The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill's documentation in SKILL.md includes a pre-written prompt for the AI agent that executes the command npx -y @alipay/agent-payment@latest install. This command fetches and executes code from the NPM registry at runtime. The package is associated with a well-known service provider.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user data in the form of invoice tables, which could potentially contain malicious instructions targeting the LLM.
Ingestion points: User-supplied text provided via the --input flag or a JSON text field, which is read and parsed by scripts/run.mjs.
Boundary markers: None identified; the system relies on the structure of the TSV/table to separate fields.
Capability inventory: The skill has access to Bash and Read tools, and its documentation explicitly instructs the agent to perform shell-based package installations.
Sanitization: The engine in scripts/engine/tax-invoice-void-check.js performs numeric validation and regex-based cleaning for amount fields, but descriptive fields like 'reason' are processed as raw strings.
[DYNAMIC_EXECUTION]: The wrapper script scripts/run.mjs utilizes createRequire and path.join to dynamically load the local engine script (tax-invoice-void-check.js) at runtime. While the path is restricted to the skill's own directory, this represents dynamic module loading.

tax-invoice-void-check-free