dsh-attachment-formats — Attachment Format Expansion (Codex-style)

English | 中文

A DeepSeek Harness web plugin that makes the composer accept many more attachment formats, Codex-style. Zero core-package changes: a pure plugin that reuses the harness-native image draft rail, upload limits, history rendering and model request pipeline.

Supported formats

File	Handling	Destination
PNG / JPEG / WebP / GIF	native pipeline (plugin not involved)	image draft rail (native)
PDF (with text layer)	text-layer extraction (≤40 pages via the pymupdf4llm high-fidelity engine; larger/unavailable falls back to pdfjs)	full text on a document card (merged on send); over-limit → workspace spill + index card
PDF (scanned / no text layer)	tesseract.js OCR (accepted only at confidence ≥45), falls back to page images	OCR success → text channel; failure → image draft rail (vision models only)
Word (.docx) / Excel (.xlsx) / PPT (.pptx)	text extraction — docx via mammoth HTML → turndown, tables kept as Markdown pipe tables	document card (merged on send); over-limit → spill + index card
Legacy .doc / .xls / .ppt	LibreOffice headless → docx/xlsx/pptx → standard Office pipeline (needs soffice; clear error when absent)	document card (merged on send)
epub / odt / rtf	pandoc → Markdown (probe on PATH); epub/odt fall back to jszip+turndown without pandoc; rtf requires pandoc	document card (merged on send)
TIFF (.tiff/.tif)	sharp (libvips) → PNG pages (multi-page, ≤20)	native image draft rail
txt / md / json / code	read in the browser (UTF-8, GB18030 fallback)	document card (merged on send); over-limit → spill + index card
BMP / ICO / AVIF / SVG etc.	browser decode → canvas → PNG	native image draft rail
iWork / audio-video / archives	— (not yet supported; explicit notice, skipped)	—

Document cards (Codex-style mounting, composer stays clean)

Text-like attachments that are dragged in or picked are not stuffed into the input box: their content mounts as a document card above the composer (file name + character count + full-text/index label, individually removable), while images keep flowing into the native image draft rail. You type normally, and at the moment of sending the plugin merges the card content into the message (with [attachment: <file name>] provenance markers) before the native submit — your prompt always stays on top and no content is lost:

• each card has a send button: send documents even without typing anything;
• pressing Enter / the native send button merges the cards first, then submits;
• cards are not merged while the model is mid-reply (they stay put for later).

Long documents (index-card mode, never silently truncated)

Text beyond 80k characters and long multi-page PDFs are not stuffed into the message. Instead:

1. the host spills them into the session workspace .dsh-attachments/<sha-16>/ (content-addressed, reused on re-drop, auto-cleaned after ~7 days of no access):
   • doc.md — PDF text layer assembled per page (leading <!-- pN --> markers), Office-extracted text, long text as-is (long JSON is prettified to doc.json);
   • pages/pNN.png — rendered page images (≤100 pages, for vision models via read_image; rendered lazily, only when the index-card path needs them);
   • manifest.json — source, page/line/char counts, engine, full source SHA-256 and the converter-policy fingerprint (engine/OCR/doc-server switches invalidate the cache automatically);
   • INDEX.md (cache root) — the aggregated list of every spilled document in this workspace.
2. the message carries only a few-hundred-token index card: page/line/char counts, an outline (PDF heading heuristics, md headings, JSON first-level key tree) and reading pointers.
3. the model reads page-by-page with the stock read tool (offset/limit, line numbers as coordinates) — full summaries read through (no dropped tails), targeted lookups jump by outline; missing content is an explicit tool failure, never silent loss.

Design rationale and evidence: docs/design-longdoc.md; comparison with similar work: docs/alternatives.md. Upgrades for current limitations (researched GitHub solutions and v0.6 roadmap): docs/upgrade-v6.md.

Engines & OCR (v3)

• PDF text engine: auto (default) → the venv's pymupdf4llm for ≤40 pages (high-fidelity tables/headings); pdfjs (seconds) for larger documents or when the venv is missing. Env: DSH_ATTACH_ENGINE=auto|python|builtin.
• Scanned-PDF OCR: python (PyMuPDF, needs system tesseract) → tesseract.js (pure JS; first use downloads the ~24MB eng/chi_sim language data into vendor/tessdata/). Confidence below 45 falls back to page images with a clear reason. Env: DSH_ATTACH_OCR=auto|baidu|tesseract-js|off (see below).

Fidelity & format coverage

• DOCX tables: mammoth HTML → turndown + GFM plugin — tables survive as Markdown pipe tables (replaces the old cell-by-cell reading order).
• TIFF: decoded by sharp (libvips prebuilt binary) into PNG pages, multi-page supported (≤20 pages per file).
• epub / odt / rtf: pandoc (probed on PATH) converts to Markdown; without pandoc, epub/odt fall back to in-process jszip + turndown, rtf reports a clear install hint.
• Legacy .doc / .xls / .ppt: LibreOffice headless (soffice, probed on PATH plus the usual Windows install locations) converts to the modern OOXML format first, then the standard Office pipeline runs. Each run uses an isolated UserInstallation profile to avoid lock conflicts.
• PDF outlines: bookmark TOCs (get_toc / pdfjs getOutline) now feed the index card's outline first; the font-size heuristic is only the fallback. Empty-bookmark PDFs are unaffected.

Cloud OCR & content-adaptive engine (zero new heavyweight deps)

• Baidu OCR API (preferred scanned-PDF OCR, free tier: 1,000 calls/month for personal accounts / 2,000 for enterprise on both 标准版 and 高精度版, per the official free-quota page): pages are sent as JPEG via plain HTTPS — no new dependencies. Configure via env:
  • BAIDU_OCR_API_KEY / BAIDU_OCR_SECRET (console → 文字识别 → create app);
  • DSH_ATTACH_OCR=auto|baidu|tesseract-js|off (auto = Baidu when credentials exist, else local tesseract.js);
  • DSH_ATTACH_OCR_ACCURATE=1 for the high-accuracy endpoint (separate free quota). Quota exhausted / API failure → automatic fallback to local tesseract.js with a note; forced baidu mode reports the reason instead.
• Remote VLM OCR (optional, token-billed): DSH_ATTACH_VLM_BASE / DSH_ATTACH_VLM_MODEL (+ optional DSH_ATTACH_VLM_KEY) point at any OpenAI-compatible vision endpoint (olmOCR-2, GLM-4V, Qwen-VL…). Pages are transcribed one by one via chat/completions. OCR chain: Baidu → VLM → tesseract.js (or force with DSH_ATTACH_OCR=vlm).
• Content-adaptive PDF engine: documents of 41–160 pages now let the Python engine decide by vector density (sampled get_drawings) — text-heavy manuals skip the slow high-fidelity pass and go straight to the fast pdfjs engine, while table/graphic-heavy documents still get pymupdf4llm. ≤40 pages are unchanged.

External doc server, cache page & workspace zero-copy

• External document parser (optional): DSH_ATTACH_DOC_SERVER=<base URL> points at a parser service (PP-StructureV3 paddleocr serve, MinerU, or any shim). Contract: POST {base}/convert with multipart field file → { "ok": true, "markdown": "..." }. When configured, PDFs go to the server first; any failure falls through to the local engine chain.
• Attachment cache settings page: Settings → 附件缓存 lists every spilled document (size/engine/time) with per-item delete and clear-all, backed by GET /api/attach-formats/cache + POST .../cache/delete + POST .../cache/clear.
• Workspace zero-copy: text files between 512KB and 16MB are first resolved against the session workspace: the browser reads the file locally to compute its full SHA-256, then GET /api/attach-formats/resolve asks the host to confirm a same-source file by name + size + full SHA-256 (bounded ~2.5s walk skipping dependency dirs). A hit mounts a 📎 reference card — the content is not uploaded (only the name, size and hash are sent); the model reads the path with its read tool. A miss falls back to the normal upload pipeline. Files over 16MB are rejected outright (no zero-copy attempt).

Context adaptation & full-text command (v2b)

• Adaptive merge limit: the client reads the token-meter contextPressure projection (model context window × current usage); the full-text merge limit becomes min(80k chars, headroom × 1.5) — when headroom is short, the card automatically turns into an index card with a status-bar note, so merged content can never blow up the context and get silently truncated by the API. A missing projection falls back to the fixed 80k threshold.
• /attach command (composer slash menu, host-registered):
  • /attach list — list the spilled documents in this workspace (id/name/size/engine);
  • /attach full <id|name> — merge the full text into model context as a next-step message (takes effect on the next message, current turn untouched); 300k-char cap with an explicit truncation notice — never silent loss. read still works afterwards for line-precise lookup.

Interactions

• Paperclip button: composer tool row (conversation.input.left), opens a multi-select file picker whose accept list covers every format in the table above.
• Drag & drop: drop a PDF / Office / text file anywhere on the page.
• Paste: copy a file and Ctrl+V into the composer (or the whole page).

Native image drag/paste stays on the harness built-in pipeline; when a single drop mixes other formats in, the plugin takes over the whole batch (converts first, then hands the produced images back to the built-in draft rail as a "synthetic drop").

Architecture

dsh-attachment-formats/
├── lib/
│   ├── index.js          # host half: POST /api/attach-formats/convert + engine routing
│   ├── client.js         # browser half: button/drop interception/synthetic drop/text injection/status bar
│   ├── cache.js          # workspace .dsh-attachments spill/manifest/INDEX.md/cleanup
│   ├── py/pymupdf4llm_convert.py  # venv high-fidelity engine (subprocess call)
│   └── convert/
│       ├── util.js       # magic-byte sniffing (pdf/tiff/OLE/rtf/zip), base64, truncation
│       ├── provider.js   # engine/binary detection (venv python, pandoc, LibreOffice) + subprocess bridges
│       ├── pdftext.js    # pdfjs text-layer extraction: line assembly/header-footer dedup/bookmark TOC
│       ├── outline.js    # md heading outline, JSON first-level key tree
│       ├── ocr.js        # tesseract.js OCR (traineddata download cache/confidence)
│       ├── pdf.js        # pdfjs-dist + @napi-rs/canvas → PNG/JPEG pages
│       ├── docx.js       # mammoth HTML → turndown+GFM → Markdown (tables preserved)
│       ├── xlsx.js       # exceljs → tab-separated text
│       ├── pptx.js       # jszip + a:t text runs → per-slide text
│       ├── tiff.js       # sharp (libvips) → PNG pages
│       ├── pandoc.js     # pandoc → Markdown + epub/odt zip fallback
│       └── libreoffice.js # legacy .doc/.xls/.ppt → modern OOXML
├── .venv/                # (optional) pymupdf4llm engine (generated by setup, not committed)
├── vendor/tessdata/      # OCR language-data cache (downloaded on first use, not committed)
├── docs/                 # design-longdoc.md / alternatives.md / upgrade-v6.md
├── scripts/smoke-*.mjs   # five offline smoke suites (converters/router/client/OCR/P0)
└── cordis.patch.yml

• The host route re-sniffs magic bytes and never trusts the client-declared kind; 160MB request cap and 64MB per-file cap; cwd is read by the client from session state and sent with the request (it decides where the spill lands).
• Tiered thresholds: full-text merge cap 80k chars (v2b lowers it adaptively by context headroom); spill page images ≤100 pages (1100px wide; PNG over the per-image byte budget falls back to JPEG); scanned-page image cap follows the deployment limit; OCR ≤20 pages per run (2000px wide), confidence <45 falls back to page images.
• Document-card content is merged into the React controlled input through a DOM event bridge at send time (same path as the native submit); the image path is fully independent and untouched.
• Conversion progress/errors show in a temporary status bar above the composer (conversation.input.dock); success auto-hides after 6s, errors can be dismissed.

Installation

From GitHub (recommended):

dsh plugin --profile web add github:linkingoscar/dsh-attachment-formats

Local development:

cd path\to\dsh-attachment-formats
npm install            # host dependencies (first time)
# optional: high-fidelity PDF engine (pymupdf4llm, self-contained venv)
python -m venv .venv
.\.venv\Scripts\python.exe -m pip install pymupdf4llm
npm run smoke          # offline smoke tests (optional)
dsh plugin --profile web add link:path\to\dsh-attachment-formats

Restart dsh web (close the page → the desktop shortcut auto-restarts, or re-run dsh web) and refresh the browser. OCR language data downloads automatically on the first scanned-PDF recognition (≈24MB, cached in vendor/tessdata/, offline-ready afterwards).

Known limitations

• OCR (tesseract.js) quality is limited on low-resolution scans and complex tables; insufficient confidence falls back to page images with an explicit note — garbled text is never injected. Higher-quality OCR (RapidOCR/MinerU/PaddleOCR) can be added as pluggable backends later (see docs/upgrade-v6.md).
• The pymupdf4llm high-fidelity engine handles ≤40-page PDFs only (larger documents use the fast pdfjs engine); table/formula reconstruction is good but not typesetting-grade — layout details can be cross-checked against page images.
• Scanned PDFs without a text layer can only go the page-image route when OCR is unavailable or fails (vision models can read them).
• Legacy .doc/.xls/.ppt require LibreOffice (soffice); rtf requires pandoc; epub/odt work out of the box but pandoc (if installed) gives better fidelity. Missing binaries produce clear, actionable errors — nothing is silently dropped.
• DOCX formulas and embedded images are not extracted (tables, headings and text are).
• XLSX outputs displayed text/results only; charts and comments are not extracted.
• Outlines prefer bookmark TOCs; PDFs without bookmarks fall back to font-size heuristics (weak on documents without strong heading styling) — the index card still carries line/page counts and reading pointers.
• iWork and archives are not converted yet.
• Attachments are attributed to the shell's current conversation (the one being viewed). Text/document cards therefore land in the dialog you are looking at. Converted page images go through the harness's native drop pipeline: if the current conversation is mid-reply it temporarily refuses drops, so the plugin waits for it to become idle before feeding the images. With several conversations open at once, other idle conversations may also accept that same synthetic drop — a harness-level behavior the plugin cannot scope; prefer attaching images with a single conversation open (text/code files are unaffected: they always stay in the current dialog).
• The "merge on send" for document cards bridges into the React controlled input over DOM events — an adaptation to an unpublished harness API; if a core upgrade breaks it, the symptom is "card content didn't enter the message", and the card's send button is the fallback (synthetic Enter path). The image path is never affected.

Releases

• v0.6.4 (latest) — session-correct attachments & verified zero-copy: attachments now attribute to the shell's current conversation (no more cards/images landing in another dialog); converted images wait for the current conversation to become idle before the synthetic drop; workspace zero-copy is confirmed by name + size + full SHA-256 (no silent substitution), >16MB is rejected outright; INDEX.md cells are escaped, INDEX rebuilds are serialized per workspace, cache hits keep the source-count fields, legacy-Office manifests carry the libreoffice+builtin engine label.
• v0.6.3 — cache lifecycle hardening: v0.6.1 8-hex cache dirs are now swept by cleanup/clear (no invisible orphans), JSON spill keeps source vs artifact sizes separate (tiering uses the spilled doc.* size), page images materialize lazily when a cache hit downgrades to index mode, INDEX.md is fully rebuilt from live manifests (no ghost rows, populated timestamps), legacy .doc/.xls/.ppt cache keys use the original OLE bytes so hits skip LibreOffice, atomic manifest/INDEX writes.
• v0.6.2 — cache correctness & fast path: 16-hex cache ids with full SHA-256 in the manifest, converter-policy fingerprint (engine/OCR/doc-server switches invalidate the cache), index cards rebuilt from structured metadata on every hit (no filename bleed-through), TTL counts model read access via file atime, page images rendered lazily (clean small PDFs skip rasterization), 2–16 MB text files reach the host spill instead of being rejected, React key warnings eliminated, Node >=20, CI actions upgraded to v7.
• v0.6.1 — correctness & engineering fixes: attachment-dock crash fix (useCallback reference), converters no longer pre-truncate (never-silent-truncation restored end-to-end), session-derived workspace authority for all routes, XLSX empty-column coordinate fix, true conversion cache keyed by source hash, cache TTL based on last access, verified merge into the composer draft; added ESLint, CI (Node 20/22) and component-level smoke tests.
• v0.6.0 — fidelity & format coverage (DOCX tables, TIFF, epub/odt/rtf, legacy Office, PDF bookmark outlines), Baidu OCR API + remote VLM OCR + external doc server, content-adaptive engine, attachment cache settings page, workspace zero-copy references.
• v0.5.0 — document cards, index-card spill, /attach list|full, adaptive merge limit, pymupdf4llm/pdfjs engines, tesseract.js OCR.

License

Apache-2.0 © 2026 linkingoscar