The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fetches and installs several third-party Python packages from PyPI at runtime using the uv tool in isolated environments.
SKILL.md and scripts/extract_paper.py invoke uv run to install pymupdf4llm==1.28.0.
scripts/render_math.py utilizes uv run to install latex2mathml==3.78.1 and defusedxml==0.7.1.
[COMMAND_EXECUTION]: The skill performs network operations to clone external Git repositories for source code auditing.
SKILL.md (Step 5) and references/code-audit.md instruct the agent to perform a "shallow temporary clone" of authoritative repositories linked in papers to verify technical claims.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external academic papers (PDFs, HTML pages, and pasted text), creating a potential surface for indirect prompt injection.
Ingestion points: Inputs include PAPER.pdf, official full-text HTML URLs, and pasted text provided by the user or fetched from the web.
Boundary markers: The skill uses structured analysis phases and Markdown coordinates (C for claim, E for evidence, L for limitation) to anchor results, though it lacks explicit "ignore instructions" delimiters for the raw source text.
Capability inventory: The agent has capabilities for subprocess execution (uv, python, git), file system writes (generating reports), and network access (cloning repos).
Sanitization: The skill employs defensive measures, such as using defusedxml to safely parse MathML and html.escape for HTML report generation, reducing the risk to the end-user viewing the reports.
[DYNAMIC_EXECUTION]: The skill uses dynamic loading techniques to import conversion libraries in its auxiliary scripts.
scripts/render_math.py uses importlib.import_module() to dynamically load the latex2mathml.converter module after verifying its availability.

paper-reading