feat: v1.6.7 - response flicker fix, code block preview in tool cards, deploy automation

Author: Brendan Gray

electron-main.js

@@ -29,7 +29,9 @@ app.commandLine.appendSwitch('disable-gpu-vsync'); // Reduce input latency
 
 // ── V8 Performance Flags ──
 // Enable V8 code caching for faster require() on subsequent launches
-app.commandLine.appendSwitch('js-flags', '--optimize-for-size --max-old-space-size=4096');
+// NOTE: --optimize-for-size is intentionally absent — it shrinks compiled machine code
+// at the cost of execution speed, which is the wrong trade-off for a dev tool.
+app.commandLine.appendSwitch('js-flags', '--max-old-space-size=4096');
 
 // ─── Single Instance Lock ─────────────────────────────────────────────
 // Prevent multiple instances from competing for config files, model files,

main/agenticChat.js

@@ -785,7 +785,10 @@ function register(ctx) {
       const hwContextSize = modelStatus.modelInfo?.contextSize || 32768;
 
       // Helper functions (defined early — needed for budget calculation)
-      const estimateTokens = (text) => Math.ceil((text || '').length / 4);
+      // /3.5 gives ~14% more conservative token estimate than /4 — real LLM tokenizers
+      // produce 3–3.5 chars/token for code and JSON (shorter than English prose).
+      // This prevents buildStaticPrompt + buildDynamicContext from overcommitting budget.
+      const estimateTokens = (text) => Math.ceil((text || '').length / 3.5);
 
       // ── ModelProfile-driven budgeting ──
       // The ModelProfile registry provides effective context size, response reserve %,
@@ -931,11 +934,13 @@ function register(ctx) {
       // Dynamic context: memory, RAG, file, error — changes between iterations.
       // Injected into user message instead of system context to avoid KV cache invalidation.
       // Chat mode: skip ALL dynamic context to maximize conversation space.
-      const buildDynamicContext = (taskTypeOverride) => {
+      // budgetOverride: optional cap for dynamic context tokens — used by overflow retry
+      // to shed memory/RAG/file context while preserving tools and preamble.
+      const buildDynamicContext = (taskTypeOverride, budgetOverride) => {
         const effectiveTaskType = taskTypeOverride || taskType;
         // Chat mode: no dynamic context injection — keep the full context for conversation
         if (effectiveTaskType === 'chat') return '';
-        let tokenBudget = Math.floor(maxPromptTokens * 0.4); // Reserve budget for dynamic context
+        let tokenBudget = budgetOverride !== undefined ? budgetOverride : Math.floor(maxPromptTokens * 0.4); // default: 40% of prompt budget
         let prompt = '';
 
         const appendIfBudget = (text, label) => {
@@ -1469,9 +1474,20 @@ function register(ctx) {
               try { await llmEngine.resetSession(true); } catch (_) {}
               sessionJustRotated = true;
               const rotatedBase = buildStaticPrompt();
+              // Fix C: use 10% of prompt budget for dynamic context on retry — drops memory/RAG/file
+              // context but keeps tools and preamble fully intact. Prevents repeat overflow on
+              // small-context models without touching the model's tool access.
+              // Fix D: if partial content was generated before the overflow, inject it so the model
+              // continues from where it left off rather than restarting the response from scratch.
+              const _firstTurnPartial = fullResponseText.trim().length > 0
+                ? fullResponseText.substring(Math.max(0, fullResponseText.length - 1500))
+                : '';
+              const _firstTurnHint = _firstTurnPartial
+                ? `\n\nYou were generating a response and the context was reset due to size constraints. Here is the end of what you wrote:\n---\n${_firstTurnPartial}\n---\nContinue directly from where you left off without repeating what you already wrote.`
+                : '';
               currentPrompt = {
                 systemContext: rotatedBase,
-                userMessage: buildDynamicContext() + '\n' + message
+                userMessage: buildDynamicContext(undefined, Math.floor(maxPromptTokens * 0.10)) + '\n' + message + _firstTurnHint
               };
               continue;
             }
@@ -1570,9 +1586,17 @@ function register(ctx) {
               }
 
               const rotatedBase = buildStaticPrompt();
+              // Fix D: include the end of what was generated so far so the model continues
+              // seamlessly rather than restarting the response after context rotation.
+              const _rotationPartial = fullResponseText.trim().length > 0
+                ? fullResponseText.substring(Math.max(0, fullResponseText.length - 1500))
+                : '';
+              const _rotationHint = _rotationPartial
+                ? `\n\nYou were generating a response and context was rotated. Here is the end of what you wrote:\n---\n${_rotationPartial}\n---\nContinue directly from where you left off without repeating what you already wrote.`
+                : `\nContext was rotated. The current user request is: ${message.substring(0, 300)}${message.length > 300 ? '...' : ''}`;
               currentPrompt = {
                 systemContext: rotatedBase,
-                userMessage: buildDynamicContext() + '\n' + convSummary + `\nContext was rotated. The current user request is: ${message.substring(0, 300)}${message.length > 300 ? '...' : ''}`
+                userMessage: buildDynamicContext() + '\n' + convSummary + _rotationHint
               };
               sessionJustRotated = true;
               lastConvSummary = convSummary;
@@ -1937,17 +1961,25 @@ function register(ctx) {
         }
 
         // ── Strip code-fence artifacts from displayed text ──
-        // Re-enabled with per-iteration offset tracking (llm-iteration-begin + iterationStartOffsetRef).
-        // The frontend prepends prior iterations' text so only the current iteration's portion is replaced.
+        // Route any conversational planning text to the thinking panel, then wipe the
+        // main chat iteration slot clean. This prevents raw JSON tool calls from flashing
+        // in the chat bubble and matches the cloud path behavior.
         if (toolResults.hasToolCalls && toolResults.results.length > 0 && mainWindow) {
-          let cleaned = responseText;
-          cleaned = cleaned.replace(/```(?:tool_call|tool|json)[^\n]*\n[\s\S]*?```/g, '');
-          cleaned = cleaned.replace(/<tool_call>[\s\S]*?<\/tool_call>/g, '');
-          cleaned = cleaned.replace(/\{\s*"(?:tool|name)"\s*:\s*"[^"]+"\s*,\s*"(?:params|arguments)"[\s\S]*?\}\s*\}/g, '');
-          cleaned = cleaned.replace(/\n{3,}/g, '\n\n').trim();
-          if (cleaned !== responseText) {
-            mainWindow.webContents.send('llm-replace-last', cleaned);
-          }
+          // Extract planning text — everything the model wrote before the first tool call indicator
+          const toolIndicators = ['{"tool":', '```tool_call', '```json\n{"tool"', '<tool_call>'];
+          let splitIdx = responseText.length;
+          for (const indicator of toolIndicators) {
+            const idx = responseText.indexOf(indicator);
+            if (idx >= 0 && idx < splitIdx) splitIdx = idx;
+          }
+          const planningText = responseText.substring(0, splitIdx).trim();
+          if (planningText) {
+            // Planning text belongs in the thinking panel, not the main chat bubble
+            mainWindow.webContents.send('llm-thinking-token', planningText);
+          }
+          // Wipe this iteration's streamed content from main chat — the final answer
+          // streams clean in the last iteration that produces no tool calls.
+          mainWindow.webContents.send('llm-replace-last', '');
         }
 
         if (!toolResults.hasToolCalls || toolResults.results.length === 0) {

package.json

@@ -1,6 +1,6 @@
 {
   "name": "guide-ide",
-  "version": "1.6.4",
+  "version": "1.6.7",
   "description": "guIDE - AI-Powered Offline IDE with local LLM, RAG, MCP tools, browser automation, and integrated terminal",
   "author": {
     "name": "Brendan Gray",

src/components/Chat/ChatPanel.tsx

@@ -2456,22 +2456,43 @@ ${e.message}`,
                               <div className="text-[11px] text-[#858585]">Completed</div>
                             </CollapsibleToolBlock>
                           ))}
-                          {executingTools.map((toolData, i) => (
-                            <CollapsibleToolBlock key={`exec-${i}`} label={getToolLabel(toolData, 'running')} icon="⟳">
-                              <div>
-                                <div className="flex items-center gap-2 mb-2">
-                                  <Loader2 size={12} className="animate-spin text-[#007acc]" />
-                                  <span className="text-[11px] text-[#858585]">Executing...</span>
+                          {executingTools.map((toolData, i) => {
+                            const isCodeWriteTool = ['write_file', 'create_file', 'edit_file', 'append_to_file'].includes(toolData.tool);
+                            const codeContent = toolData.params?.content as string | undefined;
+                            const filePath = ((toolData.params?.filePath || toolData.params?.fileName || '') as string);
+                            const ext = filePath.includes('.') ? filePath.split('.').pop()?.toLowerCase() || '' : '';
+                            const langMap: Record<string, string> = {
+                              ts: 'typescript', tsx: 'tsx', js: 'javascript', jsx: 'jsx',
+                              py: 'python', rs: 'rust', go: 'go', java: 'java', cs: 'csharp',
+                              cpp: 'cpp', c: 'c', html: 'html', css: 'css', json: 'json',
+                              yaml: 'yaml', yml: 'yaml', md: 'markdown', sh: 'bash',
+                              bat: 'batch', txt: 'text', xml: 'xml', sql: 'sql',
+                            };
+                            const language = langMap[ext] || ext || 'code';
+                            return (
+                              <CollapsibleToolBlock key={`exec-${i}`} label={getToolLabel(toolData, 'running')} icon="⟳">
+                                <div>
+                                  <div className="flex items-center gap-2 mb-2">
+                                    <Loader2 size={12} className="animate-spin text-[#007acc]" />
+                                    <span className="text-[11px] text-[#858585]">Executing...</span>
+                                  </div>
+                                  {isCodeWriteTool && codeContent ? (
+                                    <CodeBlock
+                                      code={codeContent}
+                                      language={language}
+                                      onApply={() => {}}
+                                      isToolCall={true}
+                                    />
+                                  ) : toolData.params && Object.keys(toolData.params).length > 0 ? (
+                                    <>
+                                      <div className="text-[10px] text-[#858585] mb-1 font-medium tracking-wide">PARAMETERS</div>
+                                      <pre className="whitespace-pre-wrap text-[11px] font-mono text-[#d4d4d4] bg-[#1e1e1e] rounded-md p-2">{JSON.stringify(toolData.params, null, 2)}</pre>
+                                    </>
+                                  ) : null}
                                 </div>
-                                {toolData.params && Object.keys(toolData.params).length > 0 && (
-                                  <>
-                                    <div className="text-[10px] text-[#858585] mb-1 font-medium tracking-wide">PARAMETERS</div>
-                                    <pre className="whitespace-pre-wrap text-[11px] font-mono text-[#d4d4d4] bg-[#1e1e1e] rounded-md p-2">{JSON.stringify(toolData.params, null, 2)}</pre>
-                                  </>
-                                )}
-                              </div>
-                            </CollapsibleToolBlock>
-                          ))}
+                              </CollapsibleToolBlock>
+                            );
+                          })}
                         </ToolCallGroup>
                       </div>
                     )}

src/components/Chat/hooks/useChatStreaming.ts

@@ -167,8 +167,24 @@ export function useChatStreaming(): ChatStreamingState {
     // Anti-hallucination: backend detected fake tool results
     const cleanupReplace = api.onLlmReplaceLast?.((cleanedText: string) => {
       if (streamEpochRef.current !== activeEpochRef.current) return;
-      // Preserve text from prior iterations — only replace current iteration's portion
       const prefix = streamBufferRef.current.slice(0, iterationStartOffsetRef.current);
+      // When cleanedText is empty (tool-call iteration wipe), the streamed planning text
+      // would vanish abruptly from the main chat. Promote it to a thinking segment so it
+      // transitions visually rather than disappearing — prevents the jarring flash/blank effect.
+      if (!cleanedText) {
+        const iterationText = streamBufferRef.current.slice(iterationStartOffsetRef.current).trim();
+        if (iterationText.length > 10) {
+          // Avoid duplication: backend may have already sent this text as llm-thinking-token.
+          // Only push if last thinking segment doesn't already start with the same content.
+          const lastSeg = thinkingSegmentsRef.current[thinkingSegmentsRef.current.length - 1] || '';
+          const firstChunk = iterationText.substring(0, 80);
+          if (!lastSeg.includes(firstChunk)) {
+            thinkingSegmentsRef.current.push(iterationText);
+            scheduleThinkingUpdate();
+          }
+        }
+      }
+      // Preserve text from prior iterations — only replace current iteration's portion
       streamBufferRef.current = prefix + cleanedText;
       // Jump display to buffer end — corrections show immediately, no typewriter delay
       displayPosRef.current = streamBufferRef.current.length;

website/next.config.js

@@ -8,6 +8,15 @@ const nextConfig = {
   },
   async headers() {
     return [
+      // Allow the 404 game page to be embedded in same-origin iframes
+      // (not-found.tsx renders the game in an iframe; DENY would block it)
+      {
+        source: '/404-game.html',
+        headers: [
+          { key: 'X-Frame-Options', value: 'SAMEORIGIN' },
+          { key: 'Content-Security-Policy', value: "frame-ancestors 'self'" },
+        ],
+      },
       {
         source: '/(.*)',
         headers: [

website/public/updates/latest.yml

@@ -1,8 +1,8 @@
-version: 2.4.2
+version: 1.6.7
 files:
-  - url: guIDE-Setup-2.4.2.exe
-    sha512: gRkj8j0P6rtrZbK0S8jminTw3PjE/wp7qxdwM85hEFcX4CAOP2srhaChqXy2hh9qNH36PeDr8NDuiQ+rrE3wsg==
-    size: 187949120
-path: guIDE-Setup-2.4.2.exe
-sha512: gRkj8j0P6rtrZbK0S8jminTw3PjE/wp7qxdwM85hEFcX4CAOP2srhaChqXy2hh9qNH36PeDr8NDuiQ+rrE3wsg==
-releaseDate: '2026-02-26T23:38:09.045Z'
+  - url: guIDE-Setup-1.6.7.exe
+    sha512: E1COJgnzT0cFCecAQz4R2zyAufQRSP0CK5WrNo+HaVD0ZHWMsN8ui8ZBONZZtr+YWFRgsBSz0/KTd7Gd2soEYA==
+    size: 197957257
+path: guIDE-Setup-1.6.7.exe
+sha512: E1COJgnzT0cFCecAQz4R2zyAufQRSP0CK5WrNo+HaVD0ZHWMsN8ui8ZBONZZtr+YWFRgsBSz0/KTd7Gd2soEYA==
+releaseDate: '2026-03-01T05:39:42.895Z'

website/src/app/download/page.tsx

@@ -2,7 +2,7 @@
 
 // Single source of truth for the displayed release version.
 // Updated automatically by: npm run release:deploy (from IDE root)
-const CURRENT_VERSION = '1.6.4';
+const CURRENT_VERSION = '1.6.7';
 
 import Link from 'next/link';
 import { useState } from 'react';