The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user input in the form of problem statements and strategic questions.
Ingestion points: User-provided strings are ingested via the problem_statement and task fields in SKILL.md and reference/collaboration-packets.md.
Boundary markers: The system uses structured prompt skeletons with clear headers (e.g., # Role, # Target, # Constraints) to delimit instructions from data.
Capability inventory: The skill can invoke sub-agents via platform CLI tools (codex exec, agy) and generate structured output (JSON/Markdown).
Sanitization: The skill relies on sub-agent grounding steps and structured JSON schema validation to prevent malformed or malicious data from affecting downstream processes.
[COMMAND_EXECUTION]: The skill utilizes platform-specific CLI tools for its multi-engine functionality (multi recipe).
Evidence: reference/tri-engine-reframe.md describes executing codex exec --full-auto and agy -p with specific flags. These are recognized platform utilities for multi-LLM orchestration and do not involve arbitrary shell injection from untrusted sources.
[SAFE]: The skill explicitly forbids the generation of implementation code in its Boundaries section in SKILL.md, mitigating risks associated with automated code creation.

flux