Fable 5 safeguards false-positive on routine git/diagram work forces model switch; advertised /feedback (and /bug) commands don't exist

Summary

Fable 5's safeguards flagged an entirely routine assistant message during a long professional coding session and force-switched the session to Opus 4.8 mid-task — and the feedback channels the flag banner advertises (/feedback, and the classic /bug) do not exist in this build, so there was no in-product way to report it.

What was flagged

The assistant summarizing a PlantUML diagram layout change (a C4 architecture diagram for a GPS-localization data path) and offering a git commit / git push command for a docs-only branch. No security content, no exploits, nothing dual-use — a diagram render and a git command.

The banner shown:

Fable 5's safeguards flagged this message. The safeguards are intentionally broad right now and may flag safe and routine coding, cybersecurity, or biology work. These measures let us bring you Mythos-level capabilities sooner, and we're working to refine them. Switched to Opus 4.8. Send feedback with /feedback or learn more: https://support.claude.com/en/articles/15363606

Session context

Professional embedded vehicle-control development (industrial autonomy — JAUS robotics middleware, Kalman-filter sensor fusion, CI/code-review tooling). If terms like "autonomous vehicle," "GPS," or "EKF" contributed to the trigger, that's a category error: this is standard, legitimate industrial robotics engineering, and it saturates our entire workflow — any per-message trigger on that vocabulary will fire constantly for us.

Cost

The forced model switch silently discarded the model selection in the middle of a multi-hour session with heavy context; I had to notice the banner and manually switch back with /model. A silent downgrade mid-flow is worse than a refusal — it degrades output quality without an obvious cue.

The feedback channel is broken

The banner says "Send feedback with /feedback" — but /feedback does not exist in this build (VS Code extension, WSL): the command palette returns "No matching commands." The classic /bug command doesn't exist either (only /debug, which is local session debugging). So the advertised path for reporting a false positive is a dead end, and this GitHub issue is the fallback.

Asks

1. Tune the safeguards classifier to weigh session context (hours of benign software engineering) over keyword hits.
2. Make the model switch opt-in ("this message was flagged — continue on Opus?") rather than automatic and silent-ish.
3. Ship the /feedback command the banner references (or point the banner at a channel that exists).

Environment

• Claude Code VS Code extension, WSL2 (Linux 6.6.87.2-microsoft-standard-WSL2), Windows 11
• Model: claude-fable-5 [1m]; flag occurred ~07:20 local (America/Denver), 2026-07-06