model-follow-search

English | 中文

Make DSH's web_search follow your active chat model

DeepSeek Harness ships a single built-in search backend for its web_search tool — DeepSeek's official route (deepseek-v4-flash, billed to DEEPSEEK_API_KEY) — no matter which provider your conversation actually runs on. Chat through OpenRouter and every web lookup still burns DeepSeek credits.

model-follow-search registers a model-follow search provider that takes over the seam's selection pointer:

• Toggle ON — each search live-reads the active agent model (provider + model), picks the matching gateway and credential from a mapping table, and issues the same native Anthropic-compatible web_search_20250305 server-tool request. Switch models mid-session and the very next search follows. No restarts, ever.
• Toggle OFF — searches delegate to the base provider: whatever the seam pointed at before this plugin took over (the native deepseek-official, or another search plugin's provider such as a Tavily backend), falling back to deepseek-official and then to any other registered provider. With no other search plugin installed this is byte-identical to not having the plugin installed.

A toggle lives in Settings → Plugins and as a small 🔍 pill next to the conversation input. State persists in settings.yaml.

Requirements

• Your chat model's gateway must provide an Anthropic-compatible /messages endpoint AND execute the native web_search_20250305 server tool server-side. Verified: OpenRouter (stealth/ox-alpha). Most gateways merely relay completions and cannot run server-side search.
• The matching API key stored in DSH's credential store (Models page).
• pnpm, for linking the plugin into your profile workspace.

Install

Full walkthrough in INSTALL.md (中文). Short version:

1. Copy model-follow-search/ into ~/.dsh/profiles/web/.
2. Declare it in the profile workspace (package.json + pnpm-workspace.yaml templates are in INSTALL.md), then pnpm install.
3. Append the insert entry from examples/cordis-patch-entry.yml to your profile's cordis.patch.yml.
4. Restart dsh web.

Configuration

Built-in endpoint map covers openrouter and deepseek. Add any provider via the patch entry's config.endpoints:

endpoints:
  my-provider:
    baseURL: https://anthropic-compatible.example/v1
    apiKeyEnv: MY_PROVIDER_API_KEY   # resolved from the credentials store per search

How this differs from other search plugins

The ecosystem offers fixed-backend replacements (Tavily / Brave / DashScope / MiMo / GLM …). This plugin is the only one whose backend is your current chat model: same gateway, same key, same billing surface, zero extra accounts.

It also composes with those plugins as a three-tier priority chain:

magnifier	Tavily-style plugin	search runs on
ON	any state	your active chat model
OFF	installed & enabled	that plugin's provider
OFF	disabled/removed	native deepseek-official

The pointer captured at startup becomes the fallback target when the toggle is off, and is restored automatically when this plugin is stopped or removed — so installing/removing either side never breaks search.

Known limitations

• Developed against DSH internal interfaces as of 2026-08 (settings sections, slot protocol, ctx.web); a major DSH upgrade may require adaptation.
• With the toggle off you are back on the official DeepSeek route and need a funded DEEPSEEK_API_KEY.

License

MIT

---

让内置 web_search 跟随当前对话主模型

DeepSeek Harness 的内置搜索固定走官方路由(deepseek-v4-flash, DEEPSEEK_API_KEY 计费),与你在 Models 页选的对话模型完全脱钩。

model-follow-search 注册一个「model-follow」搜索提供方并接管路由指针:

• 开关开:每次搜索实时读取当前主模型,按映射表自动选用对应网关与 API key。切换模型后下一次搜索立即跟随,零重启。
• 开关关:原样委托给原生 deepseek-official,行为与未安装插件一致。

开关位于「设置 → 插件」卡片和对话输入框旁的 🔍 小图标,状态持久化。

安装(详见 INSTALL.md)

复制插件目录到 ~/.dsh/profiles/web/ → 声明进 workspace 并 pnpm install → 把 examples/cordis-patch-entry.yml 追加进 cordis.patch.yml → 重启。

前提

网关必须支持 Anthropic 兼容 /messages 且在服务端执行原生 web_search_20250305 工具(已验证:OpenRouter)。多数仅转发对话的网关不行。