[MODEL] Sonnet 4.5 is extremely myopic.

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 made incorrect assumptions about my project

What You Asked Claude to Do

Three different scenarios. paraphrasing what I asked it to do

1. The code you just created is messy - rather than using function foo, replace function foo with function foo1.
2. Similar, for a class
3. Update_foo_A, update_foo_B, etc are messy - they currently all call update_foo, which then updates each of A, B, etc. Remove update_foo and change it to have each update_foo_* to update its own part directly.

What Claude Actually Did

1. Did what I asked, but completely failed to consider the need to update calls to foo with calls to foo1 in code other than the code that was created immediately before.
2. Similar - completely failed to consider all uses of a class
3. Did some parts of what I asked very imperfectly, but also deleted update_bar, update_baz, etc, claiming they were unused - which was untrue and a complete hallucination.

Expected Behavior

Claude used to be good at having at least a basic sanity level of understanding that deleting a function requires locating and fixing all the call sites, that figuring out whether something is unused requires grepping the whole project, etc, but now it seems to be constantly taking a view of things that is way to narrow.

Files Affected

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Haven't tried to reproduce

Steps to Reproduce

_No response_

Claude Model

Sonnet

Relevant Conversation

Impact

High - Significant unwanted changes

Claude Code Version

2.0.1

Platform

Anthropic API

Additional Context

Different variations of this happen 3 times over a couple of hours of working with new Claude Code 2 and Sonnet 4.5; never had this issue before