The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The scripts/bootstrap.py script downloads a CLI binary from GitHub releases (storytold/pdfcraft or full-aigc-skills/pdfcraft-skills). It performs SHA256 checksum verification against a hardcoded lock file (scripts/runtime.lock.json) before executing the binary to confirm its version and availability.
[COMMAND_EXECUTION]: The skill frequently executes external processes using the subprocess module. scripts/execution.py manages the lifecycle of the native PDF CLI, while scripts/ocr.py invokes the Tesseract OCR engine. These executions are scoped to specific binaries but operate outside the agent's immediate environment.
[EXTERNAL_DOWNLOADS]: scripts/bootstrap.py fetches the primary toolchain (ZIP archive) from GitHub. While the URLs are restricted to specific repositories and validated with cryptographic hashes, it represents a dependency on external hosting infrastructure.
[INDIRECT_PROMPT_INJECTION]:
Ingestion points: The skill extracts text content from PDF documents through two vectors: direct text extraction via the native CLI (scripts/verification.py) and OCR analysis of document images (scripts/ocr.py).
Boundary markers: Extracted text is stored in pageText fields within JSON receipts. There are no explicit instructions or delimiters provided to the agent to treat this extracted text as potentially untrusted content.
Capability inventory: The skill possesses the ability to execute commands (subprocess.Popen), manage local files, and perform network requests for tool installation.
Sanitization: While the skill provides automated redaction for specific secret keywords (e.g., 'api_key', 'token') in scripts/execution.py, it does not include sanitization or filtering to prevent natural language instructions within the processed PDFs from influencing the agent's logic.

pdfcraft-cli-automation