VS Code extension: streamed chat text drops individual characters at render/copy layer

Summary

In the Kimi Code VS Code extension chat panel, individual characters occasionally go missing from the assistant's rendered message text. The underlying model output is intact — verified against the session wire log — so the loss happens at the rendering (or copy-from-panel) layer.

Evidence

Session wire log (~/.kimi-code/sessions/.../agents/main/wire.jsonl) contains the full correct text, e.g.:

• В режиме «спрашивать когда нужно» — displayed/copied as В режим «спрашивать когда нужно» (missing е)
• (разрешить — она закроет...) — displayed as (разреш — она закроет...) (missing ить)
• /yolo — displayed as /olo (missing y) — especially harmful for slash commands
• «продолжай», и она пойдёт — displayed without the closing quote and comma

The dropped characters are single, scattered, and look consistent with a redraw/line-wrap race during streaming in the webview.

Environment

• Extension: moonshot-ai.kimi-code 0.7.5 (darwin-arm64), latest on the marketplace as of 2026-09-05
• VS Code on macOS (Apple Silicon)
• Model: Kimi K2

Impact

Low for prose, but dangerous for commands: a missing character in a slash command or shell snippet silently breaks copy-paste execution.

Suggestion

If the webview re-renders streamed markdown on line-wrap boundaries, consider diffing against the accumulated raw text (source of truth) rather than incremental DOM patches, or re-sync the final message text once streaming completes.

Start with the VS Code extension's chat-panel webview and its streamed markdown rendering or copy-from-panel path. Reproduce streaming text loss while comparing the rendered or copied message with the session wire log, then verify that complete text, including slash commands and punctuation, remains intact after streaming finishes.