The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill leverages the uv tool to dynamically manage and download Python dependencies required for its operation, such as docling, pdfminer.six, and PyPDF2. These are well-established libraries for document extraction and are retrieved from standard package registries.
[COMMAND_EXECUTION]: The skill executes external CLI tools including pdftotext, pdfinfo, and uv through the subprocess.run interface. The scripts correctly pass command-line arguments as an array of strings, which effectively mitigates the risk of command injection compared to using raw shell strings.
[PROMPT_INJECTION]: As a tool designed to process and analyze third-party documents, the skill is susceptible to indirect prompt injection attacks. Malicious instructions embedded in a user-provided PDF or EPUB file could attempt to hijack the agent's logic during the knowledge compilation phase.
Ingestion points: Document text is extracted and stored in local files such as cache/jobs/<job-id>/source.txt before being analyzed by the agent in segments.
Boundary markers: The skill does not currently implement explicit boundary markers (e.g., unique XML tags) or specific 'ignore' directives to separate the extracted content from the agent's instructions.
Capability inventory: The skill has the capability to read and write files in the local filesystem and execute subprocesses for text extraction tasks.
Sanitization: There is no automated sanitization or filtering logic applied to the extracted text to detect or remove potential injection patterns before processing.

knowledge-compiler