root/Jyotisha
1
0
Fork 0
Code Issues 1 Pull Requests 1 Actions Packages Projects Releases Wiki Activity

fix(consult): paced settle is opt-in for the consultation reply; one rejection path for HTTP bodies and stream events (T2/T3 follow-up)

Browse Source 
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0199rbQDTsUbCVw84wc8BTFe
This commit is contained in:
Jesse_Chen
2026-09-28 09:28:37 +08:00
co-authored by Claude Fable 5.1
parent 0bc6659052
commit 891e9c62da
6 changed files with 50 additions and 40 deletions
Download Patch File Download Diff File Expand all files Collapse all files
frontend/DESIGN.md
+1 -1
View File
@@ -753,7 +753,7 @@ or user IDs.
| 行内 / 局部等待 | 出生地解析、两个会话面时间线的 live 步、兜底活动面板的 live 行、个人报告列表与详情 | `InlineSpinner`（`inline-spin`） | 0.8s linear | `animation: none`，收成静止圆点，不要半圈圆弧 |
| 盘位骨架（仅星盘页） | D1 / 非 D1 分盘 / 西洋盘尚未返回 | `VedicChartSvg skeleton` / `WesternWheelSvg skeleton`；统一 `.chart-page-skeleton` | 2.4s ease-in-out，opacity 0.35 ↔ 0.6 | 停止动画，opacity 0.5 |
| 流式生成中 | 引导语打字、时间线 summary 与 live 行的文案 | `onboarding-caret` / `agent-activity-shimmer` | 700ms steps / 1.6s linear | 保持现有全局降级 |
| 流式正文的放字节奏 | 普通对话与校正面的回答正文（`stream-frame-buffer`） | 每帧最多 4 字（≈ 240 字/秒），积压超过 600 字才按 1/12 追赶；流结束后剩余文字按同一节奏在 1.5 s 内写完，最后一帧才算 settled；停止 / 失败 / 断线 / 页面隐藏一帧全放（2026-09-28，BUG-1075） | 逐帧 | 不涉及 CSS 动画，无需降级 |
| 流式正文的放字节奏 | 普通对话与校正面的回答正文（`stream-frame-buffer`） | 每帧最多 4 字（≈ 240 字/秒），积压超过 600 字才按 1/12 追赶；普通对话的流结束后剩余文字按同一节奏在 1.5 s 内写完（`settle({ paced: true })`），最后一帧才算 settled；停止 / 失败 / 断线 / 页面隐藏、以及校正面的流尾仍一帧全放（2026-09-28，BUG-1075） | 逐帧 | 不涉及 CSS 动画，无需降级 |
Agent 的 live 标记只有 `InlineSpinner` 一种。曾经并存的 canvas 小球（`thinking-orbs`）已移除，不得再引入第二种 live 标记。
frontend/src/hooks/use-consultation-run.ts
+13 -14
View File
@@ -808,14 +808,14 @@ export function useConsultationRun(params: ConsultationRunParams) {
// run was stopped or failed, in which case what arrived shows at once.
const settleFrames = (immediate: boolean) => {
if (immediate) {
void frames.settle({ immediate: true });
void frames.settle();
return;
}
settleRequested = true;
setStreamingReply((current) => (
current && current.sessionId === sessionId ? { ...current, settling: requestId } : current
));
void frames.settle();
void frames.settle({ paced: true });
};
try {
const response = await fetch("/api/consult", {
@@ -849,16 +849,18 @@ export function useConsultationRun(params: ConsultationRunParams) {
}),
signal: controller.signal,
});
// One rejection path for the HTTP body a pre-stream check answers with
// and for the run.failed request_rejected event the stream carries once
// it is open (BUG-1074): same status semantics either way.
const rejectRun = (status: number, message: string, code?: string): never => {
if (status === 401) window.location.assign("/login");
if (status === 402) openAccountDialog("billing", { source: "insufficient-credits" });
throw new ConsultationResponseError(status, message, code);
};
if (!response.ok) {
const contentType = response.headers.get("content-type") ?? "";
const errorPayload = contentType.includes("application/json") ? await response.json() : { message: await response.text() };
if (response.status === 401) window.location.assign("/login");
if (response.status === 402) openAccountDialog("billing", { source: "insufficient-credits" });
throw new ConsultationResponseError(
response.status,
payloadMessage(errorPayload, "服务暂时不可用"),
payloadCode(errorPayload),
);
rejectRun(response.status, payloadMessage(errorPayload, "服务暂时不可用"), payloadCode(errorPayload));
}
if (!response.body) {
throw new ConsultationResponseError(502, "浏览器未收到可读取的回答流");
@@ -966,10 +968,7 @@ export function useConsultationRun(params: ConsultationRunParams) {
// A rejection the route used to send as a JSON body with an HTTP
// status before the stream opened (BUG-1074): same status, same
// sentence, same code, so the handling below does not change.
const status = event.status ?? 503;
if (status === 401) window.location.assign("/login");
if (status === 402) openAccountDialog("billing", { source: "insufficient-credits" });
throw new ConsultationResponseError(status, payloadMessage({ message: event.message }, "服务暂时不可用"), event.reason);
rejectRun(event.status ?? 503, payloadMessage({ message: event.message }, "服务暂时不可用"), event.reason);
}
if (event.type === "run.failed") {
if (event.code === "answer_truncated") {
@@ -1086,7 +1085,7 @@ export function useConsultationRun(params: ConsultationRunParams) {
// Whatever arrived before the failure is what gets kept, not just the
// part the pacing had released so far.
settleRequested = false;
void frames.settle({ immediate: true });
void frames.settle();
const cancelled = controller.signal.aborted;
const ownsInterface = pendingConsultation.current?.requestId === requestId;
const partialReply = latestPartialReply;
frontend/src/lib/stream-frame-buffer.ts
+13 -11
View File
@@ -7,9 +7,10 @@
* happens per animation frame. Answer and thinking text are released at a
* steady per-frame pace so a burst of chunks reads as flowing text instead of
* a jump, while a large backlog (reconnect, slow tab) catches up in roughly a
* dozen frames. When the stream ends, what is still unreleased is written out
* at the same pace within a short budget instead of appearing in one frame
* (BUG-1075); only a stop, a failure or a hidden document releases at once.
* dozen frames. When a consultation stream ends, what is still unreleased is
* written out at the same pace within a short budget instead of appearing in
* one frame (a paced settle, BUG-1075); a stop, a failure, a hidden document
* and every other caller release at once.
*
* Pure release arithmetic lives in exported functions so the policy is
* testable without a DOM; scheduling is injectable for the same reason.
@@ -94,14 +95,15 @@ export type StreamFrameBuffer<Meta> = Readonly<{
/** Publish meta-only changes (timeline rows, activity) on the next frame. */
touch: () => void;
/**
* The stream has ended. By default what is still unreleased is written out
* at the typing pace within STREAM_SETTLE_MAX_FRAMES, and the last frame
* flushes with `settled: true`; the promise resolves on that frame. With
* `immediate` (stop, failure, disconnect) everything is released in one
* synchronous flush, as before BUG-1075. A hidden document always releases
* The stream has ended. By default everything unreleased goes out in one
* synchronous flush (stop, failure, disconnect, and surfaces that merge a
* snapshot right after). With `paced` (the consultation reply, BUG-1075)
* what is left is written out at the typing pace within
* STREAM_SETTLE_MAX_FRAMES, only the last frame flushes with `settled: true`,
* and the promise resolves on that frame. A hidden document always releases
* at once.
*/
settle: (options?: Readonly<{ immediate?: boolean }>) => Promise<void>;
settle: (options?: Readonly<{ paced?: boolean }>) => Promise<void>;
/** Drop everything, including scheduled work, without flushing. */
reset: (meta?: Meta) => void;
dispose: () => void;
@@ -233,7 +235,7 @@ export function createStreamFrameBuffer<Meta>(
if (settling) {
// A second settle while one is writing out: an immediate one takes
// over and flushes now; a paced one just waits for the first.
if (!settleOptions?.immediate) {
if (settleOptions?.paced) {
return new Promise<void>((resolve) => {
const previous = settling!.resolve;
settling!.resolve = () => { previous(); resolve(); };
@@ -242,7 +244,7 @@ export function createStreamFrameBuffer<Meta>(
cancelScheduled();
}
const pending = pendingChars(releasedAnswer, targetAnswer);
if (settleOptions?.immediate || scheduler.hidden() || pending <= 0) {
if (!settleOptions?.paced || scheduler.hidden() || pending <= 0) {
cancelScheduled();
releasedAnswer = targetAnswer;
releasedThinking = targetThinking;
frontend/tests/chat-navigation-a11y-contract.test.ts
+4 -1
View File
@@ -42,7 +42,10 @@ test("in-app destinations navigate client-side so the chat survives the round tr
assert.match(pageSource, /onOpenBilling: \(\) => openAccountDialog\("billing", \{ source: "account-menu" \}\)/);
assert.match(pageSource, /openAccountDialog\("billing", \{ returnTarget: event\.currentTarget, source: "credits" \}\)/);
assert.match(pageSource, /openAccountDialog\("billing", \{ source: "insufficient-credits" \}\)/);
assert.match(pageSource, /if \(response\.status === 402\) openAccountDialog\("billing", \{ source: "insufficient-credits" \}\)/);
// 原值: /if \(response\.status === 402\) openAccountDialog\("billing", \{ source: "insufficient-credits" \}\)/
// 新值: 同一句在 rejectRun 里，HTTP 体与流内 request_rejected 事件共用（状态从参数来）
// 原因: TASK-consult-first-frame-and-pacing-20260928 D2 / BUG-1074：开流后拒绝改走事件，调用点数不变（下面仍数到 2）
assert.match(pageSource, /if \(status === 402\) openAccountDialog\("billing", \{ source: "insufficient-credits" \}\)/);
// And: membership stays in-page; leaving chat is a client-side <Link> now.
assert.doesNotMatch(pageSource, /window\.location\.assign\("\/reports"\)/);
frontend/tests/consult-first-frame-and-pacing-20260928.test.tsx
+12 -6
View File
@@ -64,8 +64,9 @@ test("the consultation hook paces the write-out only for a finished run; stop, f
// The stream ended normally: paced unless the user stopped it or the server cut it.
assert.match(hook, /settleFrames\(controller\.signal\.aborted \|\| Boolean\(truncatedFailure\)\);/);
assert.match(hook, /settleFrames\(controller\.signal\.aborted\);/);
// Failure path keeps what arrived, at once.
assert.match(hook, /settleRequested = false;\s*void frames\.settle\(\{ immediate: true \}\);/);
// Failure path keeps what arrived, at once; only the finished consultation reply opts into pacing.
assert.match(hook, /settleRequested = false;\s*void frames\.settle\(\);/);
assert.equal((hook.match(/frames\.settle\(\{ paced: true \}\)/g) ?? []).length, 1);
// The paced write-out owns the display through the streaming reply's `settling` id;
// the interface completes without waiting for it and without clearing it.
assert.match(hook, /setStreamingReply\(\(current\) => \(current\?\.settling \? current : null\)\);/);
@@ -165,8 +166,10 @@ test("③ a JSON rejection becomes one run.failed request_rejected with the same
// Same precedence the client applied to the HTTP body: recovery, then message, then error.
const generation = consultationRejectionEvent(503, { error: "暂时无法生成解读", message: "咨询服务暂时不可用，请稍后再试。", recovery: "稍后重试，或换一个模型继续。" });
assert.equal(generation.type === "run.failed" && generation.message, "稍后重试，或换一个模型继续。");
assert.equal(consultationRejectionEvent(200, {}).type === "run.failed" && consultationRejectionEvent(200, {}).status, 503);
assert.equal(consultationRejectionEvent(503, null).type === "run.failed" && consultationRejectionEvent(503, null).message, STREAM_FIRST_FALLBACK_MESSAGE);
const outOfRange = consultationRejectionEvent(200, {});
assert.equal(outOfRange.type === "run.failed" ? outOfRange.status : undefined, 503);
const empty = consultationRejectionEvent(503, null);
assert.equal(empty.type === "run.failed" ? empty.message : undefined, STREAM_FIRST_FALLBACK_MESSAGE);
for (const event of [insufficient, full, generation]) consultationAgentPublicEventSchema.parse(event);
const response = streamFirstResponse({
@@ -291,8 +294,11 @@ test("the client reads smalltalk and rejections from events, with the same statu
assert.doesNotMatch(hook, /x-jyotish-response-kind/);
assert.match(hook, /if \(event\.responseKind === "smalltalk"\) \{\s*responseKind = "smalltalk";/);
assert.match(hook, /event\.type === "run\.failed" && event\.code === "request_rejected"/);
assert.match(hook, /if \(status === 402\) openAccountDialog\("billing", \{ source: "insufficient-credits" \}\);/);
assert.match(hook, /throw new ConsultationResponseError\(status, payloadMessage\(\{ message: event\.message \}, "服务暂时不可用"\), event\.reason\);/);
assert.match(hook, /const rejectRun = \(status: number, message: string, code\?: string\): never => \{\s*if \(status === 401\) window\.location\.assign\("\/login"\);\s*if \(status === 402\) openAccountDialog\("billing", \{ source: "insufficient-credits" \}\);/);
assert.match(hook, /rejectRun\(response\.status, payloadMessage\(errorPayload, "服务暂时不可用"\), payloadCode\(errorPayload\)\);/);
assert.match(hook, /rejectRun\(event\.status \?\? 503, payloadMessage\(\{ message: event\.message \}, "服务暂时不可用"\), event\.reason\);/);
// One redirect and one billing prompt for both paths (chat-navigation-a11y-contract counts them).
assert.equal((hook.match(/openAccountDialog\("billing", \{ source: "insufficient-credits" \}\)/g) ?? []).length, 2);
// The rollback / notice handling keyed on status and code is unchanged.
assert.match(hook, /caught\.code === "session_full"/);
assert.match(hook, /const reserveDidNotCommit = caught\.status === 400\s*\|\| caught\.status === 401\s*\|\| caught\.status === 402\s*\|\| caught\.status === 409;/);
frontend/tests/stream-frame-buffer.test.ts
+7 -7
View File
@@ -116,11 +116,11 @@ test("many events collapse into one flush per frame and a paced settle writes th
assert.ok(flushes.at(-1)!.answer.length < 200, "pacing is still behind the network");
// 原值: buffer.settle() 同步一帧放完剩余文字（settled: true 立即到）
// 新值: 默认 settle 按打字节奏在 STREAM_SETTLE_MAX_FRAMES 内写完，最后一帧才 settled: true；immediate 才一帧放完
// 原因: TASK-consult-first-frame-and-pacing-20260928 D3 / BUG-1075：流尾一帧全放让短回答「一下全出来」
// 新值: settle({ paced: true }) 按打字节奏在 STREAM_SETTLE_MAX_FRAMES 内写完，最后一帧才 settled: true；无参 settle() 仍一帧放完（下一条测试）
// 原因: TASK-consult-first-frame-and-pacing-20260928 D3 / BUG-1075：流尾一帧全放让短回答「一下全出来」；其他调用方（校正面的快照合并）保持旧契约
const pendingAtSettle = answer.length - flushes.at(-1)!.answer.length;
let resolved = false;
void buffer.settle().then(() => { resolved = true; });
void buffer.settle({ paced: true }).then(() => { resolved = true; });
assert.equal(flushes.at(-1)!.settled, false, "a paced settle does not flush synchronously");
let settleFrames = 0;
while (fake.scheduledFrames > 0 && settleFrames < 200) {
@@ -152,7 +152,7 @@ test("typing pace: an ordinary backlog is capped at four characters a frame, a l
assert.equal(frames, 160 / STREAM_RELEASE_MAX_CHARS, "a 160-character lump is typed out over 40 frames, not 12");
});
test("a paced settle writes 300 leftover characters within the budget; immediate, stop-style settle is one flush", () => {
test("a paced settle writes 300 leftover characters within the budget; the default settle is one flush", () => {
const fake = fakeScheduler();
const flushes: StreamFrameSnapshot<null>[] = [];
const buffer = createStreamFrameBuffer<null>({
@@ -166,7 +166,7 @@ test("a paced settle writes 300 leftover characters within the budget; immediate
const shownBefore = flushes.at(-1)!.answer.length;
assert.ok(shownBefore < 300);
void buffer.settle();
void buffer.settle({ paced: true });
let frames = 0;
while (fake.scheduledFrames > 0 && frames < 500) {
fake.tick();
@@ -187,7 +187,7 @@ test("a paced settle writes 300 leftover characters within the budget; immediate
});
immediate.setAnswer(text);
const before = flushes.length;
void immediate.settle({ immediate: true });
void immediate.settle();
assert.equal(flushes.length, before + 1);
assert.equal(flushes.at(-1)!.answer, text);
assert.equal(flushes.at(-1)!.settled, true);
@@ -205,7 +205,7 @@ test("dispose during a paced settle lets it finish, then silences the buffer", (
const text = "字".repeat(40);
buffer.setAnswer(text);
fake.tick();
void buffer.settle();
void buffer.settle({ paced: true });
buffer.dispose();
let frames = 0;
while (fake.scheduledFrames > 0 && frames < 100) {