Claude Code causes terrible UX in follow-up sessions after context loss

Problem

After a context compaction killed a session mid-commit (see #43883, #43886), the follow-up session was an absolute disaster. The model:

1. Claimed gh CLI wasn't installed when it was — just didn't check /opt/homebrew/bin/gh
2. Asked the user for permission to install gh via brew instead of just trying the existing binary
3. Filed a duplicate issue after the user had already filed one, then couldn't figure out whether it was a duplicate or not
4. Went back and forth multiple times contradicting itself about whether the issues were duplicates
5. Told the user to go set up email notifications that were already configured — wasting more time
6. Required multiple prompts to do simple things that should have been done immediately

The user had to babysit every single step. The model should have just done what was asked without hesitation, checked its assumptions before claiming something wasn't possible, and not wasted the user's time with unnecessary back-and-forth.

Expected behavior

• Try obvious paths (/opt/homebrew/bin/) before claiming a tool isn't installed
• Don't ask permission for obvious actions — just do them
• Don't contradict yourself across consecutive messages
• Check existing state before telling users to configure something

Impact

What should have been a 2-minute task (file an issue, add a comment) turned into 20+ minutes of frustration. The model actively made things worse instead of helping.