The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/rlm_repl.py uses the exec() function to run arbitrary Python code provided through command-line arguments or standard input. This is a powerful feature that allows the agent to execute any Python command in the local environment.
[DYNAMIC_EXECUTION]: The skill uses pickle.load() in scripts/rlm_repl.py to restore state from a local file (.claude/rlm_state/state.pkl). Because pickle is inherently insecure, if an attacker or a malicious process modifies this state file, it can lead to arbitrary code execution the next time the REPL is initialized or used.
[INDIRECT_PROMPT_INJECTION]: The skill's core purpose is to process large, untrusted context files, which represents a significant attack surface.
Ingestion points: The init command in rlm_repl.py reads a user-specified context file and stores its content in the persistent state.
Boundary markers: The skill does not suggest or implement specific delimiters or 'ignore' instructions to prevent the agent from obeying commands embedded within the large context files.
Capability inventory: The skill provides the agent with full Python execution (exec), directory creation, file writing (write_chunks), and generic tool access (Bash, Write, Edit) to process the data.
Sanitization: There is no evidence of sanitization or validation of the input data before it is loaded into the REPL environment or used by the agent and its subagents.

rlm