Crate somatize_core 

Core types and traits for the Soma computational graph runtime.

A graph holds two kinds of node, and this crate defines both sides of that split. A Filter computes — deterministic, content-cacheable, fit/forward. A Step decides — it polls, asks the runtime to perform Effects (a model, a tool, another graph), and returns a Transition. One NodeMeta describes either kind; one NodeOutcome says how either kind finished. Everything downstream reads the metadata, never the kind.

The contracts every other crate depends on:

• Filter — the computation unit (fit/forward)
• Step / Effect — the decision unit and what it asks for; effects are journaled, filters are cached
• Value — typed data flowing between nodes (Tensor, JSON, Bytes)
• Graph — DAG of nodes and edges, computational and effectful alike
• CacheKey / CacheStore — content-addressable caching
• DataStore — abstraction for moving data between workers
• Schema — dtype + shape for compile-time validation, on every edge
• Event — runtime lifecycle events

Re-exports

pub use cache::CacheKey;

pub use cache::CacheStore;

pub use cache::CacheTier;

pub use cache::EntryMeta;

pub use cache::Origin;

pub use control::LoopCondition;

pub use control::LoopSignal;

pub use control::read_arm_selector;

pub use control::read_loop_signal;

pub use effect::Effect;

pub use effect::EffectResult;

pub use effect::GraphEffectMode;

pub use effect::JoinPolicy;

pub use effect::LlmRequest;

pub use effect::LlmResponse;

pub use effect::NodeSpec;

pub use effect::StopReason;

pub use effect::SuspendReason;

pub use effect::ToolSpec;

pub use effect::Usage;

pub use error::Result;

pub use error::SomaError;

pub use event::Event;

pub use event::MetricRecord;

pub use event::PlanSummary;

pub use event::RunId;

pub use event::StudyId;

pub use event::TrialId;

pub use filter::Distribution;

pub use filter::Filter;

pub use filter::FilterKind;

pub use filter::FilterMeta;

pub use filter::RemoteTarget;

pub use filter::StreamMode;

pub use fingerprint::ArchitectureFingerprint;

pub use fingerprint::EdgeRef;

pub use fingerprint::pipeline_summary;

pub use fingerprint::structural_similarity;

pub use graph::Edge;

pub use graph::EdgeKind;

pub use graph::Graph;

pub use graph::Node;

pub use graph::NodeId;

pub use message::ContentBlock;

pub use message::Message;

pub use message::Messages;

pub use message::Role;

pub use schema::DataType;

pub use schema::Dimension;

pub use schema::Schema;

pub use search::Scale;

pub use search::SearchDimension;

pub use search::SearchSpace;

pub use search::Searchable;

pub use state::MemoryStateStore;

pub use state::StateStore;

pub use step::Step;

pub use step::StepCtx;

pub use step::StepMeta;

pub use step::Transition;

pub use store::DataRef;

pub use store::DataStore;

pub use store::LocalDataStore;

pub use store::StorageConfig;

pub use store::StoreMeta;

pub use store::StreamFormat;

pub use store::slice_tensor_rows;

pub use strategy::ClientSelection;

pub use strategy::CommunicationProtocol;

pub use strategy::ExploitStrategy;

pub use strategy::ExploreStrategy;

pub use strategy::FederatedAggregation;

pub use strategy::GradientAggregation;

pub use strategy::Partition;

pub use strategy::TrainingStrategy;

pub use study::CompositeObjective;

pub use study::Direction;

pub use study::Objective;

pub use study::PruningStrategy;

pub use study::Scalarizer;

pub use study::SearchStrategy;

pub use study::Study;

pub use study::Trial;

pub use study::TrialState;

pub use summary::FlagCount;

pub use summary::NodeCost;

pub use summary::RunConclusion;

pub use summary::RunOutcome;

pub use summary::RunSummary;

pub use summary::TrialSummary;

pub use summary::human_duration;

pub use tracking::EventEnvelope;

pub use tracking::EventSink;

pub use tracking::GitInfo;

pub use tracking::GraphSummaryInfo;

pub use tracking::RUN_SCHEMA_VERSION;

pub use tracking::RunKind;

pub use tracking::RunManifest;

pub use tracking::RunState;

pub use tracking::RunStatus;

pub use tracking::Tracker;

pub use value::Value;

pub use virtual_value::ValueStatus;

pub use virtual_value::VirtualValue;

pub use viz::GraphOverlay;

pub use viz::NodeOverlay;

pub use viz::NodeStatus;

Modules

action

Two-table cache model: action records + content-addressed blobs.

any

Downcasting, without every implementor writing the same three lines.

cache

Content-addressable caching — keys, traits, and metadata.

canon

Canonical, deterministic serialization for cache-key hashing.

codec

SOMA1 framed binary codec for Value.

control

Data-dependent control flow: how a loop decides to stop and how a branch picks an arm.

effect

Effects: the things a step can ask the runtime to do for it.

error

Error types for the Soma runtime.

event

Runtime lifecycle events — emitted during plan execution.

filter

The Filter trait — the fundamental computation unit in Soma.

fingerprint

Architecture fingerprints — a stable identity for a graph’s shape.

graph

Computational graph — DAG of filter nodes connected by edges.

keys

Reserved keys in a run’s output store.

message

The canonical conversation shape.

node

What the compiler and executor need to know about a node, whichever kind it is.

schema

Schema — dtype and shape for compile-time type checking between filters.

search

Search spaces for hyperparameter optimization.

state

Trained-state storage — authoritative data produced by fit().

step

The effectful counterpart to crate::filter::Filter.

store

Data Store: abstraction for moving data between workers.

strategy

Training strategies for distributed execution.

study

Study — defines an optimization experiment with objectives and strategy.

summary

What a run says about itself — the data half.

svg

Self-contained SVG rendering of a Graph — no JavaScript, no external tools.

tool

How a tool describes itself.

tracking

Experiment tracking types: run manifests, status, event envelopes, and the EventSink/Tracker traits.

util

Shared utility functions.

value

Typed values flowing between filters in a pipeline.

virtual_value

Virtual values — lazy references to data that can be materialized on demand.

viz

Rendering overlays for graph visualization.

Derive Macros

SomaFilter

Derive macro for generating config_hash() and Searchable implementations.

SomaStep

Derive config_hash for a Step.