loreguard-cli 0.14.0__tar.gz → 0.14.1__tar.gz

{loreguard_cli-0.14.0 → loreguard_cli-0.14.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: loreguard-cli
-Version: 0.14.0
+Version: 0.14.1
 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

loreguard_cli-0.14.1/loreguard.spec

@@ -0,0 +1,42 @@
+# -*- mode: python ; coding: utf-8 -*-
+from PyInstaller.utils.hooks import collect_submodules
+
+hiddenimports = ['src', 'src.config', 'src.llm', 'src.llama_server', 'src.tunnel', 'src.term_ui', 'src.models_registry', 'src.cli', 'src.npc_chat', 'httpx', 'websockets', 'aiofiles', 'pydantic']
+hiddenimports += collect_submodules('src')
+
+
+a = Analysis(
+    ['loreguard_entry.py'],
+    pathex=['.'],
+    binaries=[],
+    datas=[('templates', 'templates')],
+    hiddenimports=hiddenimports,
+    hookspath=[],
+    hooksconfig={},
+    runtime_hooks=[],
+    excludes=[],
+    noarchive=False,
+    optimize=0,
+)
+pyz = PYZ(a.pure)
+
+exe = EXE(
+    pyz,
+    a.scripts,
+    a.binaries,
+    a.datas,
+    [],
+    name='loreguard',
+    debug=False,
+    bootloader_ignore_signals=False,
+    strip=False,
+    upx=True,
+    upx_exclude=[],
+    runtime_tmpdir=None,
+    console=True,
+    disable_windowed_traceback=False,
+    argv_emulation=False,
+    target_arch=None,
+    codesign_identity=None,
+    entitlements_file=None,
+)

{loreguard_cli-0.14.0 → loreguard_cli-0.14.1}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "loreguard-cli"
-version = "0.14.0"
+version = "0.14.1"
 description = "Local inference client for Loreguard NPCs"
 readme = "README.md"
 license = "MIT"
@@ -62,3 +62,8 @@ packages = ["src"]
 [tool.ruff]
 line-length = 100
 target-version = "py310"
+
+[dependency-groups]
+dev = [
+    "pyinstaller>=6.17.0",
+]

{loreguard_cli-0.14.0 → loreguard_cli-0.14.1}/scripts/build.py

@@ -36,6 +36,8 @@ def main():
         "--clean",
         # Add src to path so imports work
         "--paths", ".",
+        # Include templates directory (jinja chat templates for llama-server)
+        "--add-data", "templates:templates",
         # Collect all src submodules
         "--collect-submodules", "src",
         # Add hidden imports that PyInstaller might miss

{loreguard_cli-0.14.0 → loreguard_cli-0.14.1}/src/cli.py

@@ -52,8 +52,11 @@ class LoreguardCLI:
         self.model_id = model_id
         self.port = port
         self.backend_url = backend_url
-        # Worker ID: use provided value, or default to hostname
-        self.worker_id = worker_id or socket.gethostname() or "worker"
+        # Worker ID: use provided value, or default to sanitized hostname.
+        # Validator requires ^[a-zA-Z0-9_-]{1,64}$ — replace dots with hyphens.
+        raw_id = worker_id or socket.gethostname() or "worker"
+        import re
+        self.worker_id = re.sub(r'[^a-zA-Z0-9_-]', '-', raw_id)[:64]
 
         self._llama = None
         self._tunnel = None
@@ -275,25 +278,8 @@ class LoreguardCLI:
             except Exception as e:
                 log.warning(f"Intent classifier error: {e}")
 
-            # Initialize dialogue act classifier
+            # Dialogue act classifier disabled
             dialogue_act_classifier = None
-            try:
-                from .dialogue_act_classifier import (
-                    DialogueActClassifier,
-                    is_dialogue_act_model_available,
-                )
-                if is_dialogue_act_model_available():
-                    log.info("Loading dialogue act classifier...")
-                    dialogue_act_classifier = DialogueActClassifier()
-                    if dialogue_act_classifier.load_model():
-                        log.info(f"Dialogue act classifier ready (device: {dialogue_act_classifier.device})")
-                    else:
-                        log.warning("Dialogue act classifier failed to load")
-                        dialogue_act_classifier = None
-                else:
-                    log.info("Dialogue act model not available, skipping")
-            except Exception as e:
-                log.warning(f"Dialogue act classifier error: {e}")
 
             # Initialize chunk detector (ADR-0023) - shares model with intent classifier
             chunk_detector = None
@@ -469,6 +455,11 @@ Available model IDs:
         action="store_true",
         help="Enable debug logging and show pipeline pass updates (in wizard mode)",
     )
+    parser.add_argument(
+        "--bundle-dir",
+        default=os.getenv("LOREGUARD_BUNDLE_DIR", ""),
+        help="Loreguard bundle directory. Auto-discovers models from manifest.txt.",
+    )
     parser.add_argument(
         "--dev",
         action="store_true",
@@ -485,6 +476,10 @@ Available model IDs:
     if args.verbose:
         logging.getLogger().setLevel(logging.DEBUG)
 
+    # Propagate --bundle-dir to env so config.py picks it up via get_bundle_dir()
+    if args.bundle_dir:
+        os.environ["LOREGUARD_BUNDLE_DIR"] = args.bundle_dir
+
     # Chat mode - test NPC chat directly via API (no model needed)
     if args.chat:
         if not args.token:
@@ -509,8 +504,17 @@ Available model IDs:
         args.token = "dev_mock_token"
         log.info("Running in DEV MODE - no backend connection")
     else:
-        # Validate token is present (server will validate format)
-        if not args.token:
+        # Local bundle backends (ws:// to localhost/127.0.0.1) run with RequireAuth=false,
+        # so any non-empty token is accepted. Only require a real token for cloud backends.
+        backend = args.backend
+        is_local_backend = (
+            backend.startswith("ws://") and
+            any(backend.startswith(f"ws://{h}") for h in ("localhost", "127.0.0.1", "[::1]"))
+        )
+        if not args.token and is_local_backend:
+            # Local bundle backends run with RequireAuth=false — any non-empty token works.
+            args.token = "local"
+        elif not args.token:
             log.error("Token required. Use --token or set LOREGUARD_TOKEN (or use --dev)")
             sys.exit(1)
 

{loreguard_cli-0.14.0 → loreguard_cli-0.14.1}/src/config.py

@@ -161,25 +161,83 @@ def load_config() -> dict:
         # Pre-shipped llama-server binary path (enterprise bundles).
         # When set, skips auto-download and uses this binary directly.
         "LLAMA_SERVER_PATH": os.getenv("LOREGUARD_LLAMA_SERVER_PATH", ""),
+
+        # Bundle directory (set by game launchers that ship a loreguard bundle).
+        # When set, the client auto-discovers models from manifest.txt inside the bundle.
+        # This is the single env var a game needs to set — no per-model configuration required.
+        "BUNDLE_DIR": os.getenv("LOREGUARD_BUNDLE_DIR", ""),
     }
 
 
+def get_bundle_dir() -> Optional[Path]:
+    """Get the loreguard bundle directory, if configured via LOREGUARD_BUNDLE_DIR.
+
+    Game launchers set this to the bundle root so the client can auto-discover
+    models from manifest.txt without any per-model configuration.
+    """
+    bundle_dir = get_config_value("BUNDLE_DIR")
+    if bundle_dir:
+        path = Path(bundle_dir)
+        if path.exists() and path.is_dir():
+            return path
+    return None
+
+
+def get_bundle_manifest() -> dict:
+    """Parse the bundle's manifest.txt into a logical-name → dir-name mapping.
+
+    Returns an empty dict if no bundle dir is configured or manifest is missing.
+
+    Manifest format:
+        nli=vectara--hallucination_evaluation_model
+        embedding=BAAI--bge-small-en-v1.5
+        reranker=cross-encoder--ms-marco-MiniLM-L-6-v2
+    """
+    bundle_dir = get_bundle_dir()
+    if not bundle_dir:
+        return {}
+    manifest_path = bundle_dir / "models" / "manifest.txt"
+    if not manifest_path.exists():
+        return {}
+    result = {}
+    for line in manifest_path.read_text().splitlines():
+        line = line.strip()
+        if not line or line.startswith("#"):
+            continue
+        if "=" in line:
+            key, _, value = line.partition("=")
+            result[key.strip()] = value.strip()
+    return result
+
+
 def get_models_dir() -> Optional[Path]:
     """Get the pre-shipped models directory, if configured (ADR-0027).
 
-    Returns None if not set, meaning models should be auto-downloaded from HF.
+    Checks LOREGUARD_MODELS_DIR first, then falls back to the bundle's models dir.
+    Returns None if neither is set, meaning models should be auto-downloaded from HF.
     """
     models_dir = get_config_value("MODELS_DIR")
     if models_dir:
         path = Path(models_dir)
         if path.exists() and path.is_dir():
             return path
+    bundle_dir = get_bundle_dir()
+    if bundle_dir:
+        path = bundle_dir / "models"
+        if path.exists() and path.is_dir():
+            return path
     return None
 
 
 def resolve_model_path(model_name: str, subdir: str = "") -> str:
     """Resolve a model path, preferring pre-shipped 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)
+
     Args:
         model_name: HuggingFace model name (e.g., 'vectara/hallucination_evaluation_model')
         subdir: Subdirectory within MODELS_DIR to check (e.g., 'hhem', 'deberta')
@@ -187,11 +245,34 @@ def resolve_model_path(model_name: str, subdir: str = "") -> str:
     Returns:
         Local path if pre-shipped model found, otherwise the original HF model name.
     """
-    models_dir = get_models_dir()
-    if models_dir and subdir:
-        local_path = models_dir / subdir
+    # 1. Explicit LOREGUARD_MODELS_DIR/<subdir>
+    explicit_dir = get_config_value("MODELS_DIR")
+    if explicit_dir and subdir:
+        local_path = Path(explicit_dir) / subdir
         if local_path.exists() and any(local_path.iterdir()):
             return str(local_path)
+
+    # 2 & 3. Bundle directory resolution
+    bundle_dir = get_bundle_dir()
+    if bundle_dir:
+        bundle_models = bundle_dir / "models"
+
+        # Try manifest.txt: find the dir name for this HF model name
+        manifest = get_bundle_manifest()
+        # The manifest uses org--model naming (/ replaced by --)
+        hf_as_dir = model_name.replace("/", "--")
+        for _key, dir_name in manifest.items():
+            if dir_name == hf_as_dir:
+                local_path = bundle_models / dir_name
+                if local_path.exists() and any(local_path.iterdir()):
+                    return str(local_path)
+                break
+
+        # Fallback: check bundle models dir using org--model convention directly
+        local_path = bundle_models / hf_as_dir
+        if local_path.exists() and any(local_path.iterdir()):
+            return str(local_path)
+
     return model_name
 
 

{loreguard_cli-0.14.0 → loreguard_cli-0.14.1}/src/llama_server.py

@@ -22,6 +22,14 @@ from typing import AsyncGenerator, Callable, Optional
 
 import httpx
 
+
+def _get_templates_dir() -> Path:
+    """Return the templates directory, handling PyInstaller onefile bundles."""
+    if getattr(sys, "frozen", False) and hasattr(sys, "_MEIPASS"):
+        return Path(sys._MEIPASS) / "templates"
+    return Path(__file__).parent.parent / "templates"
+
+
 LLAMA_VERSION = "b7789"  # Must match loreguard-engine bundle version
 
 # Download URLs for each platform
@@ -402,7 +410,7 @@ class LlamaServerProcess:
             # Llama 3.1's built-in template forces tool-calling format even without tools,
             # so we use a stripped-down template that only handles chat messages.
             "--jinja",
-            "--chat-template-file", str(Path(__file__).parent.parent / "templates" / "llama31-no-tools.jinja"),
+            "--chat-template-file", str(_get_templates_dir() / "llama31-no-tools.jinja"),
         ]
 
         # Add LoRA adapter if specified

{loreguard_cli-0.14.0 → loreguard_cli-0.14.1}/src/llm.py

@@ -797,8 +797,8 @@ class LLMProxy:
             safe_filename = self._validate_cache_filename(filename)
 
             response = await self.client.post(
-                f"{self.endpoint}/slots/{slot_id}?action=save&filename={safe_filename}",
-                json={},  # llama-server expects JSON body
+                f"{self.endpoint}/slots/{slot_id}?action=save",
+                json={"filename": safe_filename},
                 timeout=30.0,
             )
             if response.status_code == 200:
@@ -889,8 +889,8 @@ class LLMProxy:
             safe_filename = self._validate_cache_filename(filename)
 
             response = await self.client.post(
-                f"{self.endpoint}/slots/{slot_id}?action=restore&filename={safe_filename}",
-                json={},  # llama-server expects JSON body
+                f"{self.endpoint}/slots/{slot_id}?action=restore",
+                json={"filename": safe_filename},
                 timeout=30.0,
             )
             if response.status_code == 200:

{loreguard_cli-0.14.0 → loreguard_cli-0.14.1}/src/models_registry.py

@@ -52,58 +52,44 @@ class AdapterInfo:
 
 
 # Supported models for NPC inference
-# Fine-tuned Loreguard models with multi-pass pipeline training
-# Based on unsloth/Llama-3.1-8B-Instruct with Unsloth Dynamic (UD) quantization
-# UD = per-layer optimized quantization for better accuracy at same VRAM
+# Fine-tuned Loreguard NPC model based on Llama 3.1 8B Instruct
+# https://huggingface.co/beyond-logic-labs/loreguard-npc-llama3.1-8b-gguf
 # Ordered by recommendation (best first)
 SUPPORTED_MODELS: list[ModelInfo] = [
-    # GGUF models with Unsloth Dynamic quantization (cross-platform, uses llama-server)
     ModelInfo(
-        id="loreguard-vanilla-ud-q6k",
-        name="Loreguard Vanilla UD Q6_K",
-        filename="loreguard-vanilla-UD-Q6_K.gguf",
-        size_gb=6.6,
-        size_bytes=7_085_559_072,
+        id="loreguard-npc-q6k",
+        name="Loreguard NPC Llama 3.1 8B Q6_K",
+        filename="loreguard-npc-llama3.1-8b-Q6_K.gguf",
+        size_gb=6.1,
+        size_bytes=6_596_010_976,
         context_length=8192,
-        url=f"https://huggingface.co/{HF_ORG}/loreguard-vanilla-gguf/resolve/main/loreguard-vanilla-UD-Q6_K.gguf",
+        url=f"https://huggingface.co/{HF_ORG}/loreguard-npc-llama3.1-8b-gguf/resolve/main/loreguard-npc-llama3.1-8b-Q6_K.gguf",
         description="Recommended. Best quality/size balance.",
         hardware="12GB RAM • 8GB VRAM",
         recommended=True,
     ),
     ModelInfo(
-        id="loreguard-vanilla-ud-q5km",
-        name="Loreguard Vanilla UD Q5_K_M",
-        filename="loreguard-vanilla-UD-Q5_K_M.gguf",
-        size_gb=5.7,
-        size_bytes=6_145_035_552,
+        id="loreguard-npc-q8",
+        name="Loreguard NPC Llama 3.1 8B Q8_0",
+        filename="loreguard-npc-llama3.1-8b-Q8_0.gguf",
+        size_gb=7.9,
+        size_bytes=8_540_775_392,
         context_length=8192,
-        url=f"https://huggingface.co/{HF_ORG}/loreguard-vanilla-gguf/resolve/main/loreguard-vanilla-UD-Q5_K_M.gguf",
-        description="Good quality, smaller size.",
-        hardware="10GB RAM • 6-8GB VRAM",
-        recommended=False,
-    ),
-    ModelInfo(
-        id="loreguard-vanilla-ud-q4km",
-        name="Loreguard Vanilla UD Q4_K_M",
-        filename="loreguard-vanilla-UD-Q4_K_M.gguf",
-        size_gb=4.9,
-        size_bytes=5_282_912_544,
-        context_length=8192,
-        url=f"https://huggingface.co/{HF_ORG}/loreguard-vanilla-gguf/resolve/main/loreguard-vanilla-UD-Q4_K_M.gguf",
-        description="Best for 6GB VRAM. Smallest size.",
-        hardware="8GB RAM • 6GB VRAM",
+        url=f"https://huggingface.co/{HF_ORG}/loreguard-npc-llama3.1-8b-gguf/resolve/main/loreguard-npc-llama3.1-8b-Q8_0.gguf",
+        description="Maximum quality. Requires more VRAM.",
+        hardware="16GB RAM • 12GB VRAM",
         recommended=False,
     ),
     ModelInfo(
-        id="loreguard-vanilla-ud-q8",
-        name="Loreguard Vanilla UD Q8_0",
-        filename="loreguard-vanilla-UD-Q8_0.gguf",
-        size_gb=8.5,
-        size_bytes=9_177_550_624,
+        id="loreguard-npc-f16",
+        name="Loreguard NPC Llama 3.1 8B F16",
+        filename="loreguard-npc-llama3.1-8b-f16.gguf",
+        size_gb=14.9,
+        size_bytes=16_068_895_712,
         context_length=8192,
-        url=f"https://huggingface.co/{HF_ORG}/loreguard-vanilla-gguf/resolve/main/loreguard-vanilla-UD-Q8_0.gguf",
-        description="Maximum quality. Requires more VRAM.",
-        hardware="16GB RAM • 12GB VRAM",
+        url=f"https://huggingface.co/{HF_ORG}/loreguard-npc-llama3.1-8b-gguf/resolve/main/loreguard-npc-llama3.1-8b-f16.gguf",
+        description="Full precision. Research/fine-tuning use.",
+        hardware="32GB RAM • 20GB VRAM",
         recommended=False,
     ),
 ]

{loreguard_cli-0.14.0 → loreguard_cli-0.14.1}/uv.lock

@@ -600,7 +600,7 @@ wheels = [
 
 [[package]]
 name = "loreguard-cli"
-version = "0.12.2"
+version = "0.14.0"
 source = { editable = "." }
 dependencies = [
     { name = "aiofiles" },
@@ -628,6 +628,11 @@ dev = [
     { name = "ruff" },
 ]
 
+[package.dev-dependencies]
+dev = [
+    { name = "pyinstaller" },
+]
+
 [package.metadata]
 requires-dist = [
     { name = "aiofiles", specifier = ">=24.1.0" },
@@ -650,6 +655,9 @@ requires-dist = [
 ]
 provides-extras = ["dev", "build"]
 
+[package.metadata.requires-dev]
+dev = [{ name = "pyinstaller", specifier = ">=6.17.0" }]
+
 [[package]]
 name = "macholib"
 version = "1.16.4"