dsh-subagent-models
Set the provider, model and reasoning effort that DeepSeek Harness subagents run with — from
one entry, the chip beside the composer. The card it opens has two scopes over the same two
DSH-style pickers: a searchable, provider-grouped model menu and that model's own effort menu.
为 DeepSeek Harness 的子代理统一指定用哪个模型、思考强度多深：只有一个入口——输入框旁的「子代理」
胶囊；点开的卡片里用「本对话 / 全局默认」两个范围切换，两处都是同一组 DSH 风格的「模型 + 思考强度」
浮层菜单（可搜索、按提供方分组、行内勾选）。
- Host half: route store +
agent/request override (index.js)
- Web half: the composer chip + a frame-wide card (
client.js)
- No build step and no settings page. The Web half uses the host's
@deepseek-ai/dsh-client-ui-primitives when present (a frozen platform module — nothing to
install), with hand-drawn .dsm-* fallbacks on hosts that do not ship it.
Screenshots / 截图
Install / 安装
# npm
dsh plugin --profile <profile> add dsh-subagent-models
# or straight from this repository
dsh plugin --profile <profile> add github:Garybot126/dsh-subagent-models
Then restart dsh once if the plugin row was not hot-loaded.
Where to set it / 在哪里设置
| Where | What |
|---|
| Composer → 「子代理 …」 chip | The only entry. Opens the card |
| Card → 「本对话」 scope | This conversation's override; the model menu's first row is 「跟随全局」 |
| Card → 「全局默认」 scope | The profile-wide default for every conversation without an override |
The chip's label is the effective route for the conversation (e.g. 子代理 DeepSeek-V4.1-Flash · high), so the composer always says what a delegation will actually use. The card's footer repeats
it with its source: 下一次委派将使用 …（来源：本对话 / 全局默认）.
The card's two rows are DSH-style cells — click one to open a menu anchored above it. The model
menu matches the built-in composer model picker: a search box for catalogs above four models,
provider groups with sticky headings, rounded hover rows and an inline check mark on the current
choice. The effort menu lists 「模型默认」 plus only the levels that model advertises.
Both menus are driven by the live LLM catalog (ctx.llm): models are grouped by provider, and the
effort list is the one that model advertises (resolveModelInfo().reasoning.efforts), so an
unsupported level is never offered.
No settings page is registered by this plugin. The built-in 设置 → 插件 → 子智能体 page belongs to
dsh itself (recursion depth, concurrency, model whitelist) and is unrelated; see below.
Precedence / 优先级
本对话 override → 全局默认 → no route (child inherits the parent session)
When it takes effect / 什么时候生效
- Applies to the next delegation — no restart, no new conversation.
- A child that is already running picks it up on its next model request.
- With no route configured, children keep inheriting the parent session's route — exactly the
behaviour of not installing the plugin.
How the route is applied / 怎么生效
The delegation tool (subagent) is composed in the agent-plane preset, not in the Profile
patch, and it carries its own static agentOptions. A plugin cannot rewrite that composition, so
the route is applied one level lower:
agent/request (waterfall: "replace the frozen call configuration")
The listener is registered once and only rewrites an Agent that another live Agent owns — a
delegated child. A top-level Session is never touched. Consequences:
- works with the official
subagent tool, without editing the preset and without adding
model-selection parameters to the tool schema;
subagent, subagent_fork and Agent Teams members all count as delegated agents.
Data / 数据
$DSH_HOME/plugin-data/subagent-models.json (falls back to ~/.dsh/plugin-data/), hot-reloaded on
mtime, so editing it by hand also works:
{
"version": 1,
"global": {
"default": { "provider": "deepseek-official", "model": "deepseek-flash", "effort": "high" },
"presets": {}
},
"sessions": {
"<sessionId>": { "default": { "provider": "...", "model": "...", "effort": null }, "presets": {} }
}
}
global.default: null means no profile default; an absent sessions[sessionId] means the
conversation follows it.
Capability disclosure: reads and writes this one JSON file; binds no network port beyond the loopback
HTTP prefix /plugin/subagent-models used by the Web UI; never reads conversation content,
credentials, or any other file.
Limits / 限制
subagent_fork children are overridden too. A fork shares the parent's prefix, so a different
model costs its KV cache; clear the route (or keep the preset's own agentOptions) if that
matters to you.- An effort the chosen model does not support fails the adapter's validation — the plugin never
silently downgrades.
- The override is per request. Context-window and metering metadata follow the effective route,
while the child's creation options keep the preset's values.
Relationship to similar plugins / 与同类插件的关系
Several plugins in the market move subagent routing: dsh-subagent-custom (task-class routing),
dsh-subagent-model-picker (a per-session picker), dsh-subagent-default-model (settings.yaml
default), dsh-model-settings (a settings panel), dsh-orquestrator (a fixed forced route). This
one keeps the surface at one composer chip: two scopes, one pair of DSH-style pickers, and a label
that always shows the effective value.
Verified / 已验证
On dsh 0.2.1-rc.2 (desktop profile):
- global route set to an invalid model → the child failed;
deepseek-v4-pro + max → it succeeded;
- per-conversation override set to an invalid model → the child failed (the session scope is
honored), then clearing it restored the global route;
GET /plugin/subagent-models/state?sessionId=… → 200 with global, session and the catalog;POST /plugin/subagent-models/route with scope: "session" / "global" → 200 and the JSON file
updates on disk;- the composer chip and the frame-wide card register as active slot occupants; no settings section
is registered by this plugin.
License
MIT © Gary Gai
