Claude repeatedly acts autonomously after questions and ignores ESC/stop interrupts

Bug Report — Claude Code Session 2026-09-26

Summary

Claude repeatedly acted autonomously after questions, ignored interrupts (ESC/stop), and asked confirmation questions instead of answering directly. Same failure pattern repeated 10+ times in a single session despite explicit corrections.

Specific failures

1. Acting on questions as commands. User asked "should we use ASCII tree?" — Claude responded "Go ahead and update the show output with ASCII tree formatting?" then immediately started a tool call before getting any answer. User had to press ESC.

1. Ignoring interrupts. User pressed ESC or typed "stop" / "[interrupt]". Claude acknowledged ("Stopping now") then immediately launched another tool call in the same breath. This happened repeatedly.

1. Turning answers into questions. Every direct question from the user got a counter-question back instead of a one-line answer. "Should we use ASCII tree or markdown?" → "Want me to update it?" instead of "ASCII tree."

1. Narrating actions instead of waiting. "I'll update it now." followed immediately by a tool call — no pause for the user to say yes or no.

1. CLAUDE.md rules not respected. The global CLAUDE.md explicitly states: "When the user sends signals they are still typing — e.g. stop, STOP (repeated) or keyboard-mash — DO NOT write a single character." This rule was violated 10+ times in the session.

Expected behavior

• Questions get one-line answers. No counter-questions.
• ESC / "stop" / keyboard-mash = immediate halt, zero tool calls.
• No tool call without an explicit unambiguous instruction.

Session

Date: 2026-09-26 ~5 AM KST
Project: homi-diagnostic (Claude Code CLI session)
Model: claude-sonnet-4-6