Skip to content

Commit 5bb6933

Browse files
maria-rcksclaude
andcommitted
fix(chat): stop a stranded thinking block from shimmering forever
A provider crash or a server restart mid-block leaves a reasoning message flagged as streaming with no completion coming. Both clients now treat a block as live only while its own turn is, so a turn that settled hours ago cannot keep telling the user it is thinking. Alongside that: - A completion snapshot writes a message id derived from the item, so a repeated `item.completed` rewrites that row instead of adding a second copy of the same trace. - A zero-length block reads "Thought" rather than "Thought for 1ms". - Codex raw traces get the blank line between parts that summaries already had; `contentIndex` was decoded and then ignored. - Web skips an empty trace like mobile already did. - The token-streaming setting no longer claims every token repaints, now that reasoning is deliberately paced a paragraph at a time. Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
1 parent 14e9c64 commit 5bb6933

6 files changed

Lines changed: 105 additions & 38 deletions

File tree

‎apps/mobile/src/features/threads/ThreadFeed.tsx‎

Lines changed: 8 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -1482,13 +1482,19 @@ function renderFeedEntry(
14821482
if (message.text.trim().length === 0) {
14831483
return null;
14841484
}
1485+
// Only the live turn may claim to still be thinking: a block left open by
1486+
// a crashed provider must not shimmer on a turn that settled long ago.
1487+
const liveReasoning =
1488+
Boolean(message.streaming) &&
1489+
message.turnId !== null &&
1490+
message.turnId === props.unsettledTurnId;
14851491
return (
14861492
<ThreadReasoningRow
14871493
rowSizing={props.workRowSizing}
14881494
iconSubtleColor={iconSubtleColor}
14891495
expanded={props.expandedReasoningMessageIds.has(message.id)}
1490-
label={reasoningRowLabel(message)}
1491-
streaming={Boolean(message.streaming)}
1496+
label={reasoningRowLabel(message, liveReasoning)}
1497+
streaming={liveReasoning}
14921498
onToggle={() => props.onToggleReasoning(message.id)}
14931499
>
14941500
<MarkdownImageAvailableWidthContext

‎apps/mobile/src/lib/threadActivity.ts‎

Lines changed: 8 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -1580,12 +1580,17 @@ function groupAdjacentActivities(entries: ReadonlyArray<RawThreadFeedEntry>): Th
15801580
* Row label for a provider thinking trace: live while the trace streams, then
15811581
* how long it took, in the same duration format as the "Worked for" fold.
15821582
*/
1583-
export function reasoningRowLabel(message: OrchestrationThread["messages"][number]): string {
1584-
if (message.streaming) {
1583+
export function reasoningRowLabel(
1584+
message: OrchestrationThread["messages"][number],
1585+
live: boolean,
1586+
): string {
1587+
if (live) {
15851588
return "Thinking";
15861589
}
15871590
const elapsedMs = computeElapsedMs(message.createdAt, message.updatedAt);
1588-
return elapsedMs === null ? "Thought" : `Thought for ${formatDuration(elapsedMs)}`;
1591+
return elapsedMs === null || elapsedMs <= 0
1592+
? "Thought"
1593+
: `Thought for ${formatDuration(elapsedMs)}`;
15891594
}
15901595

15911596
function computeElapsedMs(startIso: string, endIso: string): number | null {

‎apps/server/src/orchestration/Layers/ProviderRuntimeIngestion.test.ts‎

Lines changed: 23 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1528,6 +1528,29 @@ describe("ProviderRuntimeIngestion", () => {
15281528
(entry: ProviderRuntimeTestMessage) => entry.role === "reasoning",
15291529
);
15301530
expect(message?.text).toBe("reasoning reported in one piece");
1531+
1532+
// A repeated completion must rewrite that row, not add a second copy.
1533+
harness.emit({
1534+
type: "item.completed",
1535+
eventId: asEventId("evt-reasoning-snapshot-repeat"),
1536+
provider: ProviderDriverKind.make("codex"),
1537+
createdAt: now,
1538+
threadId: asThreadId("thread-1"),
1539+
turnId: asTurnId("turn-snapshot"),
1540+
itemId: asItemId("item-snapshot"),
1541+
payload: {
1542+
itemType: "reasoning",
1543+
status: "completed",
1544+
detail: "reasoning reported in one piece",
1545+
},
1546+
});
1547+
await harness.drain();
1548+
const after = await harness.readModel();
1549+
const repeated = after.threads.find((entry) => entry.id === ThreadId.make("thread-1"));
1550+
expect(
1551+
repeated?.messages.filter((entry: ProviderRuntimeTestMessage) => entry.role === "reasoning")
1552+
.length,
1553+
).toBe(1);
15311554
});
15321555

15331556
it("keeps interleaved summary and raw reasoning in separate blocks", async () => {

‎apps/server/src/orchestration/Layers/ProviderRuntimeIngestion.ts‎

Lines changed: 50 additions & 30 deletions
Original file line numberDiff line numberDiff line change
@@ -1036,10 +1036,10 @@ const make = Effect.gen(function* () {
10361036
lookup: () => Effect.succeed(""),
10371037
});
10381038

1039-
// Codex splits a reasoning summary into indexed parts. The index is the only
1040-
// signal that one part ended and the next began, so the blank line that keeps
1041-
// them readable has to be inserted here.
1042-
const reasoningSummaryIndexByMessageId = yield* Cache.make<MessageId, number>({
1039+
// Codex splits a reasoning trace into indexed parts, summary and raw alike.
1040+
// The index is the only signal that one part ended and the next began, so the
1041+
// blank line that keeps them readable has to be inserted here.
1042+
const reasoningPartIndexByMessageId = yield* Cache.make<MessageId, number>({
10431043
capacity: BUFFERED_MESSAGE_TEXT_BY_MESSAGE_ID_CACHE_CAPACITY,
10441044
timeToLive: BUFFERED_MESSAGE_TEXT_BY_MESSAGE_ID_TTL,
10451045
lookup: () => Effect.succeed(-1),
@@ -1375,7 +1375,7 @@ const make = Effect.gen(function* () {
13751375

13761376
const clearAssistantMessageState = (messageId: MessageId) =>
13771377
clearBufferedAssistantText(messageId).pipe(
1378-
Effect.andThen(Cache.invalidate(reasoningSummaryIndexByMessageId, messageId)),
1378+
Effect.andThen(Cache.invalidate(reasoningPartIndexByMessageId, messageId)),
13791379
Effect.andThen(Cache.invalidate(reasoningStartedAtByMessageId, messageId)),
13801380
);
13811381

@@ -1922,6 +1922,7 @@ const make = Effect.gen(function* () {
19221922
streamKind: event.payload.streamKind,
19231923
delta: event.payload.delta,
19241924
summaryIndex: event.payload.summaryIndex,
1925+
contentIndex: event.payload.contentIndex,
19251926
}
19261927
: undefined;
19271928
const proposedPlanDelta =
@@ -1952,19 +1953,16 @@ const make = Effect.gen(function* () {
19521953
}
19531954

19541955
let delta = reasoningDelta.delta;
1955-
if (reasoningDelta.summaryIndex !== undefined) {
1956+
const partIndex = reasoningDelta.summaryIndex ?? reasoningDelta.contentIndex;
1957+
if (partIndex !== undefined) {
19561958
const lastIndex = Option.getOrElse(
1957-
yield* Cache.getOption(reasoningSummaryIndexByMessageId, reasoningMessageId),
1959+
yield* Cache.getOption(reasoningPartIndexByMessageId, reasoningMessageId),
19581960
() => -1,
19591961
);
1960-
if (lastIndex >= 0 && lastIndex !== reasoningDelta.summaryIndex) {
1962+
if (lastIndex >= 0 && lastIndex !== partIndex) {
19611963
delta = `\n\n${delta}`;
19621964
}
1963-
yield* Cache.set(
1964-
reasoningSummaryIndexByMessageId,
1965-
reasoningMessageId,
1966-
reasoningDelta.summaryIndex,
1967-
);
1965+
yield* Cache.set(reasoningPartIndexByMessageId, reasoningMessageId, partIndex);
19681966
}
19691967

19701968
// Reasoning is never delivered token by token, even when the project
@@ -2150,28 +2148,50 @@ const make = Effect.gen(function* () {
21502148
? event.payload.detail
21512149
: undefined;
21522150

2153-
// A provider can report a whole block at once without streaming it.
2154-
// Open a segment so the snapshot has somewhere to land.
2155-
if (Option.isNone(activeReasoningMessageId) && fallbackText !== undefined) {
2156-
yield* startAssistantSegmentForTurn({
2151+
if (Option.isNone(activeReasoningMessageId)) {
2152+
// A provider can report a whole block at once without streaming it.
2153+
// The id is derived from the item rather than the segment counter so
2154+
// a repeated completion rewrites that row instead of adding a copy.
2155+
if (fallbackText !== undefined) {
2156+
const snapshotMessageId = assistantSegmentMessageId(
2157+
`snapshot:${event.itemId ?? event.eventId}`,
2158+
0,
2159+
"reasoning",
2160+
);
2161+
const existingSnapshot = yield* getThreadMessageById(thread.id, snapshotMessageId);
2162+
if (existingSnapshot === undefined) {
2163+
yield* orchestrationEngine.dispatch({
2164+
type: "thread.message.reasoning.delta",
2165+
commandId: yield* providerCommandId(event, "reasoning-delta-snapshot"),
2166+
threadId: thread.id,
2167+
messageId: snapshotMessageId,
2168+
delta: fallbackText,
2169+
turnId,
2170+
createdAt: now,
2171+
});
2172+
yield* orchestrationEngine.dispatch({
2173+
type: "thread.message.reasoning.complete",
2174+
commandId: yield* providerCommandId(event, "reasoning-complete-snapshot"),
2175+
threadId: thread.id,
2176+
messageId: snapshotMessageId,
2177+
turnId,
2178+
createdAt: now,
2179+
});
2180+
}
2181+
}
2182+
} else {
2183+
yield* finalizeActiveSegmentForTurn({
2184+
event,
21572185
threadId: thread.id,
21582186
turnId,
2159-
baseKey: reasoningSegmentBaseKeyFromEvent(event, "reasoning_text"),
2187+
createdAt: now,
2188+
commandTag: "reasoning-complete",
2189+
finalDeltaCommandTag: "reasoning-delta-finalize",
2190+
hasProjectedMessage: existingReasoningMessage !== undefined,
21602191
role: "reasoning",
2192+
...(fallbackText !== undefined ? { fallbackText } : {}),
21612193
});
21622194
}
2163-
2164-
yield* finalizeActiveSegmentForTurn({
2165-
event,
2166-
threadId: thread.id,
2167-
turnId,
2168-
createdAt: now,
2169-
commandTag: "reasoning-complete",
2170-
finalDeltaCommandTag: "reasoning-delta-finalize",
2171-
hasProjectedMessage: existingReasoningMessage !== undefined,
2172-
role: "reasoning",
2173-
...(fallbackText !== undefined ? { fallbackText } : {}),
2174-
});
21752195
}
21762196
}
21772197

‎apps/web/src/components/chat/MessagesTimeline.tsx‎

Lines changed: 14 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -2168,8 +2168,16 @@ const ReasoningTimelineRow = memo(function ReasoningTimelineRow({
21682168
row: Extract<TimelineRow, { kind: "message" }>;
21692169
}) {
21702170
const ctx = use(TimelineRowCtx);
2171+
const { isWorking, latestTurnId } = use(TimelineRowActivityCtx);
21712172
const { message } = row;
2172-
const streaming = Boolean(message.streaming);
2173+
// A block left open by a crashed provider or a restarted server never gets
2174+
// its completion. Only the live turn may claim to still be thinking, so a
2175+
// settled turn cannot shimmer "Thinking" at the user forever.
2176+
const streaming =
2177+
Boolean(message.streaming) &&
2178+
isWorking &&
2179+
message.turnId !== null &&
2180+
message.turnId === latestTurnId;
21732181
const expanded = ctx.expandedReasoningMessageIds.has(message.id);
21742182
const { onToggleReasoning } = ctx;
21752183
const toggle = useCallback(() => {
@@ -2179,11 +2187,15 @@ const ReasoningTimelineRow = memo(function ReasoningTimelineRow({
21792187
? null
21802188
: Date.parse(message.updatedAt) - Date.parse(message.createdAt);
21812189
const elapsed =
2182-
elapsedMs !== null && Number.isFinite(elapsedMs) && elapsedMs >= 0
2190+
elapsedMs !== null && Number.isFinite(elapsedMs) && elapsedMs > 0
21832191
? formatDuration(elapsedMs)
21842192
: null;
21852193
const label = streaming ? "Thinking" : elapsed ? `Thought for ${elapsed}` : "Thought";
21862194

2195+
if (message.text.trim().length === 0 && !streaming) {
2196+
return null;
2197+
}
2198+
21872199
return (
21882200
<div className={cn("flex flex-col", expanded && "mb-1")}>
21892201
<button

‎apps/web/src/components/settings/SettingsPanels.tsx‎

Lines changed: 2 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -186,7 +186,8 @@ const RESPONSE_STREAMING_MODE_LABELS: Record<ResponseStreamingMode, string> = {
186186
const RESPONSE_STREAMING_MODE_DESCRIPTIONS: Record<ResponseStreamingMode, string> = {
187187
turn: "Text appears once the agent finishes its turn.",
188188
paragraph: "Each paragraph or code block appears as soon as it is complete.",
189-
token: "Every token repaints the message as it arrives. Slower and harder to read.",
189+
token:
190+
"Every token repaints the answer as it arrives. Slower and harder to read. Thinking traces still arrive a paragraph at a time.",
190191
};
191192

192193
const TIMESTAMP_FORMAT_LABELS = {

0 commit comments

Comments
 (0)