The Agent Skills Directory

[SAFE]: The skill is focused on repository maintenance and backfilling metadata. Its instructions and accompanying scripts are strictly aligned with this operational purpose.
[COMMAND_EXECUTION]: The maintenance workflow involves running local TypeScript scripts using bun and standard development utilities like vitest and tsc. These operations are limited to the project's source code and do not involve untrusted input.
[COMMAND_EXECUTION]: The scripts/extract-model-ids.ts utility uses await import() to read data from the project's own model definition files. This is a standard pattern for programmatically accessing internal TypeScript data structures.
[PROMPT_INJECTION]: The skill describes a research workflow that ingests external documentation from model vendors and Hugging Face.
Ingestion points: External technical documentation, system cards, and official model cards.
Boundary markers: The workflow requires the research agent to return structured data in a specific JSON schema.
Capability inventory: The skill provides codemod scripts (apply-cutoffs.ts, apply-family.ts) that can modify TypeScript files in the local packages/model-bank/ directory.
Sanitization: Potential risks from external data are mitigated by strict sourcing rules, a policy filter that discards data from non-authoritative sources, and an adversarial verification agent that re-validates all research claims before any code modifications are applied.

model-bank-metadata