The Agent Skills Directory

[COMMAND_EXECUTION]: Helper scripts invoke local TypeScript tools and package managers using safe subprocess calls with argument lists. The local_binary utility in scripts/local_tools.py restricts binary lookup to the current project or repository root, preventing the execution of shadowed or malicious binaries from external or ancestor directories.
[DATA_EXFILTRATION]: scripts/inspect_typescript.py and scripts/run_typecheck.py analyze project metadata and compiler output but report results primarily as aggregated counts, stable error codes, and enums. This design prevents sensitive repository file paths or raw source code from being exfiltrated to the AI context.
[PROMPT_INJECTION]: The skill manages potential indirect prompt injection from untrusted project data. Ingestion points: Reads package.json, tsconfig.json, and compiler output in scripts/inspect_typescript.py. Boundary markers: Explicitly defined in SKILL.md under the 'Security Model' section, instructing the agent to ignore instructions embedded in project evidence. Capability inventory: Executes local binaries via subprocess in scripts/run_typecheck.py and scripts/trace_perf.py. Sanitization: Paths are normalized, compiler output is summarized, and metadata is parsed into strictly defined enums before reporting to the agent context.

typescript