The Agent Skills Directory

[SAFE]: The skill is primarily instructional and does not contain executable code, scripts, or remote dependencies. It defines a conceptual framework for an orchestrator agent to delegate tasks to subagents.
[CREDENTIALS_UNSAFE]: The skill includes explicit safety instructions to prevent credential leakage. Specifically, the 'Orchestrator Responsibilities' section in SKILL.md mandates keeping secrets, credentials, and hidden reasoning out of subagent prompts.
[INDIRECT_PROMPT_INJECTION]: The skill acknowledges and mitigates the risk of indirect prompt injection by defining a strict integration workflow.
Ingestion points: The orchestrator reads and integrates output from subagents (documented in SKILL.md under 'Integration Workflow').
Boundary markers: The instructions recommend using 'bounded prompts' with explicit scopes and constraints for each subagent.
Capability inventory: The subagents are described as having capabilities for file inspection, editing, and research, though the specific tools are determined by the execution harness.
Sanitization: The 'Integration Workflow' in SKILL.md requires the orchestrator to 'Review each subagent result before trusting it' and 'Resolve contradictions' before applying changes.

subagents