Merge branch 'master' into agbenchmark-tuto

.env.template

@@ -16,13 +16,13 @@ OPENAI_API_KEY=your-openai-api-key
 ## USER_AGENT - Define the user-agent used by the requests library to browse website (string)
 # USER_AGENT="Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_4) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/83.0.4103.97 Safari/537.36"
 
-## AI_SETTINGS_FILE - Specifies which AI Settings file to use (defaults to ai_settings.yaml)
+## AI_SETTINGS_FILE - Specifies which AI Settings file to use, relative to the Auto-GPT root directory. (defaults to ai_settings.yaml)
 # AI_SETTINGS_FILE=ai_settings.yaml
 
-## PLUGINS_CONFIG_FILE - The path to the plugins_config.yaml file (Default plugins_config.yaml)
+## PLUGINS_CONFIG_FILE - The path to the plugins_config.yaml file, relative to the Auto-GPT root directory. (Default plugins_config.yaml)
 # PLUGINS_CONFIG_FILE=plugins_config.yaml
 
-## PROMPT_SETTINGS_FILE - Specifies which Prompt Settings file to use (defaults to prompt_settings.yaml)
+## PROMPT_SETTINGS_FILE - Specifies which Prompt Settings file to use, relative to the Auto-GPT root directory. (defaults to prompt_settings.yaml)
 # PROMPT_SETTINGS_FILE=prompt_settings.yaml
 
 ## OPENAI_API_BASE_URL - Custom url for the OpenAI API, useful for connecting to custom backends. No effect if USE_AZURE is true, leave blank to keep the default url
@@ -58,7 +58,7 @@ OPENAI_API_KEY=your-openai-api-key
 ## USE_AZURE - Use Azure OpenAI or not (Default: False)
 # USE_AZURE=False
 
-## AZURE_CONFIG_FILE - The path to the azure.yaml file (Default: azure.yaml)
+## AZURE_CONFIG_FILE - The path to the azure.yaml file, relative to the Auto-GPT root directory. (Default: azure.yaml)
 # AZURE_CONFIG_FILE=azure.yaml
 
 

BULLETIN.md

@@ -4,23 +4,24 @@
 📖 *User Guide*: https://docs.agpt.co.
 👩 *Contributors Wiki*: https://github.com/Significant-Gravitas/Auto-GPT/wiki/Contributing.
 
-# v0.4.5 RELEASE HIGHLIGHTS! 🚀
+# v0.4.6 RELEASE HIGHLIGHTS! 🚀
 # -----------------------------
-This release includes under-the-hood improvements and bug fixes, such as more 
-accurate token counts for OpenAI functions, faster CI builds, improved plugin 
-handling, and refactoring of the Config class for better maintainability.
+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.
 
 We have also released some documentation updates, including:
 
 - *How to share your system logs*
-  Visit [docs/share-your-logs.md] to learn how to how to share logs with us 
+  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/
 
 - *Auto-GPT re-architecture documentation*
-  You can learn more about the inner-workings of the Auto-GPT re-architecture 
+  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]
 
-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.

agbenchmark/benchmarks.py

@@ -1,3 +1,4 @@
+import os
 import sys
 from pathlib import Path
 from typing import Tuple
@@ -25,9 +26,9 @@ def bootstrap_agent(task, continuous_mode) -> Agent:
     config.continuous_mode = continuous_mode
     config.temperature = 0
     config.plain_output = True
-    command_registry = get_command_registry(config)
+    command_registry = CommandRegistry.with_command_modules(COMMAND_CATEGORIES, config)
     config.memory_backend = "no_memory"
-    config.workspace_path = Workspace.set_workspace_directory(config)
+    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",
@@ -44,16 +45,6 @@ def bootstrap_agent(task, continuous_mode) -> Agent:
     )
 
 
-def get_command_registry(config: Config):
-    command_registry = CommandRegistry()
-    enabled_command_categories = [
-        x for x in COMMAND_CATEGORIES if x not in config.disabled_command_categories
-    ]
-    for command_category in enabled_command_categories:
-        command_registry.import_commands(command_category)
-    return command_registry
-
-
 if __name__ == "__main__":
     # The first argument is the script name itself, second is the task
     if len(sys.argv) != 2:

autogpt/agents/agent.py

@@ -293,10 +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/base.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.
 
@@ -103,8 +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,
@@ -114,7 +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(
@@ -137,6 +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,
@@ -178,7 +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
@@ -195,14 +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
@@ -237,7 +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.
 
@@ -260,7 +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
@@ -275,7 +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.
 

autogpt/app/cli.py

@@ -16,7 +16,10 @@
 @click.option(
     "--ai-settings",
     "-C",
-    help="Specifies which ai_settings.yaml file to use, will also automatically skip the re-prompt.",
+    help=(
+        "Specifies which ai_settings.yaml file to use, relative to the Auto-GPT"
+        " root directory. Will also automatically skip the re-prompt."
+    ),
 )
 @click.option(
     "--prompt-settings",
@@ -129,7 +132,9 @@ def main(
             browser_name=browser_name,
             allow_downloads=allow_downloads,
             skip_news=skip_news,
-            working_directory=Path(__file__).parent.parent,  # TODO: make this an option
+            working_directory=Path(
+                __file__
+            ).parent.parent.parent,  # TODO: make this an option
             workspace_directory=workspace_directory,
             install_plugin_deps=install_plugin_deps,
             ai_name=ai_name,