Uploads now go directly from the browser to R2 through signed PUTs. Files above 16 MiB use resumable multipart uploads with 8 MiB parts. Images can be up to 100 MiB, documents 1 GiB and ZIP archives 5 GiB; cron recovers lost acknowledgements and aborts abandoned sessions.
Supported documents (text-layer PDF, Office, OpenDocument, RTF, EPUB, CSV and plain text) are converted in a browser Web Worker with AnyDoc WASM. Cron validates and indexes up to 192 KiB of Markdown per document. A server-side conversion was measured and rejected because it exceeded the 128 MiB isolate limit. The assistant and MCP read the indexed text through files_list, search, fetch and paged document_read, plus the plan-from-document prompt.
Sources: direct uploads and conversion, document tools, browser workers, bounded indexing.