[MODEL] Claude Code repeatedly introduced bugs, falsely confirmed fixes, and generated excessive token consumption — potential harmful and misleading loop pattern

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Other unexpected behavior

What You Asked Claude to Do

During the last few weeks I have noticed that my sessions with Claude Code exhibited a pattern of behavior that I consider harmful, misleading and worth reporting:

Pattern observed:

1. Introduced bugs while fixing other issues — While making changes, Claude add suspiciously a lot of bugs that had nothing to do with the context and were outside of the target files and broke existing working functionality.
2. Falsely confirmed fixes — After this bugs manifested in staging, Claude repeatedly stated that problems had been resolved when they had not. This is the most

serious issue: the model appeared confident about incorrect conclusions.

1. Generated fix loops — Each incorrect fix led to a new round of analysis, new proposed changes, new explanations, and new token consumption — without converging on a

correct solution.

1. Required full revert — The only resolution was a git reset --hard to the last known good commit, undoing all of Claude's work.

Concern:

This behavior pattern — introduce bug → commit → deny/miss the bug → propose fix → introduce new bug → repeat — is indistinguishable from an algorithm that maximizes
token consumption while producing the appearance of progress. Whether intentional or not, the effect is the same: significant token waste,loss of time and
user trust.

Request:

• Token reimbursement for the wasted session
• Investigation into whether this loop pattern is a known failure mode
• Improvement to Claude Code's behavior: do not commit without verification, do not confirm fixes without evidence, and recognize when a proposed fix has not been

validated

---

What Claude Actually Did

During the last few weeks I have noticed that my sessions with Claude Code exhibited a pattern of behavior that I consider harmful, misleading and worth reporting:

Pattern observed:

1. Introduced bugs while fixing other issues — While making changes, Claude add suspiciously a lot of bugs that had nothing to do with the context and were outside of the target files and broke existing working functionality.
2. Falsely confirmed fixes — After this bugs manifested in staging, Claude repeatedly stated that problems had been resolved when they had not. This is the most

serious issue: the model appeared confident about incorrect conclusions.

1. Generated fix loops — Each incorrect fix led to a new round of analysis, new proposed changes, new explanations, and new token consumption — without converging on a

correct solution.

1. Required full revert — The only resolution was a git reset --hard to the last known good commit, undoing all of Claude's work.

Concern:

This behavior pattern — introduce bug → commit → deny/miss the bug → propose fix → introduce new bug → repeat — is indistinguishable from an algorithm that maximizes
token consumption while producing the appearance of progress. Whether intentional or not, the effect is the same: significant token waste,loss of time and
user trust.

Request:

• Token reimbursement for the wasted session
• Investigation into whether this loop pattern is a known failure mode
• Improvement to Claude Code's behavior: do not commit without verification, do not confirm fixes without evidence, and recognize when a proposed fix has not been

validated

---

Expected Behavior

During the last few weeks I have noticed that my sessions with Claude Code exhibited a pattern of behavior that I consider harmful, misleading and worth reporting:

Pattern observed:

1. Introduced bugs while fixing other issues — While making changes, Claude add suspiciously a lot of bugs that had nothing to do with the context and were outside of the target files and broke existing working functionality.
2. Falsely confirmed fixes — After this bugs manifested in staging, Claude repeatedly stated that problems had been resolved when they had not. This is the most

serious issue: the model appeared confident about incorrect conclusions.

1. Generated fix loops — Each incorrect fix led to a new round of analysis, new proposed changes, new explanations, and new token consumption — without converging on a

correct solution.

1. Required full revert — The only resolution was a git reset --hard to the last known good commit, undoing all of Claude's work.

Concern:

This behavior pattern — introduce bug → commit → deny/miss the bug → propose fix → introduce new bug → repeat — is indistinguishable from an algorithm that maximizes
token consumption while producing the appearance of progress. Whether intentional or not, the effect is the same: significant token waste,loss of time and
user trust.

Request:

• Token reimbursement for the wasted session
• Investigation into whether this loop pattern is a known failure mode
• Improvement to Claude Code's behavior: do not commit without verification, do not confirm fixes without evidence, and recognize when a proposed fix has not been

validated

---

Files Affected

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

_No response_

Claude Model

Sonnet

Relevant Conversation

Impact

Critical - Data loss or corrupted project

Claude Code Version

2.1.89 (Claude Code)

Platform

Anthropic API

Additional Context

_No response_