The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to analyze untrusted NSFC proposal documents, which presents a surface for indirect prompt injection attacks. 1. Ingestion points: External LaTeX (.tex) files identified by list_proposal_files.py and read into the agent's context. 2. Boundary markers: Absent; the master_prompt_template.md defines roles but does not explicitly instruct the model to ignore directives embedded within the proposal text. 3. Capability inventory: The skill can execute internal Python scripts (build_parallel_vibe_plan.py, finalize_output.py), perform file management (move, delete, create), and potentially invoke the gh CLI tool for bug reporting. 4. Sanitization: Absent; the skill does not filter or sanitize proposal content before processing.
[COMMAND_EXECUTION]: The skill executes several internal Python utility scripts to automate the review and cleanup process. These scripts include robust path validation logic (e.g., in finalize_output.py) to prevent directory traversal and ensure operations remain within the designated review directory.

nsfc-reviewers