LLM model catalog

Schema: docs/schemas/runtime/llm-model-catalog.schema.yaml Data: runtime/src/mas/runtime/spec/llm-model-catalog.yaml (metadata.version) Loader: mas.runtime.engine.llm_model_catalog

Per-model context window, max output, defaults, and allowed settings — the same shape we want for a small reusable library. This repo ships a curated YAML, not a vendor dump.

Why not import LiteLLM wholesale

LiteLLM model_prices_and_context_window.json is the closest reusable community table (MIT). It is large, changes daily, and is not a MAS schema. The catalog lists it under spec.sources with reusable: true so a later mas-llm-models package can wrap that JSON behind the same ModelCatalog.get(name) API.

Other pointers (not in-process libraries):

Source	URL	Reuse
OpenRouter models API	https://openrouter.ai/api/v1/models	live HTTP (supported_parameters)
models.dev	https://models.dev	aggregated community catalog
Hugging Face model cards	https://huggingface.co/models	cards / TGI
vLLM OpenAI server	https://docs.vllm.ai/en/latest/serving/openai_compatible_server.html	think / chat_template_kwargs
Ollama API	https://github.com/ollama/ollama/blob/main/docs/api.md	think
Aider MODEL_SETTINGS	https://github.com/Aider-AI/aider/blob/main/aider/resources/model-settings.yml	Aider-specific YAML
Continue.dev	https://github.com/continuedev/continue	product configs
OpenClaw	historical MAS extract	not in this tree as a capability table — git history moved OpenClaw plugins (memory, telemetry), not a model list

Each catalog settings.<name> entry may carry supported, default, min, max, and enum (the allowed list). When the spec omits a field, catalog defaults are applied. Numeric spec values are clamped to min/max when the model is known. Reasoning knobs (reasoning.*, think, include) are deny-by-default on known models: omitting the setting means unsupported. Unknown models keep the spec as written. Do not use a default alias — it collides with last-path-segment lookup.

Shape

apiVersion: mas/v1
kind: LlmModelCatalog
metadata:
  name: standard
  version: "2026.9.2"
spec:
  sources: [...]
  models:
    gpt-5:
      aliases: [openai/gpt-5]
      context_window: 400000
      max_output_tokens: 128000
      defaults:
        reasoning.effort: medium
      settings:
        reasoning.effort: {enum: [none, minimal, low, medium, high, xhigh]}
        reasoning.mode: {enum: [standard, pro]}
        think: {supported: false}

ModelCatalog.get("openai/gpt-5") resolves aliases. Unknown models are permitted: the spec still sends what the user set; the catalog does not invent a deny-by-default filter.

Bump metadata.version when curated numbers change.