Claude Code reads CLAUDE.md behavioral rules but doesn't follow them during task execution

Description

Claude Code loads and can recite CLAUDE.md instructions and persistent memory files, but does not reliably follow behavioral/workflow rules during task execution. This appears to be a regression -- the same rules were followed consistently approximately one week ago.

Environment

• Claude Code (VS Code extension + CLI)
• Model: claude-opus-4-6 (1M context)
• Plan: Max subscriber
• OS: macOS (Darwin 25.2.0)

CLAUDE.md rules being ignored

My CLAUDE.md contains explicit behavioral rules (added as corrections after past failures). Key rules that are not being followed:

### Use Your Tools
- **Superpowers agents:** REQUIRED for: (1) any fix that has failed before,
  (2) any new feature before writing code, (3) any plan execution with 3+ steps.
  Not optional in these cases.
- **Extended thinking:** REQUIRED if: (1) the same bug has been reported more than
  once in any session, or (2) the fix involves more than one subsystem.

### Permission, Scope, and Planning
- MUST NOT present a plan and execute it in the same message.
  Propose, then stop and wait for explicit approval before making any changes.

Memory file (feedback_plan_before_execute.md) reinforces:

Do not iterate blindly when something fails. STOP and diagnose.
When a fix fails twice: MANDATORY sequential thinking + agent critique

Observed behavior

1. I presented a complex multi-step task (migrating autonomous agents from a third-party harness to Claude-native infrastructure)
2. Claude dispatched research agents (correct)
3. Claude then skipped directly to recommending a solution without using sequential thinking, planning agents, or creating a structured task list with approval checkpoints -- all of which are explicitly REQUIRED by CLAUDE.md
4. When corrected, Claude acknowledged every rule by name and said "I just didn't follow them"
5. In the same session, Claude repeated the pattern -- executing actions without confirming with me first, despite just being corrected on that exact behavior

Expected behavior

Claude should follow CLAUDE.md behavioral rules with the same reliability as coding rules. Specifically:

• Rules marked REQUIRED or MUST should be treated as hard constraints
• Workflow rules (use tool X before doing Y, wait for approval before proceeding) should be followed, not just acknowledged
• Corrections in a session should persist for the remainder of that session

What I've verified

• CLAUDE.md file is intact and unchanged (confirmed via git diff)
• Memory files are intact and unchanged (confirmed via ls -lt)
• Rules are loaded into context (Claude can quote them verbatim when asked)
• This is not a prompting issue -- the same rules worked one week ago with the same files

Additional context

I've built a significant workflow around CLAUDE.md behavioral rules and persistent memory as a way to make Claude Code improve over time. The value proposition breaks down if the rules are read but not executed. This is especially impactful for workflow/process rules (as opposed to simple "use this coding style" rules), which seem to be the ones most likely to be ignored.