polydev-ai 1.9.40 → 1.9.42

package/lib/cliManager.js

@@ -578,6 +578,8 @@ This is a known issue with @google/gemini-cli@0.3.4 and older Node.js versions.`
               success: true,
               content,
               tokens_used: this.estimateTokens(prompt + content),
+              input_tokens: Math.ceil(prompt.length / 4),
+              output_tokens: Math.ceil(content.length / 4),
               latency_ms: Date.now() - startTime,
               provider: providerId,
               mode: 'args',
@@ -651,8 +653,8 @@ This is a known issue with @google/gemini-cli@0.3.4 and older Node.js versions.`
             args = ['--model', cliModel, ...args, prompt];
           } else if (providerId === 'gemini_cli') {
             // Gemini CLI: -m for model, -p for prompt (headless mode)
-            // Add prompt prefix to prevent tool planning in non-interactive mode
-            const geminiPrompt = `Answer directly without using any tools, file operations, or searches. Do not say "I will search" or "I will look up". Provide your analysis immediately.\n\n${prompt}`;
+            // Add prompt prefix to prevent tool planning in non-interactive mode and enforce English
+            const geminiPrompt = `IMPORTANT: Always respond in English. Answer directly without using any tools, file operations, or searches. Do not say "I will search" or "I will look up". Provide your analysis immediately.\n\n${prompt}`;
             args = ['-m', cliModel, '-p', geminiPrompt];
           } else {
             // Default: just append prompt
@@ -662,8 +664,8 @@ This is a known issue with @google/gemini-cli@0.3.4 and older Node.js versions.`
           // No model specified
           if (providerId === 'gemini_cli') {
             // Gemini CLI: -p for headless mode
-            // Add prompt prefix to prevent tool planning in non-interactive mode
-            const geminiPrompt = `Answer directly without using any tools, file operations, or searches. Do not say "I will search" or "I will look up". Provide your analysis immediately.\n\n${prompt}`;
+            // Add prompt prefix to prevent tool planning in non-interactive mode and enforce English
+            const geminiPrompt = `IMPORTANT: Always respond in English. Answer directly without using any tools, file operations, or searches. Do not say "I will search" or "I will look up". Provide your analysis immediately.\n\n${prompt}`;
             args = ['-p', geminiPrompt];
           } else {
             args = [...args, prompt];
@@ -689,6 +691,8 @@ This is a known issue with @google/gemini-cli@0.3.4 and older Node.js versions.`
                   success: true,
                   content: jsonResult.content,
                   tokens_used: jsonResult.tokens_used,
+                  input_tokens: jsonResult.input_tokens,
+                  output_tokens: jsonResult.output_tokens,
                   latency_ms: Date.now() - startTime,
                   provider: providerId,
                   mode: 'args',
@@ -715,6 +719,8 @@ This is a known issue with @google/gemini-cli@0.3.4 and older Node.js versions.`
               success: true,
               content,
               tokens_used: this.estimateTokens(prompt + content),
+              input_tokens: Math.ceil(prompt.length / 4),
+              output_tokens: Math.ceil(content.length / 4),
               latency_ms: Date.now() - startTime,
               provider: providerId,
               mode: 'args',
@@ -771,6 +777,8 @@ This is a known issue with @google/gemini-cli@0.3.4 and older Node.js versions.`
                   success: true,
                   content,
                   tokens_used: this.estimateTokens(prompt + content),
+                  input_tokens: Math.ceil(prompt.length / 4),
+                  output_tokens: Math.ceil(content.length / 4),
                   latency_ms: Date.now() - startTime,
                   provider: providerId,
                   mode: 'args',
@@ -1038,17 +1046,13 @@ This is a known issue with @google/gemini-cli@0.3.4 and older Node.js versions.`
       const content = json.result;
       
       // Extract primary model from modelUsage
-      // The primary model is the one with highest cost - that's the user's configured main model
-      // (Haiku is used internally for quick tasks, but the expensive model is what the user chose)
       let primaryModel = CLI_DEFAULT_MODELS['claude_code'] || 'cli_default';
       const modelUsage = json.modelUsage || {};
       const modelNames = Object.keys(modelUsage);
       
       if (modelNames.length === 1) {
-        // Only one model used - that's the primary
         primaryModel = modelNames[0];
       } else if (modelNames.length > 1) {
-        // Multiple models - the one with highest cost is the user's configured main model
         let highestCost = -1;
         for (const [modelName, usage] of Object.entries(modelUsage)) {
           const cost = usage.costUSD || 0;
@@ -1059,13 +1063,15 @@ This is a known issue with @google/gemini-cli@0.3.4 and older Node.js versions.`
         }
       }
       
-      // Calculate total tokens (excluding cache tokens which are just infrastructure overhead)
+      // Calculate separate input and output tokens
+      let totalInputTokens = 0;
+      let totalOutputTokens = 0;
       let totalTokens = 0;
       let cacheTokens = 0;
       for (const usage of Object.values(modelUsage)) {
-        // Count actual input/output tokens
+        totalInputTokens += (usage.inputTokens || 0);
+        totalOutputTokens += (usage.outputTokens || 0);
         totalTokens += (usage.inputTokens || 0) + (usage.outputTokens || 0);
-        // Track cache tokens separately (for cost calculations, but not displayed as "tokens used")
         cacheTokens += (usage.cacheReadInputTokens || 0) + (usage.cacheCreationInputTokens || 0);
       }
       
@@ -1073,14 +1079,15 @@ This is a known issue with @google/gemini-cli@0.3.4 and older Node.js versions.`
         content,
         model_used: primaryModel,
         tokens_used: totalTokens || json.usage?.input_tokens + json.usage?.output_tokens || 0,
-        cache_tokens: cacheTokens, // Separate field for cache tokens
+        input_tokens: totalInputTokens,
+        output_tokens: totalOutputTokens,
+        cache_tokens: cacheTokens,
         cost_usd: json.total_cost_usd || 0,
         model_usage: modelUsage,
         session_id: json.session_id,
         duration_ms: json.duration_ms
       };
     } catch (e) {
-      // Not valid JSON, return null to fall back to text parsing
       return null;
     }
   }
@@ -1374,6 +1381,119 @@ This is a known issue with @google/gemini-cli@0.3.4 and older Node.js versions.`
     return this.providers.get(providerId);
   }
 
+  // ============================================
+  // Streaming CLI Methods (CLI-as-API real streaming)
+  // ============================================
+
+  /**
+   * Check if a provider supports streaming output
+   */
+  supportsStreaming(providerId) {
+    return providerId === 'claude_code' || providerId === 'gemini_cli';
+  }
+
+  /**
+   * Build CLI args for streaming mode (NDJSON output)
+   * @param {string} providerId - 'claude_code' or 'gemini_cli'
+   * @param {string} prompt - The user prompt
+   * @param {string|null} model - Optional model name
+   * @returns {string[]} CLI arguments array
+   */
+  buildStreamingArgs(providerId, prompt, model = null) {
+    if (providerId === 'claude_code') {
+      const args = ['-p', '--output-format', 'stream-json', '--verbose', '--include-partial-messages'];
+      
+      // Add model flag if specified
+      if (model) {
+        // Normalize model names to Claude CLI aliases
+        const claudeModelMap = {
+          'claude-opus-4-5': 'opus',
+          'claude-opus-4.5': 'opus',
+          'claude-4.5-opus': 'opus',
+          'claude-opus-4-5-20250514': 'opus',
+          'claude-sonnet-4-5': 'sonnet',
+          'claude-sonnet-4.5': 'sonnet',
+          'claude-4.5-sonnet': 'sonnet',
+          'claude-sonnet-4-5-20250514': 'sonnet',
+          'claude-3-5-sonnet': 'sonnet',
+          'claude-3-5-haiku': 'haiku',
+          'claude-haiku-3-5': 'haiku',
+        };
+        const cliModel = claudeModelMap[model.toLowerCase()] || model;
+        args.unshift('--model', cliModel);
+      }
+      
+      args.push(prompt);
+      return args;
+    }
+
+    if (providerId === 'gemini_cli') {
+      const args = ['-o', 'stream-json'];
+      
+      // Add model flag if specified
+      if (model) {
+        args.push('-m', model);
+      }
+      
+      // Add prompt prefix to prevent tool planning in non-interactive mode and enforce English
+      const geminiPrompt = `IMPORTANT: Always respond in English. Answer directly without using any tools, file operations, or searches. Do not say "I will search" or "I will look up". Provide your analysis immediately.\n\n${prompt}`;
+      args.push('-p', geminiPrompt);
+      return args;
+    }
+
+    throw new Error(`Provider ${providerId} does not support streaming`);
+  }
+
+  /**
+   * Spawn a CLI process in streaming mode.
+   * Returns the raw child process so the caller can read stdout line-by-line.
+   * 
+   * @param {string} providerId - 'claude_code' or 'gemini_cli'
+   * @param {string} prompt - The user prompt
+   * @param {string|null} model - Optional model name
+   * @returns {import('child_process').ChildProcess} The spawned process
+   */
+  spawnStreamingCli(providerId, prompt, model = null) {
+    const provider = this.providers.get(providerId);
+    if (!provider) {
+      throw new Error(`Unknown provider: ${providerId}`);
+    }
+
+    const args = this.buildStreamingArgs(providerId, prompt, model);
+    
+    console.log(`[Polydev CLI] Spawning streaming ${providerId}: ${provider.command} ${args.slice(0, 4).join(' ')}...`);
+
+    const child = spawn(provider.command, args, {
+      stdio: ['pipe', 'pipe', 'pipe'],
+      shell: process.platform === 'win32',
+      env: {
+        ...process.env,
+        HOME: process.env.HOME || os.homedir(),
+        CI: '1',
+        NO_COLOR: '1',
+        TERM: 'dumb',
+        NONINTERACTIVE: '1',
+        NO_BROWSER: '1',
+        BROWSER: 'echo',
+        DISPLAY: '',
+        HEADLESS: '1',
+        CODEX_DISABLE_UPDATE_CHECK: '1',
+        CLAUDE_CODE_DISABLE_UPDATE_CHECK: '1',
+        GEMINI_NO_BROWSER: '1',
+        GOOGLE_NO_BROWSER: '1',
+        npm_config_update_notifier: 'false',
+        NO_UPDATE_NOTIFIER: '1'
+      }
+    });
+
+    // Close stdin immediately to prevent hanging (critical for Gemini CLI)
+    if (child.stdin) {
+      child.stdin.end();
+    }
+
+    return child;
+  }
+
   // ============================================
   // Status Reporting Methods
   // ============================================

package/lib/tunnelClient.js

@@ -205,8 +205,13 @@ class TunnelClient {
    * Handle a single tunnel request by routing to CLI
    */
   async handleRequest(request) {
+    // Use streaming path for providers that support it
+    if (this.cliManager.supportsStreaming(request.provider)) {
+      return this.handleStreamingRequest(request);
+    }
+
     const startTime = Date.now();
-    console.error(`[Tunnel] Processing request ${request.id} → ${request.provider}`);
+    console.error(`[Tunnel] Processing request ${request.id} → ${request.provider} (non-streaming)`);
 
     try {
       const result = await this.cliManager.sendCliPrompt(
@@ -226,6 +231,8 @@ class TunnelClient {
           content: result.content || '',
           model_used: result.model || result.detectedModel || request.provider,
           tokens_used: result.tokens_used || null,
+          input_tokens: result.input_tokens || null,
+          output_tokens: result.output_tokens || null,
           latency_ms: latencyMs,
         });
       } else {
@@ -247,6 +254,319 @@ class TunnelClient {
     }
   }
 
+  /**
+   * Handle a streaming-capable tunnel request (Claude Code or Gemini CLI).
+   * Spawns CLI with stream-json flags, reads NDJSON stdout line-by-line,
+   * extracts text deltas, batches them, and POSTs chunks to /api/tunnel/stream-chunk.
+   * On completion, POSTs final response to /api/tunnel/respond for backward compat.
+   */
+  async handleStreamingRequest(request) {
+    const startTime = Date.now();
+    const requestId = request.id;
+    const provider = request.provider;
+    console.error(`[Tunnel] Processing request ${requestId} → ${provider} (streaming)`);
+
+    let child;
+    let fullContent = '';
+    let chunkIndex = 0;
+    let pendingText = '';
+    let lastFlushTime = Date.now();
+    let modelUsed = null;
+    let inputTokens = null;
+    let outputTokens = null;
+    let costUsd = null;
+
+    const BATCH_INTERVAL_MS = 150;
+    const BATCH_MIN_CHARS = 1;
+
+    // Flush accumulated text as a stream chunk
+    const flushChunk = async (force = false) => {
+      if (!pendingText) return;
+      if (!force && pendingText.length < BATCH_MIN_CHARS && (Date.now() - lastFlushTime) < BATCH_INTERVAL_MS) {
+        return;
+      }
+
+      const text = pendingText;
+      pendingText = '';
+      lastFlushTime = Date.now();
+
+      try {
+        await this.sendStreamChunk({
+          request_id: requestId,
+          chunk_index: chunkIndex++,
+          content: text,
+        });
+      } catch (err) {
+        console.error(`[Tunnel] Failed to send stream chunk for ${requestId}:`, err.message);
+      }
+    };
+
+    // Set up periodic flushing
+    const flushInterval = setInterval(async () => {
+      if (pendingText.length > 0 && (Date.now() - lastFlushTime) >= BATCH_INTERVAL_MS) {
+        await flushChunk(true);
+      }
+    }, BATCH_INTERVAL_MS);
+
+    try {
+      child = this.cliManager.spawnStreamingCli(
+        provider,
+        request.prompt,
+        request.model_requested || null
+      );
+
+      // Set up CLI timeout
+      const timeoutHandle = setTimeout(() => {
+        if (child && !child.killed) {
+          console.error(`[Tunnel] Streaming request ${requestId} timed out after ${this.CLI_TIMEOUT_MS}ms`);
+          child.kill('SIGTERM');
+          setTimeout(() => {
+            if (!child.killed) child.kill('SIGKILL');
+          }, 2000);
+        }
+      }, this.CLI_TIMEOUT_MS);
+
+      // Read stdout line-by-line (NDJSON)
+      let lineBuf = '';
+      
+      child.stdout.on('data', (data) => {
+        lineBuf += data.toString();
+        
+        // Process complete lines
+        let newlineIdx;
+        while ((newlineIdx = lineBuf.indexOf('\n')) !== -1) {
+          const line = lineBuf.slice(0, newlineIdx).trim();
+          lineBuf = lineBuf.slice(newlineIdx + 1);
+          
+          if (!line) continue;
+          
+          try {
+            const event = JSON.parse(line);
+            const delta = this.extractTextDelta(provider, event);
+            
+            if (delta) {
+              fullContent += delta;
+              pendingText += delta;
+            }
+
+            // Extract metadata from events
+            if (provider === 'claude_code') {
+              this._extractClaudeMetadata(event, (meta) => {
+                if (meta.model) modelUsed = meta.model;
+                if (meta.inputTokens != null) inputTokens = meta.inputTokens;
+                if (meta.outputTokens != null) outputTokens = meta.outputTokens;
+                if (meta.costUsd != null) costUsd = meta.costUsd;
+              });
+            } else if (provider === 'gemini_cli') {
+              this._extractGeminiMetadata(event, (meta) => {
+                if (meta.model) modelUsed = meta.model;
+                if (meta.inputTokens != null) inputTokens = meta.inputTokens;
+                if (meta.outputTokens != null) outputTokens = meta.outputTokens;
+              });
+            }
+          } catch {
+            // Skip non-JSON lines (stderr leaking into stdout, etc.)
+          }
+        }
+      });
+
+      // Capture stderr for debugging
+      let stderrBuf = '';
+      child.stderr.on('data', (data) => {
+        stderrBuf += data.toString();
+      });
+
+      // Wait for process to exit
+      await new Promise((resolve, reject) => {
+        child.on('close', (code) => {
+          clearTimeout(timeoutHandle);
+          resolve(code);
+        });
+        child.on('error', (err) => {
+          clearTimeout(timeoutHandle);
+          reject(err);
+        });
+      });
+
+      // Flush any remaining text
+      await flushChunk(true);
+      clearInterval(flushInterval);
+
+      // Send final is_final chunk marker
+      if (chunkIndex > 0) {
+        try {
+          await this.sendStreamChunk({
+            request_id: requestId,
+            chunk_index: chunkIndex++,
+            content: '',
+            is_final: true,
+          });
+        } catch (err) {
+          console.error(`[Tunnel] Failed to send final chunk for ${requestId}:`, err.message);
+        }
+      }
+
+      const latencyMs = Date.now() - startTime;
+
+      if (fullContent) {
+        console.error(`[Tunnel] Streaming request ${requestId} completed (${latencyMs}ms, ${chunkIndex} chunks, ${fullContent.length} chars)`);
+        
+        // Estimate tokens if not extracted from metadata
+        const estInputTokens = inputTokens || Math.ceil(request.prompt.length / 4);
+        const estOutputTokens = outputTokens || Math.ceil(fullContent.length / 4);
+
+        await this.sendResponse({
+          request_id: requestId,
+          content: fullContent,
+          model_used: modelUsed || request.provider,
+          tokens_used: (estInputTokens + estOutputTokens) || null,
+          input_tokens: estInputTokens,
+          output_tokens: estOutputTokens,
+          latency_ms: latencyMs,
+        });
+      } else {
+        // No content extracted from streaming — fall back to error
+        const errorMsg = stderrBuf.trim().slice(0, 500) || 'Streaming CLI produced no output';
+        console.error(`[Tunnel] Streaming request ${requestId} failed: ${errorMsg}`);
+        await this.sendResponse({
+          request_id: requestId,
+          error: errorMsg,
+          latency_ms: latencyMs,
+        });
+      }
+
+    } catch (err) {
+      clearInterval(flushInterval);
+      const latencyMs = Date.now() - startTime;
+      console.error(`[Tunnel] Streaming request ${requestId} exception: ${err.message}`);
+
+      // If we got partial content before the error, still send it
+      if (fullContent) {
+        await flushChunk(true);
+        await this.sendResponse({
+          request_id: requestId,
+          content: fullContent,
+          model_used: modelUsed || request.provider,
+          latency_ms: latencyMs,
+        });
+      } else {
+        await this.sendResponse({
+          request_id: requestId,
+          error: err.message || 'Streaming execution failed',
+          latency_ms: latencyMs,
+        });
+      }
+    }
+  }
+
+  /**
+   * Extract text delta from a parsed NDJSON event based on provider type.
+   * Returns the incremental text string, or null if this event isn't a text delta.
+   */
+  extractTextDelta(provider, event) {
+    if (provider === 'claude_code') {
+      // Claude stream_event wrapping raw API events
+      // Filter: type === 'stream_event' && event.event.type === 'content_block_delta'
+      //         && event.event.delta.type === 'text_delta'
+      if (event.type === 'stream_event' &&
+          event.event?.type === 'content_block_delta' &&
+          event.event?.delta?.type === 'text_delta') {
+        return event.event.delta.text || null;
+      }
+      return null;
+    }
+
+    if (provider === 'gemini_cli') {
+      // Gemini: type === 'message' && role === 'assistant' && delta === true
+      if (event.type === 'message' &&
+          event.role === 'assistant' &&
+          event.delta === true) {
+        return event.content || null;
+      }
+      return null;
+    }
+
+    return null;
+  }
+
+  /**
+   * Extract metadata (model, tokens, cost) from Claude Code stream events
+   */
+  _extractClaudeMetadata(event, callback) {
+    // From 'system' init event: model info
+    if (event.type === 'system' && event.subtype === 'init' && event.model) {
+      callback({ model: event.model });
+    }
+
+    // From 'result' event: usage stats
+    if (event.type === 'result') {
+      const meta = {};
+      if (event.usage) {
+        meta.inputTokens = event.usage.input_tokens || 0;
+        meta.outputTokens = event.usage.output_tokens || 0;
+      }
+      if (event.total_cost_usd != null) {
+        meta.costUsd = event.total_cost_usd;
+      }
+      // Extract model from modelUsage keys
+      if (event.modelUsage) {
+        const models = Object.keys(event.modelUsage);
+        if (models.length > 0) {
+          // Pick the model with highest cost as the primary model
+          let primary = models[0];
+          let highestCost = -1;
+          for (const [m, usage] of Object.entries(event.modelUsage)) {
+            const cost = usage.costUSD || 0;
+            if (cost > highestCost) {
+              highestCost = cost;
+              primary = m;
+            }
+          }
+          meta.model = primary;
+        }
+      }
+      callback(meta);
+    }
+  }
+
+  /**
+   * Extract metadata from Gemini CLI stream events
+   */
+  _extractGeminiMetadata(event, callback) {
+    // From 'init' event: model info
+    if (event.type === 'init' && event.model) {
+      callback({ model: event.model });
+    }
+
+    // From 'result' event: usage stats
+    if (event.type === 'result' && event.stats) {
+      callback({
+        inputTokens: event.stats.input_tokens || 0,
+        outputTokens: event.stats.output_tokens || 0,
+      });
+    }
+  }
+
+  /**
+   * Send a stream chunk to the server
+   */
+  async sendStreamChunk(chunkData) {
+    const url = `${this.serverBaseUrl}/api/tunnel/stream-chunk`;
+    const res = await fetch(url, {
+      method: 'POST',
+      headers: {
+        'Authorization': `Bearer ${this.authToken}`,
+        'Content-Type': 'application/json',
+      },
+      body: JSON.stringify(chunkData),
+    });
+
+    if (!res.ok) {
+      const text = await res.text().catch(() => '');
+      throw new Error(`Stream chunk POST failed (${res.status}): ${text}`);
+    }
+  }
+
   /**
    * Send response back to server
    */

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "polydev-ai",
-  "version": "1.9.40",
+  "version": "1.9.42",
   "engines": {
     "node": ">=20.x <=22.x"
   },