Graphorin

Appearance

Graphorin API reference v0.15.1

Graphorin API reference / @graphorin/agent / preferred-model

preferred-model ​

Per-tool / per-agent preferred-model resolution. Pure functions consulted by the agent loop AFTER the model has decided which tool(s) to call but BEFORE provider.stream(...) is invoked.

The four-step precedence ladder (highest wins):

  1. prepareStep({ provider }) - the operator's explicit per-step override always wins.
  2. Tool.preferredModel - the tool author's per-tool hint. Only the tools the model actually CALLED on the previous step are consulted - an advertised-but-uncalled hint never escalates the run. Multi-tool ties resolve to the highest cost tier ('smart' > 'balanced' > 'fast'; explicit ModelSpec is treated as the highest tier).
  3. Agent.preferredModel? - the per-agent default.
  4. Agent default provider - the v0.1-alpha behaviour.

Cost-tier resolution against Agent.modelTierMap is documented as a hint: when the requested tier is unmapped, the resolver falls through to the next precedence step rather than throwing.

Interfaces ​

InterfaceDescription
PreferredModelResolutionResult returned by resolvePreferredModel.
ResolvePreferredModelInputPure inputs to resolvePreferredModel.

Functions ​

FunctionDescription
pickTopTierAcrossToolsPick the highest-cost tier across the supplied per-tool hints. Explicit ModelSpec entries are treated as the highest tier ('smart') for tie-breaking - the conservative-correctness rule documented in DEC-169.
resolvePreferredModelWalk the precedence ladder and return the resolved provider for a single agent step. Pure function - no side effects.