Fable 5 safeguard model-switch pauses the session — add opt-in to auto-continue on the downgraded model (Opus 4.8)

Status Open
Maintainer reply None cached
Activity 0 comments · opened Jul 21, 2026

Summary

When running on Fable 5, the safety classifier sometimes flags a turn and downgrades to Opus 4.8, showing:

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. … Switched to Opus 4.8. Send feedback with /feedback or learn more

The downgrade itself is fine — the problem is that the session pauses and waits for the user to re-engage instead of continuing the turn on the downgraded model.

Impact

In an autonomous / long-running session this stalled all work for ~6 hours: the main loop was paused at the safeguard prompt, and hourly heartbeat monitors fired but could not un-pause it (only user input resumes the loop). For unattended/agentic workflows this turns a transparent model-downgrade into a multi-hour dead stop.

Current behavior

  • Safeguard flags the turn → switches to Opus 4.8 → pauses, requiring manual user re-engagement to proceed.
  • There is no settings.json key, env var, or CLI flag to change this (checked settings.md).
  • No hook event fires on the safeguard switch; the Stop hook only fires after the pause, so it can't be used to auto-resume cleanly.

Requested behavior

An opt-in (settings.json key and/or -- flag, e.g. autoContinueOnSafeguardSwitch: true) that makes the flagged turn continue automatically on the downgraded model (Opus 4.8) instead of pausing. Users who opt in accept the downgrade — they just don't want the idle stall.

Alternatives considered

  • Default to Opus 4.8 for everything (/model opus): eliminates the pause but sacrifices Fable's capability on all work, not only flagged turns — not the desired middle ground (keep Fable normally, auto-continue on downgrade only when flagged).
  • A hook event on the safeguard switch would be an acceptable alternative — it would let users script the auto-resume themselves.

Environment

  • Model: Fable 5 (Claude Code CLI), safeguard auto-downgrade to Opus 4.8.

View original on GitHub ↗