cct-cli 0.7.9.0__py3-none-any.whl

calc_terminal/native_picker.py

@@ -0,0 +1,215 @@
+"""
+CCT - native Windows file picker for Browse in Attach to Message.
+
+Reliability order (v0.8.1 root-cause fix for "Browse does nothing"):
+
+  1. tkinter askopenfilenames with root.attributes("-topmost", True).
+     The -topmost attribute GUARANTEES the dialog appears above the
+     fullscreen terminal even when the terminal owns the foreground
+     (Windows Terminal / conhost). This is the primary method because
+     the old ctypes-first path could open GetOpenFileNameW BEHIND the
+     terminal window (hwndOwner=0 in Windows Terminal returns no usable
+     owner), where it silently waited for input the user could not see —
+     exactly the reported "button bounces, nothing happens" symptom.
+
+  2. ctypes GetOpenFileNameW (comdlg32) as a fallback when tkinter is
+     unavailable. hwndOwner=0 works reliably from worker threads;
+     lpstrInitialDir sets the start directory (never lpstrFile, which
+     causes CDERR_DIALOGFAILURE).
+
+Both methods: files only (no folder selection), multi-select enabled,
+"All Files (*.*)" plus common category filters — deliberately broad.
+"""
+if __name__ == "__main__":
+    print("This is a library file.")
+    import sys; sys.exit(1)
+
+import os
+import sys
+
+try:
+    import ctypes
+    from ctypes import wintypes
+    _CTYPES_OK = True
+except Exception:
+    _CTYPES_OK = False
+
+_OFN_FILEMUSTEXIST = 0x00001000
+_OFN_PATHMUSTEXIST = 0x00000800
+_OFN_HIDEREADONLY = 0x00000004
+_OFN_EXPLORER = 0x00080000
+_OFN_ALLOWMULTISELECT = 0x00000200
+_FLAGS = (_OFN_FILEMUSTEXIST | _OFN_PATHMUSTEXIST
+          | _OFN_HIDEREADONLY | _OFN_EXPLORER | _OFN_ALLOWMULTISELECT)
+_BUFSIZE = 32768
+
+FILTER = (
+    "All Files (*.*)\0*.*\0"
+    "Code Files (*.py;*.js;*.ts;*.tsx;*.jsx;*.java;*.c;*.h;*.cpp;*.rs;*.go)\0"
+    "*.py;*.js;*.ts;*.tsx;*.jsx;*.java;*.c;*.h;*.cpp;*.rs;*.go\0"
+    "Data Files (*.json;*.yaml;*.yml;*.toml;*.xml;*.csv)\0"
+    "*.json;*.yaml;*.yml;*.toml;*.xml;*.csv\0"
+    "Documents (*.txt;*.md;*.log;*.html;*.css;*.scss;*.pdf;*.doc;*.docx)\0"
+    "*.txt;*.md;*.log;*.html;*.css;*.scss;*.pdf;*.doc;*.docx\0"
+    "Images (*.png;*.jpg;*.jpeg;*.webp;*.gif;*.svg;*.bmp;*.ico)\0"
+    "*.png;*.jpg;*.jpeg;*.webp;*.gif;*.svg;*.bmp;*.ico\0"
+    "Archives (*.zip;*.tar;*.gz;*.bz2;*.rar;*.7z)\0"
+    "*.zip;*.tar;*.gz;*.bz2;*.rar;*.7z\0"
+)
+
+TK_FILETYPES = [
+    ("All Files", "*.*"),
+    ("Code Files", "*.py *.js *.ts *.tsx *.jsx *.java *.c *.h *.cpp *.rs *.go"),
+    ("Data Files", "*.json *.yaml *.yml *.toml *.xml *.csv"),
+    ("Documents", "*.txt *.md *.log *.html *.css *.scss *.pdf *.doc *.docx"),
+    ("Images", "*.png *.jpg *.jpeg *.webp *.gif *.svg *.bmp *.ico"),
+    ("Archives", "*.zip *.tar *.gz *.bz2 *.rar *.7z"),
+]
+
+
+def is_native_supported():
+    return sys.platform == "win32"
+
+
+def parse_multi_buffer(raw):
+    if isinstance(raw, bytes):
+        raw = raw.decode("utf-16-le", errors="ignore")
+    parts = [p for p in raw.split("\x00") if p]
+    if not parts:
+        return []
+    if len(parts) == 1:
+        return parts if os.path.isabs(parts[0]) else []
+    directory = parts[0]
+    out = []
+    for name in parts[1:]:
+        candidate = os.path.join(directory, name)
+        if os.path.isabs(candidate):
+            out.append(candidate)
+    return out
+
+
+def _pick_files_tkinter(initial_dir, title):
+    """PRIMARY: tkinter dialog forced top-most, so it is always visible.
+    Returns list of paths ([] = user cancelled), None = tkinter unusable."""
+    try:
+        import tkinter as tk
+        from tkinter import filedialog
+        root = tk.Tk()
+        root.withdraw()
+        root.attributes("-topmost", True)
+        try:
+            filepaths = filedialog.askopenfilenames(
+                title=title,
+                initialdir=initial_dir,
+                filetypes=TK_FILETYPES,
+            )
+            if filepaths:
+                return list(filepaths)
+            return []  # user cancelled
+        finally:
+            try:
+                root.destroy()
+            except Exception:
+                pass
+    except Exception as e:
+        print("[NativePicker] tkinter picker failed: %s" % e, file=sys.stderr)
+        return None
+
+
+class _OPENFILENAMEW(ctypes.Structure):
+    _fields_ = [
+        ("lStructSize", wintypes.DWORD),
+        ("hwndOwner", wintypes.HWND),
+        ("hInstance", wintypes.HINSTANCE),
+        ("lpstrFilter", wintypes.LPCWSTR),
+        ("lpstrCustomFilter", wintypes.LPWSTR),
+        ("nMaxCustFilter", wintypes.DWORD),
+        ("nFilterIndex", wintypes.DWORD),
+        ("lpstrFile", wintypes.LPWSTR),
+        ("nMaxFile", wintypes.DWORD),
+        ("lpstrFileTitle", wintypes.LPWSTR),
+        ("nMaxFileTitle", wintypes.DWORD),
+        ("lpstrInitialDir", wintypes.LPCWSTR),
+        ("lpstrTitle", wintypes.LPCWSTR),
+        ("Flags", wintypes.DWORD),
+        ("nFileOffset", wintypes.WORD),
+        ("nFileExtension", wintypes.WORD),
+        ("lpstrDefExt", wintypes.LPCWSTR),
+        ("lCustData", ctypes.c_void_p),
+        ("lpfnHook", ctypes.c_void_p),
+        ("lpTemplateName", wintypes.LPCWSTR),
+    ]
+    if ctypes.sizeof(ctypes.c_void_p) == 8:
+        _fields_ += [("pvReserved", ctypes.c_void_p),
+                     ("dwReserved", wintypes.DWORD)]
+    _fields_ += [("FlagsEx", wintypes.DWORD)]
+
+
+def _pick_files_ctypes(initial_dir, title):
+    """FALLBACK: ctypes GetOpenFileNameW. Returns list of paths or None."""
+    # CRITICAL: lpstrFile buffer must start EMPTY.
+    # Setting buf.value to a directory path causes CDERR_DIALOGFAILURE (12290).
+    # Initial directory is set via lpstrInitialDir instead.
+    buf = ctypes.create_unicode_buffer(_BUFSIZE)
+
+    filter_str = FILTER + "\x00"
+    ofn = _OPENFILENAMEW()
+    ofn.lStructSize = ctypes.sizeof(_OPENFILENAMEW)
+    # hwndOwner=0: works reliably from worker threads.
+    # GetConsoleWindow() returns 0 in Windows Terminal/Textual.
+    # GetForegroundWindow() causes error 12290 (CDERR_DIALOGFAILURE).
+    ofn.hwndOwner = 0
+    ofn.lpstrFilter = filter_str
+    ofn.nFilterIndex = 1
+    ofn.lpstrFile = ctypes.cast(buf, wintypes.LPWSTR)
+    ofn.nMaxFile = _BUFSIZE
+    ofn.lpstrTitle = title
+    ofn.Flags = _FLAGS
+
+    # Set initial directory via lpstrInitialDir, NOT via buf.value
+    if initial_dir:
+        if not os.path.isdir(initial_dir):
+            initial_dir = os.path.dirname(initial_dir) or None
+        if initial_dir:
+            ofn.lpstrInitialDir = initial_dir
+
+    comdlg32 = ctypes.windll.comdlg32
+    fn = comdlg32.GetOpenFileNameW
+    fn.argtypes = [ctypes.POINTER(_OPENFILENAMEW)]
+    fn.restype = wintypes.BOOL
+
+    result = fn(ctypes.byref(ofn))
+    if result:
+        return parse_multi_buffer(buf.raw)
+    err = comdlg32.CommDlgExtendedError()
+    if err:
+        print("[NativePicker] CommDlgExtendedError=%d" % err, file=sys.stderr)
+    return None
+
+
+def pick_files(initial_dir=None, title="Select file(s) to attach"):
+    """Show a file picker. Tries the always-visible tkinter dialog first,
+    falls back to ctypes GetOpenFileNameW.
+    Returns list of paths (empty = cancelled), None = all methods failed."""
+    if not is_native_supported():
+        return None
+
+    # Method 1: tkinter (topmost — cannot open behind the terminal)
+    result = _pick_files_tkinter(initial_dir, title)
+    if result is not None:
+        return result
+
+    # Method 2: ctypes GetOpenFileNameW
+    if not _CTYPES_OK:
+        return None
+    print("[NativePicker] Trying ctypes fallback...", file=sys.stderr)
+    try:
+        result = _pick_files_ctypes(initial_dir, title)
+        if result is not None:
+            return result
+    except Exception as e:
+        print("[NativePicker] ctypes exception: %s" % e, file=sys.stderr)
+        import traceback
+        traceback.print_exc(file=sys.stderr)
+
+    return None

calc_terminal/ollama_catalog.py

@@ -0,0 +1,279 @@
+"""Ollama model catalog — Authentic, curated models with dynamic local discovery.
+Creator: Kazi Zillani (CAT Platform).
+
+Provides authentic model metadata (params, size, context, capabilities, family)
+without artificial dummy variants. Supports dynamic registration of locally installed
+models and custom user models.
+"""
+
+from typing import List, Dict, Any, Optional
+import os
+
+# Base curated catalog of authentic, modern local models
+CURATED_CATALOG: List[Dict[str, Any]] = []
+
+def _add(name: str, params: str, size_gb: float, family: str, desc: str,
+         ctx: int = 8192, caps: Optional[List[str]] = None, categories: Optional[List[str]] = None):
+    CURATED_CATALOG.append({
+        "name": name,
+        "params": params,
+        "size_gb": float(size_gb),
+        "family": family,
+        "desc": desc,
+        "context": ctx,
+        "caps": caps or ["chat"],
+        "categories": categories or ["popular"],
+    })
+
+# =====================================================================
+# 1. CORE LLAMA (Meta)
+# =====================================================================
+_add("llama3.3:70b", "70B", 42.0, "llama", "Meta Llama 3.3 70B — State of the art open weights model", 131072, ["chat", "code", "reasoning"], ["popular", "large"])
+_add("llama3.3:8b", "8B", 4.7, "llama", "Meta Llama 3.3 8B — Fast, capable general assistant", 131072, ["chat", "code"], ["popular", "coding"])
+_add("llama3.2:3b", "3B", 2.0, "llama", "Meta Llama 3.2 3B — Lightweight on-device assistant", 131072, ["chat", "code"], ["popular", "small"])
+_add("llama3.2:1b", "1B", 1.3, "llama", "Meta Llama 3.2 1B — Ultra-lightweight edge model", 131072, ["chat"], ["small"])
+_add("llama3.1:8b", "8B", 4.7, "llama", "Meta Llama 3.1 8B — High context general purpose model", 131072, ["chat", "code"], ["popular", "coding"])
+_add("llama3.1:70b", "70B", 42.0, "llama", "Meta Llama 3.1 70B — Heavyweight reasoning and coding", 131072, ["chat", "code", "reasoning"], ["large"])
+_add("llama3.1:405b", "405B", 243.0, "llama", "Meta Llama 3.1 405B — Frontier flagship open weights", 131072, ["chat", "code", "reasoning"], ["large"])
+_add("llama3:8b", "8B", 4.7, "llama", "Meta Llama 3 8B — Original Llama 3 release", 8192, ["chat", "code"], ["popular"])
+_add("llama2:7b", "7B", 3.8, "llama", "Meta Llama 2 7B — Classic foundation model", 4096, ["chat"], ["small"])
+_add("llama2:13b", "13B", 7.4, "llama", "Meta Llama 2 13B — Classic balanced model", 4096, ["chat"], ["popular"])
+
+# =====================================================================
+# 2. QWEN & QWEN CODER (Alibaba)
+# =====================================================================
+_add("qwen2.5-coder:7b", "7B", 4.7, "qwen", "Qwen 2.5 Coder 7B — Premier open-source code generation model", 32768, ["code", "chat"], ["popular", "coding"])
+_add("qwen2.5-coder:14b", "14B", 9.0, "qwen", "Qwen 2.5 Coder 14B — Deep architectural coding & refactoring", 32768, ["code", "chat"], ["coding", "popular"])
+_add("qwen2.5-coder:32b", "32B", 19.0, "qwen", "Qwen 2.5 Coder 32B — Enterprise-grade code intelligence", 32768, ["code", "chat", "reasoning"], ["coding", "large"])
+_add("qwen2.5-coder:3b", "3B", 1.9, "qwen", "Qwen 2.5 Coder 3B — Fast on-device copilot & terminal assistant", 32768, ["code", "chat"], ["coding", "small"])
+_add("qwen2.5-coder:1.5b", "1.5B", 1.0, "qwen", "Qwen 2.5 Coder 1.5B — Ultra-low latency autocomplete", 32768, ["code"], ["coding", "small"])
+_add("qwen2.5:7b", "7B", 4.4, "qwen", "Qwen 2.5 7B — Multilingual general intelligence & math", 32768, ["chat", "code"], ["popular"])
+_add("qwen2.5:14b", "14B", 9.0, "qwen", "Qwen 2.5 14B — Strong balanced reasoning and knowledge", 32768, ["chat", "code", "reasoning"], ["popular"])
+_add("qwen2.5:32b", "32B", 19.0, "qwen", "Qwen 2.5 32B — High-fidelity analytical reasoning", 32768, ["chat", "code", "reasoning"], ["large"])
+_add("qwen2.5:72b", "72B", 41.0, "qwen", "Qwen 2.5 72B — Flagship open weights generalist", 32768, ["chat", "code", "reasoning"], ["large"])
+_add("qwen2.5:3b", "3B", 1.9, "qwen", "Qwen 2.5 3B — Agile compact multilingual model", 32768, ["chat"], ["small"])
+_add("qwen2.5:1.5b", "1.5B", 0.9, "qwen", "Qwen 2.5 1.5B — Lightweight embedded model", 32768, ["chat"], ["small"])
+_add("qwen2.5:0.5b", "0.5B", 0.4, "qwen", "Qwen 2.5 0.5B — Micro model for IoT and quick parsing", 32768, ["chat"], ["small"])
+_add("qwq:32b", "32B", 20.0, "qwen", "QwQ 32B — Thinking & complex mathematical reasoning model", 32768, ["reasoning", "math", "code"], ["reasoning", "popular"])
+
+# =====================================================================
+# 3. DEEPSEEK (DeepSeek AI)
+# =====================================================================
+_add("deepseek-r1:1.5b", "1.5B", 1.1, "deepseek", "DeepSeek R1 Distill 1.5B — Micro step-by-step reasoning", 32768, ["reasoning"], ["reasoning", "small"])
+_add("deepseek-r1:7b", "7B", 4.7, "deepseek", "DeepSeek R1 Distill 7B (Qwen) — Fast chain-of-thought", 32768, ["reasoning", "code"], ["reasoning", "popular"])
+_add("deepseek-r1:8b", "8B", 4.9, "deepseek", "DeepSeek R1 Distill 8B (Llama) — Strong CoT reasoning", 32768, ["reasoning", "code"], ["reasoning", "popular"])
+_add("deepseek-r1:14b", "14B", 9.0, "deepseek", "DeepSeek R1 Distill 14B — In-depth logical deduction & math", 32768, ["reasoning", "code", "math"], ["reasoning"])
+_add("deepseek-r1:32b", "32B", 19.0, "deepseek", "DeepSeek R1 Distill 32B — High-end reasoning & complex problems", 32768, ["reasoning", "code", "math"], ["reasoning", "large"])
+_add("deepseek-r1:70b", "70B", 42.0, "deepseek", "DeepSeek R1 Distill 70B — Top-tier open reasoning model", 32768, ["reasoning", "code", "math"], ["reasoning", "large"])
+_add("deepseek-coder:6.7b", "6.7B", 3.8, "deepseek", "DeepSeek Coder 6.7B — Efficient coding specialist", 16384, ["code"], ["coding"])
+_add("deepseek-coder-v2:16b", "16B", 8.9, "deepseek", "DeepSeek Coder V2 16B (MoE) — Multi-language coding engine", 32768, ["code", "chat"], ["coding"])
+_add("deepseek-v3:671b", "671B", 400.0, "deepseek", "DeepSeek V3 671B (MoE) — Frontier flagship architecture", 65536, ["chat", "code", "reasoning"], ["large"])
+
+# =====================================================================
+# 4. GOOGLE GEMMA
+# =====================================================================
+_add("gemma3:1b", "1B", 0.8, "gemma", "Google Gemma 3 1B — Next-gen multimodal compact model", 32768, ["chat", "vision"], ["small", "vision"])
+_add("gemma3:4b", "4B", 3.3, "gemma", "Google Gemma 3 4B — State-of-the-art multimodal lightweight model", 32768, ["chat", "vision", "code"], ["popular", "vision", "small"])
+_add("gemma3:12b", "12B", 8.1, "gemma", "Google Gemma 3 12B — Advanced multimodal reasoning", 32768, ["chat", "vision", "code"], ["popular", "vision"])
+_add("gemma3:27b", "27B", 17.0, "gemma", "Google Gemma 3 27B — High-capacity multimodal reasoning", 32768, ["chat", "vision", "reasoning"], ["large", "vision"])
+_add("gemma2:2b", "2B", 1.6, "gemma", "Google Gemma 2 2B — High-efficiency on-device text generator", 8192, ["chat"], ["small"])
+_add("gemma2:9b", "9B", 5.4, "gemma", "Google Gemma 2 9B — Outstanding performance-per-parameter", 8192, ["chat", "code"], ["popular"])
+_add("gemma2:27b", "27B", 16.0, "gemma", "Google Gemma 2 27B — Powerful analytical text model", 8192, ["chat", "reasoning"], ["large"])
+_add("codegemma:7b", "7B", 5.0, "gemma", "Google CodeGemma 7B — Code completion and synthesis", 8192, ["code"], ["coding"])
+
+# =====================================================================
+# 5. MISTRAL AI
+# =====================================================================
+_add("mistral:7b", "7B", 4.1, "mistral", "Mistral 7B — The classic high-performance 7B standard", 32768, ["chat", "code"], ["popular"])
+_add("mistral-nemo:12b", "12B", 7.1, "mistral", "Mistral NeMo 12B — 128k context multilingual model with NVIDIA", 128000, ["chat", "code"], ["popular"])
+_add("codestral:22b", "22B", 12.0, "mistral", "Mistral Codestral 22B — Fast 80+ language coding model", 32768, ["code"], ["coding", "popular"])
+_add("mixtral:8x7b", "47B", 26.0, "mistral", "Mixtral 8x7B (MoE) — Fast sparse Mixture-of-Experts", 32768, ["chat", "code"], ["popular", "large"])
+_add("mixtral:8x22b", "141B", 80.0, "mistral", "Mixtral 8x22B (MoE) — High-end reasoning Mixture-of-Experts", 65536, ["chat", "code", "reasoning"], ["large"])
+_add("mathstral:7b", "7B", 4.1, "mistral", "Mistral Mathstral 7B — Mathematical reasoning and science", 32768, ["math", "reasoning"], ["reasoning"])
+
+# =====================================================================
+# 6. MICROSOFT PHI
+# =====================================================================
+_add("phi4:14b", "14B", 9.1, "phi", "Microsoft Phi-4 14B — Premier synthetic reasoning & STEM benchmark leader", 16384, ["reasoning", "math", "code"], ["popular", "reasoning"])
+_add("phi3.5:3.8b", "3.8B", 2.2, "phi", "Microsoft Phi-3.5 Mini 3.8B — 128k context high-efficiency model", 128000, ["chat", "code"], ["small", "popular"])
+_add("phi3:14b", "14B", 7.9, "phi", "Microsoft Phi-3 Medium 14B — Analytical reasoning", 16384, ["chat", "reasoning"], ["popular"])
+_add("phi3:3.8b", "3.8B", 2.2, "phi", "Microsoft Phi-3 Mini 3.8B — Compact logic model", 8192, ["chat"], ["small"])
+
+# =====================================================================
+# 7. VISION & MULTIMODAL
+# =====================================================================
+_add("llava:7b", "7B", 4.7, "vision", "LLaVA 1.5 7B — Visual question answering and image inspection", 4096, ["vision", "chat"], ["vision", "popular"])
+_add("llava:13b", "13B", 8.0, "vision", "LLaVA 1.5 13B — High-detail image understanding", 4096, ["vision", "chat"], ["vision"])
+_add("llava-llama3:8b", "8B", 5.0, "vision", "LLaVA Llama-3 8B — Multimodal assistant on Llama 3", 8192, ["vision", "chat"], ["vision", "popular"])
+_add("minicpm-v:8b", "8B", 5.5, "vision", "MiniCPM-V 2.6 8B — Outstanding OCR & high-res image analysis", 8192, ["vision", "chat"], ["vision"])
+_add("moondream:1.8b", "1.8B", 1.0, "vision", "Moondream 2 1.8B — Fast edge-ready visual model", 4096, ["vision"], ["vision", "small"])
+_add("bakllava:7b", "7B", 4.7, "vision", "BakLLaVA 7B — Mistral-backed visual assistant", 4096, ["vision"], ["vision"])
+
+# =====================================================================
+# 8. SMALL, FAST & ON-DEVICE
+# =====================================================================
+_add("smollm2:1.7b", "1.7B", 1.0, "smollm", "SmolLM2 1.7B — HuggingFace curated small model for edge devices", 8192, ["chat", "code"], ["small"])
+_add("smollm2:360m", "360M", 0.3, "smollm", "SmolLM2 360M — Ultra-compact local language model", 8192, ["chat"], ["small"])
+_add("smollm2:135m", "135M", 0.15, "smollm", "SmolLM2 135M — Micro test model with near-zero memory footprint", 2048, ["chat"], ["small"])
+_add("tinyllama:1.1b", "1.1B", 0.6, "other", "TinyLlama 1.1B — 3T token trained compact architecture", 2048, ["chat"], ["small"])
+_add("orca-mini:3b", "3B", 1.9, "other", "Orca Mini 3B — Microsoft Orca dataset instruction tuned", 4096, ["chat"], ["small"])
+_add("granite3.1:2b", "2B", 1.3, "granite", "IBM Granite 3.1 2B — Enterprise small language model", 8192, ["chat", "code"], ["small"])
+_add("granite3.1:8b", "8B", 4.9, "granite", "IBM Granite 3.1 8B — Enterprise general and tool-use model", 8192, ["chat", "code"], ["popular"])
+
+# =====================================================================
+# 9. SPECIALIZED CODING & ENTERPRISE
+# =====================================================================
+_add("codellama:7b", "7B", 3.8, "codellama", "CodeLlama 7B — Meta coding specialist", 16384, ["code"], ["coding"])
+_add("codellama:13b", "13B", 7.4, "codellama", "CodeLlama 13B — Medium-footprint coding model", 16384, ["code"], ["coding"])
+_add("codellama:34b", "34B", 19.0, "codellama", "CodeLlama 34B — High accuracy code completion", 16384, ["code"], ["coding", "large"])
+_add("starcoder2:7b", "7B", 4.0, "starcoder", "StarCoder2 7B — BigCode consortium model for 600+ languages", 16384, ["code"], ["coding"])
+_add("starcoder2:15b", "15B", 9.0, "starcoder", "StarCoder2 15B — High precision code assistant", 16384, ["code"], ["coding"])
+_add("sqlcoder:15b", "15B", 8.5, "other", "Defog SQLCoder 15B — Natural language to SQL query engine", 4096, ["code", "sql"], ["coding"])
+_add("command-r:35b", "35B", 19.0, "cohere", "Cohere Command R 35B — Optimized for RAG and tool execution", 128000, ["chat", "rag"], ["popular", "large"])
+_add("command-r-plus:104b", "104B", 60.0, "cohere", "Cohere Command R+ 104B — Frontier enterprise RAG model", 128000, ["chat", "rag", "reasoning"], ["large"])
+
+# =====================================================================
+# 10. EMBEDDINGS & RETRIEVAL (RAG)
+# =====================================================================
+_add("nomic-embed-text:latest", "137M", 0.3, "nomic", "Nomic Embed Text — 8192 context high-quality text embedding", 8192, ["embedding"], ["embedding", "small"])
+_add("bge-m3:latest", "567M", 1.2, "bge", "BAAI BGE-M3 — Multilingual, multi-granularity dense/sparse embedding", 8192, ["embedding"], ["embedding", "small"])
+_add("all-minilm:latest", "33M", 0.1, "minilm", "All-MiniLM L6 v2 — Ultra-fast embedding model for semantic search", 512, ["embedding"], ["embedding", "small"])
+_add("mxbai-embed-large:latest", "335M", 0.7, "mxbai", "MixedBread Embed Large — Top MTEB benchmark embedding model", 512, ["embedding"], ["embedding", "small"])
+
+
+# Dynamic local model registry (persisted or populated from Ollama API)
+DYNAMIC_REGISTRY: Dict[str, Dict[str, Any]] = {}
+
+
+def register_custom_model(model_dict: Dict[str, Any]):
+    """Register a custom or dynamically discovered model."""
+    name = model_dict.get("name")
+    if not name:
+        return
+    DYNAMIC_REGISTRY[name] = {
+        "name": name,
+        "params": model_dict.get("params", "Unknown"),
+        "size_gb": float(model_dict.get("size_gb", 4.0)),
+        "family": model_dict.get("family", "custom"),
+        "desc": model_dict.get("desc", f"Local model {name}"),
+        "context": int(model_dict.get("context", 8192)),
+        "caps": model_dict.get("caps", ["chat"]),
+        "categories": model_dict.get("categories", ["custom"]),
+        "installed": bool(model_dict.get("installed", False)),
+    }
+
+
+def register_installed_models(installed_names: List[str]):
+    """Ensure any locally installed model in Ollama is present in the catalog."""
+    for raw_name in installed_names:
+        name = raw_name.strip()
+        if not name:
+            continue
+        # If already in curated catalog, mark as installed in memory
+        existing = get_model(name)
+        if not existing:
+            # Estimate size / params from tag if possible
+            params = "Unknown"
+            size_gb = 4.0
+            if ":" in name:
+                tag = name.split(":")[-1].lower()
+                if tag.endswith("b") and tag[:-1].replace(".", "").isdigit():
+                    params = tag.upper()
+                    try:
+                        num = float(tag[:-1])
+                        size_gb = round(num * 0.65, 1)
+                    except Exception:
+                        size_gb = 4.0
+
+            register_custom_model({
+                "name": name,
+                "params": params,
+                "size_gb": size_gb,
+                "family": name.split(":")[0],
+                "desc": f"Installed Ollama model: {name}",
+                "context": 8192,
+                "caps": ["chat"],
+                "categories": ["installed"],
+                "installed": True,
+            })
+
+
+def all_models(include_dynamic: bool = True) -> List[Dict[str, Any]]:
+    """Return all catalog models, combining curated entries and dynamic/installed models."""
+    combined: Dict[str, Dict[str, Any]] = {}
+    for m in CURATED_CATALOG:
+        combined[m["name"]] = dict(m)
+
+    if include_dynamic:
+        for name, m in DYNAMIC_REGISTRY.items():
+            if name in combined:
+                combined[name].update(m)
+            else:
+                combined[name] = dict(m)
+
+    return list(combined.values())
+
+
+# Compatibility alias for legacy callers
+CATALOG = all_models()
+
+
+def categories() -> List[str]:
+    """Return list of standard navigation categories."""
+    return ["all", "popular", "coding", "reasoning", "vision", "small", "large", "embedding"]
+
+
+def search_models(
+    query: str = "",
+    category: Optional[str] = None,
+    family: Optional[str] = None,
+    cap: Optional[str] = None,
+) -> List[Dict[str, Any]]:
+    """Search models with optional category, family, and capability filters."""
+    q = (query or "").lower().strip()
+    cat = (category or "").lower().strip()
+    out = []
+
+    for m in all_models(include_dynamic=True):
+        if family and m.get("family") != family:
+            continue
+        if cap and cap not in m.get("caps", []):
+            continue
+        if cat and cat != "all":
+            cats = [c.lower() for c in m.get("categories", [])]
+            # Match either direct category list or family/caps
+            if cat not in cats and cat not in m.get("caps", []) and cat != m.get("family"):
+                continue
+        if q:
+            name_match = q in m.get("name", "").lower()
+            fam_match = q in m.get("family", "").lower()
+            desc_match = q in m.get("desc", "").lower()
+            params_match = q in m.get("params", "").lower()
+            if not (name_match or fam_match or desc_match or params_match):
+                continue
+        out.append(m)
+
+    return out
+
+
+def get_model(name: str) -> Optional[Dict[str, Any]]:
+    """Look up a model by exact name or base prefix."""
+    if not name:
+        return None
+    name_clean = name.strip()
+    
+    # Check dynamic first
+    if name_clean in DYNAMIC_REGISTRY:
+        return dict(DYNAMIC_REGISTRY[name_clean])
+
+    # Check curated exact match
+    for m in CURATED_CATALOG:
+        if m["name"] == name_clean:
+            return dict(m)
+
+    # Check if tag omitted (e.g. "qwen2.5-coder" matches "qwen2.5-coder:7b" or latest)
+    for m in CURATED_CATALOG:
+        if m["name"].startswith(name_clean + ":") or m["name"] == f"{name_clean}:latest":
+            return dict(m)
+
+    return None