The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes npx tsc --noEmit to verify that the applied security fixes are type-safe. This is a standard development command used within the context of the project being audited.
[INDIRECT_PROMPT_INJECTION]: The skill ingests and processes untrusted data in the form of the user's source code (convex/**/*.ts). A malicious project could potentially include instructions within code comments or strings designed to influence the agent's logic, though the use of deterministic regex-based scanning significantly mitigates this surface.
Ingestion points: Files located in the convex/ directory, excluding generated and definition files.
Boundary markers: The skill does not employ explicit boundary markers when interpolating code content into its context.
Capability inventory: The agent performs file system reads, writes (during the hardening phase), and shell command execution (npx tsc).
Sanitization: The skill relies on deterministic regex matching (/\bctx\.auth\b/, etc.) rather than purely natural language interpretation for the initial audit, which provides a layer of sanitization against basic prompt injection attempts within code.

convex-authz