Struct ProviderConfig 

pub struct ProviderConfig {

    pub kind: String,
    pub model: Option<String>,
    pub api_key_env: Option<String>,
    pub api_key: Option<String>,
    pub base_url: Option<String>,
    pub input_price_per_mtok: Option<f64>,
    pub output_price_per_mtok: Option<f64>,
    pub temperature: Option<f64>,
    pub seed: Option<u64>,
    pub context_window: Option<u64>,
    pub vision: Option<bool>,
    pub max_retries: Option<u32>,
    pub retry_after_cap_secs: Option<u64>,
    pub fallbacks: Vec<String>,
}

Fields

kind: String

anthropic | openai | local

model: Option<String>

api_key_env: Option<String>

Environment variable holding the key. Preferred over api_key.

api_key: Option<String>

Inline key. Convenient, but it lands in a file on disk — prefer the env var.

base_url: Option<String>

input_price_per_mtok: Option<f64>

Per-million-token prices, so budgets and reporting can be in dollars. Leave unset for a local model — the marginal cost really is zero.

output_price_per_mtok: Option<f64>

temperature: Option<f64>

Sampling temperature, sent verbatim by providers that accept one. Unset means the server’s default. Do not reach for 0.0 to get repeatability: measured on qwen3.6, greedy decoding walks into verbatim repetition loops that sampling noise would have broken. Pin the server’s own default value and set seed instead — same distribution, repeatable draws. The Anthropic API rejects the parameter, so setting this on an anthropic provider is a startup error rather than a silent no-op.

seed: Option<u64>

Sampling seed, for repeatable draws at a nonzero temperature. Only as deterministic as the backend: llama-server repeats exactly when requests run one at a time, and does not once concurrent requests share a batch. Rejected on anthropic for the same reason as temperature.

context_window: Option<u64>

How many tokens this model’s context holds — for a local server, the -c it was started with.

Nothing here can discover this: a provider reports how many tokens a prompt used, never how many are left. Without it the compaction threshold has to be an absolute number somebody remembers to set, and when nobody does, a long session dies on a raw exceed_context_size_error from the server with the whole run lost. With it, AgentConfig::compact_at derives a threshold and the CLI can show how much room is left.

vision: Option<bool>

Whether the model on the other end can see an image.

Declared rather than discovered, for the same reason context_window is: the Anthropic API has no endpoint that answers it. llama-server does — GET /props reports modalities.vision — so preflight checks the two against each other and warns on a disagreement in either direction. Both directions matter and for different reasons: declared-but-not-served means every image silently degrades to a line of text, and served-but-not-declared means a projector is loaded, paid for in memory, and never used.

Unset defaults to true for kind = "anthropic" (every Claude model in the family sees) and false everywhere else. False is the safe default for a local server because the failure it prevents is the expensive one: an image_url part sent to a text-only llama-server is a failed request, where an image rendered as text is merely a model that cannot see — which is what it was before.

A vision model is two files. The weights carry the language model; the vision tower ships beside them as a separate mmproj-*.gguf that --mmproj must name. --mmproj-auto is on by default and only fires for -hf downloads, so every start script here using -m <path> gets nothing from it. See docs/LLAMA-SERVER.md.

max_retries: Option<u32>

Retries per request on transient failures — 429, 5xx, transport. 0 disables. Unset means 3. Auth, billing, invalid-request and context-overflow errors are never retried: the same payload fails the same way, and overflow belongs to the compaction path.

retry_after_cap_secs: Option<u64>

A Retry-After above this many seconds is surfaced as a failure instead of slept through (default 60) — a provider can name a wait long enough that the process is simply asleep, and control never returns to a layer that could fall back instead.

fallbacks: Vec<String>

Provider entries to try, in order, when this one exhausts its retries on a transient failure. Turn-local: the next turn starts from this provider again. Each fallback answers with its own model. Empty — the default — means strict: fail rather than silently answer with a different model. mecha eval never falls back regardless: a scorecard grades the model it names.

Implementations

impl ProviderConfig

pub fn vision_enabled(&self) -> bool

Whether to render images onto this provider’s wire.

The default is per-kind rather than a flat false because the two kinds know different amounts: every model in the Anthropic family this harness speaks to has vision, and nothing about a local server is knowable from config alone.

pub fn pricing(&self) -> Option<Pricing>

Prices, if configured. Both halves are required: knowing one is worse than knowing neither, because it silently under-reports.

pub fn resolve_api_key(&self) -> Option<String>

Trait Implementations

impl Clone for ProviderConfig

fn clone(&self) -> ProviderConfig

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for ProviderConfig

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

Formats the value using the given formatter.

impl Default for ProviderConfig

fn default() -> ProviderConfig

Returns the “default value” for a type.

impl<'de> Deserialize<'de> for ProviderConfig

where ProviderConfig: Default,

fn deserialize<__D>(__deserializer: __D) -> Result<Self, __D::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Serialize for ProviderConfig

fn serialize<__S>(&self, __serializer: __S) -> Result<__S::Ok, __S::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.