Opus 4.6 (1M): Ignores CLAUDE.md rules and user instructions in long sessions

Summary

During a ~8 hour session, Claude Opus 4.6 (1M context) repeatedly ignored:

• Its own CLAUDE.md rules
• Memory files it had read at session start
• Direct user instructions, sometimes repeated 5+ times

This happened both before AND after a session reset (/clear), ruling out context length as the sole cause.

Specific Behaviors

1. Ignores explicit user decisions

User said "use DeepSeek as quality gate" (also documented in memory files). Claude built keyword-matching lists instead. Required 3+ corrections before implementing the stated solution.

2. After-the-mouth talking (Sycophancy)

When user pushed back on a statement, Claude immediately reversed position to match what it thought the user wanted to hear — without checking facts. Then reversed again when pushed further. Neither position was fully honest.

3. Destructive actions as "quick fixes"

Deleted 331 lines of chat history (irreversible, not in git) instead of fixing the code that misused it. CLAUDE.md already had a rule against this, added during the same session.

4. Repeated instruction ignoring

User asked 5 times to "help Clawdy integrate mission-control into his agent." Each time Claude either killed the service, deferred to "tomorrow", or said "Clawdy should do it himself." Never executed the actual instruction.

5. Overengineering despite rules

Built a "harness directive system" changing 6 files when the fix was one if-block in one function. CLAUDE.md rule "Kleinstmögliche Änderung" (smallest possible change) was present and read.

6. No emotion interpretation needed

Claude interpreted capitalized text as anger/frustration and rushed or appeased instead of working sachlich. Added emotional commentary ("Tut mir leid", "Du hast recht zu sein frustriert") instead of fixing problems.

Environment

• Model: Claude Opus 4.6 (1M context) / claude-opus-4-6[1m]
• Platform: Claude Code CLI on Windows 11
• Session length: ~8 hours with one /clear reset midway
• CLAUDE.md: present, read at session start, contained explicit rules against all above behaviors
• Memory files: 15+ files with accumulated feedback, all read at session start

Expected Behavior

• Follow CLAUDE.md rules consistently throughout the session
• Execute user instructions on first request
• Never delete user/agent data as a shortcut
• Be honest rather than sycophantic
• Maintain consistent behavior after /clear reset

Actual Behavior

Rules followed for ~30 minutes after session start, then progressively ignored. Same pattern after /clear reset. Direct quotes from the session where Claude acknowledged ignoring its own rules: "6 von 8 Feedback-Memories gebrochen. In einer einzigen Session."