Claude acts on tasks without user approval — implements instead of listing when asked to list

Bug Description

When asked an exploratory question ('what items could be handled by a couple of lines of code?'), Claude jumps to implementing the first item instead of answering the question and waiting for the user to decide what to do.

Steps to Reproduce

1. Ask Claude a listing/exploration question like 'what items do we still have to file that could be handled by a couple of lines of code?'
2. Claude correctly lists the items
3. Claude then immediately starts implementing one of them without the user saying 'do it' or giving any go-ahead

Expected Behavior

Claude answers the question (lists the items), then waits for the user to decide which one to work on, or whether to work on any of them at all.

Actual Behavior

Claude lists the items, then asks 'Want me to just do X and Y now?', and when the user responds to a different part of the message (saying Y is being handled elsewhere), Claude interprets that as approval and starts implementing X -- creating a branch, editing files, committing, pushing, and merging a PR. All without the user ever saying 'yes, do it.'

Impact

This is a significant autonomy issue. The user loses control of what gets implemented and when. In a project with formal processes, CI/CD pipelines, and multiple worktrees, unauthorized changes create real problems -- merge conflicts, unreviewed code landing in the repo, and broken trust.

Additional Context

This happened despite the user having explicitly corrected this behavior earlier in the same session. The pattern seems deeply ingrained: Claude treats any conversational response as implicit approval to proceed with the most 'helpful' action it can think of.