fix(agent): Fix & improve agent self-termination and resumption mechanism

- Add `AgentFinished` exception (subclass of `AgentTerminated`)
- Raise `AgentFinished` instead of `AgentTerminated` in `finish` method
- Remove resumption patch from PR #6990 in `BaseAgent`
- Clean up implementation of `finish` in `AgentProtocolServer`
- Add resumption mechanism in `run_auto_gpt` (main.py)

autogpts/autogpt/autogpt/agents/base.py

@@ -177,14 +177,6 @@ class BaseAgent(Configurable[BaseAgentSettings], ABC):
         self.legacy_config = legacy_config
         """LEGACY: Monolithic application configuration."""
 
-        # In case the agent is resumed, cursor is set to the last episode
-        if self.event_history:
-            # To prevent errors, when the last action is "finish", we register a result
-            # And move cursor to the next action
-            if self.event_history.current_episode.action.name == "finish":
-                self.event_history.register_result(ActionSuccessResult())
-                self.event_history.cursor = len(self.event_history)
-
         self.llm_provider = llm_provider
 
         self.prompt_strategy = prompt_strategy

autogpts/autogpt/autogpt/agents/utils/exceptions.py

@@ -18,6 +18,10 @@ class AgentTerminated(AgentException):
     """The agent terminated or was terminated"""
 
 
+class AgentFinished(AgentTerminated):
+    """The agent self-terminated"""
+
+
 class ConfigurationError(AgentException):
     """Error caused by invalid, incompatible or otherwise incorrect configuration"""
 

autogpts/autogpt/autogpt/app/agent_protocol_server.py

@@ -30,6 +30,7 @@ from sentry_sdk import set_user
 from autogpt.agent_factory.configurators import configure_agent_with_state
 from autogpt.agent_factory.generators import generate_agent_for_task
 from autogpt.agent_manager import AgentManager
+from autogpt.agents.utils.exceptions import AgentFinished
 from autogpt.commands.system import finish
 from autogpt.commands.user_interaction import ask_user
 from autogpt.config import Config
@@ -230,26 +231,6 @@ class AgentProtocolServer:
                 task=task, step=step, relative_path=path
             )
 
-            if step.is_last and execute_command == finish.__name__:
-                assert execute_command_args
-
-                additional_output = {}
-                task_total_cost = agent.llm_provider.get_incurred_cost()
-                if task_total_cost > 0:
-                    additional_output["task_total_cost"] = task_total_cost
-                    logger.info(
-                        f"Total LLM cost for task {task_id}: "
-                        f"${round(task_total_cost, 2)}"
-                    )
-
-                step = await self.db.update_step(
-                    task_id=task_id,
-                    step_id=step.step_id,
-                    output=execute_command_args["reason"],
-                    additional_output=additional_output,
-                )
-                return step
-
             if execute_command == ask_user.__name__:  # HACK
                 execute_result = ActionSuccessResult(outputs=user_input)
                 agent.event_history.register_result(execute_result)
@@ -261,11 +242,31 @@ class AgentProtocolServer:
                     step_id=step.step_id,
                     status="running",
                 )
-                # Execute previously proposed action
-                execute_result = await agent.execute(
-                    command_name=execute_command,
-                    command_args=execute_command_args,
-                )
+
+                try:
+                    # Execute previously proposed action
+                    execute_result = await agent.execute(
+                        command_name=execute_command,
+                        command_args=execute_command_args,
+                    )
+                except AgentFinished:
+                    additional_output = {}
+                    task_total_cost = agent.llm_provider.get_incurred_cost()
+                    if task_total_cost > 0:
+                        additional_output["task_total_cost"] = task_total_cost
+                        logger.info(
+                            f"Total LLM cost for task {task_id}: "
+                            f"${round(task_total_cost, 2)}"
+                        )
+
+                    step = await self.db.update_step(
+                        task_id=task_id,
+                        step_id=step.step_id,
+                        output=execute_command_args["reason"],
+                        additional_output=additional_output,
+                    )
+                    await agent.save_state()
+                    return step
             else:
                 assert user_input
                 execute_result = await agent.execute(

autogpts/autogpt/autogpt/app/main.py

@@ -16,8 +16,6 @@ from typing import TYPE_CHECKING, Optional
 from colorama import Fore, Style
 from forge.sdk.db import AgentDB
 
-from autogpt.file_storage import FileStorageBackendName, get_storage
-
 if TYPE_CHECKING:
     from autogpt.agents.agent import Agent
 
@@ -26,6 +24,7 @@ from autogpt.agent_factory.profile_generator import generate_agent_profile_for_t
 from autogpt.agent_manager import AgentManager
 from autogpt.agents import AgentThoughts, CommandArgs, CommandName
 from autogpt.agents.utils.exceptions import AgentTerminated, InvalidAgentResponseError
+from autogpt.commands.system import finish
 from autogpt.config import (
     AIDirectives,
     AIProfile,
@@ -35,8 +34,10 @@ from autogpt.config import (
 )
 from autogpt.core.resource.model_providers.openai import OpenAIProvider
 from autogpt.core.runner.client_lib.utils import coroutine
+from autogpt.file_storage import FileStorageBackendName, get_storage
 from autogpt.logs.config import configure_chat_plugins, configure_logging
 from autogpt.logs.helpers import print_attribute, speak
+from autogpt.models.action_history import ActionInterruptedByHuman
 from autogpt.plugins import scan_plugins
 from scripts.install_plugin_deps import install_plugin_dependencies
 
@@ -217,6 +218,21 @@ async def run_auto_gpt(
             replace_directives=override_directives,
         )
 
+        if (
+            agent.event_history.current_episode
+            and agent.event_history.current_episode.action.name == finish.__name__
+            and not agent.event_history.current_episode.result
+        ):
+            # Agent was resumed after `finish` -> rewrite result of `finish` action
+            finish_reason = agent.event_history.current_episode.action.args["reason"]
+            print(f"Agent previously self-terminated; reason: '{finish_reason}'")
+            new_assignment = await clean_input(
+                config, "Please give a follow-up question or assignment:"
+            )
+            agent.event_history.register_result(
+                ActionInterruptedByHuman(feedback=new_assignment)
+            )
+
         # If any of these are specified as arguments,
         #  assume the user doesn't want to revise them
         if not any(

autogpts/autogpt/autogpt/commands/system.py

@@ -6,7 +6,7 @@ import logging
 from typing import TYPE_CHECKING
 
 from autogpt.agents.features.context import get_agent_context
-from autogpt.agents.utils.exceptions import AgentTerminated, InvalidArgumentError
+from autogpt.agents.utils.exceptions import AgentFinished, InvalidArgumentError
 from autogpt.command_decorator import command
 from autogpt.core.utils.json_schema import JSONSchema
 
@@ -44,7 +44,7 @@ def finish(reason: str, agent: Agent) -> None:
         A result string from create chat completion. A list of suggestions to
             improve the code.
     """
-    raise AgentTerminated(reason)
+    raise AgentFinished(reason)
 
 
 @command(