Skip to content

armgit / E2B

Public Imported from GitHub
0

Open-source, secure environment with real-world tools for enterprise-grade agents.

Fix smol dev log DB updating

Tomas Valentaaee5fa1cf9be
diff --git a/api-service/agent/base.py b/api-service/agent/base.py
index 619b6b32e..e2f5a56fa 100644
--- a/api-service/agent/base.py
+++ b/api-service/agent/base.py
@@ -32,6 +32,7 @@ class AgentInteractionRequest(BaseModel):
OnLogs = Callable[[Any], Coroutine[None, None, None]]
+SetRun = Callable[[str], None]
OnInteractionRequest = Callable[[AgentInteractionRequest], Coroutine[None, None, None]]
GetEnvs = Callable[[], Coroutine[Any, Any, List[EnvVar]]]
diff --git a/api-service/agent/basic_agent.py b/api-service/agent/basic_agent.py
deleted file mode 100644
index bd37c805d..000000000
--- a/api-service/agent/basic_agent.py
+++ /dev/null
@@ -1,375 +0,0 @@
-import asyncio
-import uuid
-import chevron
-
-# Monkey patch chevron to not escape html
-chevron.render.__globals__["_html_escape"] = lambda string: string
-
-from typing import Any, Dict, List, Literal, cast
-from langchain.agents.tools import InvalidTool
-from langchain.callbacks.base import AsyncCallbackManager
-from langchain.callbacks.streaming_stdout import StreamingStdOutCallbackHandler
-from langchain.chains.llm import LLMChain
-from langchain.schema import BaseMessage, SystemMessage, HumanMessage
-from langchain.tools import BaseTool
-from langchain.prompts.chat import (
- AIMessagePromptTemplate,
- BaseMessagePromptTemplate,
- ChatPromptTemplate,
-)
-
-from agent.base import (
- AgentInteraction,
- OnLogs,
- OnInteractionRequest,
-)
-from agent.output.parse_output import ToolLog
-from agent.output.output_stream_parser import OutputStreamParser, Step
-from agent.output.token_callback_handler import TokenCallbackHandler
-from agent.output.work_queue import WorkQueue
-from agent.tools.base import create_tools
-from models.base import ModelConfig, PromptPart, get_model
-from agent.base import AgentBase, AgentInteractionRequest, GetEnvs
-from session import NodeJSPlayground
-
-
-class RewriteStepsException(Exception):
- pass
-
-
-class BasicAgent(AgentBase):
- max_run_time = 60 * 60 * 24 # in seconds
-
- def __init__(
- self,
- config: Any,
- get_envs: GetEnvs,
- on_logs: OnLogs,
- on_interaction_request: OnInteractionRequest,
- ):
- super().__init__()
- self._start_loop: asyncio.Task | None = None
- self.get_envs = get_envs
- self.config = ModelConfig(**config)
- self.on_interaction_request = on_interaction_request
- self.on_logs = on_logs
- self.should_pause = False
- self.canceled = False
- self.rewriting_steps = False
- self.llm_generation = None
- self.can_resume = asyncio.Event()
-
- @classmethod
- async def create(
- cls,
- config: Any,
- get_envs: GetEnvs,
- on_logs: OnLogs,
- on_interaction_request: OnInteractionRequest,
- ):
- return cls(
- config,
- get_envs,
- on_logs,
- on_interaction_request,
- )
-
- @staticmethod
- async def _run_tool(
- tool_log: ToolLog,
- name_to_tool_map: Dict[str, BaseTool],
- ) -> str:
- # Otherwise we lookup the tool
- if tool_log["tool_name"] in name_to_tool_map:
- tool = name_to_tool_map[tool_log["tool_name"]]
- # We then call the tool on the tool input to get an observation
- observation = await tool.arun(
- tool_log["tool_input"],
- verbose=True,
- )
- else:
- observation = await InvalidTool().arun( # type: ignore
- tool_name=tool_log["tool_name"],
- tool_input=tool_log["tool_input"],
- verbose=True,
- )
- return observation
-
- @staticmethod
- def _get_prompt_part(
- prompts: List[PromptPart],
- role: Literal["user", "system"],
- type: str,
- ):
- return (
- next(
- prompt
- for prompt in prompts
- if prompt.role == role and prompt.type == type
- )
- # Double the parenthesses to escape them because the string will be prompt templated.
- .content
- )
-
- @staticmethod
- def _create_prompt(
- config: ModelConfig,
- tools: List[BaseTool],
- instructions: Any,
- ):
- system_template = chevron.render(
- "\n\n".join(
- [
- BasicAgent._get_prompt_part(config.prompt, "system", "prefix"),
- "\n".join([f"{tool.name}: {tool.description}" for tool in tools]),
- BasicAgent._get_prompt_part(config.prompt, "system", "suffix"),
- ]
- ),
- instructions,
- )
-
- human_template = chevron.render(
- "\n\n".join(
- [
- BasicAgent._get_prompt_part(config.prompt, "user", "prefix"),
- ]
- ),
- instructions,
- )
-
- messages: List[
- BaseMessage | BaseMessagePromptTemplate | SystemMessage | HumanMessage
- ] = [
- SystemMessage(content=system_template),
- HumanMessage(content=human_template),
- AIMessagePromptTemplate.from_template("{agent_scratchpad}"),
- ]
- return ChatPromptTemplate(
- input_variables=["agent_scratchpad"],
- messages=messages,
- )
-
- async def _run(self, instructions: Any):
- playground = None
- try:
- # -----
- # Initialize callback manager and handler for controlling token stream
- callback_manager = AsyncCallbackManager([StreamingStdOutCallbackHandler()])
- self.token_handler = TokenCallbackHandler()
- callback_manager.add_handler(self.token_handler)
-
- # -----
- # Create tools
- # Create playground for code tools
- playground = NodeJSPlayground(get_envs=self.get_envs)
- await playground.open()
-
- tools = list(create_tools(playground=playground))
- self.tool_names = [tool.name for tool in tools]
- tool_map = {tool.name: tool for tool in tools}
-
- # Assign callback manager to tools
- for tool in tools:
- tool.callback_manager = callback_manager
-
- # -----
- # Create LLM from specified models
- prompt = self._create_prompt(self.config, tools, instructions)
- llm_chain = LLMChain(
- llm=get_model(self.config, callback_manager),
- prompt=prompt,
- callback_manager=callback_manager,
- verbose=True,
- )
-
- # -----
- # Create log handlers
- # Used for sending logs to db/frontend without blocking
- steps_streamer = WorkQueue[List[Step]](
- on_workload=lambda steps: self.on_logs(steps),
- )
-
- def stream_steps(steps: List[Step]):
- steps_streamer.schedule(steps)
-
- # Used for parsing token stream into logs+actions
- self._output_parser = OutputStreamParser(tool_names=self.tool_names)
-
- # This function is used to inject information from previous steps to the current prompt
- # TODO: Improve the scratchpad handling and prompts for templates
- def get_agent_scratchpad():
- steps = self._output_parser.get_steps()
-
- if (
- len(steps) == 0
- or len(steps) == 1
- and not self._output_parser._token_buffer
- ):
- return ""
-
- agent_scratchpad = ""
- for step in steps:
- agent_scratchpad += step["output"]
-
- tool_logs = (
- cast(ToolLog, action)
- for action in step["logs"]
- if action["type"] == "tool"
- )
-
- tool_outputs = "\n".join(
- log.get("tool_output", "")
- for log in tool_logs
- if log.get("tool_output")
- )
-
- if tool_outputs:
- agent_scratchpad += f"\nObservation: {tool_outputs}\nThought:"
-
- return f"This is my previous work (I will continue where I left off):\n{agent_scratchpad}"
-
- print("Generating...", flush=True)
- while True:
- try:
- await self._check_interrupt()
- # Query the LLM in background
- scratchpad = get_agent_scratchpad()
- self.llm_generation = asyncio.create_task(
- llm_chain.agenerate(
- [
- {
- "agent_scratchpad": scratchpad,
- "stop": "Observation:",
- }
- ]
- )
- )
- while True:
- await self._check_interrupt()
- # Consume the tokens from LLM
- token = await self.token_handler.retrieve_token()
-
- if token is None:
- break
-
- await self._check_interrupt()
- self._output_parser.ingest_token(token)
- stream_steps(self._output_parser.get_steps())
-
- current_step = self._output_parser.get_current_step()
-
- # Check if the run is finished
- if any(
- keyword in current_step["output"]
- for keyword in ("Final Answer", "FinalAnswer", "Finished")
- ):
- break
-
- for tool in (
- cast(ToolLog, action)
- for action in current_step["logs"]
- if action["type"] == "tool"
- ):
- await self._check_interrupt()
- tool_output = await self._run_tool(
- tool,
- name_to_tool_map=tool_map,
- )
- await self._check_interrupt()
- self._output_parser.ingest_tool_output(tool_output)
- stream_steps(self._output_parser.get_steps())
-
- except RewriteStepsException:
- print("REWRITING")
- continue
-
- except:
- raise
- finally:
- if playground is not None:
- await playground.close()
- await self.on_interaction_request(
- AgentInteractionRequest(
- interaction_id=str(uuid.uuid4()),
- type="done",
- )
- )
-
- async def _check_interrupt(self):
- if self.canceled:
- raise Exception("Canceled")
-
- if self.should_pause:
- await self.can_resume.wait()
- self.can_resume.clear()
-
- if self.rewriting_steps:
- self.rewriting_steps = False
- raise RewriteStepsException()
-
- async def _start(self, instructions: Any):
- print("Start agent run")
-
- async def start_with_timeout():
- try:
- self._start_loop = asyncio.create_task(
- self._run(instructions=instructions),
- )
-
- await asyncio.wait_for(
- self._start_loop,
- timeout=self.max_run_time,
- )
- except asyncio.TimeoutError:
- await self.stop()
- print("Timeout")
-
- asyncio.create_task(
- start_with_timeout(),
- )
-
- async def _pause(self):
- print("Pause agent run")
- self.should_pause = True
-
- async def _resume(self):
- print("Resume agent run")
- self.should_pause = False
- self.can_resume.set()
-
- async def _rewrite_steps(self, steps: List[Step]):
- print("Rewrite agent run steps")
- await self._pause()
- self.rewriting_steps = True
- if self.llm_generation:
- self.llm_generation.cancel()
- # TODO: Communicate the rewriting in model prompt - inform about what was rewritten.
- # TODO: Handle token that is generated after a rewrite - if the current step seems "finished" the model still outputs string (usually something like . or :).
- self._output_parser = OutputStreamParser(
- tool_names=self.tool_names,
- steps=steps[:-1],
- buffered_step=steps[-1],
- )
- await self._resume()
-
- async def interaction(self, interaction: AgentInteraction):
- print("Agent interaction")
- match interaction.type:
- case "pause":
- await self._pause()
- case "resume":
- await self._resume()
- case "start":
- await self._start(interaction.data["instructions"])
- case "rewrite_steps":
- await self._rewrite_steps(interaction.data["steps"])
- case _:
- raise Exception(f"Unknown interaction action: {interaction.type}")
-
- async def stop(self):
- print("Cancel agent run")
- self.canceled = True
- if self.llm_generation:
- self.llm_generation.cancel()
- if self._start_loop:
- self._start_loop.cancel()
diff --git a/api-service/agent/from_template.py b/api-service/agent/from_template.py
index 363ebc871..e88824dc3 100644
--- a/api-service/agent/from_template.py
+++ b/api-service/agent/from_template.py
@@ -1,4 +1,3 @@
-from agent.basic_agent import BasicAgent
from agent.smol_agent import SmolAgent
@@ -7,4 +6,4 @@ def get_agent_factory_from_template(template_id: str):
case "SmolDeveloper":
return SmolAgent.create
case _:
- return BasicAgent.create
+ raise ValueError(f"Unknown template {template_id}")
diff --git a/api-service/agent/smol_agent.py b/api-service/agent/smol_agent.py
index 46f068170..1f9edf0a0 100644
--- a/api-service/agent/smol_agent.py
+++ b/api-service/agent/smol_agent.py
@@ -3,7 +3,7 @@ import uuid
import os
import ast
-from typing import Any, Coroutine, List
+from typing import Any, Callable, Coroutine, List
from langchain.callbacks.base import AsyncCallbackManager
from langchain.schema import (
AIMessage,
@@ -45,6 +45,7 @@ class SmolAgent(AgentBase):
self,
config: Any,
get_envs: GetEnvs,
+ set_run_id: Callable[[str], None],
on_logs: OnLogs,
on_interaction_request: OnInteractionRequest,
model: BaseLanguageModel,
@@ -55,6 +56,7 @@ class SmolAgent(AgentBase):
self.config = config
self.on_interaction_request = on_interaction_request
self.on_logs = on_logs
+ self.set_run_id = set_run_id
self.model = model
@classmethod
@@ -62,6 +64,7 @@ class SmolAgent(AgentBase):
cls,
config: Any,
get_envs: GetEnvs,
+ set_run_id: Callable[[str], None],
on_logs: OnLogs,
on_interaction_request: OnInteractionRequest,
):
@@ -76,6 +79,7 @@ class SmolAgent(AgentBase):
return cls(
new_config,
get_envs,
+ set_run_id,
on_logs,
on_interaction_request,
model,
@@ -84,7 +88,6 @@ class SmolAgent(AgentBase):
async def generate_file(
self,
filename: str,
- run_id: str,
filepaths_string=None,
shared_dependencies=None,
prompt=None,
@@ -132,10 +135,9 @@ Begin generating the code now.
**metadata,
"result": filecode,
"model": "gpt-4",
- "run_id": run_id,
},
"type": "model",
- }
+ },
)
return filename, filecode, metadata
@@ -186,7 +188,7 @@ Begin generating the code now.
"cost": cost,
}
- async def _dev(self, run_id: str, instructions: Any):
+ async def _dev(self, instructions: Any):
with tracer.start_as_current_span("agent-run") as span:
user_prompt: str = instructions["Prompt"]
access_token: str = instructions["AccessToken"]
@@ -199,7 +201,7 @@ Begin generating the code now.
span.set_attributes(
{
- "run_id": run_id,
+ "run_id": self.run_id,
"prompt": user_prompt,
"prompt.length": len(user_prompt),
"repository": f"{owner}/{repo}",
@@ -215,10 +217,9 @@ Begin generating the code now.
"message": f"hi its me, 🐣the smol developer🐣!",
"type": "info",
"properties": {
- "run_id": run_id,
"prompt": user_prompt,
},
- }
+ },
)
playground = None
try:
@@ -230,10 +231,9 @@ Begin generating the code now.
"properties": {
"playground": "created",
"id": playground.id,
- "run_id": run_id,
},
"type": "playground",
- }
+ },
)
span.add_event(
"playground-created",
@@ -260,10 +260,9 @@ Begin generating the code now.
"properties": {
"tool": "git",
"repository": f"{owner}/{repo}",
- "run_id": run_id,
},
"type": "tool",
- }
+ },
)
span.add_event(
"repository-cloned",
@@ -302,10 +301,9 @@ Begin generating the code now.
**metadata,
"result": filepaths_string,
"model": "gpt-4",
- "run_id": run_id,
},
"type": "model",
- }
+ },
)
span.add_event(
@@ -340,10 +338,9 @@ Begin generating the code now.
"message": f"Cleaned root directory",
"properties": {
"tool": "filesystem",
- "run_id": run_id,
},
"type": "tool",
- }
+ },
)
span.add_event(
@@ -381,10 +378,9 @@ Begin generating the code now.
**metadata,
"result": shared_dependencies,
"model": "gpt-4",
- "run_id": run_id,
},
"type": "model",
- }
+ },
)
span.add_event(
@@ -409,10 +405,9 @@ Begin generating the code now.
"filename": "shared_dependencies.md",
"content": shared_dependencies,
"tool": "filesystem",
- "run_id": run_id,
},
"type": "tool",
- }
+ },
)
span.add_event(
@@ -435,7 +430,6 @@ Begin generating the code now.
with_semaphore(
self.generate_file(
name,
- run_id=run_id,
filepaths_string=filepaths_string,
shared_dependencies=shared_dependencies,
prompt=user_prompt,
@@ -464,10 +458,9 @@ Begin generating the code now.
"filename": filepath,
"content": content,
"tool": "filesystem",
- "run_id": run_id,
},
"type": "tool",
- }
+ },
)
span.add_event(
"file-saved",
@@ -493,10 +486,9 @@ Begin generating the code now.
"branch": branch,
"commit_message": commit_message,
"tool": "git",
- "run_id": run_id,
},
"type": "tool",
- }
+ },
)
span.add_event(
"pushed-repository",
@@ -521,11 +513,10 @@ Begin generating the code now.
{
"message": f"Agent run failed",
"properties": {
- "run_id": run_id,
"error": str(e),
},
"type": "error",
- }
+ },
)
span.set_status(Status(StatusCode.ERROR))
span.record_exception(e)
@@ -538,10 +529,9 @@ Begin generating the code now.
"message": f"Closed playground",
"properties": {
"playground": "closed",
- "run_id": run_id,
},
"type": "playground",
- }
+ },
)
span.add_event(
"closed-playground",
@@ -550,11 +540,9 @@ Begin generating the code now.
await self.on_logs(
{
"message": f"Agent run finished",
- "properties": {
- "run_id": run_id,
- },
+ "properties": {},
"type": "info",
- }
+ },
)
async def _dev_in_background(self, instructions: Any):
@@ -565,11 +553,12 @@ Begin generating the code now.
await self.stop()
self.run_id = str(uuid.uuid4())
+ self.set_run_id(self.run_id)
async def start_with_timeout():
try:
self._dev_loop = asyncio.create_task(
- self._dev(run_id=self.run_id, instructions=instructions),
+ self._dev(instructions=instructions),
)
await asyncio.wait_for(
self._dev_loop,
diff --git a/api-service/app.py b/api-service/app.py
index d11cac08d..00c8d6901 100644
--- a/api-service/app.py
+++ b/api-service/app.py
@@ -63,7 +63,6 @@ async def create_agent_deployment(
db_deployment["id"],
project_id,
body.config,
- db_deployment.get("logs") or [],
)
current_span = trace.get_current_span()
current_span.set_attributes(
@@ -88,7 +87,6 @@ async def create_agent_deployment(
id,
project_id,
body.config,
- [],
)
current_span = trace.get_current_span()
current_span.set_attributes(
@@ -156,7 +154,6 @@ async def interact_with_agent_deployment(
db_deployment["id"],
db_deployment["project_id"],
db_deployment["config"],
- db_deployment.get("logs") or [],
)
else:
raise HTTPException(status_code=404, detail="Deployment not found")
diff --git a/api-service/database/database.py b/api-service/database/database.py
index 1d2332646..d23f49d93 100644
--- a/api-service/database/database.py
+++ b/api-service/database/database.py
@@ -1,3 +1,4 @@
+import json
from typing import Any, List
from agent.output.output_stream_parser import Step
@@ -5,6 +6,7 @@ from database.client import Client
from session.env import EnvVar
TABLE_DEPLOYMENTS = "deployments"
+TABLE_LOGS = "log_files"
TABLE_PROJECTS = "projects"
@@ -20,7 +22,6 @@ class Database:
.limit(1)
.execute()
)
-
return None if len(response.data) == 0 else response.data[0]
async def get_deployment_by_id(self, id: str):
@@ -45,14 +46,23 @@ class Database:
async def update_deployment_logs(
self,
- id: str,
+ deployment_id: str,
+ run_id: str | None,
+ project_id: str,
logs: List[Step],
):
- await self.client.table(TABLE_DEPLOYMENTS).update(
+ if run_id is None:
+ return
+
+ await self.client.table(TABLE_LOGS).upsert(
{
- "logs": logs,
- }
- ).eq("id", id).execute()
+ "id": run_id,
+ "project_id": project_id,
+ "deployment_id": deployment_id,
+ "content": json.dumps(logs),
+ },
+ on_conflict="id",
+ ).execute()
async def update_deployment(self, id: str, enabled: bool) -> None:
await self.client.table(TABLE_DEPLOYMENTS).update(
diff --git a/api-service/deployment/in_memory.py b/api-service/deployment/in_memory.py
index 15d90b3d1..21cd60158 100644
--- a/api-service/deployment/in_memory.py
+++ b/api-service/deployment/in_memory.py
@@ -1,4 +1,4 @@
-from typing import Any, Dict, List
+from typing import Any, Dict
from opentelemetry.metrics import Observation
from database.base import db
@@ -46,15 +46,13 @@ class InMemoryDeploymentManager(AgentDeploymentManager):
id: str,
project_id: str,
config: Any,
- logs: List[Any],
):
- print("Creating deployment", config, logs)
+ print("Creating deployment", config)
deployment = await AgentDeployment.from_factory(
id,
get_agent_factory_from_template(config["templateID"]),
project_id,
config,
- logs,
)
try:
await db.create_deployment(deployment.id, project_id, config)
@@ -70,9 +68,8 @@ class InMemoryDeploymentManager(AgentDeploymentManager):
id: str,
project_id: str,
config: Any,
- logs: List[Any],
):
- print(">>>> Updating deployment", id, config, logs)
+ print(">>>> Updating deployment", id, config)
try:
await self.remove_deployment(id)
except:
@@ -82,7 +79,6 @@ class InMemoryDeploymentManager(AgentDeploymentManager):
id,
project_id,
config,
- logs,
)
async def remove_deployment(self, id: str):
diff --git a/api-service/deployment/manager.py b/api-service/deployment/manager.py
index 40e093db1..1294cb8d0 100644
--- a/api-service/deployment/manager.py
+++ b/api-service/deployment/manager.py
@@ -14,13 +14,14 @@ from agent.base import (
AgentBase,
AgentInteractionRequest,
OnLogs,
+ SetRun,
OnInteractionRequest,
GetEnvs,
)
-
AgentFactory = Callable[
- [Any, GetEnvs, OnLogs, OnInteractionRequest], Coroutine[None, None, AgentBase]
+ [Any, GetEnvs, SetRun, OnLogs, OnInteractionRequest],
+ Coroutine[None, None, AgentBase],
]
app_url = os.environ.get("APP_URL", "http://localhost:3000")
@@ -63,14 +64,25 @@ async def agent_run_cancelled(deployment_id: str):
class AgentEvents:
- def __init__(self, id: str, logs: List[Any]) -> None:
- self.id = id
- self.logs: List[Any] = logs
+ def __init__(self, deployment_id: str, project_id: str) -> None:
+ self.deployment_id = deployment_id
+ self.logs: List[Any] = []
self.interaction_requests: List[AgentInteractionRequest] = []
+ self.run_id: str | None
+
self.log_que = WorkQueue[List[Any]](
- lambda logs: db.update_deployment_logs(self.id, logs)
+ lambda logs: db.update_deployment_logs(
+ deployment_id=self.deployment_id,
+ run_id=self.run_id,
+ project_id=project_id,
+ logs=logs,
+ )
)
+ def set_run_id(self, run_id: str):
+ self.logs = []
+ self.run_id = run_id
+
async def add_log(self, log: Any):
date = datetime.now()
log["timestamp"] = str(date)
@@ -85,9 +97,11 @@ class AgentEvents:
self.interaction_requests.append(interaction_request)
match interaction_request.type:
case "done":
- await agent_run_done(self.id, interaction_request.data["prompt"])
+ await agent_run_done(
+ self.deployment_id, interaction_request.data["prompt"]
+ )
case "cancelled":
- await agent_run_cancelled(self.id)
+ await agent_run_cancelled(self.deployment_id)
# case "failed":
# # TODO: Save agent/interactions/states/run
# await agent_run_failed(self.id)
@@ -113,12 +127,12 @@ class AgentDeployment:
factory: AgentFactory,
project_id: str,
config: Any,
- logs: List[Any],
):
- event_handler = AgentEvents(deployment_id, logs)
+ event_handler = AgentEvents(deployment_id=deployment_id, project_id=project_id)
agent = await factory(
config,
lambda: db.get_env_vars(project_id),
+ event_handler.set_run_id,
event_handler.add_log,
event_handler.add_interaction_request,
)