The Agent Skills Directory

[SAFE]: The skill operates entirely within the local environment. It explicitly forbids network access, credential handling, and external transmissions in its routing rules and implementation.
[COMMAND_EXECUTION]: The skill uses bun to execute local maintenance and generation scripts (scripts/generate.mjs, scripts/validate-portable-v1-output.mjs, and scripts/validate-hermes-agent-maker.mjs). These scripts are part of the skill package and follow strict safety protocols, including ownership checks and atomic write operations.
[INDIRECT_PROMPT_INJECTION]: The skill ingests user-provided text (the summary field) which is interpolated into generated files. This is mitigated by multiple layers of defense: a character-restricting regex (safeSummary), a blocklist for sensitive terms (forbiddenTerms), and the use of draft extensions for sensitive files like USER.md and MEMORY.md. The generator also scans its own output for forbidden behavior before writing to disk.
[SAFE]: The file-writing logic in scripts/generate.mjs is highly defensive. It uses realpath and containment checks to ensure no files are written outside the designated workspace. It also implements an ownership marker system (ownership.json) to prevent overwriting files not created by the skill, and uses a transactional journaling system to handle interrupted writes.
[SAFE]: The portable plugin validation logic uses a pinned schema system that verifies the integrity of local schema files via SHA-256 hashes before use, preventing tampering with the validation oracle.

hermes-agent-maker