The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The framework has a significant exposure to indirect prompt injection.
Ingestion points: Several recipes, such as newsroom (reference/newsroom-recipe.md), eureka (reference/eureka-recipe.md), and research-grounding (reference/research-grounding.md), use field, compete, and deep-research agents to perform WebSearch and WebFetch operations on untrusted external web content.
Boundary markers: The skill employs a 'Specified Brief' mechanism (reference/specify-phase.md) to harden instructions before delegation. However, this is primarily intended for clarity and intent preservation rather than sanitizing untrusted external data.
Capability inventory: The orchestrator can execute complex code changes and system modifications through specialized implementation agents like Builder, Artisan, and Native, which are then automatically verified by separate testing agents.
Sanitization: The skill relies on 'Adversarial Refutation' (_common/ADVERSARIAL_REFUTATION.md) and 'Skeptic Panels' to detect misinformation or logical errors in the gathered content, which is a judgment-based defense rather than a deterministic security control against hidden prompt instructions.
[COMMAND_EXECUTION]: The core functionality of the skill involves executing shell commands and spawning autonomous sub-agents across Claude Code, Codex, and Antigravity (agy) environments. This is documented in detail in reference/execution-layers.md and reference/orchestration-patterns.md, and is the intended purpose of the orchestration plane.
[EXTERNAL_DOWNLOADS]: The skill references and fetches information from established technology domains and documentation sites (e.g., code.claude.com, github.com) to maintain platform compatibility and verify standards. These represent well-known and trusted services used for legitimate administrative and informational purposes.

nexus