Add pre_instruction and on_instruction hooks.

autogpt/agent/agent_manager.py

@@ -1,7 +1,7 @@
 """Agent manager for managing GPT agents"""
 from typing import List, Tuple, Union
 from autogpt.llm_utils import create_chat_completion
-from autogpt.config.config import Singleton
+from autogpt.config.config import Singleton, Config
 
 
 class AgentManager(metaclass=Singleton):
@@ -10,6 +10,7 @@ class AgentManager(metaclass=Singleton):
     def __init__(self):
         self.next_key = 0
         self.agents = {}  # key, (task, full_message_history, model)
+        self.cfg = Config()
 
     # Create new GPT agent
     # TODO: Centralise use of create_chat_completion() to globally enforce token limit
@@ -28,6 +29,10 @@ class AgentManager(metaclass=Singleton):
         messages = [
             {"role": "user", "content": prompt},
         ]
+        for plugin in self.cfg.plugins:
+            plugin_messages = plugin.pre_instruction(messages)
+            if plugin_messages:
+                messages.extend(plugin_messages)
 
         # Start GPT instance
         agent_reply = create_chat_completion(
@@ -35,9 +40,13 @@ class AgentManager(metaclass=Singleton):
             messages=messages,
         )
 
-        # Update full message history
-        messages.append({"role": "assistant", "content": agent_reply})
+        plugins_reply = agent_reply
+        for plugin in self.cfg.plugins:
+            plugin_result = plugin.on_instruction(messages)
+            if plugin_result:
+                plugins_reply = f"{plugins_reply}\n{plugin_result}"
 
+        messages.append({"role": "assistant", "content": plugins_reply})
         key = self.next_key
         # This is done instead of len(agents) to make keys unique even if agents
         # are deleted
@@ -45,7 +54,7 @@ class AgentManager(metaclass=Singleton):
 
         self.agents[key] = (task, messages, model)
 
-        return key, agent_reply
+        return key, plugins_reply
 
     def message_agent(self, key: Union[str, int], message: str) -> str:
         """Send a message to an agent and return its response
@@ -62,16 +71,26 @@ class AgentManager(metaclass=Singleton):
         # Add user message to message history before sending to agent
         messages.append({"role": "user", "content": message})
 
+        for plugin in self.cfg.plugins:
+            plugin_messages = plugin.pre_instruction(messages)
+            if plugin_messages:
+                messages.extend(plugin_messages)
+
         # Start GPT instance
         agent_reply = create_chat_completion(
             model=model,
             messages=messages,
         )
 
+        plugins_reply = agent_reply
+        for plugin in self.cfg.plugins:
+            plugin_result = plugin.on_instruction(messages)
+            if plugin_result:
+                plugins_reply = f"{plugins_reply}\n{plugin_result}"
         # Update full message history
-        messages.append({"role": "assistant", "content": agent_reply})
+        messages.append({"role": "assistant", "content": plugins_reply})
 
-        return agent_reply
+        return plugins_reply
 
     def list_agents(self) -> List[Tuple[Union[str, int], str]]:
         """Return a list of all agents