学习系统防‘工具饿死’：给关键工具配置 seed priors

• Learning/bandit mechanism is enabled and tool arm statistics are inspectable.
• You can edit learning config and reset learning DB in maintenance windows.

1. Audit current arm posteriors and identify critical tools with near-zero inclusion probability.
2. Define `seedArms` patterns for critical categories (web, exec, fs read/write) with conservative boosts.
3. Set a non-zero `min_pulls` guard so low-sample arms remain eligible during recovery phase.
4. Re-run representative tasks and compare tool selection diversity/accuracy before and after seeding.

• Over-boosted priors can bias selection and reduce adaptability for new tasks.
• Issue references ongoing implementation work; final config keys may differ (needs verification).

This tip is aggregated from community/public sources and preserved with attribution.