Fable 5 safeguard false-positives derail long agent sessions

During a legitimate dev session (applying a database migration to my own product's prod DB), Fable 5's safeguards flagged routine messages twice (02:59Z and 03:08Z on 2026-07-27). The first flag silently switched my session to Opus 4.8 (despite switchModelsOnFlag: false); the second erred every retry, and because the flagged content remained in context the session could not recover — I lost roughly an hour and two work items.

Separately: in the mobile app's live voice mode, tool calls play a loud repeating progress chime with no mute setting, which makes long-running tool loops unusable by ear.

Requests:
1. Safeguard flags should not fire on credential-indirection dev-ops text, or should degrade more gracefully than silent model switching.
2. A visible "this turn was flagged — rephrase it" recovery hint.
3. A sounds toggle for live-mode tool calls.

No files, tests, or entry points are named. Separate the safeguard behavior, flagged-turn recovery, and mobile live-mode audio requests, then locate the relevant subsystems before checking how each behavior is currently handled. Done means the reported false-positive flow recovers without unwanted model switching or repeated failure, and live tool-call sounds can be muted.

索引モデルが issue の本文から書いたものです。