DeepSeek Harness Plugin Hub

发布与管理完整 Harness Profiles,发现适合你的插件。

探索

插件目录环境预设文档中心动态

社区

发布插件联系我们报告问题

相关链接

Plugin Hub GitHubDeepSeek Harness 官方项目系统状态隐私说明
© 2026 DeepSeek Harness Plugin HubPowered byPaxTech

独立、非官方社区项目,与 DeepSeek 官方无隶属、授权或背书关系。

Input Limit — DeepSeek Harness 插件(DSH Plugin)
DeepSeek Harness Plugin Hub
ProfilesPlugins分类动态文档登录管理 Profiles
ProfilesPlugins分类动态文档登录
← Plugins
I

dsh-input-limit

Input Limit

在 dsh Web 编写器中直接设置当前模型的输入限制(上下文窗口)。

插件会安装到这里;不确定时保持 web。

npx -y @deepseek-ai/dsh plugin --profile web add github:viyu-tech/dsh-input-limit#40f54e1f62c2884ee44bc615031b9e9236a45b0e
README兼容性版本

兼容性与来源证明

Input Limit 以 dsh-input-limit 发布,当前版本为 0.1.1。Plugin Hub 会校验它的 manifest,并保存精确安装来源,便于复现安装结果。

DSH 兼容范围
*
运行环境
web
发布来源
github
Registry 更新时间
2026/9/22

版本

0.1.1stable
2026/9/22

相关插件

正在加载相关插件…

最新版
0.1.1
DSH
*
HMR
重启进程
Tree shaking
未声明可安全裁剪
解包体积
未提供
文件数
未提供
Surface
web
许可证
MIT
发布源
github
GitHub
★ 0
周下载
0
最近提交
2026/9/22
查看源码 ↗
README Badge

点击下方 Badge 复制 Markdown,粘贴到 README 即可。

这是你的 Plugin?认领权益 · 优先安全扫描

验证 package.json 声明的 GitHub 仓库,即可管理这个公开页面。认领后,Hub 会优先安排当前版本的安全扫描,并在通过后公开展示结果。

认领这个 Plugin →
报告问题

相关插件

继续浏览 models-usage 分类下经过校验的插件。

Usage@linxin666/dsh-usage用于 dsh Web GUI 的使用统计插件:检测各提供商的余额和编程套餐配额,并实时记录令牌使用情况;侧边栏 Settings 行上方显示紧凑的价格概览卡片Client Ui Model Capabilities@linxin666/dsh-client-ui-model-capabilities为自定义 pi-ai 提供商声明每个模型的能力(图像输入和推理力度),可直接在 Models 设置卡片中编辑;写入官方 llm-pi-ai 设置命名空间,不更改 DSH 源代码。Codex Subscriptiondsh-codex-subscription通过 OAuth 在 DeepSeek Harness 中使用 ChatGPT 和 Codex 订阅,支持配额管理、安全重置、网页搜索、图像和快速模式Commandcode Provider@mars-sea/dsh-commandcode-provider适用于 Command Code 的非官方 DeepSeek Harness LLM 提供商插件,移植自 pi-commandcode-provider(MIT)。注册“commandcode”提供商路由,并在 Models 页面添加卡片和实时模型目录。

README

dsh-input-limit

直接在 dsh web 对话页设置当前大模型的输入上限(上下文窗口)。

A community plugin for DeepSeek Harness (dsh) that adds a small chip to the composer tool row — right next to the model picker — showing the current model's input limit and letting you change it on the spot, without digging through the Settings page.

What it does

  • Shows the current model's effective input limit (context window) as a tiny pill in the message input bar, e.g. 输入上限 128K (Input limit 128K). A · 默认 tag means the value is inherited from the provider default. Hovering shows the exact token count.
  • Click the pill → a popover lets you type an exact token count (one unit — e.g. 131072, or 131,072) with a live ≈ 128K equivalent shown while you type, then Save, or Reset to default.
  • Persists a per-model contextWindow override into the provider's settings section (llm-pi-ai / llm-deepseek) via the same typed RPCs the Settings page uses. The change is live on the next request — it drives context pressure display and auto-compaction, so a model whose real input limit is smaller than the adapter default (e.g. 128K vs the pi-ai default 262144) is accounted for correctly.
  • Writes go through a fresh settings read and are retried once automatically when the document moved in the meantime (settings/conflict), so two open editors never clobber each other silently.

Numbers use the binary scale model capacities are conventionally quoted in (1K = 1024, 1M = 1024²): 131072 reads 128K, and the pi-ai default 262144 reads 256K.

It works for every configurable provider (pi-ai routes, the direct DeepSeek adapter), hides itself when the current model/provider is not configurable, and is a browser-only plugin: no custom host code, no server restart required beyond loading the plugin itself.

Install

Requirements: dsh CLI (tested on 0.1.5-rc.2, works with any build that ships the web UI plugin system), pnpm (the plugin installer drives it), and a configurable provider (pi-ai or deepseek-official).

The web UI is the web profile. Installing into it and restarting dsh web is all it takes.

From the Web UI (recommended)

Once the package is on npm: sidebar Plugins → Add plugin, type dsh-input-limit, pick the registry (npm by default, npmmirror as the built-in fallback), then Install → Enable now.

From npm

dsh plugin --profile web add dsh-input-limit
dsh web

From GitHub

dsh plugin --profile web add github:<you>/dsh-input-limit
dsh web

pnpm blocks build scripts of git-hosted dependencies by default; this repo ships prebuilt lib/, so a blocked prepare is harmless. To rebuild on install instead, add the key pnpm prints under allowBuilds in the profile's pnpm-workspace.yaml and re-run.

From a local checkout / tarball

# from a directory that contains the plugin package
dsh plugin --profile web add ./dsh-input-limit
dsh web          # restart (or re-open http://127.0.0.1:3080)

Uninstall

dsh plugin --profile web remove dsh-input-limit

Build from source

lib/ ships prebuilt in this repo, so installs from a checkout or GitHub work without a toolchain. To rebuild or iterate:

pnpm install
pnpm run build     # emits lib/index.mjs (host half) + lib/client.js (browser half)
pnpm run typecheck
pnpm run test

The browser bundle is emitted in the exact closure-factory format the dsh web shell's __ModuleLoader__ consumes (platform modules externalized, everything else inlined, CSS Modules compiled with lightningcss and injected as a <style data-plugin> element) — the same contract the built-in UI plugins use.

Dependency-free sanity checks (no install needed, Node ≥ 22): node tests/bundle-smoke.mjs (loads lib/client.js under a stub __ModuleLoader__), plus node --experimental-strip-types tests/capacity.verify.mjs and tests/provider.verify.mjs (logic + an end-to-end run against a mock rc.2 client wire — the Typert ctx.remote namespaces and the ctx.sessions object layer — built from a real pi-ai route layout).

How it works

  • cordis.patch.yml mounts one plugin row; the host half (src/index.ts) is a no-op whose presence makes the Host serve the browser half.
  • src/client/index.ts registers a component into the composer's conversation.input.right list seat. Services are read through ctx.get inside the plugin fiber and the concrete settings/session faces are detached there, so no component-time callback ever touches the Cordis context proxy.
  • src/client/provider.ts resolves the session's current model from its durable modelSelection projection (ctx.sessions), maps it to the provider's settings namespace via ctx.remote.settings.describe(), reads the effective context window, and writes the override with ctx.remote.settings.mutate(ns, ops, expectedRevision) (preserving the rest of the models array). The forwarded settings/document-updated event and the session's model projection refresh open chips live.

Compatibility notes

  • Tested on dsh 0.1.5-rc.2. The client plugin declares the dotted services it dereferences (remote.settings, remote.session) in its inject list — required by the Cordis context proxy; older/newer builds that rename or drop these faces will fail the plugin load loudly rather than misbehave silently.

Development & contributing

  • Product copy is Chinese (src/client/locales.ts ships a matching English dictionary); code comments are English.
  • Design deep-dive: docs/DESIGN.md — architecture, data flow, parser rules, and the Cordis pitfalls this plugin works around.
  • Release process & official-requirements checklist: docs/RELEASING.md.
  • PRs welcome. Keep the client bundle's external list to the platform modules (react/cordis/ui-slots/...); everything else must inline.

Publish to the community

To make the plugin discoverable as an official community plugin:

  1. Create a GitHub repository (dsh-input-limit) and push this directory.

  2. On the repository Settings → Topics, add the dsh-plugin topic (topic listing) — this is what the official README designates for plugin discoverability.

  3. Optionally publish to npm:

    pnpm publish     # runs `prepare` → rebuilds lib/ and ships it
    

The full release gates, artifact definitions, compliance checklist, and post-publish verification live in docs/RELEASING.md.

License

MIT