The Agent Skills Directory

[PROMPT_INJECTION]: The skill facilitates an indirect prompt injection surface by automatically extracting content from untrusted session logs to populate the agent's persistent memory. 1. Ingestion points: The session_parser.py script reads JSONL session logs from the .claude/projects directory. 2. Boundary markers: No protective delimiters or instructions to ignore embedded commands are present in the resulting MEMORY.md file. 3. Capability inventory: The skill has the capability to write to the agent's memory configuration (active_context.py) and manage a local database (search.py). 4. Sanitization: Information extraction is based on simple keyword matching in session_summary.py with no content validation, allowing potentially malicious instructions from chat logs to persist into the agent's system prompt.
[COMMAND_EXECUTION]: The skill utilizes a CLI entry point (context_manager.py) to run Python scripts that access and process sensitive chat history and project data stored within the user's home directory paths (e.g., C:\Users\renat\.claude\projects).

context-agent