Crate synapse 

synapse — LLM router and gateway.

Modules

ai_task_type

Static route-alias → AI task type table.

config

Layered runtime config (env + file paths). Env takes precedence.

embeddings

Embeddings engine: OpenAI-shaped types, the provider trait, and batch splitting.

error

Gateway error type → OpenAI-shaped JSON + HTTP status.

gateway

In-process LLM gateway: routing, fallback, native Vertex, ledger, metrics. Transport-independent core; the axum HTTP layer (server) delegates here.

guard

Configurable input guardrails backed by llm-guard. Per-route policy selection with fallback to a default policy.

jev_hybrid

Jev hybrid extraction: judge candidates with TypeSafe System One, then run a schema-pinned chat extraction per survivor on the route’s chat legs. Spec: docs/superpowers/specs/2026-09-22-jev-hybrid-extraction-design.md.

jev_native

Native TypeSafe System One (Jev) lane. Jev evaluates typed questions (choice / score / noul) against a state and returns structured decisions — it has no chat surface, so the gateway exposes it as a verbatim passthrough (POST /typesafe/v1/systemone, see server::jev_passthrough).

ledger

Pluggable cost ledger. The hot path enqueues onto a bounded channel drained by a background writer; on a full channel we drop + count, never block.

observability

OpenLLMetry gen_ai.* span attributes + Prometheus metric emission. Both lanes emit the same shape so native-Vertex calls are not blind.

pricing

Static pricing table + cost calculation.

providers

Provider catalog: genai clients + circuit breakers, keyed by provider id.

resilience

Retries + circuit breakers for synapse outbound HTTP clients.

routing

Routing: request model, lane classification, route table, executor.

server

axum surface: a thin HTTP layer that delegates to the in-process Gateway.

vertex_endpoint

Vertex AI REST host resolution for global, multi-region, and single-region locations.

vertex_native

Native Vertex REST lane: preserves cachedContents, gs:// media URIs, and strict responseSchema that the OpenAI-compatible standard lane cannot express.