dsh-web
zhu1090093659/dsh-webDeepSeek Harness(DSH)Web 插件聚合生态包 · 一切皆插件,创意工坊分发(dsh-market.com)
DeepSeek Harness plugin: fixes dsh-ui table cells painting text over the next column, and wraps long cells instead of forcing a horizontal scrollbar.
A DeepSeek Harness plugin that repairs dsh-ui table rendering:
It applies to the shipped @changfenhuang/dsh-genui 0.11.x and does not wait for an upstream release. Once upstream ships its own fix, this plugin stays harmless and can simply be removed.
dsh-ui table cells default to white-space: nowrap (the plugin's deliberate "data columns do not fold" voice). Meanwhile the host shell renders message bodies inside a .markdown subtree, and two rules from that sheet land on dsh-ui output because the components render in the same subtree:
.markdown { overflow-wrap: anywhere; } /* inherited by every cell */
.markdown :not(pre) > code { display: inline-flex; } /* every chip becomes an atomic inline box */
Add the plugin's own .inlineCode { white-space: pre-wrap } and the chip's automatic minimum size collapses: the column may be sized below its own content, while the cell's nowrap still paints the text unbroken — straight over the neighbouring column.
Only cells containing inline code are affected. A pure-text cell is nowrap, so its min-content already equals its content width and the column can never be sized below it.
Measured on a live DSH desktop page (7 tables): on 0.11.4, 4 tables and 12 cells overflowed, the worst single cell by 704px.
/* 1) Table chips must not re-open line breaking — this is the root cause. */
[class*="tableWrap"] table [class*="inlineCode"] { white-space: pre !important; }
/* 2) Single-line data cells wrap — long content no longer stretches the table. */
[class*="tableWrap"] table td:not([class*="tdMultiline"]):not([class*="tdCode"]),
[class*="tableWrap"] table th:not([class*="tdMultiline"]):not([class*="tdCode"]) {
white-space: normal !important;
overflow-wrap: anywhere;
}
The selectors use only CSS Modules class-name prefixes (tableWrap, inlineCode, tdMultiline, tdCode), never the build-time hash, so they survive plugin upgrades.
Rule 2 is a preference. Remove that block from lib/client.js if you would rather keep the upstream behaviour of a wider table plus horizontal scrolling.
The same browser fixture, in an 801px container:
| Styles | Table width | Horizontal scroll | Cross-column bleed |
|---|---|---|---|
| upstream 0.11.4 (unfixed) | 1471px | yes | 4 cells, worst 306px |
upstream main (fix #282) | 1777px | yes | none |
| this plugin | 799px | no | none |
In other words: the upstream repair makes the table wider and asks you to scroll; this plugin lets the long content wrap, so the table fits the conversation column.
The plugin is an ordinary DSH bundle: a dependency in a profile, plus one entry in that profile's bundle list.
Add it to the profile you want it in — ~/.dsh/profiles/<profile>/package.json:
{
"dependencies": {
"@adamcjm/dsh-genui-bugfix": "^0.1.0"
},
"dsh": {
"profile": {
"bundles": ["…", "@adamcjm/dsh-genui-bugfix"]
}
}
}
Install it:
cd ~/.dsh/profiles/<profile> && pnpm install
Restart the DSH app. The host-side plugin tree is assembled when the process starts, so a running instance will not pick up a new bundle.
If pnpm refuses the freshly published version because of a minimum-release-age policy, add the package name to minimumReleaseAgeExclude in that profile's pnpm-workspace.yaml.
Use a link: path instead of the version, for example "@adamcjm/dsh-genui-bugfix": "link:/path/to/dsh-genui-bugfix", then run the same two steps.
After a restart the renderer carries one extra stylesheet:
document.getElementById('dsh-genui-bugfix/table-overflow') // non-null once loaded
Visually: tables with long inline-code runs no longer print text over the next column, and their width matches the conversation column instead of scrolling.
Without DSH running you can still check both halves against the loader contract:
cd <plugin> && npm test
@adamcjm/dsh-genui-bugfix from the profile's dependencies and bundles;pnpm install;main already fixes the overlap itself (PR #282, not yet released). Rule 1 here has the same effect and coexists with it — both keep the chip at white-space: pre..markdown cascade. See issue #285 and PR #286.@adamcjm/dsh-genui-bugfix/
├── package.json # dsh.bundle.patch + dsh.client declarations
├── cordis.patch.yml # inserts one plugin row into the profile
├── lib/
│ ├── index.js # host half: node ESM, exports name/apply/inject, no behaviour
│ └── client.js # browser half: window.__ModuleLoader__.load({ id, factory })
├── test/verify.mjs # zero-dependency self-check of the loader contract
└── README.md
Both halves follow the shapes already used by shipped plugins:
name / apply / inject — the same shape as the host half of dsh-window-drag-probe, which also has no host-side behaviour;window.__ModuleLoader__.load({ id, factory }), where factory returns the exports carrying apply — the same shape as the client artifacts of @changfenhuang/dsh-genui and dsh-boot-animation. A plain ESM export function apply is not loaded, and it fails silently rather than reporting an error. This plugin has no runtime dependencies, so its factory never needs require.MIT