The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts scripts/create_container.sh and scripts/start_service.sh dynamically assemble shell commands using parameters derived from user input (such as model paths, image names, and server addresses). These commands are executed locally via eval or remotely via SSH. Unsanitized input in these fields could lead to command injection as the variables are interpolated directly into shell strings.
Ingestion points: User-provided configuration values in the '部署前准备' (Preparation) section of SKILL.md and script command-line arguments.
Boundary markers: None present in the shell scripts to isolate user input from executable command logic.
Capability inventory: Local command execution (eval), privileged Docker container management (docker run --privileged), and remote execution via SSH.
Sanitization: No evidence of parameter escaping or validation before shell interpolation.
[REMOTE_CODE_EXECUTION]: The skill uses SSH to execute administrative tasks on remote servers, including managing containers and starting background inference services when the --mode remote parameter is specified.
[EXTERNAL_DOWNLOADS]: The deployment workflow relies on fetching documentation, model configurations, and recommended settings from external repositories including GitCode (gitcode.com) and the official vLLM project documentation (docs.vllm.ai).
[DATA_EXFILTRATION]: The scripts/monitor.sh script is configured to send service logs and status updates to a vendor-internal endpoint (http://xiaoluban.rnd.huawei.com:80/) to provide notifications via the WeLink platform.
[COMMAND_EXECUTION]: The troubleshooting documentation suggests a procedure to automatically append a public key to a server's authorized_keys file to facilitate automated access. This modification of the security configuration establishes persistent remote access to the target infrastructure.

external-gitcode-ascend-vllm-ascend-deploy