fix(desktop): preserve Markdown whitespace through media extraction

Remove whole-document whitespace cleanup that erased hard breaks and fenced-code spacing across all five ingestion paths. Limit generated-image gap joining to the removed image itself. Keep soft newlines as soft breaks.

Investigated #97117 and the approaches in #97431 (@wooyongbin3-cpu) and #97175 (@Jackal991); both leave fenced-code whitespace and sibling ingestion paths exposed, so remove the destructive normalization instead.

apps/desktop/src/lib/chat-messages/parts.ts

@@ -34,8 +34,6 @@ export function renderMediaTags(text: string): string {
       (_match, lead: string, value: string, trailer: string) => `${lead}${mediaLink(value)}${trailer}`
     )
     .replace(MEDIA_TAG_RE, (_match, value: string) => mediaLink(value))
-    .replace(/[ \t]+\n/g, '\n')
-    .replace(/\n{3,}/g, '\n\n')
 }
 
 export function assistantTextPart(text: string, timestamp?: number): ChatMessagePart {

apps/desktop/src/lib/embedded-images.ts

@@ -105,13 +105,6 @@ function embeddedImageRemovalRange(text: string, dataStart: number, dataEnd: num
   return { end, start }
 }
 
-function normalizeCleanedText(text: string): string {
-  return text
-    .replace(/[ \t]+\n/g, '\n')
-    .replace(/\n{3,}/g, '\n\n')
-    .trim()
-}
-
 export function extractEmbeddedImages(text: string): EmbeddedImageExtraction {
   if (!text || !text.includes(DATA_IMAGE_PREFIX)) {
     return { cleanedText: text, images: [] }
@@ -150,7 +143,7 @@ export function extractEmbeddedImages(text: string): EmbeddedImageExtraction {
 
   pieces.push(text.slice(appendCursor))
 
-  return { cleanedText: normalizeCleanedText(pieces.join('')), images }
+  return { cleanedText: pieces.join('').trim(), images }
 }
 
 export function embeddedImageUrls(text: string): string[] {

apps/desktop/src/lib/generated-images.ts

@@ -82,17 +82,16 @@ export function stripGeneratedImageEchoes(text: string, sources: readonly string
     return text
   }
 
-  let next = text.replace(/!\[[^\]\n]*\]\([^)\n]*\)/g, '').replace(/\[[^\]\n]*\]\(\s*#media:[^)\n]*\)/g, '')
+  // Join only the gap left by an image, never normalize unrelated Markdown or code.
+  let next = text
+    .replace(/([ \t]?)!\[[^\]\n]*\]\([^)\n]*\)([ \t]?)/g, (_match, before: string, after: string) => before || after)
+    .replace(/\[[^\]\n]*\]\(\s*#media:[^)\n]*\)/g, '')
 
   for (const source of unique([...sources])) {
     next = next.replace(new RegExp(String.raw`(^|[\s([{])<?${regexEscape(source)}>?(?=$|[\s)\]},.!?])`, 'g'), '$1')
   }
 
-  return next
-    .replace(/[ \t]+\n/g, '\n')
-    .replace(/\n{3,}/g, '\n\n')
-    .replace(/[ \t]{2,}/g, ' ')
-    .trim()
+  return next.trim()
 }
 
 /** Strip generated-image echoes from text parts, dropping any part left empty.

apps/desktop/src/lib/markdown-preprocess.ts

@@ -664,7 +664,6 @@ export function preprocessMarkdown(text: string): string {
       return leading + transformed + trailing
     })
     .join('')
-    .replace(/[ \t]+\n/g, '\n')
 }
 
 /**

apps/desktop/src/lib/markdown-whitespace.test.ts

@@ -0,0 +1,25 @@
+import { describe, expect, it } from 'vitest'
+
+import { assistantTextPart, renderMediaTags } from './chat-messages/parts'
+import { extractEmbeddedImages } from './embedded-images'
+import { stripGeneratedImageEchoes } from './generated-images'
+import { preprocessMarkdown } from './markdown-preprocess'
+import { stripPreviewTargets } from './preview-targets'
+
+describe('Markdown whitespace semantics', () => {
+  it('preserves hard and soft breaks and fenced whitespace through display preprocessing', () => {
+    for (const input of ['First line  \nSecond line', 'Soft first\nSoft second', '```python\nvalue = 1  \nvalue = 2 \n```']) {
+      expect(preprocessMarkdown(input)).toBe(input)
+    }
+  })
+
+  it('preserves prose hard breaks when media and preview extraction runs', () => {
+    const prose = 'First line  \nSecond line\n\n```python\nvalue = 1  \nvalue = 2 \n\n\nvalue = 3\n```'
+    const image = 'data:image/png;base64,' + 'A'.repeat(64)
+    expect(stripPreviewTargets(prose + '\n\n[Preview: x](#preview:test)')).toContain(prose)
+    expect(renderMediaTags(prose + '\n\nMEDIA: /tmp/image.png')).toContain(prose)
+    expect(assistantTextPart(prose)).toMatchObject({ type: 'text', text: prose })
+    expect(extractEmbeddedImages(prose + '\n\n' + image).cleanedText).toContain(prose)
+    expect(stripGeneratedImageEchoes(prose + '\n\n![result](/tmp/image.png)', ['/tmp/image.png'])).toContain(prose)
+  })
+})

apps/desktop/src/lib/preview-targets.ts

@@ -1,11 +1,7 @@
 const PREVIEW_MARKDOWN_RE = /\[Preview:[^\]]+\]\((?<href>#preview[:/][^)]+)\)/gi
 
 export function stripPreviewTargets(text: string): string {
-  return text
-    .replace(PREVIEW_MARKDOWN_RE, '')
-    .replace(/[ \t]+\n/g, '\n')
-    .replace(/\n{3,}/g, '\n\n')
-    .trim()
+  return text.replace(PREVIEW_MARKDOWN_RE, '').trim()
 }
 
 export function extractPreviewTargets(text: string): string[] {

website/docs/user-guide/desktop.md

@@ -41,6 +41,7 @@ The desktop app is organized as a chat-first window with a left sidebar for navi
 The center of the app. You get:
 
 - **Streaming responses** with live tool activity and structured tool-call summaries as the agent works.
+- **Markdown line breaks** follow Markdown semantics: two trailing spaces create a hard line break; an ordinary newline stays a soft break. Media and preview extraction preserve these breaks and whitespace inside fenced code blocks.
 - **The same conversation history** as every other Hermes surface — sessions started here resume in the CLI/TUI and vice versa.
 - **Drag-and-drop files** anywhere in the chat area to attach them to your next message.
 - **A right-hand preview rail** — render web pages, files, and tool outputs side by side while you keep chatting.