posit-dev/py-shiny v1.7.0

shiny 1.7.0

on GitHub

New features

• Added test mode, enabled via the SHINY_TESTMODE=1 environment variable or the App(test_mode=) argument (which defaults to that env var). The pytest app-launch fixtures (local_app, create_app_fixture) now run apps in test mode. (#2269)

• When test mode is enabled, each session serves a JSON snapshot of its input, output, and export values at /session/{id}/dataobj/shinytest (URL available via session.get_test_snapshot_url()). App authors can surface internal reactive values with shiny.testmode.export_test_values(). The endpoint honors query params input/output/export to select specific blocks (=1 for a whole block, or a comma-separated list of keys); unlike R, requesting no blocks returns all three rather than a 400. (#2269)

• input.set_snapshot_preprocess(id, fn) (or shiny.testmode.snapshot_preprocess_input()) and my_output.snapshot_preprocess(fn) preprocess test-mode snapshot values before they are written, e.g. to scrub timestamps or temp paths. Handlers may be synchronous or asynchronous. File inputs automatically scrub each file's datapath to its basename, matching Shiny for R. export_test_values() moved from shiny.session to the new shiny.testmode module. (#2282)

• shiny.playwright.controller.AppTestValues reads a session's test-mode snapshot (input/output/export) in end-to-end tests. Its expect methods accept predicates: pass any callable (e.g. app_values.expect_input("n", is_integer)) in place of an expected value, and the expectation retries until the predicate returns a truthy value. (#2269, #2357)

• offcanvas() creates sliding Bootstrap Offcanvas panels that appear from a viewport edge. Panels can be triggered by a UI element, revealed programmatically with show_offcanvas(), or controlled by id with hide_offcanvas() and toggle_offcanvas(). (#2279)

• @render.download_button and @render.download_link pair 1:1 with ui.download_button() and ui.download_link(). (#2364)

• The shiny package now ships bundled Agent Skills under shiny/.agents/skills/ (the library-skills convention): a single shiny-for-python skill whose SKILL.md is a grouped index that routes coding agents to a per-topic reference file for each of shiny's public APIs. (#2339, #2344, #2345, #2356, #2365, #2366)

  • Foundations: reactivity (the reactive graph — reactive.value/calc/effect/event, req, isolate, poll), express (Express mode), modules-core / modules-express (reusable namespaced components), and session-lifecycle (on_ended/on_flush, request info, dynamic routes).
  • Layout and navigation: layouts (pages, cards, sidebars, columns, value boxes), navigation (navsets and page_navbar), dynamic-ui (@render.ui, update_*, insert_ui, panel_conditional), and theming (ui.Theme, presets, dark mode).
  • Outputs and rendering: plots (render.plot/render.image and interactions), data-frames (render.data_frame), and files (upload/download).
  • Feedback and interactivity: feedback (notifications, modals, progress, busy indicators), extended-tasks (non-blocking long-running work), and bookmarking (save/restore app state with bookmark_store= and session.bookmark hooks).
  • Generative AI: chat (ui.Chat) and markdown-streaming (ui.MarkdownStream).
  • Extending shiny: custom-renderers (authoring a Renderer subclass) and custom-components (custom JavaScript input/output bindings).
  • Testing and observability: testing (end-to-end tests with pytest and Playwright), debugging (inspect running apps via test mode), and otel (OpenTelemetry observability).

• shiny skills list shows each bundled skill's name and description, and shiny skills path <name> prints the skill's directory, so its SKILL.md and supporting files (references/, scripts/) can be read from the installed package. To install the skills into a coding agent, the README and the shiny skills CLI help now point at library-skills (uvx library-skills --claude). (#2340, #2367)

Improvements

• The shiny[otel] optional dependency group now includes opentelemetry-distro[otlp], so OpenTelemetry zero-code auto-instrumentation works out of the box: opentelemetry-instrument shiny run app.py. This is now the documented standard way to enable OpenTelemetry — the docs and the examples/open-telemetry/ example no longer configure providers inside the app (in-code set_tracer_provider() setup is silently ignored when a provider is already installed, e.g. under opentelemetry-instrument). The OTLP exporters are also included, making it easy to switch between gRPC and HTTP export via standard OTEL_* environment variables. Note that opentelemetry-distro pins opentelemetry-sdk to a matching minor version, so if you combine shiny[otel] with other packages that pin the OpenTelemetry SDK (e.g. logfire), the resolver may need matching versions. (#2349)

• The Playwright controllers DownloadButton and DownloadLink now share a common base class and gain expect_label(), and DownloadLink gains the width expectations previously available only on DownloadButton. (#2093)

• OutputDataFrame.set_filter() now supports multi-column filters. (#2093)

Deprecations

• @render.download is deprecated; use @render.download_button (or @render.download_link) instead. (#2364)

Bug fixes

• Fixed rare tabset ID collisions in pages with many navsets. Randomly generated data-tabsetid values were drawn from a pool of ~1 million, so a page with dozens of tabsets could occasionally produce two navsets with the same ID (a birthday collision), resulting in duplicate DOM ids and broken tab switching. IDs are now drawn from a pool of 9 trillion. (#2296)

• shiny run --launch-browser now opens the browser even when Uvicorn's INFO-level startup logs are unavailable, e.g. when logs are disabled or --log-level=warning is used. (#569)

• InputSlider.set() and InputSliderRange.set() now compute the target value's position from the slider's step configuration and drag the handle directly to it, verifying the landing position against the widget's state (finishing with arrow-key presses when the slider has more steps than the track has pixels). Previously the handle was swept one pixel at a time while polling the label, which was slow and could intermittently miss the target when the browser coalesced or dropped mouse-move events (a recurring webkit CI flake). InputSliderRange.set() orders the two handle moves so neither is clamped by its sibling, instead of parking both handles at their extremes first. Setting a value the slider cannot produce now raises an error listing the slider's actual values. (#2311, #2326)

• ui.output_data_frame() now consistently orders the filtered columns in ascending column order (#2093), and resetting a numeric range filter resets both values (#2093).

• value_box()'s id docstring now documents input.<id>_full_screen() for observing the value box's full screen state, matching card(). It previously documented the wrong reactive-value syntax, input.<id>()["full_screen"]. (#2324)

Other changes

• CI now runs the unit tests on Ubuntu with the oldest supported Python (3.10) and every runtime dependency resolved to its declared minimum version (via uv pip compile --resolution lowest-direct), so stale lower bounds in pyproject.toml are caught. As part of this, starlette and prompt-toolkit gained explicit lower bounds (>=0.17.1 and >=3.0.0), and the opentelemetry-api/opentelemetry-sdk minimums were raised from 1.20.0 to 1.24.0 — older versions could leak unsanitized error messages in span exception stack traces. Shiny's OpenTelemetry log emission (which relied on the Logger.emit() keyword form added in opentelemetry 1.38.0 and was silently dropped on older versions) now falls back to constructing the LogRecord manually, so it works across the whole supported range. (#2335)

• Optimized the test suite by avoiding network-based dataset downloads and heavy library imports during Playwright test app startup, reducing setup time significantly. (#2314)

• Packaging metadata now uses a PEP 639 SPDX license expression instead of the deprecated license table and license classifier, and stale MANIFEST.in rules were removed. Building the package (e.g. with uv build) no longer emits setuptools deprecation or manifest warnings. Wheel and sdist contents are unchanged. (#2304)

• Raised the minimum supported uvicorn version from 0.16.0 to 0.23.0 (July 2023). The old floor no longer worked in practice: running on Posit Workbench passes uvicorn the ws_per_message_deflate option, which requires uvicorn >= 0.17. (#2317)

• @add_example() gained an example_name= parameter that looks up the example in the nearest api-examples/ directory, and all ex_dir= call sites that pointed inside an api-examples/ tree were migrated to it; ex_dir= remains only for examples outside the nearest api-examples/ tree. This also fixed nine call sites that passed the example name positionally, where it was silently treated as app_file= and the example was missing from the generated docs (#2328). Docs builds now fail when an @add_example() API reference is missing its example app, instead of emitting a warning and silently omitting the example; several APIs whose examples had gone missing this way were fixed, and missing Shiny Express examples were added. (Thanks, @EltonChang1!) (#2278)

• Added api-examples for shiny.testmode.export_test_values(), shiny.testmode.snapshot_preprocess_input(), and Renderer.snapshot_preprocess(), demonstrating how to surface reactive values in the test-mode snapshot and how to scrub sensitive or nondeterministic input/output values from it (#2284), and for ui.output_code() (#2093).

• shiny.run.run_shiny_app() (and therefore shiny.pytest.create_app_fixture()) now picks random app ports from a disjoint per-worker port range when running under pytest-xdist, instead of the full 1024-49151 range shared by all workers. This prevents parallel test workers from racing to bind the same port and from reusing each other's recycled ports. When not running under pytest-xdist, the behavior is unchanged: ports are picked from random_port()'s full default range (1024-49151). (#2297)

Full Changelog: v1.6.4...v1.7.0