Mattbusel/tokio-prompt-orchestrator

Provides a five-stage bounded-backpressure DAG with deduplication, circuit breakers, rate limiting, prompt injection detection, and provider arbitrage across Anthropic, OpenAI, llama.cpp, and vLLM backends. Features a plugin system for request/response filtering, session management with context windowing, streaming token aggregation, and composable prompt transformation pipelines. Exposes REST, WebSocket, SSE, MCP, Prometheus metrics, and OpenTelemetry tracing for production LLM inference workloads.