lex-llm-mlx

LegionIO LLM provider extension for MLX-backed OpenAI-compatible servers on Apple Silicon.

This gem lives under Legion::Extensions::Llm::Mlx and depends on lex-llm >= 0.4.3 for shared provider-neutral routing, response normalization, fleet envelopes, fleet responder execution, and schema primitives.

Load it with require 'legion/extensions/llm/mlx'.

• Legion::Extensions::Llm::Mlx::Provider, exposed to legion-llm as the :mlx provider family.
• OpenAI-compatible chat, streaming, model listing, and embeddings endpoint wrappers.
• Heuristic chat, embedding, and vision capability mapping for discovered local models.
• Local-first defaults for MLX servers running on Apple Silicon hosts.
• Best-effort llm.registry event publishing through shared lex-llm registry helpers when transport is available.
• Provider-owned fleet request actor and runner backed by lex-llm.
• Shared Legion settings, JSON, and logging dependencies with full Legion::Logging::Helper integration.

Legion::Extensions::Llm::Mlx
  Mlx                          # Extension namespace, discovery metadata, default settings
  Provider                     # Health, readiness, model listing, OpenAI-compatible adapter
  Actor::FleetWorker           # Subscription actor enabled by provider instance fleet settings
  Runners::FleetWorker         # Delegates fleet execution to Legion::Extensions::Llm::Fleet::ProviderResponder
  (shared from lex-llm)
    RegistryPublisher          # Async llm.registry event publishing
    RegistryEventBuilder       # Sanitized registry envelope construction

The extension no longer writes provider adapters into the registry at require time. Loaded provider discovery metadata is consumed by legion-llm, which owns adapter creation and registry writes.

Legion::Extensions::Llm::Mlx.default_settings

Defaults target http://localhost:8000, mark the default instance as :local, allow one concurrent local request, and keep fleet participation disabled until a host opts in through extension settings.

The provider accepts the shared lex-llm configuration options:

Legion::Extensions::Llm.configure do |config|
  config.mlx_api_base = 'http://localhost:8000'
  config.mlx_api_key = ENV['MLX_API_KEY']
end

mlx_api_key is optional because most local MLX servers run without authentication. Set it when a proxy or hosted MLX gateway requires bearer authentication.

Provider discovery also reads named instances from extensions.llm.mlx.instances. Generic keys are normalized for the MLX provider:

extensions:
  llm:
    mlx:
      instances:
        local:
          base_url: http://localhost:8000
          api_key: null
          fleet:
            enabled: false
            respond_to_requests: false
            capabilities:
              - chat
              - stream_chat
              - embed

Accepted instance URL keys are base_url, api_base, endpoint, or mlx_api_base. A trailing /v1 is stripped because the shared OpenAI-compatible adapter appends endpoint paths itself.

Provider instances can opt in to consuming Legion LLM fleet requests. The provider-owned fleet actor only starts when at least one configured instance enables respond_to_requests.

extensions:
  llm:
    mlx:
      instances:
        local:
          base_url: http://localhost:8000
          fleet:
            enabled: true
            respond_to_requests: true
            capabilities:
              - chat
              - stream_chat
              - embed

• completion_url and stream_url: /v1/chat/completions
• models_url: /v1/models
• embedding_url: /v1/embeddings
• health_url: /health

The provider uses the shared Legion::Extensions::Llm::Provider::OpenAICompatible adapter so Legion routing can treat MLX, vLLM, OpenAI, and other compatible servers consistently while preserving provider-specific settings and health behavior.

When Legion::Transport and lex-llm routing are available, the provider publishes best-effort events to the llm.registry topic exchange:

• Readiness events — published asynchronously when readiness(live: true) is called.
• Model availability events — published asynchronously after list_models discovers models.

Publishing is fire-and-forget in background threads; failures never block the provider.

• readiness(live: true) calls the MLX /health endpoint and publishes readiness metadata only when the live check succeeds.
• list_models expects an OpenAI-compatible /v1/models response and publishes discovered model availability through the shared registry publisher.
• Fleet request handling is disabled unless at least one discovered instance opts in with fleet.respond_to_requests: true.
• Local instance discovery checks localhost:8080; explicitly configured instances can point at any OpenAI-compatible MLX endpoint.

bundle install
bundle exec rspec --format json --out tmp/rspec_results.json --format progress --out tmp/rspec_progress.txt
bundle exec rubocop -A