Skip to content

armgit / cindy

Public Imported from GitHub
0

Consider it done. The open-source AI agent that works out of the box · 想到,就能做到。开源、开箱即用的 AI Agent。

fix(chat): fold failed sampling output before a successful retry

shenjiachengdb6b33075442
diff --git a/apps/desktop/src/renderer/__tests__/workGroupProjectionParity.test.ts b/apps/desktop/src/renderer/__tests__/workGroupProjectionParity.test.ts
index 86c26d5d24..2780e6cd5a 100644
--- a/apps/desktop/src/renderer/__tests__/workGroupProjectionParity.test.ts
+++ b/apps/desktop/src/renderer/__tests__/workGroupProjectionParity.test.ts
@@ -18,6 +18,7 @@ interface Event {
sealed?: boolean;
autoResume?: boolean;
delivery?: 'steer' | 'turn';
+ failedSamplingAttempt?: boolean;
}
const iso = (seconds: number) => new Date(Date.UTC(2026, 0, 1) + seconds * 1000).toISOString();
@@ -76,6 +77,7 @@ function desktopItems(events: readonly Event[]): RenderItem[] {
: {}),
createdAt: iso(event.at),
turnCompleted: event.sealed,
+ failedSamplingAttempt: event.failedSamplingAttempt,
delivery: event.delivery,
...(event.autoResume ? { systemCardType: 'auto-resume' as const } : {}),
...(event.kind === 'compact' ? { systemCardType: 'compact' as const } : {}),
@@ -95,6 +97,7 @@ function normalized(events: readonly Event[]): MessageRenderNormalizedMessage[]
body: event.body ?? (event.kind === 'thinking' ? 'Thinking' : ''),
createdAt: iso(event.at),
turnCompleted: event.sealed,
+ failedSamplingAttempt: event.failedSamplingAttempt,
settledAt: event.end === undefined ? undefined : iso(event.end),
source: {
clientId: event.id,
@@ -161,6 +164,48 @@ function tree(items: readonly Projection[]): unknown[] {
}
const cases: Array<{ name: string; events: Event[]; streaming?: boolean; expected: unknown[] }> = [
+ // #6351: both platform consumers must fold failed native attempts without
+ // waiting for a successful final. Normal multi-message answers stay intact.
+ {
+ name: 'failed sampling attempts fold while the native retry is still running',
+ streaming: true,
+ events: [user(),
+ { ...answer('failed1', 1, true, '# Opening'), failedSamplingAttempt: true },
+ { ...answer('failed2', 2), failedSamplingAttempt: true }],
+ expected: [['u'], ['work-failed1', [['failed1'], ['failed2']]]],
+ },
+ {
+ name: 'exhausted or stopped sampling attempts fold without a successful final',
+ events: [user(),
+ { ...answer('failed1', 1, true, '# Opening'), failedSamplingAttempt: true },
+ { ...answer('failed2', 2), failedSamplingAttempt: true },
+ { id: 'failure', kind: 'error', at: 3, body: 'Reply interrupted' }],
+ expected: [['u'], ['work-failed1', [['failed1'], ['failed2']]], ['failure']],
+ },
+ {
+ name: 'a successful retry preserves all normal contiguous assistant answer blocks',
+ events: [user(),
+ { ...answer('failed1', 1), failedSamplingAttempt: true },
+ { ...answer('failed2', 2), failedSamplingAttempt: true },
+ answer('intro', 3), answer('final', 4, true)],
+ expected: [['u'], ['work-failed1', [['failed1'], ['failed2']]], ['intro'], ['final']],
+ },
+ {
+ name: 'manual recovery leaves the prior failed fold and error in their own user turn',
+ events: [user(),
+ { ...answer('failed1', 1), failedSamplingAttempt: true },
+ { id: 'failure', kind: 'error', at: 2, body: 'Reply interrupted' },
+ user('recovery', 3), answer('final', 4, true)],
+ expected: [['u'], ['work-failed1', [['failed1']]], ['failure'], ['recovery'], ['final']],
+ },
+ {
+ name: 'steering stays ordered between failed samples and the successful reply',
+ events: [user(),
+ { ...answer('failed1', 1), failedSamplingAttempt: true },
+ { ...user('steer', 2), delivery: 'steer' },
+ { ...answer('failed2', 3), failedSamplingAttempt: true }, answer('final', 4, true)],
+ expected: [['u'], ['work-failed1', [['failed1']]], ['steer'], ['work-failed2', [['failed2']]], ['final']],
+ },
{
name: 'running subagent remains visible between completed work groups',
events: [
diff --git a/apps/desktop/src/renderer/components/chat/messageWorkGroups.ts b/apps/desktop/src/renderer/components/chat/messageWorkGroups.ts
index 4ec5557229..57807ca2cb 100644
--- a/apps/desktop/src/renderer/components/chat/messageWorkGroups.ts
+++ b/apps/desktop/src/renderer/components/chat/messageWorkGroups.ts
@@ -171,8 +171,8 @@ export function isCompletedAssistantMessage(message: ChatMessage): boolean {
}
/** 完成态 work_group 可合并的子项:tool_segment / agent_task / thinking /
- * assistant 工作文字。运行态只通过 isWorkActivityItem 收动作,所以不会提前
- * 折叠正在输出的 assistant 文字。 */
+ * assistant 工作文字。正常流式正文不会提前折叠;已证实失败的采样输出
+ * 复用活动折叠,无需等待成功答复出现。 */
function isWorkChild(it: RenderItem): it is WorkChildItem {
return (
it.type === 'tool_segment' ||
@@ -212,8 +212,8 @@ function isWorkflowTaskItem(it: RenderItem): boolean {
);
}
-/** preview 中计为一条真实活动的 render item。assistant 进度文字
- * 始终留在主消息流,不占最近 5 条活动窗口。 */
+/** Normal assistant progress stays in the main stream. SDK-proven failed
+ * sampling output uses the existing work fold even before a retry succeeds. */
function isWorkActivityItem(it: RenderItem): it is WorkChildItem {
return (
!isRunningAgentTask(it) &&
@@ -221,7 +221,8 @@ function isWorkActivityItem(it: RenderItem): it is WorkChildItem {
!isWorkflowTaskItem(it) &&
(it.type === 'tool_segment' ||
it.type === 'agent_task' ||
- (it.type === 'message' && it.message.role === 'thinking'))
+ (it.type === 'message' && (it.message.role === 'thinking' ||
+ (it.message.role === 'assistant' && it.message.failedSamplingAttempt === true))))
);
}
@@ -491,7 +492,7 @@ function createCompletedWorkGroup(
prevBoundaryTs: number | null = null,
): WorkGroupRenderItem {
const hasAssistantText = run.some(
- (item) => item.type === 'message' && item.message.role === 'assistant',
+ (item) => item.type === 'message' && item.message.role === 'assistant' && !item.message.failedSamplingAttempt,
);
if (!hasAssistantText) return createWorkGroup(run, nextItem, false, prevBoundaryTs);
diff --git a/apps/desktop/src/renderer/features/bots/__tests__/botConversationPresentation.test.ts b/apps/desktop/src/renderer/features/bots/__tests__/botConversationPresentation.test.ts
index 9f5e43dc9b..66177c97b1 100644
--- a/apps/desktop/src/renderer/features/bots/__tests__/botConversationPresentation.test.ts
+++ b/apps/desktop/src/renderer/features/bots/__tests__/botConversationPresentation.test.ts
@@ -25,17 +25,25 @@ describe('teammate final-result presentation', () => {
// #6351: provider final_answer only closes a message, not a successful sample.
// Failed output remains in the expandable work history rather than being lifted
// beside the real answer; normal commentary is governed by the existing rules.
- it('retains failed sampling output in work history without lifting it as another successful reply', () => {
- const result = project([
- message('u', 'user'),
- message('failed', 'assistant', 'Opening from failed attempt', {
- assistantPhase: 'final_answer', failedSamplingAttempt: true, turnCostUsd: 0.1,
- }),
- message('success', 'assistant', 'Successful opening', { assistantPhase: 'final_answer', turnCompleted: true }),
- ], false);
- expect(proseIds(result)).toEqual(['success']);
- expect(allKeys(result)).toContain('msg-failed');
- });
+ it.each(['retrying', 'completed', 'exhausted', 'stopped'] as const)(
+ 'retains failed sampling output in work history when the retry is %s', (ending) => {
+ const result = project([
+ message('u', 'user'),
+ message('failed', 'assistant', 'Opening from failed attempt', {
+ assistantPhase: 'final_answer', failedSamplingAttempt: true, turnCostUsd: 0.1,
+ }),
+ message('failed-commentary', 'assistant', 'Preparing another opening', {
+ assistantPhase: 'commentary', failedSamplingAttempt: true,
+ }),
+ ...(ending === 'completed'
+ ? [message('intro', 'assistant', 'Accepted intro'),
+ message('success', 'assistant', 'Successful opening', { assistantPhase: 'final_answer', turnCompleted: true })]
+ : ending === 'retrying' ? [] : [message('error', 'error', 'Reply interrupted')]),
+ ], ending === 'retrying');
+ expect(proseIds(result)).toEqual(ending === 'completed' ? ['intro', 'success'] : []);
+ expect(allKeys(result)).toContain('msg-failed');
+ expect(allKeys(result)).toContain('msg-failed-commentary');
+ });
it('keeps coordination input hidden while process, explicit inputs, results and errors remain visible', () => {
const rows = [message('coordination', 'user', '[UI_ACTION_TRIGGER]File agreement', { isSyntheticTrigger: true }),
tool('audit'), message('visible-input', 'user', 'Requested status'),
diff --git a/apps/mobile/src/__tests__/companionConversationPresentation.test.ts b/apps/mobile/src/__tests__/companionConversationPresentation.test.ts
index b7e367501d..393f2dd49b 100644
--- a/apps/mobile/src/__tests__/companionConversationPresentation.test.ts
+++ b/apps/mobile/src/__tests__/companionConversationPresentation.test.ts
@@ -11,6 +11,30 @@ const bodies = (messages: RemoteMessage[], running: boolean) => companionConvers
const base = [row('u0', 'user', 'Help'), row('a1', 'assistant', 'Public progress'),
row('t2', 'tool_use', { toolName: 'Read', toolUseId: 't', input: { path: 'private' } }),
row('r3', 'tool_result', { toolUseId: 't', result: 'technical details' })];
+// #6351: this consumer expands work groups and infers contiguous sealed runs,
+// so grouping alone cannot prevent a failed sample becoming another reply.
+it.each(['retrying', 'completed', 'exhausted'] as const)(
+ 'does not promote failed sampling prose in a %s companion turn', (ending) => {
+ const failed = row('a1', 'assistant', '# Failed opening\n' + 'Details '.repeat(100), {
+ turnCompleted: true, turnCostUsd: 0.1, failedSamplingAttempt: {
+ threadId: 'thread', turnId: 'turn', attempt: 1, itemId: 'failed',
+ },
+ });
+ const source = [row('u0', 'user', 'Help'), failed,
+ ...(ending === 'completed' ? [row('a2', 'assistant', 'Accepted intro'),
+ row('a3', 'assistant', 'Accepted answer', { turnCompleted: true })]
+ : ending === 'exhausted' ? [row('e2', 'error', 'Model failed')] : []),
+ ];
+ const visible = bodies(source, ending === 'retrying')
+ .filter(item => item.type === 'message').map(item => item.message);
+ expect(visible.filter(item => item.kind === 'assistant').map(item => item.body))
+ .toEqual(ending === 'completed' ? ['Accepted intro', 'Accepted answer'] : []);
+ expect(visible.some(item => item.source.id === 'u0')).toBe(true);
+ expect(visible.some(item => item.source.role === 'error')).toBe(ending === 'exhausted');
+ expect(source[1]).toBe(failed);
+ expect(source[1].content).toContain('# Failed opening');
+ expect(source[1].agentMeta?.failedSamplingAttempt).toBeDefined();
+ });
it('has no process entry or commentary during generation, and retains the completed answer', () => {
const active = bodies(base, true);
expect(active.map(item => item.type)).toEqual(['message']);
diff --git a/apps/mobile/src/__tests__/messageRenderModel.test.ts b/apps/mobile/src/__tests__/messageRenderModel.test.ts
index 98c2107ceb..463ec6387b 100644
--- a/apps/mobile/src/__tests__/messageRenderModel.test.ts
+++ b/apps/mobile/src/__tests__/messageRenderModel.test.ts
@@ -39,6 +39,36 @@ function toolUse(id: string, toolName: string, input: unknown, seconds: number):
}
describe('messageRenderModel', () => {
+ // #6351: remote JSON must reach the actual work fold, not only normalization.
+ // Two completed items from failed samples must not look like accepted replies.
+ it('folds remote failed samples before success and retains their original records', () => {
+ const messages = [
+ message({ id: 'u', role: 'user', content: 'Start', createdAt: at(1) }),
+ ...[1, 2].map((attempt) => message({
+ id: `failed${attempt}`, role: 'assistant', content: '# Opening', createdAt: at(attempt + 1),
+ agentMeta: { turnCompleted: true, turnCostUsd: 0.1, failedSamplingAttempt: {
+ threadId: 'thread', turnId: 'turn', attempt, itemId: `item${attempt}`,
+ } },
+ })),
+ ];
+ for (const isSessionStreaming of [true, false]) {
+ const items = buildMobileMessageRenderItems(messages, { isSessionStreaming });
+ expect(items.map((item) => item.key)).toEqual(['message-u', 'work-failed1']);
+ expect(items[1]).toMatchObject({ type: 'work_group', children: [
+ { type: 'message', message: { source: messages[1], failedSamplingAttempt: true } },
+ { type: 'message', message: { source: messages[2], failedSamplingAttempt: true } },
+ ] });
+ }
+ const completed = buildMobileMessageRenderItems([...messages,
+ message({ id: 'intro', role: 'assistant', content: 'Accepted intro', createdAt: at(4) }),
+ message({ id: 'final', role: 'assistant', content: 'Accepted answer', createdAt: at(5),
+ agentMeta: { turnCompleted: true } }),
+ ], { isSessionStreaming: false });
+ expect(completed.map((item) => item.key)).toEqual([
+ 'message-u', 'work-failed1', 'message-intro', 'message-final',
+ ]);
+ });
+
it('keeps completed work folded when the next remote run starts before its user row arrives', () => {
const messages = [
message({ id: 'u1', role: 'user', content: 'start', createdAt: at(1) }),
diff --git a/apps/mobile/src/session/companionConversationPresentation.ts b/apps/mobile/src/session/companionConversationPresentation.ts
index 8de2f09dcc..48ae8611ad 100644
--- a/apps/mobile/src/session/companionConversationPresentation.ts
+++ b/apps/mobile/src/session/companionConversationPresentation.ts
@@ -4,7 +4,7 @@ import { collectMobileMarkdownImages } from './messageMarkdown';
import type { MobileMessageRenderItem } from './messageRenderModel';
const isDelivery = (item: MobileMessageRenderItem) => item.type === 'tool_media' && item.tools.some(tool => !!tool.media?.length || !!tool.files?.length)
- || item.type === 'message' && (item.message.companion?.kind === 'task' && item.message.companion.meta.role === 'delegation-result'
+ || item.type === 'message' && !item.message.failedSamplingAttempt && (item.message.companion?.kind === 'task' && item.message.companion.meta.role === 'delegation-result'
|| !!item.message.attachments?.length || !!item.message.media?.length
|| !!item.message.files?.length || collectMobileMarkdownImages(item.message.body).length > 0);
@@ -19,7 +19,7 @@ export function companionConversationItems(items: readonly MobileMessageRenderIt
let sealedRun = false;
let nextAssistantAt: number | null = null;
for (const item of expandWorkGroups(items).reverse()) {
- if (item.type !== 'message' || item.message.kind !== 'assistant' || !item.message.body.trim() || item.message.systemCardType) {
+ if (item.type !== 'message' || item.message.kind !== 'assistant' || item.message.failedSamplingAttempt || !item.message.body.trim() || item.message.systemCardType) {
sealedRun = false;
nextAssistantAt = null;
} else {
@@ -53,7 +53,7 @@ export function companionConversationItems(items: readonly MobileMessageRenderIt
else if (delivery) deliveredAfter.add(item.key);
}
const attached = new Set(flattened.flatMap(item => item.type === 'message'
- && item.message.kind === 'assistant' && item.message.turnCompleted === true && item.message.body.trim()
+ && item.message.kind === 'assistant' && !item.message.failedSamplingAttempt && item.message.turnCompleted === true && item.message.body.trim()
? readBotTaskResults(item.message.source.agentMeta?.botTaskResults).map(botTaskResultKey) : []));
return flattened.filter(item => {
if (item.type === 'message' && item.message.companion?.kind === 'task'
@@ -63,6 +63,9 @@ export function companionConversationItems(items: readonly MobileMessageRenderIt
|| item.type === 'subagent_group' || item.type === 'todo') return false;
if (item.type !== 'message') return true;
const message = item.message;
+ // Companion chats omit process prose; the original failed rows remain in
+ // persisted history and the ordinary task's expandable work group.
+ if (message.kind === 'assistant' && message.failedSamplingAttempt) return false;
// Desktop drops persisted synthetic resume separators from teammate chats; only the live
// reconnect card (still in progress) remains. The composer status carries recovery copy.
if (message.isSyntheticTrigger && message.systemCardType === 'auto-resume') {
diff --git a/packages/maker-core/src/agents/codex/sampling-retry.integration.test.ts b/packages/maker-core/src/agents/codex/sampling-retry.integration.test.ts
index e4450d5c93..16105c3959 100644
--- a/packages/maker-core/src/agents/codex/sampling-retry.integration.test.ts
+++ b/packages/maker-core/src/agents/codex/sampling-retry.integration.test.ts
@@ -44,12 +44,19 @@ it.skipIf(!binaryPath).each(['success', 'exhausted', 'stop', 'steer'] as const)(
for await (const chunk of req) chunks.push(Buffer.from(chunk));
requests.push(JSON.parse(Buffer.concat(chunks).toString('utf8')));
const request = requests.length;
+ const willComplete = recover || request > 2 || (request === 2 && ending === 'success');
res.writeHead(200, { 'content-type': 'text/event-stream', connection: 'close' });
sendEvent(res, { type: 'response.created', response: { id: 'response-' + request } });
+ if (willComplete) {
+ for (const [id, phase] of [['accepted-commentary-', 'commentary'], ['accepted-intro-', 'final_answer']] as const) {
+ sendEvent(res, { type: 'response.output_item.done', item: { type: 'message', role: 'assistant',
+ id: id + request, phase, content: [{ type: 'output_text', text: id + request }] } });
+ }
+ }
sendEvent(res, { type: 'response.output_item.done', item: { type: 'message', role: 'assistant',
id: 'answer-' + request, phase: 'final_answer',
content: [{ type: 'output_text', text: 'Opening from request ' + request }] } });
- if (recover || request > 2 || (request === 2 && ending === 'success')) {
+ if (willComplete) {
complete(res, request);
} else if (request === 2 && (ending === 'stop' || ending === 'steer')) {
heldResponse = res;
@@ -138,6 +145,14 @@ it.skipIf(!binaryPath).each(['success', 'exhausted', 'stop', 'steer'] as const)(
(event.data as { text?: string }).text)).toContain('Opening from request 1');
expect(events.some(event => event.type === 'error' &&
(event.data as { willRetry?: boolean }).willRetry === true)).toBe(true);
+ if (ending === 'success' || ending === 'steer') {
+ const acceptedRequest = ending === 'success' ? 2 : 3;
+ const acceptedIds = ['accepted-commentary-' + acceptedRequest, 'accepted-intro-' + acceptedRequest];
+ const textIds = events.filter(event => event.type === 'text').map(event =>
+ (event.data as { agentMessageId?: string }).agentMessageId);
+ expect(textIds).toEqual(expect.arrayContaining(acceptedIds));
+ expect(failures.flatMap(failure => failure.agentMessageIds).some(id => acceptedIds.includes(id))).toBe(false);
+ }
const failureCount = failures.length;
recover = true;
const boundary = events.length;
diff --git a/packages/maker-core/src/agents/codex/translator.test.ts b/packages/maker-core/src/agents/codex/translator.test.ts
index e285d1e32d..2f5f0e4169 100644
--- a/packages/maker-core/src/agents/codex/translator.test.ts
+++ b/packages/maker-core/src/agents/codex/translator.test.ts
@@ -122,9 +122,16 @@ describe('Codex assistant text streaming contract', () => {
translateErrorNotification(makeParams({ willRetry: true,
message: 'stream disconnected - retrying sampling request (1/5 in 212ms)...',
codexErrorInfo: { responseStreamDisconnected: {} } }), q, ctx);
- complete('successful-answer', 'Hello from attempt two', 'final_answer');
+ complete('successful-commentary', 'Preparing the accepted answer', 'commentary');
+ complete('successful-intro', 'Hello from attempt two', 'final_answer');
+ complete('successful-answer', 'The accepted answer continues', 'final_answer');
+ expect(finishCodexSamplingTurn(rt, 't1', 'turn-a', 'completed')).toBeUndefined();
const events = await collect(q);
- expect(events.filter(event => event.type === 'text')).toHaveLength(4);
+ expect(events.filter(event => event.type === 'text').map(event =>
+ (event.data as { agentMessageId?: string }).agentMessageId)).toEqual([
+ 'progress', 'failed-commentary', 'failed-answer',
+ 'successful-commentary', 'successful-intro', 'successful-answer',
+ ]);
expect(events.find(event => event.type === 'error')).toMatchObject({
data: { isTerminal: false, willRetry: true, assistantAttemptFailure: {
threadId: 't1', turnId: 'turn-a', attempt: 1,
diff --git a/packages/maker-shared/src/messageRender.ts b/packages/maker-shared/src/messageRender.ts
index 7d63cee2fe..ae9eaee328 100644
--- a/packages/maker-shared/src/messageRender.ts
+++ b/packages/maker-shared/src/messageRender.ts
@@ -105,6 +105,8 @@ export interface MessageRenderNormalizedMessage<
agentTaskStatus?: AgentTaskTerminalStatus;
/** Host 在 SDK done 边界写入。同一 user turn 的最后一次 seal 是最终答复;更早的 seal 只在带交付内容时留在工作过程外。 */
turnCompleted?: boolean;
+ /** SDK-proven failed sampling output remains readable inside the work fold. */
+ failedSamplingAttempt?: boolean;
/** tool 消息专用:配对 tool_result 提取出的产出媒体(驱动 tool_media 独立渲染项)。 */
media?: readonly MessageRenderToolMediaLike[];
files?: readonly { url: string; title: string }[];
@@ -1573,7 +1575,7 @@ function isRunningAgentTaskItem<
}
function isCompletedAssistantMessage(message: MessageRenderNormalizedMessage): boolean {
- return message.turnCompleted === true;
+ return !message.failedSamplingAttempt && message.turnCompleted === true;
}
function isAssistantAnswerCandidate<
@@ -1581,6 +1583,7 @@ function isAssistantAnswerCandidate<
>(item: MessageRenderItem<TMessage>): item is MessageRenderMessageItem<TMessage> {
return item.type === 'message'
&& item.message.kind === 'assistant'
+ && !item.message.failedSamplingAttempt
&& item.message.body.trim().length > 0;
}
@@ -1654,6 +1657,7 @@ function isDeliveryProseItem<TMessage extends MessageRenderNormalizedMessage>(
): boolean {
return item.type === 'message'
&& item.message.kind === 'assistant'
+ && !item.message.failedSamplingAttempt
&& isDeliveryProseText(item.message.body);
}
@@ -1672,12 +1676,13 @@ function isWorkChild<
);
}
-/** Assistant progress text remains visible while running and never consumes the latest-five window. */
+/** Normal progress stays visible; proven failed sampling output uses the existing fold. */
function isWorkActivityItem<TMessage extends MessageRenderNormalizedMessage>(
item: MessageRenderItem<TMessage>,
-): item is MessageRenderThinkingItem<TMessage> | MessageRenderToolGroupItem<TMessage> | MessageRenderAgentTaskItem<TMessage> {
+): item is Exclude<MessageRenderWorkChildItem<TMessage>, MessageRenderWorkGroupItem<TMessage>> {
return !isRunningAgentTaskItem(item)
- && (item.type === 'thinking' || item.type === 'tool_group' || item.type === 'agent_task');
+ && (item.type === 'thinking' || item.type === 'tool_group' || item.type === 'agent_task' ||
+ (item.type === 'message' && item.message.kind === 'assistant' && item.message.failedSamplingAttempt === true));
}
/** 边界项(用户消息 / assistant 正文)的时间戳;非 message 卡片不作为时间边界。 */
@@ -1724,7 +1729,7 @@ function createCompletedWorkGroup<TMessage extends MessageRenderNormalizedMessag
previousBoundaryMs: number | null = null,
): MessageRenderWorkGroupItem<TMessage> {
const hasAssistantText = run.some(
- (item) => item.type === 'message' && item.message.kind === 'assistant',
+ (item) => item.type === 'message' && item.message.kind === 'assistant' && !item.message.failedSamplingAttempt,
);
if (!hasAssistantText) return createWorkGroup(run, nextItem, false, previousBoundaryMs);