diff --git a/packages/agent-core-v2/docs/state-manifest.d.ts b/packages/agent-core-v2/docs/state-manifest.d.ts
index a180bc46635..1665d43d976 100644
--- a/packages/agent-core-v2/docs/state-manifest.d.ts
+++ b/packages/agent-core-v2/docs/state-manifest.d.ts
@@ -1052,6 +1052,7 @@ export interface AgentStateSnapshot {
       type: 'think';
       think: string;
       encrypted?: string;
+      detailsIndex?: number;
     } | /* ImageURLPart — packages/agent-core-v2/src/human/llm/message.ts */ {
       type: 'image_url';
       imageUrl: {
@@ -1341,6 +1342,7 @@ export interface AgentStateSnapshot {
     type: 'think';
     think: string;
     encrypted?: string;
+    detailsIndex?: number;
   } | /* ImageURLPart — packages/agent-core-v2/src/human/llm/message.ts */ {
     type: 'image_url';
     imageUrl: {
diff --git a/packages/agent-core-v2/src/agent/loop/loopService.ts b/packages/agent-core-v2/src/agent/loop/loopService.ts
index 7751e8d5add..095e927558e 100644
--- a/packages/agent-core-v2/src/agent/loop/loopService.ts
+++ b/packages/agent-core-v2/src/agent/loop/loopService.ts
@@ -809,7 +809,12 @@ export class AgentLoopService extends Disposable implements IAgentLoopService {
             );
             return;
           case 'thinking':
-            this.accumulateMachinePart(turn, { type: 'think', think: delta.delta });
+            this.accumulateMachinePart(turn, {
+              type: 'think',
+              think: delta.delta,
+              encrypted: delta.encrypted,
+              detailsIndex: delta.detailsIndex,
+            });
             void this.dispatcher.dispatch(
               new ThinkingDelta({ agentId: this.scopeContext.agentId, turnId: turn.id, delta: delta.delta }),
             );
@@ -1085,7 +1090,16 @@ export class AgentLoopService extends Disposable implements IAgentLoopService {
   }
 
   private drainMachinePartials(turn: ActiveTurn, step: MachineStepState): void {
-    for (const part of turn.partials.splice(0).filter((entry) => !isVacuousContentPart(entry))) {
+    const drained = turn.partials.splice(0).filter((entry) => !isVacuousContentPart(entry));
+    let lastCompleteThink = -1;
+    for (const [index, part] of drained.entries()) {
+      if (part.type === 'think' && part.encrypted !== undefined) {
+        lastCompleteThink = index;
+      }
+    }
+    for (const part of drained.filter(
+      (part, index) => part.type !== 'think' || index <= lastCompleteThink,
+    )) {
       this.context.appendLoopEvent({
         type: 'content.part',
         uuid: randomUUID(),
diff --git a/packages/agent-core-v2/src/agent/loop/machine/engine.ts b/packages/agent-core-v2/src/agent/loop/machine/engine.ts
index 31011d0492a..7219f46b325 100644
--- a/packages/agent-core-v2/src/agent/loop/machine/engine.ts
+++ b/packages/agent-core-v2/src/agent/loop/machine/engine.ts
@@ -21,7 +21,12 @@ import { createMachineTools, type ToolResultExtras } from './tools';
 
 export type MachineEngineDelta =
   | { readonly kind: 'assistant'; readonly delta: string }
-  | { readonly kind: 'thinking'; readonly delta: string }
+  | {
+      readonly kind: 'thinking';
+      readonly delta: string;
+      readonly encrypted?: string;
+      readonly detailsIndex?: number;
+    }
   | {
       readonly kind: 'toolCall';
       readonly toolCallId: string;
@@ -151,7 +156,12 @@ function createDeltaSplitter(): (part: StreamedMessagePart) => MachineEngineDelt
       case 'text':
         return { kind: 'assistant', delta: part.text };
       case 'think':
-        return { kind: 'thinking', delta: part.think };
+        return {
+          kind: 'thinking',
+          delta: part.think,
+          encrypted: part.encrypted,
+          detailsIndex: part.detailsIndex,
+        };
       case 'image_url':
       case 'audio_url':
       case 'video_url':
diff --git a/packages/agent-core-v2/src/human/llm/message.ts b/packages/agent-core-v2/src/human/llm/message.ts
index eb11bc446ba..9c3aa5600f5 100644
--- a/packages/agent-core-v2/src/human/llm/message.ts
+++ b/packages/agent-core-v2/src/human/llm/message.ts
@@ -16,6 +16,7 @@ export interface ThinkPart {
   type: 'think';
   think: string;
   encrypted?: string;
+  detailsIndex?: number;
 }
 
 export interface ImageURLPart {
@@ -103,6 +104,9 @@ export function mergeInPlace(target: StreamedMessagePart, source: StreamedMessag
     if (target.encrypted !== undefined) {
       return false;
     }
+    if (target.detailsIndex !== source.detailsIndex) {
+      return false;
+    }
     target.think += source.think;
     if (source.encrypted !== undefined) {
       target.encrypted = source.encrypted;
diff --git a/packages/agent-core-v2/src/human/llm/requester/bases/openai/format.ts b/packages/agent-core-v2/src/human/llm/requester/bases/openai/format.ts
index 4eb25413dac..963958b2531 100644
--- a/packages/agent-core-v2/src/human/llm/requester/bases/openai/format.ts
+++ b/packages/agent-core-v2/src/human/llm/requester/bases/openai/format.ts
@@ -36,7 +36,12 @@ import type { TokenUsage } from '#/llm/usage';
 
 import { lowerMessage, type OpenAIWireMessage } from './lower';
 import { extractToolMedia } from './patterns';
-import { DEFAULT_REASONING_KEY, extractReasoning } from './reasoning-key';
+import {
+  convertReasoningDetails,
+  DEFAULT_REASONING_KEY,
+  extractReasoning,
+  extractReasoningDetails,
+} from './reasoning-key';
 
 function responseFormatToOpenAI(format: ResponseFormat): Record<string, unknown> {
   if (format.type === 'json_object') {
@@ -265,6 +270,7 @@ export const openAIFormat: ProtocolFormat<OpenAIRequestParams, RawResponse, RawC
   },
 
   createStreamParser(options?: StreamParserOptions) {
+    const explicitReasoningKey = options?.trait?.reasoningKey?.(options.ctx);
     const bufferedToolCalls = new Map<number | string, BufferedStreamToolCall>();
 
     function convertStreamToolCall(toolCall: RawStreamToolCallDelta): StreamedMessagePart[] {
@@ -351,9 +357,17 @@ export const openAIFormat: ProtocolFormat<OpenAIRequestParams, RawResponse, RawC
       if (!delta) {
         return;
       }
-      const reasoning = extractReasoning(delta);
-      if (reasoning !== undefined) {
-        sink.onDelta({ type: 'think', think: reasoning.value });
+      const reasoningDetails =
+        explicitReasoningKey === undefined ? extractReasoningDetails(delta) : undefined;
+      if (reasoningDetails !== undefined) {
+        for (const part of convertReasoningDetails(reasoningDetails)) {
+          sink.onDelta(part);
+        }
+      } else {
+        const reasoning = extractReasoning(delta);
+        if (reasoning !== undefined) {
+          sink.onDelta({ type: 'think', think: reasoning.value });
+        }
       }
       if (typeof delta.content === 'string' && delta.content.length > 0) {
         sink.onDelta({ type: 'text', text: delta.content });
diff --git a/packages/agent-core-v2/src/human/llm/requester/bases/openai/lower.ts b/packages/agent-core-v2/src/human/llm/requester/bases/openai/lower.ts
index 865db526b9b..4f72dff3109 100644
--- a/packages/agent-core-v2/src/human/llm/requester/bases/openai/lower.ts
+++ b/packages/agent-core-v2/src/human/llm/requester/bases/openai/lower.ts
@@ -2,6 +2,7 @@ import { extractText, type ContentPart, type Message } from '#/llm/message';
 import type { ProtocolTrait, TraitContext } from '#/llm/protocol/trait';
 
 import { TOOL_RESULT_MEDIA_PLACEHOLDER } from './patterns';
+import { DEFAULT_REASONING_KEY, REASONING_DETAILS_KEY } from './reasoning-key';
 
 export type OpenAIContentPart = {
   type: 'text' | 'image_url' | 'audio_url' | 'video_url';
@@ -140,7 +141,20 @@ export function lowerMessage(message: Message, lower: OpenAILowerContext): OpenA
   } else {
     converted = { role: message.role, content: content ?? '' };
   }
-  if (hasReasoningPart || (preserveThinking && message.role === 'assistant')) {
+  const reasoningDetails: Record<string, unknown>[] = [];
+  for (const part of message.content) {
+    if (part.type !== 'think' || part.detailsIndex === undefined) continue;
+    if (part.think.length > 0) {
+      reasoningDetails.push({ type: 'summary', summary: part.think });
+    }
+    if (part.encrypted !== undefined) {
+      reasoningDetails.push({ type: 'encrypted', encrypted: part.encrypted });
+    }
+  }
+  if (reasoningDetails.length > 0) {
+    (converted as Record<string, unknown>)[REASONING_DETAILS_KEY] = reasoningDetails;
+    (converted as Record<string, unknown>)[DEFAULT_REASONING_KEY] = reasoningContent;
+  } else if (hasReasoningPart || (preserveThinking && message.role === 'assistant')) {
     (converted as Record<string, unknown>)[reasoningKey] = reasoningContent;
   }
   const hooked =
diff --git a/packages/agent-core-v2/src/human/llm/requester/bases/openai/reasoning-key.ts b/packages/agent-core-v2/src/human/llm/requester/bases/openai/reasoning-key.ts
index 064c4402cf9..bf05cafb7d0 100644
--- a/packages/agent-core-v2/src/human/llm/requester/bases/openai/reasoning-key.ts
+++ b/packages/agent-core-v2/src/human/llm/requester/bases/openai/reasoning-key.ts
@@ -1,3 +1,5 @@
+import type { StreamedMessagePart, ThinkPart } from '#/llm/message';
+
 export const KNOWN_REASONING_KEYS = [
   'reasoning_content',
   'reasoning_details',
@@ -40,3 +42,60 @@ export class ReasoningKeyDialect {
     return this._explicitKey ?? this._detected ?? DEFAULT_REASONING_KEY;
   }
 }
+
+export const REASONING_DETAILS_KEY = 'reasoning_details';
+
+export interface ReasoningDetailsElement {
+  readonly type?: string;
+  readonly index: number;
+  readonly summary?: string;
+  readonly encrypted?: string;
+}
+
+function toReasoningDetailsElement(
+  value: unknown,
+  position: number,
+): ReasoningDetailsElement | undefined {
+  if (typeof value !== 'object' || value === null) return undefined;
+  const record = value as Record<string, unknown>;
+  const type = typeof record['type'] === 'string' ? record['type'] : undefined;
+  if (type !== undefined && type !== 'summary' && type !== 'encrypted') return undefined;
+  const index = typeof record['index'] === 'number' ? record['index'] : position;
+  const summary = typeof record['summary'] === 'string' ? record['summary'] : undefined;
+  const encrypted = typeof record['encrypted'] === 'string' ? record['encrypted'] : undefined;
+  return { type, index, summary, encrypted };
+}
+
+export function extractReasoningDetails(
+  source: unknown,
+): ReasoningDetailsElement[] | undefined {
+  if (typeof source !== 'object' || source === null) return undefined;
+  const value = (source as Record<string, unknown>)[REASONING_DETAILS_KEY];
+  if (!Array.isArray(value)) return undefined;
+  const elements: ReasoningDetailsElement[] = [];
+  for (const [position, item] of value.entries()) {
+    const element = toReasoningDetailsElement(item, position);
+    if (element !== undefined) elements.push(element);
+  }
+  return elements;
+}
+
+export function convertReasoningDetails(
+  elements: readonly ReasoningDetailsElement[],
+): StreamedMessagePart[] {
+  const parts: StreamedMessagePart[] = [];
+  for (const element of elements) {
+    if (element.type !== 'encrypted' && element.summary !== undefined && element.summary.length > 0) {
+      parts.push({ type: 'think', think: element.summary, detailsIndex: element.index } satisfies ThinkPart);
+    }
+    if (element.type !== 'summary' && element.encrypted !== undefined && element.encrypted.length > 0) {
+      parts.push({
+        type: 'think',
+        think: '',
+        encrypted: element.encrypted,
+        detailsIndex: element.index,
+      } satisfies ThinkPart);
+    }
+  }
+  return parts;
+}
diff --git a/packages/agent-core-v2/src/human/test/llm/thinking.test.ts b/packages/agent-core-v2/src/human/test/llm/thinking.test.ts
index ad719f5cbf3..bd0e741bece 100644
--- a/packages/agent-core-v2/src/human/test/llm/thinking.test.ts
+++ b/packages/agent-core-v2/src/human/test/llm/thinking.test.ts
@@ -335,7 +335,7 @@ describe('openai requester thinking', () => {
     expect(client.body()['reasoning_effort']).toBe('medium');
   });
 
-  it('echoes think parts under reasoning_content by default', async () => {
+  it('echoes think parts under reasoning_content by default and restores marked reasoning_details', async () => {
     const client = stubOpenAIClient(chatCompletionChunks());
     const requester = createOpenAIRequester(undefined, { clientFactory: client.clientFactory });
     await requester.generate(
@@ -351,6 +351,33 @@ describe('openai requester thinking', () => {
     const assistant = bodyMessages(client.body())[1]!;
     expect(assistant['reasoning_content']).toBe('abc');
     expect(assistant['content']).toBe('hello');
+
+    const marked = stubOpenAIClient(chatCompletionChunks());
+    const markedRequester = createOpenAIRequester(undefined, {
+      clientFactory: marked.clientFactory,
+    });
+    await markedRequester.generate(
+      { model, thinking: { effort: 'off' } },
+      {
+        messages: [
+          createUserMessage('hi'),
+          createAssistantMessage([
+            { type: 'think', think: '第一段续', detailsIndex: 0 },
+            { type: 'think', think: '第二段', detailsIndex: 1 },
+            { type: 'think', think: '', encrypted: 'cipher', detailsIndex: 2 },
+            { type: 'text', text: 'ok' },
+          ]),
+        ],
+      },
+      { signal: new AbortController().signal },
+    );
+    const markedAssistant = bodyMessages(marked.body())[1]!;
+    expect(markedAssistant['reasoning_details']).toEqual([
+      { type: 'summary', summary: '第一段续' },
+      { type: 'summary', summary: '第二段' },
+      { type: 'encrypted', encrypted: 'cipher' },
+    ]);
+    expect(markedAssistant['reasoning_content']).toBe('第一段续第二段');
   });
 
   it('echoes an empty reasoning_content on think-less assistant messages only when keeping all', async () => {
@@ -426,6 +453,34 @@ describe('openai requester thinking', () => {
     const detectedAssistant = bodyMessages(captured[1]!)[1]!;
     expect(detectedAssistant['reasoning']).toBe('abc');
     expect('reasoning_content' in detectedAssistant).toBe(false);
+
+    const explicit = stubOpenAIClient(
+      chatCompletionChunks([
+        {
+          reasoning_details: [
+            { index: 0, type: 'summary', summary: 'ignored' },
+            { index: 1, type: 'encrypted', encrypted: 'cipher' },
+          ],
+        },
+        { content: 'ok' },
+      ]),
+    );
+    const explicitRequester = createOpenAIRequester(
+      { reasoningKey: () => 'reasoning' },
+      { clientFactory: explicit.clientFactory },
+    );
+    const explicitParts: unknown[] = [];
+    await explicitRequester.generate(
+      { model },
+      { messages },
+      {
+        signal: new AbortController().signal,
+        onEvent: (event) => {
+          if (event.type === 'llm.delta') explicitParts.push(event.part);
+        },
+      },
+    );
+    expect(explicitParts).toEqual([{ type: 'text', text: 'ok' }]);
   });
 
   it('parses reasoning from stream deltas', async () => {
@@ -470,5 +525,44 @@ describe('openai requester thinking', () => {
       { type: 'text', text: 'hi' },
       { type: 'think', think: '' },
     ]);
+    await expect(
+      collect(
+        chatCompletionChunks([
+          {
+            reasoning_content: '第一段',
+            reasoning_details: [{ index: 0, type: 'summary', summary: '第一段' }],
+          },
+          {
+            reasoning_details: [
+              { index: 0, summary: '续' },
+              { index: 1, type: 'summary', summary: '第二段' },
+            ],
+          },
+          { reasoning_details: [{ index: 2, type: 'encrypted', encrypted: 'cipher' }] },
+          { content: 'ok' },
+        ]),
+      ),
+    ).resolves.toEqual([
+      { type: 'think', think: '第一段续', detailsIndex: 0 },
+      { type: 'think', think: '第二段', detailsIndex: 1 },
+      { type: 'think', think: '', encrypted: 'cipher', detailsIndex: 2 },
+      { type: 'text', text: 'ok' },
+    ]);
+    await expect(
+      collect(
+        chatCompletionChunks([
+          {
+            reasoning_details: [
+              { index: 0, type: 'reasoning.text', text: 'foreign', format: 'unknown' },
+              { index: 1, type: 'summary', summary: 'kept' },
+            ],
+          },
+          { content: 'ok' },
+        ]),
+      ),
+    ).resolves.toEqual([
+      { type: 'think', think: 'kept', detailsIndex: 1 },
+      { type: 'text', text: 'ok' },
+    ]);
   });
 });
diff --git a/packages/agent-core-v2/test/agent/loop/loop.test.ts b/packages/agent-core-v2/test/agent/loop/loop.test.ts
index d7b931d4f5b..bd7d6bcda75 100644
--- a/packages/agent-core-v2/test/agent/loop/loop.test.ts
+++ b/packages/agent-core-v2/test/agent/loop/loop.test.ts
@@ -1572,7 +1572,7 @@ describe('interruption reminder', () => {
     expect(interruptionReminders()).toHaveLength(0);
   });
 
-  it('preserves partial thinking on user cancel', async () => {
+  it('drops unsigned thinking but keeps signed thinking on user cancel', async () => {
     ctx.mockNextResponse({ type: 'think', think: 'pondering' }, { type: 'text', text: 'answer' });
     const subscription = ctx.get(IEventBus).subscribe(ThinkingDelta, () => {
       loop.cancel();
@@ -1581,13 +1581,33 @@ describe('interruption reminder', () => {
     await expect(turn.result).resolves.toMatchObject({ type: 'cancelled' });
     subscription.dispose();
 
+    const thinkParts = ctx
+      .contextData()
+      .history.flatMap((message) => message.content)
+      .filter((part) => part.type === 'think');
+    expect(thinkParts).toEqual([]);
+    expect(interruptionReminders()).toHaveLength(1);
+
+    ctx.mockNextResponse(
+      { type: 'think', think: 'seg', encrypted: 'sig' },
+      { type: 'text', text: 'partial answer' },
+    );
+    const second = ctx.get(IEventBus).subscribe(AssistantDelta, () => {
+      loop.cancel();
+    });
+    const secondTurn = submitTurn(loop, 'Again').turn;
+    await expect(secondTurn.result).resolves.toMatchObject({ type: 'cancelled' });
+    second.dispose();
+
     expect(ctx.contextData().history).toContainEqual({
       role: 'assistant',
-      content: [{ type: 'think', think: 'pondering' }],
+      content: [
+        { type: 'think', think: 'seg', encrypted: 'sig' },
+        { type: 'text', text: 'partial answer' },
+      ],
       toolCalls: [],
       partial: true,
     });
-    expect(interruptionReminders()).toHaveLength(1);
   });
 
   it('records no partial content when the stream only produced whitespace', async () => {