Module model 

Named model registry shared by all agents.

On disk: ~/.mur/models.yaml. Schema:

schema_version: 1
models:
  anthropic_opus_4_7:
    provider: anthropic
    model: claude-opus-4-7
    secret: env:ANTHROPIC_API_KEY
    capabilities: [chat, tools]

Structs

ModelEntry

ModelRegistry

RoleEntry

Enums

BillingMode

Who pays when this model answers.

Requirement

A capability the request needs from whatever model serves it. Derived from the request itself (an image in the messages, a tool list) and never from config: a router may only substitute a model that can do the job.

Constants

CAP_CHAT

Registry capability strings. The baseline (chat) is legacy-permissive — an entry with no capabilities at all predates the field and is assumed chat-capable. Everything above the baseline is fail-closed.

CAP_TOOLS

CAP_VISION

Functions

choose_by_difficulty

Opt-in difficulty heuristic: pick frontier when the estimated input token count exceeds the threshold, else cheap. None when misconfigured (caller falls through to model_ref/global default).

pick_cheap_model

Pick the cheapest registry entry that can serve a request needing reqs, excluding exclude (the agent’s own primary). None when no qualifying entry exists → caller keeps normal candidates (fail-expensive).

resolve_model_refs

Build the ordered list of model_refs to try: [primary, ...fallback]. Priority per-agent → global. The primary is de-duplicated out of the chain (no point retrying the same ref back-to-back). Returns empty when nothing is configured, so the caller keeps today’s single-inline-model behaviour.

satisfies

Can this entry serve a request needing reqs?