Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.
01 · Chat
Route #/chat — the landing screen · File src/screens/Chat.tsx (779 lines, the largest in the app)
Chat is where the product either holds up over a long session or does not. It is the only screen a user might sit in for hours, so it carries the strictest rules in the system: the reading measure, the follow-the-stream behaviour, focus mode, and the telemetry row.
What it is today
- Header — title (editable inline), a subtitle carrying model · message count ·
token count, and two controls: a generating pill that appears only while
working, and the context-rail toggle. The
···overflow was removed last round. - Transcript —
.chat__inner, 920 px max, centre-aligned. Assistant prose is capped at 56ch, measured at 68 characters at 10, 16 and 22 px. Code blocks and tables escape the cap and use the full bubble. - Tool cards — inline approval cards with Enter/Esc, a scope badge, a result line and a duration.
- Composer —
Do anythingplaceholder, keyboard hints as the box's second line, toolbar row, Send. Glass material at the regular tier. - Telemetry — three centred, clickable groups (turns/steps/tok/s, tokens/cache hit, context %) opening detail popovers.
- Context rail — a grid track that animates 0 → 340 px, with a per-conversation breakdown.
- Focus mode —
⌘⇧Fwithdraws the chrome to an ambient bar. - Scroll behaviour — follows the stream only when already at the bottom; per-conversation scroll memory; a scroll-driven reading-progress hairline.
Where it falls short
- The header subtitle is dense and low-value.
Starter · 0.5B Q4_K_M · 2 messages · 244 tokensduplicates what the telemetry row and the sidebar engine chip already say. Three places report the same state. - The header has no model identity while idle. The generating pill only exists mid-request, so at rest the bar says nothing about what is loaded.
Assistant messages have no per-message time or cost.Checked and wrong. A two-turn conversation renders a footer under both assistant messages with identical structure (236 tokens · 56.6 tok/s · 4.2 s), because thedoneframe writestokens,tpsanddurationMsonto every message and the footer renders whenever any of the three exists. The figures were always per-turn; only the actions below are missing.No way to retry or edit a turn.Fixed. Every assistant turn carries Ask again (retryLast) and every user turn carries Edit and ask again (rewindTo). Verified: retry drops the reply and replaces it without duplicating the turn; edit rewinds the transcript and refills the composer focused. Actions reveal on hover and on focus-within, so they are reachable by keyboard. Branch and Continue are now shipped — see below.- Reasoning is a collapsed block with no affordance. The disclosure animates open but there is no indication of how long the model thought.
- The intro state is four static suggestion cards. They do not reflect the workspace, and they do not change with the loaded model.
- Nothing marks unread turns when a long generation finishes while the user is scrolled up.
- The transcript has no
content-visibility. In a 200-turn conversation every message stays in layout, which is the one place the app can get slow. - Focus mode dims earlier turns to 45% but offers no way to move between turns — no next/previous, no jump-to-latest.
The plan
Hierarchy — one hero, less restatement
- Strip the subtitle to the conversation only.
2 messages · 244 tokensmoves entirely into the telemetry popover; the subtitle becomes the model name and nothing else, or disappears when the rail toggle is present. - Add a persistent model chip to the header, left of the controls: the loaded model with its state dot. This is the DSH shape and it means the bar is informative at rest, not only mid-request.
Substance — make the transcript auditable
Per-turn actions on the existing footer — retry, edit-and-resend, branch. The figures are already correct and per-turn; what the footer lacks is anything to do with a turn after reading it.
Retry and edit-and-resend.Done. Both live in the store rather than the screen, because each has to touch the message list — doing it from a component ends up duplicating the stream handling that already lives there.Branch and continue.Done. Branch from here forks a session at any turn, leaving the original intact — verified: the branch carries only the turns up to that point, the original is byte-identical afterwards, and both are listed. Continue from here appears on the newest reply only, since an older one already has its answer beneath it, and is sent as an ordinary user turn rather than a hidden mode, because a continuation is a request and the transcript should show it.Branching exposed a real bug.
Chatresolves its session aslocation.param ?? activeId, so the URL wins when it names a session. Every other path into a session navigates (newSession(); go('chat', id)); the first version ofbranchFromonly setactiveSession, so the sidebar showed the new branch while the transcript kept showing the original indefinitely — a feature that looked broken while its data was correct.branchFromnow navigates, and the flow asserts the URL follows.Reasoning blocks show their duration (
Thought for 4.2s) in the summary row, using the disclosure already built.content-visibility: autoon message blocks withcontain-intrinsic-sizeso off-screen turns cost nothing. This is the single largest performance win available and it is invisible when it works.
Fluidity
Jump-to-latest pill.Done — and it only appears once the newest message is genuinely off screen, so it does not sit over a short transcript with nothing to jump to. When scrolled up during or after a stream, a floating capsule appears above the composer with the count of new turns. Clicking scrolls smoothly; the pill fades on a token.Unread mark on the session row in the sidebar while the user is elsewhere.
Turn navigation in focus mode.Done. Focus mode is for reading a long conversation, and reading needs a next.⌥↑/⌥↓step between your own turns and⌘↓returns to the latest; the focus bar teaches the keys, since that bar is the only chrome the mode keeps. Verified: five distinct stops going down, each press retreating on the way back, and⌘↓landing at gap 0.The first version stuck. It scrolled a turn to
offsetTop − 16and then testedoffsetTop > scrollTop + 8— which the turn it had just moved to still satisfied, so the second press did nothing. The tolerance has to exceed the offset the turn is positioned at.
Substance — the intro
- Suggestion cards become workspace-aware — derived from the actual file list rather than a fixed array, and one of them becomes "Continue where you left off" when a previous session exists.
Motion
| Interaction | Spec |
|---|---|
| Turn entering | st-enter, --st-t-local, 6 px lift — already correct |
| Jump-to-latest appear | st-enter, --st-t-control, --st-from-y: 4px |
| Reasoning disclosure | interpolate-size on ::details-content, --st-t-local — built |
| Retry / branch | No transition; the transcript re-renders. Movement here would be noise |
| Rail open/close | grid track 0 ↔ 340 px, --st-t-surface — built |
Everything is dropped to opacity-only under reduced motion, per rule 6.
Verification
- The per-turn footer assertion is already satisfied and should be kept as a
regression guard rather than written as a new check: every assistant message
with a completed stream must render
.st-msg__meta. - New assertion: with 100 seeded turns, paint time on scroll stays within 1.5×
of a 10-turn conversation — the
content-visibilityproof. npm run measuremust still report 68 characters after any header or column change.- Focus-mode turn navigation added to the shortcut table and to
flow focus.