Desktop app (Code): ~30s delay before any response on every prompt — including first message of a fresh chat. CLI on same machine is ~5s.

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Environment

• Claude Code desktop app (Code / local-agent-mode), latest version as of 2026-05-22
• Bundled CLI: 2.1.150
• OS: Windows 11
• Hardware: Ryzen 7 9800X3D, RTX 5070 Ti (not a hardware limitation)
• Subscription: Claude Max

Summary
Every prompt in the desktop app's "Code" section sits with a blinking indicator / running timer for ~30–32s of dead time before Claude begins thinking or emits a single token. This happens even on the very first message of a brand-new chat, so it is not related to conversation length.

~32s/3min of dead time before anything happens, on every prompt.

Key evidence — it is NOT the API, account, network, or context size:

• CLI on the same machine, same account, same API returns a trivial prompt in ~5.4s end-to-end (including process startup + MCP init). The desktop app takes ~32s before it even begins.
• The delay did not change after clearing cache + restarting and after disconnecting most MCP connectors / plugin bundles. Context dropped from ~130k to ~107k tokens (11% of 1M), yet the delay stayed flat (30s → 32s) — so it is decoupled from context size.
• Context breakdown at the time of a 32s delay: Messages 95k (just the conversation), Skills 16k, active MCP tools 1.6k. A fresh chat is far smaller and is still slow.
• Network latency to the API is negligible.

Conclusion
The bottleneck is the Electron desktop renderer, not inference, account routing, or network. This matches the symptoms in #31643 (18+s desktop input lag, CLI instant on same machine) and #4896 (infinite React render loop in the skills system, all_skills.json 404 retry loop). Both were closed without a confirmed fix.

Related issues: #31643, #4896

Impact
This is a paid Max subscription. A consistent ~30s startup penalty on the vendor's own desktop app, when the bundled CLI is 6x faster on identical inputs, is a significant regression. Requesting this be reopened/investigated.

What Should Happen?

First token / thinking begins within a few seconds, as it does in the CLI.

Error Messages/Logs

Steps to Reproduce

Steps to reproduce

1. Open a fresh chat in the desktop app (Code).
2. Send any prompt, even trivial (e.g. "say OK").
3. Observe ~30s of no activity before the response starts.

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

1.8089.0

Claude Code Version

2.1.150

Platform

Other

Operating System

Windows

Terminal/Shell

Other

Additional Information

I'm using Claude's desktop application in the code section and I'm not using the API but the Max plan.