Desktop App: UI renders massively duplicated content when streaming responses with code blocks #14560

Bug Description

When streaming LLM responses that contain code blocks (code fences, ASCII art, mermaid diagrams), the Desktop App UI renders massively duplicated content. A ~150-line response balloons to 2,300+ lines of visible duplicated text in the message thread.

Environment

• App: OpenCode Desktop App v1.2.10 (macOS)
• OS: macOS (Apple Silicon)
• Model: Claude Opus 4.6 via Amazon Bedrock

Reproduction Steps

1. Open the OpenCode Desktop App
2. Send a prompt that generates a response containing multiple code fences, ASCII art, or mermaid diagram blocks
3. Observe the rendered output in the message thread
4. Copy-paste the rendered thread content — duplicated content is visible

Expected Behavior

Response renders once, ~150 lines, clean and readable.

Actual Behavior

Content is duplicated dozens of times. Specific metrics from copy-pasting the rendered output:

• Expected lines: ~150
• Actual lines: 2,385
• A single line ("Input: 2") appeared 78 times

Duplication Pattern

The duplication follows a specific pattern tied to streaming inside code blocks:

Line 90:  "w = 3              ← mermaid block cuts off mid-token
Line 91:  Input:       2      ← entire previous content re-renders from top
...
Line 132: flowchart TD        ← mermaid starts again
Line 138: "w = 3 ✓"]          ← gets further this time
...
Line 160: │  "                 ← ASCII box cuts off mid-character
Line 161: Input:       2      ← content re-renders AGAIN from top

Each time a new token arrives inside a code fence, the renderer appears to re-emit all accumulated content rather than appending the new token. Every re-render is captured as additional visible content in the message thread.

Impact

• Responses are completely unreadable due to duplicated content
• Copy-pasting from the thread gives 15-20x the actual content
• Any response with code blocks, ASCII art, or mermaid diagrams triggers this

Likely Root Cause

The streaming markdown renderer re-renders the entire accumulated content block (not just the delta) on each token arrival inside code fences. This re-rendered content is being appended to the visible output instead of replacing the previous render.