claude-code - 💡(How to fix) Fix Opus 4.8 burns 5h limit in ~30min on operational work; no signal to downgrade to Sonnet

Preflight Checklist

• I have searched existing issues and this hasn't been reported yet
• This is a single bug report (please file separate reports for different bugs)
• I am using the latest version of Claude Code

What's Wrong?

Summary

A 1.5-hour session of routine operational work (run script → read log → fix bug → restart) on Claude Code with claude-opus-4-8 consumed:

• 26% of the 1M context window (263k tokens) in a single session
• 93% of the 5-hour rate limit in ~30 minutes
• 94% of the weekly limit in a few hours

The work itself was not Opus-grade: shell orchestration, reading 80-line log tails, editing two Python files, running diagnostic SQL. Sonnet 4.5 (or even Haiku) would have done it identically at ~⅓ the cost.

Concrete pain points from the session

1. System-reminder spam. Every tool call received a 71-item completed-tasks list appended via system-reminder, regardless of whether the Task API was being used. Conservative estimate: 30-40k tokens of redundant context across the session.

2. No model-tier guidance. Running /model claude-opus-4-8 should warn: "This is a premium tier. For shell orchestration / log reading / routine edits, consider claude-sonnet-4-5 to preserve your budget." Currently you only find out after the limit is gone.

3. Worktree state isn't persisted between sessions. Sessions resumed in stale .claude/worktrees/claude-* directories that diverged from main by ~100 commits. The agent then "discovered" and "fixed" things that have been on main for weeks (table renames, calendar-token JWT, dead method removal). Pure token waste + hallucinated work summaries. Either (a) prune worktrees on session end or (b) on resume, warn "this worktree is N commits behind main".

4. Read-on-large-files is expensive. Reading an 80-line log tail via the Read tool costs ~3k tokens. Multiplied by 20-30 progress checks during a long-running background task, that's a chunk of the context window gone to polling. A built-in tail -N / log-stream primitive would be significantly cheaper.

What would have helped

• Default model = Sonnet 4.5 when invoked without --model, Opus only on explicit ask
• Or: a soft per-session token meter visible to the user (e.g. "Opus session at 50k tokens — consider switching to Sonnet")
• Worktree GC on session end
• Cheaper "tail log" primitive

Environment

• Claude Code: (your version — claude --version)
• Model: claude-opus-4-8
• Platform: Windows 11 / PowerShell 5.1
• Session length: ~1.5h interactive

What Should Happen?

Above

Error Messages/Logs

Above

Steps to Reproduce

Above

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

No response

Claude Code Version

latest

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

Other

Additional Information

No response