The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data by reading and analyzing all .ts files within the convex/ directory. This creates a surface for indirect prompt injection where malicious instructions embedded in the project's source code could attempt to influence the agent's logic or the hardening process. Evidence:
Ingestion points: Scans convex/**/*.ts (SKILL.md, Step 1).
Boundary markers: Not explicitly used during file ingestion.
Capability inventory: Modifies files (Step 2) and executes npx tsc (Step 3).
Sanitization: Uses regex-based matching for detection, which provides a layer of deterministic behavior.
[COMMAND_EXECUTION]: The skill executes the npx tsc --noEmit command to verify that the hardening changes did not introduce type errors. This is a standard development workflow, but it involves executing a shell command that acts upon the current state of the project files.

convex-authz