Skip to content
CAI
Software that uses CAICheck a score

crmne/ruby_llm

60.8

Adequate · 14 November 2026

60k

lines of production code

Ruby

primary language

5

measurements over time

CAI band scale
CAI trend line
CAI lens gauges

What this system is

Features

Add .env.example with provider API key templates

The repository now includes a .env.example file that lists environment variables for all supported providers (such as Anthropic, AWS, Azure, Deepgram, OpenAI, and others) using 1Password CLI syntax as placeholders. This provides a ready-to-use template for developers to configure their local environment with the necessary API keys and endpoints without exposing real credentials.

(repo-wide) · high confidence

Add CLI entry point for provider generation

A new executable script at exe/ruby\_llm has been added to serve as the command-line interface for the library. This entry point initializes the load path and delegates execution to the RubyLLM::Generators::Provider::CLI, enabling users to run provider-related generator commands directly from the command line.

exe · high confidence

Add ElevenLabs image and video generation support

The ElevenLabs protocol now supports image and video generation flows. New modules handle image creation (including inpainting with masks and aspect ratio validation) and video generation (supporting standard text-to-video with frame inputs and the Creatify Aurora model for image-to-video animation). The implementation includes polling logic to wait for generation completion, validation of media attachments, and parsing of job statuses and results.

_lib/ruby\llm/protocols/elevenlabs/flows · high confidence

Add Hetzner Inference provider with inline image support

Introduces the Hetzner Inference provider, enabling users to send chat completions to Hetzner's API. The implementation supports image attachments by converting remote images into inline base64 data, as the Hetzner servers cannot fetch external URLs directly. Other media types like documents and audio are not supported. Also in this area: 7 commits (1 fix) modifying lib/ruby\_llm/providers.

_lib/ruby\llm/providers/hetzner · medium confidence · unverified

Add Mistral AI provider support

Users can now interact with Mistral AI models through the RubyLLM library. This addition includes full support for chat completions (with specific handling for prompt caching, tool choices, and reasoning content blocks), embeddings, document OCR, text-to-speech, and audio transcription with diarization. The provider also exposes model metadata and capabilities, allowing the library to correctly identify features like function calling and structured output for specific Mistral models.

_lib/ruby\llm/providers/mistral · high confidence

Add Mistral Conversations protocol implementation

Introduces a new protocol implementation for the Mistral 'Conversations' API endpoint, enabling users to interact with Mistral models using this specific interface. The change adds four new files: \chat.rb\ handles text completion, tool usage, and message rendering; \images.rb\ supports image generation via the same endpoint; \streaming.rb\ implements server-sent event streaming for real-time responses; and \request\_shapes.rb\ defines how conversation turns and tool steps are structured for the API. This allows the library to route Mistral requests through the Conversations API rather than the standard Chat Completions API.

_lib/ruby\llm/protocols/mistral/conversations · high confidence

Add Perplexity provider support with chat, embeddings, and document attachments

Introduces a new Perplexity provider integration, enabling users to interact with Perplexity's Sonar models for chat completions and generate embeddings. The implementation supports dynamic model listing with a static fallback for search and embedding models, handles media attachments including images and supported document types (PDF, DOC, DOCX, TXT, RTF), and correctly parses Perplexity's base64-encoded int8 embedding vectors.

_lib/ruby\llm/providers/perplexity · high confidence

Add Rake task to persist model registry to the database

A new \ruby\_llm:load\_models\ task has been added to allow users to load the selected model registry from JSON and persist it into the database. This task ensures the ActiveRecord load hook is triggered correctly, even in environments where ActiveRecord is loaded lazily (such as Rails 8.1), and provides feedback on the number of models successfully saved.

lib/tasks · high confidence

Add support for xAI batch processing

Users can now submit asynchronous batch requests to xAI through the RubyLLM library. This change introduces a new \Batches\ module within the xAI provider that handles creating batch jobs, polling for their status, canceling them, and retrieving results. The implementation supports both the newer Responses API and the legacy Chat Completions API, automatically detecting the request type and mapping xAI's specific response structures to RubyLLM's standard batch result format.

_lib/ruby\_llm/providers/xai/chat\completions · high confidence

A new symbolic link named 'contributing' has been added to the .agents/skills directory, pointing to the existing contributing skill definition located in the .claude/skills directory. This change exposes the contributing guidelines as an available skill within the agents configuration.

.agents · high confidence

Background job for streaming chat responses

The chat UI generator now includes a background job template that processes chat variables by invoking the agent's complete method with a broadcast callback, enabling real-time streaming of replies to the user interface while handling queued messages.

_lib/generators/ruby\_llm/chat\ui/templates/jobs · high confidence

Bedrock provider adds Mantle protocol support and precise capability detection

The Bedrock provider now supports the Bedrock Mantle protocol, allowing access to specific models (such as Google Gemma and OpenAI GPT-OSS variants) via the \bedrock-mantle\ endpoint with proper SigV4 signing. Additionally, a new \Capabilities\ module ensures that tool choice is correctly reported for specific models like \amazon.nova-2-lite-v1:0\, fixing previous inaccuracies where structured output or tool support was guessed incorrectly based on model IDs alone.

_lib/ruby\llm/providers/bedrock · high confidence

Class-based evaluation runner with inflector-aware discovery

The Ruby LLM task now supports running evaluations defined as classes. A new \RubyLLM::Tasks::Evaluations\ module discovers evaluation files in the \evals\ directory, using the application's inflector (Rails or Zeitwerk) to map file paths to class names. It loads each class, verifies it inherits from \RubyLLM::Evaluation\, runs it with configurable repetitions, and saves JSON reports to \tmp/evaluations\. Users can target a specific evaluation by name or run all discovered evaluations.

_lib/tasks/ruby\llm · high confidence

DeepSeek protocol adds file upload support with download restrictions

The DeepSeek protocol now supports file uploads by inheriting from the OpenAI files implementation, allowing users to attach files for vision models. However, downloaded files are explicitly unsupported; attempting to download an uploaded file raises an error, and uploaded files are marked as non-downloadable in the API response.

_lib/ruby\llm/protocols/deepseek · high confidence

GPUStack protocol adds MCP server tool aliases, tokenization, and video generation support

The GPUStack protocol implementation now includes three new capabilities: MCP server tool aliases that map internal tool names (like web\_search and code\_execution) to specific vLLM server labels with strict validation for approval settings; a tokenization module that proxies text tokenization requests to the GPUStack backend; and a video generation module that supports creating video jobs from prompts and attachments, handling job status polling and content download. These additions extend the GPUStack provider's functionality beyond basic chat completions to include tool use configuration, token counting, and media generation.

_lib/ruby\llm/protocols/gpustack · high confidence

Gemini protocol refactored into modular components with new batch, caching, and file APIs

The Gemini protocol implementation has been restructured into distinct modules (batches, caches, chat, embeddings, files, images, speech, streaming, etc.) to improve maintainability and support new capabilities. This change introduces asynchronous batch processing for both content generation and embeddings, allowing users to submit multiple requests at once and poll for results. It also adds explicit content caching via the Gemini cachedContents API to reduce latency and token costs for repeated prompts. Additionally, the protocol now supports the Gemini Files API for uploading and managing large files, enabling better handling of large attachments in chat. Other updates include support for live transcription via WebSocket, speech generation, and improved handling of media attachments and thinking models.

_lib/ruby\llm/protocols/gemini · high confidence

Generated chat UI now supports tool approvals, file attachments, and queued messages

The generated chat interface templates have been expanded to handle more complex AI interactions. Users can now approve or deny tool calls (such as code execution) directly within the chat view, with pending approvals hiding the associated tool output until resolved. The UI also displays attached files associated with the chat session and shows queued messages that are waiting to be processed. These changes are implemented via new partials for approvals, files, and queued messages, integrated into the main chat show view.

_lib/generators/ruby\_llm/chat\ui/templates/views/chats · high confidence

Initial GPUStack provider support with multimodal and speech capabilities

Users can now connect RubyLLM to a GPUStack backend to leverage its vLLM-based models. This integration supports chat completions with multimodal attachments (images, audio, video, and text files), embeddings that accept media inputs, and model discovery that automatically detects capabilities like vision, function calling, and reasoning from the GPUStack 2.x API. Additionally, the provider includes specialized handling for speech synthesis (requiring explicit streaming) and transcription (parsing Chat Completions deltas), enabling end-to-end audio workflows.

_lib/ruby\llm/providers/gpustack · high confidence

Initial Ollama provider implementation with media and capability support

This change introduces the first-class Ollama provider integration, enabling users to connect to local or cloud-hosted Ollama instances. The implementation adds support for multimodal inputs by formatting images, audio, and text file attachments via a new media module, while also allowing the system to dynamically discover model capabilities (such as vision, function calling, and reasoning) and modalities through the Ollama API. Chat interactions are handled with specific payload rendering that respects Ollama's native parameters, including a note that thinking budgets are currently ignored by the backend.

_lib/ruby\llm/providers/ollama · high confidence

Introduce Agent, Batch, and Caching abstractions alongside model registry updates

This release adds a new Agent class that lets you define reusable chat configurations (model, instructions, tools, MCP servers) as classes and instantiate them to build chats. It also introduces a Batch class for submitting provider-side batch requests of chats or embeddings, and a CachedContent class for managing provider-side prompt caches. The model registry is updated with new aliases for Claude, Gemini, and DeepSeek models, and the Attachment class now supports a resolution option to control media detail levels. Also in this area: Workflow tracking, tool namespacing, and long-running message continuation. Also in this area: 7 commits (0 fixes) modifying lib/ruby\_llm, ruby\_llm; Sandboxed code execution with isolated package environments; Workflow tracking, tool namespacing, and long-running message continuation; Include ignored mutations in the mutation testing report; RubyLLM 2.0 introduces scoped contexts, provider scaffolding, and benchmarking infrastructure; and 1 more.

(repo-wide) · medium confidence

Introduce Bedrock Converse protocol with batch, media, and reasoning support

This change adds a new Bedrock Converse protocol implementation to RubyLLM, enabling users to interact with AWS Bedrock models via the Converse API. The update introduces support for asynchronous batch inference jobs (batches.rb), allowing bulk model invocations. It expands media handling (media.rb) to accept audio, video, and document attachments alongside images and text, including support for provider-file references. The protocol also implements detailed reasoning and thinking capabilities (reasoning.rb, thinking\_stream.rb), mapping model-specific effort levels and token budgets for Claude and Nova models, while preserving streaming reasoning content. Additionally, it provides request shape parsing (request\_shapes.rb) to accurately track usage, caching, and tool interactions within the Converse payload structure. Also in this area: 25 commits (0 fixes) modifying lib/ruby\_llm; Workflow tracking, tool namespacing, and long-running message continuation.

_lib/ruby\llm/protocols/converse · medium confidence · unverified

Introduce class-based evaluation framework with dataset and reporting support

The evaluation system now supports class-based evaluations, allowing users to define evaluation logic in dedicated classes rather than inline blocks. This change introduces a structured evaluation pipeline including dataset loading from YAML, JSON, and JSONL files, case-based test execution, and detailed reporting with cost and token tracking. Users can now integrate evaluations with Minitest or RSpec test suites, run evaluations against datasets, and receive comprehensive reports showing pass/fail status, costs, and usage metrics for each trial.

_lib/ruby\llm/evaluation · high confidence

New Anthropic protocol implementation with batch, file, and streaming support

The Anthropic integration has been replaced with a new protocol structure that adds support for asynchronous Message Batches (create, find, cancel, and retrieve results), the beta Files API for large attachments, and enhanced streaming that tracks thinking blocks, citations, and server tool calls for verbatim replay. The chat implementation now supports prompt caching, context compaction, structured thinking, and tool search, while model listings are paginated and token usage is aggregated across compaction iterations.

_lib/ruby\llm/protocols/anthropic · high confidence

New Chat Completions protocol and provider-specific integrations

The library introduces a new \RubyLLM::Protocols::ChatCompletions\ module that centralizes the implementation of the OpenAI-compatible Chat Completions API. This change adds dedicated protocol files for batch processing, embeddings (including sparse vector support), image generation and editing, media attachment formatting, model listing, moderation, request shaping, reranking, speech synthesis, streaming, tool usage, and audio transcription. It also adds specific provider integrations for OpenRouter, including file management, the Responses API with hosted tools, and transcription. For users, this represents a structural shift in how API interactions are handled, consolidating these capabilities into a unified protocol layer while maintaining support for provider-specific features like OpenRouter's server tools and file APIs.

_lib/ruby\_llm/protocols/chat\_completions, lib/ruby\llm/protocols/openrouter · high confidence

New Perplexity Agent, Router, and File handling protocols

This change introduces three new protocol implementations for Perplexity: the Agent API, the Router API, and a dedicated Files handler. The Agent protocol supports preset-based model selection (including mappings for retired Sonar models), handles web search citations, and restricts attachment support to provider files, images, and text. The Router protocol enables structured output via JSON schemas with strict mode. The Files protocol allows downloading generated response files from the Agent API but explicitly disables file uploads. Also in this area: 1 commit (0 fixes) modifying lib/ruby\_llm/protocols/perplexity; 7 commits (1 fix) modifying lib/ruby\_llm/providers.

_lib/ruby\llm/protocols/perplexity · medium confidence · unverified

New Rails generator for creating LLM tool classes and view partials

The RubyLLM Rails generator now supports creating custom tool classes along with their corresponding view partials for displaying tool calls and results in the chat UI. The generator produces a base tool class in app/tools that inherits from RubyLLM::Tool, and generates Tailwind or scaffold-styled view partials for rendering tool execution details and outcomes within the message conversation flow.

_lib/generators/ruby\llm/tool/templates · high confidence

New Rails generator for creating Schematist-based schema classes

Users can now generate RubyLLM schema classes using the new \ruby\_llm:schema\ Rails generator. This tool creates a new file in \app/schemas/\ that inherits from \Schematist::Schema\, automatically handling class naming conventions (appending 'Schema' if missing) and file naming (underscored). This replaces the previous dependency on \ruby\_llm-schema\ with the \schematist\ library, providing a standardized way to define schema structures within Rails applications.

_lib/generators/ruby\llm/schema · high confidence

New Rails generators for evaluations, judges, MCP servers, and skills

Developers can now scaffold key RubyLLM components directly into their applications using new Rails generators. Running \rails g ruby\_llm:evaluation\ creates an evaluation class and its associated dataset YAML in \app/evals\, while \rails g ruby\_llm:judge\ scaffolds a judge class in \app/judges\. Additionally, \rails g ruby\_llm:mcp\ generates an MCP server class in \app/mcp\ configured via a URL or local command, and \rails g ruby\_llm:skill\ creates a skill folder with a \SKILL.md\ file in \app/skills\ following the open Agent Skills format.

_lib/generators/ruby\_llm/evaluation, lib/generators/ruby\_llm/judge, lib/generators/ruby\llm/skill · high confidence

New RubyLLM Rails install generator

The RubyLLM gem now provides a dedicated \ruby\_llm:install\ generator to streamline Rails integration. Running this generator creates the necessary database migrations for chat and message records, generates the corresponding model files, sets up an initializer for API configuration, and establishes convention directories for agents, tools, schemas, and prompts. It also handles ActiveStorage setup for file attachments (unless skipped) and provides clear next-step instructions for running migrations and starting to use the library.

_lib/generators/ruby\llm/install · high confidence

New Tailwind-based Chat UI templates with file and approval support

The generated chat interface now uses a new set of Tailwind CSS-styled ERB templates. Users will see a redesigned UI that includes a dedicated section for displaying attached files with download links, support for showing pending approvals, and the ability to queue messages during active conversations. The templates also integrate Turbo for real-time updates and provide a consistent, responsive layout for chat lists, forms, and detail views.

_lib/generators/ruby\_llm/chat\ui/templates/tailwind/views/chats · high confidence

New Tailwind-based model management views in the generated chat UI

The generated Rails chat UI now includes Tailwind-styled templates for browsing and viewing AI models. Users will see a model index page listing provider, name, context window, and per-million-token pricing, along with a show page displaying model details and capabilities. Both views include navigation links to start new chat sessions with the selected model.

_lib/generators/ruby\_llm/chat\ui/templates/tailwind/views/models · high confidence

New chat UI generator with Tailwind and Turbo streaming

The RubyLLM gem now includes a \ruby\_llm:chat\_ui\ generator that scaffolds a complete chat interface. This generator creates controllers, models, jobs, and views, supporting both standard Rails scaffold and Tailwind CSS styling via a \--ui\ option. It features Turbo streaming for real-time message updates, a Stimulus JavaScript controller for handling chat interactions, and support for tool approvals, file attachments, and queued messages. The generator also handles route setup, including namespaced model support, and injects necessary JavaScript dependencies. Also in this area: 5 commits (0 fixes) modifying lib/generators.

_lib/generators/ruby\_llm/chat\ui · medium confidence · unverified

New development and testing utility scripts

Added five new executable scripts to streamline the developer experience: \bin/console\ provides an interactive Ruby environment with pre-configured API keys for all supported providers (Anthropic, Azure, Bedrock, DeepSeek, Gemini, GPUStack, Mistral, Ollama, OpenAI, OpenRouter, Perplexity, Vertex AI, and xAI) via environment variables; \bin/setup\ automates dependency installation; \bin/test\ runs the test suite using a custom \rspec-queue\ runner that isolates databases per worker and merges SimpleCov coverage, with a fallback to standard RSpec for JRuby; \bin/rspec-queue\ implements this custom runner logic; and \bin/gitleaks-staged\ adds a pre-commit hook to scan staged files for secrets.

bin · high confidence

New model management views in the generated chat UI

The chat UI generator now includes dedicated views for managing AI models. Users will see a new Models index page listing available models with their provider, name, context window, and input/output pricing, along with a refresh action. A partial renders each model row, and a show view displays detailed model information including ID, capabilities, and max output tokens, with links to start a new chat or return to the list.

_lib/generators/ruby\_llm/chat\ui/templates/views/models · high confidence

New protocol layer for provider APIs

RubyLLM now includes a dedicated protocol layer in lib/ruby\_llm/protocols that defines the wire-format interfaces for each supported provider. This adds new protocol classes for Anthropic, OpenAI (Chat Completions and Responses), Cohere, AWS Bedrock (Converse and InvokeModel), Deepgram, ElevenLabs, Google Gemini, and System One, along with a generic Files protocol for provider-managed storage. These protocols implement the specific request shapes, streaming, tool-calling, and media handling required by each provider's API. Also in this area: 25 commits (0 fixes) modifying lib/ruby\_llm; 7 commits (1 fix) modifying lib/ruby\_llm/providers; Workflow tracking, tool namespacing, and long-running message continuation; Deferred tool search and content filter reporting.

_lib/ruby\llm/protocols · medium confidence · unverified

New provider generator scaffolds standalone gems with full test suites

The RubyLLM provider generator now creates a complete, standalone gem structure for new providers, including a core provider class, a Gemfile with development dependencies (RSpec, RuboCop, VCR, WebMock), and a comprehensive suite of live integration specs covering chat, streaming, tools, structured output, embeddings, image generation, speech, video, moderation, and reranking. The generated gem also includes CI workflows for testing across Ruby versions and publishing to RubyGems, along with configuration helpers for API keys and base URLs.

_lib/generators/ruby\llm/provider/templates/gem · high confidence

New provider generator script

A new executable script at script/generate-provider has been added to the project. This tool allows users to generate new provider implementations by invoking the RubyLLM generator scaffolding, providing a streamlined way to create core provider changes and handle existing files.

script · high confidence

OpenRouter provider now supports multimodal inputs, reasoning, and specialized generation capabilities

The OpenRouter integration has been significantly expanded to support a wider range of model capabilities. Users can now include video attachments in chat conversations, utilize prompt caching to reduce latency and cost, and leverage extended thinking (reasoning) models with detailed cost reporting. The provider also adds dedicated support for image generation, text-to-speech, audio transcription, and asynchronous video generation. Additionally, embeddings now support multimodal inputs (audio, video, PDF), and the model catalog automatically discovers and exposes capabilities like tool calling and structured outputs from OpenRouter's various model endpoints.

_lib/ruby\llm/providers/openrouter · high confidence

Rails integration now supports Action Text rich content and Active Storage attachments

The \acts\_as\_chat\ and \acts\_as\_message\ helpers in the Rails integration now fully support Action Text rich content and Active Storage attachments. Messages can include Action Text content with embedded attachments, which are automatically extracted and persisted to Active Storage. The integration handles both incoming attachments from the LLM and outgoing attachments from the user, including file uploads and URL-based attachments. This enables richer conversations with media support in Rails applications using the ActiveRecord-backed chat models. Also in this area: 20 commits (0 fixes) modifying lib/ruby\_llm/providers; 7 commits (1 fix) modifying lib/ruby\_llm/providers; Enable document URL support and refine OpenRouter vision capabilities.

_lib/ruby\_llm/providers/azure, lib/ruby\llm/providers/xai · medium confidence · unverified

Support for prompt partials and scoped variable binding

Prompt templates can now include reusable partials via the \render\ helper, allowing users to compose complex prompts from smaller, modular pieces. Additionally, local variables passed to templates are now strictly bound only if they match valid Ruby local variable names, ensuring that capitalized or non-standard Unicode locals are handled safely without causing binding errors.

_lib/ruby\llm/prompt · high confidence

System One protocol adds support for model judgments

The System One protocol now supports submitting and parsing judgment requests. This change introduces a new judgments module to render payloads for probability, choice, and score-based evaluations, a models module to list available models (marking them with 'judgment' capabilities), and a responses module to parse the resulting judgment answers, including handling probabilities, choices, and scores with confidence levels. Also in this area: 6 commits (0 fixes) modifying lib/ruby\_llm; Workflow tracking, tool namespacing, and long-running message continuation.

_lib/ruby\_llm/protocols/system\one · medium confidence · unverified

Vertex AI provider now supports Claude, Mistral, Gemini, and MaaS models with batch and media capabilities

The Vertex AI provider implementation has been expanded to support a wider range of model families beyond just Google's Gemini. It now routes Anthropic Claude models through the Anthropic protocol and Mistral models through their native protocol, in addition to the existing Gemini support. The provider also supports Meta and other publishers' MaaS models via the OpenAI-compatible endpoint. New capabilities include batch prediction for all supported model types, live and dedicated transcription, video generation (Veo), and embedding support. The model catalog listing has been improved to fetch from multiple regions and includes a fallback list of known Google models to prevent silent shrinking of available options.

_lib/ruby\llm/providers/vertexai · high confidence

Architecture

Complete provider architecture refactoring

The provider implementation has been completely refactored from a module-based structure to a class-based architecture. The previous \lib/ruby\_llm/providers/base.rb\ module has been removed and replaced with a new \Provider\ base class. All individual providers (Anthropic, Azure, Bedrock, Cohere, Deepgram, DeepSeek, ElevenLabs, Gemini, GPUStack, Mistral, Ollama, OllamaCloud, OpenAI, OpenRouter, Perplexity, VertexAI, and xAI) are now implemented as classes inheriting from this new base, utilizing a protocol-based system to define their specific API interactions, configuration options, and capabilities.

_lib/ruby\llm/providers · high confidence

Behavioural changes

1 commit (0 fixes) modifying config, lib/generators/ruby\_llm/chat\_ui/templates/javascript

A change to existing behaviour in config, lib/generators/ruby\_llm/chat\_ui/templates/javascript — 1 commit, 2 files: “Ship the chat UI's Stimulus controller with the gem (\#1527)” (refs \#1527).

_config, lib/generators/ruby\_llm/chat\ui/templates/javascript · low confidence · unverified

1 commit (0 fixes) modifying lib/ruby\_llm/protocols/meta

A change to existing behaviour in lib/ruby\_llm/protocols/meta — 1 commit, 2 files: “Add Meta transcription and Files support (\#1607)” (refs \#1607).

_lib/ruby\llm/protocols/meta · medium confidence · unverified

1 commit (0 fixes) modifying lib/ruby\_llm/protocols/zai

A change to existing behaviour in lib/ruby\_llm/protocols/zai — 1 commit, 1 file: “Add Z.ai token counts, video, OCR and uploads”.

_lib/ruby\llm/protocols/zai · medium confidence · unverified

10 commits (0 fixes) modifying lib/ruby\_llm/file\_store

A change to existing behaviour in lib/ruby\_llm/file\_store — 10 commits, 4 files: “Batch file edits and extract Office text on attach (\#1462)”; “Recognize extensionless media in directory file stores (\#1342)”; “Delete file store folders with recursive: true (\#1376)”; and 7 more (refs \#1462, \#1342, \#1376, \#1353, \#1443, …).

_lib/ruby\_llm/file\store · medium confidence · unverified

10 commits (1 fix) modifying lib/ruby\_llm/protocols/mistral

A change to existing behaviour in lib/ruby\_llm/protocols/mistral — 10 commits (1 fix), 4 files: “Send upload options and stop dropping arguments (\#1398)”; “Expand provider APIs and align framework internals”; “Raise on options operations cannot carry (\#1379)”; and 7 more (refs \#1398, \#1379, \#1428, \#1393, \#1136, …).

_lib/ruby\llm/protocols/mistral · medium confidence · unverified

12 commits (0 fixes) modifying lib/ruby\_llm/open\_telemetry

A change to existing behaviour in lib/ruby\_llm/open\_telemetry — 12 commits, 2 files: “Exercise OpenTelemetry with real calls and concurrency”; “Make tool\_calls a collection of ToolCall (\#1380)”; “Record the thinking effort as the reasoning level (\#1520)”; and 9 more (refs \#1380, \#1520, \#1522, \#1518, \#1333, …).

_lib/ruby\_llm/open\telemetry · medium confidence · unverified

12 commits (1 fix) modifying lib/ruby\_llm/protocols/elevenlabs

A change to existing behaviour in lib/ruby\_llm/protocols/elevenlabs — 12 commits (1 fix), 6 files: “Record characters, audio seconds, and search units (\#1416)”; “Send upload options and stop dropping arguments (\#1398)”; “Expand provider APIs and align framework internals”; and 9 more (refs \#1416, \#1398, \#1379, \#1353, \#1375, …).

_lib/ruby\llm/protocols/elevenlabs · medium confidence · unverified

13 commits (0 fixes) modifying lib/ruby\_llm/code\_execution

A change to existing behaviour in lib/ruby\_llm/code\_execution — 13 commits, 3 files: “Keep Ruby's long-bundled gems in code execution (\#1296)”; “Decide tool calls with permission modes”; “Keep the cause when a package install fails (\#1227)”; and 10 more (refs \#1296, \#1227, \#1103, \#1299, \#1233, …).

_lib/ruby\_llm/code\execution · medium confidence · unverified

13 commits (0 fixes) modifying lib/ruby\_llm/protocols/invoke\_model

A change to existing behaviour in lib/ruby\_llm/protocols/invoke\_model — 13 commits, 6 files: “Report input tokens for Bedrock Cohere embeddings (\#1161)”; “Expand provider APIs and align framework internals”; “Raise on options operations cannot carry (\#1379)”; and 10 more (refs \#1161, \#1379, \#1113, \#1627, \#1374, …).

_lib/ruby\_llm/protocols/invoke\model · medium confidence · unverified

15 commits (1 fix) modifying lib/ruby\_llm/protocols/deepgram

A change to existing behaviour in lib/ruby\_llm/protocols/deepgram — 15 commits (1 fix), 4 files: “Record characters, audio seconds, and search units (\#1416)”; “Send upload options and stop dropping arguments (\#1398)”; “Expand provider APIs and align framework internals”; and 12 more (refs \#1416, \#1398, \#1379, \#1114, \#1217, …).

_lib/ruby\llm/protocols/deepgram · medium confidence · unverified

16 commits (2 fixes) modifying lib/ruby\_llm/model

A change to existing behaviour in lib/ruby\_llm/model — 16 commits (2 fixs), 4 files: “Publish the model registry and make refresh! persist”; “Record characters, audio seconds, and search units (\#1416)”; “refactor(Model): Rename ModelInfo to Model::Info and update related references”; and 13 more (refs \#1416, \#854, \#859, \#1381).

_lib/ruby\llm/model · medium confidence · unverified

19 commits (0 fixes) modifying lib/ruby\_llm/accounting

A change to existing behaviour in lib/ruby\_llm/accounting — 19 commits, 2 files: “Record one-shot usage in the ledger with an owner”; “Price streaming attempts only when read”; “Expand provider APIs and align framework internals”; and 16 more (refs \#1199, \#1409, \#1471, \#1300, \#1126, …).

_lib/ruby\llm/accounting · medium confidence · unverified

19 commits (1 fix) modifying lib/ruby\_llm/protocols/bedrock

A change to existing behaviour in lib/ruby\_llm/protocols/bedrock — 19 commits (1 fix), 6 files: “Send upload options and stop dropping arguments (\#1398)”; “Expand provider APIs and align framework internals”; “Raise on options operations cannot carry (\#1379)”; and 16 more (refs \#1398, \#1379, \#1024, \#1374, \#1410, …).

_lib/ruby\llm/protocols/bedrock · medium confidence · unverified

19 commits (1 fix) modifying lib/ruby\_llm/protocols/vertexai

A change to existing behaviour in lib/ruby\_llm/protocols/vertexai — 19 commits (1 fix), 6 files: “Record characters, audio seconds, and search units (\#1416)”; “Send upload options and stop dropping arguments (\#1398)”; “Expand provider APIs and align framework internals”; and 16 more (refs \#1416, \#1398, \#1379, \#1408, \#1332, …).

_lib/ruby\llm/protocols/vertexai · medium confidence · unverified

2 commits (0 fixes) modifying app/assets

A change to existing behaviour in app/assets — 2 commits, 3 files: “Show a reply error the chat page missed (\#1544)”; “Ship the chat UI's Stimulus controller with the gem (\#1527)” (refs \#1544, \#1527).

app/assets · medium confidence · unverified

2 commits (0 fixes) modifying lib/generators/ruby\_llm/chat\_ui/templates/tailwind/views/questions, lib/generators/ruby\_llm/chat\_ui/templates/views/questions

A change to existing behaviour in lib/generators/ruby\_llm/chat\_ui/templates/tailwind/views/questions, lib/generators/ruby\_llm/chat\_ui/templates/views/questions — 2 commits, 2 files: “Submit free-form array answers from chat cards (\#1653)”; “Answer every question from one card in the chat UI (\#1541)” (refs \#1653, \#1541).

_lib/generators/ruby\_llm/chat\_ui/templates/tailwind/views/questions, lib/generators/ruby\_llm/chat\ui/templates/views/questions · medium confidence · unverified

20 commits (0 fixes) modifying lib/ruby\_llm/providers

A change to existing behaviour in lib/ruby\_llm/providers — 20 commits, 25 files: “Send video and audio on Meta, MiniMax, Kimi and Z.ai (\#1626)”; “Complete OpenRouter batches on its documented route (\#1629)”; “Add structured output on MiniMax-Text-01 (\#1621)”; and 17 more (refs \#1626, \#1629, \#1621, \#1611, \#1613, …). Also in this area: Introduce Agent, Batch, and Caching abstractions alongside model registry updates.

_lib/ruby\_llm/providers/meta, lib/ruby\_llm/providers/moonshot, lib/ruby\llm/providers/zai · medium confidence · unverified

21 commits (1 fix) modifying lib/ruby\_llm/support

A change to existing behaviour in lib/ruby\_llm/support — 21 commits (1 fix), 10 files: “Add a built-in MCP client (\#959)”; “Upload again when a provider deletes a reused file”; “Exercise OpenTelemetry with real calls and concurrency”; and 18 more (refs \#959, \#1367, \#1326, \#1548, \#1221, …).

_lib/ruby\llm/support · medium confidence · unverified

24 commits (3 fixes) modifying lib/ruby\_llm/protocols/interactions

A change to existing behaviour in lib/ruby\_llm/protocols/interactions — 24 commits (3 fixs), 7 files: “Expand provider APIs and align framework internals”; “Keep Interactions search suggestions out of history”; “Add video processing modes to attachments (\#1082)”; and 21 more (refs \#1082, \#1380, \#1360, \#1373, \#1382, …).

_lib/ruby\llm/protocols/interactions · medium confidence · unverified

25 commits (1 fix) modifying lib/ruby\_llm/protocols/openai

A change to existing behaviour in lib/ruby\_llm/protocols/openai — 25 commits (1 fix), 3 files: “Send OpenAI upload fields before the file part (\#1301)”; “Send upload options and stop dropping arguments (\#1398)”; “Raise on options operations cannot carry (\#1379)”; and 22 more (refs \#1301, \#1398, \#1379, \#1165, \#1177, …).

_lib/ruby\llm/protocols/openai · medium confidence · unverified

26 commits (0 fixes) modifying lib/ruby\_llm/protocol

A change to existing behaviour in lib/ruby\_llm/protocol — 26 commits, 8 files: “Send the phase of Responses answers back (\#1236)”; “Upload again when a provider deletes a reused file”; “Expand provider APIs and align framework internals”; and 23 more (refs \#1236, \#1241, \#1380, \#1284, \#1248, …).

_lib/ruby\llm/protocol · medium confidence · unverified

275 commits (29 fixes) modifying lib/ruby\_llm/active\_record

A change to existing behaviour in lib/ruby\_llm/active\_record — 275 commits (29 fixs), 23 files: “Approve a persisted tool call without building the chat (\#1645)”; “FINALLY. Removed debug prints”; “Rename on\_unsupported\_attachment to convert\_unsupported\_attachments”; and 271 more (refs \#1645, \#796, \#1530, \#1223, \#1205, …).

_lib/ruby\_llm/active\record · medium confidence · unverified

3 commits (0 fixes) modifying conformance

A change to existing behaviour in conformance — 3 commits, 3 files: “Exercise OpenTelemetry with real calls and concurrency”; “Split conformance by suite and drop the tour sources (\#1441)”; “Register code execution attributes for conformance (\#1516)” (refs \#1441, \#1516).

conformance · medium confidence · unverified

3 commits (0 fixes) modifying lib/generators/ruby\_llm/agent

A change to existing behaviour in lib/generators/ruby\_llm/agent — 3 commits, 1 file: “Find generated files with the app's inflections (\#1221)”; “Add Rails AI generators and refine chat UI conventions”; “Add ApplicationAgent and a ChatAgent for the chat UI” (refs \#1221).

_lib/generators/ruby\llm/agent · medium confidence · unverified

33 commits (3 fixes) modifying lib/ruby\_llm/transport

A change to existing behaviour in lib/ruby\_llm/transport — 33 commits (3 fixs), 8 files: “Add config.ssl for TLS options per context (\#1559)”; “Classify Azure's streamed rate limits as rate limits”; “Share HTTP connections between calls”; and 30 more (refs \#1559, \#968, \#1024, \#1354, \#1599, …).

_lib/ruby\llm/transport · medium confidence · unverified

36 commits (2 fixes) modifying tasks

A change to existing behaviour in tasks — 36 commits (2 fixs), 20 files: “Validate the serialized registry in models:update”; “Publish the model registry and make refresh! persist”; “Date the Available Models page and link the raw registry”; and 33 more (refs \#1421, \#1483, \#1254, \#1441, \#1549, …).

tasks · medium confidence · unverified

4 commits (0 fixes) modifying benchmarks

A change to existing behaviour in benchmarks — 4 commits, 14 files: “Add benchmarks for RubyLLM's own overhead”; “Use a registered model in the streaming benchmark”; “Drop Ruby 3.1, which reached end of life”; and 1 more (refs \#1546).

benchmarks · medium confidence · unverified

4 commits (0 fixes) modifying lib/generators/ruby\_llm/provider/templates/core

A change to existing behaviour in lib/generators/ruby\_llm/provider/templates/core — 4 commits, 2 files: “Simplify provider architecture and gem generation”; “Ask providers for their OpenTelemetry and models.dev names”; “Finish provider generator”; and 1 more.

_lib/generators/ruby\llm/provider/templates/core · medium confidence · unverified

4 commits (0 fixes) modifying lib/ruby\_llm/tools

A change to existing behaviour in lib/ruby\_llm/tools — 4 commits, 1 file: “Search deferred tools on providers without tool search (\#1283)”; “Preserve native client toolset identity and results (\#1247)”; “Make file stores safe and grep-shaped for agents (\#1443)”; and 1 more (refs \#1283, \#1247, \#1443, \#1451).

_lib/ruby\llm/tools · medium confidence · unverified

45 commits (1 fix) modifying lib/ruby\_llm/chat

A change to existing behaviour in lib/ruby\_llm/chat — 45 commits (1 fix), 11 files: “Approve a persisted tool call without building the chat (\#1645)”; “Unstick over-budget chats, widen the usage check (\#1473)”; “Queue messages that join a running conversation (\#1223)”; and 42 more (refs \#1645, \#1473, \#1223, \#1465, \#1462, …).

_lib/ruby\llm/chat · medium confidence · unverified

46 commits (4 fixes) modifying lib/ruby\_llm/protocols/cohere

A change to existing behaviour in lib/ruby\_llm/protocols/cohere — 46 commits (4 fixs), 12 files: “Record characters, audio seconds, and search units (\#1416)”; “Record Cohere live cassettes and fix two bugs they exposed”; “Preserve Cohere reasoning token counts (\#1273)”; and 43 more (refs \#1416, \#1273, \#1398, \#1379, \#1241, …).

_lib/ruby\llm/protocols/cohere · medium confidence · unverified

5 commits (0 fixes) modifying lib/generators/ruby\_llm/sandbox

A change to existing behaviour in lib/generators/ruby\_llm/sandbox — 5 commits, 4 files: “Generate MCP servers, skills, sandboxes, evals, and judges”; “Generate a system test and an API sheet for agents (\#1501)”; “Set config.sandbox and require mxc 0.4.1 (\#1487)”; and 2 more (refs \#1501, \#1487, \#1535).

_lib/generators/ruby\llm/sandbox · medium confidence · unverified

5 commits (0 fixes) modifying lib/ruby\_llm/providers/minimax

A change to existing behaviour in lib/ruby\_llm/providers/minimax — 5 commits, 3 files: “Send video and audio on Meta, MiniMax, Kimi and Z.ai (\#1626)”; “Add structured output on MiniMax-Text-01 (\#1621)”; “Route MiniMax web search to Responses for citations (\#1637)”; and 2 more (refs \#1626, \#1621, \#1637, \#1605, \#1253).

_lib/ruby\llm/providers/minimax · medium confidence · unverified

6 commits (0 fixes) modifying lib/ruby\_llm/file\_store/converters

A change to existing behaviour in lib/ruby\_llm/file\_store/converters — 6 commits, 7 files: “Batch file edits and extract Office text on attach (\#1462)”; “Read Office files as a model needs them (\#1464)”; “Find Office parts named with backslashes (\#1474)”; and 3 more (refs \#1462, \#1464, \#1474, \#1469, \#1366, …).

_lib/ruby\_llm/file\store/converters · medium confidence · unverified

7 commits (0 fixes) modifying lib/generators/ruby\_llm/agent/templates

A change to existing behaviour in lib/generators/ruby\_llm/agent/templates — 7 commits, 2 files: “Treat a blank agent prompt as no instructions”; “Rename the agent chat\_model macro to chat\_class (\#1368)”; “Give generated agents a starter prompt”; and 4 more (refs \#1368, \#1366).

_lib/generators/ruby\llm/agent/templates · medium confidence · unverified

7 commits (0 fixes) modifying lib/ruby\_llm/protocols/xai

A change to existing behaviour in lib/ruby\_llm/protocols/xai — 7 commits, 3 files: “Send upload options and stop dropping arguments (\#1398)”; “Expand provider APIs and align framework internals”; “Raise on options operations cannot carry (\#1379)”; and 4 more (refs \#1398, \#1379, \#1141, \#1393).

_lib/ruby\llm/protocols/xai · medium confidence · unverified

84 commits (2 fixes) modifying lib/ruby\_llm/protocols/responses

A change to existing behaviour in lib/ruby\_llm/protocols/responses — 84 commits (2 fixs), 10 files: “Return web search sources on the Responses API (\#1455)”; “Add the Responses API protocol and make it OpenAI's default”; “Send the phase of Responses answers back (\#1236)”; and 81 more (refs \#1455, \#1236, \#1145, \#1241, \#1005, …).

_lib/ruby\llm/protocols/responses · medium confidence · unverified

87 commits (1 fix) modifying lib/ruby\_llm/mcp

A change to existing behaviour in lib/ruby\_llm/mcp — 87 commits (1 fix), 22 files: “Give a copied chat its own state (\#1152)”; “Key MCP outages on the context's connection settings (\#1577)”; “Add a built-in MCP client (\#959)”; and 84 more (refs \#1152, \#1577, \#959, \#1432, \#1419, …).

_lib/ruby\llm/mcp · medium confidence · unverified

9 commits (0 fixes) modifying lib/ruby\_llm/providers/bedrock/mantle

A change to existing behaviour in lib/ruby\_llm/providers/bedrock/mantle — 9 commits, 4 files: “Expand provider APIs and align framework internals”; “Re-sign Bedrock retries and retry TLS errors (\#1024)”; “Drop stale thinking blocks on Claude 5.5 (\#1347)”; and 6 more (refs \#1024, \#1347, \#1641, \#1136).

_lib/ruby\llm/providers/bedrock/mantle · medium confidence · unverified

Add Vertex AI embedding request validation and response parsing

This change introduces the internal request validation and result parsing logic for Vertex AI embedding predictions. The new \Requests\ module validates that embedding batches contain non-empty lists of text strings, ensures payload keys match supported formats (instances, requests, or content), and enforces specific option restrictions for Gemini models. The \Results\ module handles parsing the provider's response, extracting embedding vectors and token counts from various response structures (including legacy and Gemini formats), and reconstructing the original request order using record keys.

_lib/ruby\_llm/protocols/vertexai/embedding\prediction · high confidence

Anthropic provider now exposes tool choice and parallel tool call capabilities

The Anthropic provider now explicitly advertises support for tool choice and parallel tool calls. Previously, these capabilities were not reported even when the underlying Anthropic API supported them; the new \Capabilities\ module augments the standard function-calling capabilities with these specific features, ensuring that applications relying on these advanced tool-use patterns are correctly informed of the provider's support.

_lib/ruby\llm/providers/anthropic · high confidence

Database schema updates for Ruby LLM 2.2 upgrade

The upgrade migration to version 2.2 introduces several database schema changes to support new features. It adds a \provider\_call\_id\ column to the \ruby\_llm\_tool\_calls\ table (renaming \provider\_id\ if it exists) and a \namespace\ column. The chat table gains \claim\_token\ and \claimed\_until\ columns to manage exclusive access, while the message table receives \phase\ and \continuation\ columns to track response states and long-form content. New columns (\workflow\_id\, \workflow\_name\, \workflow\_step\_id\, \workflow\_step\_name\) are added to the \ruby\_llm\_usages\ table with an index on \workflow\_id\ to support workflow tracking. Additionally, two new tables are created: \ruby\_llm\_queued\_messages\ for handling messages in running conversations and \ruby\_llm\_chat\_files\ for storing file metadata and content associated with chats. The migration is reversible, dropping these additions in the \down\ method.

_lib/generators/ruby\llm/upgrade/templates · high confidence

DeepSeek provider adds tool choice support and refined reasoning handling

The DeepSeek provider now explicitly supports the \tool\_choice\ capability, ensuring it is advertised in the registry when function calling is enabled. The chat implementation has been updated to handle DeepSeek-specific thinking controls, mapping the \thinking\ parameter to the API's \reasoning\_effort\ and \thinking\ fields, while also degrading unsupported JSON schema requests to \json\_object\ mode. Additionally, a new Responses protocol implementation correctly parses DeepSeek's streaming reasoning text deltas and formats assistant messages to include reasoning content as distinct parts, improving the accuracy of multi-turn conversations involving chain-of-thought outputs.

_lib/ruby\llm/providers/deepseek · high confidence

Gemini provider now explicitly declares tool choice and transcription capabilities

The Gemini provider now includes a dedicated capabilities module that augments the model's feature set based on its ID and supported modalities. Specifically, if a model supports function calling, the provider now explicitly exposes 'tool\_choice' support, allowing users to control how tools are selected. Additionally, for models that accept audio input and produce text output (and are not embedding models), the provider now declares 'transcription' capability, ensuring these features are correctly recognized by the library's capability system.

_lib/ruby\llm/providers/gemini · high confidence

Introduce skill loading tool with code execution support

Added a new \load\_skill\ tool that allows the model to retrieve skill instructions and read associated files. The tool now distinguishes between environments with and without code execution: when code execution is enabled, skill folders are mounted read-only in the run\_code working directory, allowing scripts to be executed; otherwise, scripts are noted as unavailable and the model is instructed to follow instructions without running code.

_lib/ruby\llm/skill · high confidence

Introduce typed judgment DSL and data validation

The Judge component now supports a new builder DSL that allows users to define judgment fields via method calls or blocks, automatically resolving runtime inputs from the scope. This change introduces strict validation for judgment data, ensuring that only JSON-compatible types (strings, symbols, booleans, finite numbers, arrays, and hashes) are accepted, while rejecting unsupported types and duplicate keys. Users can now write more structured and type-safe judgments with immediate feedback on invalid data structures.

_lib/ruby\llm/judge · high confidence

Major framework rewrite and API expansion

The library has been significantly refactored, replacing the previous minimal structure with a comprehensive, documented API. The main entry point now includes extensive inline documentation covering new capabilities such as image painting, video animation, speech synthesis, OCR, transcription, embeddings, and typed judgments. The codebase has been expanded to support a wide range of model providers (including Azure, Bedrock, DeepSeek, ElevenLabs, GPUStack, Hetzner, Mistral, OpenRouter, Perplexity, VertexAI, and xAI) and introduces new features like MCP client support, provider-side batching, caching, and workflow instrumentation. The internal architecture has been reorganized using Zeitwerk for autoloading with specific inflections, and the public interface now exposes methods for chat, painting, animating, speaking, transcribing, OCR, embedding, reranking, moderation, research, and judgment.

lib · high confidence

OpenAI provider adds explicit capability detection and model aliasing

The OpenAI provider now explicitly recognizes specific model capabilities such as function calling, structured output, vision, reasoning, transcription, and citations for newly listed models (including gpt-5, codex, search, and deep-research variants). It also introduces logic to map specific dated model snapshots (e.g., gpt-3.5-turbo-0125) to their base model definitions and implements the Responses API protocol with token counting and compaction support.

_lib/ruby\llm/providers/openai · high confidence

Refactored generator helpers to support namespaced models and Tailwind UI variants

The RubyLLM generator helpers have been rewritten to better support namespaced models and offer a Tailwind CSS UI option. The new \GeneratorHelpers\ module introduces logic to automatically create namespace modules when model names include namespaces (e.g., \Admin::Chat\), ensuring correct table name prefixes and association handling. It also adds a \ui\_variant\ method that detects Tailwind availability and switches generated views to Tailwind-specific templates if available, falling back to standard scaffolds otherwise. Additionally, helper methods for model names, table names, and controller class names are now dynamically derived to handle these namespaces correctly, fixing previous issues with namespaced model generation. Also in this area: 5 commits (0 fixes) modifying lib/generators.

_lib/generators/ruby\llm · medium confidence · unverified

Refined chat UI controller templates with approval and queuing support

The generated chat UI controllers have been updated to support new interaction patterns: the messages controller now queues user inputs when a conversation is already running and blocks new messages if there are pending tool-call approvals, while a new approvals controller allows users to explicitly approve or deny tool calls. Additionally, a dedicated queued messages controller enables users to remove messages from the processing queue, and the chats controller now includes a destroy action for deleting conversations.

_lib/generators/ruby\_llm/chat\ui/templates/controllers · high confidence

Refined chat UI message templates with streaming, queuing, and attachment support

The generated chat UI templates for messages have been updated to provide a more robust user experience. Assistant messages now display a 'Thinking…' state during streaming and hide the input form while pending approvals are waiting. User messages and queued messages are now rendered with distinct styles, including support for displaying file attachments (images and other files) and timestamps. The system also includes new templates for error states, tool calls, and system messages, ensuring that all message types are handled consistently in the generated interface.

_lib/generators/ruby\_llm/chat\ui/templates/views/messages · high confidence

Schema migration for RubyLLM 2.1 upgrade

The upgrade generator now provides a migration path from RubyLLM 2.0 to 2.1. This migration introduces a new \ruby\_llm\_mcp\_credentials\ table to store credentials for the built-in MCP client, supporting polymorphic owners and unique keys. It also adds a \pending\_input\ column to the existing \ruby\_llm\_tool\_calls\ table, using \jsonb\ for PostgreSQL and \json\ for other database adapters, to support new tool call input handling. Also in this area: Per-boundary cache lifetimes and improved cache pricing.

_lib/generators/ruby\llm/upgrade · high confidence

Standardized tool result display with error handling and attachments

The chat UI generator now uses a unified \\_default.html.erb.tt\ template for rendering tool results, applying consistently to both Tailwind and default view styles. This change ensures that tool results display a clear label, the tool's content (or a placeholder if empty), and any associated attachments. Crucially, it introduces error handling: if a tool execution fails, the UI renders a specific error view with the error message instead of the result content, improving visibility into tool failures for users.

_lib/generators/ruby\_llm/chat\_ui/templates/tailwind/views/messages/tool\_results, lib/generators/ruby\_llm/chat\_ui/templates/views/messages/tool\results · high confidence

Tool generator now creates correctly named view partials

The RubyLLM tool generator has been updated to produce tool call and result view partials with filenames that match the runtime lookup keys used by the chat UI. Previously, partials were generated using string manipulation that could result in mismatched names; the new implementation derives the partial name from the tool class name (stripping the 'Tool' suffix and handling namespaces) to ensure the generated views are found correctly by the application.

_lib/generators/ruby\llm/tool · high confidence

Updated Rails install generator templates for RubyLLM 2.0 schema and models

The install generator templates have been updated to reflect the RubyLLM 2.0 data model and configuration. The new schema includes dedicated tables for models, chats, messages, tool calls, MCP credentials, usage ledgers, and batches, with support for PostgreSQL JSONB and MySQL JSON types. Message models now include fields for caching, thinking text, citations, and raw content, while chat models support cancellation. The generator also creates a simplified initializer template for API key configuration and updates the message model to optionally include ActiveStorage attachments. Also in this area: Workflow tracking, tool namespacing, and long-running message continuation; Per-boundary cache lifetimes and improved cache pricing. Also in this area: 5 commits (0 fixes) modifying lib/generators.

_lib/generators/ruby\llm/install/templates · medium confidence · unverified

Updated Tailwind chat UI message templates with new message types and attachment support

The generated Tailwind chat UI templates have been updated to support a richer set of message types and interactive features. New partials allow rendering of system messages, tool calls, and queued user messages (with cancel options), while assistant and user message views now explicitly handle file attachments via Active Storage. The assistant view also includes logic to display 'Thinking…' during streaming and uses Stimulus controllers for morphing restoration, and the form view now conditionally hides input fields when pending approvals are present.

_lib/generators/ruby\_llm/chat\ui/templates/tailwind/views/messages · high confidence

Updated default tool-call display templates in generated chat UI

The generated chat UI now uses new default partials for rendering tool calls, available for both Tailwind and standard CSS styling. These templates display the tool name, arguments, and timestamp within a styled container, ensuring consistent presentation of tool interactions in the chat interface.

_lib/generators/ruby\_llm/chat\_ui/templates/tailwind/views/messages/tool\_calls, lib/generators/ruby\_llm/chat\_ui/templates/views/messages/tool\calls · high confidence

Test coverage

1 commit adding/updating tests in spec/github\_scripts; 1 commit adding/updating tests in spec/ruby\_llm/accounting/usage, spec/ruby\_llm/files; 1 commit adding/updating tests in spec/ruby\_llm/protocols/azure; 12 commits adding/updating tests in spec/ruby\_llm/providers/deepseek; 12 commits adding/updating tests in spec/ruby\_llm/providers/vertexai; 13 commits adding/updating tests in spec/ruby\_llm/protocols/elevenlabs; 14 commits adding/updating tests in spec/ruby\_llm/file\_store; 14 commits adding/updating tests in spec/ruby\_llm/protocols/bedrock; 14 commits adding/updating tests in spec/ruby\_llm/protocols/perplexity; 15 commits adding/updating tests in spec/ruby\_llm/protocols/vertexai; 16 commits adding/updating tests in spec/ruby\_llm/accounting; 16 commits adding/updating tests in spec/ruby\_llm/chat; 17 commits adding/updating tests in spec; 20 commits adding/updating tests in spec/docs; 20 commits adding/updating tests in spec/ruby\_llm/providers/xai; 292 commits adding/updating tests in spec/ruby\_llm/active\_record; 3 commits adding/updating tests in lib/generators/ruby\_llm/chat\_ui/templates/test/agents; 3 commits adding/updating tests in spec/ruby\_llm/file\_store/converters; 3 commits adding/updating tests in spec/ruby\_llm/protocols/xai; 38 commits adding/updating tests in spec/ruby\_llm/protocols; 38 commits adding/updating tests in spec/ruby\_llm/protocols/cohere; 38 commits adding/updating tests in spec/ruby\_llm/transport; 5 commits adding/updating tests in spec/docs/compare; 5 commits adding/updating tests in spec/ruby\_llm/protocols/interactions; 5 commits adding/updating tests in spec/ruby\_llm/providers/vertexai/chat\_completions; 6 commits adding/updating tests in spec/ruby\_llm/protocols/openai; 62 commits adding/updating tests in spec/ruby\_llm/protocols/responses; 66 commits adding/updating tests in spec/ruby\_llm/protocols/chat\_completions; 71 commits adding/updating tests in spec/ruby\_llm/mcp; 75 commits adding/updating tests in spec/ruby\_llm/protocols/gemini; 8 commits adding/updating tests in lib/generators/ruby\_llm/chat\_ui/templates/test/system; 8 commits adding/updating tests in spec/ruby\_llm/models; 8 commits adding/updating tests in spec/ruby\_llm/protocols/deepgram; 9 commits adding/updating tests in spec/ruby\_llm/support; Added MCP server fixture for testing; Added a dummy Rails application for integration testing; Added comprehensive test coverage for RubyLLM generators; Added comprehensive test coverage for the Agent API and core components; Added comprehensive test coverage for the OpenRouter provider; Added fixture templates for chat UI generation tests; Added test coverage for Anthropic protocol internals; Added test coverage for Azure provider protocols and media operations; Added test coverage for Bedrock Converse protocol features; Added test coverage for Bedrock Mantle protocol, model routing, and transcription capabilities; Added test coverage for Bedrock embedding and image generation protocols; Added test coverage for Mistral Conversations and Multi-Completion protocols; Added test coverage for Mistral provider capabilities, chat, models, OCR, speech, and transcription; Added test coverage for OpenRouter batch processing, Responses protocol, and transcription; Added test coverage for Rails integration and usage tracking; Added test coverage for model catalog, registry, and release workflows; Added test coverage for protocol streaming, tool call handling, and request shape analysis; Added test coverage for the RubyLLM evaluation subsystem; Added test fixtures for Agent Skills; Added test suite for gemspec constraints, logger configuration, and signed media URL sanitization; Added tests for GPUStack protocol responses, tokenization, and video generation; Added tests for Mistral Conversations protocol image handling and request shape descriptions; Added tests for Mistral batch embedding validation and job handling; Added tests for Ollama provider payload rendering and media handling; Added tests for Perplexity model listing, parsing, and error handling; Added tests for Ruby and Node.js code execution environments; Added tests for Vertex AI batch processing in Anthropic and Gemini providers; Added tests for the Landlock and MXC sandbox implementations; Added tests for the provider generator CLI and scaffold; Added tests for thinking control logic on Claude Sonnet 5.5; Added tests for xAI batch processing and status parsing; Added text and CSV fixture files for testing; Centralized test infrastructure and provider coverage matrix; Expanded test coverage for provider configurations and capabilities.

Dependencies

Expanded Rails version support to 8.1 with updated dependency constraints

The gemfiles have been updated to include a new appraisal configuration for Rails 8.1 (specifically \~\> 8.1.4), extending the library's compatibility matrix alongside existing support for Rails 7.1, 7.2, and 8.0. This change also adjusts the \json\ gem constraint for the Rails 8.1 target to allow version 3 (\>= 3, \< 4), whereas earlier Rails versions remain pinned below version 3, ensuring compatibility with the specific dependency requirements of each Rails release.

gemfiles · high confidence

RubyLLM 2.0 dependency overhaul and Rails integration restructuring

The gem's runtime dependencies have been significantly refactored for the 2.0 release: \zeitwerk\ is now pinned to \\~\> 2\, \faraday\ constraints have been adjusted (allowing \\>= 1.10.0\ via environment variable or defaulting to that range), and new dependencies including \base64\, \event\_stream\_parser\, \faraday-net\_http\, \faraday-retry\, \marcel\, and \schematist\ have been added. The \activesupport\ and \activerecord\ dependencies are now explicitly optional, and the minimum required Ruby version has been raised to 3.1.3. Additionally, the project structure now includes dedicated Gemfiles for the main library, documentation site, and a dummy Rails application used for integration testing, replacing previous ad-hoc dependency management.

(dependencies) · high confidence

Written by cca.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.

How this codebase got here

This is the PUBLIC form of this artifact. Findings are listed in full, but the details of SECURITY findings — which rule fired, in which file, on which line, and how to fix it — are deliberately withheld, and any secret-scanner results are excluded entirely. Where detail is absent here it was REMOVED FOR PUBLICATION; it is not missing from the analysis. The complete artifact is available from the repository owner.

Score

  • CAI 62 → 61 (-0.8)
  • Rubric changed (rubric-2026.10.5 → rubric-2026.11.4) — scores are not directly comparable.

Lenses

  • Code Health 69 → 68 (-0.4)
  • Architecture 76 → 76 (+0.5)
  • Maturity 63 → 65 (+1.9)
  • Readiness 72 → 71 (-0.8)
  • Security 55 → 55 (+0.0)
  • Domain Modelling 100 → 100 (+0.0)
  • Accessibility 67 → 61 (-5.5)
  • Performance 100 → 100 (+0.0)

Resolved (32)

  • Change coupling clique: embedding.rb, image.rb, transcription.rb (lib/ruby_llm/embedding.rb)
  • Change coupling: chat_methods.rb ↔ message_methods.rb (lib/ruby_llm/active_record/chat_methods.rb)
  • Duplicated block (13 lines × 2) (lib/ruby_llm/ocr.rb)
  • Duplicated block (13–14 lines × 2) (lib/ruby_llm/image.rb)
  • Duplicated block (24 lines × 2) (lib/ruby_llm/providers/hetzner/chat.rb)
  • Duplicated block (5 lines × 2) (lib/ruby_llm/prompt.rb)
  • Duplicated block (5 lines × 2) (lib/ruby_llm/protocols/cohere/tools.rb)
  • Duplicated block (6 lines × 2) (lib/ruby_llm/protocols/anthropic/chat.rb)
  • Duplicated block (6 lines × 2) (lib/ruby_llm/protocols/chat_completions/chat.rb)
  • Duplicated block (8 lines × 2) (lib/ruby_llm/embedding.rb)
  • Duplicated block (9 lines × 2) (lib/ruby_llm/protocols/anthropic/files.rb)
  • Duplicated block (9 lines × 2) (lib/ruby_llm/protocols/invoke_model/nova_embeddings.rb)
  • Duplicated block (9 lines × 3) (lib/ruby_llm/file_store/tools/search.rb)
  • Flaky test: RubyLLM::Chat provider mapping lets provider options override the mapped value
  • High: security finding (details withheld)
  • High: security finding (details withheld)
  • High: security finding (details withheld)
  • High: security finding (details withheld)
  • High: security finding (details withheld)
  • High: security finding (details withheld)
  • …and 12 more

New (90)

  • Change coupling: chat_tool_call_state.rb ↔ run_claims.rb (lib/ruby_llm/active_record/chat_tool_call_state.rb)
  • Change coupling: embedding.rb ↔ image.rb (lib/ruby_llm/embedding.rb)
  • Change coupling: embedding.rb ↔ transcription.rb (lib/ruby_llm/embedding.rb)
  • Change coupling: image.rb ↔ transcription.rb (lib/ruby_llm/image.rb)
  • Change coupling: run_claims.rb ↔ tool_call_state.rb (lib/ruby_llm/active_record/run_claims.rb)
  • Change coupling: transcription.rb ↔ provider.rb (lib/ruby_llm/protocols/chat_completions/transcription.rb)
  • Change-coupling hub: speech.rb → embedding.rb, image.rb, provider.rb, transcription.rb (lib/ruby_llm/speech.rb)
  • Duplicated block (6 lines × 2) (lib/ruby_llm/file_store/tools/edit.rb)
  • Duplicated block (6 lines × 2) (lib/ruby_llm/protocols/chat_completions/chat.rb)
  • Duplicated block (8 lines × 2) (lib/ruby_llm/protocols/invoke_model/titan_multimodal_embeddings.rb)
  • Duplicated block (9 lines × 2) (lib/ruby_llm/file_store/tools/write.rb)
  • Edited copies of one member (3 members, 11+ corresponding lines) (lib/ruby_llm/protocols/anthropic/tools.rb)
  • Edited copies of one member (3 members, 12+ corresponding lines) (lib/ruby_llm/file_store/tools/edit.rb)
  • Edited copies of one member (3 members, 2+ corresponding lines) (lib/ruby_llm/protocols/anthropic/request_shapes.rb)
  • Edited copies of one member (3 members, 3+ corresponding lines) (lib/ruby_llm/protocols/chat_completions/request_shapes.rb)
  • Edited copies of one member (3 members, 9+ corresponding lines) (lib/ruby_llm/protocols/invoke_model/nova_embeddings.rb)
  • Edited copies of one member (5 members, 10+ corresponding lines) (lib/ruby_llm/protocols/anthropic/files.rb)
  • Edited copies of one member (5 members, 12+ corresponding lines) (lib/ruby_llm/providers/azure/media.rb)
  • Edited copies of one member (7 members, 4+ corresponding lines) (lib/ruby_llm/protocols/bedrock/async_videos.rb)
  • Edited copy of a member (11 corresponding lines) (lib/ruby_llm/protocols/chat_completions/models.rb)
  • …and 70 more

Changes since last survey

  • 300 commits — 295 feature/other, 5 fixes

By area

  • lib/ruby_llm — 137 commits
  • spec/ruby_llm — 40 commits
  • docs/_advanced — 20 commits
  • spec/fixtures — 20 commits
  • lib/generators — 18 commits
  • docs/_data — 11 commits
  • docs/_core_features — 10 commits
  • docs/_reference — 9 commits
  • (root) — 7 commits
  • .github/workflows — 5 commits
  • docs/_compare — 4 commits
  • docs/assets — 3 commits
  • docs/_getting_started — 2 commits
  • docs/_includes — 2 commits
  • skills/rubyllm — 2 commits
  • app/assets — 1 commit
  • benchmarks/streaming.rb — 1 commit
  • conformance/opentelemetry — 1 commit
  • docs/Gemfile — 1 commit
  • docs/_config.yml — 1 commit

Notable commits

  • fix: Fix Gemini and Interactions attachment and usage gaps (#1373)
  • fix: Fix Hetzner, Azure and xAI requests that cannot work (#1604)
  • fix: Fix agent rescue_from, chunk types and VideoJob.find (#1361)
  • fix: Fix review follow-ups in tool calls, thinking and docs (#1386)
  • fix: Fix streamed transcripts and Cohere thinking parity (#1375)
  • change: Accept ignored seam keywords through ** (#1393)
  • change: Add Agent callback macros and subagents (#1453)
  • change: Add Attachment#partial? and Message#spilled_path (#1550)
  • change: Add Azure AI Speech fast transcription (#1523)
  • change: Add Azure judgments through the Decisions API (#1437)
  • change: Add Bedrock Nova code interpreter alias (#1624)
  • change: Add DeepSeek web search alias and usage counts (#1623)
  • change: Add Deepgram chat on its Think endpoint (#1631)
  • change: Add Kimi files, documents, OCR, token counts, web search (#1613)
  • change: Add Message#web_sources for hosted web searches (#1478)
  • change: Add Meta transcription and Files support (#1607)
  • change: Add MiniMax images, video, speech, transcription and files (#1608)
  • change: Add Mistral OCR on Vertex AI (#1632)
  • change: Add Mistral document OCR on Azure (#1635)
  • change: Add Moonshot chat batches (#1636)
  • …and 280 more

Architecture

  • Unchanged — 0 containers · 1 contexts · 0 edges

Written by cca.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.

Survey your own repository

crmne/ruby_llm was measured the same way every project in this corpus was: the same rubric, at a pinned commit, with the result published in full. Point a surveyor at a repository you know and see whether you agree with it.

About this page

  • The score is its most recent published measurement, taken on 14 November 2026 at a pinned commit. It is not a live figure and does not change until the project is measured again.
  • Measured at commit e0a6948d518ecea9f6516bdd9d71c9580099f445 — the exact code this score is about.
  • Scored under rubric-2026.11.4 — the same rubric and the same method as every other entry in this index.
  • Measured by Canine Code Analyzer (cca.canine.dev) using cca measure-2026.11.65+preprod-e1a3b69fc8ee.