NEWS

inferencer 0.2.2

Added generic list_openai_compatible_models() and query_openai_compatible() wrappers for local and self-hosted OpenAI-compatible model servers. They support optional bearer authentication and pass through JSON-schema response formats.
Updated package scope documentation to include local model inference.

Added text-query and model-list wrappers for Qwen, Zhipu, DeepSeek, Moonshot, and MiniMax.
Added query_china_fallback() for ordered fallback across providers hosted in mainland China.
Added list_china_free_models() and query_china_free_fallback(), which currently use only Zhipu's documented permanently free API model instead of treating account-specific trials as universally free.
Added configurable per-request timeouts, defaulting to 120 seconds, to the Gemini, OpenRouter, Groq, and standard fallback query helpers.
Redacted API keys from compatible-provider error messages when a provider reflects credentials in its response body.

Added direct OpenAI Responses API support through query_openai() and model discovery through list_openai_models().
Added OpenRouter logprobs and top_logprobs request parameters.
Added Groq SSE response parsing for stream = TRUE; responses are buffered and assembled before return rather than emitted incrementally.
Consolidated OpenAI-compatible chat requests behind a shared internal transport while preserving provider-specific exported wrappers.
Added opt-in live provider smoke tests for the four new transport behaviors.
Improved errors for provider failures that return non-JSON response bodies.

Added list_openrouter_benchmarks() for direct refreshes from OpenRouter's unified benchmarks endpoint, including source-native scores, pricing, provenance, and source timestamps.

Updated the Cerebras public model set used in docs and tests.
Changed query_cerebras() to accept arbitrary non-empty model IDs to better support frequent Cerebras public model refreshes.
Kept list_cerebras_models() as the recommended way to discover supported Cerebras models.

Split the optional zsh shell layer into small executable scripts under inst/shell.
Added shared shell helper modules under inst/shell/lib.
Added shell model-listing scripts and --json support for query and listing commands.
Removed the monolithic inst/shell/inferencer.sh sourced-function bundle.

Added OpenRouter video-model discovery via list_openrouter_video_models().
Added benchmark extraction support via extract_openrouter_benchmarks().
Added category-specific OpenRouter model wrappers for embeddings, image, audio, and multimodal models.

Added an R-side query_fallback() wrapper with ordered fallback across Gemini, OpenRouter, and Groq.
Added tests and documentation for the new R fallback helper.

Improved the shipped shell markdown renderer for headings, bullets, horizontal rules, and pipe tables.
Changed shell query helpers to take prompt as the first argument and optional model override as the second.
Made shell query helpers render markdown automatically on successful text responses.
Updated query_fallback() to use the default models of query_gemini(), query_openrouter(), and query_groq().

Added a zsh-oriented shell companion under inst/shell/inferencer.sh.
Added shell wrappers for Gemini, Groq, OpenRouter, Cerebras, and Ollama Cloud.
Added query_fallback() for ordered shell failover across Gemini, OpenRouter, and Groq.
Added shell documentation and README examples for sourcing and using the shipped scripts.

Changed the default OpenRouter model to openrouter/free.
Increased the default OpenRouter token budget to reduce truncation.
Improved OpenRouter response parsing and truncation error handling for alias models.

Added Gemini and OpenRouter embedding helpers.
Added Gemini and OpenRouter image-generation helpers.
Added Gemini and OpenRouter multimodal query wrappers for generic non-text inputs.