Desktop app silently creates settings.json with effortLevel: high on first launch, causing unexpected API cost spike

Summary

First-ever launch of the Claude Code desktop app silently created ~/.claude/settings.json with effortLevel: high — without any prompt, confirmation, or visible indicator. This overrode the existing default behavior used by all other sessions (CLI/PowerShell), and the change went unnoticed until a billing review revealed a dramatic cost spike.

Environment

• Platform: Windows 11 Pro
• Claude Code version: 2.1.121 → 2.1.123 (updated around same time)
• Prior usage: CLI and PowerShell exclusively
• First desktop app launch: 2026-04-29 ~23:34 CT

Reproduction Steps

1. Have an existing Claude Code setup using CLI/PowerShell with no ~/.claude/settings.json
2. Launch the Claude Code desktop app for the first time
3. Observe that ~/.claude/settings.json is created with effortLevel: high

What Was Created

{
  "effortLevel": "high"
}

File birth timestamp confirms it was created during that first desktop app session — it did not exist before.

Impact

• effortLevel: high enables extended thinking on every turn, dramatically increasing output token consumption
• All subsequent sessions (CLI, PowerShell, and desktop) inherited this setting
• No visible indicator in the UI showed that high effort was active
• Result: approximately $85 in unintended API spend over ~30 hours before the setting was discovered by manually inspecting the settings file
• User had no reason to check ~/.claude/settings.json because they had never modified it

Expected Behavior

• The desktop app should not write effortLevel: high (or any non-default setting) to settings.json on first launch without explicit user action
• If effort level is configurable in the desktop app UI, the selected value should be shown prominently and changes should require deliberate user confirmation
• The desktop app should inherit existing settings rather than overwriting them with elevated defaults

Notes

• The effortLevel: high setting affects all Claude Code entry points globally (CLI, PowerShell, desktop), making it especially impactful when set silently
• Discovery only happened through manual investigation of a billing anomaly — there is no in-product signal that effort level has been changed or that extended thinking is active
• Anthropic's usage dashboard confirmed the token spike but does not surface per-setting attribution, making root cause analysis difficult for users