The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill instructs the agent to read CLAUDE.md to extract sensitive information like WANDB_API_KEY and then use it in shell commands (e.g., ssh <server> "wandb login <WANDB_API_KEY>"). This involves active handling and potential exposure of long-lived credentials in command history.
[DATA_EXFILTRATION]: The skill accesses sensitive file paths including ~/.claude/feishu.json to read notification configurations and CLAUDE.md for server/API secrets. While used for legitimate purposes, this represents an access pattern to sensitive user data.
[COMMAND_EXECUTION]: The skill performs extensive command execution across several contexts:
Remote Execution: Uses ssh to launch background processes via screen and perform environment setup on remote servers.
Dynamic Script Generation: Step 4 describes generating a modal_launcher.py script at runtime and executing it via modal run.
Code Modification: Step 3.5 explicitly instructs the agent to modify the user's training scripts to inject wandb logging logic.
[EXTERNAL_DOWNLOADS]: The skill facilitates downloading and installing Python packages on remote instances via pip install -r requirements.txt.

aris-run-experiment