diff --git a/src/backend/drivers/ai-chat/providers/claude/ClaudeProvider.test.ts b/src/backend/drivers/ai-chat/providers/claude/ClaudeProvider.test.ts
index 285c86f76..ac5add4fa 100644
--- a/src/backend/drivers/ai-chat/providers/claude/ClaudeProvider.test.ts
+++ b/src/backend/drivers/ai-chat/providers/claude/ClaudeProvider.test.ts
@@ -364,6 +364,95 @@ describe('ClaudeProvider.complete request shape', () => {
         expect(toolUse!.input).toEqual({ q: 'puter' });
     });
 
+    it('splices round-tripped reasoning_details back in ahead of the content', async () => {
+        // The replay contract for a normalized Claude turn: the caller resends
+        // the whole message, and the thinking blocks have to reach Anthropic
+        // with their signature intact and leading the content array (Anthropic
+        // rejects both a missing signature and a trailing thinking block).
+        const { provider } = makeProvider();
+        messagesCreateMock.mockResolvedValueOnce(baseResponse);
+
+        await withTestActor(() =>
+            provider.complete({
+                model: 'claude-haiku-4-5-20251001',
+                messages: [
+                    { role: 'user', content: 'think then call a tool' },
+                    {
+                        role: 'assistant',
+                        content: 'here you go',
+                        reasoning: 'step one',
+                        refusal: null,
+                        reasoning_details: [
+                            {
+                                type: 'thinking',
+                                thinking: 'step one',
+                                signature: 'sig_1',
+                            },
+                            { type: 'redacted_thinking', data: 'ENC' },
+                        ],
+                        tool_calls: [
+                            {
+                                id: 'call_1',
+                                type: 'function',
+                                function: {
+                                    name: 'lookup',
+                                    arguments: '{\"q\":\"puter\"}',
+                                },
+                            },
+                        ],
+                    } as never,
+                ],
+            }),
+        );
+
+        const [args] = messagesCreateMock.mock.calls[0]!;
+        const assistant = args.messages[1] as Record<string, unknown>;
+        const content = assistant.content as Array<Record<string, unknown>>;
+        // Thinking blocks lead, verbatim; string content became a text block;
+        // the tool_use block is appended after.
+        expect(content).toEqual([
+            { type: 'thinking', thinking: 'step one', signature: 'sig_1' },
+            { type: 'redacted_thinking', data: 'ENC' },
+            { type: 'text', text: 'here you go' },
+            {
+                type: 'tool_use',
+                id: 'call_1',
+                name: 'lookup',
+                input: { q: 'puter' },
+            },
+        ]);
+        // Output-only fields Anthropic rejects are stripped.
+        expect('reasoning_details' in assistant).toBe(false);
+        expect('reasoning' in assistant).toBe(false);
+        expect('refusal' in assistant).toBe(false);
+    });
+
+    it('strips output-only reasoning fields even with no reasoning_details', async () => {
+        const { provider } = makeProvider();
+        messagesCreateMock.mockResolvedValueOnce(baseResponse);
+
+        await withTestActor(() =>
+            provider.complete({
+                model: 'claude-haiku-4-5-20251001',
+                messages: [
+                    {
+                        role: 'assistant',
+                        content: 'plain reply',
+                        reasoning: 'leftover',
+                        refusal: null,
+                    } as never,
+                ],
+            }),
+        );
+
+        const [args] = messagesCreateMock.mock.calls[0]!;
+        const assistant = args.messages[0] as Record<string, unknown>;
+        expect('reasoning' in assistant).toBe(false);
+        expect('refusal' in assistant).toBe(false);
+        // Content is untouched when there was nothing to splice.
+        expect(assistant.content).toBe('plain reply');
+    });
+
     it('converts a tool-role message with tool_call_id into a user-role tool_result block', async () => {
         const { provider } = makeProvider();
         messagesCreateMock.mockResolvedValueOnce(baseResponse);
diff --git a/src/backend/drivers/ai-chat/providers/claude/ClaudeProvider.ts b/src/backend/drivers/ai-chat/providers/claude/ClaudeProvider.ts
index 2f01c7968..1f589694f 100644
--- a/src/backend/drivers/ai-chat/providers/claude/ClaudeProvider.ts
+++ b/src/backend/drivers/ai-chat/providers/claude/ClaudeProvider.ts
@@ -139,6 +139,37 @@ export class ClaudeProvider implements IChatProvider {
             return message;
         });
 
+        // Splice round-tripped reasoning artifacts back into the assistant
+        // content. Anthropic rejects an extended-thinking tool-use
+        // continuation whose thinking blocks lost their `signature`, and
+        // requires those blocks to lead the content array — so they are
+        // prepended here, before the tool_use blocks are appended below.
+        // `reasoning`/`refusal` are output-only fields Anthropic rejects
+        // outright, and a caller replaying a normalized message carries them.
+        // eslint-disable-next-line @typescript-eslint/no-explicit-any
+        messages = messages.map((message: any) => {
+            const details = message.reasoning_details;
+            delete message.reasoning_details;
+            delete message.reasoning;
+            delete message.refusal;
+            if (!Array.isArray(details)) return message;
+            const blocks = details.filter(
+                (block: unknown) =>
+                    (block as { type?: string })?.type === 'thinking' ||
+                    (block as { type?: string })?.type === 'redacted_thinking',
+            );
+            if (blocks.length === 0) return message;
+            if (typeof message.content === 'string') {
+                message.content = message.content
+                    ? [{ type: 'text', text: message.content }]
+                    : [];
+            } else if (!Array.isArray(message.content)) {
+                message.content = message.content ? [message.content] : [];
+            }
+            message.content = [...blocks, ...message.content];
+            return message;
+        });
+
         // Convert OpenAI-style tool calls/results to Claude format
         // eslint-disable-next-line @typescript-eslint/no-explicit-any
         messages = messages.map((message: any) => {
diff --git a/src/backend/drivers/ai-chat/types.ts b/src/backend/drivers/ai-chat/types.ts
index 92dcb95cd..d9e83d8c2 100644
--- a/src/backend/drivers/ai-chat/types.ts
+++ b/src/backend/drivers/ai-chat/types.ts
@@ -96,9 +96,9 @@ export interface ICompleteArguments {
     max_tokens?: number;
     temperature?: number;
     reasoning?: { effort: 'low' | 'medium' | 'high' } | undefined;
-    text?: string & { verbosity?: 'concise' | 'detailed' | undefined };
+    text?: { verbosity?: 'low' | 'medium' | 'high' | undefined } | undefined;
     reasoning_effort?: 'low' | 'medium' | 'high' | undefined;
-    verbosity?: 'concise' | 'detailed' | undefined;
+    verbosity?: 'low' | 'medium' | 'high' | undefined;
     moderation?: boolean;
     custom?: unknown;
     /**
diff --git a/src/backend/drivers/ai-chat/utils/OpenAIUtil.js b/src/backend/drivers/ai-chat/utils/OpenAIUtil.js
index 6bd975d4d..7cb1800f5 100644
--- a/src/backend/drivers/ai-chat/utils/OpenAIUtil.js
+++ b/src/backend/drivers/ai-chat/utils/OpenAIUtil.js
@@ -92,7 +92,47 @@ export const process_input_messages_responses_api = async (messages) => {
     // collapsing the whole message into a single compaction item and dropping
     // the rest of its content.
     const expanded = [];
-    for (const msg of messages) {
+    for (let msg of messages) {
+        // Round-tripped reasoning artifacts become standalone `reasoning`
+        // input items — the shape the Responses API expects them back in —
+        // and precede the message they were attached to, same as compaction.
+        // `reasoning`/`refusal`/`normalized` are output-only fields the input
+        // schema rejects, and a caller replaying a normalized message carries
+        // them along with the details.
+        if (msg && typeof msg === 'object') {
+            const details = msg.reasoning_details;
+            if (
+                details !== undefined ||
+                msg.reasoning !== undefined ||
+                msg.refusal !== undefined ||
+                msg.normalized !== undefined
+            ) {
+                const {
+                    reasoning_details: _details,
+                    reasoning: _reasoning,
+                    refusal: _refusal,
+                    normalized: _normalized,
+                    ...rest
+                } = msg;
+                msg = rest;
+            }
+            if (Array.isArray(details)) {
+                for (const block of details) {
+                    if (!block || block.type !== 'reasoning') continue;
+                    expanded.push({
+                        type: 'reasoning',
+                        ...(block.id !== undefined ? { id: block.id } : {}),
+                        ...(block.encrypted_content !== undefined
+                            ? { encrypted_content: block.encrypted_content }
+                            : {}),
+                        summary: Array.isArray(block.summary)
+                            ? block.summary
+                            : [],
+                    });
+                }
+            }
+        }
+
         if (msg && Array.isArray(msg.content)) {
             const compactionBlocks = msg.content.filter(
                 (c) => c && c.type === 'compaction',
@@ -439,6 +479,25 @@ export const create_chat_stream_handler_responses_api =
                 continue;
             }
 
+            // Reasoning summaries stream as their own delta events; route
+            // them to the same `reasoning` channel the chat-completions
+            // handler uses for Deepseek/OpenRouter, so a streamed reasoning
+            // model reads identically whichever API served it.
+            if (chunk.type === 'response.reasoning_summary_text.delta') {
+                textblock.addReasoning(chunk.delta);
+                continue;
+            }
+
+            // Each summary part is a separate delta stream; separate them with
+            // a blank line, matching the non-stream handler's join.
+            if (
+                chunk.type === 'response.reasoning_summary_part.added' &&
+                chunk.summary_index > 0
+            ) {
+                textblock.addReasoning('\n\n');
+                continue;
+            }
+
             if (chunk.type === 'response.completed') {
                 last_usage = chunk.response.usage;
             }
@@ -638,11 +697,31 @@ export const handle_completion_output_responses_api = async ({
     // Reasoning models return `reasoning` output items; their human-readable
     // text only exists when the caller requested summaries via
     // `reasoning: { summary: ... }` (raw chain-of-thought is never returned).
-    const reasoningText = output
-        .filter((item) => item?.type === 'reasoning')
+    const reasoningItems = output.filter((item) => item?.type === 'reasoning');
+    const reasoningText = reasoningItems
         .flatMap((item) => (Array.isArray(item.summary) ? item.summary : []))
         .map((part) => (typeof part?.text === 'string' ? part.text : ''))
-        .join('');
+        .filter(Boolean)
+        .join('\n\n');
+
+    // The item `id` and `encrypted_content` are what let a caller replay a
+    // reasoning turn into the next request; they are opaque to us and would
+    // otherwise be lost, so they ride `reasoning_details` verbatim — the same
+    // round-trip contract as the `compaction` artifact below and as the
+    // Anthropic thinking blocks the coercer preserves.
+    const reasoningDetails = reasoningItems
+        .filter(
+            (item) =>
+                item.id !== undefined || item.encrypted_content !== undefined,
+        )
+        .map((item) => ({
+            type: 'reasoning',
+            ...(item.id !== undefined ? { id: item.id } : {}),
+            ...(item.encrypted_content !== undefined
+                ? { encrypted_content: item.encrypted_content }
+                : {}),
+            ...(Array.isArray(item.summary) ? { summary: item.summary } : {}),
+        }));
 
     const ret = {
         finish_reason: responseToolCalls.length ? 'tool_calls' : 'stop',
@@ -651,6 +730,9 @@ export const handle_completion_output_responses_api = async ({
             content: completion.output_text,
             // String-or-absent, matching every other provider's `reasoning`.
             ...(reasoningText ? { reasoning: reasoningText } : {}),
+            ...(reasoningDetails.length
+                ? { reasoning_details: reasoningDetails }
+                : {}),
             refusal: null,
             role: 'assistant',
             ...(responseToolCalls.length
diff --git a/src/backend/drivers/ai-chat/utils/OpenAIUtil.test.ts b/src/backend/drivers/ai-chat/utils/OpenAIUtil.test.ts
index fe68d0bb7..fe5a057fd 100644
--- a/src/backend/drivers/ai-chat/utils/OpenAIUtil.test.ts
+++ b/src/backend/drivers/ai-chat/utils/OpenAIUtil.test.ts
@@ -292,6 +292,58 @@ describe('process_input_messages_responses_api', () => {
         ]);
     });
 
+    it('expands round-tripped reasoning_details into reasoning input items', async () => {
+        // The replay contract documented in Objects/chatresponse.md: a caller
+        // resends the whole normalized assistant message, reasoning_details
+        // included, and the Responses input schema gets back the item shape it
+        // issued — output-only fields stripped so the request is accepted.
+        const messages: Array<Record<string, unknown>> = [
+            {
+                role: 'assistant',
+                content: 'earlier reply',
+                reasoning: 'thought',
+                refusal: null,
+                reasoning_details: [
+                    {
+                        type: 'reasoning',
+                        id: 'rs_1',
+                        encrypted_content: 'ENC',
+                        summary: [{ type: 'summary_text', text: 'thought' }],
+                    },
+                ],
+            },
+        ];
+
+        const out = (await process_input_messages_responses_api(
+            messages,
+        )) as Array<Record<string, unknown>>;
+
+        expect(out).toHaveLength(2);
+        // Reasoning item precedes the message it belonged to.
+        expect(out[0]).toEqual({
+            type: 'reasoning',
+            id: 'rs_1',
+            encrypted_content: 'ENC',
+            summary: [{ type: 'summary_text', text: 'thought' }],
+        });
+        expect(out[1]!.role).toBe('assistant');
+        // Output-only fields would be rejected by the input schema.
+        expect('reasoning_details' in out[1]!).toBe(false);
+        expect('reasoning' in out[1]!).toBe(false);
+        expect('refusal' in out[1]!).toBe(false);
+    });
+
+    it('leaves messages without reasoning artifacts alone', async () => {
+        const messages: Array<Record<string, unknown>> = [
+            { role: 'user', content: 'hi' },
+        ];
+        const out = (await process_input_messages_responses_api(
+            messages,
+        )) as Array<Record<string, unknown>>;
+        expect(out).toHaveLength(1);
+        expect(out[0]!.role).toBe('user');
+    });
+
     it('upgrades user/system text blocks to input_text', async () => {
         const messages: Array<Record<string, unknown>> = [
             {
@@ -599,6 +651,78 @@ describe('create_chat_stream_handler_responses_api', () => {
         });
     });
 
+    it('routes reasoning-summary deltas to the reasoning channel', async () => {
+        const completion = asAsyncIterable([
+            {
+                type: 'response.reasoning_summary_text.delta',
+                delta: 'first thought',
+            },
+            {
+                type: 'response.reasoning_summary_part.added',
+                summary_index: 1,
+            },
+            {
+                type: 'response.reasoning_summary_text.delta',
+                delta: 'second thought',
+            },
+            { type: 'response.output_text.delta', delta: 'answer' },
+            {
+                type: 'response.completed',
+                response: { usage: { input_tokens: 1, output_tokens: 2 } },
+            },
+        ]);
+        const init = create_chat_stream_handler_responses_api({
+            deviations: undefined,
+            completion,
+            usage_calculator: () => ({}),
+        });
+        const harness = makeCapturingChatStream();
+        await init({ chatStream: harness.chatStream });
+
+        const events = harness.events();
+        // Same event type the chat-completions handler emits, so a streamed
+        // reasoning model reads identically whichever API served it.
+        expect(
+            events
+                .filter((e) => e.type === 'reasoning')
+                .map((e) => e.reasoning)
+                .join(''),
+        ).toBe('first thought\n\nsecond thought');
+        expect(
+            events.filter((e) => e.type === 'text').map((e) => e.text),
+        ).toEqual(['answer']);
+    });
+
+    it('does not separate the first summary part with a blank line', async () => {
+        const completion = asAsyncIterable([
+            {
+                type: 'response.reasoning_summary_part.added',
+                summary_index: 0,
+            },
+            { type: 'response.reasoning_summary_text.delta', delta: 'only' },
+            { type: 'response.output_text.delta', delta: 'answer' },
+            {
+                type: 'response.completed',
+                response: { usage: { input_tokens: 1, output_tokens: 2 } },
+            },
+        ]);
+        const init = create_chat_stream_handler_responses_api({
+            deviations: undefined,
+            completion,
+            usage_calculator: () => ({}),
+        });
+        const harness = makeCapturingChatStream();
+        await init({ chatStream: harness.chatStream });
+
+        expect(
+            harness
+                .events()
+                .filter((e) => e.type === 'reasoning')
+                .map((e) => e.reasoning)
+                .join(''),
+        ).toBe('only');
+    });
+
     it('emits a compaction event when a compaction output_item completes', async () => {
         const completion = asAsyncIterable([
             {
@@ -901,6 +1025,92 @@ describe('handle_completion_output_responses_api non-stream', () => {
         expect(moderate).toHaveBeenCalledWith('questionable content');
     });
 
+    it('joins multi-part reasoning summaries with a blank line', async () => {
+        const completion = {
+            output: [
+                {
+                    type: 'reasoning',
+                    summary: [
+                        { type: 'summary_text', text: 'First thought.' },
+                        { type: 'summary_text', text: 'Second thought.' },
+                    ],
+                },
+                { role: 'assistant', type: 'message' },
+            ],
+            output_text: 'answer',
+            usage: { input_tokens: 1, output_tokens: 2 },
+        };
+        const result = await handle_completion_output_responses_api({
+            deviations: undefined,
+            stream: false,
+            completion,
+        });
+        expect(result.message.reasoning).toBe(
+            'First thought.\n\nSecond thought.',
+        );
+    });
+
+    it('carries reasoning item id and encrypted_content for replay', async () => {
+        const completion = {
+            output: [
+                {
+                    type: 'reasoning',
+                    id: 'rs_1',
+                    encrypted_content: 'ENC',
+                    summary: [{ type: 'summary_text', text: 'thought' }],
+                },
+                { role: 'assistant', type: 'message' },
+            ],
+            output_text: 'answer',
+            usage: { input_tokens: 1, output_tokens: 2 },
+        };
+        const result = await handle_completion_output_responses_api({
+            deviations: undefined,
+            stream: false,
+            completion,
+        });
+        expect(result.message.reasoning_details).toEqual([
+            {
+                type: 'reasoning',
+                id: 'rs_1',
+                encrypted_content: 'ENC',
+                summary: [{ type: 'summary_text', text: 'thought' }],
+            },
+        ]);
+        expect(result.message.reasoning).toBe('thought');
+    });
+
+    it('omits reasoning_details when there are no reasoning items', async () => {
+        const completion = {
+            output: [{ role: 'assistant', type: 'message' }],
+            output_text: 'answer',
+            usage: { input_tokens: 1, output_tokens: 2 },
+        };
+        const result = await handle_completion_output_responses_api({
+            deviations: undefined,
+            stream: false,
+            completion,
+        });
+        expect('reasoning_details' in result.message).toBe(false);
+    });
+
+    it('omits reasoning entirely when no summaries were requested', async () => {
+        const completion = {
+            output: [
+                { type: 'reasoning', summary: [] },
+                { role: 'assistant', type: 'message' },
+            ],
+            output_text: 'answer',
+            usage: { input_tokens: 1, output_tokens: 2 },
+        };
+        const result = await handle_completion_output_responses_api({
+            deviations: undefined,
+            stream: false,
+            completion,
+        });
+        expect('reasoning' in result.message).toBe(false);
+    });
+
     it('returns a stream init descriptor when stream=true', async () => {
         const completion = asAsyncIterable([]);
         const result = await handle_completion_output_responses_api({
diff --git a/src/backend/drivers/ai-chat/utils/normalizeToOpenAI.test.ts b/src/backend/drivers/ai-chat/utils/normalizeToOpenAI.test.ts
index 393dc8df5..5324b9e1f 100644
--- a/src/backend/drivers/ai-chat/utils/normalizeToOpenAI.test.ts
+++ b/src/backend/drivers/ai-chat/utils/normalizeToOpenAI.test.ts
@@ -181,11 +181,19 @@ describe('normalizeResultToOpenAI', () => {
         expect(out.finish_reason).toBe(expected);
     });
 
-    it('keeps the existing finish_reason for unknown stop_reasons', () => {
+    it('passes an unmapped vendor stop_reason through verbatim', () => {
+        // `pause_turn` means "continue this turn"; mapping it to `stop` would
+        // erase that. Objects/chatresponse.md documents the passthrough.
         const out = normalizeResultToOpenAI(
             claudeResult([{ type: 'text', text: 'x' }], 'pause_turn'),
         );
-        expect(out.finish_reason).toBe('stop');
+        expect(out.finish_reason).toBe('pause_turn');
+    });
+
+    it('falls back to the existing finish_reason when stop_reason is absent', () => {
+        const res = claudeResult([{ type: 'text', text: 'x' }]);
+        delete (res.message as Record<string, unknown>).stop_reason;
+        expect(normalizeResultToOpenAI(res).finish_reason).toBe('stop');
     });
 
     it('converts tool_use blocks into OpenAI tool_calls with stringified arguments', () => {
@@ -247,10 +255,32 @@ describe('normalizeResultToOpenAI', () => {
         expect(out.message.content).toBe('answer');
     });
 
-    it('drops redacted_thinking, compaction, and unknown blocks', () => {
+    it('preserves thinking blocks verbatim in reasoning_details for replay', () => {
+        // Anthropic rejects an extended-thinking continuation whose thinking
+        // blocks lost their signature, so the raw blocks have to survive.
+        const out = normalizeResultToOpenAI(
+            claudeResult([
+                { type: 'thinking', thinking: 'step one.', signature: 's1' },
+                { type: 'redacted_thinking', data: 'ENC' },
+                { type: 'text', text: 'answer' },
+            ]),
+        );
+        expect(out.message.reasoning_details).toEqual([
+            { type: 'thinking', thinking: 'step one.', signature: 's1' },
+            { type: 'redacted_thinking', data: 'ENC' },
+        ]);
+    });
+
+    it('omits reasoning_details when there was no reasoning', () => {
+        const out = normalizeResultToOpenAI(
+            claudeResult([{ type: 'text', text: 'answer' }]),
+        );
+        expect('reasoning_details' in (out.message as object)).toBe(false);
+    });
+
+    it('drops compaction and unknown blocks', () => {
         const out = normalizeResultToOpenAI({
             ...claudeResult([
-                { type: 'redacted_thinking', data: 'ENC' },
                 { type: 'compaction', content: 'ENC2' },
                 { type: 'server_tool_use', id: 'x', name: 'y', input: {} },
                 { type: 'text', text: 'visible' },
diff --git a/src/backend/drivers/ai-chat/utils/normalizeToOpenAI.ts b/src/backend/drivers/ai-chat/utils/normalizeToOpenAI.ts
index 39c685db1..3629cb61d 100644
--- a/src/backend/drivers/ai-chat/utils/normalizeToOpenAI.ts
+++ b/src/backend/drivers/ai-chat/utils/normalizeToOpenAI.ts
@@ -80,13 +80,26 @@ const mapStopReason = (
     stop_reason: unknown,
     fallback: string | undefined,
 ): string => {
-    if (typeof stop_reason === 'string') {
-        const mapped = STOP_REASON_TO_FINISH_REASON[stop_reason];
-        if (mapped) return mapped;
+    if (typeof stop_reason === 'string' && stop_reason !== '') {
+        // A vendor reason with no OpenAI analog passes through verbatim — the
+        // same contract the Mistral remap follows and the one
+        // Objects/chatresponse.md documents. Collapsing e.g. Anthropic's
+        // `pause_turn` to `stop` would erase a "continue this turn" signal
+        // the caller needs to act on.
+        return STOP_REASON_TO_FINISH_REASON[stop_reason] ?? stop_reason;
     }
     return fallback ?? 'stop';
 };
 
+/**
+ * Verbatim provider reasoning blocks, preserved so a normalized message can
+ * still be replayed into an extended-thinking tool-use continuation. Anthropic
+ * rejects a continuation whose thinking blocks lost their `signature`, and
+ * `redacted_thinking` is opaque but must round-trip intact. Modelled on the
+ * top-level `compaction` artifact: opaque to us, drop-in for the caller.
+ */
+type ReasoningDetail = Record<string, unknown>;
+
 type OpenAIToolCall = {
     id: unknown;
     type: 'function';
@@ -98,10 +111,11 @@ type OpenAIToolCall = {
  *
  * Returns `res` by reference when the message is already OpenAI-shaped.
  * Otherwise rebuilds `message` (text blocks joined into a string `content`,
- * `tool_use` blocks into `tool_calls`, `thinking` blocks into `reasoning`) and
- * remaps `finish_reason` from the Anthropic `stop_reason`. Everything else on
- * the result — `usage`, the top-level `compaction` artifact — passes through
- * unchanged. The caller owns the `normalized` marker.
+ * `tool_use` blocks into `tool_calls`, `thinking` blocks into `reasoning` plus
+ * verbatim `reasoning_details` for replay) and remaps `finish_reason` from the
+ * Anthropic `stop_reason`, passing an unmapped vendor reason through verbatim.
+ * Everything else on the result — `usage`, the top-level `compaction` artifact
+ * — passes through unchanged. The caller owns the `normalized` marker.
  */
 export const normalizeResultToOpenAI = (
     res: IChatMessageResult,
@@ -127,6 +141,7 @@ export const normalizeResultToOpenAI = (
 
     const textParts: string[] = [];
     const reasoningParts: string[] = [];
+    const reasoningDetails: ReasoningDetail[] = [];
     const toolCalls: OpenAIToolCall[] = [];
 
     for (const block of blocks) {
@@ -140,6 +155,12 @@ export const normalizeResultToOpenAI = (
                 if (typeof b.thinking === 'string') {
                     reasoningParts.push(b.thinking);
                 }
+                reasoningDetails.push({ ...b });
+                break;
+            // Encrypted, so there is no text to surface — but it still has to
+            // survive the round trip, so it rides `reasoning_details` too.
+            case 'redacted_thinking':
+                reasoningDetails.push({ ...b });
                 break;
             case 'tool_use':
                 toolCalls.push({
@@ -154,10 +175,10 @@ export const normalizeResultToOpenAI = (
                     },
                 });
                 break;
-            // `redacted_thinking` is encrypted, and `compaction` already
-            // rides the result's top-level `compaction` field; both — and
-            // any block type introduced later — are dropped rather than
-            // leaked into a shape that has nowhere to put them.
+            // `compaction` already rides the result's top-level
+            // `compaction` field; it — and any block type introduced
+            // later — is dropped rather than leaked into a shape that has
+            // nowhere to put it.
             default:
                 break;
         }
@@ -177,6 +198,9 @@ export const normalizeResultToOpenAI = (
             refusal: null,
             ...(toolCalls.length > 0 ? { tool_calls: toolCalls } : {}),
             ...(reasoning !== undefined ? { reasoning } : {}),
+            ...(reasoningDetails.length > 0
+                ? { reasoning_details: reasoningDetails }
+                : {}),
         },
         finish_reason: mapStopReason(native.stop_reason, res.finish_reason),
     };
diff --git a/src/docs/src/AI/chat.md b/src/docs/src/AI/chat.md
index a11a35433..76598de99 100755
--- a/src/docs/src/AI/chat.md
+++ b/src/docs/src/AI/chat.md
@@ -142,9 +142,11 @@ puter.ai.normalize = false;  // every chat() call returns the vendor-native form
 
 A `normalize` option on an individual call always overrides `puter.ai.normalize`. When neither is set, the release-date rule above decides. Normalized responses carry `normalized: true`.
 
-On a normalized response, extended-thinking output (from reasoning models that expose it) is joined into `message.reasoning`, and Anthropic stop reasons are mapped to OpenAI values (`end_turn` → `stop`, `max_tokens` → `length`, `tool_use` → `tool_calls`, `refusal` → `content_filter`).
+On a normalized response, extended-thinking output (from reasoning models that expose it) is joined into `message.reasoning`, and Anthropic stop reasons are mapped to OpenAI values (`end_turn` → `stop`, `max_tokens` → `length`, `tool_use` → `tool_calls`, `refusal` → `content_filter`). A vendor stop reason with no OpenAI equivalent — Anthropic's `pause_turn`, for instance — passes through unchanged, so treat `finish_reason` as an open set. See [`finish_reason`](/Objects/chatresponse) for the full mapping.
 
-Two caveats. The release-date rule applies to the model that actually serves the request — if a request is rerouted to a fallback provider, the served model's release date decides. And normalization drops Anthropic thinking-block signatures, so agentic loops that resend Claude extended-thinking messages in tool-use continuations should request the native format (`normalize: false`).
+Normalization does not cost you the ability to continue a reasoning turn. The opaque parts a provider needs back — Anthropic thinking-block signatures, OpenAI reasoning item ids and encrypted content — are preserved verbatim on `message.reasoning_details`. Resend that array as-is alongside the message when you continue an extended-thinking tool-use loop. The artifacts are vendor-specific and only meaningful to the model that produced them, so replay them to the same model — don't carry them across vendors.
+
+One caveat. The release-date rule applies to the model that actually serves the request — if a request is rerouted to a fallback provider, the served model's release date decides.
 
 Streaming is unaffected by normalization: streamed [`ChatResponseChunk`](/Objects/chatresponsechunk) objects already share one format across all vendors.
 
diff --git a/src/docs/src/Objects/chatresponse.md b/src/docs/src/Objects/chatresponse.md
index d0ba3383d..a37b0cfc6 100644
--- a/src/docs/src/Objects/chatresponse.md
+++ b/src/docs/src/Objects/chatresponse.md
@@ -17,7 +17,9 @@ An object containing the chat message data.
 
 - `tool_calls` (Array) - An optional array of [`ToolCall`](/Objects/toolcall) objects if the model wants to call tools.
 
-- `reasoning` (String) - Optional extended-thinking output, when the model exposes it.
+- `reasoning` (String) - Optional extended-thinking output, when the model exposes it. Multiple reasoning segments are joined with a blank line between them.
+
+- `reasoning_details` (Array) - Optional opaque reasoning artifacts, present on normalized responses from models that expose them: Anthropic `thinking`/`redacted_thinking` blocks with their `signature`, or OpenAI reasoning items with their `id` and `encrypted_content`. Treat the contents as opaque and resend the array verbatim to continue an extended-thinking turn — providers reject a continuation whose reasoning lost its signature. The human-readable text is in `reasoning`; this field is only for the round trip.
 
 - `tool_call_id` (String) - An optional identifier linking this message to the tool call it responds to.
 
@@ -27,7 +29,20 @@ An object containing the chat message data.
 
 #### `finish_reason` (String)
 
-Why generation stopped. On normalized responses, known vendor stop reasons map to `stop`, `length`, `tool_calls`, or `content_filter`; a vendor value with no OpenAI analog passes through unchanged.
+Why generation stopped. On normalized responses, known vendor stop reasons map to the OpenAI vocabulary — `stop`, `length`, `tool_calls`, or `content_filter` — and a vendor value with no OpenAI analog passes through unchanged rather than being flattened to `stop`.
+
+Anthropic models are the main source of both cases. Their stop reasons map as follows:
+
+| Anthropic `stop_reason` | Normalized `finish_reason` | Meaning |
+| --- | --- | --- |
+| `end_turn` | `stop` | The model finished its turn. |
+| `stop_sequence` | `stop` | One of your stop sequences was produced. |
+| `max_tokens` | `length` | The token limit was hit mid-answer. |
+| `tool_use` | `tool_calls` | The model wants to call a tool; see `message.tool_calls`. |
+| `refusal` | `content_filter` | The model declined to continue. |
+| `pause_turn` | `pause_turn` | A long-running server-side tool turn was paused — it has no OpenAI analog, so it passes through unchanged. Send the response back as-is to let the model continue. |
+
+Because unmapped values pass through, treat `finish_reason` as an open set: branch on the four OpenAI values you care about and handle anything else as vendor-specific rather than assuming it means `stop`.
 
 #### `normalized` (Boolean)
 
diff --git a/src/puter-js/src/modules/ai/types.js b/src/puter-js/src/modules/ai/types.js
index 5d452b3b7..a527066af 100644
--- a/src/puter-js/src/modules/ai/types.js
+++ b/src/puter-js/src/modules/ai/types.js
@@ -45,6 +45,13 @@
  * @property {{ type: string }} [cache_control]
  * @property {ImageContent[]} [images] Images attached to the message. Present on responses from
  * image-capable models.
+ * @property {string} [reasoning] Reasoning/thinking text, when the model exposes it and the
+ * request asked for it. Present on responses only.
+ * @property {object[]} [reasoning_details] Opaque provider reasoning artifacts (Anthropic thinking
+ * signatures, OpenAI reasoning item ids/encrypted content). Resend them verbatim to continue an
+ * extended-thinking turn. Present on responses only.
+ * @property {string | null} [refusal] Refusal message when the model declined, otherwise `null`.
+ * Present on responses only.
  */
 
 /**