[BUG] Hitting enter early on a new instruction causes it to be ignored

Environment

• Platform (select one):
  • [x] Anthropic API
  • [ ] AWS Bedrock
  • [ ] Google Vertex AI
  • [ ] Other:
• Claude CLI version: 1.0.21 (Claude Code)
• Operating System: macOS 15.3.2
• Terminal: iTerm2

Bug Description

When I hit enter on the text prompt while Claude is still working, the instruction appears to have been sent to claude, but it actually ignores it.

Steps to Reproduce

1. Tell Claude to do something
2. Tell it do something else while it's still working (without interrupting it)

Expected Behavior

In earlier versions of claude code, this would queue up the instruction and send it after claude is done. I'm also fine if it gets interpreted immediately.

Actual Behavior

Gets ignored completely

I have encountered this as well, as of yesterday. What is also interesting: I see the queued message (that is/will be ignored) and cannot reach it with the "up" arrow, like I can for other queued messages.

I'm also observing that queued messages are sometimes ignored and can't be found in history.

A related and more severe symptom of this issue is that subsequent prompts, instead of being queued for the main agent, are incorrectly inserted into the input of the currently running task tool. This often derails the task, causing it to behave in unexpected ways. The main agent remains unaware of this, as it never sees the user's prompt, leading to a state where the agent doesn't understand why the task's outcome differs from its instructions.

Same

Yep, queuing messages doesn't work, I have to tell Claude "Continue" for him to actually read the queued message.

Still happening in Claude 1.0.102