@vybestack/llxprt-code-core 0.7.0-nightly.251209.0061bd6bf → 0.7.0-nightly.251211.134f1920b

package/dist/src/providers/openai/OpenAIProvider.js

@@ -22,7 +22,7 @@ import crypto from 'node:crypto';
 import * as http from 'http';
 import * as https from 'https';
 import * as net from 'net';
-import { isKimiModel, getToolIdStrategy, } from '../../tools/ToolIdStrategy.js';
+import { isKimiModel, isMistralModel, getToolIdStrategy, } from '../../tools/ToolIdStrategy.js';
 import { BaseProvider, } from '../BaseProvider.js';
 import { DebugLogger } from '../../debug/index.js';
 import { ToolFormatter } from '../../tools/ToolFormatter.js';
@@ -40,6 +40,7 @@ import { buildToolResponsePayload, EMPTY_TOOL_RESULT_PLACEHOLDER, } from '../uti
 import { isLocalEndpoint } from '../utils/localEndpoint.js';
 import { filterThinkingForContext, thinkingToReasoningField, extractThinkingBlocks, } from '../reasoning/reasoningUtils.js';
 import { shouldDumpSDKContext, dumpSDKContext, } from '../utils/dumpSDKContext.js';
+import { extractCacheMetrics } from '../utils/cacheMetricsExtractor.js';
 const MAX_TOOL_RESPONSE_CHARS = 1024;
 const MAX_TOOL_RESPONSE_RETRY_CHARS = 512;
 const TOOL_ARGS_PREVIEW_LENGTH = 500;
@@ -277,13 +278,12 @@ export class OpenAIProvider extends BaseProvider {
         // This preserves meaningful whitespace in regular text chunks during streaming
         // (e.g., " 5 Biggest" should remain " 5 Biggest", not become "5 Biggest")
         if (hadReasoningTags) {
-            // Clean up multiple consecutive spaces/whitespace that may result from stripping
+            // Collapse multiple spaces/tabs but preserve newlines for proper paragraph/line breaks
             str = str.replace(/[ \t]+/g, ' ');
             str = str.replace(/\n{3,}/g, '\n\n');
-            // Only trim leading whitespace when think tags were at the beginning
-            // This prevents leading spaces from "<think>...</think>text" -> " text"
-            // but preserves trailing whitespace for streaming chunk concatenation
-            str = str.trimStart();
+            // Only trim leading horizontal whitespace (spaces/tabs), NOT newlines
+            // This preserves line breaks between think tags and content (fixes #721)
+            str = str.replace(/^[ \t]+/, '');
         }
         const afterLen = str.length;
         if (hadReasoningTags && afterLen !== beforeLen) {
@@ -438,62 +438,78 @@ export class OpenAIProvider extends BaseProvider {
      * and all tool info is only encoded in the text template.
      */
     extractKimiToolCallsFromText(raw) {
-        if (!raw || !raw.includes('<|tool_calls_section_begin|>')) {
+        // Return early only if input is null/undefined/empty
+        if (!raw) {
             return { cleanedText: raw, toolCalls: [] };
         }
         const logger = this.getLogger();
         const toolCalls = [];
         let text = raw;
-        const sectionRegex = /<\|tool_calls_section_begin\|>([\s\S]*?)<\|tool_calls_section_end\|>/g;
-        text = text.replace(sectionRegex, (_sectionMatch, sectionBody) => {
-            try {
-                const callRegex = /<\|tool_call_begin\|>\s*([^<]+?)\s*<\|tool_call_argument_begin\|>\s*([\s\S]*?)\s*<\|tool_call_end\|>/g;
-                let m;
-                while ((m = callRegex.exec(sectionBody)) !== null) {
-                    const rawId = m[1].trim();
-                    const rawArgs = m[2].trim();
-                    // Infer tool name from ID.
-                    let toolName = '';
-                    const match = /^functions\.([A-Za-z0-9_]+):\d+/i.exec(rawId) ||
-                        /^[A-Za-z0-9_]+\.([A-Za-z0-9_]+):\d+/.exec(rawId);
-                    if (match) {
-                        toolName = match[1];
-                    }
-                    else {
-                        const colonParts = rawId.split(':');
-                        const head = colonParts[0] || rawId;
-                        const dotParts = head.split('.');
-                        toolName = dotParts[dotParts.length - 1] || head;
+        // Extract tool calls from complete sections if present
+        if (raw.includes('<|tool_calls_section_begin|>')) {
+            const sectionRegex = /<\|tool_calls_section_begin\|>([\s\S]*?)<\|tool_calls_section_end\|>/g;
+            text = text.replace(sectionRegex, (_sectionMatch, sectionBody) => {
+                try {
+                    const callRegex = /<\|tool_call_begin\|>\s*([^<]+?)\s*<\|tool_call_argument_begin\|>\s*([\s\S]*?)\s*<\|tool_call_end\|>/g;
+                    let m;
+                    while ((m = callRegex.exec(sectionBody)) !== null) {
+                        const rawId = m[1].trim();
+                        const rawArgs = m[2].trim();
+                        // Infer tool name from ID.
+                        let toolName = '';
+                        const match = /^functions\.([A-Za-z0-9_]+):\d+/i.exec(rawId) ||
+                            /^[A-Za-z0-9_]+\.([A-Za-z0-9_]+):\d+/.exec(rawId);
+                        if (match) {
+                            toolName = match[1];
+                        }
+                        else {
+                            const colonParts = rawId.split(':');
+                            const head = colonParts[0] || rawId;
+                            const dotParts = head.split('.');
+                            toolName = dotParts[dotParts.length - 1] || head;
+                        }
+                        // Normalize tool name (handles Kimi-K2 style prefixes like call_functionsglob7)
+                        toolName = this.normalizeToolName(toolName);
+                        const sanitizedArgs = this.sanitizeToolArgumentsString(rawArgs);
+                        const processedParameters = processToolParameters(sanitizedArgs, toolName);
+                        toolCalls.push({
+                            type: 'tool_call',
+                            id: this.normalizeToHistoryToolId(rawId),
+                            name: toolName,
+                            parameters: processedParameters,
+                        });
                     }
-                    // Normalize tool name (handles Kimi-K2 style prefixes like call_functionsglob7)
-                    toolName = this.normalizeToolName(toolName);
-                    const sanitizedArgs = this.sanitizeToolArgumentsString(rawArgs);
-                    const processedParameters = processToolParameters(sanitizedArgs, toolName);
-                    toolCalls.push({
-                        type: 'tool_call',
-                        id: this.normalizeToHistoryToolId(rawId),
-                        name: toolName,
-                        parameters: processedParameters,
-                    });
                 }
-            }
-            catch (err) {
-                logger.debug(() => `[OpenAIProvider] Failed to parse Kimi tool_calls_section: ${err}`);
-            }
-            // Strip the entire tool section from user-visible text
-            return '';
-        });
-        if (toolCalls.length > 0) {
-            logger.debug(() => `[OpenAIProvider] Parsed Kimi tool_calls_section`, {
-                toolCallCount: toolCalls.length,
-                originalLength: raw.length,
-                cleanedLength: text.length,
+                catch (err) {
+                    logger.debug(() => `[OpenAIProvider] Failed to parse Kimi tool_calls_section: ${err}`);
+                }
+                // Strip the entire tool section from user-visible text
+                return '';
             });
+            if (toolCalls.length > 0) {
+                logger.debug(() => `[OpenAIProvider] Parsed Kimi tool_calls_section`, {
+                    toolCallCount: toolCalls.length,
+                    originalLength: raw.length,
+                    cleanedLength: text.length,
+                });
+            }
         }
+        // ALWAYS run stray token cleanup, even if no complete sections were found
+        // This handles partial sections, malformed tokens, orphaned markers, etc.
+        text = text.replace(/<\|tool_call(?:_(?:begin|end|argument_begin))?\|>/g, '');
+        text = text.replace(/<\|tool_calls_section_(?:begin|end)\|>/g, '');
         // Don't trim - preserve leading/trailing newlines that are important for formatting
         // (e.g., numbered lists from Kimi K2 that have newlines between items)
         return { cleanedText: text, toolCalls };
     }
+    /**
+     * Clean Kimi K2 tool call tokens from thinking content.
+     * Used when extracting thinking from <think> tags that may contain embedded tool calls.
+     * @issue #749
+     */
+    cleanThinkingContent(thought) {
+        return this.extractKimiToolCallsFromText(thought).cleanedText;
+    }
     /**
      * @plan:PLAN-20251023-STATELESS-HARDENING.P09
      * @requirement:REQ-SP4-002
@@ -910,9 +926,12 @@ export class OpenAIProvider extends BaseProvider {
                     }
                     else {
                         // Assistant message with tool calls
+                        // CRITICAL for Mistral API compatibility (#760):
+                        // When tool_calls are present, we must NOT include a content property at all
+                        // (not even null). Mistral's OpenAI-compatible API requires this.
+                        // See: https://docs.mistral.ai/capabilities/function_calling
                         messages.push({
                             role: 'assistant',
-                            content: text || null,
                             tool_calls: toolCalls.map((tc) => ({
                                 id: this.normalizeToOpenAIToolId(tc.id),
                                 type: 'function',
@@ -948,10 +967,16 @@ export class OpenAIProvider extends BaseProvider {
                 }
                 else {
                     for (const tr of toolResponses) {
+                        // CRITICAL for Mistral API compatibility (#760):
+                        // Tool messages must include a name field matching the function name.
+                        // See: https://docs.mistral.ai/capabilities/function_calling
+                        // Note: The OpenAI SDK types don't include name, but Mistral requires it.
+                        // We use a type assertion to add this required field.
                         messages.push({
                             role: 'tool',
                             content: this.buildToolResponseContent(tr, config),
                             tool_call_id: this.normalizeToOpenAIToolId(tr.callId),
+                            name: tr.toolName,
                         });
                     }
                 }
@@ -977,8 +1002,9 @@ export class OpenAIProvider extends BaseProvider {
         const messages = [];
         // Create a ToolIdMapper based on the tool format
         // For Kimi K2, this generates sequential IDs in the format functions.{name}:{index}
-        const toolIdMapper = toolFormat === 'kimi'
-            ? getToolIdStrategy('kimi').createMapper(filteredContents)
+        // For Mistral, this generates 9-char alphanumeric IDs
+        const toolIdMapper = toolFormat === 'kimi' || toolFormat === 'mistral'
+            ? getToolIdStrategy(toolFormat).createMapper(filteredContents)
             : null;
         // Helper to resolve tool call IDs based on format
         const resolveToolCallId = (tc) => {
@@ -1014,9 +1040,12 @@ export class OpenAIProvider extends BaseProvider {
                 const toolCalls = content.blocks.filter((b) => b.type === 'tool_call');
                 if (toolCalls.length > 0) {
                     // Assistant message with tool calls
+                    // CRITICAL for Mistral API compatibility (#760):
+                    // When tool_calls are present, we must NOT include a content property at all
+                    // (not even null). Mistral's OpenAI-compatible API requires this.
+                    // See: https://docs.mistral.ai/capabilities/function_calling
                     const baseMessage = {
                         role: 'assistant',
-                        content: text || null,
                         tool_calls: toolCalls.map((tc) => ({
                             id: resolveToolCallId(tc),
                             type: 'function',
@@ -1057,10 +1086,16 @@ export class OpenAIProvider extends BaseProvider {
                 // Convert tool responses
                 const toolResponses = content.blocks.filter((b) => b.type === 'tool_response');
                 for (const tr of toolResponses) {
+                    // CRITICAL for Mistral API compatibility (#760):
+                    // Tool messages must include a name field matching the function name.
+                    // See: https://docs.mistral.ai/capabilities/function_calling
+                    // Note: The OpenAI SDK types don't include name, but Mistral requires it.
+                    // We use a type assertion to add this required field.
                     messages.push({
                         role: 'tool',
                         content: this.buildToolResponseContent(tr, options.config),
                         tool_call_id: resolveToolResponseId(tr),
+                        name: tr.toolName,
                     });
                 }
             }
@@ -1506,9 +1541,9 @@ export class OpenAIProvider extends BaseProvider {
             // Buffer for accumulating text chunks for providers that need it
             let textBuffer = '';
             // Use the same detected format from earlier for consistency
-            const isKimiModel = model.toLowerCase().includes('kimi-k2');
+            const isKimiK2Model = model.toLowerCase().includes('kimi-k2');
             // Buffer text for Qwen format providers and Kimi-K2 to avoid stanza formatting
-            const shouldBufferText = detectedFormat === 'qwen' || isKimiModel;
+            const shouldBufferText = detectedFormat === 'qwen' || isKimiK2Model;
             // Accumulate thinking content across the entire stream to emit as ONE block
             // This handles fragmented <think>word</think> streaming from Synthetic API
             // @plan PLAN-20251202-THINKING.P16
@@ -1575,12 +1610,29 @@ export class OpenAIProvider extends BaseProvider {
                         continue;
                     // Parse reasoning_content from streaming delta (Phase 16 integration)
                     // ACCUMULATE instead of yielding immediately to handle token-by-token streaming
+                    // Extract embedded Kimi K2 tool calls from reasoning_content (fixes #749)
                     // @plan PLAN-20251202-THINKING.P16
-                    const reasoningBlock = this.parseStreamingReasoningDelta(choice.delta);
+                    // @requirement REQ-KIMI-REASONING-001.1
+                    const { thinking: reasoningBlock, toolCalls: reasoningToolCalls } = this.parseStreamingReasoningDelta(choice.delta);
                     if (reasoningBlock) {
                         // Accumulate reasoning content - will emit ONE block later
                         accumulatedReasoningContent += reasoningBlock.thought;
                     }
+                    // Accumulate tool calls extracted from reasoning_content
+                    if (reasoningToolCalls.length > 0) {
+                        for (const toolCall of reasoningToolCalls) {
+                            // Convert ToolCallBlock to accumulated format
+                            const index = accumulatedToolCalls.length;
+                            accumulatedToolCalls[index] = {
+                                id: toolCall.id,
+                                type: 'function',
+                                function: {
+                                    name: toolCall.name,
+                                    arguments: JSON.stringify(toolCall.parameters),
+                                },
+                            };
+                        }
+                    }
                     // Check for finish_reason to detect proper stream ending
                     if (choice.finish_reason) {
                         logger.debug(() => `[Streaming] Stream finished with reason: ${choice.finish_reason}`, {
@@ -1601,13 +1653,25 @@ export class OpenAIProvider extends BaseProvider {
                     }
                     // Handle text content - buffer for Qwen format, emit immediately for others
                     // Note: Synthetic API sends content that may duplicate reasoning_content.
-                    // This is the model's behavior - we don't filter it here.
+                    // We now filter duplicates by tracking when content starts matching reasoning_content.
+                    // fixes #721
                     // @plan PLAN-20251202-THINKING.P16
                     const rawDeltaContent = this.coerceMessageContentToString(choice.delta?.content);
                     if (rawDeltaContent) {
-                        const deltaContent = isKimiModel
-                            ? rawDeltaContent
-                            : this.sanitizeProviderText(rawDeltaContent);
+                        // For Kimi models, we need to buffer the RAW content without processing
+                        // because Kimi tokens stream incrementally and partial tokens would leak
+                        // through if we try to process them immediately. The buffer will be
+                        // processed when flushed (at sentence boundaries or end of stream).
+                        let deltaContent;
+                        if (isKimiK2Model) {
+                            // For Kimi: Don't process yet - just pass through and let buffering handle it
+                            // We'll extract tool calls and sanitize when we flush the buffer
+                            deltaContent = rawDeltaContent;
+                        }
+                        else {
+                            // For non-Kimi models: sanitize immediately as before
+                            deltaContent = this.sanitizeProviderText(rawDeltaContent);
+                        }
                         if (!deltaContent) {
                             continue;
                         }
@@ -1623,9 +1687,9 @@ export class OpenAIProvider extends BaseProvider {
                             });
                             // Buffer text to avoid stanza formatting
                             textBuffer += deltaContent;
-                            const hasKimiBegin = textBuffer.includes('<|tool_calls_section_begin|>');
-                            const hasKimiEnd = textBuffer.includes('<|tool_calls_section_end|>');
-                            const hasOpenKimiSection = hasKimiBegin && !hasKimiEnd;
+                            const kimiBeginCount = (textBuffer.match(/<\|tool_calls_section_begin\|>/g) || []).length;
+                            const kimiEndCount = (textBuffer.match(/<\|tool_calls_section_end\|>/g) || []).length;
+                            const hasOpenKimiSection = kimiBeginCount > kimiEndCount;
                             // Emit buffered text when we have a complete sentence or paragraph
                             // Look for natural break points, but avoid flushing mid Kimi section
                             if (!hasOpenKimiSection &&
@@ -1642,12 +1706,14 @@ export class OpenAIProvider extends BaseProvider {
                                 // @requirement REQ-THINK-003
                                 const tagBasedThinking = this.extractThinkTagsAsBlock(workingText);
                                 if (tagBasedThinking) {
+                                    // Clean Kimi tokens from thinking content before accumulating
+                                    const cleanedThought = this.cleanThinkingContent(tagBasedThinking.thought);
                                     // Accumulate thinking content - don't emit yet
                                     // Use newline to preserve formatting between chunks (not space)
                                     if (accumulatedThinkingContent.length > 0) {
                                         accumulatedThinkingContent += '\n';
                                     }
-                                    accumulatedThinkingContent += tagBasedThinking.thought;
+                                    accumulatedThinkingContent += cleanedThought;
                                     logger.debug(() => `[Streaming legacy] Accumulated thinking: ${accumulatedThinkingContent.length} chars total`);
                                 }
                                 const kimiParsed = this.extractKimiToolCallsFromText(workingText);
@@ -1709,7 +1775,10 @@ export class OpenAIProvider extends BaseProvider {
                                 // Always use sanitized text to strip <think> tags (legacy streaming)
                                 // Bug fix: Previously Kimi used unsanitized workingText
                                 // @plan PLAN-20251202-THINKING.P16
-                                if (cleanedText.trim().length > 0) {
+                                // Bug fix #721: Emit whitespace-only chunks (e.g., " " between words)
+                                // Previously we used cleanedText.trim().length > 0 which dropped spaces,
+                                // causing "list 5" to become "list5". Now we emit any non-empty cleanedText.
+                                if (cleanedText.length > 0) {
                                     yield {
                                         speaker: 'ai',
                                         blocks: [
@@ -1828,11 +1897,13 @@ export class OpenAIProvider extends BaseProvider {
                 // @plan PLAN-20251202-THINKING.P16
                 const tagBasedThinking = this.extractThinkTagsAsBlock(workingText);
                 if (tagBasedThinking) {
+                    // Clean Kimi tokens from thinking content before accumulating
+                    const cleanedThought = this.cleanThinkingContent(tagBasedThinking.thought);
                     // Use newline to preserve formatting between chunks (not space)
                     if (accumulatedThinkingContent.length > 0) {
                         accumulatedThinkingContent += '\n';
                     }
-                    accumulatedThinkingContent += tagBasedThinking.thought;
+                    accumulatedThinkingContent += cleanedThought;
                 }
                 const kimiParsed = this.extractKimiToolCallsFromText(workingText);
                 if (kimiParsed.toolCalls.length > 0) {
@@ -1891,7 +1962,10 @@ export class OpenAIProvider extends BaseProvider {
                 // Always use sanitized text to strip <think> tags (legacy final buffer)
                 // Bug fix: Previously Kimi used unsanitized workingText
                 // @plan PLAN-20251202-THINKING.P16
-                if (cleanedText.trim().length > 0) {
+                // Bug fix #721: Emit whitespace-only chunks (e.g., " " between words)
+                // Previously we used cleanedText.trim().length > 0 which dropped spaces,
+                // causing "list 5" to become "list5". Now we emit any non-empty cleanedText.
+                if (cleanedText.length > 0) {
                     yield {
                         speaker: 'ai',
                         blocks: [
@@ -1923,19 +1997,32 @@ export class OpenAIProvider extends BaseProvider {
             }
             // Emit accumulated reasoning_content as ONE ThinkingBlock (legacy path)
             // This consolidates token-by-token reasoning from Synthetic API into a single block
+            // Clean Kimi tokens from the accumulated content (not per-chunk) to handle split tokens
             // @plan PLAN-20251202-THINKING.P16
             if (accumulatedReasoningContent.length > 0) {
-                yield {
-                    speaker: 'ai',
-                    blocks: [
-                        {
-                            type: 'thinking',
-                            thought: accumulatedReasoningContent,
-                            sourceField: 'reasoning_content',
-                            isHidden: false,
-                        },
-                    ],
-                };
+                // Extract Kimi tool calls from the complete accumulated reasoning content
+                const { cleanedText: cleanedReasoning, toolCalls: reasoningToolCalls } = this.extractKimiToolCallsFromText(accumulatedReasoningContent);
+                // Emit the cleaned thinking block
+                if (cleanedReasoning.length > 0) {
+                    yield {
+                        speaker: 'ai',
+                        blocks: [
+                            {
+                                type: 'thinking',
+                                thought: cleanedReasoning,
+                                sourceField: 'reasoning_content',
+                                isHidden: false,
+                            },
+                        ],
+                    };
+                }
+                // Emit any tool calls extracted from reasoning content
+                if (reasoningToolCalls.length > 0) {
+                    yield {
+                        speaker: 'ai',
+                        blocks: reasoningToolCalls,
+                    };
+                }
             }
             // Process and emit tool calls using legacy accumulated approach
             if (accumulatedToolCalls.length > 0) {
@@ -1962,6 +2049,7 @@ export class OpenAIProvider extends BaseProvider {
                     };
                     // Add usage metadata if we captured it from streaming
                     if (streamingUsage) {
+                        const cacheMetrics = extractCacheMetrics(streamingUsage);
                         toolCallsContent.metadata = {
                             usage: {
                                 promptTokens: streamingUsage.prompt_tokens || 0,
@@ -1969,6 +2057,9 @@ export class OpenAIProvider extends BaseProvider {
                                 totalTokens: streamingUsage.total_tokens ||
                                     (streamingUsage.prompt_tokens || 0) +
                                         (streamingUsage.completion_tokens || 0),
+                                cachedTokens: cacheMetrics.cachedTokens,
+                                cacheCreationTokens: cacheMetrics.cacheCreationTokens,
+                                cacheMissTokens: cacheMetrics.cacheMissTokens,
                             },
                         };
                     }
@@ -1977,6 +2068,7 @@ export class OpenAIProvider extends BaseProvider {
             }
             // If we have usage information but no tool calls, emit a metadata-only response
             if (streamingUsage && accumulatedToolCalls.length === 0) {
+                const cacheMetrics = extractCacheMetrics(streamingUsage);
                 yield {
                     speaker: 'ai',
                     blocks: [],
@@ -1987,6 +2079,9 @@ export class OpenAIProvider extends BaseProvider {
                             totalTokens: streamingUsage.total_tokens ||
                                 (streamingUsage.prompt_tokens || 0) +
                                     (streamingUsage.completion_tokens || 0),
+                            cachedTokens: cacheMetrics.cachedTokens,
+                            cacheCreationTokens: cacheMetrics.cacheCreationTokens,
+                            cacheMissTokens: cacheMetrics.cacheMissTokens,
                         },
                     },
                 };
@@ -2051,8 +2146,10 @@ export class OpenAIProvider extends BaseProvider {
             }
             const blocks = [];
             // Parse reasoning_content from response (Phase 16 integration)
-            const reasoningBlock = this.parseNonStreamingReasoning(choice.message);
-            logger.debug(() => `[Non-streaming] parseNonStreamingReasoning result: ${reasoningBlock ? `found (${reasoningBlock.thought?.length} chars)` : 'not found'}`, {
+            // Extract embedded Kimi K2 tool calls from reasoning_content (fixes #749)
+            // @requirement REQ-KIMI-REASONING-001.2
+            const { thinking: reasoningBlock, toolCalls: reasoningToolCalls } = this.parseNonStreamingReasoning(choice.message);
+            logger.debug(() => `[Non-streaming] parseNonStreamingReasoning result: ${reasoningBlock ? `found (${reasoningBlock.thought?.length} chars)` : 'not found'}, tool calls: ${reasoningToolCalls.length}`, {
                 hasReasoningContent: 'reasoning_content' in
                     (choice.message ?? {}),
                 messageKeys: Object.keys(choice.message ?? {}),
@@ -2060,6 +2157,11 @@ export class OpenAIProvider extends BaseProvider {
             if (reasoningBlock) {
                 blocks.push(reasoningBlock);
             }
+            // Add tool calls extracted from reasoning_content
+            if (reasoningToolCalls.length > 0) {
+                blocks.push(...reasoningToolCalls);
+                logger.debug(() => `[Non-streaming] Added ${reasoningToolCalls.length} tool calls from reasoning_content`);
+            }
             // Handle text content (strip thinking / reasoning blocks) and Kimi tool sections
             const rawMessageContent = this.coerceMessageContentToString(choice.message?.content);
             let kimiCleanContent;
@@ -2163,6 +2265,7 @@ export class OpenAIProvider extends BaseProvider {
                 };
                 // Add usage metadata from non-streaming response
                 if (completion.usage) {
+                    const cacheMetrics = extractCacheMetrics(completion.usage);
                     responseContent.metadata = {
                         usage: {
                             promptTokens: completion.usage.prompt_tokens || 0,
@@ -2170,6 +2273,9 @@ export class OpenAIProvider extends BaseProvider {
                             totalTokens: completion.usage.total_tokens ||
                                 (completion.usage.prompt_tokens || 0) +
                                     (completion.usage.completion_tokens || 0),
+                            cachedTokens: cacheMetrics.cachedTokens,
+                            cacheCreationTokens: cacheMetrics.cacheCreationTokens,
+                            cacheMissTokens: cacheMetrics.cacheMissTokens,
                         },
                     };
                 }
@@ -2177,6 +2283,7 @@ export class OpenAIProvider extends BaseProvider {
             }
             else if (completion.usage) {
                 // Emit metadata-only response if no content blocks but have usage info
+                const cacheMetrics = extractCacheMetrics(completion.usage);
                 yield {
                     speaker: 'ai',
                     blocks: [],
@@ -2187,6 +2294,9 @@ export class OpenAIProvider extends BaseProvider {
                             totalTokens: completion.usage.total_tokens ||
                                 (completion.usage.prompt_tokens || 0) +
                                     (completion.usage.completion_tokens || 0),
+                            cachedTokens: cacheMetrics.cachedTokens,
+                            cacheCreationTokens: cacheMetrics.cacheCreationTokens,
+                            cacheMissTokens: cacheMetrics.cacheMissTokens,
                         },
                     },
                 };
@@ -2459,7 +2569,7 @@ export class OpenAIProvider extends BaseProvider {
                     });
                     // Dump successful streaming request if enabled
                     if (shouldDumpSuccess) {
-                        await dumpSDKContext('openai', '/v1/chat/completions', requestBody, { streaming: true }, false, baseURL || 'https://api.openai.com');
+                        await dumpSDKContext('openai', '/chat/completions', requestBody, { streaming: true }, false, baseURL || 'https://api.openai.com/v1');
                     }
                     break;
                 }
@@ -2492,7 +2602,7 @@ export class OpenAIProvider extends BaseProvider {
                     // Dump error if enabled
                     if (shouldDumpError) {
                         const dumpErrorMessage = error instanceof Error ? error.message : String(error);
-                        await dumpSDKContext('openai', '/v1/chat/completions', requestBody, { error: dumpErrorMessage }, true, baseURL || 'https://api.openai.com');
+                        await dumpSDKContext('openai', '/chat/completions', requestBody, { error: dumpErrorMessage }, true, baseURL || 'https://api.openai.com/v1');
                     }
                     // Re-throw other errors as-is
                     const capturedErrorMessage = error instanceof Error ? error.message : String(error);
@@ -2530,7 +2640,7 @@ export class OpenAIProvider extends BaseProvider {
                     }));
                     // Dump successful non-streaming request if enabled
                     if (shouldDumpSuccess) {
-                        await dumpSDKContext('openai', '/v1/chat/completions', requestBody, response, false, baseURL || 'https://api.openai.com');
+                        await dumpSDKContext('openai', '/chat/completions', requestBody, response, false, baseURL || 'https://api.openai.com/v1');
                     }
                     break;
                 }
@@ -2569,7 +2679,7 @@ export class OpenAIProvider extends BaseProvider {
                     // Dump error if enabled
                     if (shouldDumpError) {
                         const dumpErrorMessage = error instanceof Error ? error.message : String(error);
-                        await dumpSDKContext('openai', '/v1/chat/completions', requestBody, { error: dumpErrorMessage }, true, baseURL || 'https://api.openai.com');
+                        await dumpSDKContext('openai', '/chat/completions', requestBody, { error: dumpErrorMessage }, true, baseURL || 'https://api.openai.com/v1');
                     }
                     const capturedErrorMessage = error instanceof Error ? error.message : String(error);
                     const status = typeof error === 'object' &&
@@ -2599,9 +2709,9 @@ export class OpenAIProvider extends BaseProvider {
             // Buffer for accumulating text chunks for providers that need it
             let textBuffer = '';
             // Use the same detected format from earlier for consistency
-            const isKimiModel = model.toLowerCase().includes('kimi-k2');
+            const isKimiK2Model = model.toLowerCase().includes('kimi-k2');
             // Buffer text for Qwen format providers and Kimi-K2 to avoid stanza formatting
-            const shouldBufferText = detectedFormat === 'qwen' || isKimiModel;
+            const shouldBufferText = detectedFormat === 'qwen' || isKimiK2Model;
             // Accumulate thinking content across the entire stream to emit as ONE block
             // This handles fragmented <think>word</think> streaming from Synthetic API
             // @plan PLAN-20251202-THINKING.P16
@@ -2671,13 +2781,28 @@ export class OpenAIProvider extends BaseProvider {
                         continue;
                     // Parse reasoning_content from streaming delta (Pipeline path)
                     // ACCUMULATE instead of yielding immediately to handle token-by-token streaming
+                    // Extract embedded Kimi K2 tool calls from reasoning_content (fixes #749)
                     // @plan PLAN-20251202-THINKING.P16
-                    // @requirement REQ-THINK-003.1
-                    const reasoningBlock = this.parseStreamingReasoningDelta(choice.delta);
+                    // @requirement REQ-THINK-003.1, REQ-KIMI-REASONING-001.1
+                    const { thinking: reasoningBlock, toolCalls: reasoningToolCalls } = this.parseStreamingReasoningDelta(choice.delta);
                     if (reasoningBlock) {
                         // Accumulate reasoning content - will emit ONE block later
                         accumulatedReasoningContent += reasoningBlock.thought;
                     }
+                    // Add tool calls extracted from reasoning_content to pipeline
+                    if (reasoningToolCalls.length > 0) {
+                        // Get current pipeline stats to determine next index
+                        const stats = this.toolCallPipeline.getStats();
+                        let baseIndex = stats.collector.totalCalls;
+                        for (const toolCall of reasoningToolCalls) {
+                            // Add complete tool call as fragments to pipeline
+                            this.toolCallPipeline.addFragment(baseIndex, {
+                                name: toolCall.name,
+                                args: JSON.stringify(toolCall.parameters),
+                            });
+                            baseIndex++;
+                        }
+                    }
                     // Check for finish_reason to detect proper stream ending
                     if (choice.finish_reason) {
                         logger.debug(() => `[Streaming] Stream finished with reason: ${choice.finish_reason}`, {
@@ -2698,13 +2823,24 @@ export class OpenAIProvider extends BaseProvider {
                     }
                     // Handle text content - buffer for Qwen format, emit immediately for others
                     // Note: Synthetic API sends content that may duplicate reasoning_content.
-                    // This is the model's behavior - we don't filter it here.
+                    // This is the model's behavior - we don't filter it here as detection is unreliable.
                     // @plan PLAN-20251202-THINKING.P16
                     const rawDeltaContent = this.coerceMessageContentToString(choice.delta?.content);
                     if (rawDeltaContent) {
-                        const deltaContent = isKimiModel
-                            ? rawDeltaContent
-                            : this.sanitizeProviderText(rawDeltaContent);
+                        // For Kimi models, we need to buffer the RAW content without processing
+                        // because Kimi tokens stream incrementally and partial tokens would leak
+                        // through if we try to process them immediately. The buffer will be
+                        // processed when flushed (at sentence boundaries or end of stream).
+                        let deltaContent;
+                        if (isKimiK2Model) {
+                            // For Kimi: Don't process yet - just pass through and let buffering handle it
+                            // We'll extract tool calls and sanitize when we flush the buffer
+                            deltaContent = rawDeltaContent;
+                        }
+                        else {
+                            // For non-Kimi models: sanitize immediately as before
+                            deltaContent = this.sanitizeProviderText(rawDeltaContent);
+                        }
                         if (!deltaContent) {
                             continue;
                         }
@@ -2720,9 +2856,9 @@ export class OpenAIProvider extends BaseProvider {
                             });
                             // Buffer text to avoid stanza formatting
                             textBuffer += deltaContent;
-                            const hasKimiBegin = textBuffer.includes('<|tool_calls_section_begin|>');
-                            const hasKimiEnd = textBuffer.includes('<|tool_calls_section_end|>');
-                            const hasOpenKimiSection = hasKimiBegin && !hasKimiEnd;
+                            const kimiBeginCount = (textBuffer.match(/<\|tool_calls_section_begin\|>/g) || []).length;
+                            const kimiEndCount = (textBuffer.match(/<\|tool_calls_section_end\|>/g) || []).length;
+                            const hasOpenKimiSection = kimiBeginCount > kimiEndCount;
                             // Emit buffered text when we have a complete sentence or paragraph
                             // Look for natural break points, avoiding flush mid Kimi section
                             if (!hasOpenKimiSection &&
@@ -2739,12 +2875,14 @@ export class OpenAIProvider extends BaseProvider {
                                 // @requirement REQ-THINK-003
                                 const tagBasedThinking = this.extractThinkTagsAsBlock(workingText);
                                 if (tagBasedThinking) {
+                                    // Clean Kimi tokens from thinking content before accumulating
+                                    const cleanedThought = this.cleanThinkingContent(tagBasedThinking.thought);
                                     // Accumulate thinking content - don't emit yet
                                     // Use newline to preserve formatting between chunks (not space)
                                     if (accumulatedThinkingContent.length > 0) {
                                         accumulatedThinkingContent += '\n';
                                     }
-                                    accumulatedThinkingContent += tagBasedThinking.thought;
+                                    accumulatedThinkingContent += cleanedThought;
                                     logger.debug(() => `[Streaming] Accumulated thinking: ${accumulatedThinkingContent.length} chars total`);
                                 }
                                 const kimiParsed = this.extractKimiToolCallsFromText(workingText);
@@ -2806,7 +2944,10 @@ export class OpenAIProvider extends BaseProvider {
                                 // Always use sanitized text to strip <think> tags (pipeline streaming)
                                 // Bug fix: Previously Kimi used unsanitized workingText
                                 // @plan PLAN-20251202-THINKING.P16
-                                if (cleanedText.trim().length > 0) {
+                                // Bug fix #721: Emit whitespace-only chunks (e.g., " " between words)
+                                // Previously we used cleanedText.trim().length > 0 which dropped spaces,
+                                // causing "list 5" to become "list5". Now we emit any non-empty cleanedText.
+                                if (cleanedText.length > 0) {
                                     yield {
                                         speaker: 'ai',
                                         blocks: [
@@ -2906,11 +3047,13 @@ export class OpenAIProvider extends BaseProvider {
                 // @plan PLAN-20251202-THINKING.P16
                 const tagBasedThinking = this.extractThinkTagsAsBlock(workingText);
                 if (tagBasedThinking) {
+                    // Clean Kimi tokens from thinking content before accumulating
+                    const cleanedThought = this.cleanThinkingContent(tagBasedThinking.thought);
                     // Use newline to preserve formatting between chunks (not space)
                     if (accumulatedThinkingContent.length > 0) {
                         accumulatedThinkingContent += '\n';
                     }
-                    accumulatedThinkingContent += tagBasedThinking.thought;
+                    accumulatedThinkingContent += cleanedThought;
                 }
                 const kimiParsed = this.extractKimiToolCallsFromText(workingText);
                 if (kimiParsed.toolCalls.length > 0) {
@@ -2969,7 +3112,10 @@ export class OpenAIProvider extends BaseProvider {
                 // Always use sanitized text to strip <think> tags (pipeline final buffer)
                 // Bug fix: Previously Kimi used unsanitized workingText
                 // @plan PLAN-20251202-THINKING.P16
-                if (cleanedText.trim().length > 0) {
+                // Bug fix #721: Emit whitespace-only chunks (e.g., " " between words)
+                // Previously we used cleanedText.trim().length > 0 which dropped spaces,
+                // causing "list 5" to become "list5". Now we emit any non-empty cleanedText.
+                if (cleanedText.length > 0) {
                     yield {
                         speaker: 'ai',
                         blocks: [
@@ -3001,19 +3147,32 @@ export class OpenAIProvider extends BaseProvider {
             }
             // Emit accumulated reasoning_content as ONE ThinkingBlock (pipeline path)
             // This consolidates token-by-token reasoning from Synthetic API into a single block
+            // Clean Kimi tokens from the accumulated content (not per-chunk) to handle split tokens
             // @plan PLAN-20251202-THINKING.P16
             if (accumulatedReasoningContent.length > 0) {
-                yield {
-                    speaker: 'ai',
-                    blocks: [
-                        {
-                            type: 'thinking',
-                            thought: accumulatedReasoningContent,
-                            sourceField: 'reasoning_content',
-                            isHidden: false,
-                        },
-                    ],
-                };
+                // Extract Kimi tool calls from the complete accumulated reasoning content
+                const { cleanedText: cleanedReasoning, toolCalls: reasoningToolCalls } = this.extractKimiToolCallsFromText(accumulatedReasoningContent);
+                // Emit the cleaned thinking block
+                if (cleanedReasoning.length > 0) {
+                    yield {
+                        speaker: 'ai',
+                        blocks: [
+                            {
+                                type: 'thinking',
+                                thought: cleanedReasoning,
+                                sourceField: 'reasoning_content',
+                                isHidden: false,
+                            },
+                        ],
+                    };
+                }
+                // Emit any tool calls extracted from reasoning content
+                if (reasoningToolCalls.length > 0) {
+                    yield {
+                        speaker: 'ai',
+                        blocks: reasoningToolCalls,
+                    };
+                }
             }
             // Process and emit tool calls using the pipeline
             const pipelineResult = await this.toolCallPipeline.process(abortSignal);
@@ -3043,6 +3202,7 @@ export class OpenAIProvider extends BaseProvider {
                     };
                     // Add usage metadata if we captured it from streaming
                     if (streamingUsage) {
+                        const cacheMetrics = extractCacheMetrics(streamingUsage);
                         toolCallsContent.metadata = {
                             usage: {
                                 promptTokens: streamingUsage.prompt_tokens || 0,
@@ -3050,6 +3210,9 @@ export class OpenAIProvider extends BaseProvider {
                                 totalTokens: streamingUsage.total_tokens ||
                                     (streamingUsage.prompt_tokens || 0) +
                                         (streamingUsage.completion_tokens || 0),
+                                cachedTokens: cacheMetrics.cachedTokens,
+                                cacheCreationTokens: cacheMetrics.cacheCreationTokens,
+                                cacheMissTokens: cacheMetrics.cacheMissTokens,
                             },
                         };
                     }
@@ -3059,6 +3222,7 @@ export class OpenAIProvider extends BaseProvider {
             // If we have usage information but no tool calls, emit a metadata-only response
             if (streamingUsage &&
                 this.toolCallPipeline.getStats().collector.totalCalls === 0) {
+                const cacheMetrics = extractCacheMetrics(streamingUsage);
                 yield {
                     speaker: 'ai',
                     blocks: [],
@@ -3069,6 +3233,9 @@ export class OpenAIProvider extends BaseProvider {
                             totalTokens: streamingUsage.total_tokens ||
                                 (streamingUsage.prompt_tokens || 0) +
                                     (streamingUsage.completion_tokens || 0),
+                            cachedTokens: cacheMetrics.cachedTokens,
+                            cacheCreationTokens: cacheMetrics.cacheCreationTokens,
+                            cacheMissTokens: cacheMetrics.cacheMissTokens,
                         },
                     },
                 };
@@ -3221,6 +3388,7 @@ export class OpenAIProvider extends BaseProvider {
                 };
                 // Add usage metadata from non-streaming response
                 if (completion.usage) {
+                    const cacheMetrics = extractCacheMetrics(completion.usage);
                     responseContent.metadata = {
                         usage: {
                             promptTokens: completion.usage.prompt_tokens || 0,
@@ -3228,6 +3396,9 @@ export class OpenAIProvider extends BaseProvider {
                             totalTokens: completion.usage.total_tokens ||
                                 (completion.usage.prompt_tokens || 0) +
                                     (completion.usage.completion_tokens || 0),
+                            cachedTokens: cacheMetrics.cachedTokens,
+                            cacheCreationTokens: cacheMetrics.cacheCreationTokens,
+                            cacheMissTokens: cacheMetrics.cacheMissTokens,
                         },
                     };
                 }
@@ -3235,6 +3406,7 @@ export class OpenAIProvider extends BaseProvider {
             }
             else if (completion.usage) {
                 // Emit metadata-only response if no content blocks but have usage info
+                const cacheMetrics = extractCacheMetrics(completion.usage);
                 yield {
                     speaker: 'ai',
                     blocks: [],
@@ -3245,6 +3417,9 @@ export class OpenAIProvider extends BaseProvider {
                             totalTokens: completion.usage.total_tokens ||
                                 (completion.usage.prompt_tokens || 0) +
                                     (completion.usage.completion_tokens || 0),
+                            cachedTokens: cacheMetrics.cachedTokens,
+                            cacheCreationTokens: cacheMetrics.cacheCreationTokens,
+                            cacheMissTokens: cacheMetrics.cacheMissTokens,
                         },
                     },
                 };
@@ -3279,6 +3454,12 @@ export class OpenAIProvider extends BaseProvider {
             logger.debug(() => `Auto-detected 'kimi' format for K2 model: ${modelName}`);
             return 'kimi';
         }
+        // Check for Mistral models (requires 9-char alphanumeric IDs)
+        // This applies to both hosted API and self-hosted Mistral models
+        if (isMistralModel(modelName)) {
+            logger.debug(() => `Auto-detected 'mistral' format for Mistral model: ${modelName}`);
+            return 'mistral';
+        }
         const lowerModelName = modelName.toLowerCase();
         // Check for GLM-4 models (glm-4, glm-4.5, glm-4.6, glm-4-5, etc.)
         if (lowerModelName.includes('glm-4')) {
@@ -3361,57 +3542,75 @@ export class OpenAIProvider extends BaseProvider {
      * Parse reasoning_content from streaming delta.
      *
      * @plan PLAN-20251202-THINKING.P11, PLAN-20251202-THINKING.P16
-     * @requirement REQ-THINK-003.1, REQ-THINK-003.3, REQ-THINK-003.4
+     * @requirement REQ-THINK-003.1, REQ-THINK-003.3, REQ-THINK-003.4, REQ-KIMI-REASONING-001.1
+     * @issue #749
      */
     parseStreamingReasoningDelta(delta) {
         if (!delta) {
-            return null;
+            return { thinking: null, toolCalls: [] };
         }
         // Access reasoning_content via type assertion since OpenAI SDK doesn't declare it
         const reasoningContent = delta
             .reasoning_content;
         // Handle absent, null, or non-string
         if (!reasoningContent || typeof reasoningContent !== 'string') {
-            return null;
-        }
-        // Handle empty string or whitespace-only
-        if (reasoningContent.trim().length === 0) {
-            return null;
-        }
-        return {
-            type: 'thinking',
-            thought: reasoningContent,
-            sourceField: 'reasoning_content',
-            isHidden: false,
-        };
+            return { thinking: null, toolCalls: [] };
+        }
+        // Handle empty string only - preserve whitespace-only content (spaces, tabs)
+        // to maintain proper formatting in accumulated reasoning (fixes issue #721)
+        if (reasoningContent.length === 0) {
+            return { thinking: null, toolCalls: [] };
+        }
+        // Extract Kimi K2 tool calls embedded in reasoning_content (fixes issue #749)
+        const { cleanedText, toolCalls } = this.extractKimiToolCallsFromText(reasoningContent);
+        // For streaming, preserve whitespace-only content for proper formatting (issue #721)
+        // Only return null if the cleaned text is empty (length 0)
+        const thinkingBlock = cleanedText.length === 0
+            ? null
+            : {
+                type: 'thinking',
+                thought: cleanedText,
+                sourceField: 'reasoning_content',
+                isHidden: false,
+            };
+        return { thinking: thinkingBlock, toolCalls };
     }
     /**
      * Parse reasoning_content from non-streaming message.
      *
      * @plan PLAN-20251202-THINKING.P11, PLAN-20251202-THINKING.P16
-     * @requirement REQ-THINK-003.2, REQ-THINK-003.3, REQ-THINK-003.4
+     * @requirement REQ-THINK-003.2, REQ-THINK-003.3, REQ-THINK-003.4, REQ-KIMI-REASONING-001.2
+     * @issue #749
      */
     parseNonStreamingReasoning(message) {
         if (!message) {
-            return null;
+            return { thinking: null, toolCalls: [] };
         }
         // Access reasoning_content via type assertion since OpenAI SDK doesn't declare it
         const reasoningContent = message
             .reasoning_content;
         // Handle absent, null, or non-string
         if (!reasoningContent || typeof reasoningContent !== 'string') {
-            return null;
+            return { thinking: null, toolCalls: [] };
         }
-        // Handle empty string or whitespace-only
+        // Handle empty string or whitespace-only - for non-streaming complete responses,
+        // whitespace-only reasoning is unusual and should be treated as no reasoning
         if (reasoningContent.trim().length === 0) {
-            return null;
-        }
-        return {
-            type: 'thinking',
-            thought: reasoningContent,
-            sourceField: 'reasoning_content',
-            isHidden: false,
-        };
+            return { thinking: null, toolCalls: [] };
+        }
+        // Extract Kimi K2 tool calls embedded in reasoning_content (fixes issue #749)
+        const { cleanedText, toolCalls } = this.extractKimiToolCallsFromText(reasoningContent);
+        // For non-streaming, trim whitespace after extraction
+        const trimmedText = cleanedText.trim();
+        const thinkingBlock = trimmedText.length === 0
+            ? null
+            : {
+                type: 'thinking',
+                thought: trimmedText,
+                sourceField: 'reasoning_content',
+                isHidden: false,
+            };
+        return { thinking: thinkingBlock, toolCalls };
     }
 }
 //# sourceMappingURL=OpenAIProvider.js.map