The Agent Skills Directory

[DATA_EXFILTRATION]: The skill processes sensitive local session history files located in ~/.claude/projects/. It extracts conversation data, including task descriptions, assistant responses, and user-provided summaries, which are then transmitted to an external service (Google Gemini) via the llm CLI tool to generate semantic embeddings for indexing.
[COMMAND_EXECUTION]: The skill relies on executing external system commands. It calls the llm CLI to generate embeddings and uses os.execvp to replace the current process with the claude CLI when resuming a session. These calls use argument lists, which correctly mitigates common command injection risks.
[INDIRECT_PROMPT_INJECTION]: The skill indexes data from user-controlled session logs which can contain untrusted content.
Ingestion points: Reads JSONL session logs from ~/.claude/projects/.
Boundary markers: None; content is extracted and stored directly in a SQLite database.
Capability inventory: Executes system commands via subprocess.run and os.execvp.
Sanitization: Performs basic whitespace normalization and truncation but does not filter for embedded instructions that could influence the agent when search results are retrieved.

session-finder