Devika is the first open-source implementation of an Agentic Software Engineer. Initially started as an open-source alternative to Devin.
Python
import time
import json
import os
import subprocess
from jinja2 import Environment, BaseLoader
from src.agents.patcher import Patcher
from src.llm import LLM
from src.state import AgentState
from src.project import ProjectManager
from src.services.utils import retry_wrapper, validate_responses
PROMPT = open("src/agents/runner/prompt.jinja2", "r").read().strip()
RERUNNER_PROMPT = open("src/agents/runner/rerunner.jinja2", "r").read().strip()
class Runner:
def __init__(self, base_model: str):
self.base_model = base_model
self.llm = LLM(model_id=base_model)
def render(
self,
conversation: str,
code_markdown: str,
system_os: str
) -> str:
env = Environment(loader=BaseLoader())
template = env.from_string(PROMPT)
return template.render(
conversation=conversation,
code_markdown=code_markdown,
system_os=system_os,
)
def render_rerunner(
self,
conversation: str,
code_markdown: str,
system_os: str,
commands: list,
error: str
):
env = Environment(loader=BaseLoader())
template = env.from_string(RERUNNER_PROMPT)
return template.render(
conversation=conversation,
code_markdown=code_markdown,
system_os=system_os,
commands=commands,
error=error
)
@validate_responses
def validate_response(self, response: str):
if "commands" not in response:
return False
else:
return response["commands"]
@validate_responses
def validate_rerunner_response(self, response: str):
if "action" not in response and "response" not in response:
return False
else:
return response
@retry_wrapper
def run_code(
self,
commands: list,
project_path: str,
project_name: str,
conversation: list,
code_markdown: str,
system_os: str
):
retries = 0
for command in commands:
command_set = command.split(" ")
command_failed = False
process = subprocess.run(
command_set,
stdout=subprocess.PIPE,
stderr=subprocess.PIPE,
cwd=project_path
)
command_output = process.stdout.decode('utf-8')
command_failed = process.returncode != 0
new_state = AgentState().new_state()
new_state["internal_monologue"] = "Running code..."
new_state["terminal_session"]["title"] = "Terminal"
new_state["terminal_session"]["command"] = command
new_state["terminal_session"]["output"] = command_output
AgentState().add_to_current_state(project_name, new_state)
time.sleep(1)
while command_failed and retries < 2:
new_state = AgentState().new_state()
new_state["internal_monologue"] = "Oh seems like there is some error... :("
new_state["terminal_session"]["title"] = "Terminal"
new_state["terminal_session"]["command"] = command
new_state["terminal_session"]["output"] = command_output
AgentState().add_to_current_state(project_name, new_state)
time.sleep(1)
prompt = self.render_rerunner(
conversation=conversation,
code_markdown=code_markdown,
system_os=system_os,
commands=commands,
error=command_output
)
response = self.llm.inference(prompt, project_name)
valid_response = self.validate_rerunner_response(response)
if not valid_response:
return False
action = valid_response["action"]
if action == "command":
command = valid_response["command"]
response = valid_response["response"]
ProjectManager().add_message_from_devika(project_name, response)
command_set = command.split(" ")
command_failed = False
process = subprocess.run(
command_set,
stdout=subprocess.PIPE,
stderr=subprocess.PIPE,
cwd=project_path
)
command_output = process.stdout.decode('utf-8')
command_failed = process.returncode != 0
new_state = AgentState().new_state()
new_state["internal_monologue"] = "Running code..."
new_state["terminal_session"]["title"] = "Terminal"
new_state["terminal_session"]["command"] = command
new_state["terminal_session"]["output"] = command_output
AgentState().add_to_current_state(project_name, new_state)
time.sleep(1)
if command_failed:
retries += 1
else:
break
elif action == "patch":
response = valid_response["response"]
ProjectManager().add_message_from_devika(project_name, response)
code = Patcher(base_model=self.base_model).execute(
conversation=conversation,
code_markdown=code_markdown,
commands=commands,
error=command_output,
system_os=system_os,
project_name=project_name
)
Patcher(base_model=self.base_model).save_code_to_project(code, project_name)
command_set = command.split(" ")
command_failed = False
process = subprocess.run(
command_set,
stdout=subprocess.PIPE,
stderr=subprocess.PIPE,
cwd=project_path
)
command_output = process.stdout.decode('utf-8')
command_failed = process.returncode != 0
new_state = AgentState().new_state()
new_state["internal_monologue"] = "Running code..."
new_state["terminal_session"]["title"] = "Terminal"
new_state["terminal_session"]["command"] = command
new_state["terminal_session"]["output"] = command_output
AgentState().add_to_current_state(project_name, new_state)
time.sleep(1)
if command_failed:
retries += 1
else:
break
@retry_wrapper
def execute(
self,
conversation: list,
code_markdown: str,
os_system: str,
project_path: str,
project_name: str
) -> str:
prompt = self.render(conversation, code_markdown, os_system)
response = self.llm.inference(prompt, project_name)
valid_response = self.validate_response(response)
self.run_code(
valid_response,
project_path,
project_name,
conversation,
code_markdown,
os_system
)
return valid_response