Crate sim_lib_agent

Expand description

Agent runtime surfaces for SIM: agents, tools, memory, patterns, fixtures, fairness facets, and model-fabric contract re-exports.

Model placement is resolved through the model/at catalog. In-process runners and registry site exports share that catalog: a loaded native or wasm library can register an opaque site value under a placement symbol, and this crate adapts that value into an EvalSite without adding site behavior to the kernel. model/cached wraps placements in a content-addressed cassette, so successful inference replies are replayable by normalized request content, model identity, parameters, result shape, and capabilities.

Re-exports§

pub use atelier::AgentMission;
pub use atelier::AtelierAction;
pub use atelier::GuardCapability;
pub use atelier::GuardDecision;
pub use atelier::GuardEvaluation;
pub use atelier::GuardRefusal;
pub use atelier::RadarChunk;
pub use atelier::RadarError;
pub use atelier::RadarHint;
pub use atelier::RadarIndex;
pub use atelier::RadarQuery;
pub use atelier::RadarReport;
pub use atelier::RadarResult;
pub use atelier::SelfHostingScenario;
pub use atelier::SourceSpan;
pub use atelier::cassette_content_hash;
pub use atelier::evaluate_guarded_action;
pub use atelier::guard_action;
pub use atelier::retrieve_radar_hints;
pub use atelier::self_hosting_scenarios;
pub use atelier::validate_self_hosting_scenarios;

Modules§

atelier: Atelier agent-facing retrieval contracts.

Structs§

Agent: A live agent session: a named manifest of components with its own capabilities, isolation policy, and runtime state.
AgentFabric: A swarm of agents that evaluates together as a single EvalFabric, coordinating its members through an optional topology, planner, and shared blackboard.
AgentLib: The agent runtime library, registering agent functions, tools, and CLI exports as a loadable SIM Lib.
AgentManifest: The set of components composing an agent, grouped by role.
AgentPattern: Descriptor of an agent pattern: an id, description, and named slot groups.
AgentPatternSlot: One named entry within an agent-pattern slot group.
AttributionCard: Card-shaped attribution summary over an event ledger and effect ledger.
AttributionEvidence: One attribution row derived from a run event or effect record.
BlackboardMemory: Memory whose entries are shared across agents on a named blackboard.
Budget: Resource limits bounding an agent run.
Counterfactual: Minimal-change counterfactual found in a deterministic synthetic table.
CounterfactualChange: One feature delta in a counterfactual explanation.
Decomposition: Result of decomposing a goal and running the ordered subtasks.
ExternalRunnerSpec: Metadata needed to project an existing ModelRunner as an agent runner.
FairnessGroup: Binary-outcome group counts for fairness summaries.
FairnessSummary: Fairness summary backed by the F4 disparate-impact metric.
FakeAsrFixture: Scripted output for one synthetic audio reference.
FakeAsrRunner: Local runner that maps synthetic audio references to scripted transcripts.
FakeAsrSegment: One deterministic speech transcript segment.
FakeSensorFrame: One fixed sensor frame in a synthetic sensor stream.
FakeSensorStream: Fixed local stream over synthetic sensor frames.
FakeVisionFixture: Scripted output for one synthetic image reference.
FakeVisionRunner: Local runner that maps synthetic image references to scripted vision output.
FileMemory: File-backed episodic memory that persists entries to a log on disk.
ModelBid: One availability/score proposal produced for market-style runner selection.
ModelCard: Metadata describing one runner/model candidate for routing and inspection.
ModelEvent: Structured streaming event emitted by a model runner.
ModelRequest: Normalized model request used across providers and local runners.
ModelResponse: Final response returned by a runner after inference completes.
ModelUsage: Usage and accounting information attached to a model response.
PersonaMemory: Memory holding evolving persona state, optionally persisted to a path.
PlanningOutput: Output captured after running one planning task.
PlanningTask: Unit of work produced by planning combinators.
Reflection: Critique result, including a bounded retry when the runner asks for one.
SyntheticRecord: One deterministic synthetic record used for counterfactual search.
Tool: A callable tool an agent can invoke, with shape-checked arguments and results, a category, and required capabilities.
ToolSpec: Specification for projecting a runtime value into a local Tool.
VectorMemory: Vector memory supporting similarity search, optionally persisted to a path.
WorkingMemory: In-process working memory holding entries in a shared buffer.

Enums§

AgentRole: Role an agent plays in a multi-agent exchange.
ComponentKind: The role a Component plays within an agent.
MemoryKind: Category of a memory backend.
RunnerBackend: The strategy a runner component uses to produce model responses.

Constants§

AI_RUNNER_CACHE_CAPABILITY: Capability name granting an AI runner use of a response cache.
AI_RUNNER_CAPABILITY: Capability name granting an agent the right to drive an AI model runner.
AI_RUNNER_LOCAL_CAPABILITY: Capability name granting an AI runner access to a local (on-host) model.
AI_RUNNER_NETWORK_CAPABILITY: Capability name granting an AI runner outbound network access.
AI_RUNNER_RAW_LOG_CAPABILITY: Capability name granting an AI runner permission to emit raw request/response logs.
AI_RUNNER_SECRET_CAPABILITY: Capability name granting an AI runner access to secret material (e.g. API keys).

Traits§

Component: A pluggable agent building block (runner, tool, memory, planner, and so on) that participates in evaluation as an EvalSite.
MemoryBackend: Behavior shared by every memory backend: append, query, and snapshot/restore.
ModelEventSink: Sink for streaming model events emitted during inference.
ModelRunner: Executable contract for a model backend or model-like runtime surface.

Functions§

attribution_card: Builds an attribution card from the evidence for one run.
decompose: Asks a runner to split a goal into deterministic ordered subtasks.
decompose_and_run: Decomposes a goal, then executes each returned subtask in order.
external_runner_value: Wraps a provider-neutral ModelRunner in the existing agent runner value.
fairness_summary: Computes a deterministic fairness summary for two binary-outcome groups.
install_agent_lib: Installs the agent library into cx, pulling in the server lib, registering the agent address resolver and line driver, and loading AgentLib.
install_tool: Installs a tool into the registry, reusing any existing registration.
minimal_counterfactual: Finds the nearest synthetic record with the desired outcome.
reflect: Critiques an output and runs at most one retry requested by the runner.
stamp_envelope_role: Stamps the role onto a frame envelope and increments its hop count.
stamp_frame_role: Stamps the role onto a server frame’s envelope.

Type Aliases§

AgentRef: A runtime value referencing an agent.
ComponentRef: A runtime value referencing an agent component.
TopologyRef: A runtime value referencing an agent-network topology.