The Agent Skills Directory

[COMMAND_EXECUTION]: The skill contains instructions to execute shell commands that incorporate a variable ($TARGET_URL) derived from user input. This pattern creates a command injection surface if the input is not strictly validated or escaped before execution.
Evidence: The command npx lighthouse $TARGET_URL --output=json in SKILL.md directly uses the variable in a shell context.
[INDIRECT_PROMPT_INJECTION]: The skill defines a workflow that ingests an external URL and uses it in subsequent tool invocations without instructing the agent to perform sanitization.
Ingestion points: The TARGET_URL defined in the audit target identification step in SKILL.md.
Boundary markers: None provided to delimit the untrusted URL from shell command arguments.
Capability inventory: Shell command execution via npx and npm, Node.js scripting with chrome-launcher, and local file system access for report generation.
Sanitization: No validation, filtering, or escaping logic is specified for the user-provided audit target.
[EXTERNAL_DOWNLOADS]: The skill instructs the use of npx to fetch and run several common performance auditing and visualization tools from the NPM registry.
Evidence: Workflow steps involving lighthouse, serve, vite-bundle-visualizer, nuxi, webpack-bundle-analyzer, and source-map-explorer.

profiling-performance