The Agent Skills Directory

[DATA_EXPOSURE_AND_EXFILTRATION]: The skill operates on local documentation files within the project structure (specifically docs/planning/). It does not attempt to access sensitive system directories, credentials, or environment variables. No network-enabled tools or libraries are invoked to transmit data externally.
[COMMAND_EXECUTION]: The skill does not contain any shell commands, system calls, or scripts that execute outside of the provided LLM subagent configuration. It relies entirely on standard file reading and writing within its defined scope.
[REMOTE_CODE_EXECUTION]: There are no patterns involving the download of external code, scripts, or binaries. The skill uses pre-defined model aliases for text processing without fetching remote executable content.
[INDIRECT_PROMPT_INJECTION]: As the skill is designed to ingest and review user-provided documents, it naturally has a surface for indirect prompt injection. This risk is effectively mitigated by the skill's architecture: it uses a multi-agent pipeline with distinct roles (critic, arbitrator, lead-dev) and includes a mandatory human-in-the-loop checkpoint in the reviser.md stage, which requires explicit user approval before any file modifications are applied.
[PROMPT_INJECTION]: The instructions for the critic, dev, and arbitrator roles are focused on architectural review and logical consistency. They do not contain markers intended to bypass safety guardrails or override system instructions.

critic-dialogue