The Agent Skills Directory

[DATA_EXFILTRATION]: The skill accesses ~/.claude/settings.json to retrieve the MINIMAX_API_KEY. This is a sensitive platform-level configuration file that may contain other private settings or credentials.
[DATA_EXFILTRATION]: The skill transmits extensive research context, including project narrative documents and recent experiment results, to an external API (api.minimax.io). While this is a well-known service, the automated transfer of proprietary research data to a third party should be carefully considered.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection by design. It treats 'action items' returned by an external model as instructions to be implemented via file modifications and shell commands.
Ingestion points: External API responses from MiniMax chat completions (SKILL.md).
Boundary markers: None present to distinguish instructions from the external model from trusted logic.
Capability inventory: The skill utilizes Bash(*), Write, Edit, and Agent tools to modify code and execute commands.
Sanitization: No validation or sanitization process is described before implementing changes suggested by the external model.
[COMMAND_EXECUTION]: The skill performs automated script generation and execution based on external input. It is specifically instructed to deploy experiments to GPU servers via SSH and manage them in screen/tmux sessions autonomously.
[COMMAND_EXECUTION]: The instructions include a fallback mechanism to bypass tool limitations by silently using shell redirection (cat << 'EOF') for file writes without informing the user or seeking permission.

auto-review-loop-minimax