The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run, subprocess.Popen, os.execv, and os.execvp to manage its execution environment and call standard document utilities.\n
scripts/ocr.py uses os.execvp to launch worker scripts through the uv environment manager.\n
scripts/parse.py uses os.execv to re-execute itself within the specific paddleocr tool's Python interpreter, ensuring correct library versions and CUDA configurations are loaded.\n
scripts/doc.py and scripts/ocrlib/checks.py use subprocess.run to execute binary utilities like pdfinfo and pdftotext for metadata extraction and verification of parsed text against the original document layer.\n- [EXTERNAL_DOWNLOADS]: The skill performs targeted network requests to local or remote servers to manage OCR processing and provides instructions for dependency management.\n
scripts/parse.py uses urllib.request.urlopen to query available models from a user-provided or local OpenAI-compatible vLLM server endpoint (/v1/models). This is a functional requirement for offloading parsing to a specialized backend.\n
Documentation in references/install.md guides users to install dependencies from reputable sources, including official PaddlePaddle repositories.\n- [INDIRECT_PROMPT_INJECTION]: As a tool designed to ingest and parse untrusted PDF and image content, the skill has an inherent indirect prompt injection surface.\n
Ingestion points: Document files are read via scripts/parse.py and processed using the PaddleOCR engine.\n
Boundary markers: The skill generates Markdown output with clear delimiters for pages, figures, and findings (e.g., , ).\n
Capability inventory: The environment has the capability to run shell commands for PDF processing and filesystem management.\n
Sanitization: The skill mitigates risks through its design; it includes specialized auditing tools (ocr.py findings, ocr.py diff) and a "settle" workflow (detailed in references/settle.md) that explicitly instructs a reviewing agent to verify the parsed output against the ground-truth text layer, identifying hallucinations or injected text.\n- [DYNAMIC_EXECUTION]: The skill implements process replacement and dynamic environment selection to ensure reliability across different user setups.\n
The use of os.execv to jump into a uv-managed tool directory is a common and safe pattern for ensuring complex AI models run in an environment with the correct binary dependencies.

paddleocr