DeepAgent Assist

DeepAgent Assist is an interactive, terminal-based AI coding assistant built with LangGraph, LangChain, and the deepagents framework. It utilizes Google's Generative AI models (like Gemini) to assist developers with coding tasks directly from their terminal.

Features

• Interactive Chat Interface: A rich terminal UI built with click and rich for an engaging conversation experience.
• Human-in-the-Loop (HITL): Safe file operations. The agent pauses execution and requests user approval before executing sensitive actions like writing or editing files.
• DeepAgents Architecture:
  • FilesystemBackend: For workspace file manipulation (virtual mode enabled).
  • StateBackend & CompositeBackend: For structured state routing and management.
• Summarization Middleware: Automatically summarizes chat history to keep the LLM context manageable during long sessions.
• State Persistence: Built-in memory using LangGraph's MemorySaver to persist conversation threads.
• Web Search Tool: Integrated DuckDuckGo search (internet_search) allowing the agent to fetch real-time information without needing an API key.
• Skills System: Includes a skill system with a default langgraph-docs skill to fetch and reference LangGraph documentation using the standard llms.txt.

Project Structure

src/s_assist/
├── __init__.py
├── agent.py            # Agent configuration, tools, middleware, and backend setup
├── main.py             # CLI entry point, chat loop, and Human-in-the-Loop interrupt handling
├── skills/             # Agent skills (e.g., langgraph-docs/SKILL.md)
└── tools/              # Custom agent tools (e.g., web_search_tool.py)

Setup & Prerequisites

Make sure you have Python installed along with the necessary dependencies:

• google-auth
• langchain-google-genai
• deepagents
• langgraph
• click
• rich
• python-dotenv
• langchain-community
• duckduckgo-search

Authentication

The assistant requires access to Google Cloud's Generative AI via a Google Service Account or Application Default Credentials (ADC).

Environment Variables

You can configure the assistant using a .env file or environment variables:

• SA_FILE_PATH: (Optional) The absolute path to your Google Cloud Service Account JSON key file. If omitted, the application falls back to Application Default Credentials.
• MODELE_NAME: (Optional) The Google Generative AI model to use (default: gemini-3.5-flash).

Usage

Start the interactive chat loop by running the main module:

python -m src.s_assist.main [OPTIONS]

Options

• -d, --directory-name TEXT: The root workspace directory for the agent (defaults to the current working directory).
• -m, --model-name TEXT: The model name to use (overrides MODELE_NAME environment variable).

Interactive Commands

• Type your coding questions or instructions at the You: prompt.
• Validation Prompts: When the agent wants to write or edit a file, it will pause. You will be prompted to:
  • a (accepter): Approve the action.
  • r (refuser): Reject the action and provide a reason.
• Type quit, exit, or q to exit the chat session.