feat: CJK word count, pipeline timing, history modal, auto-mute, auto-dictionary

electron/config-store.ts

@@ -30,6 +30,8 @@ const DEFAULT_CONFIG: Record<string, any> = {
   repetitionElimination: true,
   toneRules: [],
   defaultTone: 'professional',
+  autoMuteOnRecord: false,
+  autoLearnDictionary: true,
   history: [],
   totalWordsThisWeek: 0,
 };

electron/llm-service.ts

@@ -3,6 +3,98 @@
  * Handles post-processing, rewriting, and connection testing.
  */
 
+/** Smart truncation: keeps beginning + end of long text, with ellipsis in middle */
+function smartTruncate(text: string, maxLen: number): string {
+  if (!text || text.length <= maxLen) return text;
+  const keepEach = Math.floor((maxLen - 20) / 2);
+  return text.slice(0, keepEach) + '\n... [truncated] ...\n' + text.slice(-keepEach);
+}
+
+/** Truncation limits for each context field (in characters) */
+const CONTEXT_LIMITS = {
+  selectedText: 500,
+  fieldText: 1500,
+  fieldTextWithMarker: 2000,  // higher to account for cursor/selection markers
+  clipboardText: 500,
+  screenContext: 400,
+  recentTranscription: 200,  // per item
+  recentTotal: 3,            // max items
+};
+
+/** Truncate text centered around the cursor position, keeping context on both sides */
+function cursorCenteredTruncate(text: string, cursorPos: number, maxLen: number): { text: string; adjustedPos: number } {
+  if (text.length <= maxLen) return { text, adjustedPos: cursorPos };
+
+  const ellipsis = '\n... [truncated] ...\n';
+  const halfWindow = Math.floor((maxLen - ellipsis.length * 2) / 2);
+  let start = Math.max(0, cursorPos - halfWindow);
+  let end = Math.min(text.length, cursorPos + halfWindow);
+
+  // If one side is shorter, give more to the other
+  if (start === 0) end = Math.min(text.length, maxLen - ellipsis.length);
+  if (end === text.length) start = Math.max(0, text.length - maxLen + ellipsis.length);
+
+  let result = '';
+  let adjustedPos = cursorPos;
+
+  if (start > 0) {
+    result = ellipsis;
+    adjustedPos = cursorPos - start + result.length;
+    result += text.slice(start, end);
+  } else {
+    result = text.slice(0, end);
+  }
+
+  if (end < text.length) {
+    result += ellipsis;
+  }
+
+  return { text: result, adjustedPos };
+}
+
+/** Build rich field context string with cursor/selection markers for the LLM */
+function buildFieldContext(context: any): string | null {
+  const fieldText: string | undefined = context.fieldText;
+  if (!fieldText) return null;
+
+  const range = context.selectionRange as { location: number; length: number } | undefined;
+  const placeholder = context.fieldPlaceholder as string | undefined;
+  const label = context.fieldLabel as string | undefined;
+  const roleDesc = context.fieldRoleDescription || context.fieldRole || 'input field';
+
+  // Build descriptor: ("Message body", text area)
+  const labelPart = label ? `"${label}", ` : '';
+  const descriptor = `(${labelPart}${roleDesc})`;
+
+  if (range && typeof range.location === 'number' && typeof range.length === 'number') {
+    const loc = range.location;
+    const len = range.length;
+
+    if (len > 0 && loc + len <= fieldText.length) {
+      // User has selected text — show [SELECTED: ...] marker
+      const { text: truncated, adjustedPos } = cursorCenteredTruncate(fieldText, loc, CONTEXT_LIMITS.fieldTextWithMarker - 30);
+      const before = truncated.slice(0, adjustedPos);
+      const selectedText = truncated.slice(adjustedPos, adjustedPos + len);
+      const after = truncated.slice(adjustedPos + len);
+      const markedText = before + '[SELECTED: ' + selectedText + ']' + after;
+
+      return `The user selected text to replace with dictation in the ${descriptor}:\n"""\n${markedText}\n"""\nThe dictated text should replace the [SELECTED: ...] portion.`;
+    } else if (len === 0 && loc <= fieldText.length) {
+      // Cursor position — show | marker
+      const { text: truncated, adjustedPos } = cursorCenteredTruncate(fieldText, loc, CONTEXT_LIMITS.fieldTextWithMarker - 10);
+      const before = truncated.slice(0, adjustedPos);
+      const after = truncated.slice(adjustedPos);
+      const markedText = before + '|' + after;
+
+      return `Existing text in the ${descriptor}:\n"""\n${markedText}\n"""\n(The "|" marks the cursor position where the dictated text will be inserted.)`;
+    }
+  }
+
+  // Fallback: no range info, show raw field text
+  const snippet = smartTruncate(fieldText, CONTEXT_LIMITS.fieldText);
+  return `Existing text in the ${descriptor}:\n"""\n${snippet}\n"""\nThe dictated text should flow naturally with this existing content.`;
+}
+
 export class LLMService {
   private getOpts(config: Record<string, any>, provider?: string) {
     const p = provider || config.llmProvider || 'siliconflow';
@@ -52,8 +144,8 @@ export class LLMService {
     return content;
   }
 
-  async process(rawText: string, config: Record<string, any>, context?: any): Promise<string> {
-    if (!rawText.trim()) return '';
+  async process(rawText: string, config: Record<string, any>, context?: any): Promise<{ text: string; systemPrompt: string }> {
+    if (!rawText.trim()) return { text: '', systemPrompt: '' };
     const opts = this.getOpts(config);
 
     const parts: string[] = [
@@ -93,24 +185,55 @@ export class LLMService {
       if (context.windowTitle) {
         parts.push(`Window title: "${context.windowTitle}"`);
       }
+      if (context.url) {
+        parts.push(`URL: ${context.url}`);
+      }
     }
 
-    if (context?.selectedText) {
-      const snippet = context.selectedText.slice(0, 500);
-      parts.push(`\nThe user was editing/viewing this text:\n"""\n${snippet}\n"""\nEnsure the dictation output is consistent and coherent with this context.`);
+    // L1: Field context with cursor/selection position markers
+    const fieldCtx = buildFieldContext(context);
+    if (fieldCtx) {
+      parts.push(`\n${fieldCtx}`);
+    } else if (context?.selectedText) {
+      // Standalone selected text (no field text available)
+      const snippet = smartTruncate(context.selectedText, CONTEXT_LIMITS.selectedText);
+      parts.push(`\nThe user had selected this text:\n"""\n${snippet}\n"""\nEnsure the dictation output is consistent and coherent with this context.`);
+    }
+
+    // Field placeholder hint
+    if (context?.fieldPlaceholder) {
+      parts.push(`The input field's placeholder reads: "${context.fieldPlaceholder}"`);
+    }
+
+    // Clipboard content
+    if (context?.clipboardText) {
+      const snippet = smartTruncate(context.clipboardText, CONTEXT_LIMITS.clipboardText);
+      parts.push(`\nClipboard content:\n"""\n${snippet}\n"""\nThis may provide additional context for the dictation.`);
+    }
+
+    // Recent transcriptions for continuity
+    if (context?.recentTranscriptions?.length > 0) {
+      const recents = context.recentTranscriptions
+        .slice(0, CONTEXT_LIMITS.recentTotal)
+        .map((t: string) => smartTruncate(t, CONTEXT_LIMITS.recentTranscription));
+      parts.push(`\nRecent transcriptions (for continuity):\n${recents.map((r: string, i: number) => `${i + 1}. ${r}`).join('\n')}`);
     }
 
     if (context?.screenContext) {
-      parts.push(`\nScreen context (from OCR): ${context.screenContext}`);
+      const snippet = smartTruncate(context.screenContext, CONTEXT_LIMITS.screenContext);
+      parts.push(`\nScreen context (from OCR): ${snippet}`);
     }
 
-    return this.call({
+    const systemPrompt = parts.join('\n');
+    const text = await this.call({
       ...opts,
       messages: [
-        { role: 'system', content: parts.join('\n') },
+        { role: 'system', content: systemPrompt },
         { role: 'user', content: rawText },
       ],
     });
+
+    return { text, systemPrompt };
   }
 
   async rewrite(selectedText: string, instruction: string, config: Record<string, any>): Promise<string> {
@@ -140,6 +263,22 @@ export class LLMService {
     const model = config.contextOcrModel || 'Qwen/Qwen2-VL-7B-Instruct';
     const opts = this.getOpts(config);
 
+    const prompt = [
+      'Analyze this screenshot to help with voice dictation context. Extract:',
+      '1. APP: Which application is open',
+      '2. TASK: What the user is working on (1 sentence)',
+      '3. KEY TERMS: List any proper nouns, brand names, technical terms, project names, or specialized vocabulary visible on screen (comma-separated)',
+      '4. TEXT CONTEXT: If there is a text input area, summarize what has been written so far (1 sentence)',
+      '',
+      'Format your response exactly as:',
+      'APP: ...',
+      'TASK: ...',
+      'KEY TERMS: ...',
+      'TEXT CONTEXT: ...',
+      '',
+      'Keep each line brief. If a field is not applicable, write "none".',
+    ].join('\n');
+
     const res = await fetch(`${opts.baseUrl}/chat/completions`, {
       method: 'POST',
       headers: {
@@ -153,11 +292,11 @@ export class LLMService {
           role: 'user',
           content: [
             { type: 'image_url', image_url: { url: dataUrl } },
-            { type: 'text', text: 'Briefly describe the content visible on screen in 1-2 sentences. Focus on what app is open and what the user is working on. Be concise.' },
+            { type: 'text', text: prompt },
           ],
         }],
-        max_tokens: 200,
-        temperature: 0.2,
+        max_tokens: 300,
+        temperature: 0.1,
       }),
     });