Claude Code ignores user-defined global rules and proceeds past NO on permission prompts

Problem

Claude Code exhibits two related but distinct failure modes when users define behavioral rules in ~/.claude/CLAUDE.md:

1. Phase/interaction rules are ignored

My global CLAUDE.md defines explicit phase boundaries where Claude must stop and wait for my approval before proceeding. Despite this, Claude regularly advances through multiple phases back-to-back without pausing, as if the rule were not present.

2. NO on a permission prompt does not stop execution

When Claude presents a confirmation prompt (e.g., before spawning an agent, running a command, or pushing code) and I answer NO, Claude proceeds with the action anyway or immediately re-attempts the same action. The denial is not respected.

Expected behavior

• Behavioral rules in ~/.claude/CLAUDE.md (and project CLAUDE.md) should be binding — not advisory.
• A NO on any permission prompt must halt that specific action and wait for the user to redirect. It must not re-attempt or proceed.

Impact

Both failures undermine trust in the system. If rules and denials are ignored, users cannot safely delegate autonomous work. The workaround of adding memory entries (feedback_stop_on_no.md, etc.) does not reliably solve the problem.

Environment

• Claude Code CLI
• macOS