The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill utilizes bootstrap scripts (tiangong-runtime-bootstrap.sh and tiangong-runtime-bootstrap.ps1) to download the Foundry runtime and required Node.js binaries directly from the vendor's official GitHub repository, ensuring that the components are sourced from a verified location.
[REMOTE_CODE_EXECUTION]: The skill downloads and executes a binary CLI tool to manage data tasks. This process is secured by the bootstrap scripts, which perform integrity checks by validating every downloaded component against pinned SHA-256 hashes defined in the bootstrap-lock.json file prior to execution.
[COMMAND_EXECUTION]: The agent interacts with the tiangong-foundry CLI through a controlled bootstrap process. This mechanism ensures that the runtime environment is restricted, arguments are passed directly without a shell, and sensitive environment variables are managed according to the vendor's specifications.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted LCA datasets and evidence from external sources, which presents an attack surface for indirect prompt injection.
Ingestion points: The agent reads dataset directories and source evidence tables provided by the user.
Boundary markers: The instructions enforce the use of structured JSON schemas (e.g., tiangong-foundry.task-start.v1) for all task operations and mandate that AI-generated inferences be clearly labeled as assumptions to distinguish them from source facts.
Capability inventory: The skill has the ability to execute the vendor's CLI, write decision files to the workspace, and perform network communication for task status and submission via the authenticated CLI.
Sanitization: The Foundry runtime includes native validation layers for schema and content integrity. Furthermore, the skill's instructions explicitly require the agent to pause for human judgment when encountering ambiguous evidence or critical decision points.

foundry-tidas-import