The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/case_store.py invokes the pdftotext system utility using subprocess.run to extract data from local files. While it uses a list of arguments rather than a shell string, it remains a point of interaction with the host operating system.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data by reading the contents of PDF files within the case directory to extract legal metadata like case numbers and dates. This introduces an attack surface where maliciously crafted documents could attempt to confuse the extraction logic or provide adversarial instructions to the agent.
Ingestion points: PDF files located in case sub-directories (e.g., `08
法院送达) are read by pdftotextand the output is parsed inscripts/case_store.py`.
Boundary markers: The skill relies on a manual --apply flag to commit extracted data, providing a user-driven boundary before state changes occur.
Capability inventory: The skill can execute local commands (pdftotext), write to the filesystem (os.replace), and update structured case data (case.yaml).
Sanitization: Text extracted from PDFs is filtered using specific regular expressions to ensure they match expected date and case number formats before being accepted.

case-progress