Fear of updating: When 'vibe coding' meets production stability

@bcherny - Your recent interview highlights the irony here

Hi Boris,

In your Peterman Pod interview (Dec 15, 2025), you said:

"You want maintainable code sometimes. You want to be very thoughtful about every line sometimes."

And that 80-90% of Claude Code is now written by Claude Code itself.

The Irony

The .claude.json architecture violates every principle you advocate for:

• No separation of concerns (config + history + state + changelog in ONE file)
• No atomic writes (corruption on concurrent access)
• No data lifecycle management (files growing to 65MB+)
• 30+ second startup delays for long-running projects

Beyond .claude.json — Systemic Stability Issues

This isn't just about config files. Recent versions have introduced critical stability problems:

• Complete system freezes — Claude Code sessions now regularly hang the entire computer (see #14674)
• Only hard reboot helps — no graceful recovery possible
• Memory leaks? — unclear what's causing these freezes, no debugging info available
• Multiple IDE instances crash — running 2+ GoLand instances with Claude Code causes total system freeze (#14674)
• Forcing users away — I'm increasingly switching to my own Go-based agent using Anthropic API directly, which is significantly more expensive but at least doesn't crash my machine

Fresh Bugs in v2.0.73 (just released!)

Here's what "thoughtful about every line" looks like in practice:

• #14674 — Complete system freeze with 2+ IDE instances (requires hard reboot!)
• #14641 — Markdown tables render incorrectly (broken layout, truncated Cyrillic text)
• #14637 — API responses truncating mid-response
• #14632 — Large blank areas where text should display (rendering broken)

All are Windows bugs in recent releases. All are regressions.

The Real Problem

Every time I need to end a session or update to a new Claude Code version, I'm genuinely afraid. Not because of the new features, but because I have no idea how many new bugs were introduced — whether by humans or by the very "vibe coding" approach you describe.

If 80-90% of your codebase is written by AI, and you admit vibe coding is only good for "throwaway code and prototypes" — who is being "very thoughtful about every line" before releases?

The Business Impact

When your tool is unstable:

• Users pay MORE to use raw API instead of Claude Code
• Trust erodes with each forced reboot
• Community contributions go ignored for 8+ months

Context

See detailed technical analysis of .claude.json issues: https://github.com/anthropics/claude-code/issues/7243

Simple Ask

Is stability on the roadmap? A one-word "yes" or "no" would be more respectful than silence.

---

Sources:

• Peterman Pod interview (Dec 2025)
• Business Insider article
• Related: #14674 #7243 #5024 #1449 #6394 #8353 #14641 #14637 #14632