added local model support with phi4-mini

Author: codingmoh

.gitignore

@@ -33,7 +33,8 @@ wheels/
 .LSOverride
 
 # Icon must end with two \r
-Icon
+Icon
+
 
 # Thumbnails
 ._*
@@ -232,3 +233,5 @@ pyrightconfig.json
 
 # End of https://www.toptal.com/developers/gitignore/api/python,macos,linux
 
+.idea/
+.vscode/

main.py

@@ -1,17 +1,19 @@
 [project]
-name = "ghost-in-the-shell"
+name = "open_codex"
 version = "0.1.7"
-description = "CLI with OpenAI integration"
+description = "CLI with OSS LLM integration"
 readme = "README.md"
 requires-python = ">=3.11"
 dependencies = [
-    "openai>=1.69.0",
+    "huggingface-hub>=0.30.2",
+    "llama-cpp-python>=0.3.8",
     "prompt_toolkit",
+    "pyperclip>=1.9.0",
 ]
 
 [project.scripts]
-ai = "main:main"
+ai = "open_codex.main:main"
 
 [build-system]
 requires = ["setuptools", "wheel"]
-build-backend = "setuptools.build_meta"
+build-backend = "setuptools.build_meta"

pyproject.toml

@@ -0,0 +1,16 @@
+
+from open_codex.agents.phi_4_mini import AgentPhi4Mini
+from open_codex.interfaces.llm_agent import LLMAgent
+
+class AgentBuilder:
+
+    @staticmethod
+    def get_agent() -> LLMAgent:
+        prompt = AgentBuilder.read_file("src/open_codex/resources/prompt.txt")
+        return AgentPhi4Mini(system_prompt=prompt)
+    
+    @staticmethod
+    def read_file(file_path: str) -> str:
+        with open(file_path, 'r') as file:
+            content = file.read()
+        return content

src/open_codex/agent_builder.py

@@ -0,0 +1,54 @@
+from typing import List, cast
+from llama_cpp import CreateCompletionResponse, Llama
+from open_codex.interfaces.llm_agent import LLMAgent
+import contextlib
+import os
+import sys
+
+
+class AgentPhi4Mini(LLMAgent):
+    def __init__(self, system_prompt: str):
+        # suppress the stderr output from llama_cpp
+        # this is a workaround for the llama_cpp library
+        # which prints a lot of warnings and errors to stderr
+        # when loading the model
+        # this is a temporary solution until the library is fixed
+        with AgentPhi4Mini.suppress_stderr():
+            self.llm: Llama = Llama.from_pretrained( # type: ignore
+                repo_id="lmstudio-community/Phi-4-mini-instruct-GGUF",
+                filename="Phi-4-mini-instruct-Q3_K_L.gguf",
+                additional_files=[],  
+            )
+        self.system_prompt = system_prompt
+
+    def one_shot_mode(self, user_input: str) -> str:
+        chat_history = [{"role": "system", "content": self.system_prompt}]
+        chat_history.append({"role": "user", "content": user_input})
+        full_prompt = self.format_chat(chat_history)
+        output_raw = self.llm(prompt=full_prompt, max_tokens=100, temperature=0.2, stream=False)
+        
+        # unfortuntely llama_cpp has a union type for the output
+        output = cast(CreateCompletionResponse, output_raw)
+        
+        assistant_reply : str = output["choices"][0]["text"].strip() 
+        return assistant_reply 
+        
+    
+    def format_chat(self, messages: List[dict[str, str]]) -> str:
+        chat_prompt = ""
+        for msg in messages:
+            role_tag = "user" if msg["role"] == "user" else "assistant"
+            chat_prompt += f"<|{role_tag}|>\n{msg['content']}\n"
+        chat_prompt += "<|assistant|>\n"
+        return chat_prompt
+    
+    @contextlib.contextmanager
+    @staticmethod
+    def suppress_stderr():
+        with open(os.devnull, 'w') as devnull:
+            old_stderr = sys.stderr
+            sys.stderr = devnull
+            try:
+                yield
+            finally:
+                sys.stderr = old_stderr

src/open_codex/agents/__init__.py

@@ -0,0 +1,7 @@
+from abc import ABC, abstractmethod
+
+class LLMAgent(ABC):
+    @abstractmethod
+    def one_shot_mode(self, user_input: str) -> str:
+        pass
+    

src/open_codex/agents/phi_4_mini.py

@@ -0,0 +1,84 @@
+import sys
+import argparse
+import subprocess
+import pyperclip
+
+from open_codex.agent_builder import AgentBuilder
+from open_codex.interfaces.llm_agent import LLMAgent
+
+
+GREEN = "\033[92m"
+RED = "\033[91m"
+BLUE = "\033[94m"
+RESET = "\033[0m"
+
+# Capture single keypress (terminal) from the user
+# and returns it as a string. It works on both Windows and Unix systems.
+# Windows
+if sys.platform == "win32":
+    import msvcrt
+
+    def get_keypress():
+        return msvcrt.getch().decode("utf-8")
+
+# Unix (Linux/macOS)
+else:
+    import termios
+    import tty
+
+    def get_keypress():
+        fd = sys.stdin.fileno()
+        old_settings = termios.tcgetattr(fd)
+        try:
+            tty.setraw(fd)
+            key = sys.stdin.read(1)
+        finally:
+            termios.tcsetattr(fd, termios.TCSADRAIN, old_settings)
+        return key
+
+def one_shot_mode(agent: LLMAgent, prompt: str):
+    try:
+        response = agent.one_shot_mode(prompt)
+        print(f"{GREEN}: {response}{RESET}") 
+        command = response.strip()
+
+        print(f"{BLUE}What do you want to do with this command?{RESET}")
+        print(f"{BLUE}[c] Copy  [e] Execute  [a] Abort{RESET}")
+        print(f"{BLUE}Press key: ", end="", flush=True)
+        choice = get_keypress().lower()
+        print(f"{RESET}")  # Clear the line after the prompt
+
+        if choice == "e":
+            print(f"{BLUE}Executing command: {command}{RESET}")
+            result = subprocess.run(command, shell=True, capture_output=True, text=True)
+            print(f"{GREEN}Command output: {result.stdout}{RESET}")
+            if result.stderr:
+                print(f"{RED}Error: {result.stderr}{RESET}")
+
+        elif choice == "c":
+            pyperclip.copy(command)
+            print(f"{GREEN}Command copied to clipboard! Paste it manually in your terminal.{RESET}")
+
+        elif choice == "a":
+            print(f"{BLUE}Aborted.{RESET}")
+        else:
+            print(f"{RED}Unknown choice. Nothing happened.{RESET}")
+
+    except Exception as e:
+        print(f"{RED}Error: {e}{RESET}")
+
+def main():
+    parser = argparse.ArgumentParser(description="Open Codex - Natural Language to CLI commands")
+    parser.add_argument("prompt", nargs="*", help="Optional prompt for one-shot mode")
+    args = parser.parse_args()
+
+    prompt = " ".join(args.prompt).strip()
+    if not prompt:
+        print("Please provide a prompt")
+        sys.exit(1)
+    
+    agent = AgentBuilder.get_agent()
+    one_shot_mode(agent, prompt)
+
+if __name__ == "__main__":
+    main()

src/open_codex/interfaces/__init__.py

@@ -0,0 +1,5 @@
+You are a helpful AI that converts natural language instructions  into valid shell commands. Always provide a concise shell command. 
+No shebang, no explanations, no extra text. Just the command. Think hard about the command you are going to provide, and make sure it is the best one.
+Stick to the easiest command possible, and if there is a simpler one, provide it.
+If the command is not straightforward, and there is an easier one, that can be run using a 3rd party tool, 
+INFORM the user about it, and provide the command to install it.