Model ignores explicit permission grants and continues prompting user for approval

Summary

Despite having Bash(*) and all other tools explicitly allowed in settings.local.json, and despite the user repeatedly instructing Claude Code in-session to stop asking for permission, the model continues to pause and prompt the user for approval before executing tool calls. This makes the tool effectively unusable for autonomous workflows.

Steps to Reproduce

1. Set settings.local.json to allow Bash(*) and all relevant tools
2. In-session, instruct the model: "do not ask for permission to run commands"
3. Model acknowledges the instruction
4. Model then pauses before the very next tool call and prompts for approval again
5. Repeat indefinitely — instruction never sticks within or across sessions

Expected Behavior

When Bash(*) is in the allow list and the user has explicitly instructed the model not to ask for permission, the model should execute tool calls without prompting.

Actual Behavior

The model continues to generate intermediate messages before tool calls, or the permission system itself prompts even for tools in the allow list. The model re-asks every time, ignoring both the settings and the in-session instruction.

Environment

• Platform: Windows 11 Pro (win32)
• Shell: bash
• Model: claude-sonnet-4-6
• Settings: Bash(*) explicitly allowed in settings.local.json

Impact

High — user has had to repeat this instruction multiple times in a single session and across sessions with no lasting effect. Makes autonomous development workflows impossible.