More prompt engineering

autogpts/autogpt/autogpt/agents/agent.py

@@ -58,10 +58,7 @@ from .base import (
 from .features.agent_file_manager import FileManagerComponent
 from .features.context import AgentContext, ContextComponent
 from .features.watchdog import WatchdogComponent
-from .prompt_strategies.code_flow import (
-    CodeFlowAgentPromptStrategy,
-    CodeFlowAgentActionProposal,
-)
+from .prompt_strategies.code_flow import CodeFlowAgentPromptStrategy
 from .protocols import (
     AfterExecute,
     AfterParse,

autogpts/autogpt/autogpt/agents/prompt_strategies/code_flow.py

@@ -5,7 +5,7 @@ from logging import Logger
 from pydantic import BaseModel, Field
 
 from autogpt.agents.base import BaseAgentActionProposal
-from autogpt.agents.prompt_strategies.one_shot import OneShotAgentPromptConfiguration
+from autogpt.agents.prompt_strategies.one_shot import OneShotAgentPromptConfiguration, AssistantThoughts
 from autogpt.config.ai_directives import AIDirectives
 from autogpt.config.ai_profile import AIProfile
 from autogpt.core.configuration.schema import SystemConfiguration
@@ -25,22 +25,6 @@ from autogpt.utils.function.model import FunctionDef
 
 _RESPONSE_INTERFACE_NAME = "AssistantResponse"
 
-
-class AssistantThoughts(ModelWithSummary):
-    observations: str = Field(
-        ..., description="Relevant observations from your last action (if any)"
-    )
-    self_criticism: str = Field(..., description="Constructive self-criticism")
-    reasoning: str = Field(..., description="Reasoning towards a new plan")
-    plan: list[str] = Field(
-        ...,
-        description="Short list that conveys your long-term plan. Parallelize where possible.",
-    )
-
-    def summary(self) -> str:
-        return self.reasoning
-
-
 class CodeFlowAgentActionProposal(BaseModel):
     thoughts: AssistantThoughts
     immediate_plan: str = Field(
@@ -66,8 +50,8 @@ FINAL_INSTRUCTION: str = (
     # "Determine exactly one command to use next based on the given goals "
     # "and the progress you have made so far, "
     # "and respond using the JSON schema specified previously:"
-    "You have to give the answer in the from of JSON schema specified previously."
-    "For the `python_code` field, you have to write Python code to execute your plan as efficiently as possible."
+    "You have to give the answer in the from of JSON schema specified previously. "
+    "For the `python_code` field, you have to write Python code to execute your plan as efficiently as possible. "
     "Your code will be executed directly without any editing: "
     "if it doesn't work you will be held responsible. "
     "Use ONLY the listed available functions and built-in Python features. "
@@ -217,10 +201,6 @@ class CodeFlowAgentPromptStrategy(PromptStrategy):
             )
         )
         assistant_reply_dict = extract_dict_from_json(response.content)
-        self.logger.debug(
-            "Parsing object extracted from LLM response:\n"
-            f"{json.dumps(assistant_reply_dict, indent=4)}"
-        )
 
         parsed_response = CodeFlowAgentActionProposal.parse_obj(assistant_reply_dict)
         if not parsed_response.python_code:
@@ -262,6 +242,7 @@ class CodeFlowAgentPromptStrategy(PromptStrategy):
                 name="execute_code_flow",
                 arguments={
                     "python_code": code_validation.functionCode,
+                    "plan_text": parsed_response.immediate_plan,
                 },
             ),
         )

autogpts/autogpt/autogpt/agents/prompt_strategies/one_shot.py

@@ -31,14 +31,17 @@ _RESPONSE_INTERFACE_NAME = "AssistantResponse"
 
 
 class AssistantThoughts(ModelWithSummary):
+    past_action_summary: str = Field(
+        ..., description="Summary of the last action you took, if ther is none, you can leave it empty"
+    )
     observations: str = Field(
-        ..., description="Relevant observations from your last action (if any)"
+        ..., description="Relevant observations from your last actions (if any)"
     )
     text: str = Field(..., description="Thoughts")
     reasoning: str = Field(..., description="Reasoning behind the thoughts")
     self_criticism: str = Field(..., description="Constructive self-criticism")
     plan: list[str] = Field(
-        ..., description="Short list that conveys the long-term plan"
+        ..., description="Short list that conveys the long-term plan considering the progress on your task so far"
     )
     speak: str = Field(..., description="Summary of thoughts, to say to user")
 

autogpts/autogpt/autogpt/commands/execute_code_flow.py

@@ -33,9 +33,14 @@ class CodeFlowExecutionComponent(CommandProvider):
                 description="The Python code to execute",
                 required=True,
             ),
+            "plan_text": JSONSchema(
+                type=JSONSchema.Type.STRING,
+                description="The plan to written in a natural language",
+                required=False,
+            ),
         },
     )
-    async def execute_code_flow(self, python_code: str) -> str:
+    async def execute_code_flow(self, python_code: str, plan_text: str) -> str:
         """Execute the code flow.
 
         Args:
@@ -52,4 +57,4 @@ class CodeFlowExecutionComponent(CommandProvider):
         }
         exec(code, result)
         result = str(await result['exec_output'])
-        return result
+        return f"Execution Plan:\n{plan_text}\n\nExecution Output:\n{result}"

autogpts/autogpt/autogpt/components/event_history.py

@@ -38,7 +38,7 @@ class EventHistoryComponent(MessageProvider, AfterParse, AfterExecute, Generic[A
             self.max_tokens,
             self.count_tokens,
         ):
-            yield ChatMessage.system(f"## Progress on your Task so far\n\n{progress}")
+            yield ChatMessage.system(f"## Progress on your Task so far\nThis is the list of the steps that you have executed previously, use this as your consideration on considering the next action!\n{progress}")
 
     def after_parse(self, result: AP) -> None:
         self.event_history.register_action(result)

autogpts/autogpt/poetry.lock

@@ -4745,7 +4745,7 @@ files = [
 name = "ptyprocess"
 version = "0.7.0"
 description = "Run a subprocess in a pseudo terminal"
-optional = true
+optional = false
 python-versions = "*"
 files = [
     {file = "ptyprocess-0.7.0-py2.py3-none-any.whl", hash = "sha256:4b41f3967fce3af57cc7e94b888626c18bf37a083e3651ca8feeb66d492fef35"},
@@ -7320,4 +7320,4 @@ benchmark = ["agbenchmark"]
 [metadata]
 lock-version = "2.0"
 python-versions = "^3.10"
-content-hash = "2ce5b3a7793bdc98bfe0bbb94ae153bbfe60d5767d9e5490420cc000c2959609"
+content-hash = "e4f170809cd2c0c6daaa3130393870ff1873efde0b5b5860ad60c25d9a64afb1"

autogpts/autogpt/pyproject.toml

@@ -77,6 +77,7 @@ google-cloud-storage = "^2.13.0"
 psycopg2-binary = "^2.9.9"
 ruff = "^0.4.4"
 pyright = "^1.1.362"
+ptyprocess = "^0.7.0"
 
 [tool.poetry.extras]
 benchmark = ["agbenchmark"]