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

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).

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

Per-key candidate control (#7819, Level 1+2)

GET /v1/auto-combo/{channel}/candidates ({channel} = the suffix after auto/, or the literal auto for the base channel) is a read-only endpoint that lists an auto/* channel's current candidate pool decorated with live reachability, reusing the existing resilience reads (never raw breaker state):

• provider circuit breaker — getCircuitBreaker(provider).getStatus() / .canExecute()
• connection cooldown — rateLimitedUntil / testStatus on the resolved provider_connections row
• model lockout — isModelLocked(provider, connectionId, model)

Each candidate also carries this API key's excluded flag. Exclusions are stored per-API-key (auto_candidate_overrides table, migration 128) — OmniRoute is single-tenant with no users table, so apiKeyId is the closest real per-caller identity — and enforced at the candidate-pool chokepoint in open-sse/services/autoCombo/virtualFactory.ts via the pure, unit-tested filterExcludedCandidates() (open-sse/services/autoCombo/candidateOverrides.ts). The filter is fail-open: an unset apiKeyId/channel or a DB lookup failure both leave the pool unfiltered, so an operator with no overrides configured sees routing byte-identical to before this feature.

Deferred to a follow-up issue: per-candidate weights + explicit ordering (Level 3 — feeds into the existing weighted/priority strategy paths) and pinning a specific combo.ts strategy per auto/* channel (Level 4). See the #7819 plan for the open question on whether overrides should stay per-API-key or become global given the single-tenant model.

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:

Combo Names That Match a Real Model Id

A combo whose name is identical to a bare model id (e.g. a combo named gpt-5.5) is an intentional, supported pattern, not a bug: it is the mechanism for per-model-id provider fallback documented in #6940. Because combo resolution is checked before bare-model-id resolution (getComboForModel() in src/sse/services/model.ts), a request for the bare id gpt-5.5 is routed through the combo's targets (e.g. acme-responses/gpt-5.5, backup-responses/gpt-5.5) instead of straight to a single provider — this reuses the combo-before-rewrite precedence built for #3227/#3233 and is regression-tested by tests/unit/responses-combo-resolution-3227.test.ts and tests/unit/combo-name-codex-responses-rewrite.test.ts.

Creating or renaming a combo to a name that shadows a real model id is never rejected — doing so would break this documented workflow. Instead (#8530), POST /api/combos and PUT /api/combos/[id] attach a non-blocking warning field to the response when the (new) name collides with a real model id:

{ "warning": { "code": "COMBO_NAME_SHADOWS_MODEL", "modelId": "gpt-5.5", "providerId": "openai" } }

At boot, scanComboModelNameCollisionsAtBoot() (src/instrumentation-node.ts) also logs a one-line [STARTUP] warning enumerating every existing combo that shadows a model id, so operators who hit this by accident (rather than intentionally, per #6940) have a signal. The detection helper lives in src/lib/combos/modelNameCollision.ts.

How It Works (Persisted Auto-Combos)

The Auto-Combo Engine dynamically selects the best provider/model for each request using a 14-factor scoring function (defined in open-sse/services/autoCombo/scoring.ts → DEFAULT_WEIGHTS). Weights form a normalized distribution (custom weights are renormalized by normalizeScoringWeights()).

Source: diagrams/auto-combo-12factor.mmd (regenerate via npm run docs:render-diagrams). The filename predates the current factor set; the diagram shows 13 of the 14 factors (missing sessionAvailability).

Sum: 0.20 + 0.15 + 0.15 + 0.12 + 0.08 + 0.05 + 0.05 + 0.05 + 0.05 + 0.05 + 0.05 + 0.05 + 0.00 + 0.00 = 1.05 as literally declared in DEFAULT_WEIGHTS; user-configured weights are renormalized into a distribution by normalizeScoringWeights() before scoring.

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).

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)

Per-Request Controls (headers) — #6023 / #6024 / #6025 / #3470

An auto combo can be steered per request via three headers, without mutating the combo's stored config. These apply only to the auto strategy and only for the request that carries them; the combo's saved modePack/budgetCap/budgetFallback are used when the header is absent.

# Force the fastest profile, cap this request at $0.05, and hard-block instead of overspending
curl -sS http://localhost:20128/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "X-OmniRoute-Mode: fast" \
  -H "X-OmniRoute-Budget: 0.05" \
  -H "X-OmniRoute-Budget-Fallback: strict" \
  -d '{"model":"auto","messages":[{"role":"user","content":"hi"}]}'

Resolution is a pure function (open-sse/services/autoCombo/requestControls.ts); the resolved values feed the engine's existing config.modePack / config.budgetCap / config.budgetFallback inputs. A combo's stored config.budgetFallback ("strict" | "cheapest") sets the persistent policy; the header overrides it for a single request.

All Routing Strategies

OmniRoute's combo engine supports 19 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.

⭐ = 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:

0. Tool-bearing bypass — a request that carries a non-empty tools array with tool_choice not explicitly "none" skips the panel entirely: it routes directly to a single model (the configured judge, or panel[0]) with tools/tool_choice passed through unmodified. Panel members have no tool access and the judge's synthesis directive discourages tool-call emission, so agentic/tool-calling clients get a real tool-call decision instead of synthesized prose (#6771).
1. Fan-out (non-tool-bearing requests only) — 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.

A panel member may also be a combo-ref step ({kind: "combo-ref", comboName: "..."}) referencing another combo — it resolves as one black-box panel voice (a full recursive dispatch into the referenced combo, not a fan-out of that combo's own targets), with the same depth/cycle protection every other combo-ref-consuming strategy already uses (#6764).

Configuration

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

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:

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

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 14-factor scoring function (open-sse/services/autoCombo/scoring.ts) treats tier membership as two signals: tierPriority (0.05) and tierAffinity (0.05). See the canonical scoring factor table above for the full DEFAULT_WEIGHTS set — the per-pack overrides (ship-fast/cost-saver/quality-first/ offline-friendly) are listed in the "Weight profiles per pack" table.

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 19 public strategies end-to-end through the real combo pipeline with a mocked upstream. Coverage includes:

• All 19 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)

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