[luannn010]

See CONTRIBUTING before starting engineering work.

Change

The web Assistant becomes a conversation-first chat screen. The old page stacked a header, a scope field, an "Activity and preferences" block, a transcript with a status line and an execution-trace section on every message, and a composer with a counter and a provider section. This replaces that with one chat column, a saved-chats rail and a quieter record of what each run did.

All changes are in web/src/screens/assistant/:

• Layout. A saved-chats rail (ChatRail.tsx) sits beside the conversation, grouped as Pinned, by project and Recent. A header button hides and shows it, and the choice is remembered in localStorage. On narrow screens (≤760px) the rail starts hidden, opens over the chat and has its own close button. Search, rename, pin, save-as-note and delete stay in the existing ChatList dialog, now reached from "Search and manage".
• Header. Upcoming, New chat and a More menu. The menu holds Search and manage chats, Name and character, Previous saved history (when present), Background, Overnight, Review and Mute delivery notifications.
• Scope at the start of a chat. A new or empty chat shows a scope picker: Personal or a project. Projects with the assistant turned off can't be picked. Each choice opens a fresh chat through the existing ?fresh= / ?project= routes. Once the chat has a message, the scope shows as a locked tag and New chat is the way to choose again. Task-scoped chats opened from a task keep that scope.
• Suggestions follow the scope and send immediately through assistant.send(text), instead of copying the text into the composer.
• Messages. The person's messages are bubbles; answers sit beside an avatar. Copy, Reply/Answer and Details are icon buttons. Proposals, reminders, evidence, turn changes and "View work" render as before.
• Work trace (Activity.tsx). While a run is attached, a live row shows "Working · Ns" (or "Waiting for you") with each distinct progress.label as it arrives. A finished turn shows "Worked for Ns", using the span of its recorded trace[].at timestamps, which expands to the traceWords lines. Non-default replyWords states such as "Run stopped" or "Changes waiting for review" join that line. The trace uses smaller muted text so it never reads as the answer.
• Word-by-word reveal. Only an answer that arrives while the page is watching is revealed a few words at a time, with a caret. Long answers speed up to finish in about 5 seconds. History renders whole, and prefers-reduced-motion skips the effect.
• Composer. Stop replaces Send while a run is attached, with the same confirmation as before. The character count appears only past 3,500 characters. The provider section becomes one line of small print, and the "no AI connection" notice moves above the conversation.
• Removed: the per-message status line, the "Execution trace" block, the Project scope field, the always-on counter and the "Provider and scope" section.

Upcoming, most urgent first (a0d013bb, 26e9d863)

Upcoming.tsx and the new priority.ts. Upcoming was a dialog with separate Goals and Routines lists. It is still a pop-up window built on the shared Dialog, so it closes with Close or Escape, but its content is reorganised. (a0d013bb tried it as a side panel; 26e9d863 returns it to the pop-up window at the requester's choice.)

• One list, most urgent first (upcomingItems):
  • Need you now: saved chats with active: "needs_you", active goals whose target_date has passed in the planner timezone, and routines whose last_result.error is set. The worker and run code write that field when a scheduled run can't start, can't complete or is interrupted.
  • Today: goals due today, and unpaused routines whose next_run_at falls today in the routine's timezone.
  • This week: within 7 days.
  • Later: further out, or no target date.
  • Paused and done are folded away.
• Counts and filters. Counts at the top jump to each group, and filters narrow to goals, routines or waiting chats.
• One quick action per row:
  • Goal: Plan next step, which opens the existing GoalWork.
  • Routine: Pause/Resume, or Check history after a failure.
  • Waiting chat: Answer, which opens that chat.
  • Edit, check-ins, plan page, run history and delete keep their existing dialogs and confirmations, now in each row's menu.
• Permissions view. Saved permissions and admission controls move behind a "Permissions and admission" button, with "Back to Upcoming" to return.
• Permissions are choices (9d6449fd). Each change kind has a select, Ask each time or Always allow, saved through the existing setAssistantApprovalScopes. Previously the only control was Revoke on an existing grant. Both directions confirm first, and an existing goal- or routine-scoped grant shows as its own option. Checked in the browser: all seven kinds load as Ask each time, and choosing Always allow on Tasks opens "Allow tasks changes without asking?". Cancelling leaves it on Ask each time. No grant was actually saved.

Waiting questions and approvals, failed-answer restore, reply targets, selection replies, Draft project, Upcoming, saved history and identity keep their existing owners and behavior.

Checks and limits

Run on feat/assistant-chat-redesign at 2f46f409:

• npx tsc --noEmit in web/: exit 0.
• npx prettier --check src/screens/assistant/ in web/: passes.

Follow-up ad684440 (3 Nov 2026): Prettier and vite build pass; tsc --noEmit passes with the ES2023 array lib. The stock npm run build still stops at the pre-existing Array.findLast/ES2022 error in library/page/Source.tsx. No browser, live AI run, or screen-reader pass was run for the follow-up. The follow-up covers the failed-send replay, paint-before-reveal flash, step-only live announcements, retaining rail refresh, pressed-button scope group, filter-specific empty states, overdue routine bucket, --control target sizes, restored delivery notice, design-guide sync, and aria-haspopup on the new menus.

Rendered checks in the built-in Chromium browser against the local Vite dev server, with seeded demo data. These ran before the commit was rebased onto feat/web-rebuild, on a tree that also had #254's two fixes. The rebase touched only these files, and the typecheck was rerun after it.

• 1440×900: the rail renders beside the chat. Choosing a project sets ?project=…, marks it selected and changes the suggestions and placeholder. Opening a saved chat shows the locked scope tag, the bubbles and "Worked for 4s · Outcome not verified", which expands to its steps.
• 375×812: no horizontal overflow, and the rail starts hidden.

Upcoming, at a0d013bb: tsc and prettier pass again. upcomingItems was run with tsx on sample data (fixed "now" 2026-11-03 14:00 Melbourne) and grouped a waiting chat, a failed routine and an overdue goal under Need you now; a goal due today and a routine at 18:00 under Today; a Friday goal under This week; a December routine and an undated goal under Later; and paused and done items as quiet. In the browser at 1440×900 the side-panel version showed the demo data (2 goals under Later, 2 folded items), and the Routines filter worked. At 375×812 it filled the width with no horizontal overflow. After 26e9d863, at 1440×900: the header button opens the pop-up with the same content; the permissions view and Back work; and Close and a real Escape key both close it. The pop-up was not rechecked at phone width. The demo data has nothing urgent, so the Need you now and Today groups were not seen rendered, and the row actions and menus were not clicked in the browser.

Not verified:

• No live run. No message was sent, to avoid calling a configured AI provider. The live "Working" row, the progress steps arriving, the word-by-word reveal and Stop were not exercised. How often the steps update depends on how often the backend reports progress.
• The rail's narrow-screen close button. The page reloaded during that test, so the button was not seen working.
• Real Safari, the desktop webview, keyboard-only and screen-reader passes, and mobile (Expo) clients.
• No automated tests were added; the web package has no test script.

Workspace

• Used the outer Orbyn Engineering workspace with this target nested under repo/, and read the relevant project guides.

This contributor confirmation is not an automated claim that a host loaded policies.

Agent disclosure

• Tools: Claude Code (desktop app), built-in Chromium browser pane, a design canvas and inline mock previews for design review, git, gh, tsc, prettier.
• Model: Claude Opus 5.5 (claude-opus-5-5); effort as set by the session.
• Human role: @luannn010 set the requirements, reviewed several style mocks, chose and amended one (rail toggle, Upcoming, locked scope that drives suggestions, timed trace), approved implementing it and requested this PR.
• External sources: none.

🤖 Generated with Claude Code

Follow-up fixes (3 Nov 2026)

• Tools: Freebuff Desktop coding agent, terminal, git, gh, npm, tsc, Prettier, Vite.
• Model: selected by the host and not exposed to the session; exact identifier unavailable.
• Effort: host default; not separately set.
• Human role: maintainer requested the review fixes and the feat(web): tidy workspace menus, dock and sidebar controls #256 rebase; no human code edits.
• External sources: none.

[kinqsradio]

Reviewed #255 against feat/web-rebuild (cadf0db6). The chat-first redesign, the saved-chats rail and the Upcoming priority model are coherent, and the scope picker/locked-scope behavior is well thought through. I would request changes before merge: the inline comments cover the failed-send replay, the reveal timing, the rail refetch, the live-region and radiogroup accessibility issues, the filtered Upcoming empty state, and the routine priority edge case.

Two smaller regressions that do not have a single changed line to attach to:

• The old deliveryAvailable === false notice was removed. When conversation notifications are unsupported, the mute menu item simply disappears; a one-line “Conversation notifications are unavailable on this server” status would preserve the previous explanation.
• docs/design.md:42 says to center the empty Chat/Work composer and dock it only once conversation begins, and docs/design.md:117 says Assistant starter prompts wrap in two columns. This PR docks the composer always and uses a single flex row (horizontal scrolling on mobile) for the prompts. If the redesign is intentional, please update the guide; otherwise align the UI. Several new clickable rows are also 30–32px tall (assistant.css:420/682/791) instead of the 36px desktop action token.

Verification gap: the PR body says no live run, no Stop/reveal exercise, no narrow-screen rail close, and no keyboard/screen-reader pass. docs/verification.md asks for a scoped acceptance receipt for substantial UI changes covering failure/recovery, narrow layout and both themes. The failed-send replay below is in that unexercised path; please test a failed send and a successful run before merging.

Local checks: tsc --noEmit -p web/tsconfig.json, Prettier on the changed files and npm run build -w web pass. No CI checks are reported for this PR, and I did not run a browser or a live AI run.

[kinqsradio]

This infers “a new answer arrived” from assistant.sending going true → false, then takes the last display item. On a failed send, send() removes the optimistic user turn and sets sending false; the last item is then the previous assistant answer, so this calls setFresh(oldText) and re-runs the reveal on an old answer. The same can happen for a stopped run that adds no new turn. The PR's claim that only an answer arriving while the page watches is revealed is therefore not true on the failure path. Track the last assistant turn id before sending (or mark a fresh turn id in accept()) and only reveal when that id changes.

[kinqsradio]

role="status" is an implicit polite live region, and it wraps the elapsed timer that updates every second. The step list is aria-live="off", so screen-reader users hear repeated “Working · Ns” announcements but miss the progress labels. Move the timer outside the live region (or make it aria-live="off") and announce step changes in a polite region instead.

[kinqsradio]

This reset runs in useEffect, after paint. When fresh changes, the first render still has the previous shown value — often Number.MAX_SAFE_INTEGER after a null text — so done is true and the full answer paints once before the effect resets it to 0. Use a state keyed by the text, a useLayoutEffect, or derive the initial shown value synchronously to avoid the flash.

[kinqsradio]

useRead clears its data and sets loading whenever a dependency changes. Passing [active] means every chat switch blanks the rail and refetches up to 50 chats even though only the highlight changed. Remove active from the dependencies, or add a retaining reload mode so the rail stays visible while it refreshes.

[kinqsradio]

This declares a radiogroup, but the buttons have no arrow-key navigation or roving tabindex; every option is a tab stop. The checked task option also has no click handler and is not disabled, so it looks interactive but does nothing. Either implement the radio keyboard model (one tab stop plus arrow keys) or use aria-pressed buttons in a plain group; render the task scope as a non-interactive tag.

[kinqsradio]

items.every(...) is true for an empty array, so filtering to a category with no rows (for example “Waiting chats” when there are none) shows “Nothing is coming up. You're caught up.” even when other goals or routines still need attention. Gate the global caught-up message on filter === "all" (or on all.length === 0) and use a category-specific empty message otherwise.

[kinqsradio]

gap <= 0 means any next_run_at in the past is labelled “Today ”, even if the schedule is days stale and last_result.error is not set. That can show a long-past time as today's work. Consider treating gap < 0 as overdue/now, or only treating it as Today within a small scheduling-delay window.

[kinqsradio]

Review fixes pushed in ad684440:

• Failed sends no longer replay the previous answer; the reveal is keyed to the new turn id and starts before paint.
• The live region announces progress steps, not the elapsed timer.
• ChatRail keeps its list visible while refreshing instead of blanking on every active-chat change.
• The scope picker is a valid pressed-button group; the locked task scope is no longer a dead button.
• Upcoming shows a filter-specific empty state, and past-due routines land in Need you now.
• Clickable targets use --control; the delivery-unavailable notice is restored; the design guide now matches the Assistant layout.
• New menu triggers report aria-haspopup="menu".

The new Assistant menus also depend on #254 for WebKit anchoring. Checks: Prettier and vite build pass; tsc --noEmit passes with the pre-existing Array.findLast lib workaround. No live AI run or screen-reader/browser pass was run.

[kinqsradio]

lgtm