The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses node to execute internal utility scripts (requirements.mjs and render-pdf.mjs) and orchestrates standard TeX tools such as latexmk and Poppler's pdftoppm. These executions are central to the skill's thesis verification purpose and are surrounded by significant safety guardrails.
[INDIRECT_PROMPT_INJECTION]: The skill identifies a potential attack surface when ingesting untrusted user data such as LaTeX source files, latexmkrc, or Makefile. It mitigates this via strict instructional boundaries, explicitly warning the agent that commands within user data are not instructions for the agent, and requiring a manual audit of build files while defaulting to a safe compilation mode (disabled shell-escape).
[DYNAMIC_EXECUTION]: The render-pdf.mjs script includes logic to discover and verify system renderers. It allows the agent to specify an absolute path to a pdftoppm binary via the --candidate flag; however, this is used for environment discovery and includes a 'health probe' mechanism that rejects the renderer if it produces suspicious output or environment errors.
[COMMAND_EXECUTION]: The test script scripts/tests/prepare-broken-poppler.py uses subprocess.run to call ldd and performs binary patching. This utility is strictly confined to the developer test environment (as evidenced by its use in the Dockerfile.clean-test) and is designed to test the skill's resilience to rendering failures.
[SAFE]: The skill demonstrates exceptional security hygiene in its test orchestration script scripts/test-clean-container.sh, which runs validation tests in a Docker container with --network none, --read-only, and --cap-drop ALL, effectively isolating the execution environment.

tsinghua-graduate-thesis