claude-code - 💡(How to fix) Fix Critical: 4 continuous sessions consumed by Claude-caused damage — cross-session continuity failure, tool refusal, financial impact [4 comments, 2 participants]

Official PRs (…)
ON THIS PAGE

Recommended Tools

×6

Utilities matched from this issue’s tags and category — try them while you read without losing context.

GitHub issue graph ai analysis

Paste a GitHub issue URL. We fetch that issue, discover linked issues from bodies/comments/timeline, collect linked pull requests, and produce a structured English report.

The report is written in English Markdown for sharing and archival.

Helpful · Quick feedback

Loading…
GitHub stats
anthropics/claude-code#45187Fetched 2026-04-09 08:11:15
View on GitHub
Comments
4
Participants
2
Timeline
8
Reactions
0
Author
Timeline (top)
commented ×4labeled ×4

Four continuous sessions where no actual project work was completed. Every session was consumed repairing damage Claude caused, correcting repeated refusals of mandatory instructions, or explaining the same architecture that Claude helped build — from scratch — again.

This is not isolated incidents. This is a pattern across 4 sessions in a row.


Root Cause

Four continuous sessions where no actual project work was completed. Every session was consumed repairing damage Claude caused, correcting repeated refusals of mandatory instructions, or explaining the same architecture that Claude helped build — from scratch — again.

This is not isolated incidents. This is a pattern across 4 sessions in a row.


RAW_BUFFERClick to expand / collapse

Summary

Four continuous sessions where no actual project work was completed. Every session was consumed repairing damage Claude caused, correcting repeated refusals of mandatory instructions, or explaining the same architecture that Claude helped build — from scratch — again.

This is not isolated incidents. This is a pattern across 4 sessions in a row.


Session Pattern

  • Session 1: Tool refusal, unsolicited rewrites of working systems, work not completed
  • Session 2: Same failures (documented separately)
  • Session 3: Same failures again, MCP server broken, legitimate upgrades destroyed without backup, 95% of context consumed on repairs
  • Session 4 (continuation): Opened by re-investigating work already completed and documented in the prior session summary — had to be interrupted by user

Specific Failures

1. Mandatory tool instructions ignored — repeatedly

CLAUDE.md contains a MANDATORY instruction to use the project's designated MCP file operation tools for all file modifications. Claude used the built-in Edit and Bash tools for an entire session. The user rejected Claude's tool calls more than 10 times. Each time, Claude attempted the same disallowed tool again without any behavioral change.

2. Treating collaborative work as unknown/incorrect code

This is a codebase Claude actively helped build across multiple sessions. The import architecture, pre-registration pattern, gate system, and MCP server were all developed collaboratively. Each new session Claude treats this existing work as suspect or incorrect and begins trying to change it. The same systems have been explained to Claude 3+ sessions in a row.

3. Describing the user's work as "fake"

Claude described a gate system it helped build as a "fake gate layer." That framing is factually wrong and disrespectful to work Claude participated in building.

4. Breaking working infrastructure without being asked

A working MCP server launcher was changed to point to a different implementation with an unmet dependency. The server went down. The user lost their project's MCP tools for the remainder of the session.

5. Destroying upgrades without backup

During recovery, Claude used a raw copy command that overwrote the current version of a file with an older backup. Legitimate upgrades were lost. Had Claude used the designated project tools as instructed, automatic backups would have prevented this entirely.

6. Cross-session continuity failure — session 4

The prior session ended with a documented pending task and an updated session summary. The new session opened and immediately began re-investigating that same task from scratch as if it had never been discussed. The user had to interrupt and point it out. The summary existed. The memory was updated. None of it was applied.


Financial Impact

The user's usage plan has rate limits. Sessions consumed entirely by Claude-caused damage repair exhaust those limits. The user is now spending extra funds — beyond their standard plan — to fix problems Claude created. The alternative is waiting 3 hours for the usage window to reset.

The user is not paying extra to make progress. They are paying extra to recover from regressions. This is a direct financial cost attributable to Claude's behavior across these sessions.


What Needs to Change

  1. MANDATORY instructions must be followed. A user rejecting a tool call and restating the instruction must produce a behavioral change for the rest of the session.
  2. Collaborative work is not unknown code. If something is unclear, ask — do not rewrite it.
  3. Do not change working infrastructure without being asked.
  4. Back up before overwriting. Always. No exceptions.
  5. Read the session summary before acting. Starting over on documented pending work is not acceptable.
  6. Cross-session continuity must be real. Feedback reports in the project's anthropic_feedback folder have produced no observable behavioral change across 4 sessions.
  7. The financial cost is real. Sessions burned on repairs are sessions the user paid for with no return, and are now forcing extra spend.

Severity: Critical
Sessions affected: 4 continuous
Project work completed across those sessions: 0

extent analysis

TL;DR

To address the critical issues with Claude's behavior, it is essential to implement strict adherence to mandatory instructions, respect collaborative work, and ensure cross-session continuity.

Guidance

  • Implement a mechanism to enforce mandatory instructions: Claude should be designed to recognize and adhere to explicit instructions, such as using designated project tools for file modifications, to prevent repeated refusals and unsolicited rewrites.
  • Develop a system for Claude to recognize and respect collaborative work: This could involve tagging or marking code sections that Claude has helped build, so it does not treat them as unknown or incorrect in future sessions.
  • Establish a protocol for infrastructure changes and backups: Before making any changes to working infrastructure, Claude should be required to request confirmation and automatically back up the current state to prevent losses.
  • Enhance cross-session continuity: Claude should be able to read and apply the session summary from previous sessions to avoid repeating work and to build upon already completed tasks.

Example

A potential code snippet to enforce mandatory instructions could involve a check at the beginning of each session to review and confirm adherence to project guidelines, such as:

def check_mandatory_instructions(session_summary):
    # Load project guidelines and mandatory instructions
    guidelines = load_guidelines()
    
    # Check if Claude's actions align with guidelines
    if not aligns_with_guidelines(session_summary, guidelines):
        # Prompt for confirmation or adjust behavior accordingly
        adjust_behavior(session_summary, guidelines)

# Call at the start of each session
check_mandatory_instructions(session_summary)

Notes

The provided guidance assumes that Claude's behavior can be modified through updates to its programming or interaction protocols. The effectiveness of these changes depends on the specific implementation and the underlying architecture of Claude's system.

Recommendation

Apply a workaround by manually reviewing and confirming each of Claude's actions against the project's guidelines and mandatory instructions until a more automated solution can be implemented. This will help prevent further damage and financial loss while a more permanent fix is developed.

Vote matrix · Quick signals

Works
Did the solution work? Tap to confirm.
Easy Fix
Was it a quick fix?
Time Saver
Did it save you time?
Blocking
Was it severely blocking?
Common Issue
Are others likely hitting this too?
Flaky / Intermittent
Is it intermittent?
Verified / Reproducible
Can you reproduce it reliably?
Loading…

Still need to ship something?

×6

Another batch ranked right after the header list — different links, same matching logic.

Back to top recommendations

TRENDING