Struct Agent 

pub struct Agent {
    pub config: DecisionConfig,
    pub persist: bool,
    /* private fields */
}

The orchestrator.

The trait objects carry Send + Sync because scema-daemon shares one Arc<Agent> across connection threads. Constructing an agent per request would reload the Deep Q* checkpoint every time, and a Mutex would serialise every observation behind whichever request is currently walking a large tree. Every implementation in this workspace is already thread-safe — they hold plain data and take no locks.

Fields

config: DecisionConfig

persist: bool

When false, cycle computes and returns everything but writes nothing. The default for scema simulate, which is explicitly a counterfactual and must not leave a trace that reads like a decision the agent made.

Implementations

impl Agent

pub fn new(root: impl Into<PathBuf>, dqstar_checkpoint: Option<String>) -> Self

An agent rooted at a state directory, with the default observers and evaluators.

dqstar_checkpoint is the sniper’s scematica-nn-agent.json when there is one. A missing file is not an error — the evaluator reports it through its applicability, which is where an operator will actually see it.

pub fn memory(&self) -> &MemoryStore

pub fn records(&self) -> &RecordStore

pub fn observers(&self) -> &[Box<dyn Observer + Send + Sync>]

pub fn evaluators(&self) -> &[Box<dyn Evaluator + Send + Sync>]

pub fn observe(&self, locator: &str) -> Result<WorldState>

Perceive an environment.

pub fn hypothesize(&self, world: &WorldState, goal: &Goal) -> Vec<Hypothesis>

Propose branches from every hypothesiser, in a stable order.

Duplicate ids are dropped, first proposer wins. Two hypothesisers arriving at the same branch is a real occurrence — a memory of a procedure that a signal rule would also propose — and ranking the same branch twice would inflate its apparent support.

pub fn cycle(&self, locator: &str, goal: Goal) -> Result<Cycle>

Run one full pass.

pub fn cycle_over(&self, world: WorldState, goal: Goal) -> Result<Cycle>

Run a pass over a world that was already observed.

Split out so the CLI can observe once and simulate several goals against the same world, and so a test can drive the loop over a constructed world without a filesystem.