The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill contains a hardcoded dummy API key within its test suite. Specifically, tests/test_mask_for_model.py uses sk-live-9f2b71ce4a0011 as a sample value to verify the functionality of the PII masking logic. This is dummy test data and does not represent a real credential leak.
[EXTERNAL_DOWNLOADS]: The scripts/liteparse_extract.py utility uses the liteparse library, which may download Tesseract language models (e.g., tessdata_best) from GitHub repositories (specifically from the tesseract-rs or LlamaIndex ecosystem) during the first execution if not provided locally. The skill documents this behavior and provides a --tessdata flag for offline operation.
[COMMAND_EXECUTION]: The skill invokes the standard system utility pdftotext to extract text from PDF files. This is performed via subprocess.run in scripts/split_packet.py and tests/test_encoding_crlf.py. The calls use a list of arguments to target specific local file paths, minimizing command injection risks.
[INDIRECT_PROMPT_INJECTION]: As the skill is designed to ingest and process untrusted PDF and OCR data from various sources (legal documents, scans), it possesses an inherent attack surface for indirect prompt injection.
Ingestion points: Untrusted data enters the agent context through scripts/normalize.py and scripts/routing_gate.py when reading PDF or text-based documents.
Boundary markers: The VLM prompt template in references/prompt_vlm_ocr_pl.md utilizes explicit delimiters (=== POCZATEK SZABLONU === and === KONIEC SZABLONU ===) to scope the task for vision-language models.
Capability inventory: The skill has the capability to write files (via split_packet.py using the --eksport flag) and execute local commands (via pdftotext).
Sanitization: The skill includes proactive sanitization via scripts/mask_for_model.py, which masks Polish PII (PESEL, NIP, IBAN) and machine secrets (API keys, Bearer tokens) before data is sent to external LLMs.
[SAFE]: The skill implements several high-quality security best practices, including a routing gate (scripts/routing_gate.py) that explicitly detects and rejects 'Zip Bombs' (via decompression ratio checks) and 'Billion Laughs' XML entity expansion attacks in OOXML files.

doc-intel-contract-pl