The Agent Skills Directory

[DATA_EXFILTRATION]: The skill accesses the agent's internal history log and memory storage located at ~/.claude/history.jsonl and ~/.claude/projects/*/memory/*.md. This behavior is transparently documented and serves the primary purpose of cross-session information retrieval.
[COMMAND_EXECUTION]: The skill utilizes standard Unix utilities such as grep, ls, and head to perform searches and summarize files. It provides specific command templates for the agent to follow when executing these searches.
[PROMPT_INJECTION]: The skill exhibits a surface for indirect prompt injection by processing potentially untrusted data from past sessions or project files.
Ingestion points: Reads from memory summaries, history logs, and local project files (e.g., .py, .ts, .md) as defined in SKILL.md.
Boundary markers: Employs a structured output format with clear sections for 'Source' and 'Snippet' to isolate retrieved content from the agent's instructional context.
Capability inventory: Limited to read-only shell commands (grep, ls, head) for data retrieval; no code execution or file modification tools are utilized by this skill.
Sanitization: No explicit sanitization or escaping of the content found within search snippets is described.

hermes-search