The Agent Skills Directory

[COMMAND_EXECUTION]: The skill extensively uses the Bash tool to execute arbitrary commands for running experiments, monitoring processes, and managing files. It specifically includes logic to deploy and run code on remote GPU servers via SSH and session managers like screen or tmux.
[REMOTE_CODE_EXECUTION]: In Phase C, the skill autonomously writes or modifies experiment scripts and model code, which it then proceeds to execute. This cycle of automated code generation and execution is a core feature but represents a significant security risk if the logic is subverted.
[DATA_EXPOSURE]: The skill attempts to read the file ~/.claude/feishu.json. This file is located in a sensitive application configuration directory and is likely to contain API tokens, webhooks, or other credentials used for the Feishu/Lark messaging platform.
[DATA_EXFILTRATION]: The skill is designed to send research scores, weaknesses, and pipeline status updates to an external Feishu endpoint if configured. This creates a pathway for data to leave the local environment.
[INDIRECT_PROMPT_INJECTION]: The skill possesses an indirect prompt injection surface. In Phase B, it parses assessment results and "action items" from the output of an external tool (mcp__codex__codex). These items are then used to drive code changes and command execution in Phase C.
Ingestion points: Output from the mcp__codex__codex tool (Phase B).
Boundary markers: None identified for the parsed reviewer feedback.
Capability inventory: Full Bash access, file writing/editing, and remote SSH execution.
Sanitization: No evidence of sanitization or validation of the action items before implementation.
[EXTERNAL_DOWNLOADS]: The skill uses curl to fetch bibliographic information from academic services including dblp.org and doi.org during the citation verification process.

auto-review-loop