Struct ModelConfig 

pub struct ModelConfig {
    pub model: String,
    pub base_url: String,
    pub api_backend: ApiBackend,
    pub auth: AuthScheme,
    pub anthropic_version: String,
    pub betas: Vec<String>,
    pub extra_headers: Vec<(String, String)>,
    pub max_tokens_cap: Option<u32>,
    pub developer_rendering: DeveloperRendering,
    pub provider_prefs: Option<Value>,
}

The per-model wire configuration record (ADR-0007 + plan §3.1/§9).

Fields

model: String

The model id (native claude-…; OpenRouter anthropic/claude-…).

base_url: String

The endpoint base URL ({base_url}/v1/messages is the request path).

api_backend: ApiBackend

The endpoint family — selects the auth header and backend quirks.

auth: AuthScheme

The resolved credential.

anthropic_version: String

The anthropic-version header value.

betas: Vec<String>

anthropic-beta values, latched per session (plan §4.8). Defaults to INTERLEAVED_THINKING_BETA (plan §9.3).

extra_headers: Vec<(String, String)>

Extra headers appended to every request (gateway auth quirks etc.).

max_tokens_cap: Option<u32>

Optional ceiling clamped onto SamplingArgs.max_tokens. None (the default) sends the caller’s budget through untouched.

A clamp here is silent by construction — it is a min, so a caller who deliberately asks for more gets less with no error and no way to notice. ADR-0007 already rejects that for reasoning_effort (“never silently clamp — that would corrupt eval comparisons”); the same reasoning applies to the output budget, so the wire now forwards the request and lets the API’s own error surface when a model will not accept it.

Set it to Some(n) only to pin a model whose real ceiling is lower than DEFAULT_MAX_TOKENS — e.g. claude-3-haiku at 4096 — where the clamp turns a 400 into a working request.

developer_rendering: DeveloperRendering

How Developer messages are rendered.

provider_prefs: Option<Value>

OpenRouter provider routing preferences. None → the default trio (ignore: ["amazon-bedrock"], allow_fallbacks: false, require_parameters: true, matching cc-reverse-proxy) when the backend is OpenRouter; ignored on other backends.

Vertex is deliberately not excluded — it is a production-relevant Anthropic provider (user decision, 2026-07-18). Note from the Task-12 live smoke: Vertex honours the thinking config only when the interleaved-thinking beta header reaches it, and cross-provider routing between turns forfeits prompt-cache reads; pin only: ["anthropic"] here when cache-hit determinism matters more than provider failover.

Implementations

impl ModelConfig

pub fn new( model: impl Into<String>, base_url: impl Into<String>, key: impl Into<String>, ) -> ModelConfig

Build a config for model against base_url with key, detecting the backend and choosing the matching auth scheme.

pub fn from_env() -> Result<ModelConfig, ProviderError>

Resolve the common case from the environment: LOCODE_API_KEY (required), LOCODE_BASE_URL (default DEFAULT_BASE_URL); the model defaults (default DEFAULT_MODEL).

Errors

ProviderError::Auth when LOCODE_API_KEY is unset or empty.

pub fn interleaved_thinking(&self) -> bool

Whether the interleaved-thinking beta is active (waives the thinking budget clamp, plan §9.3).

pub fn effective_provider_prefs(&self) -> Option<Value>

The OpenRouter provider preferences to inject, if any (plan §9.2): the configured value, or the default trio on the OpenRouter backend.

Trait Implementations

impl Clone for ModelConfig

fn clone(&self) -> ModelConfig

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for ModelConfig

fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error>

Formats the value using the given formatter.