[MODEL] Cannot handle > 1 bug fix in a conversation

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

Claude modified files I didn't ask it to modify

What You Asked Claude to Do

I am using Claude to assist in coding a full stack application, often in my unit testing I find the following:

1. Agents cannot locate the issue, yet there is clearly a bug. The agents are more of a syntax checker rather than finding deeper issues for example here is a snippet:

**● The agent's analysis on items 2 and 3 was too shallow — "no bug found" when the user is clearly seeing bugs means I
need to look deeper. Let me read the actual files.

Searched for 2 patterns, read 6 files (ctrl+o to expand)**

1. When finding multiple errors to analyze, fix Claude 'Forgets about errors 2 thru N', and immediately starts a 'Chain of Thought' fix for error # 1, and forgets about the remaining issues, prompting me to waste time/tokens to get a full fix of the errors identified. My work around is to force Claude to create a 'Punchlist' so it does not forget what it is doing. This should be default behavior of the model rather than the user having to 'prethink' how to phrase the question to Claude. For Example here is a snippet:

**On your meta-question — honest answers:

Do you have to ask for a punch list every time?
No. I'll save this as a standing instruction in memory right now so it applies automatically to every future session.

Is it a design issue?
Partially yes, partially workflow. The real behavior is: I don't "forget" mid-session, but I have a strong bias toward
starting to code the moment I have enough context to act. On multi-bug requests, that eagerness causes me to dive into item 1 before fully analyzing items 2-N. A punch-list-first rule suppresses that bias. Whether that bias should be a default is a fair product feedback point.**

What Claude Actually Did

1. Claude proceeds to fix bug # 1 and it forgets about the rest, telling me to retest.

Expected Behavior

1. Parse my list of errors.
2. Create a punch list of issues to correct. OR ask me any questions of not clear on the issue.
3. Proceed to fix each of the identified errors reducing cycle time back and forth to the AI.

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

1. Give claude 6 or 7 bugs from any source code.
2. do NOT ask for a punch list
3. Claude's default will be to fix issue # 1 maybe # 2 and then 'forget' the remaining tasks.
4. Instead of going back and unit testing 1 - 7, I have to test 1 & 2 and give Claude the remaining bugs as a new conversation.

Claude Model

Sonnet

Relevant Conversation

Impact

Critical - Data loss or corrupted project

Claude Code Version

Claude Code v2.1.107 ▝▜█████▛▘ Sonnet 4.6 · Claude Pro

Platform

Anthropic API

Additional Context

_No response_