[KKTTSPJ]

Summary

With --ui-management, the WebUI still assumes a server without a config. Most of the config-mode behaviour added for #234 exists only without UI management. This PR makes a configured model behave the same in both modes:

1. Configured models are offered with UI management too.
2. A configured model's memory can be released from the WebUI in both modes.
3. A model with built-in voices lists the same voices in both modes.

What stays as it is: without UI management, the WebUI offers only the configured models, with no catalog, downloads or model switching (the request in #234). POST /v1/models/unload stays management-only.

It pairs with #971, which adds the configured field to GET /v1/models and keeps a config model's settings when the WebUI loads it. Without that PR, item 1 adds nothing and items 2 and 3 still apply.

Background

I am sending this because, while working on #941, I kept running into places where UI management and config mode disagree for no clear reason. So I traced where the two modes split.

• WebUI - Initial native Svelte WebUI embedded inside audiocpp_server.exe — no Python, Node.js, or Gradio needed #185, the original design. It introduced the embedded WebUI as one UI for both setups:

  • Started with --ui and no config, it is a standalone host with model management.
  • With a config, "the same UI can front an existing server config".
  • --ui-management is added "when that instance should permit model switching".
  • The code was a single catalog-driven UI. The modes differed only in the management features (downloads, models folder, package sizes) and in whether loading was allowed.

• Confusion about web UI and existing configuration #234 and Track native UI configuration-mode gaps #236. Confusion about web UI and existing configuration #234 asked for a UI limited to the configured models when not managing. Track native UI configuration-mode gaps #236 answered it:

  • without UI management, the Studio lists the server's configured models;
  • the model button shows "Configured";
  • the voice list shows the server's voices.

  These were added as if (!server.ui_management) branches. The management side kept WebUI - Initial native Svelte WebUI embedded inside audiocpp_server.exe — no Python, Node.js, or Gradio needed #185's assumption of a server without a config. So a config combined with --ui-management never got the configured-model handling, although the server supports that combination and the README documents it.

• Later changes followed the same split.

  • Use shared eSpeak frontend for Kokoro English #515 made a model with built-in voices list only those voices with UI management. That fit then, because the management list held only the embedded demo recordings, which such a model cannot clone.
  • webui: list the server's voices with UI management enabled #941 (mine) changed that list to the server's voices but did not revisit the built-in-voice branch.

The result today, with a config and --ui-management:

• A config model under an id the catalog does not have is not offered at all.
• A config model under a catalog id is reloaded from the catalog. That reload drops its config settings; server: keep config model settings on WebUI loads, and offer demo voices with --ui #971 fixes this on the server side.
• A model with built-in voices hides the server's presets, for example a preset with a voice_id.

Without UI management, the WebUI cannot release a configured model's memory. Yet POST /v1/tasks/unload_models (#197) does exactly that and needs no UI management. Only POST /v1/models/unload, which is part of model switching, is management-only.

Changes

• Configured models with UI management.
  • The model list adds the config's models whose id is not in the catalog, using configured from GET /v1/models.
  • These entries show their id beside the name, because they often share a catalog entry's name.
  • A config model under a catalog id stays that catalog entry, so other formats of it can still be loaded.
• One notion of a configured entry.
  • isConfiguredEntry() is true for every entry without UI management, and for the entries added above with it.
  • Where the WebUI branched on ui_management for the selected model, it now branches on isConfiguredEntry(). That covers the model path, the installed check, load and ensure-loaded, settings import, the Arena and the LFM2 panel.
  • A configured entry therefore behaves the same in both modes: it loads on its first request and is never reloaded from the catalog.
• Release.
  • When a configured entry is loaded, its model button releases it through POST /v1/tasks/unload_models.
  • After a request loads such a model, the WebUI refreshes the model list, so the button shows the model as loaded. Before, it stayed "Configured" until the next refresh.
• Voice list for models with built-in voices.
  • Both modes now list the server's voices (presets and library names) and the built-in voices, leaving out the embedded demo recordings.
  • Before, the list held all server voices plus the built-in ones without UI management, and only the built-in ones with it.
  • Library WAVs still appear for such a model, as they already did without UI management.
• The management-only parts are unchanged: the Models tab, downloads, the models folder and package choices.
• dist/index.html is rebuilt.

Testing

• npm run check: 0 errors, 0 warnings. npm run build.
• Setup:
  • In a browser, against a CUDA build of upstream main (5b772fe0) with this PR and server: keep config model settings on WebUI loads, and offer demo voices with --ui #971 applied. "main" is the same build without the two PRs.
  • The config is the one our packaged Windows launcher writes: 10 models, some under catalog ids (irodori-tts) and some not (irodori-tts-anime), and a voice_dir with one WAV.
  • Added to it: my-irodori (not a catalog id, preset narr2), a preset narr on irodori-tts, and kokoro-test. kokoro-test is a kokoro_tts entry registered without its files, used only to look at its voice list.

With UI management:

Check	main	This PR
irodori-tts-anime, my-irodori and kokoro-test in the model list	not offered	offered, with the id beside the name
my-irodori: voice list	—	the library WAV and narr2
my-irodori: generate with narr2, then the model button	—	"Configured", then "Loaded" right after the request; a click releases it (loaded: false)
kokoro-test: voice list	—	heart (preset), the library WAV, then the built-in voices
irodori-tts (catalog entry): generate with narr, which makes the WebUI load it	—	works; narr is still listed afterwards

Without UI management:

Check	main	This PR
Model list, Models tab	configured models only, no Models tab	same
my-irodori after a request	"Configured" until a refresh, then "Loaded" and not clickable	"Loaded" right after the request; a click releases it (loaded: false)
kokoro-test: voice list	heart, the library WAV, the built-in voices	same

I did not run this PR without #971. Without it, configured is absent, so the code adds no entries for item 1.

🤖 Generated with Claude Code

[0xShug0]

Thanks, @KKTTSPJ PR Merged!