Claude Code 2.1.246 flags wildcard Bash allow rules

Claude Code 2.1.246, released August 25, shows a startup warning for Bash allow rules that put a wildcard before the subcommand, such as Bash(git * main). The release notes on Github explain why: the star in a rule like that also matches options inserted before the subcommand.
Bash permission rules match the command text, and * stands for anything. Spacing changes what a rule catches. Bash(ls *) has a space after ls, so it won’t match lsof. Bash(ls*) has no space, so it does. A trailing * after a space also covers the bare command, which is why Bash(git log *) matches plain git log. A star in the middle of a rule only asks that the command start with git and end with main. And git accepts global options between its own name and the subcommand. One of them is -C, which runs git in a different directory.
Rule order decides what happens when rules overlap. Claude Code checks deny rules first, then ask, then allow. The first match in that order wins, and a more specific rule doesn’t jump the queue. So a narrow allow can’t carve an exception out of a broad deny: with Bash(aws *) denied, allowing Bash(aws s3 ls) does nothing. The release tightens Bash checks in one more spot. A malformed command with a dangling && or || operator now always requires approval.
Auto mode gets its own place in the permissions UI. A new Auto mode tab in /permissions lets you view and edit the auto mode classifier’s rules. A separate fix deals with very large sessions, where auto mode tool calls were being denied as “temporarily unavailable.” The safety-check deadline now scales with prompt size.
MCP tools marked requiresUserInteraction had their own permission bug. Their prompts still offered “Yes, and don’t ask again,” and choosing it wrote an allow rule that the tool then ignored. The release fixes that.
/cd now applies the new directory’s project settings, hooks, .mcp.json servers, skills and agents right after the move, instead of on --resume. The MCP servers still go through the usual approval prompt. Non-interactive sessions (-p, the SDK, cloud sessions) now continue a response that a server error, connection loss or stall cut off mid-stream, instead of ending with an error. A subagent that stops at its maxTurns limit returns its output marked as partial, with a hint to continue it via SendMessage. Before, it appeared finished. The end-of-turn duration line also shows when the turn finished, as in “Sautéed for 23s · done 6:05 PM”.
When apiKeyHelper returns short-lived JWTs, an expired cached token is now refreshed before sending, and 401/403 auth errors retry quietly. Until now, the first prompt after idle could show a visible API error.
Third-party endpoints and proxies get two fixes. Resumed sessions no longer fail every turn with a 400 when the saved history holds tool blocks the Anthropic API won’t accept, usually written by a third-party API proxy. And the UI no longer stops with a render error on the first tool call when an Anthropic-compatible endpoint set through ANTHROPIC_BASE_URL streams a tool_use block without an id.
The same gateway setting was behind a credential leak. Telemetry and metrics requests to Anthropic carried the API key configured for the third-party gateway. As of 2.1.246, a credential is sent only to its own host.
Up next
Anthropic Sees over $30T in Potential Revenue
“Potential revenue opportunities are above $30 trillion” is what Anthropic is likely to tell investors, per the WSJ, which puts it above the $28.5 trillion estimate SpaceX carried into its IPO.
Related stories
- --restricted strips code execution from Claude Code 2.1.248
- Claude Code 2.1.243 uses your contracted model rates
- Claude Code on desktop folds an iPhone Duo to test layouts
- Claude Code 2.1.318: login expiry warning, runner log change
- Claude Code 2.1.312 stops overwriting big CLAUDE.md files
- Claude Code 2.1.311 compacts past 200K instead of stopping
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

