[yaacovcorcos]

Word previews lost parts of real documents, Office files had no zoom, search or page controls and reloaded after every agent command, and the PDF reader redrew and reloaded more than it needed to, which hurts most on slow Windows laptops. This PR makes Office files read like PDFs and takes over the PDF findings from the 2026-11-05 review.

Platforms: tested on macOS (Apple silicon) only, in a dev app and browser tests. Windows and Linux were not run. Nothing here is platform-specific. GPU drawing (enableHWA) falls back to software where Chromium has no GPU. The file revision uses the inode, which Node reports on Windows too. The pinch zoom reads Control+wheel, which Windows and Linux touchpads also send. Still to do: a before/after on a slow Windows laptop (the same long PDF at fit width and at 150%), since the browser CPU throttle does not slow the engines' workers.

Office

• Word previews are complete (fix(web): size the Word preview limits…). T3's preparation step was sized for a renderer on the main thread. A table of a few hundred rows cut the document off there, a body stopped at 5,000 paragraphs, and long style or list definitions were dropped. Three of six real documents previewed wrong; one showed 6 of 21 pages. The new WORD_LIMITS are measured against the @betteroffice engine, which runs in a worker:

  • a paragraph's text is limited, because line breaking grows with its square;
  • so is one element's markup, because a table is laid out in one go;
  • so is a header's length: a longer header is left out alone, and the others, the footers and the notes stay. Text in CDATA counts toward a paragraph's limit (both found by review on the T3 copy).

  All six documents now open whole, with the first page in 0.3–0.7 s as before. A document past 1,000 pages shows its first 1,000 with "Preview limited" instead of an error.

• The PDF reader's controls on Word and PowerPoint: page number and steps, zoom (buttons, keys, pinch, fit width, actual size), a sidebar of pages, and search with every match marked. Each engine searches in its own worker. Word searches its document model, so no page is built to search it. Slides gain a selectable text layer, word by word from their laid-out glyphs. Search: 95 matches in 30 ms on a 21-page report, 63 ms on a 69-slide deck, 2.2 s on a 385-page document (mostly waiting for its layout).

• An open file stays put while the agent works. The file panel remounted every document preview (Word, PowerPoint, archives, fonts, binaries) after each command, edit, tool call or subagent in the thread, so an open report reloaded and went back to page 1 every few seconds. The preview is now keyed by its file; each possible change checks the file's revision (which the server now reports for every exact workspace file, not only PDFs), and only a real change reloads it. A reload you ask for (Reload file from disk, or refresh in Files) still always reloads it, so a failed preview can be retried. Offered to T3 as fix(web): keep an open document preview while other files change pingdotgg/t3code#32670 (issue #32669).

• A Word file or slide deck reopens where it was left: page, point on the page and zoom, for the window's session, as PDFs already do. That also covers a file that changed and reloaded. A jump to a page past those mounted so far (a quote far into a long document) used to stop short; it now arrives.

• Ask in chat on a Word page or slide. The new OfficeCitation quote kind sits beside the file, PDF and Compute kinds. It keeps the file, the page or slide number and the text. The agent gets the path, page and text with a note that the page number is the reader's. The chip reopens the page with the quote marked, and an exported conversation keeps it as a page quote.

PDF reader

• Pages draw on the GPU (enableHWA). The experimental quarter-resolution mode and the halved canvas limit are gone: at 300% the base canvas is twice as sharp (0.14 → 0.27 of screen resolution).
• Opening a PDF no longer extracts every page's text: 70 calls → 1 on a 69-page PDF. Search still prepares when it opens.
• A renewed signed URL no longer reloads the PDF. The server now reports a PDF's file revision (device, inode, size, mtime), and the reader keeps the loaded document while that revision is unchanged.
• Data loss fix: Save a copy kept none of the values typed into a form, yet reported success. It now saves them, and page edits start from them, while saves are still checked against the file on disk.
• Thumbnails take their turn in pdf.js's rendering queue, after the pages in view, and copy pages already drawn. While a PDF opens, only the pages in view are checked.

Upstream seams

Scient code is in apps/web/src/scient/office, packages/shared/src/officeCitations.ts and the existing Scient quote modules. T3 files change only in marked SCIENT-FORK seams:

• officeArchive.ts (limits), OfficePages, WordDocument, Presentation, both workers and their protocols, and word-text;
• AssetAccess, FilePreviewPanel and AssistantCitationChip.

Presentation drops T3's previous/next slide bar, since the reader's page controls replace it. T3's limit tests in AttachmentFilePreview.browser.test.tsx now size their inputs from WORD_LIMITS. UPSTREAM.md records both ("Office preview limits", "Office reader"). The limits are offered to T3 as pingdotgg/t3code#32040, so this copy can go if it merges.

Verification

• Codex reviews: one of the six PDF commits, one of the three Office commits, a re-check of those fixes, one of the Word limits as offered to T3, a design review of the reload and reopen work, a review of it and a re-check. T3's own review bots found the header and CDATA problems in #32040. The earlier rounds found eight real problems, all fixed here: form edits saved against the wrong revision, a thumbnail copy ignoring CSS rotation, find activating a header match (twice), thumbnail paints retaining built Word pages, and in the Word paragraph count, an empty text-box paragraph that turned the count off and indentation or field codes counted as text. The last two rounds found four more, also fixed: a document that failed to open did not retry once its file appeared; a remembered page past the end of a now shorter file was never reached; an older page jump could scroll back over a newer one; and a test that depended on timing. T3's review bot on #32670 found that a manual refresh no longer retried a failed preview, and the live check found a reload returning one page short; both fixed.

• New tests:

  • officeArchive.scient.test.ts: each limit path.
  • officeFindPlacement.test.ts: Word match placement and slide words over glyphs.
  • ScientOfficeReader.browser.test.tsx: find, zoom, sidebar, Ask in chat and reveal; reopening at the left place, far into a long document, in a shorter file, and a newer jump over a waiting one.
  • ScientFileRevisionBoundary.browser.test.tsx: an unchanged file stays mounted across checks; a changed one reloads; a failed open retries.
  • AssetAccess.test.ts: a workspace document reports its revision, and a changed one a new revision.
  • officeCitations.test.ts: link format and what the agent gets.
  • URL renewal: in useScientPdfReader.browser.test.tsx and AssetAccess.test.ts.
  • Filled-form copy, with real pdf.js: pdfRendering.scient.browser.test.tsx.
  • Edit revision: pdfEditSession.browser.test.ts.

  Each regression test fails without its fix.

• Gate: vp fmt --check, vp lint --report-unused-disable-directives, pnpm run typecheck, pnpm run build, pnpm run test:desktop-smoke and git diff --check pass. pnpm run test: everything passes except two suites that fail the same way on main on this machine: provider-fx skills, which reads ~/.fx, and scripts release-build. A notebook and an audio browser test flaked once under load and pass alone. After the reload and reopen work and the merge of main (481 commits, no conflicts): pnpm run typecheck, vp fmt --check and vp lint pass, and the file, Office, PDF, asset and file-surface browser tests (744) and the server asset tests (71) pass; the full suite runs in CI.

• Dev app: Scient (Dev) · officeviewers, with a fresh profile, real Word, PowerPoint and PDF files and screenshots. It found two layout bugs, fixed here: a 3 px horizontal overflow, and fit width not following a resize. Not run there: Save a copy's native dialog, the 30-minute URL renewal, and sending to a provider. Each is covered by the tests above. Second session (2026-11-07), same app: a 21-page Word report open at page 8 stayed mounted at page 8, at the same scroll position, while a Codex agent in the thread ran four commands; switching to a slide deck and back returned to page 8; widening the panel from 277 to 533 px kept page 8; touching the file on disk reloaded it at page 8. That session found the long-document jump that stopped short, fixed here.

Not in this PR

• Word equations (OMML) still render as flat text, and numbers beside Hebrew in PowerPoint land on the wrong side. Both are in the engine.
• Excel keeps T3's viewer.
• pdf.js still keeps about ten drawn pages. Raising that costs memory and needs the Windows measurement first.

Documentation impact: Updated — docs/user/file-previews.md, docs/user/composer.md, UPSTREAM.md.

Model and harness: Claude Opus 5.5 in Claude Code (Scient); reviews by Codex gpt-6.1-sol.

🤖 Generated with Claude Code

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	5.5 KiB	5.5 KiB	0 B (0.0%)	6.8 KiB	✅
Codex	Thread snapshot wire	4.2 KiB	4.2 KiB	0 B (0.0%)	4.9 KiB	✅
Codex	Live turn WebSocket wire	1.3 KiB	1.3 KiB	0 B (0.0%)	2.0 KiB	✅
Codex	Live turn WebSocket decoded	22.3 KiB	22.3 KiB	0 B (0.0%)	24.4 KiB	✅
Codex	Live turn messages	1	1	0 (0.0%)	8	✅
Claude	Total thread wire	5.5 KiB	5.5 KiB	0 B (0.0%)	6.8 KiB	✅
Claude	Thread snapshot wire	4.2 KiB	4.2 KiB	0 B (0.0%)	4.9 KiB	✅
Claude	Live turn WebSocket wire	1.3 KiB	1.3 KiB	0 B (0.0%)	2.0 KiB	✅
Claude	Live turn WebSocket decoded	22.7 KiB	22.7 KiB	0 B (0.0%)	24.4 KiB	✅
Claude	Live turn messages	1	1	0 (0.0%)	8	✅

Baseline: c098727 · PR result: 36d2907 · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.
Bounded compact HTTP snapshot with live cursor.

• Codex decoded thread snapshot: 88.2 KiB
• Claude decoded thread snapshot: 88.4 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.