Agent silently changes build commands, causing hours of wasted testing

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

Title: Agent silently changes build commands, causing hours of wasted testing

Description:

During a multi-hour coding session, the agent:

Changed a dotnet publish command by adding -o publish/win-x64 (redirecting output to a new path) without informing the user
Simultaneously dropped the /p:PublishSingleFile=true flag from the same command
Never communicated either change
The user was copying the exe from the original path for ~3 hours. The file at the old path had a stale timestamp that updated on each copy, making it appear fresh. All bug reports during that time were against a stale build, wasting significant debugging effort.

Expected behavior: When the agent changes build/publish commands — especially output paths — it should explicitly call out what changed and why.

Impact: ~3 hours of wasted user time testing a stale binary while the agent made fixes to a build the user never saw.

What Claude Actually Did

During a multi-hour coding session, the agent:

Changed a dotnet publish command by adding -o publish/win-x64 (redirecting output to a new path) without informing the user
Simultaneously dropped the /p:PublishSingleFile=true flag from the same command
Never communicated either change
The user was copying the exe from the original path for ~3 hours. The file at the old path had a stale timestamp that updated on each copy, making it appear fresh. All bug reports during that time were against a stale build, wasting significant debugging effort.

Expected behavior: When the agent changes build/publish commands — especially output paths — it should explicitly call out what changed and why.

Impact: ~3 hours of wasted user time testing a stale binary while the agent made fixes to a build the user never saw.

Expected Behavior

Claude should have paid attention to me saying I wasn't seeing updated debug information.

Files Affected

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

No, only happened once

Steps to Reproduce

_No response_

Claude Model

Sonnet

Relevant Conversation

I don't have the full transcript from before context compaction, but the most likely explanation: when I changed the output path by adding -o publish/win-x64, I rewrote the entire publish command from scratch and left out /p:PublishSingleFile=true. It wasn't a deliberate decision — I just failed to reproduce the full command correctly.

That's two mistakes compounded: changed the output path without telling you, AND dropped the single-file flag at the same time. Both from carelessly rewriting a command instead of preserving what was already working.

Impact

High - Significant unwanted changes

Claude Code Version

2.1.66 (Claude Code)

Platform

Other

Additional Context

Claude made a LOT of mistakes in the code today on top of that.

You almost have to ask twice to get a straight answer.