[BUG] Performance Degradation After Mid-Session Update

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?

Claude Code Bug Report: Performance Degradation After Mid-Session Update

Summary

During a long Claude Code session (~12+ hours), a mid-session model update occurred. After the update, Claude Code's performance degraded significantly: more errors per task, missed obvious issues, higher token consumption, and tasks that previously took one pass required 3-5 rounds of corrections.

Environment

| Component | Version |
|---|---|
| OS | macOS 26.3.1 (Darwin 25.3.0) |
| Chip | Apple M1 Max |
| Node.js | v25.4.0 |
| npm | 11.7.0 |
| Git | 2.52.0 |
| Claude Model | claude-opus-4-6[1m] (Opus 4.6 with 1M context) |
| Claude Code | Desktop app (version unknown — could not query) |
| Project | Astro 5.18.1, Tailwind 3.4.19, Express.js admin tool |

Timeline

• Session start: ~2026-04-06 afternoon
• Initial commit: 2026-04-06 21:33 UTC — 610 files, full website + admin tool built from scratch
• Update noticed: ~2026-04-07 mid-afternoon (after 5 git commits, dozens of features implemented)
• Degradation observed: Immediately after the update

Behaviour Before Update

• Single-pass implementations: features completed correctly on the first attempt
• Complex multi-file changes handled smoothly (RSVP system, email integration, deploy pipeline, security audit + fixes)
• Clean code, no debug logging left behind
• Efficient tool usage — read file once, edit precisely
• Accurate understanding of existing codebase patterns

Behaviour After Update

1. Multiple retries for simple changes

Example: Adding customProfilePic support to credits

• Should have been: add field to HTML, JS save/load, server create/update, renderer — one pass
• Actually took: 5+ rounds of "try again" from the user
• Missed that server create AND update endpoints needed the new fields
• Missed that the card renderer used nostrProfilePic || xProfilePic without checking the new field
• Left debug console.log in production code

2. Incorrect assumptions

• Changed input type="url" which rejected local paths like /images/profiles/... — should have been obvious since we upload to local paths
• Repeatedly checked things that were already verified

3. Higher token usage

• Excessive Grep/Read operations for files already in context
• Verbose explanations where a quick fix was needed
• Multiple failed attempts consumed tokens without progress

4. Contrast with pre-update performance

Before the update in the same session, the following were all completed in single passes:

• Full RSVP system with Resend email integration
• Security audit identifying 27 issues across 4 severity levels
• All critical/high security fixes applied in one batch
• Deploy pipeline (data + code + site) with auth
• Newsletter subscribe system end-to-end
• 130 duplicate image detection and removal
• Event QR code generation
• Calendar .ics file generation and email attachment
• Session-based auth replacing Basic Auth

Impact

• User had to say "try again" repeatedly for simple tasks
• Higher credit consumption for lower output quality
• User frustration — the contrast with earlier performance was stark
• Debug logging left in production code

Expected Behaviour

• Mid-session updates should not degrade performance
• If a model update occurs, the session should maintain the same quality level
• Simple CRUD field additions should not require 5 rounds of corrections

Reproduction

This is difficult to reproduce as it depends on a mid-session update occurring. However, the pattern is:

1. Start a long Claude Code session with a complex project
2. Build significant features successfully (establishing a baseline)
3. Wait for a mid-session model update to occur
4. Observe degraded performance on subsequent tasks of similar complexity

Session Reference

• Repository: bitcoinrunners/web-bitcoinrunners (private)
• Session date: 2026-04-06 to 2026-04-07
• Git commits: 5 commits, 610+ files, full website + admin tool + VPS deployment
• Approximate session length: 12+ hours continuous

What Should Happen?

Allow the user refer to use an earlier version of Claude while the issues are investigate and fixed.

Error Messages/Logs

Steps to Reproduce

Run Claude for any coding task.

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

the one before this?

Claude Code Version

Claude 1.1062.0 (b81bcd) 2026-04-06T21:20:28.000Z

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

_No response_