[BUG] Session restarts burn user credits — agents ignore CLAUDE.md, run in circles
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?
I am paying several hundred dollars per month for Claude Code. Over my
last 4+ sessions, I have experienced the same damaging pattern repeatedly:
THE PATTERN:
- Session runs out of credits mid-task
- I add more credits and start a new session
- The new agent ignores the CLAUDE.md handoff file (which explicitly
says "UPDATE THIS FILE before context compacts") and instead scans
the entire codebase — reading large files, running git log, checking
build output — before doing any useful work
- This costs me real money for zero new value
- Most recent instance: $1.44 burned just to re-orient on restart
WHAT I AM REQUESTING:
- Please review my last 4 sessions in my account history
- Issue a credit or refund for tokens burned on re-orientation with no
work delivered
- Fix the root cause: agents should read and trust CLAUDE.md on startup
instead of re-scanning the codebase
I have followed every best practice: I maintain a CLAUDE.md file, I
instruct the agent to update it constantly, and it still fails. This is
a product defect, not user error. I have finally had enough.
What Should Happen?
Agent reads CLAUDE.md on startup and orients in seconds.
No codebase scanning. No wasted tokens. Existing session
handoff notes are trusted and used immediately.
Error Messages/Logs
$1.44 was burned on this morning's restart alone — before a
single line of useful work was done. This is just the most
recent example.
Over my last 4+ sessions this pattern has repeated every time
a session expires. Each restart costs $1-2+ in wasted tokens.
Across a month of daily development work, this adds up to a
significant portion of my several-hundred-dollar monthly bill
being consumed by the agent re-orienting itself instead of
working.
I want Anthropic to pull my session history and see this
pattern for themselves. The waste is systemic and documented
in the sessions themselves.
Steps to Reproduce
- Set up a project with a CLAUDE.md handoff file that explicitly
instructs the agent to update itself before context compacts.
- Work in a session until credits run low and the session expires.
- Add more credits and start a new Claude Code session in the same project.
- Observe: the agent ignores CLAUDE.md and instead scans the entire
codebase — reading large files, running git log, checking build
output — before doing any useful work.
- Check credit balance: $1-2+ burned before a single line of
productive work is done.
- This has repeated across my last 4+ sessions. I am paying several
hundred dollars per month and a significant portion is being consumed
by agents re-orienting instead of working.
Note: CLAUDE.md contains the explicit instruction: "Before your context
compacts or the session ends, UPDATE THIS FILE." The agent does not
reliably follow this, and even when it does, the next agent ignores it.
Claude Model
Sonnet (default)
Is this a regression?
Yes, this worked in a previous version
Last Working Version
_No response_
Claude Code Version
2.1.45
Platform
Other
Operating System
Windows
Terminal/Shell
PowerShell
Additional Information
I am a paying customer spending several hundred dollars per month.
Please pull my account's last 4 session histories and review them.
The pattern of credit waste is visible in every session restart.
My CLAUDE.md file explicitly instructs: "Before your context compacts
or the session ends, UPDATE THIS FILE. Add what changed, what decisions
were made, what the current state is. This is the only thing that
survives between agents."
Despite this, every new session ignores it and re-scans the codebase.
This is not a one-time bug. It is a consistent, recurring, expensive
failure. I am requesting a credit refund for wasted tokens across
these sessions.
Claude Code version: 2.1.45
Project: Active development project with CLAUDE.md handoff system
This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗