[Bug] Fable5 model_refusal_fallback triggered during routine code review and bug fix tasks
Bug Description
Fable5 switched on normal engineering work. I had Fable5 do a code review, it found bugs and was in the process of fixing those bugs and it flagged a security concern and switched to Opus. I had opus review the logs and this is what it said: That screenshot confirms it exactly — that's the model_refusal_fallback event I found in the transcript, rendered as the inline CLI banner: "This model's safeguards flagged this message… Switched to Opus 4.8." So the diagnosis holds: Fable 5's safety layer false-positived on ordinary code-fix work and the harness failed over to Opus 4.8, retracting the in-flight message. Not the Bash classifier, not quota.
Two actionable things, one of which is visible in your screenshot and I'd missed:
- /config — the grey tip line under the banner ("You can configure model switch behavior in /config") is the knob. That's where you set whether a safeguard flag auto-falls-back to Opus, prompts you, or stops. If you'd rather it not silently switch mid-task, that's where to change it.
- /feedback — worth firing on this one, since the flagged content was benign engineering work; that's how the false positive gets reported back.
Environment Info
- Platform: linux
- Terminal: tmux
- Version: 2.1.198
- Feedback ID: f22e77fa-45b5-48d6-bac7-d5cdfe27a07c
Errors
[]This issue has 1 comment on GitHub. Read the full discussion on GitHub ↗