The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions and reference documents utilize standard development commands (npm, pnpm, git, node) for package management and CI workflows. These are well-integrated into the intended purpose of the skill and include multiple safety gates.
[DATA_EXFILTRATION]: The skill includes explicit security boundaries prohibiting the publication of artifacts containing credentials, .npmrc files, or session logs. Network operations identified in the smoke-test recipes (using curl and Invoke-WebRequest) are restricted to local host testing (127.0.0.1) to verify authentication protections.
[REMOTE_CODE_EXECUTION]: External resources are limited to trusted repositories, such as deepseek-ai/deepseek-harness. The installation patterns follow standard package manager behaviors without suspicious piping of remote content to shell interpreters.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from package.json and plugin manifests. While this represents a theoretical attack surface, the risk is mitigated by the use of a dedicated validation script (verify-release.mjs) that performs strict regex-based parsing and semantic validation of the input data.

plugin-release