The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill manages an attack surface where untrusted data (such as system requirements, specifications, or code feedback) is ingested via the input_data field in the Task dataclass found in reference/task-management.md and SKILL.md.
Ingestion points: Untrusted external data enters the agent context through the input_data and metadata fields of the Task objects.
Boundary markers: There are no explicit boundary markers or delimiters shown in the provided examples to isolate untrusted data from agent instructions.
Capability inventory: The orchestrator manages agents with capabilities for code implementation, architecture design, and security reviews as defined in reference/agent-taxonomy.md.
Sanitization: No explicit sanitization or validation logic for external content is included in the task processing logic.
[DATA_EXFILTRATION]: No patterns of sensitive data harvesting or unauthorized network transmission were detected. The workflow orchestration is entirely logical and structural.
[COMMAND_EXECUTION]: The skill focuses on workflow management and coordination using Python objects and does not utilize dangerous shell commands or system-level calls.
[REMOTE_CODE_EXECUTION]: The skill does not download or execute scripts from remote sources. All logic is contained within the provided reference files.

Subagent-Driven Development