Module inference_agent 

Re-exports

pub use crate::modalities::logic::n3_compiler::AgentIntentFrame;

pub use crate::modalities::logic::n3_compiler::N3OutputMode;

pub use crate::modalities::logic::n3_compiler::MAX_CONTEXT_NAMESPACE_SLOTS;

pub use crate::modalities::logic::n3_compiler::MAX_INTENT_SCOPE_SLOTS;

Structs

AgentIntent

Structured intent message from LLM → Webizen. Every call must declare what it intends to do — the Webizen validates this against the Rights Ontology BEFORE the LLM ever sees the user’s semantic graph.

AgentOutput

The raw output from the LLM, before Webizen validation.

LocalLlmAgent

The concrete local inference agent. Uses a mock inference path for now; swap infer_local_model for an actual llama.cpp FFI call.

Enums

AgentBackend

Describes where inference actually runs.

AgentError

WebizenVerdict

The Webizen VM’s ruling on an AgentIntent or AgentOutput.

Constants

INFERENCE_TIMEOUT_MS

Default maximum milliseconds for a local inference call (interactive). Batch/overnight profile raises this via llm_bench::inference_timeout_ms().

LLM_MEMORY_BUDGET_BYTES

Hard memory ceiling for the LLM runtime within the 512MB system floor. Leaves the remaining 384MB for the Webizen VM, SLG Arena, and WASM stack.

LLM_RULE_INTENT_FRAME_MISMATCH

LLM_RULE_NO_ADVERSARIAL_CONDUCT

LLM_RULE_NO_OUTBOUND_TELEMETRY

LLM_RULE_NO_SANCTUARY_ACCESS

LLM_RULE_PROFILE_VIOLATION

LLM_RULE_PROVENANCE_REQUIRED

LLM_RULE_REMOTE_CONSENT

LLM_RULE_TOKEN_BUDGET

MAX_OUTPUT_TOKENS

Maximum tokens the agent may generate in a single turn. Enforces deterministic compute cost — no runaway generation that blocks the edge device.

SANCTUARY_SCOPE_WEBIZEN

Special webizen hash marking a Sanctuary-flagged graph scope.

Traits

AgentRuntime

The core abstraction. All LLM backends MUST implement this. The trait is object-safe so it can be boxed and swapped at runtime.