Much faster agent profile switch and app startup

[dobby-d-elf]

Summary

• Performance-only improvements to make app loading and agent profile switching much faster
• Apply selected agent profile defaults immediately, refresh the visible/current screen before clearing the switch animation, and schedule non-visible workspace/model metadata in the background after the current view settles.
• Keep profile catalog refreshes profile-generation-aware so stale background responses cannot overwrite the selected agent profile model/provider.
• Start first-load session list rendering before unrelated workspace/onboarding metadata settles, with a smooth row-level flip animation for first load and profile changes.

No new or changed functionality is intended. This is scoped to perceived performance and smoothness only.

Measurements

• 25x faster agent profile switch -> 10s reduced to about 0.3s. The 0.3s figure is from a moderate-size install, and users with show_cli_sessions=true plus very large state.db files may still see longer waits due to the underlying bug(cli-sessions): show_cli_sessions can make /api/sessions hit the 30s WebUI timeout on large state.db #2628 API latency.
• Significantly faster session rows & chat - measured around ~0.3s to load session list on cold boot vs a few seconds before

Review follow-up

• Boot synthetic model options now carry data-provider, and /api/settings exposes the default model provider so early sends before model-catalog hydration keep the provider.
• Existing exact default-model options also get provider metadata, avoiding fuzzy fallback from gpt-5.4 to gpt-5.4-mini while the catalog is still hydrating.
• The model catalog starts loading in the background when the agent profile is switched. This is one of the slowest operations in switching a profile which means the model-picking-menu will not be available for the first 1-5 seconds after an agent profile is switched. If a user clicks model picker menu while it is still loading, it will open when available. Picker open is capped so it cannot appear hung behind a slow catalog refresh. Mentioning this as an edge case/limitation - most users will not encounter it but if someone quickly switches agent profile AND then tries to switch the default model via the model picker, the menu can take a few extra seconds.
• Workspace tree loading now has an inline comment explaining why hidden panels are fire-and-forget while visible panels are awaited.
• Normal UI profile switching closes the profile picker before selecting a profile; stale background catalog responses are ignored if a profile switch happened while they were in flight.

Tests

• Automated and manual verification performed

[nesquena-hermes]

Summary

Reading PR head e89ef6ce against origin/master end-to-end — boot.js (39 lines), panels.js (87), sessions.js (69), api/profiles.py (3), plus the test edits — this is a well-instrumented perf PR. The framing "no new functionality, only perceived perf" matches what I see in the diff. The measurements in the body (10s → 0.3s on default→pepper) are plausible given that the master path awaits Promise.all([populateModelDropdown(), loadWorkspaceList()]) before the switch animation clears, and both of those hit network on every switch.

The shape of the change is sound: apply profile defaults from /api/profile/switch response immediately, kick off catalog refreshes in the background guarded by a generation counter, and lazy-hydrate the model catalog only when the picker opens.

Code reference

The new lazy-hydration trigger at static/ui.js:1453:

if(typeof window._ensureModelDropdownReady==='function') window._ensureModelDropdownReady();
const ready=window._modelDropdownReady;
if(ready&&typeof ready.then==='function'){
  try{await ready;}catch(_){}

Paired with the boot factory at static/boot.js:1547-1552:

const _startBootModelDropdown=()=>{
  const ready=window._modelDropdownReady;
  if(ready&&typeof ready.then==='function') return ready;
  const next=_hydrateBootModelDropdown();
  window._modelDropdownReady=next;
  return next;
};

And the generation guard in switchToProfile at static/panels.js:4769:

const _switchGen = ++_profileSwitchGeneration;
...
if (_switchGen !== _profileSwitchGeneration) return;

This is the correct fix for the stale-promise hazard that opens up the moment you stop awaiting in-line.

Diagnosis

1. Backend payload addition. api/profiles.py:917,974 adds default_model_provider to the /api/profile/switch response. That's exactly the right minimal backend change — the client can now apply both model and provider without an extra round-trip to /api/models. The provider comes from model_cfg.get('provider') on the per-profile config.yaml, which is the same source as the agent's resolver, so the two should stay in sync.

2. Workspace duplicate-write elimination. The sessionInProgress branch in master makes two round-trips: await newSession(false) (which POSTs /api/session/new with inheritWs), then await api('/api/session/update', ...) to set the workspace again. The PR removes the second call because S._profileSwitchWorkspace is set at panels.js:4807 before newSession() runs, and newSession() already consumes that flag at sessions.js:460:

const switchWs=S._profileSwitchWorkspace;
S._profileSwitchWorkspace=null;
const inheritWs=switchWs||(S.session?S.session.workspace:null)||(S._profileDefaultWorkspace||null);

So one fewer POST per switch, with the same observable result. The updated test_sprint40_ui_polish.py correctly inverts the old assertion (assertNotIn('/api/session/update', block)).

3. FLIP animation correctness. _captureSessionListFlipPositions() at sessions.js:1926 is called BEFORE list.innerHTML='' (which clears the DOM) — verified at sessions.js:2876-2880. The captured Map<sid, top> is then used to compute delta after re-render. Standard FLIP technique, implemented correctly. prefers-reduced-motion is honored at sessions.js:1944 and style.css:746-749. Good.

4. Generation counter. _profileSwitchGeneration properly handles the case where a user clicks Profile A, then quickly clicks Profile B — the older promise's continuations all short-circuit via if (_switchGen !== _profileSwitchGeneration) return;. There are five such early-returns in switchToProfile, plus the catch and finally arms gate their visible effects on the same comparison. This is the right pattern.

Concerns

1. Boot-time synthetic option lacks data-provider. At static/boot.js:1428-1437:

const opt=document.createElement('option');
opt.value=s.default_model;
opt.textContent=typeof getModelLabel==='function'?getModelLabel(s.default_model):s.default_model;
opt.dataset.custom='1';
sel.querySelectorAll('option[data-custom]').forEach(o=>o.remove());
sel.appendChild(opt);

No opt.dataset.provider. If the first /api/chat/start POST happens before populateModelDropdown() resolves (because the user types and hits send within ~50ms of boot), the model will be sent without a provider, and the agent has to fall back to the legacy provider resolver. Compare with the equivalent code in switchToProfile at panels.js:4801-4805 which does set opt.dataset.provider. Suggest adding:

opt.dataset.provider = window._activeProvider || '';

right before sel.appendChild(opt) in the boot branch.

2. _refreshProfileSwitchBackground clobbers window._modelDropdownReady to null at panels.js:4502. If the picker is already open at the moment the user triggers a profile switch (e.g. they had it open to verify the current model), the dropdown's current _modelData reflects the old profile until the user closes and reopens. That might be fine — the dropdown re-queries the dropdown options via populateModelDropdown on next open via _ensureModelDropdownReady — but worth confirming the open-dropdown UX manually. The model chip on the topbar gets the new default immediately, so this is probably only a visible inconsistency in the (rare) open-dropdown-during-switch case.

3. The awaitWorkspaceLoad option on newSession() is a clean way to make the workspace panel render synchronously when visible and async when hidden. The implementation at sessions.js:521-523 returns the promise from loadDir('.') and conditionally awaits. Worth a brief comment near that gate explaining why the asymmetry exists, otherwise the next reader will assume it's a missing await. Something like:

// loadDir('.') is fire-and-forget when the workspace panel is closed —
// awaiting it would block the new-session flow for users who never look
// at the file tree. When the panel is visible we await so the file
// list renders synchronously with the session swap.

4. Test surface. The new tests are entirely source-text assertions (look for specific JS strings). That's fine for catching accidental reverts but won't catch a refactor that preserves the strings but breaks the behavior. The FLIP test in particular (test_profile_refresh_captures_row_positions) asserts getBoundingClientRect().top is present in the file — true even if you delete the call site and stash it in a comment somewhere. Not blocking, but a small puppeteer-style integration test that loads the page and asserts the session list animates on profile change would be more robust. Skip if puppeteer isn't already in CI.

5. Measurements need a baseline anchor. The body says "10s → 0.3s" but doesn't note install size. The 10s case is presumably consistent with #2628's diagnosis (get_cli_sessions scanning a large state.db) — if the user has show_cli_sessions=true and a 2.7GB state.db, this PR will help perceived switch speed but not the underlying API latency. Worth mentioning in the PR body that the 0.3s figure is on a moderate-size install and that show_cli_sessions=true users may still see longer waits due to #2628.

Recommendation

Approve with the four concerns above noted as follow-ups (none blocking). The generation guard pattern is exactly right for an "apply optimistically, refresh in background" perf path, the FLIP animation is implemented properly with reduced-motion honored, and the duplicate workspace-update elimination is a clean cleanup. Add opt.dataset.provider on the boot synthetic option (concern 1) before merge if convenient — that's the only one with a behavioral risk.

[dobby-d-elf]

moving to draft while I iron out a few issues

[dobby-d-elf]

1. Boot-time synthetic model options now carry provider metadata via opt.dataset.provider = window._activeProvider || '', so an early /api/chat/start before catalog hydration still sends the provider instead of falling back to legacy resolution.
2. For profile switching and the model picker, we now invalidate and immediately warm the model catalog in the background after a profile switch. The picker itself stays conservative: if the user clicks before the catalog finishes, it still waits rather than showing stale options.
3. Added the loadDir('.') comment explaining why hidden workspace trees are fire-and-forget while visible workspace panels are awaited during session/profile swaps.
4. Tightened the FLIP test coverage without adding Puppeteer/Playwright, since browser automation is not already in CI. The test now verifies capture-before-render and play-after-render ordering instead of only checking for broad source strings.
5. Updated the PR body measurement caveat: the 0.3s figure is from a moderate-size install, and users with show_cli_sessions=true plus very large state.db files may still see longer waits due to the underlying bug(cli-sessions): show_cli_sessions can make /api/sessions hit the 30s WebUI timeout on large state.db #2628 API latency.

[dobby-d-elf]

@nesquena-hermes this is ready to review again. Tested and addressed review comments