coding_agent_lib

A small Python library wrapping a LangGraph + Ollama coding agent. Instead of the original interactive input() loop, you import it and call agent.run(query) directly from your own code.

Install

pip install -e . --break-system-packages

This pulls in langgraph, langchain-ollama, and langchain-core.

You also need a running Ollama server with a model pulled locally, e.g.:

ollama pull qwen2.5-coder:7b
ollama serve

Usage

from coding_agent_lib import CodingAgent

agent = CodingAgent(model="qwen2.5-coder:7b")  # base_url defaults to localhost:11434

# Generate new code -- the agent infers intent, task, and file path from
# your natural-language query.
result = agent.run(
    "generate code for finding factorial of a number and save it to factorial.py"
)

if result.success:
    print("Written to:", result.file_path)
    print(result.code)
else:
    print("Error:", result.error)

# Debug an existing file
result = agent.run("debug factorial.py, it crashes on negative input")

CodingAgent(model=None, base_url=None, temperature=0.0, workdir=None)

• model — Ollama model name. Defaults to env var OLLAMA_MODEL or "qwen2.5-coder:7b".
• base_url — Ollama server URL. Defaults to env var OLLAMA_BASE_URL or "http://localhost:11434".
• temperature — sampling temperature (default 0, deterministic).
• workdir — if set, relative file paths extracted from queries are resolved against this directory instead of the current working directory.

agent.run(query: str) -> AgentResult

Runs a single request end-to-end (parse → generate/debug → save to disk) and returns an AgentResult:

Notes

• The agent uses the LLM itself (not regex/keyword matching) to figure out intent, task, and file path from free-form queries.
• "debug" mode reads the existing file, asks the LLM to fix it, and overwrites the file with the corrected version.
• Each call to run() is independent/stateless; call it repeatedly (in a loop, a web handler, a CLI, etc.) instead of using the old blocking input() prompt.