Prevent Fable 5 from switching down to Opus 4.8

My notes

Summary

A short tutorial on using two Claude Code sessions together: one instance runs the actual work inside a tmux pane, while a second instance watches it and automatically intervenes if the first session gets silently downgraded from a higher-tier model to a lower one. The “fix” involves editing the recent conversation before resending it to avoid re-triggering the downgrade.

Key Insight

• Claude Code can auto-downgrade the active model mid-session (in this case from a higher tier down to a lower one), apparently triggered by certain content categories appearing in the conversation, including security/cybersecurity topics.
• Tmux lets a second CLI agent instance both read and write (send-keys) into the terminal pane of a first, independent instance, effectively turning one agent into a supervisor for another running in the same environment.
• The supervisor pattern: poll on an interval (the video uses a recurring/loop-style prompt), tail the primary session’s JSONL transcript file for the specific downgrade warning string, and when found, take action on the primary pane.
• The corrective action shown is to strip out security/cybersecurity-related references from the recent transcript before re-injecting a prompt, on the theory that removing the trigger content prevents the downgrade from re-occurring.
• Net effect: an automated loop that detects an unwanted model-tier change and pushes the session back to the desired model without manual intervention.

Caveat worth flagging: this is a single, unverified TikTok source describing what looks like an attempt to route around an internal model-selection/safety mechanism by scrubbing trigger keywords. Treat the underlying “downgrade cause” theory as anecdotal, and note that deliberately stripping safety-relevant content to avoid a guardrail is the ethically/operationally risky part of the technique, not the tmux orchestration itself.