The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill uses bunx @rayhanadev/truffler to analyze the repository. bunx (similar to npx) downloads the package from the npm registry and executes it, which involves running code from a third-party developer that is not the skill author or a trusted vendor.
[REMOTE_CODE_EXECUTION]: The instructions note that agents fetch canonical prompts from react.doctor/prompts/... at runtime. This indicates the agent dynamically loads and follows instructions from a remote network source.
[COMMAND_EXECUTION]: The skill executes various local shell commands, including rg (ripgrep) for searching and nr (a project script runner) to execute tasks like typecheck, test, lint, changeset, and smoke:json-report.
[PROMPT_INJECTION]: The skill has an indirect prompt injection surface when evaluating repository content.
Ingestion points: Reads code content from the packages directory via rg and @rayhanadev/truffler.
Boundary markers: No delimiters or safety warnings are implemented to distinguish between repository code and instructions.
Capability inventory: File system reads, command execution via nr and bunx, and telemetry emission via Sentry.
Sanitization: There is no evidence of sanitization or filtering of the content ingested from source code files before the agent processes it.

product-thinking