Session recovery wastes millions of tokens — auto-save before model switches, allow subscription on external platforms

Resolved 💬 3 comments Opened Mar 27, 2026 by FerniFernDev Closed Mar 30, 2026

Complaint: Session Recovery Cost is Unacceptable — Subscription Should Work on OpenClaw

The Situation

I'm building a complex, multi-session project (LuminSelf — a full-stack social media platform). Claude Code is my primary development tool. I use it through a subscription (not API keys).

During an active session where significant work had been done (production deployment setup, Backblaze B2 configuration, Railway environment variables, domain purchases, and more), the session was lost when switching models. None of the progress was saved before the session ended.

The Cost of Recovery

The next session — this one — was spent entirely on recovery:

  • Duration: 1 hour 50 minutes
  • Tokens consumed: ~39,000,000+
  • Cost: $13+ at subscription rates
  • Productive work accomplished: ZERO

All 39 million tokens went to:

  1. Figuring out what was done in the lost session
  2. Auditing the codebase to verify what exists
  3. Checking Railway dashboard to verify what was configured
  4. Re-saving progress to memory files
  5. Fixing mistakes made during recovery (wrong screenshot tool, wrong minimize approach, etc.)

Not a single line of code was written. Not a single feature was built. Not a single bug was fixed.

The Broader Problem

1. Session persistence is fragile

Switching models, hitting a context limit, or any interruption can kill a session. When that happens, ALL unsaved context is lost. The memory system helps, but only if progress was saved BEFORE the crash. If the session dies mid-work, everything since the last save is gone.

2. Recovery is absurdly expensive

Recovering context requires:

  • Reading dozens of memory files
  • Auditing the actual codebase (thousands of files)
  • Verifying external services (Railway, Backblaze, domains)
  • Taking screenshots to check browser state
  • Re-establishing all the context that was in the previous conversation

This easily consumes millions of tokens — and produces nothing of value.

3. API key costs would be devastating

If I were using Claude through API keys (e.g., on OpenClaw for my AI assistant Alexis), this single recovery session could cost $50-100+ in API fees. That's $50-100 to get back to where I already was, with zero forward progress.

The subscription model absorbs this cost. But the subscription doesn't work on OpenClaw. This means:

  • On Claude Code (subscription): Recovery is "free" but wastes 2 hours of my time
  • On OpenClaw (API keys): Recovery would cost real money AND waste 2 hours

4. The ask

  • Allow subscriptions to be used on OpenClaw — or provide a reasonable API credit system for subscribers. Forcing API key pricing for the same model that's available on subscription is unfair, especially when so many tokens are wasted on non-productive recovery.
  • Auto-save session state before model switches — if the user switches models, the current session should automatically save a handoff note before dying.
  • Better session persistence — the memory system is good but manual. Automatic checkpointing would prevent entire sessions from being lost.

Summary

I lost a session. It took 2 hours and 39 million tokens just to figure out where I was. No work was done. If I were on API keys, this would have cost real money for zero value. The subscription should work across platforms, and session persistence should be automatic, not manual.

---
Submitted by a user who has spent 37 sessions and hundreds of hours building with Claude Code. The tool is powerful. The persistence model is not.

View original on GitHub ↗

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