Document Reader

Reads bounded text from PDF, DOCX, XLSX, PPTX, TXT, CSV, MD, JSON and RTF files on Linux, macOS and native Windows. Sources are request attachments or files inside the configured agent work directory and workspace. Request-specific path restrictions and attachment identities retain their existing meaning.

The shared limits remain 10 MiB per input file, 40,000 characters per requested window, and the existing archive expansion and PDF worker resource limits. Results describe pagination and omitted text. The reader does not execute document macros or modify the source.

Windows requirements

The SDK reads through built-in Windows PowerShell and its C# helper. No Companion, WSL, Docker, elevation or execution-policy bypass is needed. A blocked reader returns document_safe_read_unavailable and instructs the model not to retry or bypass the restriction. Request cancellation stops the read and extraction. PDF and Office parsing use the same packaged libraries as the other platforms; no Office installation or OCR is required.

PDF parsing runs in a separate Node process on Windows to keep native parser faults out of the Runtime. Linux/macOS retain their worker-thread boundary. Both use the same parser, page/text limits and 30-second deadline. Windows passes documents through pipes, without temporary document files, and waits for process exit on cancellation. The V8 old-generation limit is 128 MiB, young-generation at most 16 MiB (4 MiB semi-spaces), and stack 4 MiB. These are JavaScript engine limits, not a cap on native allocations. Explicit Node CLI flags enforce the heap and stack limits; operator-provided Node configuration remains inherited.

If Windows cannot start or stop the PDF process, the tool returns document_pdf_process_unavailable and asks the model not to retry or bypass the restriction.