Struct TurnRequest 

pub struct TurnRequest<'a> {
    pub storage: Arc<dyn StorageAdapter>,
    pub llm: LlmConfig,
    pub max_iterations: u32,
    pub conversation_id: &'a str,
    pub request_id: &'a str,
    pub user_message: &'a str,
    pub access: AccessContext,
    pub llm_provider: Option<Arc<dyn LlmProvider>>,
    pub reranker: Option<Arc<dyn Reranker>>,
}

Everything one streaming turn needs. Bundled into a struct so the call sites (the reference server’s handle_send_message and the lambda’s send_message) stay readable and the security-critical access field can never be silently dropped from a positional argument list.

Fields

storage: Arc<dyn StorageAdapter>

The storage seam (conversations / messages / sessions / knowledge).

llm: LlmConfig

The resolved LLM config for this turn.

max_iterations: u32

Agent-loop iteration cap.

conversation_id: &'a str

The conversation this turn belongs to.

request_id: &'a str

The protocol request id (streaming correlation).

user_message: &'a str

The inbound user message.

access: AccessContext

The requester’s document-level entitlements. Retrieval (the auto-injected [Relevant knowledge] context AND the knowledge_search tool) reads through storage.knowledge_for_access(&access), so a restricted document is never surfaced to a requester who lacks the entitlement. An AccessContext::anonymous sees only org-public docs (fail closed for ACL’d content).

llm_provider: Option<Arc<dyn LlmProvider>>

Optional test-injected LLM surface (a MockLlmClient) so the turn runs deterministically offline. None in production (a live client is built from llm).

reranker: Option<Arc<dyn Reranker>>

Optional post-retrieval reranker (feature gap G8). When Some, the knowledge_search tool overfetches candidates and reorders the top-K with this reranker before they reach the model. None (the default) keeps the retrieval order unchanged, so default behavior is byte-for-byte the same. Selected by build_reranker.