[BUG] Don't ask again" answer selected not persisting across similar Bash commands

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?

## Bug Description

When using the "Don't ask again" checkbox on command approval prompts, the system continues to request approval for subsequent similar commands, even though they should be covered by the pre-authorized patterns.

## Steps to Reproduce

1. Execute a Bash command like git commit -m "message 1"
2. When prompted for approval, check "Don't ask again"
3. Approve the command
4. Execute another similar command like git commit -m "message 2"
5. Expected: Command executes without prompt (already covered by pattern)
6. Actual: System requests approval again

## Environment

• Claude Code version: [your version]
• Platform: Windows 10
• Working directory contained git repository: Yes

## Additional Context

The pre-authorized patterns defined at the start of the conversation include:
Bash(git add:)
Bash(git commit:)
Bash(dotnet build:*)

These patterns with :* wildcards should cover ALL variations of these commands, but the approval system continues to prompt even after checking "Don't ask again".

This happens with:

• Different commit messages
• Combined commands (e.g., cd "path" && git commit ...)
• Commands with different parameters but same base command

## Expected Behavior

Once "Don't ask again" is checked for a command pattern, all subsequent commands matching that pattern (or covered by the wildcard pre-authorization) should execute without additional prompts.

## Actual Behavior

The system re-prompts for approval on similar commands, even when:

1. "Don't ask again" was previously checked
2. Commands are covered by pre-authorized patterns like Bash(git commit:*)

This significantly impacts workflow efficiency during iterative development sessions.

## Additional Bug: Wasteful Token Usage During Bug Reporting

While reporting this bug, the assistant used WebFetch to retrieve the GitHub issues URL, consuming user tokens unnecessarily.

Issue: The assistant should know standard GitHub URL patterns (/issues/new) without needing to fetch them.

Impact: Users are charged tokens for the assistant's meta-actions (like finding where to report bugs), which should be free/built-in knowledge.

Expected: Common platform URLs (GitHub, Stack Overflow, docs sites) should be part of the assistant's base knowledge without requiring web fetches.

Actual: Assistant makes unnecessary HTTP requests that consume user tokens for basic navigation.

What Should Happen?

1. claude shoud run the commands without asking for permission
2. send the bug report without consuming my tokens when reporting a bug.

Error Messages/Logs

Steps to Reproduce

Steps to Reproduce

1. Open Claude Code (version 2.0.32 or later)
2. Start a conversation in a git repository
3. Execute: git add file.txt
4. When the approval dialog appears:

• Check the ✅ "Don't ask me again" checkbox
• Click "Allow"

1. Execute: git add file2.txt (different file, same command pattern)

Expected: No prompt (covered by "don't ask again")
Actual: Approval prompt appears again

Alternative reproduction:

1. Execute: git commit -m "first commit"
2. Check "Don't ask me again" and approve
3. Execute: git commit -m "second commit" (different message)

Expected: No prompt (same base command)
Actual: Prompt appears again

## What Makes It Worse

Even though pre-authorized patterns exist at session start:
Bash(git add:)
Bash(git commit:)

The "don't ask again" checkbox doesn't seem to:

• Persist the pattern across similar commands
• Respect the wildcard patterns already defined
• Save preferences between invocations with different parameters

## Impact

In a typical development session with 10-20 git operations, users must approve each one individually despite repeatedly checking "don't ask again", making the checkbox essentially non-functional.

Claude Model

Sonnet (default)

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.0.32 (Claude Code)

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

Windows Terminal

Additional Information

_No response_