Devika is the first open-source implementation of an Agentic Software Engineer. Initially started as an open-source alternative to Devin.
Python
from jinja2 import Environment, BaseLoader
from src.llm import LLM
PROMPT = open("src/agents/planner/prompt.jinja2").read().strip()
class Planner:
def __init__(self, base_model: str):
self.llm = LLM(model_id=base_model)
def render(self, prompt: str) -> str:
env = Environment(loader=BaseLoader())
template = env.from_string(PROMPT)
return template.render(prompt=prompt)
def validate_response(self, response: str) -> bool:
return True
def parse_response(self, response: str):
result = {
"project": "",
"reply": "",
"focus": "",
"plans": {},
"summary": ""
}
current_section = None
current_step = None
for line in response.split("\n"):
line = line.strip()
if line.startswith("Project Name:"):
current_section = "project"
result["project"] = line.split(":", 1)[1].strip()
elif line.startswith("Your Reply to the Human Prompter:"):
current_section = "reply"
result["reply"] = line.split(":", 1)[1].strip()
elif line.startswith("Current Focus:"):
current_section = "focus"
result["focus"] = line.split(":", 1)[1].strip()
elif line.startswith("Plan:"):
current_section = "plans"
elif line.startswith("Summary:"):
current_section = "summary"
result["summary"] = line.split(":", 1)[1].strip()
elif current_section == "reply":
result["reply"] += " " + line
elif current_section == "focus":
result["focus"] += " " + line
elif current_section == "plans":
if line.startswith("- [ ] Step"):
current_step = line.split(":")[0].strip().split(" ")[-1]
result["plans"][int(current_step)] = line.split(":", 1)[1].strip()
elif current_step:
result["plans"][int(current_step)] += " " + line
elif current_section == "summary":
result["summary"] += " " + line.replace("```", "")
result["project"] = result["project"].strip()
result["reply"] = result["reply"].strip()
result["focus"] = result["focus"].strip()
result["summary"] = result["summary"].strip()
return result
def execute(self, prompt: str, project_name: str) -> str:
prompt = self.render(prompt)
response = self.llm.inference(prompt, project_name)
return response