ruby_llm-providers-lms 0.1.1 → 0.2.0

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: cdaf33f5245f3b78bcfd33a13c7f2a0838149de0998b09c7f8e4256e447cda69
-  data.tar.gz: 285c1c5ddff4227b90b435755b8cc29b1a28bb695fa33c7f1c6510ce3a5853cb
+  metadata.gz: 3644c34d63679aa27bdbb8efc413a9ddc0a6d60dfcd7a93504b14d10dae910c7
+  data.tar.gz: 13d0cc4ac01da254166c9169fb3dc5b8daf7fffd4d4b869b58848b26dfc33a1b
 SHA512:
-  metadata.gz: 548a88a50a88795f9b5b7cd11c64ee1c43e76406c7730c34c42cf90bf2e748eded77029f38f9994f5ab8820a05d7e0479118e8a43f3dbb165a3bc11e0297a892
-  data.tar.gz: d01419d61951513cc00dcd72e546f7a56e2023fa59ff1bcadc99e0b117f1bdfab6b06c541078b0a984dede41f025e9668fd00921086b3163c5590654d1d99bc2
+  metadata.gz: 4cdd3109e72648dfd1294ea842a5d480b07974780ee2eee791f7a4e40b2a981a914cad85106ac6fba7290836a8bce32018307fd66998f8d413f21538b948a36a
+  data.tar.gz: c43720ead841a2a2bc7e7557a6209c83d2f485f65e5e3197c4794e0e894de5c76a636a3316879b5182dc9cd175025a771b3eb9a43b274c16e25d3fb05c86de70

data/CHANGELOG.md

@@ -0,0 +1,121 @@
+# Changelog
+
+All notable changes to this project are documented in this file.
+
+The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
+and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
+
+## [0.2.0] - 2026-09-14
+
+### Added
+
+- `:native_chat` protocol for LM Studio's native chat API (`POST /api/v1/chat`).
+  The endpoint is stateful — LM Studio stores each response and continues from a
+  `response_id` — so the protocol sends only the messages newer than the last
+  assistant response and points `previous_response_id` at it, which makes ordinary
+  multi-turn `RubyLLM::Chat` work on top of it. The id rides on
+  `Message#raw_content`, which RubyLLM persists, so a chat reloaded from the
+  database continues rather than refusing to replay. Streaming included.
+  Native-only options reach the request through `with_provider_options`:
+  `integrations:` (MCP plugins or ephemeral MCP servers the server runs itself),
+  `context_length:`, `store: false`, and the sampling settings the native schema
+  accepts (`top_p`, `top_k`, `min_p`, `repeat_penalty`). The endpoint reports no
+  stop reason of its own, so a generation that runs into `max_output_tokens` is
+  reported as `finish_reason: :length` from the token counts.
+- `:native_v0` protocol — LM Studio's `POST /api/v0/chat/completions`. Same wire
+  shape as the OpenAI-compatible endpoint, but the response carries the `stats`
+  block that `/v1/chat/completions` leaves empty (tokens per second, time to first
+  token, generation time, stop reason) plus `model_info` and `runtime`, reachable
+  through `Message#raw`.
+- `:anthropic` protocol — LM Studio also serves Anthropic's Messages API, driven
+  by RubyLLM's stock Anthropic protocol with only the request path adjusted.
+- Model management on the provider, backed by LM Studio's native REST API:
+  `native_models`, `loaded_models`, `load_model` (with an optional
+  `context_length:` override), `unload_model`, and `download_model`. All three
+  writes are sent non-idempotent: RubyLLM retries a POST that times out, and
+  LM Studio answers a second load of a resident model by starting a second
+  instance, so a replayed load would hold the weights in memory twice.
+- Reasoning control across every protocol that supports it. LM Studio reports
+  each model's reasoning vocabulary in its own spelling (`off`, `low`, `medium`,
+  `high`, `xhigh`, `on`); the provider translates it into the vocabulary
+  RubyLLM's registry reads — `off` becomes the `none` effort, `on` becomes a
+  `:toggle` option — so `with_thinking`, `with_thinking(false)` and
+  `with_thinking(effort:)` all resolve against what the model actually supports.
+  The native endpoint takes a different spelling from the OpenAI ones, so
+  `:native_chat` translates `none` back to `off` on the way out.
+- `reasoning` and `tool_choice` capabilities on listed models, alongside the
+  `function_calling` and `vision` already derived from the native listing.
+- Model listings read LM Studio's native v1 listing (`/api/v1/models`) when the
+  server has one and fall back to the v0 listing otherwise, adding display name,
+  description, parameter count, size on disk, quantization bit width, downloaded
+  variants and the selected one, reasoning options, and the loaded instance's
+  context length and remaining idle TTL to each model's metadata.
+- Embeddings always go to `POST /v1/embeddings`, whichever chat dialect
+  `lms_protocol` selects — LM Studio serves them nowhere else, and resolving
+  them through the configured chat protocol meant `:anthropic` refused them
+  outright.
+- `RubyLLM::Providers::LMS::VERSION` and `RubyLLM::Providers::LMS.version`.
+- Examples 11-15: native inference stats, the Anthropic protocol, the native chat
+  protocol, model management, and a protocol capability matrix.
+
+### Fixed
+
+- The `:responses` protocol dropped reasoning text. OpenAI never returns raw
+  reasoning, only an optional summary, so RubyLLM's stock protocol reads a
+  reasoning item's `summary` and the matching
+  `response.reasoning_summary_text.delta` event. LM Studio returns the reasoning
+  itself in `content` as `reasoning_text`, streamed as
+  `response.reasoning_text.delta`, leaving `summary` empty — so the response
+  spent reasoning tokens and `message.thinking` came back nil. A
+  `LMS::Responses` subclass now reads both shapes, on the sync and streaming
+  paths alike.
+- The published gem no longer ships `spec/` and its VCR cassettes, which are
+  recorded conversations with the author's own models.
+
+### Changed
+
+- Models are named by their LM Studio display name when the server reports one,
+  falling back to the model id.
+- The `:responses` protocol is now `RubyLLM::Providers::LMS::Responses` rather
+  than RubyLLM's stock `Protocols::Responses`. Behaviour is unchanged apart from
+  the reasoning fix above; code that referenced the stock class by name should
+  follow.
+- The gemspec reads the version from `lib/ruby_llm/providers/lms/version.rb`
+  instead of carrying its own literal.
+- A failing `:live` spec no longer deletes its cassette, which let a real
+  regression pass on the next run by re-recording itself green. Pass
+  `RERECORD=1` to re-record deliberately.
+
+## [0.1.1] - 2026-09-11
+
+### Added
+
+- A friendly `RubyLLM::Error` when the LM Studio server is not running.
+  Connection refusals bypass RubyLLM's error middleware, which only maps HTTP
+  status codes, so the raw `Faraday::ConnectionFailed` used to surface instead.
+- Examples 01-10, covering basic usage, streaming, tools, structured output,
+  embeddings, the model catalog, the `:responses` protocol, the connection guard,
+  provider options, and provider introspection.
+
+### Fixed
+
+- VCR cassettes replay without `LMS_API_BASE` set, so CI can run them.
+
+### Changed
+
+- CI tests only non-EOL Rubies (3.3, 3.4, 4.0).
+
+## [0.1.0] - 2026-09-11
+
+### Added
+
+- Initial release: a `:lms` provider for RubyLLM that talks to LM Studio's local
+  OpenAI-compatible server (`lms server start`, default
+  `http://localhost:1234/v1`) with no API key or configuration required.
+- `:chat_completions` and `:responses` protocols.
+- Model listings enriched from LM Studio's native REST API, plus a packaged
+  `models.json` fallback catalog.
+
+[0.2.0]: https://github.com/madbomber/ruby_llm-providers-lms/compare/v0.1.1...v0.2.0
+[0.1.1]: https://github.com/madbomber/ruby_llm-providers-lms/compare/v0.1.0...v0.1.1
+[0.1.0]: https://github.com/madbomber/ruby_llm-providers-lms/releases/tag/v0.1.0