The Agent Skills Directory

[SAFE]: The skill defines a logical workflow for task delegation based on complexity, which is a standard operational practice for multi-agent systems.
[SAFE]: The instructions include explicit guardrails such as 'Do not call every agent by default' and 'Do not skip review for non-trivial code changes,' which promote the principle of least privilege and human-in-the-loop verification.
[SAFE]: The 'Standard workflow' mandates security and review gates (Phase 5: Verify), demonstrating a security-conscious design for handling significant tasks like architectural changes or issue fixes.
[INDIRECT_PROMPT_INJECTION]: While the skill processes user-supplied tasks which could contain malicious instructions, the workflow mitigation strategy includes:
Ingestion points: User tasks and request parameters (SKILL.md).
Boundary markers: None specified in the high-level workflow definition.
Capability inventory: Uses codebase_investigator, delegates to specialists, and can update GitHub/docs.
Sanitization: Explicitly includes QA, security, and review gates in the 'Standard workflow' to validate outputs before completion.

orchestration-workflow