Kind: Type
Source: atloria-monorepo/libs/agent-core/src/runtime/model-router.ts
Model Router — Phase 2
Maps task categories to model deployment names. Enables cost-optimized routing: cheap models for simple tasks, full models only where quality matters.
Resolution order per call:
- Explicit model passed at spawn time (highest priority — always wins)
- Category route from routing table
- Default model (fallback — zero behavior change when no table configured)
All models must be deployments on the same Azure endpoint. The router only returns model names — the provider config is unchanged.
Default routing table for atlas-docs-resource (East US 2): explore / synthesize / tool_dispatch → gpt-5.6-luna (150M TPM quota) analyze / review / chat → gpt-5.6-luna (10M TPM quota) plan / implement → gpt-5.6-terra (full reasoning, 1M context) default (uncategorized) → gpt-5.6-terra (no behavior change)
Definition
ts| 'explore' // file search, directory listing, reading code — nano
| 'analyze' // code understanding, grep, web search — mini
| 'plan' // architecture, design decisions — full
| 'implement' // writing/editing code — full
| 'review' // code review, quality chec…
Was this page helpful?