welsione/dsh-model-router

Unified model routing plugin: one logical ModelID backed by multiple providers' equivalent models, organized into plans (each with three tiers tier1/2/3) — first-token failover with cooldown, health-aware ranking, per-candidate reasoning effort; manual per-session tier picking in the chat window, and a settings panel that auto-saves every change to routes and model capabilities.

📦 Published on npm: @welsione/dsh-model-router

中文说明见 README.md。

dsh plugin --profile web add @welsione/dsh-model-router

Screenshots

Chat window · plan & manual tier picker	Settings panel · route management

Left: the chat-window "plan · tier" picker — per-plan custom tier names (e.g. 穷鬼套餐: 夯 / NPC / 拉完了), click to pick a tier manually, and the current provider/model + reasoning effort shown at the bottom. Right: the "统一模型路由" settings panel — candidate chains per plan with per-candidate failure/success counts (health), reasoning effort, switch stats (280 requests · 9 switches), cooldown and model-capability editing, all auto-saved.

Overview

Capability	One line
Multi-plan routing	Multiple plans (Route Groups) coexist, each with its own three-tier candidate chain and tier names
Auto failover	Pre-first-token failures (rate limit / quota / auth / network / unknown model / empty response) switch to the next candidate; the failed one enters graded cooldown (short for rate-limit, medium for server, long for auth) with exponential backoff (capped at 30 min)
Health-aware ranking	Sliding-window time-decayed, error-code-weighted scores reorder chains — recent results weigh more, server-class failures cost more than rate-limit ones; toggle off in one click
Three tiers + manual tier	Per-plan tier1 light · tier2 standard · tier3 powerful, auto-selected by purpose with downgrade; per-session manual tier in the chat window
Reasoning effort	Per-candidate reasoningEffort, prechecked with a real request at save time — only host-accepted levels allowed
Tier names	Per-plan custom tier display names (routes.<id>.tierNames), click-to-rename colored capsules, synced to the chat plan menu
Model capability write-back	Edit custom-provider model capabilities (reasoning effort / contextWindow / maxTokens) in the panel, written back to the host llm-pi-ai, hot-reloaded
Management panel	Built-in 模型路由 card in DSH Settings with route stats / cooldown / health / capability editing, auto-save; live routing status in the chat toolbar
UI language	Panel copy goes through the host locale registry (model-router namespace) with shipped zh/en dictionaries and reload-free switching; third-party language packs can register more languages for the same namespace
Session safety	Recoverable route events, automatic replayState sanitization across providers

Compatibility

• DSH 0.1.0-rc.x – 0.1.5-rc.x (verified on 0.1.5-rc.1; compatible with the settings.installSection API introduced in 0.1.2) · Node.js ≥ 22 (React 18/19) · Last verified 2026-09-11.

Install / Uninstall

Published on npm as @welsione/dsh-model-router; dsh plugin add pulls it from npm automatically:

dsh plugin --profile web add @welsione/dsh-model-router    # install
dsh plugin --profile web remove @welsione/dsh-model-router  # uninstall

npm unreachable? Fall back to the GitHub source: dsh plugin --profile web add github:welsione/dsh-model-router.

After install, a 模型路由 card appears in Settings; the chat model picker becomes a plan picker with tier switching. See docs/usage.md.

Quick start

Open Settings → 模型路由, add a unified ModelID (e.g. deepseek-v4-flash) and configure tier1/2/3 candidates — any change auto-saves instantly. Minimal config and full examples: docs/usage.md.

Configuration

Visual editing for enabled / cooldownMs / maxSwitchesPerStep / healthRanking / reasoningEffortsFallback / routes (including per-plan tierNames). Full config table, llm-pi-ai capability write-back and panel API: docs/usage.md.

Permissions & data

No outbound calls, no workspace file access, no telemetry, no API-key handling; runtime state is in-memory, manual tiers persist to host settings.

Troubleshooting

Missing 模型路由 card / candidate not in catalog / auto-save failed / constant switching / session stuck at step 1 → quick checklist: docs/usage.md.

Development

npm test                 # unit tests (node:test, zero deps)
npm test && npm pack --dry-run   # pre-publish gate

Pure logic in lib/core.mjs + wiring in lib/index.js + Web UI in lib/client.js; compliance evidence: docs/self-check.md.

Publishing a new version (auto-publish to npm)

Push a v<version> tag and the GitHub Action (npm-publish.yml) verifies the tag matches package.json version, then runs npm publish:

npm version patch   # bump version (patch/minor/major, update CHANGELOG)
git push && git push --tags

Publishing requires the repo's NPM_TOKEN secret (an Automation token for the welsione npm account).

License

MIT. Report security issues privately through the repository.