The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fetches 'What's New' documentation from docs.python.org when the target codebase's Python version exceeds the skill's current verified version (3.14). This is a well-known and trusted service used solely for documentation retrieval.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface because it is designed to ingest and analyze untrusted Python source code and PR diffs.
Ingestion points: Untrusted source code is interpolated into the {{TARGET_CONTENT_OR_PATHS}} placeholder in the references/reviewer-prompt.md template.
Boundary markers: The review template utilizes markdown headers and explicit scoping language to delineate instructions from untrusted content.
Capability inventory: The reviewer subagent's capabilities are restricted to generating structured text verdicts; it lacks access to the host file system, network tools, or external command execution.
Sanitization: The 'blind reviewer' protocol effectively mitigates risk by processing the code in a self-contained subagent prompt, preventing untrusted input from influencing the primary agent's broader context or conversation history.

adversarial-python