Mixr routing and cost

The harness’s Mixr meta-agent picks { model, provider, maxTokens, reason } per node or job. It is Cursor Auto-shaped in role, not in implementation: policy and catalog live in D inside the harness; classify calls hit a local OpenAI-compatible router when possible.

On-device router (default)

MIXR_ROUTER = on-device | api | auto
MIXR_ROUTER_BASE_URL = http://127.0.0.1:11434/v1
MIXR_ROUTER_MODEL = <local router model id>
MIXR_ROUTER_RAM_MIN_MB = 2048

  • on-device (default): Mixr POSTs a short classify prompt to $MIXR_ROUTER_BASE_URL/chat/completions.

  • api: skip local classify; use catalog/router-self for the Mixr self-call only.

  • auto: same as on-device when available RAM ≥ MIXR_ROUTER_RAM_MIN_MB and the local server answers; otherwise api/heuristic catalog.

Override available RAM for tests with HARNESS_AVAIL_RAM_MB.

Weights may come from anywhere (including models once published via Hugging Face) as long as they run under a non-Python local runner (GGUF/ONNX/llama.cpp, etc.). What the harness refuses is in-process RouteLLM / PyTorch / Transformers glue.

Suppress lists

MIXR_SUPPRESS_PROVIDERS = anthropic,...
MIXR_SUPPRESS_MODELS = ...
MIXR_ALLOW_SUPPRESSED = off | explicit-only

Defaults suppress anthropic. Automatic routes never select a suppressed provider or model. Only an explicit per-invoke override (allowSuppressed) may use them when MIXR_ALLOW_SUPPRESSED=explicit-only.

Catalog tiers (heuristic)

| Tier | Role | Default examples | | --- | --- | --- | | 1 | Flash / diffs / explore | GLM-5.3-Flash, DeepSeek chat | | 2 | Agentic multi-file | MiniMax M3, Kimi K3, GLM-5.3 | | 3 | Architecture / scaffolding | GLM-5.3 at high reasoning budget |

Anthropic is not in the default catalog.

Offline cost lanes (not interactive Mixr)

Batch APIs, prompt-cache reads, and surplus open-weight endpoints help background indexing and bulk jobs. They are not the interactive Mixr path — latency and dropouts make them wrong for turn loops.

OpenAI-compat seam

harness serve exposes:

  • GET /api/health — readiness + Mixr status

  • GET /v1/models — catalog ids

  • POST /v1/chat/completions — returns a Mixr RoutePlan as assistant content (stub until live completions wire)

  • POST /api/provider/session / POST /api/provider/turn — t3code HarnessDriver seam

Related