Skip to content

Devika is the first open-source implementation of an Agentic Software Engineer. Initially started as an open-source alternative to Devin.

Python

added token usage track in agent state

k1070641bfd2ed3
diff --git a/devika.py b/devika.py
index 57b5f93..1beab9e 100644
--- a/devika.py
+++ b/devika.py
@@ -160,8 +160,9 @@ def calculate_tokens():
@app.route("/api/token-usage", methods=["GET"])
@route_logger(logger)
def token_usage():
- from src.llm import TOKEN_USAGE
- return jsonify({"token_usage": TOKEN_USAGE})
+ project_name = request.args.get("project_name")
+ token_count = AgentState().get_latest_token_usage(project_name)
+ return jsonify({"token_usage": token_count})
@app.route("/api/real-time-logs", methods=["GET"])
diff --git a/src/agents/action/action.py b/src/agents/action/action.py
index f61f7ba..5544d60 100644
--- a/src/agents/action/action.py
+++ b/src/agents/action/action.py
@@ -39,15 +39,15 @@ class Action:
else:
return response["response"], response["action"]
- def execute(self, conversation: list) -> str:
+ def execute(self, conversation: list, project_name: str) -> str:
prompt = self.render(conversation)
- response = self.llm.inference(prompt)
+ response = self.llm.inference(prompt, project_name)
valid_response = self.validate_response(response)
while not valid_response:
print("Invalid response from the model, trying again...")
- return self.execute(conversation)
+ return self.execute(conversation, project_name)
print("===" * 10)
print(valid_response)
diff --git a/src/agents/agent.py b/src/agents/agent.py
index 204360c..165bf23 100644
--- a/src/agents/agent.py
+++ b/src/agents/agent.py
@@ -93,7 +93,8 @@ class Agent:
Formatter Agent is invoked to format and learn from the contents
"""
results[query] = self.formatter.execute(
- browser.extract_text()
+ browser.extract_text(),
+ project_name
)
"""
@@ -118,7 +119,7 @@ class Agent:
Decision making Agent
"""
def make_decision(self, prompt: str, project_name: str) -> str:
- decision = self.decision.execute(prompt)
+ decision = self.decision.execute(prompt, project_name)
for item in decision:
function = item["function"]
@@ -134,7 +135,7 @@ class Agent:
elif function == "generate_pdf_document":
user_prompt = args["user_prompt"]
# Call the reporter agent to generate the PDF document
- markdown = self.reporter.execute([user_prompt], "")
+ markdown = self.reporter.execute([user_prompt], "", project_name)
_out_pdf_file = PDF().markdown_to_pdf(markdown, project_name)
project_name_space_url = project_name.replace(" ", "%20")
@@ -154,10 +155,10 @@ class Agent:
elif function == "coding_project":
user_prompt = args["user_prompt"]
# Call the planner, researcher, coder agents in sequence
- plan = self.planner.execute(user_prompt)
+ plan = self.planner.execute(user_prompt, project_name)
planner_response = self.planner.parse_response(plan)
- research = self.researcher.execute(plan, self.collected_context_keywords)
+ research = self.researcher.execute(plan, self.collected_context_keywords, project_name)
search_results = self.search_queries(research["queries"], project_name)
code = self.coder.execute(
@@ -177,7 +178,7 @@ class Agent:
conversation = ProjectManager().get_all_messages_formatted(project_name)
code_markdown = ReadCode(project_name).code_set_to_markdown()
- response, action = self.action.execute(conversation)
+ response, action = self.action.execute(conversation, project_name)
ProjectManager().add_message_from_devika(project_name, response)
@@ -188,7 +189,8 @@ class Agent:
if action == "answer":
response = self.answer.execute(
conversation=conversation,
- code_markdown=code_markdown
+ code_markdown=code_markdown,
+ project_name=project_name
)
ProjectManager().add_message_from_devika(project_name, response)
elif action == "run":
@@ -238,7 +240,7 @@ class Agent:
self.patcher.save_code_to_project(code, project_name)
elif action == "report":
- markdown = self.reporter.execute(conversation, code_markdown)
+ markdown = self.reporter.execute(conversation, code_markdown, project_name)
_out_pdf_file = PDF().markdown_to_pdf(markdown, project_name)
@@ -261,7 +263,7 @@ class Agent:
if project_name_from_user:
ProjectManager().add_message_from_user(project_name_from_user, prompt)
- plan = self.planner.execute(prompt)
+ plan = self.planner.execute(prompt, project_name_from_user)
print(plan)
print("=====" * 10)
@@ -288,7 +290,7 @@ class Agent:
self.update_contextual_keywords(focus)
print(self.collected_context_keywords)
- internal_monologue = self.internal_monologue.execute(current_prompt=plan)
+ internal_monologue = self.internal_monologue.execute(current_prompt=plan, project_name=project_name)
print(internal_monologue)
print("=====" * 10)
@@ -296,7 +298,7 @@ class Agent:
new_state["internal_monologue"] = internal_monologue
AgentState().add_to_current_state(project_name, new_state)
- research = self.researcher.execute(plan, self.collected_context_keywords)
+ research = self.researcher.execute(plan, self.collected_context_keywords, project_name)
print(research)
print("=====" * 10)
diff --git a/src/agents/answer/answer.py b/src/agents/answer/answer.py
index 6c60951..94e2575 100644
--- a/src/agents/answer/answer.py
+++ b/src/agents/answer/answer.py
@@ -40,14 +40,14 @@ class Answer:
else:
return response["response"]
- def execute(self, conversation: list, code_markdown: str) -> str:
+ def execute(self, conversation: list, code_markdown: str, project_name: str) -> str:
prompt = self.render(conversation, code_markdown)
- response = self.llm.inference(prompt)
+ response = self.llm.inference(prompt, project_name)
valid_response = self.validate_response(response)
while not valid_response:
print("Invalid response from the model, trying again...")
- return self.execute(conversation, code_markdown)
+ return self.execute(conversation, code_markdown, project_name)
return valid_response
diff --git a/src/agents/coder/coder.py b/src/agents/coder/coder.py
index cd2b0d7..6584644 100644
--- a/src/agents/coder/coder.py
+++ b/src/agents/coder/coder.py
@@ -102,13 +102,13 @@ class Coder:
project_name: str
) -> str:
prompt = self.render(step_by_step_plan, user_context, search_results)
- response = self.llm.inference(prompt)
+ response = self.llm.inference(prompt, project_name)
valid_response = self.validate_response(response)
while not valid_response:
print("Invalid response from the model, trying again...")
- return self.execute(step_by_step_plan, user_context, search_results)
+ return self.execute(step_by_step_plan, user_context, search_results, project_name)
print(valid_response)
diff --git a/src/agents/decision/decision.py b/src/agents/decision/decision.py
index d11dba3..6a98445 100644
--- a/src/agents/decision/decision.py
+++ b/src/agents/decision/decision.py
@@ -32,14 +32,14 @@ class Decision:
return response
- def execute(self, prompt: str) -> str:
+ def execute(self, prompt: str, project_name: str) -> str:
prompt = self.render(prompt)
- response = self.llm.inference(prompt)
+ response = self.llm.inference(prompt, project_name)
valid_response = self.validate_response(response)
while not valid_response:
print("Invalid response from the model, trying again...")
- return self.execute(prompt)
+ return self.execute(prompt, project_name)
return valid_response
\ No newline at end of file
diff --git a/src/agents/feature/feature.py b/src/agents/feature/feature.py
index fa4a273..cf4f88a 100644
--- a/src/agents/feature/feature.py
+++ b/src/agents/feature/feature.py
@@ -103,7 +103,7 @@ class Feature:
project_name: str
) -> str:
prompt = self.render(conversation, code_markdown, system_os)
- response = self.llm.inference(prompt)
+ response = self.llm.inference(prompt, project_name)
valid_response = self.validate_response(response)
diff --git a/src/agents/formatter/formatter.py b/src/agents/formatter/formatter.py
index 9cd062a..f6e3e7a 100644
--- a/src/agents/formatter/formatter.py
+++ b/src/agents/formatter/formatter.py
@@ -16,7 +16,7 @@ class Formatter:
def validate_response(self, response: str) -> bool:
return True
- def execute(self, raw_text: str) -> str:
+ def execute(self, raw_text: str, project_name: str) -> str:
raw_text = self.render(raw_text)
- response = self.llm.inference(raw_text)
+ response = self.llm.inference(raw_text, project_name)
return response
\ No newline at end of file
diff --git a/src/agents/internal_monologue/internal_monologue.py b/src/agents/internal_monologue/internal_monologue.py
index 48460ed..6a81820 100644
--- a/src/agents/internal_monologue/internal_monologue.py
+++ b/src/agents/internal_monologue/internal_monologue.py
@@ -33,15 +33,15 @@ class InternalMonologue:
else:
return response["internal_monologue"]
- def execute(self, current_prompt: str) -> str:
+ def execute(self, current_prompt: str, project_name: str) -> str:
current_prompt = self.render(current_prompt)
- response = self.llm.inference(current_prompt)
+ response = self.llm.inference(current_prompt, project_name)
valid_response = self.validate_response(response)
while not valid_response:
print("Invalid response from the model, trying again...")
- return self.execute(current_prompt)
+ return self.execute(current_prompt, project_name)
return valid_response
diff --git a/src/agents/patcher/patcher.py b/src/agents/patcher/patcher.py
index 3be452f..b3108dd 100644
--- a/src/agents/patcher/patcher.py
+++ b/src/agents/patcher/patcher.py
@@ -115,7 +115,7 @@ class Patcher:
error,
system_os
)
- response = self.llm.inference(prompt)
+ response = self.llm.inference(prompt, project_name)
valid_response = self.validate_response(response)
diff --git a/src/agents/planner/planner.py b/src/agents/planner/planner.py
index 3b77e65..3f24177 100644
--- a/src/agents/planner/planner.py
+++ b/src/agents/planner/planner.py
@@ -65,7 +65,7 @@ class Planner:
return result
- def execute(self, prompt: str) -> str:
+ def execute(self, prompt: str, project_name: str) -> str:
prompt = self.render(prompt)
- response = self.llm.inference(prompt)
+ response = self.llm.inference(prompt, project_name)
return response
diff --git a/src/agents/reporter/reporter.py b/src/agents/reporter/reporter.py
index 43790f3..8cfc98d 100644
--- a/src/agents/reporter/reporter.py
+++ b/src/agents/reporter/reporter.py
@@ -28,16 +28,17 @@ class Reporter:
def execute(self,
conversation: list,
- code_markdown: str
+ code_markdown: str,
+ project_name: str
) -> str:
prompt = self.render(conversation, code_markdown)
- response = self.llm.inference(prompt)
+ response = self.llm.inference(prompt, project_name)
valid_response = self.validate_response(response)
while not valid_response:
print("Invalid response from the model, trying again...")
- return self.execute(conversation, code_markdown)
+ return self.execute(conversation, code_markdown, project_name)
return valid_response
diff --git a/src/agents/researcher/researcher.py b/src/agents/researcher/researcher.py
index 309a806..c9c2574 100644
--- a/src/agents/researcher/researcher.py
+++ b/src/agents/researcher/researcher.py
@@ -42,16 +42,16 @@ class Researcher:
"ask_user": response["ask_user"]
}
- def execute(self, step_by_step_plan: str, contextual_keywords: List[str]) -> str:
+ def execute(self, step_by_step_plan: str, contextual_keywords: List[str], project_name: str) -> str:
contextual_keywords = ", ".join(map(lambda k: k.capitalize(), contextual_keywords))
step_by_step_plan = self.render(step_by_step_plan, contextual_keywords)
- response = self.llm.inference(step_by_step_plan)
+ response = self.llm.inference(step_by_step_plan, project_name)
valid_response = self.validate_response(response)
while not valid_response:
print("Invalid response from the model, trying again...")
- return self.execute(step_by_step_plan, contextual_keywords)
+ return self.execute(step_by_step_plan, contextual_keywords, project_name)
return valid_response
\ No newline at end of file
diff --git a/src/agents/runner/runner.py b/src/agents/runner/runner.py
index 8e61c3a..0ec84d4 100644
--- a/src/agents/runner/runner.py
+++ b/src/agents/runner/runner.py
@@ -136,7 +136,7 @@ class Runner:
error=command_output
)
- response = self.llm.inference(prompt)
+ response = self.llm.inference(prompt, project_name)
valid_response = self.validate_rerunner_response(response)
@@ -233,7 +233,7 @@ class Runner:
project_name: str
) -> str:
prompt = self.render(conversation, code_markdown, os_system)
- response = self.llm.inference(prompt)
+ response = self.llm.inference(prompt, project_name)
valid_response = self.validate_response(response)
diff --git a/src/llm/__init__.py b/src/llm/__init__.py
index 6c2bf5c..5ab9f9c 100644
--- a/src/llm/__init__.py
+++ b/src/llm/__init__.py
@@ -1 +1 @@
-from .llm import LLM, TOKEN_USAGE
\ No newline at end of file
+from .llm import LLM
\ No newline at end of file
diff --git a/src/llm/llm.py b/src/llm/llm.py
index 4080b5e..c5c4843 100644
--- a/src/llm/llm.py
+++ b/src/llm/llm.py
@@ -4,9 +4,10 @@ from .ollama_client import Ollama
from .claude_client import Claude
from .openai_client import OpenAI
+from src.state import AgentState
+
import tiktoken
-TOKEN_USAGE = 0
TIKTOKEN_ENC = tiktoken.get_encoding("cl100k_base")
class Model(Enum):
@@ -38,15 +39,14 @@ class LLM:
models.update(ollama_models)
return models
- def update_global_token_usage(self, string: str):
- global TOKEN_USAGE
- TOKEN_USAGE += len(TIKTOKEN_ENC.encode(string))
- print(f"Token usage: {TOKEN_USAGE}")
+ def update_global_token_usage(self, string: str, project_name: str):
+ token_usage = len(TIKTOKEN_ENC.encode(string))
+ AgentState().update_token_usage(project_name, token_usage)
def inference(
- self, prompt: str
+ self, prompt: str, project_name: str
) -> str:
- self.update_global_token_usage(prompt)
+ self.update_global_token_usage(prompt, project_name)
model = self.model_id_to_enum_mapping()[self.model_id]
@@ -59,6 +59,6 @@ class LLM:
else:
raise ValueError(f"Model {model} not supported")
- self.update_global_token_usage(response)
+ self.update_global_token_usage(response, project_name)
return response
diff --git a/src/state.py b/src/state.py
index 323146f..412b7ad 100644
--- a/src/state.py
+++ b/src/state.py
@@ -132,4 +132,27 @@ class AgentState:
agent_state = session.query(AgentStateModel).filter(AgentStateModel.project == project).first()
if agent_state:
return json.loads(agent_state.state_stack_json)[-1]["completed"]
- return None
\ No newline at end of file
+ return None
+
+ def update_token_usage(self, project: str, token_usage: int):
+ with Session(self.engine) as session:
+ agent_state = session.query(AgentStateModel).filter(AgentStateModel.project == project).first()
+ print(agent_state)
+ if agent_state:
+ state_stack = json.loads(agent_state.state_stack_json)
+ state_stack[-1]["token_usage"] += token_usage
+ agent_state.state_stack_json = json.dumps(state_stack)
+ session.commit()
+ else:
+ state_stack = [self.new_state()]
+ state_stack[-1]["token_usage"] = token_usage
+ agent_state = AgentStateModel(project=project, state_stack_json=json.dumps(state_stack))
+ session.add(agent_state)
+ session.commit()
+
+ def get_latest_token_usage(self, project: str):
+ with Session(self.engine) as session:
+ agent_state = session.query(AgentStateModel).filter(AgentStateModel.project == project).first()
+ if agent_state:
+ return json.loads(agent_state.state_stack_json)[-1]["token_usage"]
+ return 0
\ No newline at end of file