The Agent Skills Directory

[SAFE]: The skill is designed with a strong security posture for subagent orchestration, emphasizing isolation and verification. Isolation is achieved through worker contracts that require an 'Explicit File Ownership Boundary' (whitelist), strictly prohibiting workers from modifying files outside their assigned scope. The skill includes specific failure taxonomies and decision rules to handle ownership collisions, semantic conflicts, and scope expansion, ensuring the parent agent maintains control over the workspace integration.
[INDIRECT_PROMPT_INJECTION]: The skill processes data from subagents, which represents an input surface for potential indirect prompt injection, however this is mitigated by strict verification protocols. * Ingestion points: Worker results are ingested via TOON-formatted return packets as defined in SKILL.md and templates/return-packet.toon. * Boundary markers: The protocol uses Token-Oriented Object Notation (TOON) with explicit length tags to ensure message integrity and structural enforcement. * Capability inventory: The parent agent performs integration tests and diff verification, while subagents are restricted to disjoint file whitelists. * Sanitization: The protocol mandates the inspection of concrete artifacts (diffs, command results) and the rejection of prose summaries as valid evidence, providing a defense against deceptive subagent behavior.

fable-delegate