[Bug] Reasoning block renders raw Markdown / jumps while streaming in Desktop

Description

When the model is generating reasoning/thinking in OpenCode Desktop, the reasoning block is rendered through the Markdown streaming projection. While the message is incomplete the live projection splits the text at the last complete token and heals the trailing tail into plain text. Incomplete Markdown syntax (headings, emphasis, lists, reference-style links) renders as raw text and the layout jumps on every streaming update. Once the message completes the content is re-rendered as full Markdown and looks correct, but the in-flight state is broken/noisy.

OpenCode version

1.18.31 (Desktop, Windows 11)

Steps to reproduce

1. Open a session in OpenCode Desktop with a model that emits reasoning (e.g. a provider that streams a reasoning part).
2. Start a new assistant turn that produces a longer reasoning block containing headings / lists / emphasis.
3. Watch the reasoning block while it streams: the trailing text alternates between plain and raw Markdown artifacts and the block jumps.

Screenshot and/or share link

N/A

Operating System

Windows 11

Terminal

N/A