Code Patch Agent

Demo Link - https://code-patch-agent-agentic-56ejnziy0.vercel.app/

Code Patch Agent is an agentic pair-programming assistant that automates code analysis, feature planning, code generation, and iterative safety/logic reviews. By combining a multi-agent workflow with a Retrieval-Augmented Generation (RAG) code indexing database, Code Patch Agent can index arbitrary Git repositories, read relevant source files, draft patches, and validate them prior to application.

Warning

Active Development Warning: This project is under active development. Some core components—specifically syntax validators for non-Python languages, multi-file validation loops, and complex multi-agent iterations—are experimental. Certain tasks may trigger validation errors or logic review rejections. If the reviewer agent rejects a draft patch throughout all cycles, adjust your request or prompt and run the task again.

---

1. System Architecture & Workflows

Code Patch Agent utilizes a four-tier architecture:

1. Configuration Layer (config.py): Central settings provider built with Pydantic.
2. RAG / Code Search Layer (rag/): Language-aware splitting and Max Marginal Relevance (MMR) retrieval using ChromaDB.
3. Multi-Agent Coding Pipeline (agents/): Specialized LLM nodes constrained by Pydantic structured output validation.
4. Deterministic Validation Layer (reviewer/): In-memory syntax checks and block validation.

Repository Indexing Workflow

The ingestion pipeline processes files incrementally using content hashes to minimize indexing costs:

flowchart TD
    subgraph Ingestion Phase
        A[Git Repository URL] -->|RepoManager| B[Local Clone/Pull]
        B -->|RepositoryLoader| C[Filter Extensions & Sizes]
        C -->|manifest.py| D{Compare Content Hashes}
        D -->|New/Modified| E[CodeSplitter: Language-Aware]
        D -->|Deleted| F[Delete from ChromaDB]
        D -->|Unchanged| G[Skip]
        E -->|Stable ID Generation| H[Insert into ChromaDB]
        H -->|Save Manifest| I[JSON Manifest File]
    end

Loading

Multi-Agent Code Generation & Review Loop

The task execution flow leverages a Planner, Reader (with solution suggestions block checks), Writer, and a Reviewer that performs both static syntax checking and semantic checks:

flowchart TD
    UserRequest([User Request]) --> PlannerAgent[PlannerAgent]
    PlannerAgent -->|Plan| ReaderAgent[ReaderAgent]
    ReaderAgent -->|MMR Retrieval Query| VectorDB[(ChromaDB)]
    VectorDB -->|Relevant Code Chunks| ReaderAgent
    ReaderAgent -->|Compliance Check: Verify No Code Suggestions| ReaderAgent
    ReaderAgent -->|ReaderResult + Context| WriterAgent[WriterAgent]
    WriterAgent -->|WriterResult: Proposed Patches| ReviewerAgent[ReviewerAgent]
    
    ReviewerAgent -->|Deterministic Validator| StaticValidator[StaticValidator]
    StaticValidator -->|Sequential Patch Applier| SyntaxChecker[SyntaxChecker]
    SyntaxChecker -->|Syntax Check Results| StaticValidator
    
    StaticValidator -->|Validation Pass/Fail| ReviewerAgent
    ReviewerAgent -->|LLM Review Prompt| LLMReview[LLM Logic & Safety Review]
    LLMReview -->|ReviewResult| ReviewerDecision{Approved?}
    
    ReviewerDecision -->|No: Blocking Issues| WriterAgentRevise[WriterAgent.revise]
    WriterAgentRevise -->|Revised WriterResult| ReviewerAgent
    ReviewerDecision -->|Yes| Done([Approved Patch!])

Loading

---

2. Directory Structure

├── agents/             # Multi-agent role implementations (Planner, Reader, Writer, Reviewer)
├── api/                # FastAPI web server and SSE log broadcaster
├── git_utils/          # Local repository clone, pull, and workspace managers
├── prompts/            # Structured prompts and schemas for Pydantic LLM outputs
├── public/             # Web client UI (index.html, style.css, app.js)
├── rag/                # Code indexing, loaders, splitters, and retrievers
├── reviewer/           # In-memory code patch validation and syntax checks
├── tests/              # Test suites
├── vercel.json         # Vercel Serverless routing deployment config
└── pyproject.toml      # Dependency specifications

---

3. Getting Started

Prerequisites

• Python 3.10 or 3.11
• uv (recommended Python package manager)

Installation

1. Clone this repository locally.
2. Initialize the virtual environment and install the required dependencies:

   uv venv
   .venv\Scripts\activate      # Windows PowerShell/Command Prompt
   # or source .venv/bin/activate (macOS/Linux)
   
   uv pip install -r requirements.txt

Configuration

Create a .env file in the root directory to declare your defaults:

MISTRAL_API_KEY=your_mistral_api_key_here
# Optional:
OPENAI_API_KEY=your_openai_api_key_here
GEMINI_API_KEY=your_gemini_api_key_here
GROQ_API_KEY=your_groq_api_key_here

Running the Web Server

Launch the FastAPI application locally:

.venv\Scripts\uvicorn.exe api.index:app --port 8000

Open http://localhost:8000 in your web browser.

---

4. Key Web Client Features

• Visual Agent Pipeline Tab: Tracks agent workflows (Planner, Reader, Writer, Reviewer) step-by-step in real time. Shows evaluations, goals, code context matches, and review cycle decisions.
• Configuration Manager: Customize settings dynamically at runtime (LLM model name, provider, chunk limits, MMR ratios, review cycle limits).
• API Key Local Persistence: Enter your custom provider API key inside the settings panel. It gets written directly to the project's local .env file and reloaded instantly.
• System logs: Monitor stream console logging outputs in a separate tab, with automated log scrubbing to protect secrets like API tokens and passwords.
• Diff Patches: Displays search-and-replace diff listings in a green/red visual format.
• Task Alerts: Displays warning banners with prompt improvement recommendations if a task run terminates without approval from the Reviewer Agent.