[Bug] Fable 5 bio-category false positives on infrastructure terminology / request for better programmatic model recovery options when model is downgraded
Bug Description
Two related issues from a long Claude Code session on Fable 5 (which fell back to Opus 5)
1) Bio-category false positives on routine infrastructure work.
Fable 5 flagged three messages as [bio] in 26 hours. All three were ordinary software/devops work with no biological content whatsoever:
2026-08-25 09:44:25 PDT fable-5 -> opus-5 [bio]
2026-08-25 13:35:57 PDT fable-5 -> opus-5 [bio]
2026-08-26 01:56:33 PDT fable-5 -> opus-5 [bio]
The third followed a shell command running git worktree prune and listing file permissions. The flagged message read: "all ten files in the settings family are now owner-only, and the worker is the healthy pid". The words family, worker, healthy, prune, and live all appear in the exchange as software terms. Whatever the actual trigger, common infrastructure vocabulary appears to be reaching and tripping the biology classifier.
2) No supported way to programmatically return to the configured model after a fallback.
In a world where model demotion is a mandate, the existing automatic switch is the less-evil behavior I want. With auto-switching disabled, the conversation pauses and requires a human to act, which defeats long unattended sessions. But after a switch, the model stays on the fallback for the rest of the conversation, and there is no programmatic way back:
- Hooks receive a model field only on SessionStart, and there is no hook output field that can set the model.
- Rewriting settings.json affects future sessions, not the running one.
For an unattended session, this means silently running on a different, less-capable, and less-appropriate model for hours. The only notification mechanism is a single line that scrolls past in the chat text, among tool output. I discovered the 01:56 switch several hours later the following morning by looking at the CLI and then scrolling back, and could not initially find when the demotion of the Fable model to Opus 5 happened.
Requests:
- Tune the bio classifier against software infrastructure vocabulary.
- Provide a supported way to intentionally automate restoration of the configured model: a hook output field, a CLI call, or an opt-in "return to the configured model after the flagged turn" setting, perhaps with a rollback to before the offending turn and an instruction to not do the thing that tripped the protection. Right now the only option I have is an external detector and tmux workaround, which is not a proper product solution.
Related feature requests: anthropics/claude-code issues #47275, #31342, #17772.
Environment Info
- Platform: darwin
- Terminal: tmux
- Version: 2.1.245
- Feedback ID: 63250c70-8aab-47f6-84bd-db71fc7e81c7
This issue has 2 comments on GitHub. Read the full discussion on GitHub ↗