[BUG] --dangerously-skip-permissions flag does not suppress permission prompts

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

--dangerously-skip-permissions does not prevent Claude Code from prompting "Do you want to proceed?" on every Bash command. The permission prompt appears on every single command regardless of:

1. Using the --dangerously-skip-permissions CLI flag
2. 2. Selecting "Yes, allow reading from [project] from this project" when prompted
3. 3. Including explicit instructions in the prompt ("DO NOT ask for permission", "Execute all commands without asking for confirmation")

This makes autonomous task execution impossible and forces the user to manually approve dozens of commands per session. A typical multi-step debugging task requires approving 30-50+ individual commands, turning what should be a 10-minute autonomous task into 2-4 hours of babysitting.

Example - flag is active, explicit "do not ask" instructions given, still prompts on every command:

> claude --dangerously-skip-permissions

> restart back-end and test the UI in the browser. DO NOT ask for 
  permission until completely finished. 
  RULES:
  - Execute all commands without asking for confirmation
  -   - Do not stop to ask yes/no at any point
  -   -   - Complete the entire task autonomously
Bash command:
  lsof -i :5001 -P | grep LISTEN | awk '{print $2}' | xargs kill ...

Do you want to proceed?
> 1. Yes
> 2.   2. No
> 3. ```
This occurs on literally every command in the session, despite the flag AND the explicit prompt instructions.

### What Should Happen?

With `--dangerously-skip-permissions`, all Bash commands and file operations should execute automatically without any user confirmation prompts. The user should be able to give a multi-step task and have Claude Code execute it autonomously from start to finish, only returning when complete with results.

### Error Messages/Logs

```shell

Steps to Reproduce

1. Launch Claude Code with: claude --dangerously-skip-permissions
2. 2. When first prompted, select "Yes, allow reading from [project] from this project"
3. 3. Give any multi-step task, e.g.: "restart the backend server and verify the API returns correct data"
4. 4. Observe that every subsequent Bash command still shows the permission prompt: "Do you want to proceed? 1. Yes 2. No"
5. 5. Additionally, include explicit no-permission instructions in the prompt such as: "DO NOT ask for permission. Execute all commands without asking for confirmation. Complete the entire task autonomously."
6. 6. Observe that the permission prompt STILL appears on every single command

Environment: macOS, Claude Code latest version, Opus model

Impact: A typical debugging/testing session requires 30-50+ manual approvals, turning a 10-minute autonomous task into 2-4 hours of babysitting. This completely negates the value of agentic coding for multi-step workflows.

Claude Model

Sonnet (default)

Is this a regression?

No, this never worked

Last Working Version

_No response_

Claude Code Version

Latest as of March 9, 2026

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

This issue has persisted across multiple sessions over the past week. The flag has no observable effect — behavior is identical with and without it.

Attempts to work around the issue by including explicit natural language instructions in the prompt (e.g., "DO NOT ask for permission", "Execute all commands without asking for confirmation", "Complete the entire task autonomously, only return when finished") also have no effect.

Selecting "Yes, allow reading from [project] from this project" when initially prompted also does not prevent subsequent prompts.

This makes Claude Code unusable for any autonomous multi-step workflow. The entire value proposition of agentic coding is lost when the user must manually approve every individual command.