Release v0.4.7 (#5094)

Author: lc0rp

.gitignore

@@ -160,3 +160,10 @@ openai/
 
 # news
 CURRENT_BULLETIN.md
+
+# AgBenchmark
+agbenchmark/reports/
+
+# Nodejs
+package-lock.json
+package.json

BULLETIN.md

@@ -4,24 +4,18 @@
 📖 *User Guide*: https://docs.agpt.co.
 👩 *Contributors Wiki*: https://github.com/Significant-Gravitas/Auto-GPT/wiki/Contributing.
 
-# v0.4.6 RELEASE HIGHLIGHTS! 🚀
+# v0.4.7 RELEASE HIGHLIGHTS! 🚀
 # -----------------------------
-This release includes under-the-hood improvements and bug fixes, including better UTF-8
-special character support, workspace write access for sandboxed Python execution,
-more robust path resolution for config files and the workspace, and a full restructure
-of the Agent class, the "brain" of Auto-GPT, to make it more extensible.
+This release introduces initial REST API support, powered by e2b's agent 
+protocol SDK (https://github.com/e2b-dev/agent-protocol#sdk). 
 
-We have also released some documentation updates, including:
+It also includes improvements to prompt generation and support 
+for our new benchmarking tool, Auto-GPT-Benchmarks
+(https://github.com/Significant-Gravitas/Auto-GPT-Benchmarks).
 
-- *How to share your system logs*
-  Visit [docs/share-your-logs.md] to learn how to how to share logs with us
-  via a log analyzer graciously contributed by https://www.e2b.dev/
+We've also moved our documentation to Material Theme, at https://docs.agpt.co.
 
-- *Auto-GPT re-architecture documentation*
-  You can learn more about the inner-workings of the Auto-GPT re-architecture
-  released last cycle, via these links:
-  * [autogpt/core/README.md]
-  * [autogpt/core/ARCHITECTURE_NOTES.md]
+As usual, we've squashed a few bugs and made some under-the-hood improvements.
 
-Take a look at the Release Notes on Github for the full changelog!
+Take a look at the Release Notes on Github for the full changelog:
 https://github.com/Significant-Gravitas/Auto-GPT/releases.

autogpt/core/runner/cli_web_app/client/__init__.py renamed to agbenchmark/__init__.py

@@ -0,0 +1,54 @@
+import os
+import sys
+from pathlib import Path
+from typing import Tuple
+
+from autogpt.agents import Agent
+from autogpt.app.main import run_interaction_loop
+from autogpt.commands import COMMAND_CATEGORIES
+from autogpt.config import AIConfig, Config, ConfigBuilder
+from autogpt.memory.vector import get_memory
+from autogpt.models.command_registry import CommandRegistry
+from autogpt.prompts.prompt import DEFAULT_TRIGGERING_PROMPT
+from autogpt.workspace import Workspace
+
+PROJECT_DIR = Path().resolve()
+
+
+def run_specific_agent(task, continuous_mode=False) -> Tuple[str, int]:
+    agent = bootstrap_agent(task, continuous_mode)
+    run_interaction_loop(agent)
+
+
+def bootstrap_agent(task, continuous_mode) -> Agent:
+    config = ConfigBuilder.build_config_from_env(workdir=PROJECT_DIR)
+    config.debug_mode = True
+    config.continuous_mode = continuous_mode
+    config.temperature = 0
+    config.plain_output = True
+    command_registry = CommandRegistry.with_command_modules(COMMAND_CATEGORIES, config)
+    config.memory_backend = "no_memory"
+    config.workspace_path = Workspace.init_workspace_directory(config)
+    config.file_logger_path = Workspace.build_file_logger_path(config.workspace_path)
+    ai_config = AIConfig(
+        ai_name="Auto-GPT",
+        ai_role="a multi-purpose AI assistant.",
+        ai_goals=[task],
+    )
+    ai_config.command_registry = command_registry
+    return Agent(
+        memory=get_memory(config),
+        command_registry=command_registry,
+        ai_config=ai_config,
+        config=config,
+        triggering_prompt=DEFAULT_TRIGGERING_PROMPT,
+    )
+
+
+if __name__ == "__main__":
+    # The first argument is the script name itself, second is the task
+    if len(sys.argv) != 2:
+        print("Usage: python script.py <task>")
+        sys.exit(1)
+    task = sys.argv[1]
+    run_specific_agent(task, continuous_mode=True)

agbenchmark/benchmarks.py

@@ -0,0 +1,4 @@
+{
+  "workspace": "auto_gpt_workspace",
+  "entry_path": "agbenchmark.benchmarks"
+}

agbenchmark/config.json

@@ -0,0 +1,24 @@
+{
+    "TestBasicCodeGeneration": {
+        "difficulty": "basic",
+        "dependencies": [
+            "TestWriteFile"
+        ],
+        "data_path": "agbenchmark/challenges/code/d3"
+    },
+    "TestBasicMemory": {
+        "difficulty": "basic",
+        "data_path": "agbenchmark/challenges/memory/m1"
+    },
+    "TestReadFile": {
+        "difficulty": "basic",
+        "dependencies": [
+            "TestWriteFile"
+        ],
+        "data_path": "agbenchmark/challenges/interface/read_file"
+    },
+    "TestWriteFile": {
+        "dependencies": [],
+        "data_path": "agbenchmark/challenges/interface/write_file"
+    }
+}

agbenchmark/regression_tests.json

@@ -17,6 +17,7 @@
 from autogpt.llm.utils import count_string_tokens
 from autogpt.logs import logger
 from autogpt.logs.log_cycle import (
+    CURRENT_CONTEXT_FILE_NAME,
     FULL_MESSAGE_HISTORY_FILE_NAME,
     NEXT_ACTION_FILE_NAME,
     USER_INPUT_FILE_NAME,
@@ -109,6 +110,13 @@ def on_before_think(self, *args, **kwargs) -> ChatSequence:
             self.history.raw(),
             FULL_MESSAGE_HISTORY_FILE_NAME,
         )
+        self.log_cycle_handler.log_cycle(
+            self.ai_config.ai_name,
+            self.created_at,
+            self.cycle_count,
+            prompt.raw(),
+            CURRENT_CONTEXT_FILE_NAME,
+        )
         return prompt
 
     def execute(
@@ -285,10 +293,10 @@ def execute_command(
         # Handle non-native commands (e.g. from plugins)
         for command in agent.ai_config.prompt_generator.commands:
             if (
-                command_name == command["label"].lower()
-                or command_name == command["name"].lower()
+                command_name == command.label.lower()
+                or command_name == command.name.lower()
             ):
-                return command["function"](**arguments)
+                return command.function(**arguments)
 
         raise RuntimeError(
             f"Cannot execute '{command_name}': unknown command."

autogpt/agents/agent.py

@@ -1,7 +1,8 @@
 from __future__ import annotations
 
+import re
 from abc import ABCMeta, abstractmethod
-from typing import TYPE_CHECKING, Any, Optional
+from typing import TYPE_CHECKING, Any, Literal, Optional
 
 if TYPE_CHECKING:
     from autogpt.config import AIConfig, Config
@@ -23,6 +24,8 @@
 class BaseAgent(metaclass=ABCMeta):
     """Base class for all Auto-GPT agents."""
 
+    ThoughtProcessID = Literal["one-shot"]
+
     def __init__(
         self,
         ai_config: AIConfig,
@@ -91,6 +94,7 @@ def __init__(
     def think(
         self,
         instruction: Optional[str] = None,
+        thought_process_id: ThoughtProcessID = "one-shot",
     ) -> tuple[CommandName | None, CommandArgs | None, AgentThoughts]:
         """Runs the agent for one cycle.
 
@@ -103,9 +107,8 @@ def think(
 
         instruction = instruction or self.default_cycle_instruction
 
-        prompt: ChatSequence = self.construct_prompt(instruction)
-        prompt = self.on_before_think(prompt, instruction)
-
+        prompt: ChatSequence = self.construct_prompt(instruction, thought_process_id)
+        prompt = self.on_before_think(prompt, thought_process_id, instruction)
         raw_response = create_chat_completion(
             prompt,
             self.config,
@@ -115,7 +118,7 @@ def think(
         )
         self.cycle_count += 1
 
-        return self.on_response(raw_response, prompt, instruction)
+        return self.on_response(raw_response, thought_process_id, prompt, instruction)
 
     @abstractmethod
     def execute(
@@ -138,6 +141,7 @@ def execute(
 
     def construct_base_prompt(
         self,
+        thought_process_id: ThoughtProcessID,
         prepend_messages: list[Message] = [],
         append_messages: list[Message] = [],
         reserve_tokens: int = 0,
@@ -179,7 +183,11 @@ def construct_base_prompt(
 
         return prompt
 
-    def construct_prompt(self, cycle_instruction: str) -> ChatSequence:
+    def construct_prompt(
+        self,
+        cycle_instruction: str,
+        thought_process_id: ThoughtProcessID,
+    ) -> ChatSequence:
         """Constructs and returns a prompt with the following structure:
         1. System prompt
         2. Message history of the agent, truncated & prepended with running summary as needed
@@ -196,14 +204,86 @@ def construct_prompt(self, cycle_instruction: str) -> ChatSequence:
         cycle_instruction_tlength = count_message_tokens(
             cycle_instruction_msg, self.llm.name
         )
-        prompt = self.construct_base_prompt(reserve_tokens=cycle_instruction_tlength)
+
+        append_messages: list[Message] = []
+
+        response_format_instr = self.response_format_instruction(thought_process_id)
+        if response_format_instr:
+            append_messages.append(Message("system", response_format_instr))
+
+        prompt = self.construct_base_prompt(
+            thought_process_id,
+            append_messages=append_messages,
+            reserve_tokens=cycle_instruction_tlength,
+        )
 
         # ADD user input message ("triggering prompt")
         prompt.append(cycle_instruction_msg)
 
         return prompt
 
-    def on_before_think(self, prompt: ChatSequence, instruction: str) -> ChatSequence:
+    # This can be expanded to support multiple types of (inter)actions within an agent
+    def response_format_instruction(self, thought_process_id: ThoughtProcessID) -> str:
+        if thought_process_id != "one-shot":
+            raise NotImplementedError(f"Unknown thought process '{thought_process_id}'")
+
+        RESPONSE_FORMAT_WITH_COMMAND = """```ts
+        interface Response {
+            thoughts: {
+                // Thoughts
+                text: string;
+                reasoning: string;
+                // Short markdown-style bullet list that conveys the long-term plan
+                plan: string;
+                // Constructive self-criticism
+                criticism: string;
+                // Summary of thoughts to say to the user
+                speak: string;
+            };
+            command: {
+                name: string;
+                args: Record<string, any>;
+            };
+        }
+        ```"""
+
+        RESPONSE_FORMAT_WITHOUT_COMMAND = """```ts
+        interface Response {
+            thoughts: {
+                // Thoughts
+                text: string;
+                reasoning: string;
+                // Short markdown-style bullet list that conveys the long-term plan
+                plan: string;
+                // Constructive self-criticism
+                criticism: string;
+                // Summary of thoughts to say to the user
+                speak: string;
+            };
+        }
+        ```"""
+
+        response_format = re.sub(
+            r"\n\s+",
+            "\n",
+            RESPONSE_FORMAT_WITHOUT_COMMAND
+            if self.config.openai_functions
+            else RESPONSE_FORMAT_WITH_COMMAND,
+        )
+
+        use_functions = self.config.openai_functions and self.command_registry.commands
+        return (
+            f"Respond strictly with JSON{', and also specify a command to use through a function_call' if use_functions else ''}. "
+            "The JSON should be compatible with the TypeScript type `Response` from the following:\n"
+            f"{response_format}\n"
+        )
+
+    def on_before_think(
+        self,
+        prompt: ChatSequence,
+        thought_process_id: ThoughtProcessID,
+        instruction: str,
+    ) -> ChatSequence:
         """Called after constructing the prompt but before executing it.
 
         Calls the `on_planning` hook of any enabled and capable plugins, adding their
@@ -238,7 +318,11 @@ def on_before_think(self, prompt: ChatSequence, instruction: str) -> ChatSequenc
         return prompt
 
     def on_response(
-        self, llm_response: ChatModelResponse, prompt: ChatSequence, instruction: str
+        self,
+        llm_response: ChatModelResponse,
+        thought_process_id: ThoughtProcessID,
+        prompt: ChatSequence,
+        instruction: str,
     ) -> tuple[CommandName | None, CommandArgs | None, AgentThoughts]:
         """Called upon receiving a response from the chat model.
 
@@ -261,7 +345,9 @@ def on_response(
         )  # FIXME: support function calls
 
         try:
-            return self.parse_and_process_response(llm_response, prompt, instruction)
+            return self.parse_and_process_response(
+                llm_response, thought_process_id, prompt, instruction
+            )
         except SyntaxError as e:
             logger.error(f"Response could not be parsed: {e}")
             # TODO: tune this message
@@ -276,7 +362,11 @@ def on_response(
 
     @abstractmethod
     def parse_and_process_response(
-        self, llm_response: ChatModelResponse, prompt: ChatSequence, instruction: str
+        self,
+        llm_response: ChatModelResponse,
+        thought_process_id: ThoughtProcessID,
+        prompt: ChatSequence,
+        instruction: str,
     ) -> tuple[CommandName | None, CommandArgs | None, AgentThoughts]:
         """Validate, parse & process the LLM's response.