The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: As a complex orchestrator, the skill has a significant indirect prompt injection surface because it ingests untrusted data from repository artifacts (such as commit history and PR bodies) and external web content to drive subsequent agent tasks. The skill identifies ingestion points in files like reference/abide-recipe.md and reference/newsroom-recipe.md. To mitigate this risk, it employs context markers (_AGENT_CONTEXT, NEXUS_HANDOFF) and a mandatory SPECIFY phase designed to harden instructions and remove load-bearing ambiguity before delegating work to specialist agents.
[COMMAND_EXECUTION]: The skill executes shell commands to coordinate and monitor external agent CLI tools like agy and codex. This includes the use of sensitive flags such as --dangerously-skip-permissions to enable autonomous operation in headless environments, which the skill manages by requiring explicit user notification and pre-flight checks.
[DYNAMIC_EXECUTION]: The framework dynamically generates and executes multi-agent workflows based on real-time task classification and context analysis. It manages state and recovery across these sessions using a hierarchical guardrail system.

nexus