perf(client): инкрементальный рендер стрима ответа (#492)

Путь ответа ассистента теперь рендерится инкрементально, по образцу
StreamingPlainText из ветки reasoning (#492, волна C эпика #497).

Раньше MarkdownPart прогонял ВЕСЬ накопленный ответ через канонический
конвейер (markdownToProseMirrorSync → PMNode.fromJSON → DOMSerializer →
DOMPurify) на КАЖДОМ throttled-тике (~20 Гц). На синтетическом потоке
~100 КБ это 394 вызова renderChatMarkdown — O(числа тиков), причём каждый
вызов заново парсит всю растущую строку.

Теперь:
- StreamingMarkdownText делит текст на блоки по безопасному срезу
  (splitPlainChunks — та же append-only-инвариантность, что у reasoning):
  СТАБИЛИЗИРОВАННЫЕ блоки идут через канонический конвейер и мемоизируются
  (каждый блок парсится РОВНО ОДИН раз), живой ХВОСТ — дешёвый plain-text
  (React-escaped, без парсера/санитайзера/innerHTML) до стабилизации.
- На финализации (флип state → done или конец хода) — ОДИН полный
  канонический рендер всего текста: побайтовая визуальная паритетность с
  прежним выводом (включая <li><p>-обёртки схемы и scoped-CSS из #498).
- Гейт liveness тот же, что у ReasoningBlock: streaming =
  turnStreaming && part.state === "streaming".

Также цикл рендера частей переведён на ИСЧЕРПЫВАЮЩИЙ switch по видам частей
с never-проверкой в default (вместо прежнего WARNING-комментария): новый
закрытый вид части в UIMessagePart теперь ошибка компиляции.

Тесты:
- perf-smoke: на ~100 КБ потоке число вызовов renderChatMarkdown ≤ blocks+2
  и ≪ ticks (O(блоков), не O(тиков)); мутация (снять memo с MarkdownChunk)
  краснит ассерт (78631 вызовов).
- visual-regression: финальный рендер побайтово равен renderChatMarkdown
  всего текста (в т.ч. <li><p>), инкрементальный вид сходится к нему на
  финализации; учтён neutralizeInternalLinks.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
2026-07-12 04:35:06 +03:00
parent 25a31e6c0d
commit 2f8c5d9a98
4 changed files with 395 additions and 45 deletions
@@ -27,6 +27,7 @@ vi.mock("@/features/ai-chat/utils/markdown.ts", async () => {
import MessageItem from "./message-item";
import { messageSignature } from "@/features/ai-chat/utils/message-signature.ts";
import { splitPlainChunks } from "./streaming-plain-text";
// matchMedia (read by MantineProvider) is stubbed globally in vitest.setup.ts.
@@ -114,3 +115,89 @@ describe("MessageItem markdown memoization", () => {
expect(queryByText("streamed answer")).not.toBeNull();
});
});
// PERF SMOKE (#492): the whole point of the incremental streaming render is that
// the ANSWER path costs O(number of markdown blocks), NOT O(number of throttled
// ~20Hz ticks). Pre-#492 the finalized MarkdownPart re-parsed the WHOLE growing
// answer on every delta — a synthetic ~100 KB stream measured 394 renderChatMarkdown
// calls (one per tick). With the incremental render each STABILIZED block is parsed
// exactly once (memoized in MarkdownChunk) and the live tail is cheap plain text, so
// the call count collapses to ~= the block count regardless of tick granularity.
describe("MessageItem streaming answer render is O(blocks), not O(ticks)", () => {
// ~100 KB answer. Each section is a heading + a paragraph — TWO blank-line
// delimited markdown blocks — so the safe-cut block count is ~2× the section
// count. The perf claim is about the BLOCK count (the memoization granularity),
// measured directly with splitPlainChunks below, not the section count.
const buildAnswer = () => {
const SECTIONS = 100;
const paragraphs: string[] = [];
for (let i = 0; i < SECTIONS; i++) {
paragraphs.push(`## Section ${i}\n\n` + "lorem ipsum dolor ".repeat(55));
}
const full = paragraphs.join("\n\n");
// The number of memoized markdown blocks the incremental render splits into
// (all but the live tail are parsed once each).
return { full, blocks: splitPlainChunks(full).length };
};
const streamMsg = (text: string, state: "streaming" | "done"): UIMessage =>
({
id: "m1",
role: "assistant",
parts: [{ type: "text", text, state }],
}) as UIMessage;
it("parses each block ~once over a 100KB stream (≈blocks, ≪ ticks)", () => {
renderChatMarkdownSpy.mockClear();
const { full, blocks } = buildAnswer();
const CHUNK = 128; // a realistic ~20Hz throttled delta size
const ticks = Math.ceil(full.length / CHUNK);
let msg = streamMsg(full.slice(0, CHUNK), "streaming");
const { rerender } = render(
<MantineProvider>
<MessageItem
message={msg}
signature={messageSignature(msg)}
turnStreaming
/>
</MantineProvider>,
);
for (let end = 2 * CHUNK; end < full.length; end += CHUNK) {
msg = streamMsg(full.slice(0, end), "streaming");
rerender(
<MantineProvider>
<MessageItem
message={msg}
signature={messageSignature(msg)}
turnStreaming
/>
</MantineProvider>,
);
}
// Finalize: the streaming→done flip renders the whole answer through ONE
// canonical pass (visual parity), so the finished DOM matches the pre-#492
// output. This is the single extra parse on top of the per-block ones.
const done = streamMsg(full, "done");
rerender(
<MantineProvider>
<MessageItem message={done} signature={messageSignature(done)} />
</MantineProvider>,
);
const calls = renderChatMarkdownSpy.mock.calls.length;
// Sanity: the stream really had far more ticks than blocks (else the test is
// vacuous — the point is that calls scale with blocks, not ticks).
expect(ticks).toBeGreaterThan(blocks * 3);
// O(blocks): each stabilized block parsed once + the single final whole-text
// parse. A small constant absorbs the finalize render and the live-tail block;
// the load-bearing claim is the bound below.
expect(calls).toBeLessThanOrEqual(blocks + 2);
// ≪ ticks — and, non-vacuously, the blocks WERE parsed (not skipped entirely).
expect(calls).toBeLessThan(ticks / 3);
expect(calls).toBeGreaterThan(blocks / 2);
// MUTATION-VERIFY (documented, not run here): dropping the `memo()` wrapper on
// MarkdownChunk (so every stable block re-parses each tick) drives `calls`
// toward `ticks` (~394), reddening both upper-bound assertions above.
});
});
@@ -0,0 +1,112 @@
import { describe, expect, it, vi } from "vitest";
import { render } from "@testing-library/react";
import { MantineProvider } from "@mantine/core";
import type { UIMessage } from "@ai-sdk/react";
// Stub react-i18next (the component reads `useTranslation`). Mirrors the other
// message-item specs.
vi.mock("react-i18next", () => ({
useTranslation: () => ({ t: (key: string) => key }),
}));
import MessageItem from "./message-item";
import { messageSignature } from "@/features/ai-chat/utils/message-signature.ts";
// The REAL canonical renderer (NOT the spy the memo test installs): this file
// exercises the actual markdown output so the visual-regression assertions below
// compare against genuine HTML (incl. the schema's `<li><p>` wrappers).
import { renderChatMarkdown } from "@/features/ai-chat/utils/markdown.ts";
import classes from "./ai-chat.module.css";
const msg = (
parts: UIMessage["parts"],
extra?: Partial<UIMessage>,
): UIMessage =>
({ id: "m1", role: "assistant", parts, ...extra }) as UIMessage;
const renderRow = (message: UIMessage, turnStreaming = false) =>
render(
<MantineProvider>
<MessageItem
message={message}
signature={messageSignature(message)}
turnStreaming={turnStreaming}
/>
</MantineProvider>,
);
// A rich multi-block answer that exercises headings, a list (the `<li><p>` case
// the scoped CSS tightens), inline emphasis, and multiple paragraphs.
const ANSWER = [
"# Заголовок",
"",
"Первый абзац с **жирным** и `кодом`.",
"",
"- пункт один",
"- пункт два",
"",
"Второй абзац.",
].join("\n");
describe("MessageItem final render — visual parity with the canonical pipeline", () => {
it("a finalized text part renders exactly renderChatMarkdown(text)", () => {
const { container } = renderRow(
msg([{ type: "text", text: ANSWER, state: "done" }]),
);
const block = container.querySelector(`.${classes.markdown}`);
expect(block).not.toBeNull();
// Byte-for-byte the canonical output (the SAME whole-text pass the pre-#492
// MarkdownPart produced), including `<li><p>…</p></li>` wrappers.
expect(block!.innerHTML).toBe(renderChatMarkdown(ANSWER, {}));
// The list wrapper is really present (guards against a vacuous empty render).
expect(container.querySelectorAll("li p").length).toBe(2);
});
it("the streaming incremental view CONVERGES to the canonical render on finish", () => {
// Mount mid-stream (live tail) — the DOM here is the incremental view.
const { container, rerender } = render(
<MantineProvider>
<MessageItem
message={msg([{ type: "text", text: ANSWER, state: "streaming" }])}
signature={messageSignature(
msg([{ type: "text", text: ANSWER, state: "streaming" }]),
)}
turnStreaming
/>
</MantineProvider>,
);
// Finish the turn: state flips to done AND the turn is no longer streaming.
const done = msg([{ type: "text", text: ANSWER, state: "done" }]);
rerender(
<MantineProvider>
<MessageItem message={done} signature={messageSignature(done)} />
</MantineProvider>,
);
// After finish there is exactly ONE canonical markdown container whose HTML is
// the whole-text render — identical to the non-streaming path above.
const blocks = container.querySelectorAll(`.${classes.markdown}`);
expect(blocks.length).toBe(1);
expect(blocks[0].innerHTML).toBe(renderChatMarkdown(ANSWER, {}));
});
it("neutralizeInternalLinks is honored on the finalized render", () => {
const linkAnswer = "См. [страницу](/p/abc).";
const { container } = render(
<MantineProvider>
<MessageItem
message={msg([{ type: "text", text: linkAnswer, state: "done" }])}
signature={messageSignature(
msg([{ type: "text", text: linkAnswer, state: "done" }]),
)}
neutralizeInternalLinks
/>
</MantineProvider>,
);
const block = container.querySelector(`.${classes.markdown}`);
expect(block!.innerHTML).toBe(
renderChatMarkdown(linkAnswer, { neutralizeInternalLinks: true }),
);
// The internal link was made inert (no href) by the neutralization flag.
const a = container.querySelector("a");
expect(a?.hasAttribute("href")).toBe(false);
});
});
@@ -4,6 +4,7 @@ import { useTranslation } from "react-i18next";
import type { UIMessage } from "@ai-sdk/react";
import ToolCallCard from "@/features/ai-chat/components/tool-call-card.tsx";
import ReasoningBlock from "@/features/ai-chat/components/reasoning-block.tsx";
import { StreamingMarkdownText } from "@/features/ai-chat/components/streaming-markdown-text.tsx";
import ChatErrorAlert from "@/features/ai-chat/components/chat-error-alert.tsx";
import ChatStoppedNotice from "@/features/ai-chat/components/chat-stopped-notice.tsx";
import { ToolUiPart, isToolPart } from "@/features/ai-chat/utils/tool-parts.tsx";
@@ -86,17 +87,39 @@ interface MessageItemProps {
* One assistant text part rendered as sanitized markdown. Memoized on its inputs
* so a finalized text part is NOT re-parsed on every streamed delta: during a
* turn only the actively-growing tail part changes its `text`, so every earlier
* part hits the memo and skips the expensive marked + DOMPurify pass. Props are
* primitives, so React.memo's default shallow compare is exactly right (the
* `text` string is compared by value).
* part hits the memo and skips the expensive canonical parse + DOMPurify pass.
* Props are primitives, so React.memo's default shallow compare is exactly right
* (the `text` string is compared by value).
*
* Streaming gate (#492) — mirrors ReasoningBlock:
* - `streaming` (this is the live, actively-growing tail part of an in-flight
* turn): render incrementally via StreamingMarkdownText — the stabilized blocks
* go through the canonical pipeline (each parsed ONCE, memoized) and only the
* live tail is cheap plain text. This makes the per-tick cost O(new blocks),
* not the pre-#492 O(ticks) whole-answer re-parse on every ~20Hz delta.
* - finalized (the common case, and the turn-end flip): render the WHOLE text
* through ONE canonical pass — byte-identical to the pre-#492 output (visual
* parity). The row re-renders on the streaming→done flip because
* `messageSignature` tracks each part's `state` (and `turnStreaming` flips at
* turn end), so the incremental view always converges to this single render.
*/
const MarkdownPart = memo(function MarkdownPart({
text,
neutralizeInternalLinks,
streaming,
}: {
text: string;
neutralizeInternalLinks: boolean;
streaming: boolean;
}) {
if (streaming) {
return (
<StreamingMarkdownText
text={text}
neutralizeInternalLinks={neutralizeInternalLinks}
/>
);
}
const html = renderChatMarkdown(text, { neutralizeInternalLinks });
if (html) {
return (
@@ -179,47 +202,10 @@ function MessageItem({
{resolveAssistantName(assistantName) ?? t("AI agent")}
</Text>
{message.parts.map((part, index) => {
if (part.type === "reasoning") {
// Reasoning ("thinking") -> a collapsible block with its own token
// count. Empty/whitespace reasoning with no authoritative count carries
// nothing to show, so skip it (avoids an empty 0-token block).
const text = (part as { text?: string }).text ?? "";
if (!text.trim() && !(reasoningTokens && reasoningTokens > 0))
return null;
// Absent state (persisted rows) and "done" both mean finalized.
// `messageSignature` already includes each part's `state`, so the
// streaming→done flip changes the row signature and re-renders this
// row — which is what lets ReasoningBlock switch from chunked plain
// text to its one-time markdown parse (see reasoning-block.tsx).
// ALSO require the turn to be live: a part stranded at
// `state:"streaming"` after the turn ended (no `reasoning-end` — see
// the `turnStreaming` prop doc) must still finalize and parse.
const streaming =
turnStreaming && (part as { state?: string }).state === "streaming";
return (
<ReasoningBlock
key={index}
text={text}
tokens={reasoningTokens}
streaming={streaming}
/>
);
}
if (part.type === "text") {
// Skip empty/whitespace-only text parts (a streaming message often
// starts with an empty text part before the first token arrives); the
// typing indicator covers that gap until real content streams in.
if (!part.text.trim()) return null;
return (
<MarkdownPart
key={index}
text={part.text}
neutralizeInternalLinks={neutralizeInternalLinks}
/>
);
}
// Tool parts (`tool-*` / `dynamic-tool`) are template-literal kinds, so
// they cannot be a `switch` case; the runtime guard handles them, and the
// switch below covers every CLOSED (literal-typed) part kind with a
// compile-time exhaustiveness check in its default.
if (isToolPart(part.type)) {
return (
<ToolCallCard
@@ -232,7 +218,76 @@ function MessageItem({
);
}
return null;
switch (part.type) {
case "reasoning": {
// Reasoning ("thinking") -> a collapsible block with its own token
// count. Empty/whitespace reasoning with no authoritative count
// carries nothing to show, so skip it (avoids an empty 0-token block).
const text = part.text ?? "";
if (!text.trim() && !(reasoningTokens && reasoningTokens > 0))
return null;
// Absent state (persisted rows) and "done" both mean finalized.
// `messageSignature` already includes each part's `state`, so the
// streaming→done flip changes the row signature and re-renders this
// row — which is what lets ReasoningBlock switch from chunked plain
// text to its one-time markdown parse (see reasoning-block.tsx).
// ALSO require the turn to be live: a part stranded at
// `state:"streaming"` after the turn ended (no `reasoning-end` — see
// the `turnStreaming` prop doc) must still finalize and parse.
const streaming = turnStreaming && part.state === "streaming";
return (
<ReasoningBlock
key={index}
text={text}
tokens={reasoningTokens}
streaming={streaming}
/>
);
}
case "text": {
// Skip empty/whitespace-only text parts (a streaming message often
// starts with an empty text part before the first token arrives); the
// typing indicator covers that gap until real content streams in.
if (!part.text.trim()) return null;
// The live, actively-growing tail part of the in-flight turn renders
// incrementally (see MarkdownPart); a finalized part (persisted, or
// the turn-end flip) renders the whole text through one canonical
// pass. Same liveness rule as the reasoning branch above.
const streaming = turnStreaming && part.state === "streaming";
return (
<MarkdownPart
key={index}
text={part.text}
neutralizeInternalLinks={neutralizeInternalLinks}
streaming={streaming}
/>
);
}
case "source-url":
case "source-document":
case "file":
case "step-start":
// Not surfaced in the chat bubble (v1) — same as the pre-#492 default.
return null;
default: {
// Compile-time exhaustiveness over the CLOSED union members: every
// literal-typed part kind is handled above, so the only kinds that
// can reach here are the OPEN template-literal ones (`tool-*` — caught
// by the guard at runtime — and `data-*`) plus `dynamic-tool`. Adding
// a NEW closed part kind to UIMessagePart makes this assignment fail
// to compile, forcing it to be handled instead of silently ignored
// (this replaces the pre-#492 fall-through `return null` + WARNING).
const _exhaustive:
| `tool-${string}`
| "dynamic-tool"
| `data-${string}` = part.type;
void _exhaustive;
return null;
}
}
})}
{/* A persisted turn error (server stored it in metadata.error). Rendered
here so it survives a thread remount and shows in reopened history. */}
@@ -0,0 +1,96 @@
import { memo, useMemo } from "react";
import { splitPlainChunks } from "@/features/ai-chat/components/streaming-plain-text.tsx";
import { renderChatMarkdown } from "@/features/ai-chat/utils/markdown.ts";
import classes from "@/features/ai-chat/components/ai-chat.module.css";
/**
* One STABILIZED markdown block, rendered through the canonical pipeline and
* memoized on its string prop. During streaming only the TAIL chunk grows (the
* `splitPlainChunks` append-only invariant guarantees every earlier chunk is
* byte-identical across deltas), so React skips every stable block and each one
* is parsed by `renderChatMarkdown` EXACTLY ONCE — turning the pre-#492
* "re-parse the whole accumulated answer on every ~20Hz tick" (O(ticks)) into
* O(number of blocks). The markup is DOMPurify-sanitized inside renderChatMarkdown
* before it reaches `dangerouslySetInnerHTML`.
*
* NOTE (transient streaming-only artifact): a safe cut is a blank-line boundary,
* so a construct that legitimately contains a blank line (e.g. a fenced code block
* with an empty line) can be split across chunks and render oddly WHILE it is still
* streaming. This is cosmetic and self-heals: the moment the part finalizes,
* MarkdownPart renders the WHOLE text through one canonical pass (visual parity
* with the pre-#492 output). The reasoning path makes the same trade (plain text
* while streaming, one markdown parse at the end).
*/
const MarkdownChunk = memo(function MarkdownChunk({
text,
neutralizeInternalLinks,
}: {
text: string;
neutralizeInternalLinks: boolean;
}) {
const html = renderChatMarkdown(text, { neutralizeInternalLinks });
if (html) {
return (
<div
className={classes.markdown}
// Sanitized by renderChatMarkdown (DOMPurify) before insertion.
dangerouslySetInnerHTML={{ __html: html }}
/>
);
}
// Malformed/unsupported markdown could not render synchronously: raw text.
return (
<div className={classes.markdown} style={{ whiteSpace: "pre-wrap" }}>
{text}
</div>
);
});
/**
* The cheap streaming-time stand-in for the finalized answer's one-time markdown
* parse (see MarkdownPart in message-item.tsx). Mirrors StreamingPlainText's
* chunked-memo pattern but renders the STABILIZED prefix as real markdown (each
* block parsed once, memoized) and only the LIVE tail as flat plain text — so the
* user sees formatted output for everything up to the last safe cut, and the not-
* yet-stable tail (which markdown-parsing every tick would make O(ticks)) stays a
* single cheap escaped text node until it stabilizes into a new block.
*
* `splitPlainChunks` yields chunks where, under append-only growth, every chunk
* except the LAST is immutable; the last chunk is the live tail. Index keys are
* therefore stable (a given index never changes to a different chunk's content).
*/
export function StreamingMarkdownText({
text,
neutralizeInternalLinks,
}: {
text: string;
neutralizeInternalLinks: boolean;
}) {
const chunks = useMemo(() => splitPlainChunks(text), [text]);
return (
<>
{chunks.map((chunk, index) =>
index < chunks.length - 1 ? (
<MarkdownChunk
key={index}
text={chunk}
neutralizeInternalLinks={neutralizeInternalLinks}
/>
) : (
// The live tail: flat, React-escaped plain text (no markdown parse, no
// sanitizer, no innerHTML). `pre-wrap` preserves its newlines; trailing
// separator newlines are dropped at display time so the block gap comes
// from the markdown margins, not a doubled empty line (mirrors
// PlainChunk in streaming-plain-text.tsx).
<div
key={index}
className={classes.markdown}
style={{ whiteSpace: "pre-wrap" }}
>
{chunk.replace(/\n+$/, "")}
</div>
),
)}
</>
);
}