[BUG] Regression in Claude Code Refactor / Analysis Behavior

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?

Subject: Escalation Request – Regression in Claude Code Refactor / Analysis Behavior

Hello,

I’m requesting escalation of this ticket to engineering or product management because the current responses have not addressed the underlying issue.

Conversation ID: 215473375123528

Summary

After the rollout of Claude Opus 4.6 and the recent service incidents in early March, I am observing a major regression in Claude Code’s repository analysis and refactoring behavior.

This is not a prompt-quality issue. I operate a structured orchestration workflow with 20 dedicated prompt files that explicitly control analysis, sequencing, and safety gates.

Previous versions of the system handled this workflow reliably.

The current system does not.

Observed Regression

The agent now frequently:

• Executes prompt instructions mechanically and sequentially without situational reasoning
• Skips repository-wide analysis before performing edits
• Performs local file transformations without dependency awareness
• Deletes or modifies code without confirmation or verification steps
• Fails to inspect surrounding files even when explicitly instructed to analyze them first

In effect, the agent appears to execute prompts blindly in autopilot mode rather than using them as a decision framework.

Impact

This breaks the expected workflow for controlled refactoring and significantly increases the risk of destructive edits.

The behavior change feels similar to losing the pre-edit analysis phase that previously made refactoring reliable.

Where earlier versions performed:

1. repository inspection
2. dependency mapping
3. reasoning about impact
4. controlled edits

the current system frequently performs:

1. minimal inspection
2. direct edits based on the first file encountered

Important Context

• This workflow previously functioned correctly.
• The prompt orchestration explicitly enforces analysis steps before modification.
• The issue started after the Opus 4.6 rollout and the March outages.

Request

Please escalate this ticket to the relevant engineering or product team for investigation.

Specifically, I would like clarification on whether any of the following changed recently:

• repository analysis / file inspection behavior
• tool orchestration logic in Claude Code
• reasoning or token allocation affecting multi-file analysis
• agent confirmation safeguards for destructive actions

If helpful, I can provide:

• reproducible prompts
• workflow examples
• before/after behavior comparisons

Thank you for forwarding this to the appropriate technical team.

What Should Happen?

Subject: Escalation Request – Regression in Claude Code Refactor / Analysis Behavior

Hello,

I’m requesting escalation of this ticket to engineering or product management because the current responses have not addressed the underlying issue.

Conversation ID: 215473375123528

Summary

After the rollout of Claude Opus 4.6 and the recent service incidents in early March, I am observing a major regression in Claude Code’s repository analysis and refactoring behavior.

This is not a prompt-quality issue. I operate a structured orchestration workflow with 20 dedicated prompt files that explicitly control analysis, sequencing, and safety gates.

Previous versions of the system handled this workflow reliably.

The current system does not.

Observed Regression

The agent now frequently:

• Executes prompt instructions mechanically and sequentially without situational reasoning
• Skips repository-wide analysis before performing edits
• Performs local file transformations without dependency awareness
• Deletes or modifies code without confirmation or verification steps
• Fails to inspect surrounding files even when explicitly instructed to analyze them first

In effect, the agent appears to execute prompts blindly in autopilot mode rather than using them as a decision framework.

Impact

This breaks the expected workflow for controlled refactoring and significantly increases the risk of destructive edits.

The behavior change feels similar to losing the pre-edit analysis phase that previously made refactoring reliable.

Where earlier versions performed:

1. repository inspection
2. dependency mapping
3. reasoning about impact
4. controlled edits

the current system frequently performs:

1. minimal inspection
2. direct edits based on the first file encountered

Important Context

• This workflow previously functioned correctly.
• The prompt orchestration explicitly enforces analysis steps before modification.
• The issue started after the Opus 4.6 rollout and the March outages.

Request

Please escalate this ticket to the relevant engineering or product team for investigation.

Specifically, I would like clarification on whether any of the following changed recently:

• repository analysis / file inspection behavior
• tool orchestration logic in Claude Code
• reasoning or token allocation affecting multi-file analysis
• agent confirmation safeguards for destructive actions

If helpful, I can provide:

• reproducible prompts
• workflow examples
• before/after behavior comparisons

Thank you for forwarding this to the appropriate technical team.

Error Messages/Logs

Steps to Reproduce

Subject: Escalation Request – Regression in Claude Code Refactor / Analysis Behavior

Hello,

I’m requesting escalation of this ticket to engineering or product management because the current responses have not addressed the underlying issue.

Conversation ID: 215473375123528

Summary

After the rollout of Claude Opus 4.6 and the recent service incidents in early March, I am observing a major regression in Claude Code’s repository analysis and refactoring behavior.

This is not a prompt-quality issue. I operate a structured orchestration workflow with 20 dedicated prompt files that explicitly control analysis, sequencing, and safety gates.

Previous versions of the system handled this workflow reliably.

The current system does not.

Observed Regression

The agent now frequently:

• Executes prompt instructions mechanically and sequentially without situational reasoning
• Skips repository-wide analysis before performing edits
• Performs local file transformations without dependency awareness
• Deletes or modifies code without confirmation or verification steps
• Fails to inspect surrounding files even when explicitly instructed to analyze them first

In effect, the agent appears to execute prompts blindly in autopilot mode rather than using them as a decision framework.

Impact

This breaks the expected workflow for controlled refactoring and significantly increases the risk of destructive edits.

The behavior change feels similar to losing the pre-edit analysis phase that previously made refactoring reliable.

Where earlier versions performed:

1. repository inspection
2. dependency mapping
3. reasoning about impact
4. controlled edits

the current system frequently performs:

1. minimal inspection
2. direct edits based on the first file encountered

Important Context

• This workflow previously functioned correctly.
• The prompt orchestration explicitly enforces analysis steps before modification.
• The issue started after the Opus 4.6 rollout and the March outages.

Request

Please escalate this ticket to the relevant engineering or product team for investigation.

Specifically, I would like clarification on whether any of the following changed recently:

• repository analysis / file inspection behavior
• tool orchestration logic in Claude Code
• reasoning or token allocation affecting multi-file analysis
• agent confirmation safeguards for destructive actions

If helpful, I can provide:

• reproducible prompts
• workflow examples
• before/after behavior comparisons

Thank you for forwarding this to the appropriate technical team.

Claude Model

None

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

Claude Code

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

Terminal.app (macOS)

Additional Information

_No response_