 
 [[package]]