The Agent Skills Directory

Local Command Execution: The skill uses the git CLI via execFileSync to cross-reference session logs with commit history. This is restricted to local repository metadata and is essential for the report's functionality.
Indirect Prompt Injection Defense: The instructions explicitly warn the model that project names, which are sourced from local directory names, must be treated as inert data. This prevents potential injection attempts where a user or attacker might name a directory with strings like "ignore previous instructions" to manipulate the report output.
CSV Injection Prevention: The provided Node.js script includes logic to neutralize spreadsheet formulas in the exported CSV data by prefixing sensitive characters with an apostrophe, protecting users when opening the report in applications like Excel or Google Sheets.
Secure File Handling: The mining script implements defensive measures when writing to temporary directories, such as using the wx file flag and checking for symbolic links to prevent potential race conditions or symlink attacks.
Local Data Isolation: All data mining is performed offline on the local filesystem (~/.claude/projects) and local git history. The core processing script makes no network calls, ensuring that sensitive transcript data remains on the user's machine.

receipts