Repository navigation
Desktop UI audit: fix the broken screens, cut repeated text, one layout throughout - #1282
Conversation
Alpha showed live-looking controls with a 'Not available yet' footnote for General (Startup, Notifications, Updates), Models (Defaults, Providers), Connections > Agents, Privacy > Data, and rows on Access and Sessions. The catalogue is the one owner of what Settings lists, so those entries, their tabs and their empty categories are gone, along with the catalogue's pending flag and the group note it drove. About reads the version from package.json instead of '0.1.0 (prototype)', and no page header has a subtitle (Advanced's dek is gone). UI audit #4, #11. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01JcBpmVHPWxktA9QurvUMjM
The dialog put long segmented controls beside their names, so the controls ran past the dialog's edge and the names wrapped a word per line. Each setting now stacks its name over its control, the choices are short (This Mac / Boat / Any), and Auto-stop is the kit's segmented control instead of a native select. The two explanation sentences are gone. The stress channel step checks the dialog fits. UI audit #1. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01JcBpmVHPWxktA9QurvUMjM
…e only A turn resumed after a restart or a usage limit was shown with the resumption notice in front of the person's words, as if they had typed it. The SDK now strips its own notices from a successor turn's request (`resumed_request`), and the view marks that message `retried`, so the desktop draws the person's words with a small "Retried" mark under them. A session's list preview read a reply that opened with a table or a code block as raw pipes and fences. The preview now skips fenced code and table rows and takes the first prose line. Stress: the restart and usage-limit steps check the retried mark and that no notice text sits in the person's bubble; the table step checks the list preview is the prose line. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01JcBpmVHPWxktA9QurvUMjM
- Session rows are one line; only "Needs you" rows keep a summary. An
empty channel says "No sessions yet" with no button beside the
composer.
- The chat heading has no subtitle unless something is unusual (another
channel, another place, a lease that is not live).
- The permission card says "{agent} wants to run" over the command, with
Deny, Always allow and Allow once; the other answers sit under ⋯.
- Access and Native hooks say each thing once: no heading that repeats
the row, no footnotes, one hooks warning under the group, "None yet"
for an empty list.
- Voice key: the row says Connected / Not connected with one button; the
key is typed in a dialog. A call refused for want of a key offers
"Add key" beside the notice instead of sending the person to Settings.
Stress: the voice step saves the key through the call's Add key offer;
selectors follow the new button names.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JcBpmVHPWxktA9QurvUMjM
- Boat in Accounts is one row: Connected / Not connected / Locked and one
button; the token is pasted in a dialog Connect… opens, in the sans
font. Auto-stop is a row of its own beside it. Change passphrase and
Allow use the same row button as every other action.
- A voice call refused because Nessa is locked offers Unlock beside the
notice and then starts the call; one refused for want of a key offers
Add key (the key dialog).
- Devices: no line under the title, one Add device, and Boat says where
it stands once.
- The Agents peek says where a session stands in one line ("Waiting for
you", or the source's line of what it is doing), with no "Earlier in
this turn".
- nessa-client accepts a message's `retried` mark (true only); without
it every read of a resumed turn was refused.
Stress: the voice step goes through Unlock and Add key in place; the
devices step checks one Add, one Boat word and the page header; the
Accounts check finds one button per account, no fields, and Auto-stop in
its own row. settings-page and responsive follow the new copy.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JcBpmVHPWxktA9QurvUMjM
- Every top-level page (Settings, Devices and its pages, New channel, Agents) starts with the kit's PageHeader (nessa_ui#156, pinned here): one title size, actions on its line, no subtitle, the same distance below the titlebar (--desktop-page-top), in one centred width (--desktop-page-width). - Settings' sidebar is the app's flat sidebar material, not a floating card, and Settings groups are flat (kit `surface="flat"`): rows on the page with hairlines, their words lined up with the title. The bar's title and the column are measured against one container, so they stay aligned at narrow widths. - New channel is a flat layer with no dotted band, its field right under the question. - One brand name: Settings' way back reads "nessa Studio". - The Agents peek's permission card wears the chat card's fill and edge; sample answers use the real wording. Stress: Settings' sidebar is checked flat; Add a phone's rows flat; the New channel field sits under its question; Linked devices unlocks the vault Accounts asks for first. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01JcBpmVHPWxktA9QurvUMjM
|
Important
This repository does not receive automatic reviews because it has fewer than 10 stars.
|
The stress chaos check timed out on 'Pending execution contradicts its queued message': a successor turn still in the queue listed its pending preview with the model's resumption notice in front, while its message showed the person's words alone, and the client refuses the two disagreeing. Both now read one helper (person_text). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01JcBpmVHPWxktA9QurvUMjM
The card now names its answers itself (Deny, Always allow, Allow once) and keeps the rest under its menu, and an agent's card no longer repeats the agent's title. The scripted and stress checks still compared button words with the review's labels and found a card by its title, so they failed on a card that offers exactly what the review does. Each answer button now carries its option id (data-option), each menu trigger the ids behind it (data-options), and each card its review key (data-review). unofferedAnswers compares ids, the stress finds a review's card by its key, and scripted-scenarios checks the agent's card does not repeat the title. The app-review fixture lists its options' ids. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01JcBpmVHPWxktA9QurvUMjM
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01JcBpmVHPWxktA9QurvUMjM
Connections opens on Accounts, and with the vault locked its unlock dialog covers the tabs, so the passphrase step timed out clicking a tab that was already selected. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01JcBpmVHPWxktA9QurvUMjM
|
CI status on this PR, as of Failures this PR caused, now fixed
Failing on main too
Generated by Claude Code |
Hiding Connections › Agents left Accounts as the page Connections opens on. Accounts reads the vault, so with Nessa locked, choosing Connections (say, to reach MCP servers) asked for the passphrase before the person picked a page, and its dialog covered the tabs. The CI stress mcp step timed out on it. MCP servers is now first; the stress mcp step fails if opening Connections asks for the passphrase (it does with the old order). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01JcBpmVHPWxktA9QurvUMjM
… lined up - #22 The Agents peek replies with the composer in its compact size (`.desktop-composer[data-size="compact"]`): same surface, Send at its usual size, instead of a pill of its own. - #23 The overview's first group starts right under its header, the list says Loading (kit LoadingText) until its rows are drawn, and a row is its title plus one line: an agent's command, or an app's who-asks-for-what. - #24 (kit) A trace never cuts the verb, cuts a path from its start, and draws a failed call's words in their usual tone. - #25 (kit) A table's copy button sits above its top-right corner. - #26 The access menu's right edge meets the composer's (kit `alignTo`). - #27 Every pane's composer has the same gutter (`--workspace-dock-gutter`, the 8px narrow-pane override is gone), and a home beside other panes is titled "New session". - #28 Jump to Latest sits on the composer's top edge, half over it, not on the last line of the reply. nessa-ui-revision pins nessa_ui ef3187f (branch head of nessa_ui#156). Stress: menu checks the access menu edge, the overview gap, two-line rows and the compact reply; table checks the copy button and Jump to Latest; window checks one gutter and the named home beside; trace checks the verb, the cut side and the failed row's tone. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01JcBpmVHPWxktA9QurvUMjM
|
Two Windows server shards failed on
Generated by Claude Code |
nessa_ui#156 merged as 2b5bba0, which holds the same tree as the branch head this PR was checked against (the last commit there only added a registry dependency). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01JcBpmVHPWxktA9QurvUMjM
Keeps main's saved-logins switch in Devices and its login stress step beside this branch's page headers and voice step. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01JcBpmVHPWxktA9QurvUMjM
Codex Review Summary: ✅ CompletedThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 9bd3e7c247
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| onClick={answer(allowOnce)} | ||
| > | ||
| {option.label} | ||
| {answerNames.once} |
There was a problem hiding this comment.
Preserve persistent harness options' meaning in button labels
When a Claude or Codex review offers only a harness allow_always option, or lists it before allow_once, this button now says “Allow once” but sends the persistent option's ID. approvalOptions in gateway-views.ts deliberately maps harness allow_always options to choice: "once" because they do not create a Nessa project rule; that classification does not mean the harness forgets the answer. The SDK explicitly accepts persistent-only reviews (claude_and_codex_keep_harness_persistent_choices_and_can_add_project_rules), so this can cause someone intending to authorize one call to grant persistent approval instead. Preserve the provider's label for these options, or carry enough scope information to distinguish them before assigning standardized labels; apply the same correction to the compact rendering.
Useful? React with 👍 / 👎.
The title-only masthead shipped in 8fc0370 (#1282), but smoke still required a subtitle. Reuse the exact-title and extra-copy contract from fe7551a (#893) for #1327 qualification. The unchanged full browser check holds33/33, injected extra copy fails all4 Settings profiles, and restoration holds33/33.
The title-only masthead shipped in 8fc0370 (#1282), but smoke still required a subtitle. Reuse the exact-title and extra-copy contract from fe7551a (#893) for #1327 qualification. The unchanged full browser check holds33/33, injected extra copy fails all4 Settings profiles, and restoration holds33/33.
Requested by Saurav · project thread
This PR fixes UI audit findings #1–#28 from the desktop UI audit of 2026-11-09: every P0, P1 and P2 item. Q items #29–#31 are left for Saurav to decide. Each commit is one pass and can be reviewed alone. It builds on nessa_ui#156 (merged), which adds the kit's
PageHeader, a flatSettingsGroup, the ToolTrace, table and access-menu fixes, andComposerAccessMode.alignTo.nessa-ui-revisionpins that PR's squash commit,2b5bba0.Before
After
Background
Saurav asked for an audit of the desktop screens, because they had too much text, too many subtitles and layouts that did not match. The audit found 31 issues. This PR fixes them in order: things that are broken first, then too much text, then mismatched layouts.
The problem
The solution
cdc3e16a442679b9160fdaa3resumed_request), and the message is markedretried, so the person's words show with a small "Retried" mark. List previews skip code and tables.187f758b433fc7955796a27ePageHeader, at the same top inset (--desktop-page-top) and width (--desktop-page-width). Settings' sidebar and groups are flat. New channel is flat, with its field right under the question. One brand name ("nessa Studio"). The peek's permission card matches the chat's.959a1f22person_text), so it matches its message. Before this, the client refused the view as inconsistent and the chaos check timed out. The new projection testa_queued_retry_previews_the_persons_wordscovers it.fa70f011,691fbf90,dadd06dbdata-option,data-options) and the card by its review key (data-review), not by its words. The words changed in #7, and they will change again.0ef11da7mcpstep now fails if opening Connections asks for the passphrase.d370176d.desktop-composer[data-size="compact"]). The overview's first group sits right under the header, the list shows LoadingText until rows are drawn, and a row is its title plus one line (an agent's command, or an app's who-asks-for-what). The access menu passes the composer asalignTo. Every pane's dock uses--workspace-dock-gutter, and the 8 px narrow-pane override is gone. A home pane beside others is titled "New session". Jump to Latest sits centred on the composer's top edge. Trace and table fixes come with the kit pin.da4257bcnessa-ui-revisionto nessa_ui#156's squash commit2b5bba0.9bd3e7c2PageHeaderbeside main'sDeviceLoginsSwitch, and the stress doc keeps main'sloginstep beside this branch'svoicetext. The changed-file list against main is the same as before the merge.Page layout before and after:
flowchart LR subgraph Before S1[Settings: 28px title, 620px card column, floating sidebar card] A1[Agents: 20px title at x=284] D1[Devices: 20px title + subtitle, 1024px wide, two Add buttons] N1[New channel: 22px title + subtitle in an inset card with a dotted band] end subgraph After PH[kit PageHeader: one size, actions on its line, no subtitle] PH --> S2[Settings: flat sidebar, flat groups] PH --> A2[Agents: filter as the action] PH --> D2[Devices: Add device as the action] PH --> N2[New channel: Close as the action, field under the question] T[--desktop-page-top / --desktop-page-width] --> S2 & A2 & D2 & N2 endVoice refusal before and after:
flowchart LR R[Call refused] -->|before| M[Sentence pointing at Settings, no action] R -->|after: VoiceStartRefused.fix| K{fix} K -->|key| AK[Add key dialog, then dismiss] K -->|unlock| U[Unlock vault, then start the call]Resumed turn text before and after:
Composer and its menus, before and after #22, #26, #27 and #28:
flowchart TB subgraph Before P1[Agents peek: .agents-reply-pill, own surface + 26px Send] AM1[Access menu: align start at the shield] G1[Dock: 16px, 8px under 420px wide] J1[Jump: 12px above the scroller foot, on the text] end subgraph After C[.desktop-composer] --> P2["compact size: data-size=compact"] C --> AM2["ComposerAccessMode alignTo=form: end edges meet"] TK[--workspace-dock-gutter: 16px] --> G2[every pane's dock] J2[Jump: centred on the scroller foot = composer top edge] endnessa-client's strict validator now accepts
retried(true only). Without that change, every read of a resumed turn was refused.How we checked it
0ef11da7passed every check. CI ond370176dpassed Required checks after re-runs of two Windows Rust timing tests (vault cleanup and lease) that this PR does not touch. CI on the merge9bd3e7c2is running.tscandeslintare clean,node --checkpasses on the stress script, and the devices, settings and overview suites pass (457 tests).pnpmunit tests: 5706 passed on the earlier commits. Ford370176d, the overview, panes, transcript andsrc/desktop/uisuites pass (402 tests). The overview test now checks that an agent's row is its title and its command, and that an app's row has no command line.tsc,eslintandprettierare clean.person_textand passes with it.verification/desktop/scripts/conversation-stress.mjs, chromium):d370176d:menus=1 tables=1 traces=1 windows=1held 8/8. New checks:settings-pagepassed every check.responsivewas 15/15.shared-controlswas 6/8.segmentedtimed out, androwsfailed with the "kit row under its glyph" fill. Project notes record both failing on main in the cloud sandbox. I did not re-run them on main this session.Tradeoffs and follow-ups
Rollout
Desktop only. No data changes. nessa_ui#156 is merged and pinned, so this can merge once CI is green.
🤖 Generated with Claude Code
https://claude.ai/code/session_01JcBpmVHPWxktA9QurvUMjM