The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests and audits local source code files in the convex/ directory. This creates a potential surface where malicious instructions embedded in code comments could attempt to influence the agent's behavior, though this risk is mitigated by the skill's requirement for a deterministic, regex-based scan as the primary step.
Ingestion points: Project source files matching convex/**/*.ts.
Boundary markers: No specific markers are used to delimit code from instructions during ingestion.
Capability inventory: The skill can modify source files and execute the TypeScript compiler via shell commands.
Sanitization: There is no explicit sanitization of the ingested code content before the audit logic is applied.
[COMMAND_EXECUTION]: The skill uses npx tsc --noEmit to verify that the injected authorization logic is type-safe and does not introduce regressions. This is a standard and expected use of development tools for this type of skill.
[DYNAMIC_EXECUTION]: The skill dynamically generates a new utility file (convex/model/auth.ts) and rewrites existing handler functions to include authorization checks. These modifications are generated from local templates and specific audit hits rather than external sources.

convex-authz