The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill accesses sensitive authentication and configuration files at ~/.vibes/auth.json and ~/.vibes/.env to manage deployment tokens and API keys.
[CREDENTIALS_UNSAFE]: The OPENROUTER_API_KEY is passed as a command-line argument (--ai-key) to the deploy-cloudflare.js script, which can expose the secret to other users or processes on the system.
[COMMAND_EXECUTION]: The skill executes dynamic JavaScript code blocks using bun -e and runs various local scripts for server hosting, code assembly, and deployment.
[PROMPT_INJECTION]: Potential command injection vulnerability exists where the skill instructs the agent to interpolate user-provided prompts directly into shell commands (e.g., bun ... --prompt "USER_PROMPT_HERE") without sanitization.
[PROMPT_INJECTION]: Indirect prompt injection surface identified:
Ingestion points: User input appPrompt is collected via AskUserQuestion in SKILL.md and passed to a sub-agent.
Boundary markers: Absent; the user input is directly interpolated into the prompts/builder.md template.
Capability inventory: The builder sub-agent is granted broad tool access including Read, Write, Edit, and Bash to perform its generation tasks.
Sanitization: No validation or escaping is performed on the appPrompt before it is processed by the sub-agent.

launch