The Agent Skills Directory

[DATA_EXFLLTRATION]: The skill performs discovery and health checks by reading sensitive application configuration files and databases located in the user's home directory. Evidence includes access to ~/.claude/settings.json, ~/.codex/config.toml, ~/.claude/plugins/installed_plugins.json, and the local SQLite database at ~/.claude-mem/claude-mem.db.
[COMMAND_EXECUTION]: The skill executes bundled diagnostic scripts and standard CLI tools to perform its tasks. Evidence includes the execution of python3 <skill-root>/scripts/memory_preflight.py and the use of rg (ripgrep) for searching the codebase.
[INDIRECT_PROMPT_INJECTION]: The skill processes historical session data (memory) from external sources, which creates a surface for indirect prompt injection if those logs contain malicious instructions from previous untrusted contexts.
Ingestion points: MEMORY.md file, claude-mem.db SQLite database, and results from MCP search tools.
Boundary markers: The instructions include explicit directives to label claims as confirmed, stale, or contradicted and to verify retrieved facts against current source code and primary documentation.
Capability inventory: Subprocess execution of bundled Python scripts, file system searching via rg, and standard file editing/writing for engineering tasks.
Sanitization: Instructions mandate that the agent treat recalled information as hypotheses until verified against authoritative local state.

gpt-engineer-mem