Overview
The model router inspects the `taskType` field on each check and maps it to a model family (haiku/sonnet/opus). It then picks the cheapest provider that satisfies the tier and has remaining budget.
Tier Table
| Task type | Family | Anthropic model | OpenAI model | Cost / 1 M tokens |
|---|---|---|---|---|
| `file-analysis` | haiku | claude-haiku-4-5 | gpt-4o-mini | $0.80 |
| `code-generation` | sonnet | claude-sonnet-4-5 | gpt-4o | $3.00 |
| `code-review` | sonnet | claude-sonnet-4-5 | gpt-4o | $3.00 |
| `architecture-decision` | opus | claude-opus-4-5 | gpt-4o | $15.00 |
| `security-review` | opus | claude-opus-4-5 | gpt-4o | $15.00 |
Budget Enforcement
Set `budgetUsd` in your DAG or pass `--budget 1.00` to the CLI. When the remaining budget falls below the cost of the next check, the router falls back to a cheaper model tier. If no tier fits, the lane is suspended and a `budget-exceeded` event fires.
Supported Providers
- Anthropic — Claude 3 / 4 family via SSE - OpenAI — GPT-4o family via SSE + stream_options - VS Code Sampling — Copilot routing for in-editor workflows - Mock — Built-in deterministic, zero-cost, zero-key provider - Ollama — Local models (Enterprise) - Bedrock — AWS-hosted Claude (Enterprise) - Gemini — Google AI (Enterprise)
