The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill documents a pattern for dynamic tool creation where an LLM generates Python code based on a description, which is then written to a file and executed.
Evidence: The MCPAgent.create_tool method generates code using llm.generate, writes it to a file via f.write(code), and then dynamically loads it using connect_server.
[COMMAND_EXECUTION]: The skill provides patterns for executing shell commands, some of which use potentially unsafe methods.
Evidence: SandboxedExecution.execute_sandboxed utilizes subprocess.run with shell=True, which is prone to command injection if the command string is not properly sanitized.
Evidence: CheckpointManager._capture_workspace uses subprocess.getoutput to run git commands in the workspace directory.
[EXTERNAL_DOWNLOADS]: The skill implements a pattern for downloading content from arbitrary URLs to provide context to the agent.
Evidence: ContextManager.add_url uses the requests.get method to fetch data from remote locations.
[DATA_EXFILTRATION]: The skill includes patterns for reading local file and directory contents and injecting them into the agent's processing context.
Evidence: ContextManager.add_file and add_folder facilitate reading filesystem data into the self.context list, which could lead to exposure if transmitted to external services.

autonomous-agent-patterns