Struct ModelProviderRuntimeOptions 

pub struct ModelProviderRuntimeOptions {

    pub auth_profile_override: Option<String>,
    pub provider_kind: Option<String>,
    pub provider_api_url: Option<String>,
    pub zeroclaw_dir: Option<PathBuf>,
    pub secrets_encrypt: bool,
    pub reasoning_enabled: Option<bool>,
    pub reasoning_effort: Option<String>,
    pub reasoning_effort_passthrough: bool,
    pub provider_timeout_secs: Option<u64>,
    pub extra_headers: HashMap<String, String>,
    pub api_path: Option<String>,
    pub provider_max_tokens: Option<u32>,
    pub merge_system_into_user: bool,
    pub provider_extra: Option<Value>,
    pub replay_assistant_reasoning: Option<bool>,
    pub thinking_passthrough: bool,
    pub cache_passthrough: bool,
    pub cache_ttl: Option<CacheTtl>,
    pub native_tools: Option<bool>,
    pub wire_api: Option<String>,
    pub think: Option<bool>,
    pub vision: Option<bool>,
    pub chat_template_kwargs: Option<Value>,
    pub tls_ca_cert_path: Option<String>,
    pub multimodal: MultimodalConfig,
    pub tool_result_image_policy: ToolResultImagePolicy,
}

Fields

auth_profile_override: Option<String>

provider_kind: Option<String>

Explicit provider implementation from [providers.models.<family>.<alias>].kind. When unset, provider resolution falls back to the configured family.

provider_api_url: Option<String>

zeroclaw_dir: Option<PathBuf>

secrets_encrypt: bool

reasoning_enabled: Option<bool>

reasoning_effort: Option<String>

reasoning_effort_passthrough: bool

Forward the runtime-configured reasoning_effort to every model on OpenAI-compatible providers, bypassing the OpenAI-reasoning-family name filter. Propagated from ModelProviderConfig::reasoning_effort_passthrough.

provider_timeout_secs: Option<u64>

HTTP request timeout in seconds for LLM model_provider API calls. None uses the model_provider’s built-in default (120s for compatible model_providers).

extra_headers: HashMap<String, String>

Extra HTTP headers to include in model_provider API requests.

api_path: Option<String>

Custom API path suffix for OpenAI-compatible model_providers (e.g. “/v2/generate” instead of the default “/chat/completions”).

provider_max_tokens: Option<u32>

Maximum output tokens for LLM model_provider API requests. None uses the model_provider’s built-in default.

merge_system_into_user: bool

When true, system messages are merged into the first user message before sending. Propagated from ModelProviderConfig::merge_system_into_user.

provider_extra: Option<Value>

Extra JSON parameters merged into API request bodies at the top level. Propagated from ModelProviderConfig::provider_extra.

replay_assistant_reasoning: Option<bool>

When Some(false), strip assistant reasoning fields from outbound history replay. None honours provider default.

thinking_passthrough: bool

Forward Anthropic extended thinking through OpenAI-compatible providers: inject the runtime thinking params as an Anthropic-shaped thinking request object and normalize gateway thinking responses for replay. Propagated from ModelProviderConfig::thinking_passthrough.

cache_passthrough: bool

Forward Anthropic prompt caching through OpenAI-compatible providers: inject cache_control breakpoints (system prompt; rolling last message) into request bodies and capture gateway-reported cache usage. Propagated from ModelProviderConfig::cache_passthrough.

cache_ttl: Option<CacheTtl>

Prompt-cache entry lifetime for providers that place Anthropic cache markers (native Anthropic; compatible ones behind cache_passthrough). None keeps the 5-minute default. Propagated from ModelProviderConfig::cache_ttl.

native_tools: Option<bool>

When set, the provider is asked to use its native tool-calling schema instead of OpenAI-compat tool calls. Generic across families.

wire_api: Option<String>

Wire protocol to use for this provider. Some("responses") routes the provider through the OpenResponses /v1/responses API instead of chat_completions. None uses the provider’s built-in default (chat_completions for most providers).

think: Option<bool>

Enable or disable chain-of-thought thinking. Forwarded as enable_thinking in the request body. None lets the model decide.

vision: Option<bool>

Override the provider’s vision (image input) capability. None keeps the provider family’s built-in default; Some(false) marks a text-only model served by a vision-capable family as non-vision. Mapped from ModelProviderConfig::vision.

chat_template_kwargs: Option<Value>

Forwarded verbatim as a top-level chat_template_kwargs object in the request body for OpenAI-compatible providers (folded into extra_body by apply_compat_options). Consumed by chat-template-aware backends such as vLLM, SGLang, and llama.cpp.

tls_ca_cert_path: Option<String>

Path to a custom CA certificate file for TLS connections.

multimodal: MultimodalConfig

The configured [multimodal] policy.

Providers normalize image markers on their own boundary (a channel can call chat with raw [IMAGE:<path>] markers that never passed through the runtime), and that normalization now decodes pixels and applies max_images / max_image_size_mb. Carrying the configured policy here keeps that second pass on the same rules the runtime already applied, instead of silently reverting to defaults and re-trimming history a configured request had legitimately accepted.

tool_result_image_policy: ToolResultImagePolicy

How compatible chat-completions providers handle images a native role=tool result declared in its attachments array; legacy tool results (no array key) pass verbatim under both settings.

Trait Implementations

impl Clone for ModelProviderRuntimeOptions

fn clone(&self) -> ModelProviderRuntimeOptions

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for ModelProviderRuntimeOptions

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

Formats the value using the given formatter.

impl Default for ModelProviderRuntimeOptions

fn default() -> Self

Returns the “default value” for a type.