[Tracking][UX] main/0.2.0 product QA follow-ups

[HaoxuanLiTHUAI]

Scope

Product/UX findings reproduced on current main 51b384c using Safari Computer Use, Playwright Chromium/WebKit, desktop and 390 × 844 mobile viewports, plus the isolated 208 QA deployment.

Issues

Composer and first run

• [UX] Combine provider, model, and reasoning selection in the composer #481 — combine provider, model, and reasoning selection; compact no-provider onboarding
• [Bug][UX] Reloading an unsent draft loses attachment chips but leaves staged files orphaned #482 — reload loses draft attachments while staged files remain
• [Bug][UX] @ mention picker omits persistent-library files despite advertising file mentions #483 — persistent files missing from @ mentions
• [Bug][UX] Agent/provider failures lack actionable recovery and retry affordances #485 — failed runs lack actionable retry/recovery

Interaction correctness

• [Bug][UX] ask_user option clicks submit immediately despite a separate Submit action #484 — ask_user option clicks submit unexpectedly
• [Bug][UX] Turn duration remains attached to a stale run after stop and resume #489 — stale turn duration after stop/resume

Information architecture

• [UX] Replace Knowledge Base setup internals with a guided PDF workflow #486 — replace Knowledge Base runtime internals with a guided PDF flow
• [UX] Use progressive disclosure for simple Agent and Trace sessions #488 — progressive disclosure for simple Agent/Trace sessions

Responsive and performance

• [Bug][UX] 390px regressions clip Settings navigation and Live Demo content #487 — 390px Settings and Live Demo clipping regressions
• [Performance][UX] Serve hashed SPA assets with compression and immutable caching #490 — compression/cache/code-splitting for SPA assets

Coordination

• An isolated fix is currently being prepared for [UX] Combine provider, model, and reasoning selection in the composer #481 only.
• Please comment on an issue before starting work so concurrent QA contributors do not duplicate implementation.
• Keep fixes in separate PRs by issue/scope; do not mix release-version changes into these UX PRs.
• Real-provider and causal-rollback validation will be reported separately after a runnable provider is confirmed.

Historical references

Some symptoms are regressions/follow-ups to closed issues #45, #97, #99, #108, #272, #316, #318, and #321.

[HaoxuanLiTHUAI]

Additional real-provider findings:

• [Bug][UX] Workspace restore is invisible in chat and leaves the visible narrative on the reverted version #492 — restore succeeds on disk but is invisible in Chat / leaves the visible narrative on v2
• [Bug][UX] Principal can claim Trace was recorded before the Trace Agent accepts or rejects it #493 — Principal can claim Trace was recorded before asynchronous Trace acceptance

Both were reproduced with qwen3.6-35b-a3b on the isolated 208 QA environment; checkpoint restore itself succeeded and returned the file to VERSION_ONE.

[HaoxuanLiTHUAI]

Coordination update: #481 is now implemented in draft PR #494. Other issues in this tracker remain unclaimed by this branch; please continue to claim them individually before implementation.

[HaoxuanLiTHUAI]

Implementation update: #482 is addressed by draft PR #495. The draft/attachment recovery path has passed production Web build, 25 targeted tests, the full 1063-test suite, and browser close/reopen + cleanup verification. Other contributors can treat #482 as claimed and review/test #495 instead of starting a parallel fix.

[HaoxuanLiTHUAI]

Implementation update: #483 is addressed by draft PR #496. Persistent-library mentions and the preview-to-composer action passed Web/full suites plus a real Safari upload and immediate-discovery flow. Other contributors can treat #483 as claimed and coordinate review on #496.

[HaoxuanLiTHUAI]

Implementation update: #484 is addressed by draft PR #497. Pointer, number-key, arrow, Enter, Submit, and free-text paths were verified against a real Runtime ask_user flow in Safari, in addition to full automated suites. Other contributors can treat #484 as claimed and review #497.

[HaoxuanLiTHUAI]

QA clarification for #481 / PR #494: the oversized no-provider panel seen during #497 testing came from that PR’s independent main-based branch. Safari verification on #494 confirms the state is a single content-width capsule above the composer. Keep PRs independent for review, but use a layered local integration build for subsequent product screenshots so already-fixed UX does not appear to regress.

[HaoxuanLiTHUAI]

Implementation update: #485 is addressed by draft PR #498. Real-provider 401/404/429/408/503 and recovery flows passed in Safari, alongside full automated suites. Terminal error duplication found during replay testing was fixed in the same scoped PR. Other contributors can treat #485 as claimed and review #498.

[HaoxuanLiTHUAI]

Implementation update: #486 is addressed by draft PR #499 and deployed to the 208 cloud Preview (fresh state, commit 2323268a). Direct PDF upload, guided status, actionable duplicate error, reload persistence, and container-restart persistence passed in Safari. Other contributors can treat #486 as claimed and review #499.

[HaoxuanLiTHUAI]

UX #487 is actively fixed in Draft PR #500, so please avoid duplicate implementation: #500

208 Preview + exact 390×844 WebKit verification are complete for Chinese/English layout, overflow, touch target geometry, and keyboard focus traversal. Safari Computer Use will be added after the locked test Mac is manually unlocked.

[HaoxuanLiTHUAI]

Additional Safari + real-provider QA on the 208 Preview found two non-duplicate follow-ups:

• [Bug][UX][Safari] Closing Settings loses focus instead of returning to the trigger #501 — Closing Settings loses focus instead of returning to the trigger.
• [Bug][UX] Real provider reply is rendered twice in Chat and Live Demo #502 — A single real provider reply is rendered twice in Chat and Live Demo.

Both are filed separately from Draft PR #500 so contributors can avoid duplicate work. I will address them as independent fixes.

[HaoxuanLiTHUAI]

#502 is now implemented in independent Draft PR #503: #503

Root cause is a false-positive Principal delegation guard classification for literal-output smoke-test text, not UI content duplication. Runtime suite and real 208 provider history pass; Safari UI pass remains pending due lock screen.

[HaoxuanLiTHUAI]

#501 is implemented in independent Draft PR #507: #507

The final 208 Safari pass covers both Sidebar Settings close paths and the provider-management entry point. Please avoid duplicate implementation.

[HaoxuanLiTHUAI]

#492 is implemented in independent Draft PR #508: #508

The deployed 208 real-provider rollback flow now shows one durable localized restore fact across Chat, Trace, and Files, including after reload. Please avoid duplicate implementation.

[HaoxuanLiTHUAI]

#493 is implemented in independent Draft PR #510: #510

record_trace now returns authoritative submitted/accepted/rejected state; only accepted permits a persistence claim. Real rejected and accepted paths are verified on 208. Please avoid duplicate implementation.