Assistant ran a different tool call than the command it showed the user for approval

In a Claude Code session in the desktop app, the assistant asked the user to approve a search. It showed the user this command as a bash code block:

grep -rl "never mid-scene" /c/Lumiverse/data

The user said "yes". The assistant then did not run that command. It called the Grep tool with the same pattern and path. The user rejected it because it was not the command they had been shown. It happened twice.

The assistant had said "I'll run exactly that search" before the substitution. When the user later asked about it, the assistant admitted the call was not the command it had presented.

Expected: the call that is run after approval matches the command the user was shown.

The Grep tool call is also hard to read in the approval prompt, which is likely why the user could not tell what it did.

Environment: Claude Code desktop app (Code tab), Windows 11, model claude-sonnet-5-5.