The Shiny workbench

The workbench uses the same R objects and numerical functions. Its synthetic example predicts continuous response with subject-grouped folds. Opening the app does not install runtimes or download weights.

state <- hm_app_state()
if (!isTRUE(all(identical(state$status, "empty")))) {
  cli::cli_abort("Example assertion failed.", class = "hm_example_error")
}
x <- hm_example()
if (!isTRUE(all(methods::validObject(x)))) {
  cli::cli_abort("Example assertion failed.", class = "hm_example_error")
}

hm_app(launch = TRUE, host = "127.0.0.1", port = NULL)

Use hm_app(launch = FALSE) to construct a shiny.appobj without starting a server. Import an RDS HMDataset/HMSequence or start with the synthetic example, inspect wavelength and grouping metadata, choose the supported task/model and training recipe, run, inspect the held-out metrics, and export the reproducible configuration/results. A changed input or configuration invalidates stale results. Keep subject IDs in split metadata; individual spectra are not independent biological units.

Uploads are limited to 50 MiB while the app runs; stopping it restores the previous Shiny upload limit. Use only trusted RDS files. Bounded fitting and runtime setup require Linux with taskset, prlimit and timeout. On other hosts the workbench shows this limitation before configuration and disables those actions. Import, inspection and configuration remain available; use the ordinary hm_fit() API for local R models.

hm_submit_job() starts an owned background job; hm_job_status() inspects it and hm_cancel_job() cancels that owned job. hm_export_result() writes a new export directory. Default exports omit individual-level data; private exports require an explicit request. Read persisted package runs with hm_read_run() and reproduce their task/workflow/split in an R script. The export becomes complete only after its reproduction script and optional private model files are written. A failed export keeps its reserved directory for inspection and cannot be read as complete; retry with a new directory.

UI/server tests and real browser workflows are separate evidence. Full keyboard, focus, both colour modes, narrow layout, cancellation/recovery, two-session isolation and app-to-script equality must be inspected in a real browser before claiming browser qualification. Optional app dependencies are Shiny, bslib, callr and processx. The shared website uses themakR; the app uses bslib directly.

The current workbench supports grouped regression/classification with R engines. Prepare ROI reductions through the API before import; an interactive annotation editor is outside this workbench. Runtime metadata resolution, terms review and installation are explicit user actions. Each new runtime resolution clears the terms checkbox. Review its displayed notices and check it again before installing; review of an earlier resolution does not authorize the new set of artifacts.