Merge branch 'Significant-Gravitas:master' into a-slip-of-a-pen

agbenchmark/benchmarks.py

@@ -1,3 +1,4 @@
+import os
 import sys
 from pathlib import Path
 from typing import Tuple
@@ -25,7 +26,7 @@ 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.init_workspace_directory(config)
     config.file_logger_path = Workspace.build_file_logger_path(config.workspace_path)
@@ -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/main.py

@@ -134,36 +134,9 @@ def run_auto_gpt(
     config.file_logger_path = Workspace.build_file_logger_path(config.workspace_path)
 
     config.plugins = scan_plugins(config, config.debug_mode)
-    # Create a CommandRegistry instance and scan default folder
-    command_registry = CommandRegistry()
-
-    logger.debug(
-        f"The following command categories are disabled: {config.disabled_command_categories}"
-    )
-    enabled_command_categories = [
-        x for x in COMMAND_CATEGORIES if x not in config.disabled_command_categories
-    ]
 
-    logger.debug(
-        f"The following command categories are enabled: {enabled_command_categories}"
-    )
-
-    for command_category in enabled_command_categories:
-        command_registry.import_commands(command_category)
-
-    # Unregister commands that are incompatible with the current config
-    incompatible_commands = []
-    for command in command_registry.commands.values():
-        if callable(command.enabled) and not command.enabled(config):
-            command.enabled = False
-            incompatible_commands.append(command)
-
-    for command in incompatible_commands:
-        command_registry.unregister(command)
-        logger.debug(
-            f"Unregistering incompatible command: {command.name}, "
-            f"reason - {command.disabled_reason or 'Disabled by current config.'}"
-        )
+    # Create a CommandRegistry instance and scan default folder
+    command_registry = CommandRegistry.with_command_modules(COMMAND_CATEGORIES, config)
 
     ai_config = construct_main_ai_config(
         config,

autogpt/app/setup.py

@@ -83,6 +83,7 @@ def prompt_user(
                 "Falling back to manual mode.",
                 speak_text=True,
             )
+            logger.debug(f"Error during AIConfig generation: {e}")
 
             return generate_aiconfig_manual(config)
 

autogpt/commands/__init__.py

@@ -3,5 +3,5 @@
     "autogpt.commands.file_operations",
     "autogpt.commands.web_search",
     "autogpt.commands.web_selenium",
-    "autogpt.commands.task_statuses",
+    "autogpt.commands.system",
 ]

autogpt/commands/execute_code.py

@@ -1,4 +1,8 @@
-"""Execute code in a Docker container"""
+"""Commands to execute code"""
+
+COMMAND_CATEGORY = "execute_code"
+COMMAND_CATEGORY_TITLE = "Execute Code"
+
 import os
 import subprocess
 from pathlib import Path
@@ -251,9 +255,9 @@ def execute_shell(command_line: str, agent: Agent) -> str:
     "execute_shell_popen",
     "Executes a Shell Command, non-interactive commands only",
     {
-        "query": {
+        "command_line": {
             "type": "string",
-            "description": "The search query",
+            "description": "The command line to execute",
             "required": True,
         }
     },

autogpt/commands/file_operations.py

@@ -1,6 +1,10 @@
-"""File operations for AutoGPT"""
+"""Commands to perform operations on files"""
+
 from __future__ import annotations
 
+COMMAND_CATEGORY = "file_operations"
+COMMAND_CATEGORY_TITLE = "File Operations"
+
 import contextlib
 import hashlib
 import os
@@ -228,22 +232,6 @@ def write_to_file(filename: str, text: str, agent: Agent) -> str:
         return f"Error: {err}"
 
 
-@command(
-    "append_to_file",
-    "Appends to a file",
-    {
-        "filename": {
-            "type": "string",
-            "description": "The name of the file to write to",
-            "required": True,
-        },
-        "text": {
-            "type": "string",
-            "description": "The text to write to the file",
-            "required": True,
-        },
-    },
-)
 @sanitize_path_arg("filename")
 def append_to_file(
     filename: str, text: str, agent: Agent, should_log: bool = True

autogpt/commands/git_operations.py

@@ -1,4 +1,7 @@
-"""Git operations for autogpt"""
+"""Commands to perform Git operations"""
+
+COMMAND_CATEGORY = "git_operations"
+COMMAND_CATEGORY_TITLE = "Git Operations"
 
 from git.repo import Repo
 

autogpt/commands/image_gen.py

@@ -1,4 +1,8 @@
-""" Image Generation Module for AutoGPT."""
+"""Commands to generate images based on text input"""
+
+COMMAND_CATEGORY = "text_to_image"
+COMMAND_CATEGORY_TITLE = "Text to Image"
+
 import io
 import json
 import time

autogpt/commands/task_statuses.py → autogpt/commands/system.py

@@ -1,6 +1,10 @@
-"""Task Statuses module."""
+"""Commands to control the internal state of the program"""
+
 from __future__ import annotations
 
+COMMAND_CATEGORY = "system"
+COMMAND_CATEGORY_TITLE = "System"
+
 from typing import NoReturn
 
 from autogpt.agents.agent import Agent

autogpt/commands/web_search.py

@@ -1,6 +1,10 @@
-"""Google search command for Autogpt."""
+"""Commands to search the web with"""
+
 from __future__ import annotations
 
+COMMAND_CATEGORY = "web_search"
+COMMAND_CATEGORY_TITLE = "Web Search"
+
 import json
 import time
 from itertools import islice