@lovstudio/dsh-search-model

Search-enabled composer model seat for the DeepSeek Harness web GUI.

The built-in composer model seat (conversation.input.model) only lists provider-grouped models. This plugin takes over that single seat cell with a ModelSelect that adds a pinned search field and fuzzy matching while keeping the built-in package running underneath — the /model popup and the shared per-session model directory still come from @deepseek-ai/dsh-client-ui-model-selection.

What the takeover adds:

• Sticky search — the field is pinned above the scrolling option list, so typing filters while rows scroll and the field never leaves the viewport.
• Fuzzy matching — each whitespace-separated term matches as a plain substring, a separator-free spelling (v4pro finds deepseek-v4-pro), or an ordered subsequence within one provider/model name or id (v4p still finds deepseek-v4-pro); order is preserved, never permuted. Descriptions are intentionally not searchable (this surface never shows them), and a query with no hits shows a distinct "no matching models" message instead of the catalog-empty copy.
• Keyboard + ARIA — entering the pane focuses the search; ArrowUp/Down enter the visible rows at the correct edge; Escape is staged (clear query → back to the Model/Effort root → close). The model list owns its own role="menu" region so the textbox never sits inside a menu.

How it takes over the seat

conversation.input.model is a single slot; occupants coexist at distinct priorities and the lowest live entry renders. The built-in registers at the default priority 0; this plugin registers the same seat at priority -100, so its ModelSelect renders while the built-in package keeps providing the modelDirectories service and the /model command. No data logic is copied: the seat's injected directory/load/select face is built over the shared ctx.modelDirectories service, so the takeover shares one catalog and one session projection with the /model popup.

Removing the plugin restores the built-in seat exactly.

Install

Install the prebuilt GitHub package into the DSH web profile:

npx @deepseek-ai/dsh plugin --profile web add -w github:lovstudio/dsh-search-model

Then restart the web profile so the new client bundle is served:

npx @deepseek-ai/dsh web

For a pinned release, use github:lovstudio/dsh-search-model#v0.2.1. The repository includes its built entry points, so installation does not run a prepare script or require build approvals.

Version 0.2.1 fixes a missing remote.session service dependency that could crash the composer model seat when opening an uncached session.

Load gate

After installing, open the web GUI, click the composer model seat (e.g. DeepSeek-V4-Flash · High), drill into the model list and confirm the search field with the placeholder Search by name or ID appears above the groups. Typing v4p should narrow to DeepSeek-V4-Pro rows while other models stay hidden — that proves the takeover bundle, not the built-in seat, is rendering.