LGPL-3.0 TypeScript DSH bundle plugin that gives vision, compaction, review, title and image-generation calls their own model routes, with reconcile-based live reconfiguration and no change to the main session route.
DSH integration
Native runtime
Author-claimed
Safety audit
Unaudited
Last verified
2026-11-16
License
LGPL-3.0
01What can it help you accomplish?
Move one narrow call category — context compaction, session titles, approval reviews — onto its own cheap or capable model, without switching the model that runs the conversation
Those calls land on the provider/model pair set in the plugin card; only provider/model are swapped, so timeouts, retries, fallback and the main session route stay harness-owned
People running an expensive main model who want compaction summaries and session titles on a cheap route, or auto-review on a stronger one
Keep pictures usable in a session whose main model only accepts text
The image passes admission, the text-only route receives a durable `[image: {…}]` text reference, and the model resolves it by calling `describe_image`; image-capable routes still receive the original `ImageBlock`s
Sessions on text-only main routes where screenshots and pasted images still have to be understood
Produce or edit images from inside the session, through an images API the profile already authorises
PNG files under `.dsh/generated/` in the working directory, returned as paths the main model can check with `inspect_image`
Profiles with an OpenAI-compatible images provider configured, where illustrations should be created and verified in the same conversation
Retune any route — provider, model, handoff, compression threshold — from the web UI while the profile keeps running
The next requests use the new route: each feature's `reconcile*()` pair registers or disposes exactly the pieces whose conditions changed
Anyone adjusting cost against quality per feature on a live profile instead of restarting it
02How to install into DeepSeek Harness
Prerequisites
- Node.js 24.9+ and dsh-loader 2.x — the plugin consumes the 2.x facades
- A DSH release in one of the six explicit families: `0.1.5`, `0.1.6`, `0.1.7`, `0.2.0`, `0.2.1`, `0.2.2`, prereleases included
- A profile to install into — the README's install block takes `<PROFILE>`, or asks which one
- Provider/model route IDs already configured on the official Models page; the plugin card only points at existing routes
- A vision-capable model on the chosen vision provider, and for `generate_image` a model whose upstream endpoint actually generates images
Installation steps
- 01
Add the dependency to that profile: `dsh plugin --profile <PROFILE> add @dsh-plugin/dsh-auxiliary` (or the equivalent plugin-manager command for the profile)
$ dsh plugin --profile <PROFILE> add @dsh-plugin/dsh-auxiliary
- 02
From 0.4.1 the package is a bundle plugin, so the add also appends it to `dsh.profile.bundles` — no manual `cordis.patch.yml` insert
- 03
If pnpm stops with `ERR_PNPM_IGNORED_BUILDS`, approve the `prepack` build in the profile's `pnpm-workspace.yaml` (`allowBuilds`) and retry the add
- 04
Remove any pre-0.4.1 manual insert row for this plugin from `cordis.patch.yml`, otherwise the bundle layer and the row mount it twice
- 05
Confirm `node_modules/@dsh-plugin/dsh-auxiliary` holds `lib/index.js` and `lib/client.js`; if the artifacts are missing, run `npm run build` in the plugin directory and re-add
$ npm run build
- 06
Open the loader-managed plugin configuration, enter the provider/model pairs, and mark the models to use under Settings → Models
Verify the integration
- `node_modules/@dsh-plugin/dsh-auxiliary` contains a built `lib/` directory with at least `lib/index.js` and `lib/client.js`
- Ask the agent to run `inspect_image` on a Host-readable path — the README's example is `Use inspect_image to analyze screenshots/error.png`
- The install block ends with verify only: do not start the profile, report what changed instead
03DSH integration and capability boundaries
DSH bundle plugin on the harness LLM seam: `dsh plugin --profile <PROFILE> add @dsh-plugin/dsh-auxiliary` mounts it through `dsh.bundle.patch`, its `llm/stream` waterfall listeners reroute purpose-keyed calls through `ctx.llm`, and `ctx.tools.register` adds the `inspect_image`, `describe_image` and `generate_image` tools
Purpose-keyed model routing
a provider/model pair entered in the plugin card, pointing at a route already configured on the Models page→the matched call re-enters `ctx.llm.stream()` with a frozen replacement config; every other call is passed through untouched
the `llm/stream` listener stays installed once the feature is on and passes calls through untouched until a complete route existsthe replacement keeps the same route marker, so an equality check stops the router matching its own re-entryVision understanding with `inspect_image`
a Host-readable image path (absolute or workspace-relative) and an optional question→a text answer from the selected vision model, optionally truncated at `vision.maxTokens`
commits the image file through the `ctx.attachments` seamthe tool is registered only while `tool.enabled` is on and a complete vision route exists`tool.maxImageBytes` caps each file and `tool.timeoutMs` is the cooperative tool-call budgetImage handoff with `describe_image`
a chat request carrying images on a route whose declared capabilities are text-only→a durable text reference in place of the image blocks, plus a `describe_image` call that answers with a text description
a runtime wrapper claims image input for text-only main routes while handoff is active, so the admission preflight passesthe text reference replaces the image blocks in that request, and nested images in tool results are rewritten toothe rewrite is request-local: aborts, errors, early iterator closure and plugin disposal keep the host stream semanticsCompaction route and optional compression engine
a provider/model pair in the Context compaction group, plus `engine.enabled` and `engine.compressPrompt` for the engine→every `purpose: 'compaction'` summarization call served by the auxiliary model; with the engine on, summaries follow an explicit compression instruction and fire at `engine.thresholdRatio`
`engine.enabled: true` replaces the stock compaction backend, and the plugin skips its engine with a warning when `@deepseek-ai/dsh-compaction-basic` is loadedthe threshold must stay above the retention ratio (`engine.retainRatio`, 16% by default)Approval and session-title routes
provider/model pairs in the Approval model and Title model groups→official auto-review calls and `purpose: 'session-title'` calls answered by the chosen models, leaving the provider's own config and the main session route untouched
only provider/model are replaced; policy, inputs, Session attribution, cancellation and decisions stay nativethe official auto-review plugin and its Auto permission mode still have to be enabled separatelythe session-title provider keeps its own deployment-level config; only its `purpose: 'session-title'` calls moveImage generation with `generate_image`
a prompt, optional `size` and `n`, and an optional `reference` — one path or an array of paths for multi-reference edits→PNG files under `.dsh/generated/` in the working directory, returned as paths
calls `POST {baseURL}/images/generations` or `POST {baseURL}/images/edits` on the provider's OpenAI-compatible images APIresolves `apiKeyEnv` through the harness credential seam on every call — never `process.env`writes the resulting images under `.dsh/generated/` in the working directory
04Who is it for? When not to use it?
Good for
- People running an expensive main model who want compaction summaries and session titles on a cheap route, or auto-review on a stronger one
- Sessions on text-only main routes where screenshots and pasted images still have to be understood
- Profiles with an OpenAI-compatible images provider configured, where illustrations should be created and verified in the same conversation
- Anyone adjusting cost against quality per feature on a live profile instead of restarting it
Not for
- Compatibility is pinned to Node.js 24.9+, dsh-loader 2.x and six explicit DSH release families — `0.1.5`, `0.1.6`, `0.1.7`, `0.2.0`, `0.2.1`, `0.2.2`, prereleases included. The README lists no other versions.
- `engine.enabled: true` replaces the stock compaction backend, so it must not run alongside `@deepseek-ai/dsh-compaction-basic`: the plugin detects the conflict and skips its engine with a warning, which leaves the custom compression prompt inactive.
- The `Allow image input` checkbox only writes a model's canonical `input` declaration — it cannot add vision to a text-only model, so it is safe only when the upstream endpoint really accepts images.
- Preferences live at `<profile>/config/dsh-auxiliary/config.json`; the old 1.x `dsh-auxiliary` settings namespace is not imported automatically, so previous values have to be copied by hand, and manual file edits need a plugin reload or restart.
05Compatibility, maintenance and safety notes
- Compatibility is pinned to Node.js 24.9+, dsh-loader 2.x and six explicit DSH release families — `0.1.5`, `0.1.6`, `0.1.7`, `0.2.0`, `0.2.1`, `0.2.2`, prereleases included. The README lists no other versions.
- `engine.enabled: true` replaces the stock compaction backend, so it must not run alongside `@deepseek-ai/dsh-compaction-basic`: the plugin detects the conflict and skips its engine with a warning, which leaves the custom compression prompt inactive.
- The `Allow image input` checkbox only writes a model's canonical `input` declaration — it cannot add vision to a text-only model, so it is safe only when the upstream endpoint really accepts images.
- Preferences live at `<profile>/config/dsh-auxiliary/config.json`; the old 1.x `dsh-auxiliary` settings namespace is not imported automatically, so previous values have to be copied by hand, and manual file edits need a plugin reload or restart.
LGPL-3.0 · no release yet · active, last push 2026-11-13
06Frequently asked questions
How do I install dsh-auxiliary into DeepSeek Harness?
Run `dsh plugin --profile <PROFILE> add @dsh-plugin/dsh-auxiliary` in the profile you want (or let the agent ask which one). Since 0.4.1 the package is a bundle plugin that declares `dsh.bundle.patch`, so the add also appends it to `dsh.profile.bundles` and no manual `cordis.patch.yml` row is needed; approve the `prepack` build in `pnpm-workspace.yaml` (`allowBuilds`) if pnpm stops with `ERR_PNPM_IGNORED_BUILDS`, and remove any pre-0.4.1 insert row so the plugin is not mounted twice.
Does it change the model that runs my main conversation?
No. Every route matches exactly one call category — `purpose: 'compaction'`, `purpose: 'session-title'`, the official auto-review call, or the tools it registers — and re-enters `ctx.llm.stream()` with only provider/model replaced. The listeners are pure pass-throughs until a complete route exists, so an unconfigured feature changes nothing.
Where are the routes configured inside a DSH session?
In the loader-managed plugin configuration: one collapsed card groups vision, compaction, approval, title, image generation and the compression engine. Enter provider/model IDs already configured on the official Models page; saves apply to subsequent requests without a restart. Preferences live at `<profile>/config/dsh-auxiliary/config.json`, the old 1.x settings namespace is not imported, and manual file edits need a plugin reload or restart.
Which DSH versions and runtimes does it support?
Node.js 24.9+ and dsh-loader 2.x, with six explicit release families — `0.1.5`, `0.1.6`, `0.1.7`, `0.2.0`, `0.2.1`, `0.2.2`, prereleases included. Before `0.1.6-alpha.2` the form lives under Settings → Plugin configuration; from that release it is on the plugin detail page.
What happens to images when the main model is text-only?
With Image handoff on (the default), a runtime wrapper on `ctx.llm.resolveModelInfo` lets the image pass admission, the text-only request receives a durable `[image: …]` reference, and the model calls `describe_image` to get a description from the vision route. Image-capable routes keep their original image blocks. Disable `vision.handoff` to restore the original admission and stream behaviour, and remember that the `Allow image input` checkbox cannot add vision to a text-only model.
07Related DSH workflows
dsh-agent-teams
by nanmicoder
AgentTeams 多智能体团队协作插件:自然语言驱动建团队、拉成员、拆任务声明依赖、成员间消息直达,Web 右上角活动面板
dsh-blender-plugin
by sixtysevenlf
Adds tools for Blender scene editing, viewport capture, render profiling, headless work, and verification.
dsh-free-search
by dddmuc
DSH 免费搜索插件,执行联网检索。
dsh-bili-scroll
by c1overzz
Schedules Bilibili discovery, reads subtitles, writes daily reports, and optionally posts a comment tagging you.
08Data and sources
`dsh-auxiliary` is a [DeepSeek Harness](https://deepseek-harness.github.io/deepseek-harness/) plugin that layers auxilia…
Add the plugin dependency: `dsh plugin --profile <PROFILE> add @dsh-plugin/dsh-auxiliary` (or the equivalent plugin-mana…
Since 0.4.1 the package is a bundle plugin (declares `dsh.bundle.patch`), so `dsh plugin add` also appends it to `dsh.pr…
Tools are registered with `ctx.tools.register(defineTool(...))` and announced to the model through `ctx.systemPrompt.sec…
This page is generated from the project’s public documentation, repository metadata and a structured parse of DSH Plugins; last verified on 2026-11-16. Found an error? Submit a correction.
Best DeepSeek Harness Plugins
Twelve plugins worth installing first — picked from the whole catalog, across every category.
