---
version: 1.9.4
date: 2026-04-07
title: Customizable Magic Prompts
---

## App

### Improvements
- Chat/Performance: reduced streaming render churn across the app.
- Chat/Tool Output: LSP diagnostics now render directly in tool output (thanks to @yulia-ivashko).
- Usage: refreshed MiniMax CN coding-plan quota data (thanks to @nzlov).

### Fixes
- Settings/Magic Prompts: added a dedicated Magic Prompts page with editable templates for commit/PR generation, PR and issue reviews, failed-check/comment analysis, and merge/cherry-pick conflict resolution.
- Chat/Scrolling: fixed jumpy follow behavior and restored stable bottom-resume/live-compaction updates.
- Reliability/Streaming: improved reconnect, retry, and directory-aware event routing to reduce stuck session/subagent states after transient disconnects (thanks to @jwcrystal, @daveotero).
- Models: added defensive handling for missing model pricing/capability metadata (thanks to @Chi-square-test).
- Desktop/Performance: removed costly window translucency and reduced duplicate notification triggers for a cooler, less noisy desktop experience.
- Startup/Remote: restored remote provider startup behavior and tightened host/port detection to reduce false startup failures.

## VS Code

### New
- Reliability/Streaming: added loading timeouts, automatic SSE reconnect, and message retry behavior (thanks to @jwcrystal).

### Improvements
- Reliability/Windows: normalized workspace path handling in SSE event lookup to keep live session updates working consistently on Windows (thanks to @widipa).
- Chat/Performance: reduced streaming re-render fanout and status-row churn for smooth long responses in the editor panel.
- Chat/Tool Output: LSP diagnostics now render directly in tool output (thanks to @yulia-ivashko).

### Fixes
- Sessions/Streaming: fixed directory-aware event routing and post-reconnect session resync (thanks to @daveotero, @jwcrystal).
- Chat/Scrolling: fixed scroll jumps and stabilized follow-to-latest behavior.
- Models: added defensive fallbacks for missing model cost/capability metadata (thanks to @Chi-square-test).