The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill instructions reference sensitive SSH key paths (~/.ssh/id_rsa) and guide the user through managing authentication credentials for cluster-wide access.
[COMMAND_EXECUTION]: The provided scripts (e.g., multi-node-test.sh, pre-test-check.sh) perform remote command execution across multiple nodes via SSH, defaulting to the root user and disabling standard security features like StrictHostKeyChecking, which increases vulnerability to Man-in-the-Middle (MitM) attacks.
[PERSISTENCE]: The skill provides instructions and automation for modifying persistence mechanisms, specifically adding public keys to ~/.ssh/authorized_keys and appending environment configurations to ~/.bashrc to maintain access and environment state across sessions.
[PRIVILEGE_ESCALATION]: The operational model assumes high-level privileges, encouraging execution as the root user and requiring privileged mode for Docker containers to facilitate direct hardware (NPU) access.
[INDIRECT_PROMPT_INJECTION]: The skill contains data ingestion surfaces where it parses external files and command outputs (e.g., version.cfg, npu-smi output) using regex in parse-hccl-result.py without explicit sanitization, while possessing high-impact capabilities like remote command execution.
Ingestion points: scripts/parse-hccl-result.py (reads from logs/stdin), scripts/pre-test-check.sh (reads remote config files).
Boundary markers: Absent.
Capability inventory: Remote shell execution (SSH), file writing, and network operations.
Sanitization: Limited to basic regex patterns for log extraction.
[EXTERNAL_DOWNLOADS]: Fetches software dependencies from well-known sources including MPICH's official site and the vendor's own documentation portals.

hccl-test