避免 Claude Fable 被降級成 Opus 的多 Root 架構

• URL: https://www.threads.com/share/BACq0wNIeu/
• Date Saved: 2026-08-01
• Source: Threads (@yuliang615)
• Tags: ai-engineering

Summary

A technique to prevent Claude Fable from being downgraded to Opus (when safety classifiers trigger a fallback). Uses a multi-root agent architecture:

Architecture: Root 0 / Root 1 / Root 2

• Root 0 (runs as Fable/Opus): Acts as a relay/dispatcher only. Does NOT do project planning itself. Two jobs:

  1. Pass tasks to Root 1
  2. Monitor whether Root 1 has been downgraded to Opus

• Root 1 (runs as Fable): The actual “brain” that does all the real work

• Root 2 (optional): Assigned by Root 1 for miscellaneous tasks (saves tokens, like using Opus subagents for grunt work)

Auto-recovery flow: If Root 0 detects Root 1 got downgraded to Opus:

1. Summarize Root 1’s conversation
2. Modify the prompt to Root 1 (remove aggressive/flagged wording)
3. Re-dispatch Root 1 with cleaned prompt

This creates a fully automated degradation-avoidance loop.

Caveat: Only works if you’re not actually building something malicious — just cases where legitimate work gets falsely flagged.

Context from Related Threads

• After 7/7 pricing change, users developed strategies to minimize Fable usage (use Opus for most tasks, only call Fable for what Opus can’t do)
• Fable 5 triggers safety classifiers → falls back to Opus 4.8
• Open-source alternatives (GLM-5.2, DeepSeek V4) approaching closed-model performance at 1/4-1/8 cost