refactor(ui): unify streaming and settled answers onto a single render path (#642) - #647
Merged
Merged
Conversation
…642) The in-flight answer no longer renders in a separate `.maka-turn-streaming` section that gets unmounted and swapped for a committed `TurnView` at settle. Instead the streaming buffer is injected into the tail turn's own TurnView (the same `key={turnId}` node the committed turn settles into), so live→settled is a data-source swap on a stable DOM node, not a two-subtree handoff. - chat-view.tsx: `TurnView` takes an optional `liveStreaming` prop, passed only to the tail turn (`turns[last]` while streaming). The separate streaming section is deleted; a tiny fallback covers the rare "streaming before the optimistic user turn materialized" case. Per-frame work stays scoped to the tail turn — `materializeTurns` is still memoized on committed state only, so sibling TurnViews keep their identity and skip re-render. - Footer is now hidden by default and revealed on hover / focus-within of the answer block (both the assistant turn footer and the user-message copy), replacing the quiet-0.72 + settle fade-in. While a turn is live its footer slot is a reserved-height placeholder, never an actionable regenerate/branch (a live tail's derived status is `completed`). - maka-tokens.css: re-key the content-visibility override onto `.maka-turn[data-live-streaming="true"]`; drop the dead streaming box/margin rule and the `maka-footer-fade-in` keyframe. - session-events: textless / thinking-only completion now refreshes the committed message in BEFORE clearing the live buffer, so the shared node's answer block never unmounts before the committed turn lands. Tests: rewrite the footer-placeholder / marker-cascade / motion-allowlist contracts for the single-node shape; add behavioral coverage for single-node streaming, footer suppression while live, hover-gated settled footer, multi-step and tail-selection, and textless refresh-before-clear.
…642) The time under a user message was a self-refreshing relative label ("2小时前") that stayed always-visible. Per design review it should read as absolute wall-clock time and share the copy action's hover reveal. - chat-display-helpers: add `formatClockTime` (24-hour `HH:mm`). - chat-view: the user meta row now shows `HH:mm` (full date stays on the `title`) and the whole row — time + copy — is hover-gated on the user bubble (`group/usermsg`, opacity-0 → group-hover/focus-within), matching the assistant footer's hover reveal. Drops the now-unused `RelativeTime` import.
…ar label (#642) Three manual-test findings on the folded streaming/settled chrome: 1. DeepThinking live→settled height jump: the live body is a bounded `max-h-64 overflow-y-auto` scroll box, but the settled body was unbounded, so a long reasoning expanded to a full vertical strip and jumped taller the frame it settled. Bound the settled body to the same 256px scroll box so the live→settled swap is height-neutral. 2. Chevron pinned far-right: every folded trow (深度思考, tool group, tool row) pushed its chevron to the trigger's trailing edge via `ml-auto`. Drop `ml-auto` so the chevron (and the tool row's duration) sits right after the label text (gap-2), near the words instead of巨右. 3. Running-tool shimmer: verified the trow shimmer is correctly wired — a running/pending tool renders the same TextShimmer light-band the 深度思考 title uses, and the animation runs live in-build. Locked it in with a render regression test (running + pending shimmer, completed does not).
Astro-Han
changed the title
Two P2 regressions found by the #642 review (codex + pi + ChatGPT): P2-A — deferred textless clear could wipe a newer turn. The #642 refresh-before-clear made the textless / thinking-only completion clear the live buffer AFTER an async refresh resolves, but `clearStreaming` has no slot identity guard (unlike the text path's `clearSettledAssistantStreamSlot`). If the refresh resolved after the user started the next turn, the deferred clear blanked turn-2's answer + reasoning. Snapshot the turn's identity (messageId + thinking) at schedule time and clear only the parts still current, via a new `clearStreamingIfCurrent`. Adds a synced `thinkingBySessionRef` so the snapshot can read live thinking (which has no per-turn id in state). P2-B — a tool-only step showed an actionable footer on a still-running answer. `streamingActive` was derived from streamingText/thinkingText only, so a `tool_start` with no answer text yet fell through to the settled branch, whose derived status defaults to `completed` — rendering a clickable regenerate/branch on an in-flight answer. Extend liveness to in-flight tools (running/pending/waiting_permission); LiveStreamingEntries still no-ops on empty text so the tool-only tail shows just its timeline tool, no empty bubble. Regression tests: newer-turn slot + thinking survival and a positive control for P2-A; footer suppression while a tool runs + settled-turn control for P2-B. codex P2 (tail selection) was a false positive — pi verified the optimistic user-message commit keeps the tail == in-flight invariant. Touch a11y left as the user's deliberate hover-gating choice (focus-within fallback kept).
jackwener
approved these changes
Jul 8, 2026
jackwener
left a comment
jackwener
left a comment
Member
There was a problem hiding this comment.
Thorough review passed — this is careful work. Verified the core claim (the two-subtree seam is genuinely gone: live answer renders as trailing timeline entries inside the tail turn's own TurnView, keyed by the stable turnId, so settle is a data-source swap on one node) and the three risk layers: no re-animation on settle (committed Markdown renders without streamFade; the visibleMessages filter closes the double-render window and the slot-clear releases it in a single state update), keying stays stable mid-stream (memo'd TurnView + referentially-stable undefined for non-tail turns), and materializeTurns deps keep pure-text streaming out of the hot path. Contract hygiene is exemplary: every touched contract re-pinned in its new form, +8 behavioral streaming-handoff tests, and the removed streaming CSS/keyframes have zero dangling references.
Two conscious sign-offs recorded: (1) the user-message time is now hover/focus-gated (was always-visible for touch/AT) — accepting as the deliberate design shift from 3c79b37; full timestamp remains on the bubble title. (2) The chevron moves + hover-gated footer opacity are #642 detail-audit changes riding along in a refactor PR — fine, but noted so visual-baseline reviewers expect them.
Verified on today's main: desktop suite 2258/2258 green, CDP visual smoke of streaming-answer + turn-narrative fixtures clean.
7 tasks
jackwener
added a commit
that referenced
this pull request
Jul 8, 2026
…643 by @hqhq1025) — adopted with review fixes F1-F3+F5 (#656) Adopts PR #643 (unified Automation: heartbeat + cron, single tool) onto current main as a squash, preserving main's ToolCardBody dispatch and the #647 tool-activity refactors, with the maintainer's review fixes: - F1 (HIGH, automation-scheduler.ts): the busy-session defer budget is a ~45min wall-clock window (DEFER_WINDOW_MS, equivalent to the old wakeup-scheduler's 5s→5min exponential backoff), replacing the ~120s 24-retry cap. A transient busy window no longer terminally expires a `once` automation — skipFire only runs when the window is exhausted. - F2 (HIGH, automation-can-fire.ts): 'waiting_for_user' joins HEARTBEAT_IDLE_STATUSES (#639 decision — the wakeup's home scenario is starting a turn in place of the user); the desktop canfire test now pins the new set instead of the opposite. - F3 (MEDIUM, automation-state.ts): computeJitter ported verbatim from the old wakeup-scheduler and wired into computeNextFire — recurring (interval/cron) re-schedules get up to 10% delay jitter capped at 15min, one-shot fires landing on :00/:30 get up to 90s early jitter, computed on the actual fire timestamp. Mirrored the old jitter unit tests and added wiring coverage; AutomationManagerDeps gains an injectable `random`. - F5 (MEDIUM, packages/ui/tool-activity.tsx): AutomationResultPreview card (created/deleted/listed, localized) replaces the deleted CronJobResultPreview inside the shared ToolCardBody dispatch, plus a new desktop contract test binding the runtime tool's real output strings to the rendered card. - LOW: MAX_TERMINAL_KEPT 5 → 50 (old wakeup history cap) and the model-facing list now surfaces deferred fire attempts (deferredFireCount), mirroring the old CronList fire_attempts. - Fix: the CLI scheduler tick timer is unref()ed — the PR's always-on 5s tick kept the Node event loop alive, hanging any bootstrap consumer that exits without close() (this hung the CLI test suite indefinitely). Also allow-listed the PR's CLI persistence warnings in check-console. Original feature authored by @hqhq1025 in #643.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Closes #642
Why
A single assistant answer used to be rendered by two unrelated React subtrees: the settled
turns.map(TurnView)and a separate live<section className="maka-turn maka-turn-streaming">. The stream-to-settled handoff worked by unmounting one subtree and mounting the other, and that seam forced a whole set of hand-tuned special cases (streaming-shell margin, footer placeholder, fade-in, always-mounted status cluster, a 1000ms grace window). Worse, the textless / thinking-only completion path cleared the live region before the settled turn mounted, leaving a non-atomic flicker.The key opening: the user message commits synchronously the moment it is sent, so during streaming
materializeTurns(messages)already emits the tail turn (user bubble + correct turnId + an empty assistant timeline), and the runtime assigns the assistant rows the same turnId. Injecting the live content into that already-present tail turn turns the handoff into a data-source swap on one stablekey={turnId}node — no unmount.What changed
TurnViewgains an optionalliveStreamingprop; only theturn.turnId === tailTurnIdTurnView receives a fresh object per delta (so only it re-renders), every sibling gets a stableundefinedand skips via memo. The standalonemaka-turn-streamingsubtree and its alignment special-cases are deleted. The perf invariant holds: exactly one TurnView re-renders per frame, same as before.group-hover/answer,group-hover/usermsg, withfocus-withinpreserving a11y). Height is reserved throughout, so the handoff is height-neutral. The user-message time is now an absoluteHH:mm.content-visibility: visibleoverride moves from.maka-turn-streamingto.maka-turn[data-live-streaming="true"]; dead rules removed.drainAssistantStreamingnow refreshes committed history before clearing the live buffer (refresh-before-clear), holding the live answer until the committed message lands — no unmount gap.max-h-64scroll bound as the live<pre>, so an expanded long reasoning no longer jumps taller the frame it settles; the chevron on all three folded trows (deep-thinking, tool group, tool row) dropsml-autoso it sits next to the label text instead of the far edge.Deferred (split into #646, does not block this PR)
Real-time status feedback across the communication stages: a "connecting" state while the LLM connects, an "outputting" state during streaming, and a clear running-to-done seam for tools. The tool shimmer is correctly wired and verified to animate, but it flashes by on sub-second tools and lacks an explicit running-to-done transition.
Verification
npm run typecheckclean;@maka/ui46 +@maka/desktop2257 tests pass (including the new refactor(ui): unify streaming (live) and committed answer onto one render path #642 behavior tests and a tool-shimmer render regression test).maka-turn-streamingnode, footer and user time reveal only on hover. Issue 1 (a long settled reasoning is capped at a 256px scroll box, no height jump) and issue 2 (chevron sits 8px after the label, not far-right) both confirmed live.