The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted repository data (roadmaps and plan files) to autonomously formulate goals and objectives for child tasks, creating a surface for indirect instructions to hijack the agent pipeline.
Ingestion points: The skill reads authoritative planning documents including AGENTS.md, ROADMAP.md, and various markdown files within the plans/ directory (e.g., plans/import.md).
Boundary markers: While the skill generates a structured 'Goal contract' as a routing envelope, it does not explicitly sanitize the natural language content of the cited plan files which are passed into the context of the spawned child task.
Capability inventory: The skill possesses significant capabilities including the ability to commit changes to the Git repository, create new autonomous threads or tasks (create_thread), and set native objectives for those tasks (create_goal).
Sanitization: There is no evidence of instruction filtering, escaping, or validation of the semantic content of the ingested plans before they are used to define the 'Outcome' and 'Included results' of the generated goal contract.

next-goal