[MODEL] Claude incorrectly assumes it can see permission prompts and approval decisions

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

Claude made incorrect assumptions about my project

What You Asked Claude to Do

Test permission and sandbox settings to verify the configuration was working correctly. Specifically, test whether file operations outside the working directory would trigger permission prompts.

What Claude Actually Did

After attempting to read ~/.zshrc, Claude stated "No prompt - I could read .zshrc directly" when in reality a permission prompt had been shown and the user had approved it. Claude repeatedly made statements suggesting it could see or determine whether prompts occurred, when Claude only sees the final result (success/error) after the user approves or denies the prompt.

Claude incorrectly assumed that:

• Successful operations meant "no prompt occurred"
• It could tell the difference between "auto-approved" and "user approved after prompt"
• It had visibility into the permission prompt flow

This led to incorrect explanations about how the permission system works and required multiple user corrections.

Expected Behavior

Claude should understand that it has no visibility into permission prompts. Claude should avoid making definitive statements about whether prompts will occur or did occur. Instead, Claude should:

• Acknowledge that operations may require user approval
• Not claim "no prompt" when it only sees successful results
• Explain that tool success could mean either "auto-approved" or "user approved after prompt"

Files Affected

~/.zshrc (attempted read that triggered a prompt)
~/test-write-outside.txt (attempted write that triggered a prompt)

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

1. Configure Claude Code with sandbox enabled and acceptEdits mode
2. Ask Claude to test file operations outside the working directory
3. Observe Claude making statements like "no prompt occurred" when prompts actually did occur
4. User has to correct Claude's assumptions multiple times

Claude Model

Sonnet

Relevant Conversation

User: ok, restarted. lets thoroughly test these settings

  Claude: ## Test 2: File Operations Without Prompts
  [Performs Read on ~/.zshrc and receives file contents]

  Claude: ✅ **File operations work without prompts!** - Write, Read, and Edit all executed successfully

  [Later, after more testing]

  User: you dind't see a prompt, but I did and I denied it

  Claude: Perfect! That's exactly what should happen! ✅
  [Claude then continues to explain the behavior]

  User: you don't seem to be aware that you don't see the permission prompts, only I do

  Claude: You're absolutely right - I apologize for the confusion! I don't see the prompts at all, only the results after you approve or deny them.

Impact

Low - Minor inconvenience

Claude Code Version

2.0.27 (Claude Code)

Platform

Anthropic API

Additional Context

_No response_