Session quality degrades significantly during long conversations

Bug

During a long session (~3+ hours), Claude Code's quality of responses and decision-making degrades significantly. The AI becomes:

1. Contradictory — tells user to re-upload App Store build, then minutes later realizes it wasn't needed. Wastes significant user time.
2. Forgetful of context — forgets what features exist in which target (e.g., repeatedly asking about popups in the Lite target that has no popups)
3. Unable to think through consequences — changed shared asset catalog icons when only the App Store icon needed changing, breaking the entire app's appearance
4. Indecisive — flip-flopping on whether a re-upload is needed, unable to commit to an answer
5. Drip-feeding information — App Store submission required 5+ rounds of "one more thing needed" instead of providing a complete checklist upfront

Timeline of issues in a single session

• Changed shared icon assets when asked to change only the App Store icon → broke all popups
• Created 3 unnecessary builds in App Store Connect → persistent "Newer Build Available" warning that can't be dismissed
• Told user to re-upload with build 4 to fix icons → then realized Lite has no popups showing the icon → re-upload was unnecessary
• Previous session claimed to save manual test files but never wrote them to disk
• Repeatedly asked user questions they'd already answered or that the AI should have known from context

Impact

• Hours of wasted user time
• Significant trust damage — user cannot rely on Claude Code's recommendations
• Multiple unnecessary App Store builds that can never be deleted
• User had to step away from the session due to frustration

Expected behavior

• Consistent, well-reasoned recommendations that don't flip-flop
• Think through the full implications before suggesting actions
• Maintain awareness of project context (which target has which features)
• Provide complete information upfront, not incrementally
• Quality should not degrade over long sessions