@appkit/llamacpp-cli 2.1.0 → 2.2.0

package/dist/lib/router-server.js

@@ -42,6 +42,216 @@ const fs = __importStar(require("fs/promises"));
 const path = __importStar(require("path"));
 const file_utils_1 = require("../utils/file-utils");
 const router_logger_1 = require("./router-logger");
+/**
+ * Parse Qwen3-Coder XML tool calls from text content.
+ * Handles: <tool_call><function=NAME\n<parameter=P>V</parameter></function></tool_call>
+ * Returns extracted tool calls and cleaned text (XML removed).
+ */
+function parseXmlToolCalls(text) {
+    const toolCalls = [];
+    const skippedNames = [];
+    const toolCallRegex = /<tool_call>([\s\S]*?)<\/tool_call>/g;
+    let match;
+    while ((match = toolCallRegex.exec(text)) !== null) {
+        const inner = match[1];
+        const funcMatch = /<function=(\w+)/.exec(inner);
+        if (!funcMatch)
+            continue;
+        const name = funcMatch[1];
+        const input = {};
+        const paramRegex = /<parameter=(\w+)>([\s\S]*?)<\/parameter>/g;
+        let paramMatch;
+        while ((paramMatch = paramRegex.exec(inner)) !== null) {
+            input[paramMatch[1]] = paramMatch[2].trim();
+        }
+        // Skip malformed tool calls with no parameters (model generation failure)
+        if (Object.keys(input).length === 0) {
+            skippedNames.push(name);
+            continue;
+        }
+        toolCalls.push({ name, input });
+    }
+    const cleanText = text.replace(/<tool_call>[\s\S]*?<\/tool_call>/g, '').trim();
+    return { toolCalls, cleanText, skippedNames };
+}
+function generateToolUseId() {
+    return 'toolu_' + Array.from({ length: 16 }, () => Math.floor(Math.random() * 16).toString(16)).join('');
+}
+/**
+ * Count how many consecutive recent user messages contained only error tool_results.
+ * Used to detect infinite error-feedback loops: if >= 2, stop sending error feedback and strip.
+ */
+function countConsecutiveErrorCycles(requestBody) {
+    try {
+        const body = JSON.parse(requestBody);
+        const messages = body.messages ?? [];
+        let count = 0;
+        for (let i = messages.length - 1; i >= 0; i--) {
+            const msg = messages[i];
+            if (msg.role === 'assistant')
+                continue; // skip assistant turns
+            if (msg.role !== 'user')
+                break;
+            const content = Array.isArray(msg.content) ? msg.content : [];
+            const toolResults = content.filter((c) => c.type === 'tool_result');
+            if (toolResults.length === 0)
+                break; // non-tool user message, stop
+            if (toolResults.every((c) => c.is_error)) {
+                count++;
+            }
+            else {
+                break; // mixed or all-success results, stop counting
+            }
+        }
+        return count;
+    }
+    catch {
+        return 0;
+    }
+}
+function emitSseEvent(res, eventType, data) {
+    res.write(`event: ${eventType}\ndata: ${JSON.stringify(data)}\n\n`);
+}
+/**
+ * Emit a fully reconstructed SSE stream from parsed block state.
+ * Used when the original stream needs modification (XML tool calls or thinking-only).
+ */
+function emitReconstructedSseStream(res, messageStartData, blocks, stopReason, outputTokens) {
+    if (messageStartData) {
+        emitSseEvent(res, 'message_start', messageStartData);
+    }
+    let idx = 0;
+    for (const block of blocks) {
+        if (block.type === 'text') {
+            if (!block.content)
+                continue;
+            emitSseEvent(res, 'content_block_start', { type: 'content_block_start', index: idx, content_block: { type: 'text', text: '' } });
+            emitSseEvent(res, 'content_block_delta', { type: 'content_block_delta', index: idx, delta: { type: 'text_delta', text: block.content } });
+            emitSseEvent(res, 'content_block_stop', { type: 'content_block_stop', index: idx });
+        }
+        else if (block.type === 'thinking') {
+            if (!block.content)
+                continue;
+            emitSseEvent(res, 'content_block_start', { type: 'content_block_start', index: idx, content_block: { type: 'thinking', thinking: '' } });
+            emitSseEvent(res, 'content_block_delta', { type: 'content_block_delta', index: idx, delta: { type: 'thinking_delta', thinking: block.content } });
+            emitSseEvent(res, 'content_block_stop', { type: 'content_block_stop', index: idx });
+        }
+        else if (block.type === 'tool_use') {
+            emitSseEvent(res, 'content_block_start', { type: 'content_block_start', index: idx, content_block: { type: 'tool_use', id: block.id, name: block.name, input: {} } });
+            emitSseEvent(res, 'content_block_delta', { type: 'content_block_delta', index: idx, delta: { type: 'input_json_delta', partial_json: JSON.stringify(block.input ?? {}) } });
+            emitSseEvent(res, 'content_block_stop', { type: 'content_block_stop', index: idx });
+        }
+        idx++;
+    }
+    emitSseEvent(res, 'message_delta', { type: 'message_delta', delta: { stop_reason: stopReason, stop_sequence: null }, usage: { output_tokens: outputTokens } });
+    emitSseEvent(res, 'message_stop', { type: 'message_stop' });
+}
+/**
+ * Make a backend HTTP request, buffer the full SSE stream, and return parsed state.
+ * Used for the initial request AND for retries when the model generates malformed output.
+ */
+function bufferSseRequest(options, requestBody) {
+    return new Promise((resolve, reject) => {
+        const req = http.request(options, (res) => {
+            let sseBuffer = '';
+            const rawEvents = [];
+            let messageStartData = null;
+            let messageDeltaData = null;
+            const parsedBlocks = {};
+            let outputTokens = 0;
+            res.on('data', (chunk) => {
+                sseBuffer += chunk.toString();
+                const parts = sseBuffer.split('\n\n');
+                sseBuffer = parts.pop() ?? '';
+                for (const part of parts) {
+                    if (!part.trim())
+                        continue;
+                    rawEvents.push(part);
+                    let dataStr = '';
+                    for (const line of part.split('\n')) {
+                        if (line.startsWith('data: '))
+                            dataStr = line.slice(6);
+                    }
+                    try {
+                        const data = JSON.parse(dataStr);
+                        if (data.type === 'message_start')
+                            messageStartData = data;
+                        else if (data.type === 'content_block_start') {
+                            const idx = data.index ?? 0;
+                            parsedBlocks[idx] = { type: data.content_block?.type ?? 'text', content: '', name: data.content_block?.name, id: data.content_block?.id };
+                        }
+                        else if (data.type === 'content_block_delta') {
+                            const block = parsedBlocks[data.index];
+                            if (block) {
+                                if (data.delta?.type === 'text_delta')
+                                    block.content += data.delta.text ?? '';
+                                else if (data.delta?.type === 'thinking_delta')
+                                    block.content += data.delta.thinking ?? '';
+                            }
+                        }
+                        else if (data.type === 'message_delta') {
+                            messageDeltaData = data;
+                            outputTokens = data.usage?.output_tokens ?? 0;
+                        }
+                    }
+                    catch { /* non-JSON SSE (ping etc.) */ }
+                }
+            });
+            res.on('end', () => {
+                if (sseBuffer.trim())
+                    rawEvents.push(sseBuffer);
+                resolve({
+                    rawEvents,
+                    messageStartData,
+                    stopReason: messageDeltaData?.delta?.stop_reason ?? 'end_turn',
+                    blocks: Object.values(parsedBlocks),
+                    outputTokens,
+                });
+            });
+            res.on('error', reject);
+        });
+        req.on('error', reject);
+        req.write(requestBody);
+        req.end();
+    });
+}
+/**
+ * Apply Qwen3 model quirk fixes to a buffered SSE result.
+ * Returns the action to take (what to emit) without actually emitting anything.
+ */
+function classifyBufferedResult(result) {
+    const { blocks, stopReason } = result;
+    const textBlocks = blocks.filter(b => b.type === 'text');
+    const thinkingBlocks = blocks.filter(b => b.type === 'thinking');
+    const allText = textBlocks.map(b => b.content).join('');
+    const { toolCalls, cleanText, skippedNames } = parseXmlToolCalls(allText);
+    if (toolCalls.length > 0) {
+        const newBlocks = [
+            ...blocks.filter(b => b.type !== 'text'),
+            ...(cleanText ? [{ type: 'text', content: cleanText }] : []),
+            ...toolCalls.map(tc => ({ type: 'tool_use', content: '', id: generateToolUseId(), name: tc.name, input: tc.input })),
+        ];
+        return { action: 'fix1', newBlocks, stopReason: 'tool_use' };
+    }
+    if (allText && !cleanText) {
+        return { action: 'fix3', skippedNames };
+    }
+    // fix4: text + malformed tool calls — emit text and empty tool_use blocks for error feedback
+    if (cleanText && skippedNames.length > 0) {
+        const newBlocks = [
+            ...blocks.filter(b => b.type !== 'text'),
+            { type: 'text', content: cleanText },
+            ...skippedNames.map(name => ({ type: 'tool_use', content: '', id: generateToolUseId(), name, input: {} })),
+        ];
+        return { action: 'fix4', newBlocks, stopReason: 'tool_use', skippedNames };
+    }
+    if (stopReason === 'end_turn' && thinkingBlocks.length > 0 && !textBlocks.some(b => b.content)) {
+        const thinkingText = thinkingBlocks.map(b => b.content).join('\n');
+        const newBlocks = [...thinkingBlocks, { type: 'text', content: thinkingText }];
+        return { action: 'fix2', newBlocks, stopReason: 'end_turn' };
+    }
+    return { action: 'raw' };
+}
 /**
  * Router HTTP server - proxies requests to backend llama.cpp servers
  */
@@ -110,6 +320,9 @@ class RouterServer {
             else if (url === '/health' && method === 'GET') {
                 await this.handleHealth(req, res);
             }
+            else if (url.startsWith('/props') && method === 'GET') {
+                await this.handleProps(req, res, url);
+            }
             else if (url === '/v1/models' && method === 'GET') {
                 await this.handleModels(req, res);
             }
@@ -149,6 +362,51 @@ class RouterServer {
             timestamp: new Date().toISOString(),
         }));
     }
+    /**
+     * Proxy llama.cpp's /props to a backend server. Pass ?model=<name> to
+     * select which backend; otherwise picks the first running server.
+     * Used by clients (e.g. lcode) to discover the loaded n_ctx.
+     */
+    async handleProps(req, res, url) {
+        const query = new url_1.URL(url, 'http://localhost').searchParams;
+        const requestedModel = query.get('model');
+        const servers = await this.getAllServers();
+        const running = servers.filter((s) => s.status === 'running');
+        const target = requestedModel
+            ? await this.findServerForModel(requestedModel)
+            : running[0] ?? null;
+        if (!target || target.status !== 'running') {
+            this.sendError(res, 404, 'Not Found', requestedModel
+                ? `No running server for model: ${requestedModel}`
+                : 'No running servers');
+            return;
+        }
+        const host = target.host === '0.0.0.0' ? '127.0.0.1' : target.host;
+        const backendReq = http.request({
+            hostname: host,
+            port: target.port,
+            path: '/props',
+            method: 'GET',
+            timeout: this.config.requestTimeout,
+        }, (backendRes) => {
+            res.writeHead(backendRes.statusCode || 200, {
+                'Content-Type': backendRes.headers['content-type'] ?? 'application/json',
+            });
+            backendRes.pipe(res);
+        });
+        backendReq.on('error', (err) => {
+            if (!res.headersSent) {
+                this.sendError(res, 502, 'Bad Gateway', `Backend /props failed: ${err.message}`);
+            }
+        });
+        backendReq.on('timeout', () => {
+            backendReq.destroy();
+            if (!res.headersSent) {
+                this.sendError(res, 504, 'Gateway Timeout', 'Backend /props did not respond in time');
+            }
+        });
+        backendReq.end();
+    }
     /**
      * List models endpoint - aggregate from all running servers
      */
@@ -278,6 +536,20 @@ class RouterServer {
                 await this.logRequest(modelName, '/v1/messages', statusCode, timer.elapsed(), errorMsg, undefined, promptPreview);
                 return;
             }
+            // Inject tool call guidance when tools are present (Qwen3-Coder workaround:
+            // the model sometimes generates tool calls with no parameters when context is long)
+            if (anthropicRequest.tools && anthropicRequest.tools.length > 0) {
+                const guidance = 'When using tools, always include ALL required parameters with their complete values. Never omit parameters from tool calls.';
+                if (typeof anthropicRequest.system === 'string' && anthropicRequest.system) {
+                    anthropicRequest.system = guidance + '\n\n' + anthropicRequest.system;
+                }
+                else if (Array.isArray(anthropicRequest.system)) {
+                    anthropicRequest.system = [{ type: 'text', text: guidance }, ...anthropicRequest.system];
+                }
+                else {
+                    anthropicRequest.system = guidance;
+                }
+            }
             // Find server for model
             const server = await this.findServerForModel(modelName);
             if (!server) {
@@ -335,30 +607,262 @@ class RouterServer {
             const backendReq = http.request(options, (backendRes) => {
                 // Handle streaming vs non-streaming
                 if (anthropicRequest.stream) {
-                    // For streaming, set SSE headers and pipe response
                     res.writeHead(backendRes.statusCode || 200, {
                         'Content-Type': 'text/event-stream',
                         'Cache-Control': 'no-cache',
                         'Connection': 'keep-alive',
                     });
-                    // Pipe response directly (llama.cpp sends correct Anthropic SSE format)
-                    backendRes.pipe(res);
+                    // Buffer the full SSE stream so we can detect and fix Qwen3 model quirks before
+                    // forwarding to the client. Headers are sent above but NO events are emitted until
+                    // we've finished processing (enabling transparent retry for Fix 3).
+                    let sseBuffer = '';
+                    const rawEvents = [];
+                    let messageStartData = null;
+                    let messageDeltaData = null;
+                    const parsedBlocks = {};
+                    let outputTokens = 0;
+                    backendRes.on('data', (chunk) => {
+                        sseBuffer += chunk.toString();
+                        const parts = sseBuffer.split('\n\n');
+                        sseBuffer = parts.pop() ?? '';
+                        for (const part of parts) {
+                            if (!part.trim())
+                                continue;
+                            rawEvents.push(part);
+                            let dataStr = '';
+                            for (const line of part.split('\n')) {
+                                if (line.startsWith('data: '))
+                                    dataStr = line.slice(6);
+                            }
+                            try {
+                                const data = JSON.parse(dataStr);
+                                if (data.type === 'message_start') {
+                                    messageStartData = data;
+                                }
+                                else if (data.type === 'content_block_start') {
+                                    const idx = data.index ?? 0;
+                                    parsedBlocks[idx] = {
+                                        type: data.content_block?.type ?? 'text',
+                                        content: '',
+                                        name: data.content_block?.name,
+                                        id: data.content_block?.id,
+                                    };
+                                }
+                                else if (data.type === 'content_block_delta') {
+                                    const block = parsedBlocks[data.index];
+                                    if (block) {
+                                        if (data.delta?.type === 'text_delta')
+                                            block.content += data.delta.text ?? '';
+                                        else if (data.delta?.type === 'thinking_delta')
+                                            block.content += data.delta.thinking ?? '';
+                                    }
+                                }
+                                else if (data.type === 'message_delta') {
+                                    messageDeltaData = data;
+                                    outputTokens = data.usage?.output_tokens ?? 0;
+                                }
+                            }
+                            catch {
+                                // Non-JSON SSE data (e.g. ping) — still buffered in rawEvents
+                            }
+                        }
+                    });
                     backendRes.on('end', async () => {
+                        if (sseBuffer.trim())
+                            rawEvents.push(sseBuffer);
+                        const firstResult = {
+                            rawEvents,
+                            messageStartData,
+                            stopReason: messageDeltaData?.delta?.stop_reason ?? 'end_turn',
+                            blocks: Object.values(parsedBlocks),
+                            outputTokens,
+                        };
+                        let classified = classifyBufferedResult(firstResult);
+                        let finalResult = firstResult;
+                        if (classified.action === 'fix3') {
+                            const skipped = classified.skippedNames ?? [];
+                            // Only retry for single empty call glitches (random sampling failure).
+                            // If 2+ empty calls were generated the model is in a stuck pattern — retry
+                            // would just double the wait time with the same degenerate result.
+                            if (skipped.length === 1) {
+                                console.error(`[Router] Retrying single malformed XML call (attempted: ${skipped.join(', ')})`);
+                                try {
+                                    const retryResult = await bufferSseRequest(options, requestBody);
+                                    const retryClassified = classifyBufferedResult(retryResult);
+                                    if (retryClassified.action !== 'fix3') {
+                                        classified = retryClassified;
+                                        finalResult = retryResult;
+                                        console.error(`[Router] Retry succeeded (action: ${retryClassified.action})`);
+                                    }
+                                    else {
+                                        console.error(`[Router] Retry also malformed, giving up`);
+                                    }
+                                }
+                                catch (err) {
+                                    console.error('[Router] Retry request failed:', err);
+                                }
+                            }
+                            else {
+                                console.error(`[Router] Skipping retry — model stuck generating ${skipped.length} malformed calls (${skipped.join(', ')})`);
+                            }
+                        }
+                        if (classified.action === 'fix1') {
+                            console.error(`[Router] Converting ${classified.newBlocks.filter(b => b.type === 'tool_use').length} XML tool call(s) to tool_use blocks`);
+                            emitReconstructedSseStream(res, finalResult.messageStartData, classified.newBlocks, classified.stopReason, finalResult.outputTokens);
+                        }
+                        else if (classified.action === 'fix2') {
+                            console.error('[Router] Injecting fallback text block (thinking-only response detected)');
+                            emitReconstructedSseStream(res, finalResult.messageStartData, classified.newBlocks, classified.stopReason, finalResult.outputTokens);
+                        }
+                        else if (classified.action === 'fix3') {
+                            const skipped = classified.skippedNames ?? [];
+                            const errorCycles = countConsecutiveErrorCycles(requestBody);
+                            if (errorCycles >= 2) {
+                                // Already tried error feedback twice — model is stuck, strip to avoid infinite loop
+                                console.error(`[Router] Stripping fix3 after ${errorCycles} error cycles (${skipped.join(', ')})`);
+                                const newBlocks = finalResult.blocks.filter(b => b.type === 'thinking');
+                                emitReconstructedSseStream(res, finalResult.messageStartData, newBlocks, finalResult.stopReason, finalResult.outputTokens);
+                            }
+                            else {
+                                // Send empty tool_use blocks so Claude Code returns parameter errors for model self-correction
+                                console.error(`[Router] Forwarding ${skipped.length} empty tool_use block(s) for error feedback [cycle ${errorCycles + 1}] (${skipped.join(', ')})`);
+                                const emptyToolBlocks = skipped.map(name => ({
+                                    type: 'tool_use',
+                                    content: '',
+                                    name,
+                                    id: generateToolUseId(),
+                                    input: {},
+                                }));
+                                const newBlocks = [
+                                    ...finalResult.blocks.filter(b => b.type === 'thinking'),
+                                    ...emptyToolBlocks,
+                                ];
+                                emitReconstructedSseStream(res, finalResult.messageStartData, newBlocks, 'tool_use', finalResult.outputTokens);
+                            }
+                        }
+                        else if (classified.action === 'fix4') {
+                            // Text + malformed tool calls
+                            const skipped = classified.skippedNames ?? [];
+                            const errorCycles = countConsecutiveErrorCycles(requestBody);
+                            if (errorCycles >= 2) {
+                                // Already tried error feedback twice — strip malformed calls, return just the text
+                                console.error(`[Router] Stripping fix4 malformed call(s) after ${errorCycles} error cycles (${skipped.join(', ')})`);
+                                const textOnlyBlocks = classified.newBlocks.filter(b => b.type !== 'tool_use');
+                                emitReconstructedSseStream(res, finalResult.messageStartData, textOnlyBlocks, 'end_turn', finalResult.outputTokens);
+                            }
+                            else {
+                                console.error(`[Router] Text + ${skipped.length} malformed tool call(s), forwarding empty tool_use for error feedback [cycle ${errorCycles + 1}] (${skipped.join(', ')})`);
+                                emitReconstructedSseStream(res, finalResult.messageStartData, classified.newBlocks, 'tool_use', finalResult.outputTokens);
+                            }
+                        }
+                        else {
+                            // Raw passthrough
+                            for (const event of finalResult.rawEvents) {
+                                res.write(event + '\n\n');
+                            }
+                        }
+                        res.end();
                         await this.logRequest(modelName, '/v1/messages', backendRes.statusCode || 200, timer.elapsed(), undefined, `${server.host}:${server.port}`, promptPreview);
                         resolve();
                     });
                 }
                 else {
-                    // For non-streaming, collect response and forward
+                    // Non-streaming: collect full response then apply fixes
                     let responseData = '';
                     backendRes.on('data', (chunk) => {
                         responseData += chunk.toString();
                     });
                     backendRes.on('end', async () => {
-                        res.writeHead(backendRes.statusCode || 200, {
-                            'Content-Type': 'application/json',
-                        });
-                        res.end(responseData);
+                        let finalResponse = responseData;
+                        try {
+                            const responseObj = JSON.parse(responseData);
+                            if (Array.isArray(responseObj.content)) {
+                                const textBlocks = responseObj.content.filter((c) => c.type === 'text');
+                                const allText = textBlocks.map((c) => c.text ?? '').join('');
+                                const { toolCalls, cleanText, skippedNames } = parseXmlToolCalls(allText);
+                                if (toolCalls.length > 0) {
+                                    // Fix 1: XML tool calls
+                                    console.error(`[Router] Converting ${toolCalls.length} XML tool call(s) to tool_use blocks`);
+                                    const newContent = responseObj.content.filter((c) => c.type !== 'text');
+                                    if (cleanText)
+                                        newContent.push({ type: 'text', text: cleanText });
+                                    for (const tc of toolCalls) {
+                                        newContent.push({ type: 'tool_use', id: generateToolUseId(), name: tc.name, input: tc.input });
+                                    }
+                                    responseObj.content = newContent;
+                                    responseObj.stop_reason = 'tool_use';
+                                    finalResponse = JSON.stringify(responseObj);
+                                }
+                                else if (allText && !cleanText) {
+                                    const errorCycles = countConsecutiveErrorCycles(requestBody);
+                                    // Fix 3: error feedback with loop detection
+                                    if (errorCycles >= 2) {
+                                        console.error(`[Router] Stripping fix3 after ${errorCycles} error cycles (${skippedNames.join(', ')})`);
+                                        responseObj.content = responseObj.content.filter((c) => c.type !== 'text');
+                                        finalResponse = JSON.stringify(responseObj);
+                                    }
+                                    else {
+                                        console.error(`[Router] Forwarding ${skippedNames.length} empty tool_use block(s) for error feedback [cycle ${errorCycles + 1}] (${skippedNames.join(', ')})`);
+                                        const emptyToolUseBlocks = skippedNames.map(name => ({
+                                            type: 'tool_use',
+                                            id: generateToolUseId(),
+                                            name,
+                                            input: {},
+                                        }));
+                                        responseObj.content = [
+                                            ...responseObj.content.filter((c) => c.type !== 'text'),
+                                            ...emptyToolUseBlocks,
+                                        ];
+                                        responseObj.stop_reason = 'tool_use';
+                                        finalResponse = JSON.stringify(responseObj);
+                                    }
+                                }
+                                else if (cleanText && skippedNames.length > 0) {
+                                    const errorCycles = countConsecutiveErrorCycles(requestBody);
+                                    // Fix 4: text + malformed tool calls with loop detection
+                                    if (errorCycles >= 2) {
+                                        console.error(`[Router] Stripping fix4 malformed call(s) after ${errorCycles} error cycles (${skippedNames.join(', ')})`);
+                                        responseObj.content = [
+                                            ...responseObj.content.filter((c) => c.type !== 'text'),
+                                            { type: 'text', text: cleanText },
+                                        ];
+                                        finalResponse = JSON.stringify(responseObj);
+                                    }
+                                    else {
+                                        console.error(`[Router] Text + ${skippedNames.length} malformed tool call(s), forwarding empty tool_use for error feedback [cycle ${errorCycles + 1}] (${skippedNames.join(', ')})`);
+                                        const emptyToolUseBlocks = skippedNames.map(name => ({
+                                            type: 'tool_use',
+                                            id: generateToolUseId(),
+                                            name,
+                                            input: {},
+                                        }));
+                                        responseObj.content = [
+                                            ...responseObj.content.filter((c) => c.type !== 'text'),
+                                            { type: 'text', text: cleanText },
+                                            ...emptyToolUseBlocks,
+                                        ];
+                                        responseObj.stop_reason = 'tool_use';
+                                        finalResponse = JSON.stringify(responseObj);
+                                    }
+                                }
+                                else {
+                                    // Fix 2: Thinking-only
+                                    const hasText = responseObj.content.some((c) => c.type === 'text' && c.text);
+                                    const thinkingBlocks = responseObj.content.filter((c) => c.type === 'thinking');
+                                    if (!hasText && thinkingBlocks.length > 0) {
+                                        console.error('[Router] Injecting fallback text block (thinking-only response detected)');
+                                        const thinkingText = thinkingBlocks.map((c) => c.thinking ?? '').join('\n');
+                                        responseObj.content.push({ type: 'text', text: thinkingText });
+                                        finalResponse = JSON.stringify(responseObj);
+                                    }
+                                }
+                            }
+                        }
+                        catch {
+                            // Not valid JSON or unexpected shape — forward original
+                        }
+                        res.writeHead(backendRes.statusCode || 200, { 'Content-Type': 'application/json' });
+                        res.end(finalResponse);
                         await this.logRequest(modelName, '/v1/messages', backendRes.statusCode || 200, timer.elapsed(), undefined, `${server.host}:${server.port}`, promptPreview);
                         resolve();
                     });