test(app): benchmark populated chat Back and document UX review

This commit is contained in:
2026-10-24 11:24:47 +01:00
parent 877c2f1acd
commit ac75da12e4
6 changed files with 197 additions and 3 deletions
+55
View File
@@ -0,0 +1,55 @@
# UI/UX review: chat Back latency and save feedback
User requested another navigation/animation/feedback review, specifically chat
Back taking several seconds. No production chats/runs or personal devices used.
| Finding | Change/source | Focused verification | Landing | Delivery |
| --- | --- | --- | --- | --- |
| Unchanged answer parts reparsed/rebuilt during streaming | da44001a caches each Markdown render subtree until text or inherited appearance changes; link callbacks remain current | Deterministic reuse regression failed before fix; text/theme/link tests pass; selection element retained | Main | Pending |
| Usage-only updates rebuild transcript | 6d623197 separates transcript invalidation from usage/Info notifications | Both-host regression failed before fix; Info still updates live | Main | Pending |
| Open phone conversation becomes unread | 399d520c separates active reading identity from underlying list chrome | Both-host regression failed before fix; other chats still become unread | Main | Pending |
| Tabs are unmounted/revealed only after a detail finishes closing | 877c2f1a keeps the native destination mounted under its pushed route | Real Navigator regression; detail covers tabs, Back immediately reveals laid-out destination | Main | Pending |
| Server save closes editor before persistence succeeds | 25437b31 keeps draft/error/retry in editor, shows saving state, prevents duplicate sends | Agent: 88 tests, four pre-fix failures; 12 browser checks/6 captures across both hosts | Merged main | Pending |
## Measured Back behavior
`tool/chat_back_latency_probe.mjs` serves the normal built app against the
loopback protocol fixture. Android user-agent selects Flutter's actual Android
theme/transition. Chrome CPU throttling is 4x, viewport 390x844. Metrics measure
from before pointer dispatch to the first frame reporting the returned chat-list
route. Route completion/ownership semantics are unchanged between builds.
| Controlled scenario | Before route settled | After route settled | Before/after maximum frame gap |
| --- | ---: | ---: | ---: |
| 100 normal user/assistant pairs, quiet | 563 ms | 531 ms | 100 / 33 ms |
| 100 assistant messages folded into one turn, live text at 10 Hz | 1,153 ms | 563 ms | 400 / 117 ms |
| 199 assistant messages plus one user (one maximum history page), folded/live | 2,108 ms | 857 ms | 800 / 217 ms |
These are controlled browser samples, not device timings or statistical claims.
The largest case reproduced the reported multi-second delay. Initial open/settle
for that sample was about 4.05s for both builds; the change avoids repeating that
rendering work while streaming/closing rather than deferring it to opening.
- Before: installed Docker web bundle, source from prior release.
- After: `build/ux_review_web` from current source.
- Captures and raw JSON: `build/ux-review-before-android-{heavy,large-turn,max-page}/`
and matching `build/ux-review-after-android-*` directories.
- No explicit multi-second wait/network await was found in Back. Flutter's
normal Android toolbar/system-button fallback is 450ms, which cannot alone
explain the reproduced delay. Native transitions/predictive gestures retained.
- `PageRoute.completed` ownership guard retained to prevent re-push during exit.
- Optional `MAX_BACK_MS`/`MAX_FRAME_GAP_MS` probe budgets support repeatable lab
regression checks without imposing flaky wall-clock assertions on unit tests.
## Checks and limits
- 439 chat/shell/navigation tests passed, including predictive cancel/commit,
disclosure identity, paging, focus, text, drafts, follow/read position and Info.
- Full gate: clean analysis and **1,848 tests pass**, including added system
text-scaling coverage for cached Markdown.
- Settings evidence: `settings-server-editor-save-ux.md` and isolated
`tool/server_editor_save_probe.{dart,mjs}`. Root remount during successful
activation remains the root's existing behavior; no network wait added to Back.
- Native Android performance remains unverified: no ADB, device install or capture.
No transition-duration shortcuts, backend edits, workflow changes or production
mutation tests. Changes remain confined to app/.