shell-llm 0.1.4

Interactive shell with LLM-powered features

LLM Shell Assistant

An intelligent shell that combines traditional shell capabilities with natural language processing. Features include command generation from natural language, automatic error explanations, and command completion.

Prerequisites

• Python 3.8 or higher
• A Google API key for Gemini AI (specifically configured for gemini-2.0-flash model)
• GCC compiler (for the C core)

Setup

Method 1: Install from PyPI (Recommended)

Install directly using pip:

pip install shell-llm

Set up your Google API key:

echo "GEMINI_API_KEY=your_api_key_here" > .env

Method 2: Install from Source

1. Clone the repository:

git clone https://github.com/yourusername/shell-llm.git
cd shell-llm

2. Create and activate a virtual environment:

python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

3. Install the package in development mode:

pip install -e .

4. Set up your Google API key:

echo "GEMINI_API_KEY=your_api_key_here" > .env

Running the Shell

After installation, simply run:

shell-llm

If installed from source with a virtual environment, make sure your virtual environment is activated:

source venv/bin/activate  # On Windows: venv\Scripts\activate
shell-llm

Usage

• Regular shell commands work as normal
• Start with # for natural language queries:

  #how do I find large files
  

• Add verbosity flags for more information:
  • -v: Show brief command explanation
  • -vv: Show detailed command explanation with options and examples

  #how do I copy files with scp -vv
  

Error Handling

The shell provides intelligent error handling with AI-powered explanations:

1. Direct Error Capture

   • Errors are captured directly from command execution in the C core
   • No re-running of commands to capture errors
   • Handles both execution errors and usage messages

2. Error Processing

   • Errors are immediately displayed with a red "Error:" prefix
   • The LLM analyzes the error using the Gemini model
   • Responses are cached to improve performance

3. Structured Solutions The error handler provides a consistent format:

   Error: <original error message>
   
   Problem: <one-line explanation of what went wrong>
   
   Solution:
     • <step-by-step instructions>
     • <clear actionable items>
     • <relevant suggestions>
   

Example outputs:

$ mkdir root/pass
Error: mkdir: cannot create directory 'root/pass': No such file or directory

Problem: The parent directory 'root' does not exist

Solution:
  • First, create the parent directory 'root' using the command `mkdir root`
  • Verify that the 'root' directory was created successfully using `ls -l`
  • Then, create the 'pass' directory inside 'root' using `mkdir root/pass`

$ scp ssh
Error: usage: scp [-346ABCOpqRrsTv] [-c cipher] [-D sftp_server_path] [-F ssh_config]
           [-i identity_file] [-J destination] [-l limit] [-o ssh_option]
           [-P port] [-S program] [-X sftp_option] source ... target

Problem: The scp command was invoked with incorrect arguments

Solution:
  • Review the scp command syntax: scp [options] source target
  • Specify a source file/directory to copy from
  • Specify a destination where you want to copy to
  • Add any needed options like -r for directories

Technical Implementation

Architecture Overview

The project follows a hybrid architecture combining Python and C for optimal performance:

1. Core Shell Implementation (core.c)

   • Written in C for performance-critical operations
   • Handles direct command execution
   • Manages working directory changes
   • Implements pipeline execution
   • Exposed to Python through Cython bindings

2. Python Shell Wrapper (shell.py)

   • Main shell interface implementation
   • Handles user input/output with rich formatting
   • Manages LLM integration
   • Implements command history and completion
   • Error handling and explanation generation

3. LLM Integration (llm.py)

   • Manages communication with Google's Gemini AI
   • Implements caching for faster responses
   • Handles structured command generation
   • Provides error explanation capabilities

4. Command Completion (completions.py)

   • Custom command completion engine
   • Combines traditional shell completion with LLM suggestions

Key Components

Command Response Schema

{
    "command": str,        # The shell command to execute
    "explanation": str,    # Brief explanation of the command
    "detailed_explanation": str  # Detailed breakdown with examples
}

LLM Client Features

• Lazy initialization of API client
• Response caching with JSON persistence
• Structured output parsing
• Error handling and explanation generation
• Command generation with context awareness

Shell Features

• Asynchronous command execution
• Pipeline support with error handling
• Rich terminal output with custom formatting
• History management with file persistence
• Intelligent error handling with LLM-powered explanations

Project Structure

llm_shell/
├── shell.py           # Main shell implementation
├── core.c            # C core implementation
├── core.pyx          # Cython interface
├── llm.py            # LLM client implementation
├── completions.py    # Command completion engine
├── setup.py          # Build configuration
├── requirements.txt  # Python dependencies
└── .env             # Environment configuration

Performance Optimizations

1. C Core Integration

   • Direct system calls for command execution
   • Efficient pipeline handling
   • Minimal Python-C context switching

2. LLM Response Caching

   • JSON-based persistent cache
   • Cache invalidation on model updates
   • Lazy loading of expensive resources

3. Asynchronous Operations

   • Non-blocking command execution
   • Asynchronous LLM API calls
   • Responsive UI during long operations

Security Considerations

1. API Key Management

   • Environment-based configuration
   • No hardcoded credentials
   • Secure key storage recommendations

2. Command Execution

   • Sanitized command inputs
   • Controlled execution environment
   • Error containment and reporting

3. Cache Security

   • Local-only cache storage
   • No sensitive data in cache
   • Proper file permissions

Contributing

1. Fork the repository
2. Create a feature branch
3. Implement your changes
4. Add tests if applicable
5. Submit a pull request

License

[Insert License Information]

Core Shell Implementation (core.c)

The C core provides high-performance shell operations through direct system calls. Here are the key components:

1. Shell Context Management

typedef struct {
    char *cwd;           // Current working directory
    char **env;          // Environment variables
    int last_exit_code;  // Last command's exit code
    int interactive;     // Whether shell is interactive
} ShellContext;

ShellContext* shell_init(void) {
    ShellContext *ctx = malloc(sizeof(ShellContext));
    ctx->cwd = getcwd(NULL, 0);
    // ... initialize environment and other fields
    return ctx;
}

The Shell Context maintains the state of the shell environment. The ShellContext struct holds:

• cwd: Dynamically allocated string storing the current working directory path
• env: Array of environment variables as "NAME=VALUE" strings
• last_exit_code: Exit status of the most recently executed command
• interactive: Flag indicating if the shell is running in interactive mode

During initialization, the context:

1. Allocates memory for the structure

2. Gets the current working directory using getcwd()

3. Copies the current environment variables

4. Sets default values for exit code and interactive mode

5. Command Execution Pipeline

int shell_execute_pipeline(ShellContext *ctx, const char **commands, int num_commands) {
    int pipes[num_commands-1][2];
    pid_t pids[num_commands];

    // Create pipes for command communication
    for (int i = 0; i < num_commands-1; i++) {
        pipe(pipes[i]);
    }

    // Fork processes for each command
    for (int i = 0; i < num_commands; i++) {
        pids[i] = fork();
        if (pids[i] == 0) {
            // Child: Setup pipes and execute command
            setup_pipes(i, pipes, num_commands);
            execute_command(commands[i]);
        }
    }
    // Parent: Wait for completion
    wait_for_children(pids, num_commands);
}

The pipeline execution system implements Unix-style command chaining:

1. Creates an array of pipes (N-1 pipes for N commands)

2. For each command in the pipeline:

   • Creates a new process using fork()
   • Child process:
     • Sets up input/output pipes
     • Redirects stdin/stdout to appropriate pipe ends
     • Executes the command
   • Parent process:
     • Keeps track of child PIDs
     • Manages pipe file descriptors

3. Parent waits for all child processes to complete

4. Handles errors and returns the last command's exit status

5. Directory Management

int shell_cd(ShellContext *ctx, const char *path) {
    if (chdir(path) != 0) return -1;
    free(ctx->cwd);
    ctx->cwd = getcwd(NULL, 0);
    return 0;
}

The directory management system:

1. Attempts to change directory using chdir()

2. On success:

   • Frees the old working directory string
   • Gets and stores the new working directory

3. Handles special cases like:

   • cd with no arguments (goes to HOME)
   • Relative paths
   • Symbolic links

4. Returns -1 on error (e.g., directory not found)

5. Environment Variable Handling

int shell_setenv(ShellContext *ctx, const char *name, const char *value) {
    char *new_var;
    asprintf(&new_var, "%s=%s", name, value);
    // Update or add environment variable
    update_environment(ctx, new_var);
}

Environment variable management:

1. Creates a new environment string in "NAME=VALUE" format
2. Searches existing environment for the variable
3. If found: replaces the old value
4. If not found: adds to the environment array
5. Handles memory allocation and deallocation
6. Maintains null termination of the environment array

Python Shell Wrapper (shell.py)

The Python wrapper provides high-level functionality and LLM integration:

1. Asynchronous Command Handling

async def handle_command(self, query: str):
    if query.startswith('#'):
        # Natural language query
        response = await self.llm_client.generate_command(query[1:])
        self.console.print(f"[bold bright_red]{response['command']}[/bold bright_red]")
        
        if '-vv' in query:
            self.console.print(f"[green_yellow]{response['detailed_explanation']}[/green_yellow]")
        elif '-v' in query:
            self.console.print(f"[green_yellow]{response['explanation']}[/green_yellow]")
    else:
        # Direct shell command
        await self.execute_shell_command(query)

The command handler processes two types of inputs:

1. Natural Language Queries (starting with #):

   • Strips the # prefix
   • Sends query to LLM for command generation
   • Handles verbosity flags:
     • No flag: Shows only the command
     • -v: Adds brief explanation
     • -vv: Adds detailed explanation with examples

2. Direct Shell Commands:

   • Passes directly to the C core for execution
   • Captures and handles errors
   • Maintains asynchronous operation

3. Rich Terminal Output

def get_prompt(self):
    cwd = self.core_shell.get_cwd()
    return HTML(
        f'<ansigreen>{self.username}@{self.hostname}</ansigreen>:'
        f'<ansiblue>{cwd}</ansiblue>$ '
    )

The prompt system provides:

1. Color-coded components using HTML-style formatting

2. Dynamic current directory display

3. Username and hostname information

4. ANSI color support with fallback

5. Customizable prompt structure

6. Error Handling with LLM Explanations

async def execute_shell_command(self, command: str):
    try:
        result = self.core_shell.execute(command)
        if result != 0:
            error_msg = os.popen(f"{command} 2>&1").read().strip()
            explanation = await self.llm_client.explain_error(error_msg)
            self.console.print(f"[bright_yellow]{explanation}[/bright_yellow]")
    except Exception as e:
        await self.handle_error(e)

The error handling system:

1. Executes commands through the C core
2. Captures both stdout and stderr
3. On non-zero exit codes:
   • Captures the error message
   • Sends to LLM for explanation
   • Formats and displays user-friendly explanation
4. Handles Python exceptions separately
5. Maintains async operation throughout

LLM Integration (llm.py)

The LLM integration layer manages all interactions with Google's Gemini AI:

1. Structured Command Generation

COMMAND_SCHEMA = {
    "type": "object",
    "properties": {
        "command": {
            "type": "string",
            "description": "The shell command to execute"
        },
        "explanation": {
            "type": "string", 
            "description": "Brief explanation of what the command does"
        },
        "detailed_explanation": {
            "type": "string",
            "description": "Detailed explanation including options and examples"
        }
    },
    "required": ["command", "explanation", "detailed_explanation"]
}

The command schema ensures:

1. Consistent response structure
2. Required fields are always present
3. Clear field descriptions for the LLM
4. Validation of response format
5. Easy parsing and handling of responses

Advanced Caching System

The caching system uses a multi-level approach to optimize performance:

1. Cache Structure

class LLMClient:
    def __init__(self, api_key: str):
        self.cache_file = Path.home() / '.llm_shell_cache.json'
        self._cache = {}
        self._load_cache()
        
    def _load_cache(self):
        """Load the persistent cache from disk."""
        try:
            if self.cache_file.exists():
                with open(self.cache_file, 'r') as f:
                    self.persistent_cache = json.load(f)
            else:
                self.persistent_cache = {}
        except Exception:
            self.persistent_cache = {}

The caching system implements:

1. Two-tier caching:

   • In-memory LRU cache for fastest access
   • Persistent JSON file for long-term storage

2. Version-aware caching:

   • Cache keys include version information
   • Automatic invalidation on prompt changes

3. Type-specific caching:

   • Different handling for commands vs. explanations
   • Structured vs. plain text responses

4. Cache Key Generation

def _cache_key(self, query_type: str, text: str) -> str:
    version = "v2"  # Increment when changing prompts
    return hashlib.sha256(f"{version}|{query_type}|{text}".encode()).hexdigest()

Key generation ensures:

1. Unique keys for each query type and text

2. Version-based cache invalidation

3. Consistent hashing across sessions

4. Collision-free storage

5. Secure key generation

6. Memory Cache Management

@lru_cache(maxsize=1000)
def _get_from_memory_cache(self, cache_key: str) -> Optional[str]:
    return self.persistent_cache.get(cache_key)

def _add_to_cache(self, cache_key: str, response):
    self.persistent_cache[cache_key] = response
    self._get_from_memory_cache.cache_clear()
    self._save_cache()

The memory cache:

1. Uses Python's LRU cache decorator

2. Limits memory usage (1000 entries)

3. Automatically evicts least recently used entries

4. Synchronizes with persistent storage

5. Handles cache invalidation

6. Cache Persistence

def _save_cache(self):
    """Save the persistent cache to disk."""
    try:
        with open(self.cache_file, 'w') as f:
            json.dump(self.persistent_cache, f)
    except Exception:
        pass  # Fail silently if we can't save cache

Persistence features:

1. Atomic file writing

2. Error handling for disk operations

3. JSON format for human readability

4. Automatic recovery from corruption

5. Silent failure to prevent disruption

6. Cache Security

def clear_cache(self):
    """Clear both memory and persistent cache."""
    self._get_from_memory_cache.cache_clear()
    self.persistent_cache = {}
    if self.cache_file.exists():
        self.cache_file.unlink()

Security considerations:

1. Cache file permissions
2. No sensitive data storage
3. Secure deletion option
4. Error recovery
5. Version-based invalidation

Contributing

1. Fork the repository
2. Create a feature branch
3. Implement your changes
4. Add tests if applicable
5. Submit a pull request

License

[Insert License Information]