The Agent Skills Directory

[SAFE]: The skill implements a proactive security policy regarding credential handling. It explicitly instructs the agent never to ask for, print, or forward sensitive tokens such as access tokens, refresh tokens, or passwords, and directs the user to perform authentication in a trusted terminal using the tiangong-lca auth login command.
[SAFE]: The skill uses a Node.js wrapper script (scripts/run-embedding-ft.mjs) to execute a vendor-specific CLI tool. The script safely handles arguments and delegates execution to a utility library (cli-launcher.mjs) without using dangerous dynamic evaluation patterns.
[SAFE]: External configuration is managed through environment variables (TIANGONG_LCA_API_BASE_URL, TIANGONG_LCA_SUPABASE_PUBLISHABLE_KEY, TIANGONG_LCA_OAUTH_CLIENT_ID), which is a standard and safe practice for secret management.
[INDIRECT_PROMPT_INJECTION]: The skill processes job definitions from JSON files (e.g., assets/example-jobs.json) which contain metadata about database tables and functions. While this represents an ingestion surface for untrusted data, it is the intended primary purpose of the skill, and the execution is performed by the underlying CLI tool rather than the agent directly.
Ingestion points: --input parameter in scripts/run-embedding-ft.mjs and the assets/example-jobs.json file.
Boundary markers: Absent in the script-level processing.
Capability inventory: The runTiangongCommand function in scripts/run-embedding-ft.mjs executes the CLI, which has capabilities to interact with remote Supabase functions and database systems.
Sanitization: Not explicitly performed in the wrapper script; security relies on the validation logic within the tiangong-lca CLI and the receiving Supabase Edge Function.

embedding-ft