dsh-finalize

English | 简体中文

A DeepSeek Harness plugin that runs one finalize pass the moment a file is delivered — and nowhere else. It strips the fingerprints document libraries leave behind, recomputes the statistics that are actually derivable, and puts everything it will not fabricate into a side panel instead of writing it into the file.

Design rule: fix what is provably wrong, never invent evidence. Edit time, revision history, comments and "was written by a human" claims are reported, not written. Application is removed rather than replaced with a fake Microsoft Office Word 16.0 string — absence is unremarkable, a lie contradicts the rest of the package.

Why this exists

Documents produced by python-docx, python-pptx, openpyxl or LibreOffice carry their own birth certificate: a creator field naming the library, a 2013 template date, Words=0 on a forty-page report, a leftover template thumbnail. Those are bugs, not style choices, and anyone who unzips the file can see them.

The usual advice — forge the edit time, invent a revision count — makes things worse. It contradicts the fields you cannot change (the rsid trail, the real save history) and collapses the moment someone looks. This plugin takes the honest half: normalize what is wrong, report what only a human can supply.

What it does

Format	Automatic fixes	Reported, never fixed
docx	creator and description, template date, Application/AppVersion, word/character/paragraph/line counts, template thumbnail	TotalTime=0, page count, punctuation, unresolved comments, unaccepted revisions, hidden and white text
pptx	same core properties, slide/notes/hidden-slide counts, word counts	no speaker notes, cropped images still in ppt/media, hidden slides and shapes, comments, speaker notes
xlsx	core properties, TitlesOfParts aligned to sheet names	zero formulas, pivot cache retaining deleted rows, external links, hidden sheets, rows and columns, comments
odt ods odp	meta:generator, creator, dates, meta:document-statistic, thumbnail and its manifest entry	meta:editing-duration=PT0S, single save cycle, annotations, hidden text and sheets
pdf	nothing — read-only by design	Producer/Creator/XMP, incremental updates, missing bookmarks, signatures, punctuation in extractable text

Several files in one delivery are processed independently. Several deliveries in one session merge into a single report for 30 minutes instead of overwriting each other.

Text checks

When a document is delivered, its paragraphs are also checked for punctuation that does not match their language. The language is detected per paragraph from script counts (Han / Kana / Hangul / Latin), and Latin paragraphs are separated into English, French and German by diacritics and stop words.

Language	Expected	Flagged example
Chinese	，。！？；：、（）“”	这是测试,包含逗号
Japanese	、。「」『』	テスト，です
Korean	ASCII , . ? !	테스트，입니다
English / French / German	ASCII, plus each language's own quotes (« », „ “)	sentence，with

Also flagged: spaces before punctuation, three or more repeated marks, and unbalanced “ ” / 「 」 / 『 』.

Findings are reported, never rewritten — a blind character swap damages code, URLs, formulas and citations far more often than it helps. In the panel every finding group can be expanded into its individual occurrences (up to 50, each with the original snippet and the suggested replacement). URLs, e-mail addresses, paths, decimals and list numbers (1., 3.2) are skipped entirely; spreadsheet cells are only checked when longer than eight characters.

PDF text is decoded from the text layer: Flate content streams, Tj/TJ operators and per-font ToUnicode maps, including Identity-H CID fonts, so Word and LibreOffice exports are covered. Scanned pages, fonts without a ToUnicode map and text drawn as outlines cannot be read — those files are reported as not checked rather than guessed at.

Only your own text is judged. When the delivered document reuses paragraphs from a file the user supplied, those paragraphs are treated as the template: they are never listed as problems, and their punctuation style becomes the accepted style for new text — a template that writes 一: makes the assistant's 二: correct, not an error. Text the user edited themselves falls into the same bucket: noted, never changed.

Detection is automatic: the newest sibling Office file that is older than the delivery and shares at least a quarter of its paragraphs (minimum three) becomes the baseline. textCheck.templates pins explicit paths, textCheck.baseline: false turns detection off.

Review state

A document carries more than its visible pages. Before delivery the plugin reports what still travels with the file:

Finding	Detail
Unresolved comments	count, authors, date and text of each comment (docx, pptx, xlsx, odf)
Unaccepted revisions	count, authors, and the text that was deleted (docx)
Hidden content	hidden text, white text, hidden slides and shapes, speaker notes, hidden sheets, rows and columns

Every finding is listed per file and expands into individual items in the panel. Nothing is deleted or accepted automatically — whether a comment is an internal note or an intentional annotation is your decision. A clean document gets one line: 审阅状态：无批注、无修订、无隐藏内容.

Text shading

Two different things hide behind "the text has a background":

• Highlight (w:highlight, a:highlight) — the marker pen. A deliberate author choice: never removed, always reported so you can confirm it.
• Run shading (w:shd w:fill="…") — a background colour glued to a text run. The classic AI/paste artifact is a neutral grey (F2F2F2, E7E6E6, EDEDED, F3F3F3, EEEEEE); the plugin removes exactly those, and reports every other fill colour instead of guessing.

autoFix.stripTextShading: false turns removal off and keeps everything report-only.

Install

The plugin is a DSH bundle — a plain package with a dsh.bundle.patch; there is no build step and no runtime dependency.

git clone https://github.com/hengyun721/dsh-finalize.git
dsh plugin --profile desktop add /absolute/path/to/dsh-finalize

Or from inside the Harness, install the checkout directory with the plugin manager (install_bundle, target = the absolute path above). Restart once if it reports restart-required.

Usage

Nothing to call. Deliver a file with present (or a configured transport) and the pass runs before the tool executes, so the file on disk is already normalized when it reaches the recipient. The text check runs only for delivered documents — ordinary chat replies are never scanned.

Relative paths are resolved against the session's working directory, not the Host process directory. A file that cannot be read is reported as a failure in the panel, counted in the badge, and shown with the original path next to the resolved one — it never disappears silently.

A shield icon appears in the message action row next to copy / like / dislike / branch:

State	Meaning
grey	no delivery in this session yet
amber with a count	that many items need a human
click	panel with fixed / open / read-only rows per file

The panel also offers Scan recent files, which runs the pass on the newest Office files in the working directory without a delivery — handy when a document was produced by a script rather than through present.

Configuration

- id: dsh-finalize
  name: 'dsh-finalize'
  config:
    author: ''
    triggerTools: [present, qqbot_send_file]
    formats: [docx, pptx, xlsx, pdf, odt, ods, odp]
    scanLimit: 5
    autoFix:
      stripGeneratorFields: true
      dropApplicationField: true
      recomputeStats: true
      stripTextShading: true
    textCheck:
      punctuation: true
      languages: [zh, ja, en, fr, ko, de]
      templates: []
      baseline: true
    report:
      surface: panel

Key	Type	Default	Description
author	string	''	Written into dc:creator, lastModifiedBy and meta:initial-creator. Empty clears generator names without inventing a person.
triggerTools	string[]	present, qqbot_send_file	Delivery tools that trigger the pass.
formats	string[]	all seven	Adapter whitelist; anything else is reported as skipped.
scanLimit	number	5	How many recent files Scan recent files processes.
autoFix.stripGeneratorFields	boolean	true	Remove library names and generated by … descriptions.
autoFix.dropApplicationField	boolean	true	Remove Application/AppVersion instead of faking an Office version.
autoFix.recomputeStats	boolean	true	Recompute word/character/paragraph counters from the body.
autoFix.stripTextShading	boolean	true	Remove neutral grey/blue run shading left by paste or conversion. Highlights and other fills are only reported.
textCheck.punctuation	boolean	true	Run the punctuation check on document text.
textCheck.languages	string[]	all six	Language codes the checker may flag: zh, ja, ko, en, fr, de.
textCheck.templates	string[]	[]	Explicit template paths, relative to the delivered file.
textCheck.baseline	boolean	true	Auto-detect the template by paragraph overlap when no explicit path is set.
report.surface	panel | silent	panel	Whether the action-row button renders.

Repository layout

Path	Responsibility
index.js	Host half. Registers the tools/pre-execute hook on the delivery tools, routes each file to an adapter, merges per-session reports, and serves /api/finalize.report and /api/finalize.scan. Registers no tool and no prompt section.
client.js	Browser half. Registers the shield button into conversation.chat.assistant-actions and renders the report panel (fixed overlay, theme tokens, no DOM writes outside its own tree).
lib/office.js	Format engine, dependency-free. ZIP reader/writer (node:zlib), OOXML adapters (docx/pptx/xlsx), ODF adapter (odt/ods/odp), read-only PDF inspector.
lib/punctuation.js	Text checker. Per-paragraph language detection (zh/ja/ko/en/fr/de) and the punctuation rules; returns grouped findings with snippets and suggested replacements.
lib/pdf-text.js	Heuristic PDF text extraction: object streams, Flate content streams, Tj/TJ decoding and ToUnicode CMaps. Used only for the punctuation check; it never writes.
cordis.patch.yml	Bundle patch: inserts the dsh-finalize row and the default config. This is what install_bundle applies to the profile.
package.json	Package manifest: dsh.bundle.patch, the dsh.client web half (platform, immediately, inject), the export map, and the check / test scripts.
locale/en.json, locale/zh.json	Display metadata read by the Plugin Manager and Settings without activating the plugin: meta.title, meta.description.
icon.svg	Original plugin icon (document + check), used in the plugin inventory and bundle details.
test/smoke.test.mjs	Four node:test cases: ODF rewrite, docx rewrite, PDF inspector, non-ZIP rejection. Fixtures are built in memory; nothing on disk.
.github/workflows/ci.yml	CI on push and pull request: Node 20, npm run check, then npm test.
CHANGELOG.md	Version history in Keep-a-Changelog format.
LICENSE	MIT.
README.md	This document.
README.zh.md	Chinese translation.

Engine map

lib/office.js is deliberately one file with a flat API:

Symbol	Role
readZip(buffer)	Parse the central directory into { name, data, time, date } entries. No ZIP64.
writeZip(entries)	Rebuild the archive with deflateRaw. Entries named mimetype are stored uncompressed and keep their position, as ODF requires.
finalizeDocx / finalizePptx / finalizeXlsx	OOXML adapters; they share fixCore, stripApplication, dropThumbnail and commonUnresolved.
finalizeOdf	ODF adapter over meta.xml and META-INF/manifest.xml.
inspectPdf	Regex-level PDF inspector; returns observations only and never rewrites bytes.
finalizeOffice(bytes, kind, options)	Dispatcher used by the Host half.

Compatibility

Built to sit beside other plugins without changing their behaviour.

• No tool, no prompt section. Nothing is added to the model's tool list or system prompt, so token accounting and tool selection are untouched.
• One passive listener. tools/pre-execute runs only for the configured delivery tools, always returns next(), and swallows its own failures, so it can neither block nor rewrite another tool's call.
• No transport coupling. It hooks the delivery tool, not a transport, so QQ, Web and CLI transports keep working.
• Namespaced surface. Routes are /api/finalize.report and /api/finalize.scan; the slot id is dsh-finalize. Neither collides with the shipped deliverables, changes.* or presented routes.
• Opt-in debug only. Nothing is written outside the delivered files unless DSH_FINALIZE_DEBUG=1.

Limits

• Page counts need a layout engine; the plugin reports rather than guesses.
• ODF support rewrites meta.xml and the manifest thumbnail entry, nothing else. Verified structurally; opening in LibreOffice is not covered by CI.
• Legacy binary formats (doc, xls, ppt), rtf and ipynb are not implemented. doc stores a real accumulated edit time in PIDSI_EDITTIME — worth adding before reaching for anything cosmetic.
• This is not an authorship detector and does not claim a file "looks human". It removes noise and reports contradictions.

Debugging

Set DSH_FINALIZE_DEBUG=1 before starting the Harness to write dsh-finalize.log and dsh-finalize-last.json into the OS temp directory. Off by default; nothing is written to your workspace.

Nothing else touches the disk: document bytes are read into memory and written back in place, there are no temporary copies, no sidecar files and no cache. On Windows the debug files land in %TEMP% (C:\Users\<user>\AppData\Local\Temp), on macOS and Linux in $TMPDIR and /tmp.

Development

npm run check   # node --check on the three runtime files
npm test        # four unit tests, no dependencies

npm test runs the test file in-process so it also works inside sandboxed terminals where the test runner cannot spawn children.

License

MIT — see LICENSE.