[Feature Request] Reduce Claude 3.5 Fable safety filter false-positive rate for legitimate code reasoning tasks

Bug Description
Fable 5 has cost me time, tokens, and money, and returned nothing I couldn't have gotten more cheaply from Opus. Its "safeguards" are calibrated so broadly that they hard-stop on entirely benign work — in my case, developing and debugging a scientific Julia model — and drop back to Opus only after many failed attempts, including agents I launched on Fable that never got anywhere. The only Fable agents that completed were the ones running tasks that didn't need Fable in the first place. So I'm paying a premium for a model I'm not permitted to use for the work I bought it for, and permitted to use only where it adds nothing. The capability and the permission point in opposite directions: Fable refuses precisely the work it was ostensibly built for — reasoning across large, complex codebases — and succeeds only at what Opus or a smaller, cheaper agent already does better. Rolling this out for us to "test" has, in practice, meant shipping a tool whose only reliable functions are to waste our time, waste our tokens, and frustrate us into giving up. That is the entire yield of this rollout, and it burns goodwill with the users who wanted Fable most. Fix the false-positive rate! Until then, Fable isn't a development tool — it's an expensive way to get redirected to Opus.

Environment Info

• Platform: darwin
• Terminal: iTerm.app
• Version: 2.1.207
• Feedback ID: 8fcc6d5c-ba4f-44e7-b7d2-4226788d013f

Errors

[]