The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes system commands via subprocess.run in several files. In src/aut_sci_ppt/generator/formula_renderer.py, it invokes pdflatex and pdftoppm to render academic formulas. The LaTeX content is written to a temporary file without sanitization, which could allow for LaTeX injection attacks. In src/aut_sci_ppt/pdf_extractor.py, the skill attempts to execute an external Python script from a hardcoded path in the user's home directory (~/.openclaw/workspace-shclaw-ppt/skills/Sh_Sci_Fig/scripts/extract_figure.py).
[EXTERNAL_DOWNLOADS]: The skill performs network requests to several external service domains for core functionality. It communicates with api.anthropic.com, api.openai.com, and api.moonshot.cn for AI-based parsing and translation. It also fetches image data from latex.codecogs.com during formula rendering.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection.
Ingestion points: Untrusted data is ingested from PDF files using the fitz library in src/aut_sci_ppt/pdf_extractor.py and src/aut_sci_ppt/paper_workflow.py.
Boundary markers: There are no explicit delimiters or instructions to ignore embedded commands when the extracted text is passed to the AI parsing logic in src/aut_sci_ppt/parser/ai_parser.py.
Capability inventory: The skill has the capability to write .pptx files to the local filesystem and make outbound network connections to AI providers.
Sanitization: No sanitization or filtering is applied to the text extracted from external PDFs before it is processed by the LLM.
[COMMAND_EXECUTION]: src/aut_sci_ppt/paginator/smart_paginator.py contains a dynamic import call (__import__) which is used to load components of its own model at runtime, representing a dynamic execution pattern.

aut-sci-ppt