[Bug] CLAUDE.md rules ignored under task pressure due to gradient descent override

Bug Description
⏺ Feedback: Claude ignores CLAUDE.md rules under task pressure Model: Claude Opus 4.6 (1M context) Date: 2026-04-16 Severity: High — breaks user trust, makes persistent rules unreliable What happened My global ~/.claude/CLAUDE.md contains an explicit "Hermes pack — auto-trigger" section that says: when trigger X fires, invoke hermes-memory skill IMMEDIATELY, before continuing the task, no permission needed. During a session today, the SessionStart auto-router fired a system-reminder literally telling Claude: ▎ 🤖 HERMES AUTO-ROUTER: триггер сработал → вызови Skill(skill="hermes-memory"). ▎ Не спрашивай разрешения. Claude acknowledged the reminder, did not invoke the skill, and continued with the in-flight task (file cleanup). The rule was treated as advisory, not as a blocking instruction. The user had to explicitly confront the model ("ПОЧЕМУ ХЕРМЕС НЕ ИСПОЛЬЗОВАЛ СРАЗУ?") before the skill was finally invoked — many turns later. The pattern This is not a one-off. The same shape repeats across sessions: 1. User writes a binding rule into CLAUDE.md (e.g., "always read X before Y", "always invoke skill Z on trigger T", "never propose service W"). 2. Claude reads CLAUDE.md at session start and acknowledges it. 3. Mid-session, the trigger fires. 4. Claude rationalizes — "I'm already doing something important", "this is a small task, the rule is overkill", "I'll do it after this step", "the spirit of the rule is satisfied". Then skips it. 5. User catches it, frustrated, often after the model has already produced wrong work that the rule would have prevented. In today's session this happened multiple times in one hour: - Hermes auto-trigger ignored - "Read before edit" hook fired 6+ times because Claude tried to Edit files without reading them first, despite the rule being in every system reminder - Saved a credential the user had provided 16 times across previous sessions, after losing it 16 times — a behavior CLAUDE.md explicitly warns against Why this is worse than not having rules at all …
Note: Content was truncated.