Struct LlmCallConfig 

pub struct LlmCallConfig {

    pub model: String,
    pub temperature: Option<f32>,
    pub max_tokens: Option<u32>,
    pub tools: Vec<ToolDefinition>,
    pub reasoning_effort: Option<String>,
    pub speed: Option<String>,
    pub verbosity: Option<String>,
    pub metadata: HashMap<String, String>,
    pub previous_response_id: Option<String>,
    pub tool_search: Option<ToolSearchConfig>,
    pub prompt_cache: Option<PromptCacheConfig>,
    pub openrouter_routing: Option<OpenRouterRoutingConfig>,
    pub parallel_tool_calls: Option<bool>,
    pub volatile_suffix_len: usize,
}

Configuration for an LLM call

Fields

model: String

temperature: Option<f32>

max_tokens: Option<u32>

tools: Vec<ToolDefinition>

reasoning_effort: Option<String>

Reasoning effort level (for models that support it: low, medium, high)

speed: Option<String>

Speed (service tier) for this call: “flex”, “default”, or “priority”. Serialized as OpenAI service_tier; omitted when None so the provider keeps its default (“auto”) routing.

verbosity: Option<String>

Verbosity for this call: “low”, “medium”, or “high”. Serialized as OpenAI verbosity; omitted when None so the provider keeps its default (“medium”) output length.

metadata: HashMap<String, String>

Metadata to send with the API request for tracking and debugging. Keys and values are strings. Both OpenAI and Anthropic support metadata fields. Typically includes: session_id, agent_id, org_id, turn_id, exec_id.

previous_response_id: Option<String>

Previous response ID for stateful continuation (OpenAI Responses API). When set, the provider can skip re-encoding cached context.

tool_search: Option<ToolSearchConfig>

Tool search configuration for deferred tool loading

prompt_cache: Option<PromptCacheConfig>

Prompt caching configuration for provider-specific cache controls.

openrouter_routing: Option<OpenRouterRoutingConfig>

OpenRouter-only model fallback and provider routing controls.

parallel_tool_calls: Option<bool>

Request-level parallel tool calling preference (EVE-598).

Serialized onto the provider request when Some(_): OpenAI sets parallel_tool_calls; Anthropic maps Some(false) → tool_choice.disable_parallel_tool_use = true. None preserves provider defaults (no field sent).

volatile_suffix_len: usize

Number of trailing messages that are volatile (regenerated every turn) and must not anchor a message-level prompt-cache breakpoint.

ReasonAtom sets this to the count of live <facts> messages it appends at the conversation tail. Drivers that place a message cache breakpoint on the last block (Anthropic) skip this many trailing messages so the breakpoint lands on the last stable block — otherwise a tail that changes each turn would evict the conversation-history cache. 0 (the default) preserves the previous behavior exactly.

Implementations

impl LlmCallConfig

pub fn resolved_parallel_tool_calls(&self, supported: bool) -> Option<bool>

Resolve the effective wire value for parallel_tool_calls, gated by whether the driver/model can express it on the request.

Returns None (omit the field, keep the provider default) when the preference is unset or supported is false. Drivers call this with self.supports_parallel_tool_calls(&config.model) so the preference is only serialized where the provider has a control for it. The local tool scheduler honors the preference independently, so Some(false) still serializes execution even when this returns None.

Trait Implementations

impl Clone for LlmCallConfig

fn clone(&self) -> LlmCallConfig

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for LlmCallConfig

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

Formats the value using the given formatter.