The Agent Skills Directory

[COMMAND_EXECUTION]: The skill generates and executes shell commands using the quark-cli tool. However, this is the primary purpose of the skill and is protected by multiple security controls:
Execution Gate: Checkpoint 3 requires explicit user approval ("yes", "run it") before any command execution occurs.
Path Sanitization: Instructions explicitly mandate resolving local model and output directories to absolute paths and quoting all user-controlled paths and values to prevent injection.
Trusted Tooling: The skill restricts itself to using the public quark-cli installed in the environment and forbids generating or patching custom runners.
Remote Code Protection: The skill defaults to --no_trust_remote_code and requires explicit user consent before omitting this flag for models that require custom repository code.
[DYNAMIC_EXECUTION]: The skill uses small Python scripts to inspect the environment and model metadata. These scripts are restricted to read-only metadata inspection (e.g., importlib.metadata, AutoConfig.from_pretrained with trust_remote_code=False) and do not execute untrusted external code.
[INDIRECT_PROMPT_INJECTION]: While the skill ingests model configuration files (config.json), it mitigates risks by using standard Hugging Face AutoConfig parsers and providing structured risk analysis for ambiguous or unknown model types in the model_analysis.json artifact.

quark-torch-llm-ptq