pub struct LlmQuery {
pub id: QueryId,
pub prompt: String,
pub system: Option<String>,
pub max_tokens: u32,
pub grounded: bool,
pub underspecified: bool,
pub cache_breakpoint: Option<String>,
}Expand description
LLM request emitted during execution. Transport-agnostic (no channel, HTTP, or MCP Sampling details).
Fields§
§id: QueryId§prompt: String§system: Option<String>§max_tokens: u32§grounded: boolWhen true, the host should ground the response in external evidence (web search, code reading, documentation, etc.) rather than relying solely on LLM internal knowledge. The host decides the means.
underspecified: boolWhen true, the prompt’s preconditions depend on intent/goal definitions that exist outside the current context and cannot be inferred by the LLM. The host decides the resolution means (user query, RAG, DB lookup, delegated agent, etc.).
cache_breakpoint: Option<String>Optional opaque hint to the host on where to place a prompt-cache
boundary (e.g. Anthropic cache_control block). The engine does not
interpret the value — it forwards it verbatim on the paused-session
JSON so the host can map it to the provider-specific cache API.
Recommended values (host-defined; engine is opaque):
"context"— cache everything up to and including the system prompt / long shared context so that repeated calls (panel / moa / ucb) hit the cache"prompt"— cache the full prompt including user turn
Absent when not set; hosts that do not implement prompt caching MUST ignore the field.