Code Generation Agent

Autonomous code generation agent for DPAIA tasks using Claude AI and MCP tools.

Prerequisites

• Python 3.8+
• Anthropic API key
• Generated LIB.py file (must be created separately)
• MCP server URL

Setup

1. Install dependencies:

pip install -r requirements.txt

2. Create .env file:

cp .env.example .env
# Edit .env and add your ANTHROPIC_API_KEY

Running the Agent

python codegen_dpaia_agent.py \
  --lib-path /path/to/LIB.py \
  --mcp-url http://your-mcp-server-url

Options

• --lib-path: Path to generated LIB.py file (required)
• --mcp-url: MCP server URL (required)
• --config: Path to JSON configuration file
• --repo: Repository in format owner/repo
• --issue-url: GitHub issue URL
• --issue-title: Issue title
• --issue-body: Issue description
• --model: Claude model to use (default: claude-sonnet-4-20250514)
• --max-tokens: Max tokens per request (default: 8000)
• --execute: Execute generated code and iteratively improve based on errors
• --max-iterations: Maximum number of iterations when using --execute (default: 12)
• --output: Output file path for generated code (see Output Behavior below)
• --instance-id: Custom instance ID (default: timestamp)
• --log-level: Logging level: DEBUG, INFO, WARNING, ERROR (default: INFO)

Example

python codegen_dpaia_agent.py \
  --lib-path ./LIB.py \
  --mcp-url http://localhost:8080 \
  --repo owner/repo \
  --issue-title "Fix N+1 query problem" \
  --issue-body "Detect and resolve N+1 SELECT queries" \
  --execute \
  --max-iterations 10 \
  --output generated_code.py

Pipeline Description

The agent follows a simple iterative pipeline:

1. Repository Setup: Clones the target repository into an isolated instance

2. LIB Loading: Loads the generated LIB.py file containing MCP tool functions

3. Code Generation (two modes):

   • Without execution: Generates thinking → generates code → saves results
   • With execution (--execute flag): Iteratively:
     • Generates thinking about the task
     • Generates code based on thinking and previous context
     • Executes the code
     • Analyzes execution results and errors
     • Uses error feedback to improve in the next iteration
     • Repeats until success or max iterations reached

4. Output: Saves generated code, execution logs, and iteration history:

   • results_codegen/ - execution logs
   • generated_code_iterations/<instance_id>/iteration_01.py, iteration_02.py, etc. - individual iteration code files
   • --output file (if specified):
     • Without --execute: Saves the single iteration's code
     • With --execute: Saves all iterations concatenated into one file with headers (# === ITERATION 1 ===, # === ITERATION 2 ===, etc.)