Struct RuntimeOptions 

pub struct RuntimeOptions {

    pub provider: Box<dyn LLMProvider>,
    pub execution_plane: Option<Box<dyn ExecutionPlane>>,
    pub session_log: Option<Arc<dyn SessionLog>>,
    pub compression_store: Option<Arc<dyn ArchiveStore>>,
    pub payload_store: Option<Arc<dyn PayloadStore>>,
    pub kernel_reliability: Option<KernelReliability>,
    pub session_id: Option<String>,
    pub max_tokens: u32,
    pub max_turns: Option<u32>,
    pub timeout_ms: Option<u64>,
    pub extensions: Option<Value>,
    pub agent_id: Option<String>,
    pub memory_scope: Option<MemoryScope>,
    pub pre_query_memory: Option<Arc<dyn Fn(&str) -> Vec<MemoryQuery> + Send + Sync>>,
    pub system_prompt: Option<String>,
    pub initial_memory: Vec<String>,
    pub skill_dir: Option<PathBuf>,
    pub memory_store: Option<Box<dyn MemoryStore>>,
    pub knowledge_source: Option<Box<dyn KnowledgeSource>>,
    pub signal_source: Option<Box<dyn SignalSource>>,
    pub governance: Option<Arc<Mutex<Governance>>>,
    pub os_profile: Option<OsProfile>,
    pub governance_policy: Option<GovernancePolicy>,
    pub signal_policy: Option<SignalPolicy>,
    pub scheduler_policy: Option<SchedulerPolicyConfig>,
    pub resource_quota: Option<ResourceQuota>,
    pub memory_policy: Option<MemoryPolicy>,
    pub tokenizer: Option<String>,
    pub enable_plan_tool: Option<bool>,
    pub on_tool_suspend: Option<ToolSuspendHandler>,
    pub on_permission_request: Option<PermissionRequestHandler>,
    pub milestone_policy: MilestonePolicy,
    pub milestone_contract: Option<MilestoneContract>,
    pub run_spec: Option<AgentRunSpec>,
    pub allowed_tool_ids: Option<Vec<String>>,
    pub baseline_tool_ids: Option<Vec<String>>,
    pub on_turn_metrics: Option<OnTurnMetricsHandler>,
    pub stable_core_tool_ids: Vec<String>,
    pub on_milestone_evaluate: Option<MilestoneEvaluationHandler>,
}

Configuration for a RuntimeRunner (aligned with Node/Python RuntimeOptions).

Fields

provider: Box<dyn LLMProvider>

execution_plane: Option<Box<dyn ExecutionPlane>>

session_log: Option<Arc<dyn SessionLog>>

compression_store: Option<Arc<dyn ArchiveStore>>

payload_store: Option<Arc<dyn PayloadStore>>

Storage for canonical opaque external payload locators.

kernel_reliability: Option<KernelReliability>

Bounded recovery and replay policy. Omitted fields retain kernel defaults.

session_id: Option<String>

When set, execute reuses this session id.

max_tokens: u32

max_turns: Option<u32>

timeout_ms: Option<u64>

extensions: Option<Value>

agent_id: Option<String>

memory_scope: Option<MemoryScope>

Required by host-generated memory queries and semantic page-out writes.

pre_query_memory: Option<Arc<dyn Fn(&str) -> Vec<MemoryQuery> + Send + Sync>>

I4: optional run-start memory pre-fetch hook. The runner calls this once per run, before the first LLM turn, with the goal string; each returned query becomes a memory_store.search and the resulting hits page into the knowledge partition before turn 1. Mirrors the Node SDK preQueryMemory. Sync-only in Rust today — async hosts can pre-compute. Errs-open when memory_store or agent_id is missing.

system_prompt: Option<String>

initial_memory: Vec<String>

skill_dir: Option<PathBuf>

memory_store: Option<Box<dyn MemoryStore>>

knowledge_source: Option<Box<dyn KnowledgeSource>>

signal_source: Option<Box<dyn SignalSource>>

governance: Option<Arc<Mutex<Governance>>>

os_profile: Option<OsProfile>

governance_policy: Option<GovernancePolicy>

signal_policy: Option<SignalPolicy>

scheduler_policy: Option<SchedulerPolicyConfig>

resource_quota: Option<ResourceQuota>

memory_policy: Option<MemoryPolicy>

Opt-in long-term memory policy (set_memory_policy), enforced at the kernel memory traps.

tokenizer: Option<String>

enable_plan_tool: Option<bool>

on_tool_suspend: Option<ToolSuspendHandler>

on_permission_request: Option<PermissionRequestHandler>

milestone_policy: MilestonePolicy

How to handle EvaluateMilestone actions. Default: RequireVerifier.

milestone_contract: Option<MilestoneContract>

run_spec: Option<AgentRunSpec>

allowed_tool_ids: Option<Vec<String>>

The run’s exposure ceiling — the outer bound on what this run may EVER advertise. Not a static profile: an INTERSECTION applied every turn (exposed ⊆ ceiling), so baseline_tool_ids, stable_core_tool_ids, and skill allowed_tools all narrow within it and none can widen past it. The kernel meta-tools (skill/memory/knowledge/update_plan/read_result) are exempt on the id axis; the KIND axis still applies. Lowers to the same capability_filter sub-agents use; byte-stable across the run, so it never busts the prompt-cache prefix. Augments run_spec’s filter when both are set; synthesizes a minimal top-level spec otherwise. None/empty means no ceiling. Exposure still starts from the minimal baseline.

baseline_tool_ids: Option<Vec<String>>

The PRE-ACTIVATION exposure surface, selected from under the allowed_tool_ids ceiling (AgentRunSpec::exposure_baseline). Makes narrow→wide progressive disclosure expressible: exposed = meta ∪ ((baseline ∪ stable_core ∪ ⋃ active skills' allowed_tools) ∩ ceiling). None and Some(vec![]) both select the minimal surface (meta-tools + stable-core only). Entries outside the ceiling silently intersect away.

on_turn_metrics: Option<OnTurnMetricsHandler>

P0-C: optional per-turn metrics sink for tool-gating telemetry (see TurnMetrics). Pure observation; invoked once per LLM turn. Panics are not caught — keep the sink trivial.

stable_core_tool_ids: Vec<String>

P1-B/D stable-core: tool ids always exposed under skill gating. Empty ⇒ skills narrow to exactly their declared tools + meta-tools. Opt-in: no skill declaring tools ⇒ never engages.

on_milestone_evaluate: Option<MilestoneEvaluationHandler>