dsh-shift-router

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-v4-flash] → fix the failing test
🧭 judging…
🧠 [deepseek-v4-pro]   ← "design the auth flow" → upgraded instantly
⚠️ deepseek-v4-flash 429 → cooldown, retrying on glm-5.2 — retry in 1m
🦾 [glm-5.2]           ← same-tier failover

Features

• Instant upgrades, trend-gated downgrades — one smart verdict switches to the strong tier immediately; coming back down requires a sliding-window majority (default 5 turns, ≥60%, low-confidence votes ignored).
• Cache-aware routing — when Fast and Smart share a provider, the router raises the downgrade threshold (0.9) and holds off while the prompt cache is warm, so switching to a cheaper model never costs more than staying put.
• Runtime failover — 429 / 5xx / quota failures put the model into exponential-backoff cooldown (1m → 4m → 16m → 1h → 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 (a smart verdict) 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, each failed worker result counts 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/throughput tracking and an optional USD pricing table (/router stats shows "what this session would have cost on the Smart model").
• 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).

Install

As a bundle (recommended)

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

The bundle's cordis.patch.yml inserts the plugin into any profile that lists it. The plugin loads without any configuration (all defaults are safe); tier models come from the settings panel or the patch row.

Installing from git (dsh plugin --profile <name> add github:green-dalii/dsh-shift-router) builds dist/ automatically via the package's prepare script. pnpm ≥ 10 refuses git dependencies' prepare scripts by default — add this to the profile's pnpm-workspace.yaml and re-add if the build is skipped:

allowBuilds:
  dsh-shift-router: true

This grants the package permission to run its build script at install time. For a fully lock-down install, use npm run build on a source checkout (below) instead.

From source (local development)

Point the profile's patch layer at the built entry:

# ~/.dsh/profiles/<name>/cordis.patch.yml
- insert:
    - id: shift-router
      name: '/absolute/path/to/dsh-shift-router/dist/index.js'
      config:
        tiers:
          fast:
            models:
              - { provider: opencode-go, model: deepseek-v4-flash, priority: 1 }
          smart:
            models:
              - { provider: opencode-go, model: deepseek-v4-pro, priority: 1 }

Hot reload

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 panel 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.

   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

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

Field	Default	Description
enabled	true	Master switch
tiers.fast.models	[]	Fast-tier chain (provider/model + priority); also the Judge's model chain
tiers.smart.models	[]	Smart-tier chain
routing.mode	auto	auto (default): judge + routing + failover + orchestration; manual: no judge, only explicit /route-force overrides; off: fully passive for model selection (commands/telemetry still work)
routing.judgeTimeout	5000	Judge call timeout (ms)
routing.judgeMaxTokens	4000	Max output tokens for a single Judge call
routing.judgePromptCap	6000	Max prompt characters sent to the Judge (bounds Judge cost)
routing.window.size	5	Downgrade sliding window size
routing.window.threshold	0.6	Fast-majority ratio required to downgrade
routing.window.minConfidence	0.5	Ignore judge verdicts below this confidence
routing.cacheAware.enabled	true	Same-provider cache protection
routing.cacheAware.sameFamilyThreshold	0.9	Downgrade threshold when tiers share a provider
routing.cacheAware.idleBoundaryMs	300000	Idle gap before a warm cache is considered cold
orchestration.mode	auto	auto: complex → Smart CTO; off: plain two-tier routing
orchestration.maxRounds	3	Delegate→review rounds hard cap (enforced: each subagent delegation counts one round; at the cap the subagent tool is denied)
orchestration.escalationThreshold	2	Failed worker results before Smart must take over the phase (enforced: each isError subagent result counts)
orchestration.requireSmartModel	true	Skip orchestration if the Smart model can't be resolved
failover.baseMs	60000	Cooldown base delay for 5xx failures (1m)
failover.maxMs	21600000	Hard cap on the backoff ladder (6h)
failover.startAttempts4xx	3	4xx (429/quota) failures start at this attempt (16m), client limits usually outlive server blips
failover.speedWindowSize	5	Recent tokens/sec readings kept for the /router stats average
telemetry.callLogCap	1000	Max per-message attribution records kept for baseline cost computation
ux.routerLogVerbose	false	Print router decisions to the harness log
pricing	[]	Optional {provider, model, input, output, cacheRead?, cacheWrite?} USD-per-1M-token table for cost telemetry

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

Commands

Command	Effect
/router	Compact status
/router status / /router stats	Full status: tiers, window, transitions, cooldowns, tokens, cost telemetry
/router on / /router off	Enable / disable (session-scoped)
/router verbose	Toggle verbose router logging
/router orchestrate auto|off	Orchestration mode
/router config	Show effective config + available providers/models + usage
/router config set <path> <value>	Set one field (persisted), e.g. set routing.judgeTimeout 8000, set tiers.fast.models [...]
/router config set-fast <provider/model>	Replace the Fast tier chain with one model
/router config set-smart <provider/model>	Replace the Smart tier chain with one model
/router config reset	Restore the composition default
/route-force <fast|smart|auto|provider/model>	Force the next turn to a tier/model (one-shot)

How it works (DSH integration)

Capability	DSH mechanism
Turn-start classification	agent/pre-step waterfall (step === 1, top-level agents only)
Model switching	agent/request waterfall (per-step provider/model override)
Runtime failover	agent/request-error waterfall (cooldown + {kind:'retry'} same-tier retry)
Judge LLM calls	ctx.llm.stream() — reuses the harness's adapters, credentials, and JSON-mode enforcement
Orchestrator instruction	ctx.systemPrompt.section() rendered per agent while orchestration is active
Orchestration hard caps	tools/pre-execute denies the subagent tool at the cap; tools/result counts failed workers; the prompt section switches to a "wrap up" notice
Config (GUI + commands)	dsh-settings namespace shift-router (same store for both surfaces)
Usage telemetry / cooldown recovery	session/event assistant/message (TokenUsage; a successful message clears the model's cooldown)
Commands	ctx.commands.register()
Tier-chain prompt variables	{{shift_router_fast_chain}} / {{shift_router_smart_chain}}

Subagents are never routed. Workers spawned by the subagent tool carry session.header.origin === 'subagent' and keep their pinned model; the router only drives top-level agents.

Orchestration and the DSH subagent tool

The original pi plugin delegated through pi-subagents with agent: "worker", context: "fresh", and a per-call model pin. DSH's subagent tool differs:

• The tool takes description + prompt (and run_in_background); a worker runs in its own fresh session — the prompt is its world.
• The worker model is pinned by deployment configuration (dsh-tool-subagent's agentOptions), not by the tool call. By default a worker inherits the parent's model.
• Therefore the orchestrator prompt instructs the CTO to delegate with precise task contracts, review/iterate/escalate within the hard caps, and lists the Fast-tier chain the deployment should have pinned in tool-subagent.agentOptions for cost parity.

The caps are enforced by the router, not just described: every subagent tool call while an orchestration turn is active increments orchestration.rounds; every failed (isError) subagent result increments orchestration.escalations; once capHit() is true the subagent tool is denied at tools/pre-execute and the orchestrator prompt section is replaced by a "wrap up now" notice. /router status shows the live counters (round x/max, esc y/threshold).

Development

npm run build       # tsc → dist/
npm test            # vitest (62 tests: routing / failover / judge parsing / orchestration / config schema)
npm run typecheck

End-to-end test (no credentials)

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

# after creating a scratch profile with this bundle + @deepseek-ai/dsh-headless:
dsh --profile <tmp> --patch e2e/overlay.yml "design a migration plan for our billing system"
# → ROUTER-E2E: turn ran on fake/fake-smart   (judge said smart → upgraded to the Smart tier)

The e2e also verifies the settings namespace persists (e2e/settings-probe.mjs).

Architecture

src/
├── index.ts        # plugin entry: event wiring, per-agent state, judge, orchestration section
├── config.ts       # Schemastery schema + deep-merge normalization
├── types.ts        # shared types + defaults
├── router.ts       # pure routing engine (upgrade/downgrade/window/cache-aware)
├── judge.ts        # LLM Judge via ctx.llm.stream() + reply parsing
├── failover.ts     # exponential-backoff cooldown state machine
├── tier.ts         # tier model resolution + display
├── orchestrate.ts  # orchestrator prompt + lifecycle + caps
├── stats.ts        # telemetry snapshot (tokens / throughput / cost estimate)
└── commands.ts     # /router and /route-force

Pure logic (router / failover / judge parsing / orchestration) is unit-tested in isolation; DSH wiring is exercised by the headless e2e.

License

MIT © 2026 green-dalii and contributors.