Psyclone 2.2.0: LLM Connections in CMSDK
Modern real-time systems increasingly need to talk to large language models — for reasoning, classification, generation or supervision. Psyclone 2.2.0 makes that a first-class capability with LLM Connections in CMSDK: a native connection primitive for calling language-model endpoints directly from your components.
Declare it, don’t wire it
A new <llm> / <llmvendor> primitive lets you declare a connection in your spec. LLMConnection parses the vendor and model configuration into typed errors and a stats struct, connects over the CMSDK HTTP client, and calls the endpoint via interact() with JSON path extraction to pull exactly the fields you need out of the response.
Request/response and real streaming
Two modes are supported. Mode-A is a straightforward request/response call. Mode-B is true token streaming, backed by a real transport stack: an SSE event parser, an HTTP chunked decoder, a bounded byte-source pump, and an event-stream frame parser that validates the AWS event-stream prelude and per-message CRCs. Streaming carries full statistics accounting, a typed cancel error, and clean reset() parity, plus an owner-end lifecycle callback hook so your component always knows where the connection stands.
Secrets stay out of your specs
API keys never have to live in a spec file. Credentials resolve from environment-variable or file references, and inline secrets are actively rejected. On Windows, SSL client verification now imports the Windows ROOT system store, so certificate verification against public endpoints works correctly on end-user machines — not just on the box where CMSDK was built.
LLM Connections bring language-model integration into the same sub-millisecond fabric that carries the rest of your system’s messages — supervised, observable, and built to production standards.
Psyclone 2.2.0 is available now on the download page. Full detail in the CMSDK documentation.
