English · 简体中文

Routine turns shouldn't cost flagship money. The turns that matter shouldn't be left to a cheap model.

Before every turn of a top-level agent, a small LLM Judge (running on your Fast-tier model chain) classifies the user's message as fast (routine) or smart (consequential). The chosen tier then drives the whole turn — thinking, tool calls, code edits — through the harness's own agent/request pipeline. The Judge only classifies; it never does the work.

🦾 [deepseek-flash]     → fix the failing test
🧭 judging…
🧠 [deepseek-v4-pro]    ← "design the auth flow" → upgraded instantly
⚠️ deepseek-flash 429 → cooldown, same-tier failover — retry in 1m
🦾 [deepseek-v4-flash]  ← next healthy model in the Fast chain

• Expected-cost routing — the turn runs Smart iff pSmart ≥ θ, with θ = 1/reworkPenalty: the bar is price-independent, so the one knob is how badly a wrong downgrade hurts. Upgrades are instant; coming back down needs downgradeMemory consecutive decisive fast turns. A Judge outage or an unsure verdict is a hold — the router keeps its position instead of guessing.
• Cache-aware routing — when Fast and Smart share a provider, the decision bar is divided by sameFamilyPenalty (default 1.5) and downgrades are held off while the prompt cache is warm, so switching to a cheaper model never costs more than staying put.
• Runtime failover — 429 / 402 / 5xx / quota / usage-limit / unsupported-model failures put the model into exponential-backoff cooldown (1m → 4m → 16m → 1h04m → 4h16m → 6h cap; client-side limits start at 16m) and re-resolve the same tier to the next healthy model — same-turn retry, never cross-tier.
• Task-level orchestration — complex tasks run the Smart tier as a CTO that plans, delegates implementation to Fast engineer subagents via the harness's subagent tool, reviews each result, and iterates. The hard caps are enforced by the plugin, not just prompted: each delegation counts a round, consecutive worker failures count an escalation, and once a cap is hit the subagent tool is denied outright and the system prompt switches to a "wrap up now" notice.
• Cost telemetry — per-tier token tracking and an optional USD pricing table (/router status shows "what this session would have cost on the Smart model").
• Visible when it acts — a tier or model switch is written into the transcript as a [shift-router] Fast → Smart · … notice (the harness has no status-bar seat for plugins), so an enabled router is never silently invisible. Set ux.routerLogVerbose to get one notice per judged turn, not just per switch.
• Zero-config startup — a no-op until you configure tiers; then routing just works. Configuration is editable live via the GUI settings panel and /router config commands (persisted, no restart).

This package is a DSH bundle: its cordis.patch.yml inserts the plugin row into any profile that lists it. Every channel below ends in the same dsh plugin --profile <name> add …, which forwards to pnpm inside the profile directory.

dsh plugin --profile web add dsh-shift-router

Installs the prebuilt artifact — no build script runs on your machine, so there is nothing to authorize.

npm pack      # or download the release tarball
dsh plugin --profile web add ./dsh-shift-router-0.6.0.tgz

Prebuilt as well, and the option to reach for when a registry is not available.

dsh plugin --profile web add github:green-dalii/dsh-shift-router#v0.6.0

A git install fetches source, not artifacts, so the package's prepare script builds dist/. pnpm ≥ 10 refuses a git dependency's prepare until you allow it — if the first add fails, copy the exact key pnpm prints into the profile's pnpm-workspace.yaml and re-run:

allowBuilds:
  dsh-shift-router: true

This authorizes the package's code to execute on your machine at install time, outside the agent sandbox. Pin a tag or commit (…#v0.6.0, …#<sha>) so a later push cannot change what you actually run.

git clone https://github.com/green-dalii/dsh-shift-router.git
cd dsh-shift-router && npm install && npm run build
dsh plugin --profile web add /path/to/dsh-shift-router

dsh --profile web --dump-config | grep -A3 'id: shift-router'

The plugin loads with no configuration at all (every default is safe); tier models come from the settings card (SPEC §12) or from the profile's patch row. Later layers win, and a patch replaces the target row's whole config value — a row that overrides this one must restate every key it needs (SPEC §10).

DeepSeek Harness supports hot reload through @deepseek-ai/cordis-plugin-hmr, but two things are worth knowing:

1. The official Web bundle ships the shared HMR row disabled (packages/bundle/web-app/cordis.patch.yml has - id: hmr, disabled: true, upstream TODO: "Re-enable shared HMR for Web after its reload lifecycle is tested"). Re-enable it in your profile patch — this is the documented override mechanism:

   # ~/.dsh/profiles/<name>/cordis.patch.yml
   - id: hmr
     disabled: false

2. What hot-reloads and what doesn't (verified against the current implementation):

   • ✅ Configuration changes — editing this profile patch (or the home patch) re-runs the affected plugin's apply() with the new config, no restart. The plugin's own config is also hot through the settings namespace (/router config set and the GUI card apply live without HMR at all).
   • ❌ Module (code) changes — the HMR accepted-dependency graph currently covers the harness's own modules only; editing an external plugin's compiled files (e.g. dist/index.js) does not trigger a reload in the current release, so code changes still require a restart. This is the untested "reload lifecycle" the upstream TODO refers to, not a limitation of this plugin.
   • ❌ Client package metadata — the dsh.client manifest and exports["./client"] are cached in-process; adding/fixing them requires a profile restart (only dist/client.js content changes ride the client HMR rebuild chain).

   In practice: configure with /router config / the settings panel (always live), switch models by editing the patch (live once HMR is on), and restart only when you change plugin code.

Configuration lives in the shift-router settings namespace: edit it in the GUI (Settings → Plugins → Plugin configuration — the "Shift-Router" card), with /router config commands, or via the profile patch row. All fields have safe defaults.

Tier models are the only thing you must choose: docs/MODELS.md covers how to pick a Fast and a Smart model (the Fast chain also serves the Judge), what makes a good fallback, and which models can take images.

All numeric fields are range-validated by the schema (e.g. window.minConfidence must be in [0,1], window.size a positive integer); invalid values are rejected at load / on set, never silently accepted.

Upgrading from v0.5.0: routing decisions change immediately with no config edit — the decision rule went from counting window votes to weighing expected cost, and two legacy knobs changed meaning. See SPEC.md §15 and the [0.6.0] section of CHANGELOG.md.

The package ships a browser-side (client) module that registers a "Shift-Router" card in the GUI settings page:

• Where: Settings → Plugins → Plugin configuration (that page is provided by the official dsh-client-ui-settings-plugins; the card registers into the settings.plugin.item slot).
• What: a form over every scalar leaf field (booleans, numbers, enums) plus the two tier model chains, grouped into seven sections (General / Models / Routing / Orchestration / Failover / Telemetry / Logs & UX) with sub-groups for the routing section (Judge / Decision window / Cache-aware). Scalar fields use the compact settings-row pattern — label + hint on the left, control right-aligned on the same line — so each field is one tight row instead of three stacked lines. Controls use the host-plane design tokens: toggle switches (contrast-safe in light and dark themes), a styled select for enums, unit suffixes inside numeric inputs (ms, tokens, 0–1, …), and an ordered row editor for model chains — the row order is the in-tier fallback order, so the first available model wins and the rest are its fallbacks. The provider/model dropdowns are auto-loaded from DSH's runtime model catalog (llm.models — the same catalog the DSH settings surface reads): only providers with a currently advertised model list appear, no dormant-directory noise, and nothing is hardcoded, so the card works with any deployment's configured models. A "Custom…" escape covers values outside the catalog. Staged saving, per-field reset to default, and override markers work exactly like the official cards.
• Boundary: only pricing (the optional USD cost table) stays with /router config and the profile patch; the tier model chains are editable in the card.
• Build: npm run build emits both the host artifact (dist/index.js) and the client bundle (dist/client.js). The client module is discovered through the dsh.client manifest by dsh-client-modules, which requires the plugin to be mounted by package name (dsh-shift-router) — a source-checkout patch (name: '/path/dist/index.js') does not serve the card.

Harness builds up to 0.1.0-rc.x filtered a third-party settings namespace out of the browser's settings.describe response unless it appeared in WEB_SETTINGS_NAMESPACES — which is why scripts/expose-gui-settings.mjs exists: it patches the profile's installed dsh-host-apiproxy (idempotent; re-run after a dependency upgrade). From 0.1.5-rc.2, this project's baseline, that package and its whitelist are gone and the namespace is exposed natively, so the script reports "not needed" and exits 0. It is not part of the published package (files), so it is only reachable from a source checkout.

The mechanism-by-mechanism mapping is normative and lives in SPEC.md §1.1. Two things are worth stating here because they are not in that table:

• Throughput is deliberately absent. DSH renders tok/s natively (chat footer, trajectory panel) from decode time; the router owns routing decisions and spend, not rate display.
• Subagents are never routed. Workers spawned by the subagent tool carry session.header.origin === 'subagent' and keep their pinned model; the router drives top-level agent turns only.

Turn-level routing decides which model runs a turn; task-level orchestration decides how a complex task is executed. When the Judge says smart and orchestration.mode is auto (the default), the router hands the turn to the Smart tier as a CTO: it plans, delegates implementation to Fast workers through the harness's subagent tool, reviews each result and iterates — then an independent acceptance audit checks the claim. A fast verdict never triggers any of this.

1. Enter — the decision tier is Smart, the subagent tool is present and the mode is auto; the orchestrator instruction is injected as a system-prompt section (SPEC §7.2).
2. Delegate — the CTO calls subagent; each call is one round, counted at dispatch (a dispatched delegation has already spent budget, so this keeps maxRounds a true ceiling).
3. Review and iterate — a failed worker advances the failure streak, and the convergence protocol requires every re-delegation to carry a structured ## Failure report (what failed, where, which acceptance test re-checks it); repeating the same feedback triggers takeover (SPEC §7.2.1).
4. Stop — at the cap the subagent tool is denied at tools/pre-execute and the prompt section becomes a wrap-up notice; agent/turn-stopping releases the state. Orchestration is single-turn: nothing spans turns, and a leaked run is swept at the next turn's start.

Caps stop a run from flying away; they cannot stop a CTO from claiming acceptance it never verified. A run that actually delegated is therefore audited: deterministic checks always run (every dispatched worker reported, a CTO summary exists, no cap ended the run), and — when orchestration.audit.enabled — one small Fast-tier review reads the goal, the summary and the worker results. It never blocks a turn: the LLM half runs detached and its verdict appears as Last audit: in /router status.

Upstream calls the per-worker model pin mandatory: a worker that inherits the parent's model runs on Smart mid-orchestration, which collapses the economics. In DSH that pin sits behind the harness's own allowlist (subagent-model-selection, default off). /router allow-workers writes the Fast chain into it for you; when it is not authorised the router says so in a Worker delegation: line rather than pretending otherwise (SPEC §7.4).

/router status shows 🪄 active (round x/y, esc a/b, fail streak n), then Orchestration spend: $X · N/M workers reported once workers report, and Last audit: … after the audit settles. Entry also writes a route notice into the transcript (SPEC §13.1), so the shape of the turn is visible without reading counters.

A fast verdict; orchestration.mode: off (or /router orchestrate off); the subagent tool absent from the composition; or no Smart model resolvable — in which case the turn simply runs on Smart, with no delegation and no audit.

npm run build       # tsc (host → dist/) + tsc client + tsdown (client bundle → dist/client.js)
npm test            # vitest (336 tests across 18 files: EV routing / failover signatures / judge parsing + prompt contract / orchestration / config schema + migration / telemetry / route notices / config registries + GUI form model + card UX + model catalog / packaged-install contract)
npm run typecheck

e2e/ contains a fake LLM adapter that registers the fake provider, so the whole routing pipeline can be exercised without any API key:

npm run test:e2e

That builds a throwaway DSH_HOME, installs this checkout as a bundle into a derived headless profile, runs one turn through the fake adapter, and asserts:

• ROUTER-E2E: turn ran on fake/fake-smart notice=yes — the Judge ran, the EV rule escalated, the wire model was actually switched to the Smart tier, and the [shift-router] route notice reached the model request (SPEC §13.1);
• the same outcome under e2e/legacy-config-overlay.yml, a profile patched with the pre-alignment config (the legacy knobs at their old defaults plus the removed requireSmartModel key) — i.e. the upgrade path, not just a fresh install;
• the same outcome under e2e/orchestration-overlay.yml, which uses the plugin's default orchestration mode (auto) and mounts the web-only subagent-model-selection-settings row beside it — the composition that once aborted the boot;
• the shift-router settings namespace round-trips a write, and the Host model catalog advertises the deployment's configured routes (e2e/settings-probe.mjs);
• a packed install (npm pack → tarball → second scratch profile) boots with the plugin loaded, its browser half offered to the client module loader, and no @deepseek-ai/* copy installed into the profile — devDependencies are absent there, so a runtime import that is not declared for the consumer fails in the e2e instead of in a user's install.

It never touches your real DSH_HOME and cleans up after itself (--keep to inspect). To run the same thing by hand, install the bundle into a scratch profile and apply the overlay:

DSH_HOME=/tmp/scratch dsh plugin --profile tmp add /path/to/dsh-shift-router
DSH_HOME=/tmp/scratch dsh --profile tmp --patch e2e/overlay.yml "design a migration plan"

Repository layout — the module-by-module map, the pure-logic/glue split and the test layering — lives in CONTRIBUTING.md § Repository layout, so there is one copy to keep accurate. In short: src/ is the host half (pure decision modules plus DSH wiring in index.ts), src/client/ is the browser half (the settings card), tests/ covers both, and e2e/ boots scratch profiles — including the packed artifact.

• pi-shift-router — the upstream project this plugin is adapted from: the same two-tier architecture (LLM Judge, fallback chains, exponential-backoff failover, task-level orchestration) for pi-coding-agent. Behaviour parity and the deliberate divergences are tracked in ALIGNMENT.md.
• dsh-plugin-dev-skill — the agent skill for building DSH plugins: tools (defineTool), LLM adapters, services, events, config and packaging, with the Cordis mental model and a verification checklist. This project follows it; install it in DSH, Claude Code or Codex.
• obsidian-llm-wiki — an Obsidian plugin that turns notes and PDFs into a linked, queryable knowledge base (entity and concept pages, graph-powered Q&A, local-first, no backend). Same author's other line of work.

MIT © 2026 green-dalii and contributors.