Skip to contents

shinygenui (development version)

Tool arguments

  • Tools now set convert = FALSE, so ellmer passes the model’s JSON arguments to genui_dispatch() unchanged. ellmer used to convert them first, and dispatch then rejected valid calls: an array of enums arrived as a factor and an array of objects as a data frame. In particular, a data_table call that named its columns always failed (#29).
  • Components now receive arguments in one documented form however the call was made: an array of strings, numbers, or other single values is an atomic vector, any other array is an unnamed list, and an object is a named list without NULL fields. See the “Argument values” section of ?genui_component (#29).
  • update_component now replaces each supplied argument as a whole. An array of objects can be updated, where before the update was rejected or, in a replay, ignored. A supplied object replaces the previous one instead of being merged with it field by field (#29).
  • A required argument that the model omits is now reported as missing instead of as an invalid NA. Errors quote a rejected value as JSON, so arrays, objects, and nulls remain recognizable to the model (#29).

Dependencies

  • Tested with ellmer 0.5.2. The ellmer requirement remains >= 0.4.1. Since ellmer 0.5.1, chat_openai_compatible() and most providers built on it no longer use strict tool schemas, so the model can omit required arguments or send values outside an enum. shinygenui validates every call itself and returns such mistakes to the model as tool errors (#29).

shinygenui 0.2.0

CRAN release: 2026-09-22

Chat integration

  • Require shinychat >= 0.5.0 and ellmer >= 0.4.1. Streaming, cancellation, and attachments are now delegated to shinychat::chat_server(), which relies on ellmer’s stream controller. Conversation history stays disabled until it can restore the canvas and registry together. The package no longer imports {promises} directly (#17, #21).
  • Greetings now use the shinychat welcome message API and disappear after the first user message. They remain display only and are not sent to the model (#21).
  • Use shinychat::tool_result_display() for tool results and present tense titles for running tools. The mtcars example uses the input toolbar to keep its prompt buttons above the composer (#21).

Examples

  • Example and acceptance test apps now read their model and reasoning effort from the required SHINYGENUI_MODEL and SHINYGENUI_EFFORT environment variables instead of relying on defaults (#9).
  • The 01-mtcars-explorer example now offers prompt buttons above the chat composer. A click fills and focuses the composer without submitting (#10).

Documentation

  • The README now includes a recording of the mtcars explorer example (#11).
  • Code examples now consistently use the same model provider while continuing to support any ellmer provider (#8).

Branding

  • Add a hex sticker logo with a p5.brush watercolor polygon background (#14, #15, #24).

shinygenui 0.1.0

CRAN release: 2026-09-09

New features

  • Use genui_component() to define a component and genui_catalog() to collect components into a catalog. Each component becomes an ellmer tool with validated arguments.
  • Use genui_canvas() and genui_server() to let the model add components to a running app. The included tools let it update or remove a component, clear the canvas, and inspect the current canvas.
  • Every call from the model passes through genui_dispatch() before it can change the canvas. If a call fails, the model receives the error and the Shiny session keeps running.
  • genui_components_bslib() provides a value box, Markdown card, data table, scatter plot, and histogram with a slider for the number of bins. genui_card_row() provides a container for arranging these components.
  • genui_trace() records each successful call. genui_replay() can rebuild the canvas from that record without an LLM.