OmniRoute/docs/routing/AUTO-COMBO.md

title, version, lastUpdated

title	version	lastUpdated
OmniRoute Auto-Combo Engine	3.8.31	2026-06-20

OmniRoute Auto-Combo Engine

For Users: Looking for a quick start? See the Auto-Combo User Guide for simple explanations and examples.

Self-managing model chains with adaptive scoring + zero-config auto-routing

Zero-Config Auto-Routing (auto/ prefix)

NEW: No combo creation required. Use auto/ prefix directly in any client.

Quick Examples

Model ID	Variant	Behavior
auto	default	All connected providers, LKGP strategy, balanced weights
auto/coding	coding	Quality-first weights, suitable for code generation
auto/fast	fast	Low-latency weighted selection
auto/cheap	cheap	Cost-optimized routing (lowest cost first)
auto/offline	offline	Favors providers with highest quota availability
auto/smart	smart	Quality-first + higher exploration rate (10%) for better model discovery
auto/lkgp	lkgp	Explicit LKGP (same as default auto)

Category × Tier Composition (auto/<category>:<tier>)

OpenRouter-style suffixes separate what kind of route (category) from how to optimize it (tier), so you can compose them freely (#4235 Phase B, open-sse/services/autoCombo/suffixComposition.ts):

• Categories (filter the candidate pool by capability): coding · reasoning · vision · chat · multimodal. vision/multimodal keep vision-capable models; reasoning keeps reasoning/thinking models.
• Tiers (pick the scoring weights / pool filter): fast (ship-fast) · cheap (alias floor, cost-saver) · reliable (circuit-breaker health + latency stability) · free / pro (filter the pool by model tier via classifyTier — free-tier vs. premium).

Example	Resolves to
auto/coding:fast	coding pool, low-latency weights
auto/coding:cheap	coding pool, cost-optimized (alias auto/coding:floor)
auto/reasoning:pro	reasoning/thinking models only, premium tier
auto/vision	vision-capable models (no tier → balanced weights)
auto/multimodal:free	multimodal-capable models, free tier only

Any valid auto/<category>[:<tier>] resolves on demand; a curated subset is advertised in /v1/models and the dashboard (AUTO_SUFFIX_VARIANTS in open-sse/services/autoCombo/builtinCatalog.ts). Filtering is fail-open — if a constraint matches no connected models, the full pool is used so routing never breaks. The core scorer (combo.ts) is unchanged; the category/tier filter is applied in buildAutoCandidates.

Live model intelligence: auto-routing fitness is informed by live Arena ELO rankings + models.dev tier data when the ARENA_ELO_SYNC_ENABLED flag is on (falls back to the static fitness map otherwise).

How to use:

# Any IDE or CLI tool that supports OpenAI format
Base URL: http://localhost:20128/v1
API Key:  <your-endpoint-key>

# In your code/config, set model to:
model: "auto"                 # balanced default
model: "auto/coding"          # best for coding tasks
model: "auto/fast"            # fastest available
model: "auto/cheap"           # cheapest per token

What happens:

1. OmniRoute detects auto/ prefix in src/sse/handlers/chat.ts
2. Queries all active provider connections from the database
3. Filters to those with valid credentials (API key or OAuth token)
4. Determines the model per connection (connection.defaultModel or provider's first model)
5. Builds a virtual combo in-memory (not stored in DB)
6. Routes using the selected variant's weight profile + LKGP strategy

Key properties:

• ✅ Always-on: No toggle, no combo creation, no configuration needed
• ✅ Dynamic: Reflects current connected providers automatically
• ✅ Session stickiness: LKGP ensures last successful provider is prioritized
• ✅ Multi-account aware: Each provider connection becomes a separate candidate
• ✅ No DB writes: Virtual combo exists only for the request, zero persistence overhead

Behind the scenes:

Request: { model: "auto/coding" }
   ↓
src/sse/handlers/chat.ts detects prefix
   ↓
createVirtualAutoCombo('coding') → candidatePool from active connections
   ↓
handleComboChat (same engine as persisted combos)
   ↓
Auto-scoring selects best provider/model per request

Implementation files:

File	Purpose
open-sse/services/autoCombo/autoPrefix.ts	Prefix parser (parseAutoPrefix)
open-sse/services/autoCombo/virtualFactory.ts	Creates virtual AutoComboConfig objects
open-sse/services/autoCombo/providerRegistryAccessor.ts	Test hook for mocking provider registry
src/sse/handlers/chat.ts	Integration: auto prefix short-circuit
src/shared/constants/providers.ts	SYSTEM_PROVIDERS.auto system entry

How It Works (Persisted Auto-Combos)

The Auto-Combo Engine dynamically selects the best provider/model for each request using a 9-factor scoring function (defined in open-sse/services/autoCombo/scoring.ts → DEFAULT_WEIGHTS). All weights sum to 1.0.

Source: diagrams/auto-combo-9factor.mmd

Factor	Default Weight	Description
health	0.22	Health score from circuit breaker (CLOSED=1.0, HALF_OPEN=0.5, OPEN=0.0)
quota	0.17	Remaining quota / rate-limit headroom [0..1]
costInv	0.17	Inverse blended cost (60% input + 40% output token price, normalized) — cheaper = higher score
latencyInv	0.13	Inverse p95 latency normalized to pool — faster = higher score
taskFit	0.08	Task-type fitness (coding, review, planning, analysis, debugging, docs)
specificityMatch	0.08	Match between request specificity (manifest hint) and model tier
stability	0.05	Variance-based stability (low latency stdDev / error rate)
tierPriority	0.05	Account-tier priority — Ultra=1.0, Pro=0.67, Standard=0.33, Free=0.0
tierAffinity	0.05	Affinity between the candidate's tier and the manifest-recommended tier

Sum: 0.22 + 0.17 + 0.17 + 0.13 + 0.08 + 0.08 + 0.05 + 0.05 + 0.05 = 1.0 (validated by validateWeights()).

Mode Packs

Four pre-defined weight profiles in open-sse/services/autoCombo/modePacks.ts. Each pack overrides the default weights to bias selection toward a specific goal. Below are the full weight tables per pack (each row sums to 1.0).

Factor	ship-fast	cost-saver	quality-first	offline-friendly
quota	0.14	0.14	0.10	0.37
health	0.28	0.19	0.18	0.28
costInv	0.05	0.37	0.05	0.10
latencyInv	0.32	0.05	0.05	0.05
taskFit	0.10	0.10	0.37	0.00
stability	0.00	0.05	0.15	0.10
tierPriority	0.05	0.05	0.05	0.05

Notes:

• tierAffinity and specificityMatch are not set in mode packs — calculateScore() treats them as ?? 0 when absent.
• Each pack's emphasis at a glance:
  • ship-fast → latencyInv 0.32 + health 0.28 (low-latency, healthy connections)
  • cost-saver → costInv 0.37 (cheapest tokens win)
  • quality-first → taskFit 0.37 + stability 0.15 (best model for the task, consistent)
  • offline-friendly → quota 0.37 + health 0.28 (max headroom regardless of speed/cost)

All Routing Strategies

OmniRoute's combo engine supports 17 routing strategies (declared in src/shared/constants/routingStrategies.ts → ROUTING_STRATEGY_VALUES). The Auto Combo engine itself is exposed under the auto strategy; the others are available for persisted combos.

Strategy	Description
priority	First-target ordered list with explicit priority
weighted	Weighted random by per-target weight
round-robin	Cycle through targets in order
context-relay	Hand off context across targets (long conversations)
fill-first	Fill each target's quota before moving to next
p2c	Power-of-2-choices random load balancing
random	Uniform random selection
least-used	Pick target with lowest current load
cost-optimized	Minimize $ per request given catalog pricing
reset-aware ⭐	Prioritize by quota reset time — short reset windows ranked higher
reset-window	Prefer targets whose quota window resets soonest
headroom	Pick the target with the most remaining quota headroom
strict-random	Random without deduplication of repeats
auto	Use Auto Combo scoring (9-factor) — recommended
lkgp	Last-Known-Good Path (sticky route to last successful target)
context-optimized	Pick target with best fit for current context size
fusion 🧬	Fan out to a panel of models in parallel, then synthesize one answer via a judge (see below)

⭐ = New in v3.8.0 · 🧬 = New in v3.8.36

Fusion Strategy

fusion is the one strategy that does not pick a single target. It fans the prompt out to every panel model in parallel, then a configurable judge model synthesizes a single final answer from all panel responses. Ported from upstream decolua/9router (OpenRouter's Fusion design); implementation in open-sse/services/fusion.ts.

How it works:

1. Fan-out — the prompt is sent to every panel model at once, forced non-streaming with tools stripped (the judge needs complete prose to synthesize).
2. Quorum-grace collection — as soon as minPanel answers arrive, a short grace timer starts for the stragglers, then fusion proceeds with whatever was collected. This caps the slowest model's penalty on wall time, bounded by a hard timeout.
3. Judge synthesis — panel answers are anonymized (Source 1, Source 2, … — so the judge weighs substance, not model brand) and handed to the judge, which analyzes consensus / contradictions / partial coverage / unique insights / blind spots, then writes one authoritative answer. The judge call keeps the client's original stream flag + tools, so streaming and downstream tool use still work.
4. Graceful degradation — 0 panel answers → 503; exactly 1 survivor → that answer is returned directly (nothing to fuse); a single-model panel answers directly.

Configuration

Configured on the combo's config blob (no schema migration — it reuses the existing combos table):

Field	Type	Default	Purpose
config.judgeModel	string	first panel model	Model that synthesizes the final answer
config.fusionTuning.minPanel	number	2	Successful answers required before the grace timer starts (clamped to [2, panelSize])
config.fusionTuning.stragglerGraceMs	number	8000	How long to wait for laggards once quorum is reached
config.fusionTuning.panelHardTimeoutMs	number	90000	Absolute cap so one hung model can't stall the request

Defaults live in FUSION_DEFAULTS (open-sse/services/fusion.ts).

Example

curl -X POST http://localhost:20128/api/combos \
  -H "Authorization: Bearer <key>" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "fusion-panel",
    "strategy": "fusion",
    "targets": [
      { "model": "cc/claude-opus-4-7" },
      { "model": "cx/gpt-5.5" },
      { "model": "glm/glm-5.1" }
    ],
    "config": {
      "judgeModel": "cc/claude-opus-4-7",
      "fusionTuning": { "minPanel": 2, "stragglerGraceMs": 8000, "panelHardTimeoutMs": 90000 }
    }
  }'

Then call it like any combo: {"model":"fusion-panel","messages":[...]}.

Virtual Auto-Combo Factory

The Auto Combo engine doesn't require pre-defined combos. Instead, open-sse/services/autoCombo/virtualFactory.ts builds candidates on-the-fly:

1. Pulls getProviderConnections({ isActive: true }) (all enabled connections)
2. Filters to those with valid credentials (API key or non-expired OAuth token via hasUsableOAuthToken())
3. Cross-references with getProviderRegistry() for model availability + pricing
4. For each tuple (provider, model, connection), builds a VirtualAutoComboCandidate
5. Picks connection.defaultModel (or the registry's first model) as the dispatch target
6. Scores each candidate using the 9-factor scorePool() and the variant's weight pack
7. Returns the resulting in-memory AutoComboConfig for handleComboChat() — never persisted to DB

This means adding a new provider with auto/* enabled automatically expands the candidate pool — no manual combo editing needed. The virtual combo is rebuilt per request, so newly-added or newly-healthy connections are picked up immediately.

Self-Healing

• Temporary exclusion: Score < 0.2 → excluded for 5 min (progressive backoff, max 30 min)
• Circuit breaker awareness: OPEN → auto-excluded; HALF_OPEN → probe requests
• Incident mode: >50% OPEN → disable exploration, maximize stability
• Cooldown recovery: After exclusion, first request is a "probe" with reduced timeout

Bandit Exploration

5% of requests (configurable) are routed to random providers for exploration. Disabled in incident mode.

API

There is no dedicated POST /api/combos/auto endpoint — Auto-Combo is consumed in two ways:

1. Zero-config (recommended): Send any chat completion request with model: "auto" or model: "auto/<variant>". The virtual factory builds the combo per request — no persistence, no API calls needed.

2. Persisted combo with strategy: "auto": Create a regular combo via POST /api/combos and set strategy: "auto" plus config.auto.weights / config.auto.candidatePool. The same scoring engine is used; the combo is stored in combos and reusable by ID.

For discovery, GET /api/combos/auto lists every variant with its resolved candidate pool plus context_length / max_output_tokens — the MAX across the candidate pool's windows. Clients (e.g. the opencode plugin) must advertise these values instead of 0: a zero context disables opencode's auto-compaction entirely, letting sessions grow until the gateway's history purge destroys context. MAX is safe to advertise because the auto-combo context pre-filter routes oversized requests to large-window candidates.

# Zero-config usage (no combo creation)
curl -X POST http://localhost:20128/v1/chat/completions \
  -H "Authorization: Bearer <key>" \
  -H "Content-Type: application/json" \
  -d '{"model":"auto/coding","messages":[{"role":"user","content":"Hello"}]}'

# Persisted auto combo via the regular combos endpoint
curl -X POST http://localhost:20128/api/combos \
  -H "Content-Type: application/json" \
  -d '{"id":"my-auto","name":"Auto Coder","strategy":"auto","config":{"auto":{"candidatePool":["anthropic","google","openai"],"weights":{"quota":0.15,"health":0.3,"costInv":0.05,"latencyInv":0.35,"taskFit":0.1,"stability":0,"tierPriority":0.05}}}}'

Auto router strategies

Persisted strategy: "auto" combos can set config.routerStrategy (or legacy config.auto.routerStrategy) to one of:

• rules — default weighted scoring
• cost / eco — cheapest healthy provider
• latency / fast — lowest p95 latency with reliability penalty
• sla-aware / sla — prefer candidates that satisfy p95 latency, error-rate, and optional cost SLOs
• lkgp — last known good provider first

Router strategies in detail

The auto-combo engine exposes 5 pluggable RouterStrategy implementations that you can swap via config.routerStrategy (or the legacy config.auto.routerStrategy). Each strategy picks one provider from the candidate pool, given a RoutingContext (task type, tool/vision hints, token estimate, optional SLA policy, optional last-known-good provider).

1. rules (default) — 6-factor weighted scoring

Wraps the existing scoring engine. Filters out OPEN circuit-breaker candidates, then runs scorePool() with the current task type and getTaskFitness(), picking the top-scoring provider.

class RulesStrategyImpl implements RouterStrategy {
  readonly name = "rules";
  readonly description =
    "6-factor weighted scoring: quota, health, cost, latency, taskFit, stability";

  select(pool, context) {
    const eligible = pool.filter((c) => c.circuitBreakerState !== "OPEN");
    const ranked = scorePool(eligible.length > 0 ? eligible : pool, context.taskType, undefined, getTaskFitness);
    return { provider: ranked[0].provider, /* ... */ };
  }
}

When to use: Default. Use when you want a balanced trade-off across all signals.

Alias: rules (no alias)

---

2. cost / eco — cheapest healthy provider

Sorts the candidate pool by costPer1MTokens (ascending) and picks the cheapest. Filters out OPEN candidates first.

class CostStrategyImpl implements RouterStrategy {
  readonly name = "cost";
  readonly description = "Always selects cheapest available provider";

  select(pool, context) {
    const healthy = pool.filter((c) => c.circuitBreakerState !== "OPEN");
    const sorted = [...healthy].sort((a, b) => a.costPer1MTokens - b.costPer1MTokens);
    return { provider: sorted[0].provider, /* ... */ };
  }
}

When to use: Cost-sensitive workloads, batch processing, or background jobs.

Aliases: cost, eco

---

3. latency / fast — lowest p95 latency with reliability penalty

Sorts by p95LatencyMs + (errorRate * 1000). The error-rate penalty ensures unreliable providers are ranked lower even if their nominal latency is low.

class LatencyStrategyImpl implements RouterStrategy {
  readonly name = "latency";
  readonly description = "Prioritizes lowest p95 latency with reliability weighting";

  select(pool, context) {
    const healthy = pool.filter((c) => c.circuitBreakerState !== "OPEN");
    const sorted = [...healthy].sort((a, b) =>
      (a.p95LatencyMs + a.errorRate * 1000) - (b.p95LatencyMs + b.errorRate * 1000)
    );
    return { provider: sorted[0].provider, /* ... */ };
  }
}

When to use: Latency-sensitive workloads like real-time chat, autocomplete, or interactive coding assistants.

Aliases: latency, fast

---

4. sla-aware / sla — latency/error/cost SLO compliance

Scores each candidate by how well it satisfies the configured SLO policy:

Factor	Weight	Formula
Latency score	35%	threshold / max(value, ε)
Error score	35%	threshold / max(value, ε)
Health score	15%	1.0 (CLOSED) / 0.5 (HALF_OPEN) / 0.0 (OPEN)
Cost score	10%	threshold / max(value, ε) or inverse normalized
Stability score	5%	inverse normalized latency stddev

When hardConstraints: true, candidates are sorted primarily by violation score (how far they exceed any SLO), then by composite score. Otherwise it's just the composite score.

class SLAStrategyImpl implements RouterStrategy {
  readonly name = "sla-aware";
  readonly description = "Selects the provider most likely to satisfy latency, error-rate, and cost SLOs";

  select(pool, context) {
    // ... scores each candidate against policy: { targetP95Ms, maxErrorRate, maxCostPer1MTokens, hardConstraints }
  }
}

SLA fields (set on the combo config):

{
  "strategy": "auto",
  "config": {
    "routerStrategy": "sla-aware",
    "slaTargetP95Ms": 1500,
    "slaMaxErrorRate": 0.05,
    "slaMaxCostPer1MTokens": 5,
    "slaHardConstraints": true
  }
}

When to use: Production workloads with strict latency, error-rate, or cost budgets.

Aliases: sla-aware, sla

---

5. lkgp — last known good provider first

Tries the last known good provider (if set) first, then falls back to the rules strategy. Useful for session stickiness — the same provider handles follow-up requests in a conversation.

class LKGPStrategyImpl implements RouterStrategy {
  readonly name = "lkgp";
  readonly description = "Tries last known good provider first, then falls back to rules";

  select(pool, context) {
    if (context.lkgpEnabled === false) {
      return getStrategy("rules").select(pool, context);
    }

    if (context.lastKnownGoodProvider) {
      const candidates = pool.filter(
        (c) => c.provider === context.lastKnownGoodProvider && c.circuitBreakerState !== "OPEN"
      );
      if (candidates.length > 0) {
        return { provider: candidates[0].provider, /* ... */ };
      }
    }

    // Fallback to rules strategy
    return getStrategy("rules").select(pool, context);
  }
}

When to use: Multi-turn conversations where you want the same provider to handle follow-up requests (e.g., for caching, context continuity, or pricing consistency).

Alias: lkgp (no alias)

---

Custom router strategies

You can register your own RouterStrategy implementation via the public API:

import { registerStrategy, type RouterStrategy } from "@omniroute/open-sse/services/autoCombo/routerStrategy";

class MyCustomStrategy implements RouterStrategy {
  readonly name = "my-custom";
  readonly description = "My custom routing strategy";

  select(pool, context) {
    // Your routing logic here
    return {
      provider: pool[0].provider,
      model: pool[0].model,
      strategy: this.name,
      reason: "MyCustomStrategy: ...",
      candidatesConsidered: pool.length,
      finalScore: 1.0,
    };
  }
}

registerStrategy("my-custom", new MyCustomStrategy());

Then use it:

{
  "strategy": "auto",
  "config": {
    "routerStrategy": "my-custom"
  }
}

---

Router strategy selection guide

Use case	Strategy	Reason
Balanced workload	rules	Default — considers all factors
Minimize cost	cost	Always picks cheapest
Minimize latency	latency	Picks fastest reliable provider
Strict SLOs	sla-aware	Filters by p95/error/cost thresholds
Multi-turn chat	lkgp	Session stickiness

SLA-aware fields:

{
  "strategy": "auto",
  "config": {
    "routerStrategy": "sla-aware",
    "slaTargetP95Ms": 1500,
    "slaMaxErrorRate": 0.05,
    "slaMaxCostPer1MTokens": 5,
    "slaHardConstraints": true
  }
}

Task Fitness

30+ models scored across 6 task types (coding, review, planning, analysis, debugging, documentation). Supports wildcard patterns (e.g., *-coder → high coding score).

Auto Variants Recap

Including the bare auto (default) plus the 6 AutoVariant values declared in autoPrefix.ts, there are 7 invokable model IDs:

auto, auto/coding, auto/fast, auto/cheap, auto/offline, auto/smart, auto/lkgp

(AutoVariant itself enumerates 6 values; the 7th option is "no variant" — bare auto — handled by parseAutoPrefix() as variant: undefined.)

How tiers fit Auto-Combo

The 9-factor scoring function (open-sse/services/autoCombo/scoring.ts) treats tier membership as one signal via the tierPriority weight. Default weights (from DEFAULT_WEIGHTS):

Factor	Default weight	Notes
Tier priority	0.05	Tier 1 premium → higher score
Latency (p50 inverse)	0.35	Fastest wins
Cost ($/1M inverse)	0.20	Cheapest blended price wins (60% input + 40% output ratio)
Recent health/error rate	0.15	Unhealthy deprioritized
Quota remaining	0.10	Near-exhausted deprioritized
Context window match	0.08	Penalizes short windows
Task fitness	0.10	Coding → coding-specialist models
Stability	0.00	Disabled by default

Tier alone does not force Tier 1 first — if Tier 1 latency is bad or cost-vs-quality is suboptimal, Tier 2 wins. To force tier ordering, use combo strategy priority and arrange providers by tier.

To strongly favor Tier 1 (subscription), increase tierPriority weight:

{
  "strategy": "auto",
  "config": { "auto": { "weights": { "tierPriority": 0.3, "costInv": 0.05 } } }
}

See docs/marketing/TIERS.md for tier definitions and provider classification.

Testing & Coverage

Deterministic routing-decision matrix (npm run test:combo:matrix)

tests/integration/combo-matrix/*.test.ts proves the routing decision of all 17 public strategies end-to-end through the real combo pipeline with a mocked upstream. Coverage includes:

• All 17 ROUTING_STRATEGY_VALUES strategies (ordered, weighted, cost, context, fusion, …).
• quota-share (internal) end-to-end: DRR fairness + saturation deprioritization via the real selectQuotaShareTarget seam (registerQuotaFetcher / setLKGP / __setHeadroomSaturationFetcherForTests).
• context-relay universal-handoff coverage across every target count.

This suite runs in CI (test:integration job) with --test-concurrency=1 and --test-force-exit so it is deterministic and does not require live credentials.

Gated live smoke (NOT in CI — real providers)

Command	What it does
npm run test:combo:live	In-process real routing with RUN_COMBO_LIVE=1; snapshots a live OmniRoute DB
npm run test:combo:live:vps	HTTP calls against a live OmniRoute server (set COMBO_LIVE_BASE_URL)
npm run test:combo:live:vps:failover	Same, with deliberate failover scenarios

These smoke tests exercise the real wire path (combo → provider → completion). They are intentionally excluded from CI because they require live credentials and VPS access.

---

Files

File	Purpose
open-sse/services/autoCombo/scoring.ts	9-factor scoring function, DEFAULT_WEIGHTS, pool norm
open-sse/services/autoCombo/taskFitness.ts	Model × task fitness lookup
open-sse/services/autoCombo/engine.ts	Selection logic, bandit, budget cap
open-sse/services/autoCombo/selfHealing.ts	Exclusion, probes, incident mode
open-sse/services/autoCombo/modePacks.ts	4 weight profiles (ship-fast, cost-saver, quality-first, offline-friendly)
open-sse/services/autoCombo/autoPrefix.ts	auto/ prefix parser + 6 variants
open-sse/services/autoCombo/virtualFactory.ts	Builds in-memory AutoComboConfig from live connections
open-sse/services/autoCombo/providerRegistryAccessor.ts	Test hook for mocking provider registry
src/shared/constants/routingStrategies.ts	ROUTING_STRATEGY_VALUES (17 strategies)
src/sse/handlers/chat.ts	Integration: auto-prefix short-circuit