@snowamberx/dsh-role-router
按角色路由 DeepSeek Harness 代理请求：计划模式使用规划器模型，否则使用默认模型；子代理使用子代理模型。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add @snowamberx/dsh-role-router@0.3.2
Role Router 以 @snowamberx/dsh-role-router 发布，当前版本为 0.3.2。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。
Tired of manually switching models between planning and execution?
dsh-role-router routes every DeepSeek Harness agent request by role:
planner model in plan mode, the default model otherwise, and a dedicated
model for subagents — no manual intervention needed.
default, planner and subagent are configured
independently; an unconfigured role passes the request through untouched
(the official layered selection applies). planner is triggered
automatically by plan mode (/plan and friends)./model — the host's live, provider-grouped catalog), plus a
composer summary pill.role-router settings namespace (user layer, which takes precedence);
saved settings apply to the next request without a restart.Every model request is routed by role; the listeners are registered on the root context, so they observe top-level agents and every in-process subagent:
| Role | Requests | Model source |
|---|---|---|
default | top-level agents outside plan mode | Configured → forced; unset → pass-through (official layered selection) |
planner | top-level agents while plan mode is active | Configured → forced; unset → pass-through (official layered selection) |
subagent | every in-process subagent request (any depth) | Configured → forced; unset → pass-through (official layered selection) |
An unset role leaves the request untouched, so the harness's official
per-session selection applies with its usual precedence — in-session
switches (composer / /model) > the session's latest logged request > the
global default. In-session model switches therefore take effect for unset
roles on the next turn (the selection is snapshotted at prompt assembly, so
a mid-turn switch never splits the running turn).
Switching models drops an inherited adapter-owned reasoningEffort unless
the role configures an explicit one (the routed model may not support the
previous model's effort; prepareCall rejects unsupported explicit efforts);
pass-through requests keep everything the official layer assembled,
including its effort. Plan mode is folded from the durable plan/mode
session events; ctx.planMode is consulted first when
visible. Auxiliary model calls (compaction, session-title) and out-of-process
subagent providers (acp, codex, …) are unaffected.
Known interplay: a forced planner route is persisted into the session's request header, and the official "latest logged request" layer treats that as the session's current model. To keep an unset default (follow-official) from inheriting the planner model after plan mode ends, the plugin snapshots the official route at the plan-entry edge and restores it once at the plan-exit edge, then resumes pure pass-through; a fixed default role wins over the restore, and an unconfigured planner never snapshots.
dsh plugin --profile web add @snowamberx/dsh-role-router
Restart dsh web for the change to take effect.
This package is a standard DSH community bundle: package.json declares the
dsh.bundle manifest ("patch": "./cordis.patch.yml" — the configuration
layer dsh plugin add recognizes and stacks into dsh.profile.bundles) and
a dsh.client web half (platform: "web" + inject dependency edges), with
exports["./client"] shipping the browser bundle at lib/client.js (a
standard window.__ModuleLoader__.load({ id, factory }) closure-factory
artifact served by client-modules under /plugins/). The build runs tsc
(node half + types) and tsdown with the same clientBundle preset the
official packages/client/* packages use.
- id: model-router
name: '@snowamberx/dsh-role-router'
config:
default: # optional; omit the key to keep it unset (pass-through)
provider: deepseek-official
model: deepseek-v4-flash
planner: # optional
provider: deepseek-official
model: deepseek-v4-pro
reasoningEffort: max # optional
subagent: # optional
provider: deepseek-official
model: deepseek-v4-flash
Unknown keys and blank provider/model/reasoningEffort values fail loud at load. All three roles are optional; unconfigured roles pass requests through to the official layered selection.
role-router namespace: { default?, planner?, subagent? }, each role being
{ provider, model, reasoningEffort? }. Settings-document values win over the
composition layer.
BSD-3-Clause. See LICENSE.