Strip !pip installs

vision_agent/agent/vision_agent_coder_prompts.py

@@ -230,7 +230,7 @@ def find_text(image_path: str, text: str) -> str:
 
 
 FIX_BUG = """
-**Role** As a coder, your job is to find the error in the code and fix it. You are running in a notebook setting so you can run !pip install to install missing packages.
+**Role** As a coder, your job is to find the error in the code and fix it. You are running in a notebook setting but do not run !pip install to install new packages.
 
 **Documentation**:
 This is the documentation for the functions you have access to. You may call any of these functions to help you complete the task. They are available through importing `from vision_agent.tools import *`.

vision_agent/agent/vision_agent_coder_prompts_v2.py

@@ -77,7 +77,7 @@
 
 
 FIX_BUG = """
-**Role**: As a coder, your job is to find the error in the code and fix it. You are running in a notebook setting so you can run !pip install to install missing packages.
+**Role** As a coder, your job is to find the error in the code and fix it. You are running in a notebook setting but do not run !pip install to install new packages.
 
 **Task**: A previous agent has written some code and some testing code according to a plan given to it. It has introduced a bug into it's code while trying to implement the plan. You are given the plan, code, test code and error. Your job is to fix the error in the code or test code.
 

vision_agent/agent/vision_agent_planner_prompts_v2.py

@@ -1,7 +1,7 @@
 PLAN = """
 **Role**: You are an expert planning agent that can understand the user request and search for a plan to accomplish it.
 
-**Task**: As a planning agent you are required to understand the user's request and search for a plan to accomplish it. Use Chain-of-Thought approach to break down the problem, create a plan, and then provide a response. Esnure your response is clear, concise, and helpful. You can use an interactive Pyton (Jupyter Notebok) environment, executing code with <execute_python>, each execution is a new cell so old code and outputs are saved.
+**Task**: As a planning agent you are required to understand the user's request and search for a plan to accomplish it. Use Chain-of-Thought approach to break down the problem, create a plan, and then provide a response. Esnure your response is clear, concise, and helpful. You can use an interactive Pyton (Jupyter Notebok) environment but do not !pip install packages, execute code with <execute_python>, each execution is a new cell so old code and outputs are saved.
 
 **Documentation**: this is the documentation for the functions you can use to accomplish the task:
 {tool_desc}

vision_agent/agent/vision_agent_planner_v2.py

@@ -21,6 +21,7 @@
     extract_tag,
     print_code,
     print_table,
+    remove_installs_from_code,
 )
 from vision_agent.agent.types import AgentMessage, InteractionContext, PlanContext
 from vision_agent.agent.vision_agent_planner_prompts_v2 import (
@@ -180,6 +181,7 @@ def run_critic(
 
 
 def code_safeguards(code: str) -> str:
+    code = remove_installs_from_code(code)
     if "get_tool_for_task" in code:
         lines = code.split("\n")
         new_lines = []