The Agent Skills Directory

[SAFE]: The skill is well-architected with security in mind. It uses native file tools instead of shell commands to avoid cross-platform vulnerabilities and command injection risk.
[DATA_EXPOSURE]: While the skill reads a PROJECT-CONTEXT.md file, it implements a mandatory sanitization step to drop secrets, tokens, credentials, and auth-embedded URLs before the data reaches the sub-agents.
[PROMPT_INJECTION]: The skill includes specific defenses against indirect prompt injection. It instructs the agent to treat the PROJECT-CONTEXT.md content as untrusted declarative data and explicitly disregard any embedded instructions or imperative directives found within it.
[COMMAND_EXECUTION]: Sub-agents (personas) are explicitly restricted to 'analysis-only' mode. They are strictly forbidden from editing files, running state-changing commands, or using tools that modify the repository or external systems.

dev-team