Struct AgentHarnessOptions 

pub struct AgentHarnessOptions {

    pub system_prompt: String,
    pub model: Option<Model>,
    pub thinking_level: ThinkingLevel,
    pub skills: Vec<Skill>,
    pub prompt_templates: Vec<PromptTemplate>,
    pub tools: Vec<Arc<dyn AgentTool>>,
    pub session: Session,
    pub observer: Arc<dyn RuntimeObserver>,
    pub observation_context: ObservationContext,
    pub observation_parent: Option<OperationId>,
    pub stream_fn: Option<StreamFn>,
    pub get_api_key: Option<GetApiKey>,
    pub compaction: CompactionSettings,
    pub compact_algorithms: Arc<CompactAlgorithmRegistry>,
    pub before_tool_call: Option<BeforeToolCallHook>,
    pub after_tool_call: Option<AfterToolCallHook>,
    pub on_control_plane_prompt: Option<OnControlPlanePromptHook>,
    pub budget_cap_usd: Option<f64>,
    pub reload_skills_fn: Option<ReloadSkillsFn>,
    pub on_turn_end: Option<OnTurnEndHook>,
    pub turn_continuation_cap: Option<u32>,
    pub max_iterations: Option<u32>,
    pub runtime_extensions: Arc<dyn RuntimeExtensionPort>,
    pub runtime_extension_model_context: ExtensionModelContextProjection,
    pub runtime_extension_cwd: String,
    pub runtime_extension_has_interactive_client: bool,
}

Fields

system_prompt: String

Base system prompt prepended to the rendered skill catalog.

model: Option<Model>

Active model. None when no model is configured yet (daemon launches model-less and the client injects one per-session via set_model); the turn loop errors with “Agent has no model set” until one is assigned.

thinking_level: ThinkingLevel

skills: Vec<Skill>

prompt_templates: Vec<PromptTemplate>

tools: Vec<Arc<dyn AgentTool>>

session: Session

observer: Arc<dyn RuntimeObserver>

Embedder-owned runtime observer shared with the inner bare Agent.

observation_context: ObservationContext

Session/run/job/node correlation inherited by inner operations.

observation_parent: Option<OperationId>

Optional parent operation for the inner agent run.

stream_fn: Option<StreamFn>

get_api_key: Option<GetApiKey>

Optional per-provider credential resolver applied to every model call.

compaction: CompactionSettings

Auto-compaction thresholds. Defaults to DEFAULT_COMPACTION_SETTINGS.

compact_algorithms: Arc<CompactAlgorithmRegistry>

Custom compaction algorithm registry. The builtin algorithm is always available.

before_tool_call: Option<BeforeToolCallHook>

Optional tool hooks supplied by the embedding runtime.

after_tool_call: Option<AfterToolCallHook>

on_control_plane_prompt: Option<OnControlPlanePromptHook>

Control-plane prompt resolver. None keeps prompt-classified tool calls fail-closed.

budget_cap_usd: Option<f64>

Per-session USD cap. None disables the check.

reload_skills_fn: Option<ReloadSkillsFn>

Embedder-owned loader used by AgentHarness::reload_skills_from_disk.

on_turn_end: Option<OnTurnEndHook>

Optional cross-prompt turn completion hook.

turn_continuation_cap: Option<u32>

Maximum hook-driven continuations for one prompt cycle.

max_iterations: Option<u32>

Hard cap on inner agent loop iterations. None is unbounded.

runtime_extensions: Arc<dyn RuntimeExtensionPort>

Engine-independent lifecycle port supplied by the embedding runtime.

runtime_extension_model_context: ExtensionModelContextProjection

Reconstructed, de-duplicated persistent model context for this branch.

runtime_extension_cwd: String

Working directory included in runtime-extension lifecycle context.

runtime_extension_has_interactive_client: bool

Whether this harness is currently attached to an interactive client.

Implementations

impl AgentHarnessOptions

pub fn new(model: impl Into<Option<Model>>, session: Session) -> Self

model may be a concrete Model (wrapped into Some) or an Option<Model> (kept as-is) so both the model-less daemon path and the existing concrete-model callers build through one constructor.