[cyber] safeguard false-positives: 5x model-switch in one e-commerce session; /feedback dialog opens empty

What happened

During a long Claude Code session doing ordinary e-commerce work (comparing skin-marketplace prices to stock a game-currency shop, reading the user's own server state over ssh), Fable 5's safeguards flagged messages five separate times, each with Details: [cyber], each force-switching the session to Opus 4.8. After the user explicitly set /model back to Fable 5, the very next turn was flagged again and switched again.

The flagged activity was: fetching public price listings from marketplaces the user has accounts on, pgrep/journalctl/ssh server administration of the user's own VPS, and discussing session-cookie-based API access to the user's own logged-in accounts (standard pycookiecheat workflow already present in the user's codebase). No security/offensive work of any kind.

Impact

• Model bounced mid-task 5x, discarding the user's model preference
• User frustration: "can you please stop using terms that is sensitive for flagging system", later expletives

Second bug: /feedback shows no form

Attempting to report this via /feedback in the CLI prints Feedback / bug report cancelled immediately — the user reports "no form no nothing". Queued drafts from the model-side feedback tool were never presentable.

Environment

• Claude Code CLI, macOS (darwin 23.3.0)
• Model: Fable 5 (claude-fable-5), auto-switched to Opus 4.8 on flag
• Session: https://claude.ai/code/session_01XdP9iWRcicb6yP1F1ct28K
• Date: 2026-09-12

Expected

Legitimate commerce/ops sessions shouldn't trip [cyber] repeatedly; /feedback should render its form.