The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill performs scans on local environment files, specifically searching .env, .env.local, and .env.example. While it uses sed to redact the actual values of discovered keys to prevent them from entering the agent's context, the direct access to these sensitive file paths for auditing and modification is a security consideration.
[COMMAND_EXECUTION]: The skill instructs the agent to execute multiple shell commands including grep, mktemp, diff, and npm uninstall. These are used to discover patterns in the codebase, compare package versions, and modify the project's dependency tree. It also directs the agent to delete specific markdown files identified as scaffolding, which is a destructive action consistent with the skill's cleanup purpose.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface because it ingests data from untrusted project sources (including README files, source code, and package manifests) and uses this data to drive subsequent cleanup actions. The analysis identified the following evidence chain:
Ingestion points: Reads package.json, .env, index.html, and the src/ directory (SKILL.md).
Boundary markers: No explicit delimiters or instructions are used to distinguish between file content and agent instructions during these reads.
Capability inventory: Performs npm uninstall, grep, sed, and file deletions across various scripts.
Sanitization: Includes a redaction step for environment variable values using sed to prevent credential exposure.

lovable-cleanup