Claude ignores user steering messages sent during active tool execution

Bug Description

When the user sends steering messages (corrections, "wait", "stop") while Claude is actively executing tool calls, Claude ignores them and continues executing the original plan. This occurs in both normal sessions and fork sessions.

Steps to Reproduce

1. Start a Claude Code session (normal or fork)
2. Give Claude a multi-step task that involves multiple sequential tool calls, e.g.:

• "Create 3 config files named a.yaml, b.yaml, c.yaml, each with a model field set to 'test-model'"

1. While Claude is actively running tool calls (Edit, Write, Bash), send a steering message:

• "wait, change the model field to 'correct-model' instead"

1. Observe: Claude ignores the message and continues creating files with 'test-model'

Expected Behavior

Claude should pause or stop the current task when the user sends a steering/interruption message mid-execution, then acknowledge the correction and adjust its plan accordingly.

Actual Behavior

Claude continues executing the original plan without acknowledging the user's input. The steering messages appear to be buffered but never injected into the model's next decision context. The user must manually interrupt (Esc / stop) to halt execution.

Environment

• OS: Windows 11
• Shell: PowerShell
• Claude Code: 2.1.259
• Reproducibility: Observed in both normal and fork sessions

Root Cause Hypothesis

This is a hypothesis, not confirmed.

The CLI execution loop may not check for new user messages between tool calls. Messages sent during active execution appear to be queued but not surfaced to the model until the current batch of tool calls completes or the user manually interrupts.

Impact

• Users cannot correct or redirect Claude mid-execution
• Unwanted file modifications may occur before the user can stop Claude
• Particularly problematic in long-running autonomous sessions
• Breaks the expected real-time collaboration workflow between user and agent