The Agent Skills Directory

[COMMAND_EXECUTION]: The skill performs file deletions using the Bash tool with a command constructed via simple string interpolation: rm '{path}'. This implementation is vulnerable to command injection if a filename contains a single quote (e.g., '; [command]; '). An attacker or malicious process could potentially execute arbitrary code by placing a file with a crafted name into the targeted memory directories.
[COMMAND_EXECUTION]: The skill automates destructive file operations (rm) across multiple directories, including .claude/agent-memory/ and .claude/projects/. While this is intended for memory maintenance, the automated nature and lack of path validation increase the risk of accidental or malicious data loss.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it processes untrusted content from various memory files that may have been populated with data from external sources in previous sessions.
Ingestion points: The skill uses Glob and Read to process all .md files within the .claude/ memory hierarchy.
Boundary markers: None are employed; the skill parses the full body of the files for word-overlap and contradiction detection.
Capability inventory: The skill has access to Bash (used for rm), Write (used for updating the MEMORY.md index), and Read/Glob tools.
Sanitization: No sanitization, escaping, or path validation is performed on the data extracted from the files before it is used in shell commands.

dream