The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads the uxc binary from a third-party repository on GitHub (holon-run/uxc).
Evidence: scripts/install-uxc.zsh fetches the 0.22.0 release from GitHub via curl.
[REMOTE_CODE_EXECUTION]: The skill installs and executes the uxc binary. It employs strict verification by comparing the SHA-256 hashes of both the tarball and the binary against hardcoded pins defined in scripts/lib/uxc-release.zsh.
Evidence: Hash validation logic is implemented in scripts/lib/uxc-owned-binary.zsh and enforced during installation.
[DYNAMIC_EXECUTION]: The skill dynamically generates a shell script at runtime to act as a command-line facade for the Ardot tool.
Evidence: scripts/lib/uxc-link-contract.zsh renders a shell wrapper, including validation logic that prevents shell injection by rejecting single quotes or newlines in configuration variables.
[INDIRECT_PROMPT_INJECTION]: The skill processes external design files from ardot.tencent.com, which represents an untrusted data source.
Ingestion points: External URLs processed by the ardot-mcp-cli and readiness checks in scripts/uxc-readiness.zsh.
Boundary markers: The skill defines specific status identifiers like FACADE_READY and TASK_ACCEPTANCE to distinguish between stages of execution.
Capability inventory: The skill can execute the uxc binary, write manifest and link files to local directories, and communicate with the Ardot HTTP endpoint.
Sanitization: Readiness results are sanitized using jq to filter the JSON envelope and discard non-essential data before providing output to the agent.

ardot-mcp