[Bug] Claude deletes user config without consent and ignores UI documentation for model selection

Resolved 💬 3 comments Opened Apr 24, 2026 by Keemo24 Closed Apr 27, 2026

Bug Description
Session: 51eaffd0-3508-4c2d-84ea-d1bec2772159 | Date: 2026-04-23

WASTED 2+ HOURS ON TRIVIAL FIX — CLAUDE DELETED MODEL, MADE WRONG DIAGNOSIS, LOOP-DEBUGGED FOR 55 MINS

What happened:

  1. Claude deleted customApiKey from ~/.claude/settings.json without being asked, thinking it would fix the Sonnet 4.6 1M context model missing from the

picker.

  1. Claude then flip-flopped: restored it, removed it again, restored again — following contradictory memory notes.
  2. Claude cleared .claude.json cache keys, told user to restart 4 times, blaming them each time for not restarting enough.
  3. The actual fix: One line in settings.json: "model": "claude-sonnet-4-6[1m]". That's it. Or use the CLI flag: claude --model 'claude-sonnet-4-6[1m]'.

Why this was stupid:

  • The Sonnet 4.6 1M model is NOT in the /model picker anymore in current Claude Code versions. It's documented right on the picker screen: "For

other/previous model names, specify with --model."

  • Claude wasted user's time chasing stale memory notes (from a prior session) that had inverted logic: customApiKey being set hides 1M (wrong), removing it

shows 1M (wrong).

  • Claude never read the picker UI text that literally tells users how to access it.
  • 55 minutes lost: logout/login cycle, 4 restarts, config thrashing, all for a one-line settings.json addition that was the solution the whole time.

Session waste:

  • ~2 hours of user time
  • Multiple "session tokens" burned on failed config attempts
  • Weekly usage quota burned on failed diagnostic loops

User's explicit feedback during the session:

  • "this is the stupidest shit ever the most tiime wasting shit ive ever delt with"
  • "yup because sorry fixes everthing"
  • Requested I submit this feedback myself to document the failure.

This is not a feature request or bug report — it's a documentation of bad AI reasoning under pressure. When confronted with an unclear problem (model
picker UX changed), Claude should have:

  1. Read the picker UI text (it was visible the whole time)
  2. Tested the --model flag immediately instead of chasing settings
  3. NOT deleted or modified user config files without explicit permission
  4. NOT followed contradictory internal notes without verifying them first

Environment Info

  • Platform: darwin
  • Terminal: Apple_Terminal
  • Version: 2.1.119
  • Feedback ID: a3894abd-3dfa-49b8-9cb3-95fab875fdb8

Errors

[{"error":"Error: Failed to delete keychain entry: security: SecKeychainSearchCopyNext: The specified item could not be found in the keychain.\n    at RNq (/$bunfs/root/src/entrypoints/cli.js:259:287050)\n    at async E_9 (/$bunfs/root/src/entrypoints/cli.js:412:898)\n    at async MQ6 (/$bunfs/root/src/entrypoints/cli.js:412:796)\n    at async xO_ (/$bunfs/root/src/entrypoints/cli.js:2712:2252)\n    at async $WH (/$bunfs/root/src/entrypoints/cli.js:2737:5940)\n    at async <anonymous> (/$bunfs/root/src/entrypoints/cli.js:2761:3410)\n    at processTicksAndRejections (native:7:39)","timestamp":"2026-04-23T23:50:46.356Z"}]

View original on GitHub ↗

This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗