The Agent Skills Directory

[COMMAND_EXECUTION] (HIGH): The skill explicitly defines and executes destructive shell commands including rm -rf .local/ .cache/ .upm/ .breakpoints.
Evidence: Phase 4.1 in SKILL.md and the 'Execute Phase' in diagram.html show direct calls to rm -rf on common directory names.
Risk: Without strict path validation or scoping, there is a high risk of deleting critical user data if the agent's current working directory is misaligned.
[INDIRECT_PROMPT_INJECTION] (MEDIUM): The skill processes and modifies untrusted external content from an exported project, including source code, package.json, and vite.config.ts.
Ingestion Points: package.json, vite.config.ts, and all **/*.{ts,tsx,js,jsx} files.
Boundary Markers: Absent. The tool reads file contents directly for pattern matching and replacement.
Capability Inventory: Subprocess execution (rm, rm -rf) and file write operations.
Sanitization: Absent. The skill uses line-based filtering and JSON parsing but does not sanitize content against potential agent-steering instructions embedded in comments or metadata.
[DATA_EXPOSURE] (LOW): While the tool primarily deletes data, it scans the entire project structure for specific patterns, which involves reading sensitive configuration files like .replit and package.json.

ln-724-replit-cleaner