The Agent Skills Directory

[PROMPT_INJECTION]: The skill instructions are purely functional and describe a structured workflow for code review. There are no patterns suggesting overrides of safety guidelines or hidden instructions to bypass model constraints.
[DATA_EXFILTRATION]: The skill reads project-local files like 开发计划.md and PRD documents and writes results to a local documentation directory (docs/05-ADR/). It does not contain any network-facing commands (e.g., curl, wget) or hardcoded credentials.
[REMOTE_CODE_EXECUTION]: No remote code execution patterns or external package installations were detected. The sub-agent logic described is an internal orchestration pattern for LLM contexts and does not involve executing untrusted scripts.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data by reading markdown files (开发计划.md, PRDs). While this presents an attack surface where instructions embedded in these docs could influence the agent, the skill's capabilities are limited to reading/writing markdown reports, presenting a very low risk. No specific sanitization or boundary markers are mentioned, which is common for internal development tools.
[OBFUSCATION]: Analysis for Base64 encoding, zero-width characters, homoglyphs, and hidden text patterns returned no findings. The content is plain text markdown.

code-review