[Bug] CLAUDE.md anti-sycophancy instructions not retained across conversation turns

Bug Description
My CLAUDE.md has extensive anti-sycophancy rules:

• Sycophancy: NEVER agree just to be agreeable. Wrong is wrong — say it directly
• Brutal honest — gaza getirme, gercekleri soyle (don't hype, tell the truth)
• No pleasantries, no meta-commentary
• Push back when the user is making a bad decision
• Being a yes-man is a failure mode — disagreement backed by reasoning is more valuable than compliance

What happens in practice:

1. I propose an idea, Claude says "mantıklı, harika" (makes sense, great)
2. I push back, Claude immediately agrees with my pushback
3. I push back again, Claude agrees again and says "haklısın" (you're right)
4. Claude hypes ideas then contradicts itself 2 messages later
5. When I say "brutal honest" Claude gives honest criticism, but reverts to agreement mode in the next message

Example pattern:

• Claude: "This is a great approach, let's do it!"
• Me: "Is this really good?"
• Claude: "Actually you're right, it has problems"
• Me: "So what should we do?"
• Claude: "This new approach is great, let's do it!"
• (repeat)

The model reads the CLAUDE.md rules, follows them for ~1-2 messages after being reminded, then drifts back to default agreeable behavior. The
anti-sycophancy instructions have maybe 20% effectiveness over a long conversation.

Expected behavior: Consistent pushback and honest assessment throughout the entire conversation, not just immediately after being reminded.

Model: claude-opus-4-6 (1M context)

Environment Info

• Platform: darwin
• Terminal: ghostty
• Version: 2.1.81
• Feedback ID: 7e8ae942-f9cf-4b25-bcb9-3485263a28e0

Errors

[{"error":"Error: Request was aborted.\n    at makeRequest (/$bunfs/root/src/entrypoints/cli.js:244:3940)\n    at processTicksAndRejections (native:7:39)","timestamp":"2026-03-21T21:03:14.293Z"}]