Architecture

The gem is a single RubyLLM provider class, RubyLLM::Providers::LMS, with one protocol class per LM Studio dialect. RubyLLM does the rest: chats, streaming, tools, embeddings and judgments all run through RubyLLM's own machinery, and the provider decides which LM Studio endpoint each one reaches.

Routing

RubyLLM asks the provider to resolve a protocol for every operation. The provider answers by operation, not just by configuration:

Operation	Goes to	Chosen by
chat (ask, streaming, tools)	the chat protocol	lms_protocol, default :chat_completions
embed	:chat_completions → POST /v1/embeddings	always — LM Studio serves embeddings nowhere else
judge	:decisions or :system_one	lms_judgment_protocol, default :decisions
list models	:chat_completions, enriched from the native listing	always — RubyLLM lists through the provider's default protocol, whatever lms_protocol says
load / unload / download	the provider's own ModelManagement methods	—

The judgment route is the provider's private resolve_protocol override: when RubyLLM resolves with operation: :judge, it substitutes lms_judgment_protocol. Without it, lms_protocol = :anthropic would send judgments to a protocol that has none.

Source layout

File	Responsibility
lib/ruby_llm/providers/lms.rb	The provider: protocol registry, API base and auth headers, embedding and judgment routing, configuration options; registers :lms with RubyLLM at load time
lms/connection_guard.rb	Wraps RubyLLM's connection so a refused connection raises a RubyLLM::Error telling you to run lms server start
lms/models.rb	Model listing: OpenAI /v1/models enriched with native details; capabilities, modalities and metadata
lms/native_v1.rb	Rewrites the native v1 listing into the shape models.rb reads, including reasoning options and decision fields
lms/endpoints.rb	Native listing paths, spelled relative to the /v1 base
lms/responses.rb	:responses — reads LM Studio's reasoning_text as well as OpenAI's summaries
lms/anthropic_messages.rb	:anthropic — RubyLLM's Anthropic protocol with LM Studio's path
lms/native_chat.rb, native_chat/*	:native_chat — stateful /api/v1/chat, with conversation continuation and streaming
lms/decisions.rb	:decisions — OpenAI Decisions, answers matched to questions by name
lms/system_one.rb	:system_one — System One with the /v1-relative path and image support
lms/model_management.rb	native_models, loaded_models, load_model, unload_model, download_model
lms/version.rb	RubyLLM::Providers::LMS::VERSION

:chat_completions and :native_v0 are small classes defined in lms.rb itself; :native_v0 is the Chat Completions protocol pointed at /api/v0/chat/completions.

Design rules

• Everything lives in RubyLLM::Providers::LMS. The gem adds no top-level constants. Archspec.rb enforces this, and that nothing under lib/ references RSpec, WebMock or VCR.
• Subclass RubyLLM, don't fork it. Each protocol is RubyLLM's stock protocol with the smallest override LM Studio needs — a path, a response field, a matching rule.
• Local means no ceremony. configuration_requirements is empty, so RubyLLM asks for no API key; local? is true, so it assumes any model id exists; the auth header is sent only when lms_api_key is set.
• Paths are relative to the /v1 base. Native endpoints are reached as ../api/v1/... so a non-default lms_api_base moves everything together.