---

What it does

You stop choosing a model. The plugin reads each prompt, asks the allaigate semantic router what kind of task it is and how hard it is, and sends the call to the model your rules name for that combination.

Your prompt	Router says	Answered by
"Привет, как дела?"	chitchat, low	a fast, cheap model
"Translate to French: Hello world"	translation, low	a fast, cheap model
"Write a Python one-liner that reverses a string"	code, low	a fast, cheap model
"Design a lock-free MPMC queue in Rust and prove linearizability"	code, high	your strongest model

Simple work stops paying premium prices, hard work stops being answered by a model that cannot do it.

---

Install

dsh plugin --profile web add cortiq-dsh-llm-router

No global dsh? Use npx -p @deepseek-ai/dsh dsh plugin --profile web add cortiq-dsh-llm-router.

One prerequisite. The router picks among the models your other provider plugins already serve — llm-deepseek, llm-pi-ai, or any other registered provider. Keep at least one configured: this plugin never talks to a model vendor itself, it only decides who should answer.

---

Quick start

1. Get a key. Sign up at allaigate.com and copy your cortiq_… key.
2. Store it. In the harness: Settings → Models → Cortiq Router, paste the key. Or export it: export CORTIQ_ROUTER_KEY=cortiq_…
3. Pick the model Cortiq Router · Auto in the chat model selector.

That is the whole setup. The defaults route across DeepSeek models; point them at your own model ids when yours differ (see Routing rules).

---

How it works

your prompt
    │
    ▼
┌──────────────────────────────┐
│ 1. Text extraction           │  only the last user turn by default —
└───────────┬──────────────────┘  the rest of the conversation stays local
            │  POST /v1/route
            ▼
┌──────────────────────────────┐
│ 2. Semantic router           │  task = code, complexity = 0.78 → high
└───────────┬──────────────────┘  router.allaigate.com
            │
            ▼
┌──────────────────────────────┐
│ 3. Your routing table        │  code @ high → ["your-strong-model", …]
└───────────┬──────────────────┘  an ordered chain, not a single name
            │
            ▼
┌──────────────────────────────┐
│ 4. The provider that serves  │  llm-deepseek, llm-pi-ai, …
│    that model answers        │  streamed back untouched
└──────────────────────────────┘

Nothing here is a single point of failure. No key, an unreachable router, a timeout — the call still goes out, on the fallback model. A candidate whose provider refuses (missing credential, unknown model) is stepped over and the next candidate answers. Only a provider that has already started streaming keeps the call.

---

Model picker entries

Four entries, all of them routing. The last three pin the policy for that conversation without a trip to settings:

Entry	Behaviour
Auto	Uses the policy profile from settings
Auto · cost saver	Prefers the cheap end of every chain
Auto · balanced	The middle ground
Auto · quality first	Escalates to strong models sooner

---

Routing rules

Two layers, checked in order: a per-task rule first, then the tier chain.

llm-cortiq-router:
  # Where your low / medium / high tiers sit on the router's 0…1 score
  complexityBands:
    low: 0.35
    medium: 0.65

  # Fallback chains per tier — tried in order
  globalTiers:
    low:    [deepseek-v4-flash]
    medium: [deepseek-v4-pro]
    high:   [deepseek-v4-pro]

  # Per-task overrides, applied before the chains above
  taskRules:
    code:
      low:  [deepseek-v4-flash]
      high: [claude-opus-5]
    translation: [deepseek-v4-flash]

  # Last resort: router unreachable, unpaid, or nothing else registered
  defaultModel: deepseek-v4-flash

Use the model ids your provider actually advertises. An OpenRouter route names DeepSeek models deepseek/deepseek-v4-flash, while llm-deepseek names the same model deepseek-v4-flash. A bare id is looked up in every registered provider's catalog; write provider:model (e.g. openrouter:deepseek/deepseek-v4-pro) to name a route exactly.

---

Settings

Setting	Default	What it does
apiKeyEnv	CORTIQ_ROUTER_KEY	Credential holding the allaigate key
routerUrl	https://router.allaigate.com	Classifier endpoint; point it at your own deployment
routerProfile	balanced	cost-saver / balanced / quality-first
textStrategy	last_user	What gets classified: last user turn, + system prompt, or the whole thread
maxChars	4000	Hard cap on characters sent to the classifier
timeoutMs	15000	Hard prompts escalate to an oracle (~10 s); on timeout the call runs on the fallback
complexityBands	0.35 / 0.65	Where low / medium / high sit on the score
globalTiers, taskRules, defaultModel	see above	Which model answers
language	auto	Copy language: en, ru, zh, de, fr, es, tr
echoRouting	true	Log every decision: task, tier, score, chosen provider/model

---

Self-hosting the classifier

Point routerUrl at your own deployment and prompts never leave your infrastructure. The endpoint must serve POST /v1/route over TLS that Node accepts — a private CA works when NODE_EXTRA_CA_CERTS names it; a self-signed certificate does not (Node's fetch has no per-request opt-out).

---

Links

• npm: cortiq-dsh-llm-router
• Plugin docs: packages/dsh-llm-cortiq-router (English · Русский)
• Changelog: CHANGELOG.md
• Plugin catalog: DSH 1024Store
• Get a key: allaigate.com

License

MIT