@vybestack/llxprt-code-core 0.6.1 → 0.7.0-nightly.251206.43b97dbf4

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

@@ -22,9 +22,11 @@ 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 { BaseProvider, } from '../BaseProvider.js';
 import { DebugLogger } from '../../debug/index.js';
 import { ToolFormatter } from '../../tools/ToolFormatter.js';
+import { convertToolsToOpenAI } from './schemaConverter.js';
 import { GemmaToolCallParser } from '../../parsers/TextToolCallParser.js';
 import { processToolParameters } from '../../tools/doubleEscapeUtils.js';
 import { getCoreSystemPromptAsync } from '../../core/prompts.js';
@@ -36,6 +38,7 @@ import { ensureJsonSafe } from '../../utils/unicodeUtils.js';
 import { ToolCallPipeline } from './ToolCallPipeline.js';
 import { buildToolResponsePayload, EMPTY_TOOL_RESULT_PLACEHOLDER, } from '../utils/toolResponsePayload.js';
 import { isLocalEndpoint } from '../utils/localEndpoint.js';
+import { filterThinkingForContext, thinkingToReasoningField, extractThinkingBlocks, } from '../reasoning/reasoningUtils.js';
 const MAX_TOOL_RESPONSE_CHARS = 1024;
 const MAX_TOOL_RESPONSE_RETRY_CHARS = 512;
 const TOOL_ARGS_PREVIEW_LENGTH = 500;
@@ -218,6 +221,278 @@ export class OpenAIProvider extends BaseProvider {
         }
         return new OpenAI(clientOptions);
     }
+    /**
+     * Coerce provider "content" (which may be a string or an array-of-parts)
+     * into a plain string. Defensive for OpenAI-compatible providers that emit
+     * structured content blocks.
+     */
+    coerceMessageContentToString(content) {
+        if (typeof content === 'string') {
+            return content;
+        }
+        if (Array.isArray(content)) {
+            const parts = [];
+            for (const part of content) {
+                if (!part)
+                    continue;
+                if (typeof part === 'string') {
+                    parts.push(part);
+                }
+                else if (typeof part === 'object' &&
+                    part !== null &&
+                    'text' in part &&
+                    typeof part.text === 'string') {
+                    parts.push(part.text);
+                }
+            }
+            return parts.length ? parts.join('') : undefined;
+        }
+        return undefined;
+    }
+    /**
+     * Strip provider-specific "thinking" / reasoning markup from visible text.
+     * This prevents DeepSeek / Kimi-style <think> blocks from leaking into
+     * user-visible output or tool arguments.
+     */
+    sanitizeProviderText(text) {
+        if (text === null || text === undefined) {
+            return '';
+        }
+        const logger = this.getLogger();
+        let str = typeof text === 'string' ? text : String(text);
+        const beforeLen = str.length;
+        const hadReasoningTags = /<(?:think|thinking|analysis)>|<\/(?:think|thinking|analysis)>/i.test(str);
+        // DeepSeek / generic <think>...</think> blocks.
+        // Replace with a single space to preserve word spacing when tags appear mid-sentence.
+        // This prevents "these<think>...</think>5" from becoming "these5" instead of "these 5".
+        // Multiple consecutive spaces will be collapsed below.
+        str = str.replace(/<think>[\s\S]*?<\/think>/gi, ' ');
+        // Alternative reasoning tags some providers use.
+        str = str.replace(/<thinking>[\s\S]*?<\/thinking>/gi, ' ');
+        str = str.replace(/<analysis>[\s\S]*?<\/analysis>/gi, ' ');
+        // Clean up stray unmatched tags - replace with space to preserve word separation.
+        str = str.replace(/<\/?(?:think|thinking|analysis)>/gi, ' ');
+        // Only clean up whitespace if we had reasoning tags to strip
+        // 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
+            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();
+        }
+        const afterLen = str.length;
+        if (hadReasoningTags && afterLen !== beforeLen) {
+            logger.debug(() => `[OpenAIProvider] Stripped reasoning tags`, {
+                beforeLen,
+                afterLen,
+            });
+        }
+        return str;
+    }
+    /**
+     * Extract thinking content from <think>, <thinking>, or <analysis> tags
+     * and return it as a ThinkingBlock. Returns null if no thinking tags found.
+     *
+     * This must be called BEFORE sanitizeProviderText which strips these tags.
+     *
+     * Handles two formats:
+     * 1. Standard: <think>Full thinking paragraph here...</think>
+     * 2. Fragmented (Synthetic API): <think>word</think><think>word</think>...
+     *
+     * For fragmented format, joins with spaces. For standard, joins with newlines.
+     *
+     * @plan PLAN-20251202-THINKING.P16
+     * @requirement REQ-THINK-003
+     */
+    extractThinkTagsAsBlock(text) {
+        if (!text) {
+            return null;
+        }
+        // Collect all thinking content from various tag formats
+        // Note: We only trim leading/trailing whitespace from each part, not internal newlines
+        // This preserves formatting like numbered lists within thinking content
+        const thinkingParts = [];
+        // Match <think>...</think>
+        const thinkMatches = text.matchAll(/<think>([\s\S]*?)<\/think>/gi);
+        for (const match of thinkMatches) {
+            const content = match[1];
+            if (content?.trim()) {
+                // Preserve internal newlines but remove leading/trailing whitespace
+                thinkingParts.push(content.trim());
+            }
+        }
+        // Match <thinking>...</thinking>
+        const thinkingMatches = text.matchAll(/<thinking>([\s\S]*?)<\/thinking>/gi);
+        for (const match of thinkingMatches) {
+            const content = match[1];
+            if (content?.trim()) {
+                thinkingParts.push(content.trim());
+            }
+        }
+        // Match <analysis>...</analysis>
+        const analysisMatches = text.matchAll(/<analysis>([\s\S]*?)<\/analysis>/gi);
+        for (const match of analysisMatches) {
+            const content = match[1];
+            if (content?.trim()) {
+                thinkingParts.push(content.trim());
+            }
+        }
+        if (thinkingParts.length === 0) {
+            return null;
+        }
+        // Detect fragmented format: many short parts (likely token-by-token streaming)
+        // If average part length is very short (< 10 chars) and we have many parts,
+        // it's likely fragmented and should be joined with spaces
+        const avgPartLength = thinkingParts.reduce((sum, p) => sum + p.length, 0) /
+            thinkingParts.length;
+        const isFragmented = thinkingParts.length > 5 && avgPartLength < 15;
+        // Join with space for fragmented, newlines for standard multi-paragraph thinking
+        const combinedThought = isFragmented
+            ? thinkingParts.join(' ')
+            : thinkingParts.join('\n\n');
+        this.getLogger().debug(() => `[OpenAIProvider] Extracted thinking from tags: ${combinedThought.length} chars`, { tagCount: thinkingParts.length, isFragmented, avgPartLength });
+        return {
+            type: 'thinking',
+            thought: combinedThought,
+            sourceField: 'think_tags',
+            isHidden: false,
+        };
+    }
+    /**
+     * Normalize tool name by stripping Kimi-K2 style prefixes.
+     *
+     * Handles malformed tool names where the model concatenates prefixes like
+     * "functions" or "call_functions" with the actual tool name:
+     * - "functionslist_directory" -> "list_directory"
+     * - "call_functionslist_directory6" -> "list_directory"
+     * - "call_functionsglob7" -> "glob"
+     */
+    normalizeToolName(name) {
+        let normalized = (name || '').trim();
+        // Strip Kimi-K2 style prefixes where model concatenates "functions" or "call_functions"
+        // with the actual tool name (e.g., "functionslist_directory" -> "list_directory")
+        // Pattern: (call_)?functions<actual_tool_name><optional_number>
+        const kimiPrefixMatch = /^(?:call_)?functions([a-z_]+[a-z])(\d*)$/i.exec(normalized);
+        if (kimiPrefixMatch) {
+            const originalName = normalized;
+            normalized = kimiPrefixMatch[1];
+            this.getLogger().debug(() => `[OpenAIProvider] Stripped Kimi-style prefix from tool name: "${originalName}" -> "${normalized}"`);
+        }
+        return normalized.toLowerCase();
+    }
+    /**
+     * Sanitize raw tool argument payloads before JSON parsing:
+     * - Remove thinking blocks (<think>...</think>, etc.).
+     * - Strip Markdown code fences (```json ... ```).
+     * - Try to isolate the main JSON object if wrapped in prose.
+     */
+    sanitizeToolArgumentsString(raw) {
+        if (raw === null || raw === undefined) {
+            return '{}';
+        }
+        let text;
+        if (typeof raw === 'string') {
+            text = raw;
+        }
+        else {
+            try {
+                text = JSON.stringify(raw);
+            }
+            catch {
+                text = String(raw);
+            }
+        }
+        text = text.trim();
+        // Strip fenced code blocks like ```json { ... } ```.
+        if (text.startsWith('```')) {
+            text = text.replace(/^```[a-zA-Z0-9_-]*\s*/m, '');
+            text = text.replace(/```$/m, '');
+            text = text.trim();
+        }
+        // Remove provider reasoning / thinking markup.
+        text = this.sanitizeProviderText(text);
+        // If provider wrapped JSON in explanation text, try to isolate the object.
+        const firstBrace = text.indexOf('{');
+        const lastBrace = text.lastIndexOf('}');
+        if (firstBrace !== -1 && lastBrace !== -1 && lastBrace > firstBrace) {
+            const candidate = text.slice(firstBrace, lastBrace + 1).trim();
+            if (candidate.startsWith('{') && candidate.endsWith('}')) {
+                return candidate;
+            }
+        }
+        return text.length ? text : '{}';
+    }
+    /**
+     * Parse Kimi-K2 `<|tool_calls_section_begin|> ... <|tool_calls_section_end|>`
+     * blocks out of a text string.
+     *
+     * - Returns cleanedText with the whole section removed.
+     * - Returns ToolCallBlock[] constructed from the section contents.
+     *
+     * This is used for HF/vLLM-style Kimi deployments where `tool_calls` is empty
+     * and all tool info is only encoded in the text template.
+     */
+    extractKimiToolCallsFromText(raw) {
+        if (!raw || !raw.includes('<|tool_calls_section_begin|>')) {
+            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;
+                    }
+                    // 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,
+            });
+        }
+        // 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 };
+    }
     /**
      * @plan:PLAN-20251023-STATELESS-HARDENING.P09
      * @requirement:REQ-SP4-002
@@ -685,6 +960,113 @@ export class OpenAIProvider extends BaseProvider {
         // This ensures each tool message has a corresponding tool_calls in previous message
         return this.validateToolMessageSequence(messages);
     }
+    /**
+     * Build messages with optional reasoning_content based on settings.
+     *
+     * @plan PLAN-20251202-THINKING.P14
+     * @requirement REQ-THINK-004, REQ-THINK-006
+     */
+    buildMessagesWithReasoning(contents, options, toolFormat) {
+        // Read settings with defaults
+        const stripPolicy = options.settings.get('reasoning.stripFromContext') ??
+            'none';
+        const includeInContext = options.settings.get('reasoning.includeInContext') ?? false;
+        // Apply strip policy first
+        const filteredContents = filterThinkingForContext(contents, stripPolicy);
+        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)
+            : null;
+        // Helper to resolve tool call IDs based on format
+        const resolveToolCallId = (tc) => {
+            if (toolIdMapper) {
+                return toolIdMapper.resolveToolCallId(tc);
+            }
+            return this.normalizeToOpenAIToolId(tc.id);
+        };
+        // Helper to resolve tool response IDs based on format
+        const resolveToolResponseId = (tr) => {
+            if (toolIdMapper) {
+                return toolIdMapper.resolveToolResponseId(tr);
+            }
+            return this.normalizeToOpenAIToolId(tr.callId);
+        };
+        for (const content of filteredContents) {
+            if (content.speaker === 'human') {
+                // Convert human messages to user messages
+                const textBlocks = content.blocks.filter((b) => b.type === 'text');
+                const text = textBlocks.map((b) => b.text).join('\n');
+                if (text) {
+                    messages.push({
+                        role: 'user',
+                        content: text,
+                    });
+                }
+            }
+            else if (content.speaker === 'ai') {
+                // Convert AI messages with optional reasoning_content
+                const textBlocks = content.blocks.filter((b) => b.type === 'text');
+                const text = textBlocks.map((b) => b.text).join('\n');
+                const thinkingBlocks = extractThinkingBlocks(content);
+                const toolCalls = content.blocks.filter((b) => b.type === 'tool_call');
+                if (toolCalls.length > 0) {
+                    // Assistant message with tool calls
+                    const baseMessage = {
+                        role: 'assistant',
+                        content: text || null,
+                        tool_calls: toolCalls.map((tc) => ({
+                            id: resolveToolCallId(tc),
+                            type: 'function',
+                            function: {
+                                name: tc.name,
+                                arguments: this.normalizeToolCallArguments(tc.parameters),
+                            },
+                        })),
+                    };
+                    if (includeInContext && thinkingBlocks.length > 0) {
+                        const messageWithReasoning = baseMessage;
+                        messageWithReasoning.reasoning_content =
+                            thinkingToReasoningField(thinkingBlocks);
+                        messages.push(messageWithReasoning);
+                    }
+                    else {
+                        messages.push(baseMessage);
+                    }
+                }
+                else if (textBlocks.length > 0 || thinkingBlocks.length > 0) {
+                    // Plain assistant message
+                    const baseMessage = {
+                        role: 'assistant',
+                        content: text,
+                    };
+                    if (includeInContext && thinkingBlocks.length > 0) {
+                        const messageWithReasoning = baseMessage;
+                        messageWithReasoning.reasoning_content =
+                            thinkingToReasoningField(thinkingBlocks);
+                        messages.push(messageWithReasoning);
+                    }
+                    else {
+                        messages.push(baseMessage);
+                    }
+                }
+            }
+            else if (content.speaker === 'tool') {
+                // Convert tool responses
+                const toolResponses = content.blocks.filter((b) => b.type === 'tool_response');
+                for (const tr of toolResponses) {
+                    messages.push({
+                        role: 'tool',
+                        content: this.buildToolResponseContent(tr, options.config),
+                        tool_call_id: resolveToolResponseId(tr),
+                    });
+                }
+            }
+        }
+        // Validate tool message sequence to prevent API errors
+        return this.validateToolMessageSequence(messages);
+    }
     /**
      * Validates tool message sequence to ensure each tool message has a corresponding tool_calls
      * This prevents "messages with role 'tool' must be a response to a preceeding message with 'tool_calls'" errors
@@ -698,6 +1080,18 @@ export class OpenAIProvider extends BaseProvider {
         const logger = this.getLogger();
         const validatedMessages = [...messages];
         let removedCount = 0;
+        // Debug: Log the full message sequence for tool call analysis
+        logger.debug(() => `[OpenAIProvider] validateToolMessageSequence: analyzing ${messages.length} messages`, {
+            messageRoles: messages.map((m) => m.role),
+            toolCallIds: messages
+                .filter((m) => m.role === 'assistant' &&
+                'tool_calls' in m &&
+                Array.isArray(m.tool_calls))
+                .flatMap((m) => m.tool_calls?.map((tc) => tc.id) ?? []),
+            toolResponseIds: messages
+                .filter((m) => m.role === 'tool')
+                .map((m) => m.tool_call_id),
+        });
         // Check if there are any tool_calls in conversation
         // If no tool_calls exist, this might be isolated tool response testing - skip validation
         const hasToolCallsInConversation = validatedMessages.some((msg) => msg.role === 'assistant' &&
@@ -826,13 +1220,8 @@ export class OpenAIProvider extends BaseProvider {
                 metadataKeys: Object.keys(metadata ?? {}),
             });
         }
-        // Convert IContent to OpenAI messages format
-        const configForMessages = options.config ?? options.runtime?.config ?? this.globalConfig;
-        const messages = this.convertToOpenAIMessages(contents, toolReplayMode, configForMessages);
-        if (logger.enabled && toolReplayMode !== 'native') {
-            logger.debug(() => `[OpenAIProvider] Using textual tool replay mode for model '${model}'`);
-        }
-        // Detect the tool format to use (once at the start of the method)
+        // Detect the tool format to use BEFORE building messages
+        // This is needed so that Kimi K2 tool IDs can be generated in the correct format
         const detectedFormat = this.detectToolFormat();
         // Log the detected format for debugging
         logger.debug(() => `[OpenAIProvider] Using tool format '${detectedFormat}' for model '${model}'`, {
@@ -840,8 +1229,18 @@ export class OpenAIProvider extends BaseProvider {
             detectedFormat,
             provider: this.name,
         });
-        // Convert Gemini format tools to the detected format
-        let formattedTools = toolFormatter.convertGeminiToFormat(tools, detectedFormat);
+        // Convert IContent to OpenAI messages format
+        // Use buildMessagesWithReasoning for reasoning-aware message building
+        // Pass detectedFormat so that Kimi K2 tool IDs are generated correctly
+        const messages = toolReplayMode === 'native'
+            ? this.buildMessagesWithReasoning(contents, options, detectedFormat)
+            : this.convertToOpenAIMessages(contents, toolReplayMode, options.config ?? options.runtime?.config ?? this.globalConfig);
+        if (logger.enabled && toolReplayMode !== 'native') {
+            logger.debug(() => `[OpenAIProvider] Using textual tool replay mode for model '${model}'`);
+        }
+        // Convert Gemini format tools to OpenAI format using the schema converter
+        // This ensures required fields are always present in tool schemas
+        let formattedTools = convertToolsToOpenAI(tools);
         // CRITICAL FIX: Ensure we never pass an empty tools array
         // The OpenAI API errors when tools=[] but a tool call is attempted
         if (Array.isArray(formattedTools) && formattedTools.length === 0) {
@@ -1106,16 +1505,37 @@ 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
-            // Buffer text for Qwen format providers to avoid stanza formatting
-            const shouldBufferText = detectedFormat === 'qwen';
+            const isKimiModel = model.toLowerCase().includes('kimi-k2');
+            // Buffer text for Qwen format providers and Kimi-K2 to avoid stanza formatting
+            const shouldBufferText = detectedFormat === 'qwen' || isKimiModel;
+            // 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
+            let accumulatedThinkingContent = '';
+            let hasEmittedThinking = false;
+            // Accumulate reasoning_content from streaming deltas (legacy path)
+            // Synthetic API sends reasoning token-by-token, so we accumulate to emit ONE block
+            // @plan PLAN-20251202-THINKING.P16
+            let accumulatedReasoningContent = '';
             // Track token usage from streaming chunks
             let streamingUsage = null;
+            // Track total chunks for debugging empty responses
+            let totalChunksReceived = 0;
             try {
                 // Handle streaming response
                 for await (const chunk of response) {
+                    totalChunksReceived++;
                     if (abortSignal?.aborted) {
                         break;
                     }
+                    // Debug: Log first few chunks and every 10th chunk to understand stream behavior
+                    if (totalChunksReceived <= 3 || totalChunksReceived % 10 === 0) {
+                        logger.debug(() => `[Streaming legacy] Chunk #${totalChunksReceived} received`, {
+                            hasChoices: !!chunk.choices?.length,
+                            firstChoiceDelta: chunk.choices?.[0]?.delta,
+                            finishReason: chunk.choices?.[0]?.finish_reason,
+                        });
+                    }
                     const chunkRecord = chunk;
                     let parsedData;
                     const rawData = chunkRecord?.data;
@@ -1152,6 +1572,14 @@ export class OpenAIProvider extends BaseProvider {
                     const choice = chunk.choices?.[0];
                     if (!choice)
                         continue;
+                    // Parse reasoning_content from streaming delta (Phase 16 integration)
+                    // ACCUMULATE instead of yielding immediately to handle token-by-token streaming
+                    // @plan PLAN-20251202-THINKING.P16
+                    const reasoningBlock = this.parseStreamingReasoningDelta(choice.delta);
+                    if (reasoningBlock) {
+                        // Accumulate reasoning content - will emit ONE block later
+                        accumulatedReasoningContent += reasoningBlock.thought;
+                    }
                     // Check for finish_reason to detect proper stream ending
                     if (choice.finish_reason) {
                         logger.debug(() => `[Streaming] Stream finished with reason: ${choice.finish_reason}`, {
@@ -1165,23 +1593,23 @@ export class OpenAIProvider extends BaseProvider {
                         if (choice.finish_reason === 'length') {
                             logger.debug(() => `Response truncated due to length limit for model ${model}`);
                         }
-                        // Flush any buffered text when stream finishes
-                        if (textBuffer.length > 0) {
-                            yield {
-                                speaker: 'ai',
-                                blocks: [
-                                    {
-                                        type: 'text',
-                                        text: textBuffer,
-                                    },
-                                ],
-                            };
-                            textBuffer = '';
-                        }
+                        // Don't flush buffer here on finish - let the final buffer handling
+                        // after the loop process it with proper sanitization and think tag extraction
+                        // This was causing unsanitized <think> tags to leak into output (legacy path)
+                        // @plan PLAN-20251202-THINKING.P16
                     }
                     // Handle text content - buffer for Qwen format, emit immediately for others
-                    const deltaContent = choice.delta?.content;
-                    if (deltaContent) {
+                    // Note: Synthetic API sends content that may duplicate reasoning_content.
+                    // This is the model's behavior - we don't filter it here.
+                    // @plan PLAN-20251202-THINKING.P16
+                    const rawDeltaContent = this.coerceMessageContentToString(choice.delta?.content);
+                    if (rawDeltaContent) {
+                        const deltaContent = isKimiModel
+                            ? rawDeltaContent
+                            : this.sanitizeProviderText(rawDeltaContent);
+                        if (!deltaContent) {
+                            continue;
+                        }
                         _accumulatedText += deltaContent;
                         // Debug log for providers that need buffering
                         if (shouldBufferText) {
@@ -1194,22 +1622,103 @@ 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;
                             // Emit buffered text when we have a complete sentence or paragraph
-                            // Look for natural break points
-                            if (textBuffer.includes('\n') ||
-                                textBuffer.endsWith('. ') ||
-                                textBuffer.endsWith('! ') ||
-                                textBuffer.endsWith('? ') ||
-                                textBuffer.length > 100) {
-                                yield {
-                                    speaker: 'ai',
-                                    blocks: [
-                                        {
-                                            type: 'text',
-                                            text: textBuffer,
-                                        },
-                                    ],
-                                };
+                            // Look for natural break points, but avoid flushing mid Kimi section
+                            if (!hasOpenKimiSection &&
+                                (textBuffer.includes('\n') ||
+                                    textBuffer.endsWith('. ') ||
+                                    textBuffer.endsWith('! ') ||
+                                    textBuffer.endsWith('? ') ||
+                                    textBuffer.length > 100)) {
+                                const parsedToolCalls = [];
+                                let workingText = textBuffer;
+                                // Extract <think> tags and ACCUMULATE instead of emitting immediately (legacy path)
+                                // This handles fragmented <think>word</think> streaming from Synthetic API
+                                // @plan PLAN-20251202-THINKING.P16
+                                // @requirement REQ-THINK-003
+                                const tagBasedThinking = this.extractThinkTagsAsBlock(workingText);
+                                if (tagBasedThinking) {
+                                    // 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;
+                                    logger.debug(() => `[Streaming legacy] Accumulated thinking: ${accumulatedThinkingContent.length} chars total`);
+                                }
+                                const kimiParsed = this.extractKimiToolCallsFromText(workingText);
+                                if (kimiParsed.toolCalls.length > 0) {
+                                    parsedToolCalls.push(...kimiParsed.toolCalls);
+                                    logger.debug(() => `[OpenAIProvider] Streaming buffer (legacy) parsed Kimi tool calls`, {
+                                        count: kimiParsed.toolCalls.length,
+                                        bufferLength: workingText.length,
+                                        cleanedLength: kimiParsed.cleanedText.length,
+                                    });
+                                }
+                                workingText = kimiParsed.cleanedText;
+                                const parsingText = this.sanitizeProviderText(workingText);
+                                let cleanedText = parsingText;
+                                try {
+                                    const parsedResult = this.textToolParser.parse(parsingText);
+                                    if (parsedResult.toolCalls.length > 0) {
+                                        parsedToolCalls.push(...parsedResult.toolCalls.map((call) => ({
+                                            type: 'tool_call',
+                                            id: `text_tool_${Date.now()}_${Math.random()
+                                                .toString(36)
+                                                .substring(7)}`,
+                                            name: this.normalizeToolName(call.name),
+                                            parameters: call.arguments,
+                                        })));
+                                        cleanedText = parsedResult.cleanedContent;
+                                    }
+                                }
+                                catch (error) {
+                                    const logger = this.getLogger();
+                                    logger.debug(() => `TextToolCallParser failed on buffered text: ${error}`);
+                                }
+                                // Emit accumulated thinking BEFORE tool calls or text content (legacy path)
+                                // This ensures thinking appears first in the response
+                                // @plan PLAN-20251202-THINKING.P16
+                                if (!hasEmittedThinking &&
+                                    accumulatedThinkingContent.length > 0 &&
+                                    (parsedToolCalls.length > 0 || cleanedText.trim().length > 0)) {
+                                    yield {
+                                        speaker: 'ai',
+                                        blocks: [
+                                            {
+                                                type: 'thinking',
+                                                thought: accumulatedThinkingContent,
+                                                sourceField: 'think_tags',
+                                                isHidden: false,
+                                            },
+                                        ],
+                                    };
+                                    hasEmittedThinking = true;
+                                    logger.debug(() => `[Streaming legacy] Emitted accumulated thinking: ${accumulatedThinkingContent.length} chars`);
+                                }
+                                if (parsedToolCalls.length > 0) {
+                                    yield {
+                                        speaker: 'ai',
+                                        blocks: parsedToolCalls,
+                                    };
+                                }
+                                // 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) {
+                                    yield {
+                                        speaker: 'ai',
+                                        blocks: [
+                                            {
+                                                type: 'text',
+                                                text: cleanedText,
+                                            },
+                                        ],
+                                    };
+                                }
                                 textBuffer = '';
                             }
                         }
@@ -1308,19 +1817,45 @@ export class OpenAIProvider extends BaseProvider {
             }
             // Check buffered text for <tool_call> format before flushing as plain text
             if (textBuffer.length > 0) {
-                // Try to parse <tool_call> format from buffered text
-                let parsedToolCalls = [];
-                let cleanedText = textBuffer;
+                const parsedToolCalls = [];
+                let workingText = textBuffer;
+                // Note: Synthetic API sends reasoning via both reasoning_content AND content fields.
+                // This is the model's behavior - we don't strip it since the model is the source.
+                // The user can configure reasoning display settings if they don't want duplicates.
+                // @plan PLAN-20251202-THINKING.P16
+                // Extract any remaining <think> tags from final buffer (legacy path)
+                // @plan PLAN-20251202-THINKING.P16
+                const tagBasedThinking = this.extractThinkTagsAsBlock(workingText);
+                if (tagBasedThinking) {
+                    // Use newline to preserve formatting between chunks (not space)
+                    if (accumulatedThinkingContent.length > 0) {
+                        accumulatedThinkingContent += '\n';
+                    }
+                    accumulatedThinkingContent += tagBasedThinking.thought;
+                }
+                const kimiParsed = this.extractKimiToolCallsFromText(workingText);
+                if (kimiParsed.toolCalls.length > 0) {
+                    parsedToolCalls.push(...kimiParsed.toolCalls);
+                    this.getLogger().debug(() => `[OpenAIProvider] Final buffer flush (legacy) parsed Kimi tool calls`, {
+                        count: kimiParsed.toolCalls.length,
+                        bufferLength: workingText.length,
+                        cleanedLength: kimiParsed.cleanedText.length,
+                    });
+                }
+                workingText = kimiParsed.cleanedText;
+                const parsingText = this.sanitizeProviderText(workingText);
+                let cleanedText = parsingText;
                 try {
-                    const parsedResult = this.textToolParser.parse(textBuffer);
+                    const parsedResult = this.textToolParser.parse(parsingText);
                     if (parsedResult.toolCalls.length > 0) {
-                        // Convert parsed tool calls to ToolCallBlock format
-                        parsedToolCalls = parsedResult.toolCalls.map((call) => ({
+                        parsedToolCalls.push(...parsedResult.toolCalls.map((call) => ({
                             type: 'tool_call',
-                            id: `text_tool_${Date.now()}_${Math.random().toString(36).substring(7)}`,
-                            name: call.name,
+                            id: `text_tool_${Date.now()}_${Math.random()
+                                .toString(36)
+                                .substring(7)}`,
+                            name: this.normalizeToolName(call.name),
                             parameters: call.arguments,
-                        }));
+                        })));
                         cleanedText = parsedResult.cleanedContent;
                     }
                 }
@@ -1328,14 +1863,33 @@ export class OpenAIProvider extends BaseProvider {
                     const logger = this.getLogger();
                     logger.debug(() => `TextToolCallParser failed on buffered text: ${error}`);
                 }
-                // Emit tool calls from text parsing first
+                // Emit accumulated thinking BEFORE tool calls or text content (legacy path)
+                // @plan PLAN-20251202-THINKING.P16
+                if (!hasEmittedThinking &&
+                    accumulatedThinkingContent.length > 0 &&
+                    (parsedToolCalls.length > 0 || cleanedText.trim().length > 0)) {
+                    yield {
+                        speaker: 'ai',
+                        blocks: [
+                            {
+                                type: 'thinking',
+                                thought: accumulatedThinkingContent,
+                                sourceField: 'think_tags',
+                                isHidden: false,
+                            },
+                        ],
+                    };
+                    hasEmittedThinking = true;
+                }
                 if (parsedToolCalls.length > 0) {
                     yield {
                         speaker: 'ai',
                         blocks: parsedToolCalls,
                     };
                 }
-                // Then emit any remaining cleaned text
+                // 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) {
                     yield {
                         speaker: 'ai',
@@ -1349,18 +1903,54 @@ export class OpenAIProvider extends BaseProvider {
                 }
                 textBuffer = '';
             }
+            // Emit any remaining accumulated thinking that wasn't emitted yet (legacy path)
+            // (e.g., if entire response was just thinking with no content)
+            // @plan PLAN-20251202-THINKING.P16
+            if (!hasEmittedThinking && accumulatedThinkingContent.length > 0) {
+                yield {
+                    speaker: 'ai',
+                    blocks: [
+                        {
+                            type: 'thinking',
+                            thought: accumulatedThinkingContent,
+                            sourceField: 'think_tags',
+                            isHidden: false,
+                        },
+                    ],
+                };
+                hasEmittedThinking = true;
+            }
+            // Emit accumulated reasoning_content as ONE ThinkingBlock (legacy path)
+            // This consolidates token-by-token reasoning from Synthetic API into a single block
+            // @plan PLAN-20251202-THINKING.P16
+            if (accumulatedReasoningContent.length > 0) {
+                yield {
+                    speaker: 'ai',
+                    blocks: [
+                        {
+                            type: 'thinking',
+                            thought: accumulatedReasoningContent,
+                            sourceField: 'reasoning_content',
+                            isHidden: false,
+                        },
+                    ],
+                };
+            }
             // Process and emit tool calls using legacy accumulated approach
             if (accumulatedToolCalls.length > 0) {
                 const blocks = [];
                 for (const tc of accumulatedToolCalls) {
                     if (!tc)
                         continue;
+                    const sanitizedArgs = this.sanitizeToolArgumentsString(tc.function.arguments);
+                    // Normalize tool name (handles Kimi-K2 style prefixes)
+                    const normalizedName = this.normalizeToolName(tc.function.name || '');
                     // Process tool parameters with double-escape handling
-                    const processedParameters = processToolParameters(tc.function.arguments || '', tc.function.name || '');
+                    const processedParameters = processToolParameters(sanitizedArgs, normalizedName);
                     blocks.push({
                         type: 'tool_call',
                         id: this.normalizeToHistoryToolId(tc.id),
-                        name: tc.function.name || '',
+                        name: normalizedName,
                         parameters: processedParameters,
                     });
                 }
@@ -1400,6 +1990,40 @@ export class OpenAIProvider extends BaseProvider {
                     },
                 };
             }
+            // Detect and warn about empty streaming responses (common with Kimi K2 after tool calls)
+            // Only warn if we truly got nothing - not even reasoning content
+            if (_accumulatedText.length === 0 &&
+                accumulatedToolCalls.length === 0 &&
+                textBuffer.length === 0 &&
+                accumulatedReasoningContent.length === 0 &&
+                accumulatedThinkingContent.length === 0) {
+                // Provide actionable guidance for users
+                const isKimi = model.toLowerCase().includes('kimi');
+                const isSynthetic = (baseURL ?? this.getBaseURL())?.includes('synthetic') ?? false;
+                const troubleshooting = isKimi
+                    ? isSynthetic
+                        ? ' To fix: use streaming: "disabled" in your profile settings. Synthetic API streaming does not work reliably with tool calls.'
+                        : ' This provider may not support streaming with tool calls.'
+                    : ' Consider using streaming: "disabled" in your profile settings.';
+                logger.warn(() => `[OpenAIProvider] Empty streaming response for model '${model}' (received ${totalChunksReceived} chunks with no content).${troubleshooting}`, {
+                    model,
+                    baseURL: baseURL ?? this.getBaseURL(),
+                    isKimiModel: isKimi,
+                    isSyntheticAPI: isSynthetic,
+                    totalChunksReceived,
+                });
+            }
+            else {
+                // Log what we DID get for debugging
+                logger.debug(() => `[Streaming legacy] Stream completed with accumulated content`, {
+                    textLength: _accumulatedText.length,
+                    toolCallCount: accumulatedToolCalls.length,
+                    textBufferLength: textBuffer.length,
+                    reasoningLength: accumulatedReasoningContent.length,
+                    thinkingLength: accumulatedThinkingContent.length,
+                    totalChunksReceived,
+                });
+            }
         }
         else {
             // Handle non-streaming response
@@ -1425,22 +2049,57 @@ export class OpenAIProvider extends BaseProvider {
                 }
             }
             const blocks = [];
-            // Handle text content
-            if (choice.message?.content) {
-                blocks.push({
-                    type: 'text',
-                    text: choice.message.content,
-                });
+            // 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'}`, {
+                hasReasoningContent: 'reasoning_content' in
+                    (choice.message ?? {}),
+                messageKeys: Object.keys(choice.message ?? {}),
+            });
+            if (reasoningBlock) {
+                blocks.push(reasoningBlock);
+            }
+            // Handle text content (strip thinking / reasoning blocks) and Kimi tool sections
+            const rawMessageContent = this.coerceMessageContentToString(choice.message?.content);
+            let kimiCleanContent;
+            let kimiToolBlocks = [];
+            if (rawMessageContent) {
+                // Extract <think> tags as ThinkingBlock BEFORE stripping them
+                // Only do this if we didn't already get reasoning from reasoning_content field
+                // @plan PLAN-20251202-THINKING.P16
+                // @requirement REQ-THINK-003
+                if (!reasoningBlock) {
+                    const tagBasedThinking = this.extractThinkTagsAsBlock(rawMessageContent);
+                    if (tagBasedThinking) {
+                        blocks.push(tagBasedThinking);
+                        logger.debug(() => `[Non-streaming] Extracted thinking from <think> tags: ${tagBasedThinking.thought.length} chars`);
+                    }
+                }
+                const kimiParsed = this.extractKimiToolCallsFromText(rawMessageContent);
+                kimiCleanContent = kimiParsed.cleanedText;
+                kimiToolBlocks = kimiParsed.toolCalls;
+                // Always sanitize text content to remove <think> tags
+                // Bug fix: Previously Kimi-K2 used unsanitized kimiCleanContent,
+                // which caused <think> tags to leak into visible output
+                // @plan PLAN-20251202-THINKING.P16
+                const cleanedText = this.sanitizeProviderText(kimiCleanContent);
+                if (cleanedText) {
+                    blocks.push({
+                        type: 'text',
+                        text: cleanedText,
+                    });
+                }
             }
             // Handle tool calls
             if (choice.message?.tool_calls && choice.message.tool_calls.length > 0) {
                 // Use the same detected format from earlier for consistency
                 for (const toolCall of choice.message.tool_calls) {
                     if (toolCall.type === 'function') {
-                        // Use tool name directly without normalization for legacy compatibility
-                        const toolName = toolCall.function.name || '';
+                        // Normalize tool name (handles Kimi-K2 style prefixes)
+                        const toolName = this.normalizeToolName(toolCall.function.name || '');
+                        const sanitizedArgs = this.sanitizeToolArgumentsString(toolCall.function.arguments);
                         // Process tool parameters with double-escape handling
-                        const processedParameters = processToolParameters(toolCall.function.arguments || '', toolName);
+                        const processedParameters = processToolParameters(sanitizedArgs, toolName);
                         blocks.push({
                             type: 'tool_call',
                             id: this.normalizeToHistoryToolId(toolCall.id),
@@ -1450,42 +2109,49 @@ export class OpenAIProvider extends BaseProvider {
                     }
                 }
             }
+            // Add any tool calls parsed from Kimi inline sections
+            if (kimiToolBlocks.length > 0) {
+                blocks.push(...kimiToolBlocks);
+                this.getLogger().debug(() => `[OpenAIProvider] Non-stream legacy added Kimi tool calls from text`, { count: kimiToolBlocks.length });
+            }
             // Additionally check for <tool_call> format in text content
-            if (choice.message?.content &&
-                typeof choice.message.content === 'string') {
-                try {
-                    const parsedResult = this.textToolParser.parse(choice.message.content);
-                    if (parsedResult.toolCalls.length > 0) {
-                        // Add tool calls found in text content
-                        for (const call of parsedResult.toolCalls) {
-                            blocks.push({
-                                type: 'tool_call',
-                                id: `text_tool_${Date.now()}_${Math.random().toString(36).substring(7)}`,
-                                name: call.name,
-                                parameters: call.arguments,
-                            });
-                        }
-                        // Update the text content to remove the tool call parts
-                        if (choice.message.content !== parsedResult.cleanedContent) {
-                            // Find the text block and update it
-                            const textBlockIndex = blocks.findIndex((block) => block.type === 'text');
-                            if (textBlockIndex >= 0) {
-                                blocks[textBlockIndex].text =
-                                    parsedResult.cleanedContent;
-                            }
-                            else if (parsedResult.cleanedContent.trim()) {
-                                // Add cleaned text if it doesn't exist
-                                blocks.unshift({
-                                    type: 'text',
-                                    text: parsedResult.cleanedContent,
+            if (kimiCleanContent) {
+                const cleanedSource = this.sanitizeProviderText(kimiCleanContent);
+                if (cleanedSource) {
+                    try {
+                        const parsedResult = this.textToolParser.parse(cleanedSource);
+                        if (parsedResult.toolCalls.length > 0) {
+                            // Add tool calls found in text content
+                            for (const call of parsedResult.toolCalls) {
+                                blocks.push({
+                                    type: 'tool_call',
+                                    id: `text_tool_${Date.now()}_${Math.random().toString(36).substring(7)}`,
+                                    name: this.normalizeToolName(call.name),
+                                    parameters: call.arguments,
                                 });
                             }
+                            // Update the text content to remove the tool call parts
+                            if (choice.message.content !== parsedResult.cleanedContent) {
+                                // Find the text block and update it
+                                const textBlockIndex = blocks.findIndex((block) => block.type === 'text');
+                                if (textBlockIndex >= 0) {
+                                    blocks[textBlockIndex].text =
+                                        parsedResult.cleanedContent;
+                                }
+                                else if (parsedResult.cleanedContent.trim()) {
+                                    // Add cleaned text if it doesn't exist
+                                    blocks.unshift({
+                                        type: 'text',
+                                        text: parsedResult.cleanedContent,
+                                    });
+                                }
+                            }
                         }
                     }
-                }
-                catch (error) {
-                    const logger = this.getLogger();
-                    logger.debug(() => `TextToolCallParser failed on message content: ${error}`);
+                    catch (error) {
+                        const logger = this.getLogger();
+                        logger.debug(() => `TextToolCallParser failed on message content: ${error}`);
+                    }
                 }
             }
             // Emit the complete response as a single IContent
@@ -1613,14 +2279,8 @@ export class OpenAIProvider extends BaseProvider {
         }
         // Determine tool replay mode for model compatibility (e.g., polaris-alpha)
         const toolReplayMode = this.determineToolReplayMode(model);
-        // Convert IContent to OpenAI messages format
-        const configForMessages = options.config ?? options.runtime?.config ?? this.globalConfig;
-        const messages = this.convertToOpenAIMessages(contents, toolReplayMode, configForMessages);
-        // Log tool replay mode usage for debugging
-        if (logger.enabled && toolReplayMode !== 'native') {
-            logger.debug(() => `[OpenAIProvider] Using textual tool replay mode for model '${model}'`);
-        }
-        // Detect the tool format to use (once at the start of the method)
+        // Detect the tool format to use BEFORE building messages
+        // This is needed so that Kimi K2 tool IDs can be generated in the correct format
         const detectedFormat = this.detectToolFormat();
         // Log the detected format for debugging
         logger.debug(() => `[OpenAIProvider] Using tool format '${detectedFormat}' for model '${model}'`, {
@@ -1628,8 +2288,19 @@ export class OpenAIProvider extends BaseProvider {
             detectedFormat,
             provider: this.name,
         });
-        // Convert Gemini format tools to the detected format
-        let formattedTools = toolFormatter.convertGeminiToFormat(tools, detectedFormat);
+        // Convert IContent to OpenAI messages format
+        // Use buildMessagesWithReasoning for reasoning-aware message building
+        // Pass detectedFormat so that Kimi K2 tool IDs are generated correctly
+        const messages = toolReplayMode === 'native'
+            ? this.buildMessagesWithReasoning(contents, options, detectedFormat)
+            : this.convertToOpenAIMessages(contents, toolReplayMode, options.config ?? options.runtime?.config ?? this.globalConfig);
+        // Log tool replay mode usage for debugging
+        if (logger.enabled && toolReplayMode !== 'native') {
+            logger.debug(() => `[OpenAIProvider] Using textual tool replay mode for model '${model}'`);
+        }
+        // Convert Gemini format tools to OpenAI format using the schema converter
+        // This ensures required fields are always present in tool schemas
+        let formattedTools = convertToolsToOpenAI(tools);
         // CRITICAL FIX: Ensure we never pass an empty tools array
         // The OpenAI API errors when tools=[] but a tool call is attempted
         if (Array.isArray(formattedTools) && formattedTools.length === 0) {
@@ -1905,8 +2576,18 @@ 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
-            // Buffer text for Qwen format providers to avoid stanza formatting
-            const shouldBufferText = detectedFormat === 'qwen';
+            const isKimiModel = model.toLowerCase().includes('kimi-k2');
+            // Buffer text for Qwen format providers and Kimi-K2 to avoid stanza formatting
+            const shouldBufferText = detectedFormat === 'qwen' || isKimiModel;
+            // 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
+            let accumulatedThinkingContent = '';
+            let hasEmittedThinking = false;
+            // Accumulate reasoning_content from streaming deltas (pipeline path)
+            // Synthetic API sends reasoning token-by-token, so we accumulate to emit ONE block
+            // @plan PLAN-20251202-THINKING.P16
+            let accumulatedReasoningContent = '';
             // Track token usage from streaming chunks
             let streamingUsage = null;
             const allChunks = []; // Collect all chunks first
@@ -1918,6 +2599,11 @@ export class OpenAIProvider extends BaseProvider {
                     }
                     allChunks.push(chunk);
                 }
+                // Debug: Log how many chunks were received
+                logger.debug(() => `[Streaming pipeline] Collected ${allChunks.length} chunks from stream`, {
+                    firstChunkDelta: allChunks[0]?.choices?.[0]?.delta,
+                    lastChunkFinishReason: allChunks[allChunks.length - 1]?.choices?.[0]?.finish_reason,
+                });
                 // Now process all collected chunks
                 for (const chunk of allChunks) {
                     // Check for cancellation during chunk processing
@@ -1960,6 +2646,15 @@ export class OpenAIProvider extends BaseProvider {
                     const choice = chunk.choices?.[0];
                     if (!choice)
                         continue;
+                    // Parse reasoning_content from streaming delta (Pipeline path)
+                    // ACCUMULATE instead of yielding immediately to handle token-by-token streaming
+                    // @plan PLAN-20251202-THINKING.P16
+                    // @requirement REQ-THINK-003.1
+                    const reasoningBlock = this.parseStreamingReasoningDelta(choice.delta);
+                    if (reasoningBlock) {
+                        // Accumulate reasoning content - will emit ONE block later
+                        accumulatedReasoningContent += reasoningBlock.thought;
+                    }
                     // Check for finish_reason to detect proper stream ending
                     if (choice.finish_reason) {
                         logger.debug(() => `[Streaming] Stream finished with reason: ${choice.finish_reason}`, {
@@ -1973,23 +2668,23 @@ export class OpenAIProvider extends BaseProvider {
                         if (choice.finish_reason === 'length') {
                             logger.debug(() => `Response truncated due to length limit for model ${model}`);
                         }
-                        // Flush any buffered text when stream finishes
-                        if (textBuffer.length > 0) {
-                            yield {
-                                speaker: 'ai',
-                                blocks: [
-                                    {
-                                        type: 'text',
-                                        text: textBuffer,
-                                    },
-                                ],
-                            };
-                            textBuffer = '';
-                        }
+                        // Don't flush buffer here on finish - let the final buffer handling
+                        // after the loop process it with proper sanitization and think tag extraction
+                        // This was causing unsanitized <think> tags to leak into output (pipeline path)
+                        // @plan PLAN-20251202-THINKING.P16
                     }
                     // Handle text content - buffer for Qwen format, emit immediately for others
-                    const deltaContent = choice.delta?.content;
-                    if (deltaContent) {
+                    // Note: Synthetic API sends content that may duplicate reasoning_content.
+                    // This is the model's behavior - we don't filter it here.
+                    // @plan PLAN-20251202-THINKING.P16
+                    const rawDeltaContent = this.coerceMessageContentToString(choice.delta?.content);
+                    if (rawDeltaContent) {
+                        const deltaContent = isKimiModel
+                            ? rawDeltaContent
+                            : this.sanitizeProviderText(rawDeltaContent);
+                        if (!deltaContent) {
+                            continue;
+                        }
                         _accumulatedText += deltaContent;
                         // Debug log for providers that need buffering
                         if (shouldBufferText) {
@@ -2002,22 +2697,103 @@ 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;
                             // Emit buffered text when we have a complete sentence or paragraph
-                            // Look for natural break points
-                            if (textBuffer.includes('\n') ||
-                                textBuffer.endsWith('. ') ||
-                                textBuffer.endsWith('! ') ||
-                                textBuffer.endsWith('? ') ||
-                                textBuffer.length > 100) {
-                                yield {
-                                    speaker: 'ai',
-                                    blocks: [
-                                        {
-                                            type: 'text',
-                                            text: textBuffer,
-                                        },
-                                    ],
-                                };
+                            // Look for natural break points, avoiding flush mid Kimi section
+                            if (!hasOpenKimiSection &&
+                                (textBuffer.includes('\n') ||
+                                    textBuffer.endsWith('. ') ||
+                                    textBuffer.endsWith('! ') ||
+                                    textBuffer.endsWith('? ') ||
+                                    textBuffer.length > 100)) {
+                                const parsedToolCalls = [];
+                                let workingText = textBuffer;
+                                // Extract <think> tags and ACCUMULATE instead of emitting immediately
+                                // This handles fragmented <think>word</think> streaming from Synthetic API
+                                // @plan PLAN-20251202-THINKING.P16
+                                // @requirement REQ-THINK-003
+                                const tagBasedThinking = this.extractThinkTagsAsBlock(workingText);
+                                if (tagBasedThinking) {
+                                    // 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;
+                                    logger.debug(() => `[Streaming] Accumulated thinking: ${accumulatedThinkingContent.length} chars total`);
+                                }
+                                const kimiParsed = this.extractKimiToolCallsFromText(workingText);
+                                if (kimiParsed.toolCalls.length > 0) {
+                                    parsedToolCalls.push(...kimiParsed.toolCalls);
+                                    logger.debug(() => `[OpenAIProvider] Streaming buffer (pipeline) parsed Kimi tool calls`, {
+                                        count: kimiParsed.toolCalls.length,
+                                        bufferLength: workingText.length,
+                                        cleanedLength: kimiParsed.cleanedText.length,
+                                    });
+                                }
+                                workingText = kimiParsed.cleanedText;
+                                const parsingText = this.sanitizeProviderText(workingText);
+                                let cleanedText = parsingText;
+                                try {
+                                    const parsedResult = this.textToolParser.parse(parsingText);
+                                    if (parsedResult.toolCalls.length > 0) {
+                                        parsedToolCalls.push(...parsedResult.toolCalls.map((call) => ({
+                                            type: 'tool_call',
+                                            id: `text_tool_${Date.now()}_${Math.random()
+                                                .toString(36)
+                                                .substring(7)}`,
+                                            name: this.normalizeToolName(call.name),
+                                            parameters: call.arguments,
+                                        })));
+                                        cleanedText = parsedResult.cleanedContent;
+                                    }
+                                }
+                                catch (error) {
+                                    const logger = this.getLogger();
+                                    logger.debug(() => `TextToolCallParser failed on buffered text: ${error}`);
+                                }
+                                // Emit accumulated thinking BEFORE tool calls or text content
+                                // This ensures thinking appears first in the response
+                                // @plan PLAN-20251202-THINKING.P16
+                                if (!hasEmittedThinking &&
+                                    accumulatedThinkingContent.length > 0 &&
+                                    (parsedToolCalls.length > 0 || cleanedText.trim().length > 0)) {
+                                    yield {
+                                        speaker: 'ai',
+                                        blocks: [
+                                            {
+                                                type: 'thinking',
+                                                thought: accumulatedThinkingContent,
+                                                sourceField: 'think_tags',
+                                                isHidden: false,
+                                            },
+                                        ],
+                                    };
+                                    hasEmittedThinking = true;
+                                    logger.debug(() => `[Streaming pipeline] Emitted accumulated thinking: ${accumulatedThinkingContent.length} chars`);
+                                }
+                                if (parsedToolCalls.length > 0) {
+                                    yield {
+                                        speaker: 'ai',
+                                        blocks: parsedToolCalls,
+                                    };
+                                }
+                                // 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) {
+                                    yield {
+                                        speaker: 'ai',
+                                        blocks: [
+                                            {
+                                                type: 'text',
+                                                text: cleanedText,
+                                            },
+                                        ],
+                                    };
+                                }
                                 textBuffer = '';
                             }
                         }
@@ -2097,19 +2873,45 @@ export class OpenAIProvider extends BaseProvider {
             }
             // Check buffered text for <tool_call> format before flushing as plain text
             if (textBuffer.length > 0) {
-                // Try to parse <tool_call> format from buffered text
-                let parsedToolCalls = [];
-                let cleanedText = textBuffer;
+                const parsedToolCalls = [];
+                let workingText = textBuffer;
+                // Note: Synthetic API sends reasoning via both reasoning_content AND content fields.
+                // This is the model's behavior - we don't strip it since the model is the source.
+                // The user can configure reasoning display settings if they don't want duplicates.
+                // @plan PLAN-20251202-THINKING.P16
+                // Extract any remaining <think> tags from final buffer
+                // @plan PLAN-20251202-THINKING.P16
+                const tagBasedThinking = this.extractThinkTagsAsBlock(workingText);
+                if (tagBasedThinking) {
+                    // Use newline to preserve formatting between chunks (not space)
+                    if (accumulatedThinkingContent.length > 0) {
+                        accumulatedThinkingContent += '\n';
+                    }
+                    accumulatedThinkingContent += tagBasedThinking.thought;
+                }
+                const kimiParsed = this.extractKimiToolCallsFromText(workingText);
+                if (kimiParsed.toolCalls.length > 0) {
+                    parsedToolCalls.push(...kimiParsed.toolCalls);
+                    this.getLogger().debug(() => `[OpenAIProvider] Final buffer flush (pipeline) parsed Kimi tool calls`, {
+                        count: kimiParsed.toolCalls.length,
+                        bufferLength: workingText.length,
+                        cleanedLength: kimiParsed.cleanedText.length,
+                    });
+                }
+                workingText = kimiParsed.cleanedText;
+                const parsingText = this.sanitizeProviderText(workingText);
+                let cleanedText = parsingText;
                 try {
-                    const parsedResult = this.textToolParser.parse(textBuffer);
+                    const parsedResult = this.textToolParser.parse(parsingText);
                     if (parsedResult.toolCalls.length > 0) {
-                        // Convert parsed tool calls to ToolCallBlock format
-                        parsedToolCalls = parsedResult.toolCalls.map((call) => ({
+                        parsedToolCalls.push(...parsedResult.toolCalls.map((call) => ({
                             type: 'tool_call',
-                            id: `text_tool_${Date.now()}_${Math.random().toString(36).substring(7)}`,
-                            name: call.name,
+                            id: `text_tool_${Date.now()}_${Math.random()
+                                .toString(36)
+                                .substring(7)}`,
+                            name: this.normalizeToolName(call.name),
                             parameters: call.arguments,
-                        }));
+                        })));
                         cleanedText = parsedResult.cleanedContent;
                     }
                 }
@@ -2117,14 +2919,33 @@ export class OpenAIProvider extends BaseProvider {
                     const logger = this.getLogger();
                     logger.debug(() => `TextToolCallParser failed on buffered text: ${error}`);
                 }
-                // Emit tool calls from text parsing first
+                // Emit accumulated thinking BEFORE tool calls or text content
+                // @plan PLAN-20251202-THINKING.P16
+                if (!hasEmittedThinking &&
+                    accumulatedThinkingContent.length > 0 &&
+                    (parsedToolCalls.length > 0 || cleanedText.trim().length > 0)) {
+                    yield {
+                        speaker: 'ai',
+                        blocks: [
+                            {
+                                type: 'thinking',
+                                thought: accumulatedThinkingContent,
+                                sourceField: 'think_tags',
+                                isHidden: false,
+                            },
+                        ],
+                    };
+                    hasEmittedThinking = true;
+                }
                 if (parsedToolCalls.length > 0) {
                     yield {
                         speaker: 'ai',
                         blocks: parsedToolCalls,
                     };
                 }
-                // Then emit any remaining cleaned text
+                // 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) {
                     yield {
                         speaker: 'ai',
@@ -2138,6 +2959,39 @@ export class OpenAIProvider extends BaseProvider {
                 }
                 textBuffer = '';
             }
+            // Emit any remaining accumulated thinking that wasn't emitted yet
+            // (e.g., if entire response was just thinking with no content)
+            // @plan PLAN-20251202-THINKING.P16
+            if (!hasEmittedThinking && accumulatedThinkingContent.length > 0) {
+                yield {
+                    speaker: 'ai',
+                    blocks: [
+                        {
+                            type: 'thinking',
+                            thought: accumulatedThinkingContent,
+                            sourceField: 'think_tags',
+                            isHidden: false,
+                        },
+                    ],
+                };
+                hasEmittedThinking = true;
+            }
+            // Emit accumulated reasoning_content as ONE ThinkingBlock (pipeline path)
+            // This consolidates token-by-token reasoning from Synthetic API into a single block
+            // @plan PLAN-20251202-THINKING.P16
+            if (accumulatedReasoningContent.length > 0) {
+                yield {
+                    speaker: 'ai',
+                    blocks: [
+                        {
+                            type: 'thinking',
+                            thought: accumulatedReasoningContent,
+                            sourceField: 'reasoning_content',
+                            isHidden: false,
+                        },
+                    ],
+                };
+            }
             // Process and emit tool calls using the pipeline
             const pipelineResult = await this.toolCallPipeline.process(abortSignal);
             if (pipelineResult.normalized.length > 0 ||
@@ -2145,8 +2999,9 @@ export class OpenAIProvider extends BaseProvider {
                 const blocks = [];
                 // Process successful tool calls
                 for (const normalizedCall of pipelineResult.normalized) {
+                    const sanitizedArgs = this.sanitizeToolArgumentsString(normalizedCall.originalArgs ?? normalizedCall.args);
                     // Process tool parameters with double-escape handling
-                    const processedParameters = processToolParameters(normalizedCall.originalArgs || JSON.stringify(normalizedCall.args), normalizedCall.name);
+                    const processedParameters = processToolParameters(sanitizedArgs, normalizedCall.name);
                     blocks.push({
                         type: 'tool_call',
                         id: this.normalizeToHistoryToolId(`call_${normalizedCall.index}`),
@@ -2195,6 +3050,41 @@ export class OpenAIProvider extends BaseProvider {
                     },
                 };
             }
+            // Detect and warn about empty streaming responses (common with Kimi K2 after tool calls)
+            // Only warn if we truly got nothing - not even reasoning content
+            const pipelineStats = this.toolCallPipeline.getStats();
+            if (_accumulatedText.length === 0 &&
+                pipelineStats.collector.totalCalls === 0 &&
+                textBuffer.length === 0 &&
+                accumulatedReasoningContent.length === 0 &&
+                accumulatedThinkingContent.length === 0) {
+                // Provide actionable guidance for users
+                const isKimi = model.toLowerCase().includes('kimi');
+                const isSynthetic = (baseURL ?? this.getBaseURL())?.includes('synthetic') ?? false;
+                const troubleshooting = isKimi
+                    ? isSynthetic
+                        ? ' To fix: use streaming: "disabled" in your profile settings. Synthetic API streaming does not work reliably with tool calls.'
+                        : ' This provider may not support streaming with tool calls.'
+                    : ' Consider using streaming: "disabled" in your profile settings.';
+                logger.warn(() => `[OpenAIProvider] Empty streaming response for model '${model}' (received ${allChunks.length} chunks with no content).${troubleshooting}`, {
+                    model,
+                    baseURL: baseURL ?? this.getBaseURL(),
+                    isKimiModel: isKimi,
+                    isSyntheticAPI: isSynthetic,
+                    totalChunksReceived: allChunks.length,
+                });
+            }
+            else {
+                // Log what we DID get for debugging
+                logger.debug(() => `[Streaming pipeline] Stream completed with accumulated content`, {
+                    textLength: _accumulatedText.length,
+                    toolCallCount: pipelineStats.collector.totalCalls,
+                    textBufferLength: textBuffer.length,
+                    reasoningLength: accumulatedReasoningContent.length,
+                    thinkingLength: accumulatedThinkingContent.length,
+                    totalChunksReceived: allChunks.length,
+                });
+            }
         }
         else {
             // Handle non-streaming response
@@ -2220,12 +3110,22 @@ export class OpenAIProvider extends BaseProvider {
                 }
             }
             const blocks = [];
-            // Handle text content
-            if (choice.message?.content) {
-                blocks.push({
-                    type: 'text',
-                    text: choice.message.content,
-                });
+            // Handle text content (strip thinking / reasoning blocks) and Kimi tool sections
+            const pipelineRawMessageContent = this.coerceMessageContentToString(choice.message?.content);
+            let pipelineKimiCleanContent;
+            let pipelineKimiToolBlocks = [];
+            if (pipelineRawMessageContent) {
+                const kimiParsed = this.extractKimiToolCallsFromText(pipelineRawMessageContent);
+                pipelineKimiCleanContent = kimiParsed.cleanedText;
+                pipelineKimiToolBlocks = kimiParsed.toolCalls;
+                // Always use sanitized text - even Kimi-K2 should have consistent tag stripping
+                const cleanedText = this.sanitizeProviderText(pipelineKimiCleanContent);
+                if (cleanedText) {
+                    blocks.push({
+                        type: 'text',
+                        text: cleanedText,
+                    });
+                }
             }
             // Handle tool calls
             if (choice.message?.tool_calls && choice.message.tool_calls.length > 0) {
@@ -2234,8 +3134,9 @@ export class OpenAIProvider extends BaseProvider {
                     if (toolCall.type === 'function') {
                         // Normalize tool name for consistency with streaming path
                         const normalizedName = this.toolCallPipeline.normalizeToolName(toolCall.function.name, toolCall.function.arguments);
+                        const sanitizedArgs = this.sanitizeToolArgumentsString(toolCall.function.arguments);
                         // Process tool parameters with double-escape handling
-                        const processedParameters = processToolParameters(toolCall.function.arguments || '', normalizedName);
+                        const processedParameters = processToolParameters(sanitizedArgs, normalizedName);
                         blocks.push({
                             type: 'tool_call',
                             id: this.normalizeToHistoryToolId(toolCall.id),
@@ -2245,42 +3146,48 @@ export class OpenAIProvider extends BaseProvider {
                     }
                 }
             }
+            if (pipelineKimiToolBlocks.length > 0) {
+                blocks.push(...pipelineKimiToolBlocks);
+                this.getLogger().debug(() => `[OpenAIProvider] Non-stream pipeline added Kimi tool calls from text`, { count: pipelineKimiToolBlocks.length });
+            }
             // Additionally check for <tool_call> format in text content
-            if (choice.message?.content &&
-                typeof choice.message.content === 'string') {
-                try {
-                    const parsedResult = this.textToolParser.parse(choice.message.content);
-                    if (parsedResult.toolCalls.length > 0) {
-                        // Add tool calls found in text content
-                        for (const call of parsedResult.toolCalls) {
-                            blocks.push({
-                                type: 'tool_call',
-                                id: `text_tool_${Date.now()}_${Math.random().toString(36).substring(7)}`,
-                                name: call.name,
-                                parameters: call.arguments,
-                            });
-                        }
-                        // Update the text content to remove the tool call parts
-                        if (choice.message.content !== parsedResult.cleanedContent) {
-                            // Find the text block and update it
-                            const textBlockIndex = blocks.findIndex((block) => block.type === 'text');
-                            if (textBlockIndex >= 0) {
-                                blocks[textBlockIndex].text =
-                                    parsedResult.cleanedContent;
-                            }
-                            else if (parsedResult.cleanedContent.trim()) {
-                                // Add cleaned text if it doesn't exist
-                                blocks.unshift({
-                                    type: 'text',
-                                    text: parsedResult.cleanedContent,
+            if (pipelineKimiCleanContent) {
+                const cleanedSource = this.sanitizeProviderText(pipelineKimiCleanContent);
+                if (cleanedSource) {
+                    try {
+                        const parsedResult = this.textToolParser.parse(cleanedSource);
+                        if (parsedResult.toolCalls.length > 0) {
+                            // Add tool calls found in text content
+                            for (const call of parsedResult.toolCalls) {
+                                blocks.push({
+                                    type: 'tool_call',
+                                    id: `text_tool_${Date.now()}_${Math.random().toString(36).substring(7)}`,
+                                    name: this.normalizeToolName(call.name),
+                                    parameters: call.arguments,
                                 });
                             }
+                            // Update the text content to remove the tool call parts
+                            if (choice.message.content !== parsedResult.cleanedContent) {
+                                // Find the text block and update it
+                                const textBlockIndex = blocks.findIndex((block) => block.type === 'text');
+                                if (textBlockIndex >= 0) {
+                                    blocks[textBlockIndex].text =
+                                        parsedResult.cleanedContent;
+                                }
+                                else if (parsedResult.cleanedContent.trim()) {
+                                    // Add cleaned text if it doesn't exist
+                                    blocks.unshift({
+                                        type: 'text',
+                                        text: parsedResult.cleanedContent,
+                                    });
+                                }
+                            }
                         }
                     }
-                }
-                catch (error) {
-                    const logger = this.getLogger();
-                    logger.debug(() => `TextToolCallParser failed on message content: ${error}`);
+                    catch (error) {
+                        const logger = this.getLogger();
+                        logger.debug(() => `TextToolCallParser failed on message content: ${error}`);
+                    }
                 }
             }
             // Emit the complete response as a single IContent
@@ -2338,19 +3245,25 @@ export class OpenAIProvider extends BaseProvider {
     }
     /**
      * Detects the tool call format based on the model being used
-     * @returns The detected tool format ('openai' or 'qwen')
+     * @returns The detected tool format ('openai', 'qwen', or 'kimi')
      */
     detectToolFormat() {
         // Auto-detect based on model name if set to 'auto' or not set
-        const modelName = (this.getModel() || this.getDefaultModel()).toLowerCase();
+        const modelName = this.getModel() || this.getDefaultModel();
         const logger = new DebugLogger('llxprt:provider:openai');
+        // Check for Kimi K2 models (requires special ID format: functions.{name}:{index})
+        if (isKimiModel(modelName)) {
+            logger.debug(() => `Auto-detected 'kimi' format for K2 model: ${modelName}`);
+            return 'kimi';
+        }
+        const lowerModelName = modelName.toLowerCase();
         // Check for GLM-4 models (glm-4, glm-4.5, glm-4.6, glm-4-5, etc.)
-        if (modelName.includes('glm-4')) {
+        if (lowerModelName.includes('glm-4')) {
             logger.debug(() => `Auto-detected 'qwen' format for GLM-4.x model: ${modelName}`);
             return 'qwen';
         }
         // Check for qwen models
-        if (modelName.includes('qwen')) {
+        if (lowerModelName.includes('qwen')) {
             logger.debug(() => `Auto-detected 'qwen' format for Qwen model: ${modelName}`);
             return 'qwen';
         }
@@ -2421,5 +3334,61 @@ export class OpenAIProvider extends BaseProvider {
         }
         return shouldRetry;
     }
+    /**
+     * 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
+     */
+    parseStreamingReasoningDelta(delta) {
+        if (!delta) {
+            return null;
+        }
+        // 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,
+        };
+    }
+    /**
+     * 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
+     */
+    parseNonStreamingReasoning(message) {
+        if (!message) {
+            return null;
+        }
+        // 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;
+        }
+        // Handle empty string or whitespace-only
+        if (reasoningContent.trim().length === 0) {
+            return null;
+        }
+        return {
+            type: 'thinking',
+            thought: reasoningContent,
+            sourceField: 'reasoning_content',
+            isHidden: false,
+        };
+    }
 }
 //# sourceMappingURL=OpenAIProvider.js.map