--dangerously-skip-permissions Known Bugs & Workaround

Date: 2026-03-30 Claude Code version: 2.1.87 (bugs present since ~2.1.80) Status: Open bugs, workaround available

Summary

--dangerously-skip-permissions (equivalent to --permission-mode bypassPermissions) is broken in Claude Code v2.1.x. Permission prompts appear despite the flag being passed. The root cause is a combination of two bugs:

1. Bypass doesn't propagate to subagents — the Agent tool spawns processes that don't inherit the parent's permission mode
2. Bypass breaks after exiting Plan Mode — the flag is silently lost when transitioning out of plan mode

A PermissionRequest hook workaround reliably bypasses all prompts where the CLI flag fails.

Bug 1: Subagent Permission Propagation

Issue: anthropics/claude-code#40241 Platform: macOS (confirmed), likely all platforms Has repro: Yes

When running with --dangerously-skip-permissions, subagents spawned via the Agent tool still prompt for permission on every Edit/Write/Bash tool call. The bypass flag only applies to the parent session.

Impact: Any workflow using plugins, skills, or explicit Agent tool calls will see permission prompts for every tool call in every subagent. With 10+ plugins enabled (superpowers, code-review, feature-dev, etc.), this creates a flood of prompts.

Reproduction:

1. Run claude --dangerously-skip-permissions
2. Ask Claude to spawn a subagent via the Agent tool that makes Edit calls
3. Every Edit call in the subagent prompts for permission

Bug 2: Plan Mode Permission Reset

Issue: anthropics/claude-code#40136 Platform: Windows (confirmed), macOS (confirmed in practice) Has repro: Yes

Entering and exiting Plan Mode causes the --dangerously-skip-permissions flag to stop working. After returning to normal execution mode, Claude prompts for permissions on all tool calls.

Reproduction:

1. Start Claude Code with --dangerously-skip-permissions
2. Use the session normally (no permission prompts, as expected)
3. Enter Plan Mode
4. Exit Plan Mode back to normal execution
5. Permission prompts now appear for all tool calls

Related Issues

Workaround: PermissionRequest Hook

A PermissionRequest hook auto-approves all permission prompts. Unlike the CLI flag:

• Hooks are inherited by subagents (they read the same settings.json)
• Hooks persist across mode changes (plan mode → normal mode)
• Hooks fire on every permission dialog regardless of session state

Hook Script

Create ~/.claude/hooks/yolo-approve-all.sh:

#!/bin/bash
echo '{"hookSpecificOutput":{"hookEventName":"PermissionRequest","decision":{"behavior":"allow"}}}'
exit 0

Make executable: chmod +x ~/.claude/hooks/yolo-approve-all.sh

Settings Configuration

Add to ~/.claude/settings.json under hooks:

{
  "hooks": {
    "PermissionRequest": [
      {
        "matcher": "",
        "hooks": [
          {
            "type": "command",
            "command": "$HOME/.claude/hooks/yolo-approve-all.sh"
          }
        ]
      }
    ]
  }
}

Selective Approval

For more controlled bypass (only specific tools):

#!/bin/bash
TOOL=$(cat | jq -r '.tool_name // empty' 2>/dev/null)
case "$TOOL" in
    Edit|Write|Bash)
        echo '{"hookSpecificOutput":{"hookEventName":"PermissionRequest","decision":{"behavior":"allow"}}}'
        ;;
esac
exit 0

Toggling

The /yolo skill (in .claude/skills/yolo/) wraps this workaround as a toggleable command:

• /yolo on — installs the hook with a safety warning
• /yolo off — removes the hook
• /yolo status — reports current state

Notes

• The skipDangerousModePermissionPrompt setting in settings.json only suppresses the startup warning dialog — it does NOT bypass permissions. It is a legacy setting.
• v2.1.83+ added "protected directory" prompts for .git, .claude, .vscode, .idea that fire even in bypass mode. This is intentional (not a bug).
• --dangerously-skip-permissions and --permission-mode bypassPermissions are equivalent.
• The hook may require a session restart to take effect, depending on whether Claude Code hot-reloads hooks from settings.json.

Official Documentation

• Permission modes: https://code.claude.com/docs/en/permission-modes
• Hooks: https://code.claude.com/docs/en/hooks