The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documents a data modeling and hook architecture where external or user-provided data is processed through ctx.input and ctx.result in references/data-hooks.md. This ingestion of untrusted data into hook logic creates a potential surface for indirect prompt injection, although the documentation promotes the use of Zod for structural validation and CEL for logical constraints.\n
Ingestion points: Data enters the context via ctx.input (payload) and ctx.previous/ctx.result (record state) within lifecycle hooks described in references/data-hooks.md.\n
Boundary markers: The skill relies on ObjectSchema and Zod definitions (referenced in references/_index.md) to define data boundaries, and CEL P tags for logical predicates.\n
Capability inventory: Hooks can be granted api.read, api.write, and log capabilities, allowing for database operations and diagnostic output, as detailed in references/data-hooks.md.\n
Sanitization: The skill provides structural sanitization through Zod schemas and the os validate CLI utility to check metadata integrity.\n- [DYNAMIC_EXECUTION]: The references/data-hooks.md file describes a mechanism for executing JavaScript code snippets (body) within a sandboxed QuickJS VM. While this is gated by a capability model and isolation, it represents a dynamic code execution path within the documented framework.

objectstack-data