mirror of
https://github.com/getpaseo/paseo.git
synced 2026-11-04 23:44:41 +00:00
Mohamed Boudra
4f7813d3aa
* Keep a disclosure's trigger in place on web Expanding or collapsing a turn summary, tool-call group, tool call details or todo badge now holds the row that owns the trigger. Rows inside the timeline get a disclosure context; each toggle runs its state change through `useDisclosureHold()`, and the viewport's new `holdRow(rowId, apply)` replaces `holdReadingPosition()`. On web the hold pins the reading anchor to that row, stops following output, and resumes following when the change leaves the list at its end. Native keeps its previous behaviour (detach, then apply) until the anchored toggle lands. * Render the native chat timeline with LegendList in forward order The native timeline was an inverted FlatList. Disclosure toggles moved their trigger, and a cold open settled visibly. It is now a forward-order LegendList that stays virtualized, and one owner moves the scroll offset at a time: - Before placement: LegendList's initial scroll, behind its hidden reveal. - While following: the end pin (native-end-pin.tsx). This is a zero-height first child at the bottom of the content, which native maintainVisibleContentPosition keeps in place. A content shrink or growth never scrolls from JS; only a viewport resize or an explicit jump does. - During a disclosure hold and while detached: LegendList MVCP with shouldRestorePosition (native-disclosure-hold.ts), so a toggle keeps its trigger at the same y and paints one layout change. - Every newly mounted row starts at a per-kind estimate (new-row-estimates.ts). PaseoScrollView defers React Native's content clamp while MVCP is set. A shrink below an unmoved anchor no longer draws a frame past the end (RN #58578, PR #58695). VisibleContentPositionClampTest covers this. LegendList 3.7.1 patch, one "Paseo patch" comment per hunk: - The content height comes from React state on the new architecture. - Rows get a per-row size estimate. - The MVCP anchor lock runs on native too. Upstream gated it to web in d24a9d93 without a native reason. Without the lock, the size pass for prepended history re-anchors on a prepended row and moves the rows on screen. Native ends the lock when its row leaves the view, not on the 300 ms clock. - Header and footer reveal together with the rows. - Only lists that keep their end preserve the initial end. - A remeasure within one pixel plus the 1/8 dp floor of the known size is noise. Upstream compares the raw size with the floored one, so the live turn's assistant row (1380 -> 1381 px) read as a resize and moved every row a frame after the reveal. - The Android silent initial scroll retry sends its 1 dp nudge and the return to the target in one task. Upstream returns on the next frame, but its 16 ms check can finish the scroll first and cancel that frame, so the list revealed 3 px short of its end and moved into place a frame later. - The top-padding shrink hold applies on iOS only. Android, emulator. Evidence is in .artifacts/legend-impl. - J1, J2 at the end, J3, J4, J5, J6, J7: pass. - J2 while streaming: not reachable on the Ten second stream fixture; its turns never fold. - J8: no blank frames over 10 fresh flings, at host load 32-37. J4: still. Before these two hunks, 6 of 7 cold opens moved every row one frame after the reveal. After them, 9 valid cold opens show no shift. 5 of those were traced through the nudge path. Runs that hit the dev client's startup SIGSEGV in MountingCoordinator::pullTransaction were left out. That crash is in the emulator's dropbox from before this work. * Revise a history row when its turn summary line changes LegendList renders a row again only when its item changes. The forward-order timeline in d1f2ae352 folds a finished turn into its host user message, but the host's item does not change when the fold appears, so a turn that folded live lost its summary line until something else revised the row. collapseTurns now reports the hosts whose summary line appeared, changed or went away since the previous fold (updatedHostIds), and the history row revision treats them as changed content. * Stop following the end while a user scroll owns the viewport The bottom anchor controller kept sticky-bottom for the whole of a drag and its momentum and only decided at the end. While it is sticky the native list mounts the end pin, so a drag started during a stream had every new line move the rows under the finger: on the iOS simulator the offset jumped by each line's height (+22, +35, +22 pt) mid-drag, and Android does the same with the same code. beginUserScroll now detaches. endUserScroll resumes following when the gesture ends at the end (scrolling there without animation) and otherwise stays detached and drops any pending anchor, as before. The re-stick is not animated: on Android a line that mounts during a smooth scroll cancels it, and the list then followed 40 pt short of the end until the controller detached it. Measured after the change, frame by frame: - iOS and Android, drag up mid-stream: the content tracks the finger and momentum with no line steps; the stream grows below; tapping the pill re-sticks and follows again. - A short drag at the end, mid-stream: following resumes; the release lands on the end in one frame (the drag distance, here 8 pt on iOS and 21 dp on Android). A line already mounting when the drag begins can still move the rows once in its first frame, because the detach reaches native one commit after the touch. * Correct the iOS scroll view under maintainVisibleContentPosition Phase 2 of the LegendList timeline. The native timeline relies on maintainVisibleContentPosition for the end pin and the disclosure hold, and React Native 0.86 on iOS gets two cases wrong. modules/paseo-scroll now registers PaseoScrollViewComponentView, a subclass of RCTScrollViewComponentView, for the iOS scroll view, like PaseoScrollView on Android. - A mount that removes its own anchor: detaching removes the end pin in the same mount that opens the tapped disclosure, and the offset jumped by the removed view's frame. The subclass keeps the offset when the anchor's tag is no longer registered, the check React Native 0.88 adds in https://github.com/react/react-native/pull/57294. That PR's extra superview check is not copied because it skips anchors detached by removeClippedSubviews (https://github.com/react/react-native/issues/58910). - A shrink at the end: UIKit clamps the offset to the new end during the mount, then React Native adds the anchor's movement, so a turn folding while following landed above the end (https://github.com/react/react-native/issues/58578). The subclass restores the pre-mount offset for the adjustment and clamps after it, as https://github.com/react/react-native/pull/58695 does. The dev client was rebuilt with the module. Verified on the iPhone 16 Pro simulator, frame by frame: the summary far up and the latest turn's summary while following each keep the trigger at 0 px with one layout change per toggle, the todo card the same, the cold open lands still, older history pages keep the visible rows in place, and the keyboard, send-time shrink and streaming stay pinned to the end. Freshly mounted markdown on iOS paints one frame after its layout, because enriched-markdown renders asynchronously (https://github.com/software-mansion/enriched-markdown/issues/1131). A synchronous first render removed that frame but raised fling frame time at p90 and p99, so it is not part of this change. docs/agent-stream-disclosure.md now describes the clamp correction on both platforms. * Keep the rows under the finger at a drag's start and its release Two one-frame row movements remained around a user scroll while the timeline streams. A drag's first frames. The end pin stays mounted until JS stops following, one round trip after the drag begins, so a line mounting in those frames scrolled the rows under the finger by its height. Both scroll view subclasses now skip the content-position correction while the user drags and the anchor is the end pin, which they recognise by its native ID ("paseo-end-pin"). Android overrides scrollToPreservingMomentum, the position keeper's only scroll call, from the begin-drag event until the finger lifts; iOS skips the adjustment while the scroll view is dragging. The skip is limited to the pin: LegendList's own anchoring for the detached list also uses the native correction, and a history page landing mid-drag needs it. It starts at the drag rather than at touch-down: a tap or long press while following would otherwise leave the list stranded short of its end, still following, once the finger lifts. The release. A gesture ending within 32 pt of the end re-stuck with an instant scroll, a snap of up to 56 px. Following now resumes only when the finger lifts at or past the end, or momentum stops there, within 1 px, the distance web resumes at (strategy-web.tsx). Anything else stays detached with the pill. The check runs when the gesture is decided, so output that arrived below meanwhile counts. A release past the end counts as at the end: on iOS it bounces back, and a line landing during the bounce scrolls in as following would. endUserScroll takes isAtEnd, like releaseDisclosure. Tests: VisibleContentPositionClampTest gains growth at the end during a drag (red before the change: the offset moved by the 120 px growth, 352 -> 472) and growth without a touch (the pin still follows); the module suite passes on emulator-5558, 11 tests. bottom-anchor-controller.test.ts: a release near but short of the end stays detached. Device, frame scans while streaming (dev clients rebuilt): - Android, 8 slow drags with a pill tap between: no frame moves against the finger once a drag starts; the stream follows between drags. Three short releases 50 px short of the end stay put until the pill; three drags back past the end resume following. - iOS simulator, 6 drags: no line step during any drag or its momentum. Two short releases stay detached with no snap; three drags past the end resume following within 0.5 s of the bounce. * Document the LegendList timeline and remove what the inverted list left The native section of docs/agent-stream-disclosure.md described the FlatList hold (anchor neighbour, native child index, frame spacing, render window). It now covers why the timeline is forward order, who owns the offset in each state, the hold and its sync layout trigger, user scroll and release, draw distance numbers, the LegendList patch and the scroll view corrections. Draw distance stays at 2000. The spike measured it on a chat with a 15,000 px turn expanded: 500 blanked the viewport 1.0 to 1.5 s per fling and 2000 not at all. A re-measure on the collapsed chat (500, 1000, 1500, 2000, interleaved) blanked at no value and its jank followed host load, which stayed between 19 and 40, so it could not separate 1000 or 1500 from 2000. The patch file now opens with a header listing its eight hunks by purpose and linking LegendApp/legend-list#593, which reports each one upstream. The markdown doc records the iOS first-paint delay (software-mansion/enriched-markdown#1131). Removed with the inverted list: PaseoScrollView's child-reveal override (a no-op without the inverting transform) and ScrollSelectionTest, the invertedWrapper style, and the inverted-list passages in android.md, native-text-selection.md, mobile-testing.md, floating-panels.md and agent-stream-performance.md. model.test, layout.test and footer-spacing.test still asserted the inverted native order and failed; they now assert forward order, and footer-spacing uses the real strategies instead of an inverted stub. * fix(app): show a reply's footer when the next prompt arrives on native LegendList re-renders a mounted row only when its item or extraData changes. When a new prompt arrives, the previous reply's footer moves from the live auxiliary slot into that reply's layout item (completedFooter). The reply's StreamItem keeps its identity, and renderItem is a stable event, so the row never rendered its new layout and the footer (copy, timing, fork) disappeared until the row remounted. FlatList had the same gap. The history renderer is now LegendList's extraData. Its identity changes with the history layout map, and StreamRow's memo limits the pass to rows whose layout item changed. * test(app): run the web disclosure hold cases in a real browser The three hold cases simulated layout in jsdom: hand-written row heights, scroll clamping and resize callbacks. They could pass while Chromium paints the trigger somewhere else. They now run in Vitest browser mode with real layout, ResizeObserver and scroll clamping. The two hold cases fail when the hold is bypassed. The browser project pre-bundles @tanstack/react-virtual, so its first import no longer reloads the test mid-run. The lucide stub gains ArrowDown, which pane-find imports. * fix(app): let a drag end a disclosure hold on native A drag detached the bottom controller but kept the disclosure hold. Every item size change restarts the hold's quiet frames, including a streaming row's, so the hold can outlive the toggle. While it lasts, shouldRestorePosition accepts only the held row. Once the drag scrolls that row away, a history page arriving mid-gesture gets no anchor, and the rows the reader was on move by the page's height. The drag now cancels the hold without releasing it. The controller keeps no hold state of its own beyond detached, and the gesture's end decides whether following resumes. * perf(app): keep the scroll-to-bottom button's state out of the timeline Opening a tool call or turn at the end of the chat now holds its trigger in place, so the timeline detaches and the scroll-to-bottom button appears. That visibility was useState in AgentStreamView, so showing the button re-rendered the whole timeline view for a change only the button reads. The viewport now writes near-bottom into a per-agent store that only the button subscribes to. * fix(app): show the scroll-to-bottom button on native whenever following has stopped A gesture now resumes following only when it ends within 1 px of the end, but the button still hid within 32 px. A release 2-32 px short of the end left the list detached with no button to return. While detached the button now hides only at the end; while following it keeps the 32 px tolerance so content growth that the offset trails for a frame does not flash it. The button also updates when following starts or stops without a scroll. * fix(app): stop the timeline jumping when the keyboard opens or closes on native The keyboard's settled height was committed to the list after the motion, as paddingTop on Android and contentInset.top on iOS, to keep the top of the history reachable under the translated dock. While following, the end pin absorbed it in the same mount. While detached or holding a disclosure, nothing did: Android's native anchor is LegendList's ScrollAdjust view, which padding does not move, and LegendList compensates a padding change only on iOS. Every row jumped by the keyboard height 0.2-1 s after the keyboard settled, and back after it closed. On send, the same commit raced the jump to the end. The keyboard now only translates the list with the composer. The cost: while the keyboard is open, the composer viewport clips the top keyboard-height of the list, so the start of a short conversation is hidden until it closes. * chore(app): describe the native timeline as LegendList in stream comments * fix(app): keep a disclosure's trigger in place when its row is cut off at the top on native LegendList's idsInView starts at the first fully visible row. When a turn host's prompt bubble sat above the viewport, a disclosure hold found no anchor in the data pass (shouldRestorePosition rejects every other row), and the size pass anchored on the row below the host instead, so the trigger moved by the host's size change. The patch appends the partly visible rows to idsInView after the fully visible ones, so the held row anchors while every pass that does not restrict the anchor keeps the first fully visible row. The browser test renders the patched native build, which vitest now aliases instead of the package's unpatched web build. * fix(app): let a turn's summary line collapse again after it opens The React Compiler cached StreamRow's `renderStreamItem(layoutItem)` on the layout item and the renderer. The revised item that useRevisedHistoryRows hands a row whose display state changed only fed the row id, so the row reused its cached content. Opening a turn usually inserts its rows directly below the host, which also gives the host a new layout item. When another row sits between them, such as a compaction marker, the host kept its layout item, its summary line stayed collapsed, and every later tap asked to open the turn again. The renderer now takes the item it renders, so the compiler's cache keys on the revision. StreamRow moves to its own module, which the tests compile the way Expo does. * test(app): alias the patched LegendList build in browser tests only Unit tests load @legendapp/list outside Vite, so the native build's react-native import skipped the react-native-web alias and reached React Native's Flow source, failing every unit suite that imports the native strategy. * fix(app): keep a size-only disclosure's trigger in place when its row is cut off at the top on native LegendList's prepareMVCP picks the anchor for a size change with no data change as the first row in view, without asking shouldRestorePosition. A disclosure that only grows its own row (a todo card) while that row's top was above the viewport therefore held the row below it, and the trigger moved by the row's size change. The patch now applies shouldRestorePosition to size changes as it already does to data changes. * test(app): render the timeline row revision test in a real browser The repo's unit tests do not mount components in JSDOM. The test still compiles stream-row.tsx with the React Compiler, so it fails when a revised row ignores the item it was handed.
76 lines
8.5 KiB
Markdown
76 lines
8.5 KiB
Markdown
# Agent stream performance
|
||
|
||
How assistant text gets from a provider to the screen, and why it is paced on the way. Read this before changing `packages/server/src/server/agent/agent-stream-coalescer.ts`, the reducer queue in `packages/app/src/timeline/session-stream-reducers.ts`, or the reveal in `packages/app/src/hooks/use-revealed-text.ts`.
|
||
|
||
For terminal output, which is a separate pipeline with separate budgets, see [terminal-performance.md](terminal-performance.md).
|
||
|
||
## The pipeline
|
||
|
||
```
|
||
provider deltas (every provider streams incrementally)
|
||
→ AgentStreamCoalescer (daemon, leading + trailing, ≤1 message per 60ms per agent)
|
||
→ recordTimeline: one canonical row per flushed item
|
||
→ agent_stream ws message
|
||
→ reducer queue (app, one commit per frame) → session store
|
||
→ source-item plugin transforms → Markdown presentation / tool grouping
|
||
→ paced reveal (app, per displayed item) → paint
|
||
```
|
||
|
||
Every provider delivers incremental text, so there is no provider that needs special handling: Claude via `includePartialMessages`, Codex via `agent_message_delta`, ACP agents via `agent_message_chunk`, Pi and OMP via `text_delta`.
|
||
|
||
## Why the reveal is paced
|
||
|
||
Arrival is lumpy and there is no fixing that at the source. A 60ms coalescing window carries however many characters the model produced in those 60ms, which swings by an order of magnitude within a single turn. Painting each delta as it lands makes the size of those lumps visible, and that is what reads as jagged.
|
||
|
||
So arrival sets a _target_ and the reveal rate is derived from the backlog instead. A burst makes the text catch up faster; it does not make the text jump. Shrinking the coalescing window does not fix this — it makes the lumps smaller and more frequent, at the cost of message rate on a daemon loop that already contends with terminal frames and per-message relay encryption.
|
||
|
||
## Invariants
|
||
|
||
- **The coalescer is leading + trailing.** The first delta after an idle window flushes synchronously; only the rest of the burst waits for the trailing timer. Reverting to trailing-only adds a full window to the first character of every turn. Same shape and the same reason as `TerminalOutputCoalescer`.
|
||
- **The leading flush adds a canonical row, and that is fine.** A burst's first chunk lands as its own timeline row. `mergeAssistantChunks` / `mergeReasoningChunks` in `timeline-projection.ts` join contiguous same-turn rows, and clients read the projected timeline, so history is unaffected. Tests that assert on raw rows have to account for the extra row; tests that assert on what a client sees do not.
|
||
- **The store holds the full text; only the rendered slice is paced.** Copy, selection, the chat outline, and scroll geometry all read the same string the user can see. Pacing the store instead would leave the bottom anchor chasing a content height that is ahead of the reveal.
|
||
- **One source assistant message is one display row.** Presentation projects plugin output and caches unchanged history. The shared [Markdown renderer](markdown-renderer.md) owns rendering on every platform. Chat paces the visible source and applies a 32,000-character render cap on web; native keeps the complete message. The native renderer keeps closed blocks in a stable head and sends the paced suffix to a separate tail view during streaming; completion reunites them for selection. This stays inside the renderer, without promoting blocks into timeline history.
|
||
- **Message identity survives completion.** Assistant row ids are source message ids. Plugin rows can replace a source, so find, scroll-to-message and history reveal still use `getStreamItemMessageId`. Width-specific measurements describe whole rendered messages.
|
||
- **First sight of a text is revealed whole.** Only growth is paced. This is what makes history hydration, timeline replay, a virtualized row remounting on scroll, and an already-finished message all render complete on first paint without a special case for each.
|
||
- **Leaving `phase: "streaming"` snaps the reveal.** A completed turn must never be left holding characters. `layoutStream` sets the phase, so anything outside the live head with an active turn is already complete.
|
||
- **The reducer queue commits on a frame, with a timer as the ceiling.** A frame callback never fires in a hidden tab, so a timer races it and wins when nothing is painting — the store has to keep advancing either way.
|
||
- **A history row re-renders only when its item or layout item identity changes.** Every
|
||
coalesced tick hands the list a new layout. Without a memo boundary each tick re-rendered every
|
||
mounted row (measured on the inverted FlatList native used before LegendList: about 50 rows on a
|
||
phone, 100–250 ms of JS per tick). `layoutStream` keeps a layout item's identity when nothing about it changed,
|
||
`useRevisedHistoryRows` hands a fresh item identity to rows whose tool-call group, expanded state,
|
||
or breakpoint changed, and `StreamRow` memoizes on both. Every viewport runs its history
|
||
through that hook; the web viewport once skipped it and history hosts of a live tool group went
|
||
stale. A new field on `StreamLayoutItem` must be added to `areLayoutItemsEquivalent`, or sharing
|
||
silently stops.
|
||
- **Streaming cost must not grow with revealed history.** Reading older messages must not
|
||
make the next reply progressively more expensive to display. Reuse unchanged history work
|
||
while text arrives, and release obsolete window caches as the reader reveals more history;
|
||
window caches must stay bounded while their source timeline remains loaded.
|
||
|
||
## Native parsing cost
|
||
|
||
The native library reparses and measures its entire `markdown` prop on each update. During streaming, the renderer splits source at an mdast block boundary and memoizes the head, so per-frame updates parse only the tail. Boundary discovery runs when a source line completes, rather than on every reveal frame. The unfinished line and final parsed block stay in the tail, preserving list continuation, setext headings, tables and open fences. Reference-dependent source and HTML retain a larger tail because later source can change their meaning. A single long block therefore still carries its full parsing cost.
|
||
|
||
Completion restores one native text view per Mermaid-delimited segment, so selection crosses prose paragraphs again. The head retains its trailing block margin while split. JS reveal remains the pacing source; native opacity animation stays disabled.
|
||
|
||
## Measuring
|
||
|
||
- **Smoothness (user-perceived):** `packages/app/e2e/browser/agent-stream-smoothness.spec.ts`, gated behind `PASEO_AGENT_STREAM_PERF_E2E=1`. Drives the mock provider's `bursty-stream` model and reports coefficient of variation of characters painted per frame (smoothness) plus p95 gap between visible updates (stalls). Both numbers are needed: a stalled stream is perfectly smooth.
|
||
- **Reproducing bursty arrival:** the `bursty-stream` model in `mock-load-test-agent.ts` emits uneven runs of tokens separated by idle gaps. Burst sizes come from a seeded generator, so a run repeats exactly.
|
||
- **Rate policy in isolation:** `computeRevealStep` in `packages/app/src/agent-stream/text-reveal.ts` is pure; `text-reveal.test.ts` covers convergence and burst flattening without a renderer.
|
||
- **Per-delta cost on a long timeline:** `npx vitest bench --project unit src/agent-stream/stream-tick.bench.ts` from `packages/app` replays 200 deltas over a 2,000-item timeline through the pure model and layout pipeline, with the default mounted window and with history fully revealed. It measures the JS work per delta, not paint.
|
||
|
||
Healthy numbers (2026-08, Expo web against a local dev daemon, real Claude Haiku agent, ~8.5s samples during active streaming). Setting `TEXT_REVEAL_HORIZON_MS` to 0 makes the reveal paint on arrival, which is how the baseline column was taken:
|
||
|
||
| | paint on arrival | paced |
|
||
| -------------------------------- | ---------------- | ----- |
|
||
| frames that advanced the text | 6% | 87% |
|
||
| chars-per-frame CV | 4.11 | 1.86 |
|
||
| gap between visible updates, p50 | 317ms | 17ms |
|
||
| gap between visible updates, p95 | 383ms | 17ms |
|
||
|
||
Total characters painted is roughly the same either way — the reveal changes when they land, not how many arrive.
|
||
|
||
Measure the **total** length across every `assistant-message` element, not the last one. A turn emits many assistant messages, so the tail element keeps changing identity and its length is not monotonic; sampling only the tail reads those handovers as resets and reports almost no growth. `sampleStreamFrames` does this correctly.
|