Fallback: when a provider runs dry, the next one answers
A provider can now name models to use when it can't take a request — a coding plan out of quota, a rate limit, an overloaded or failing vendor.
- Set it in the provider editor (the new Fallback field: pick models from your other providers, reorder with Up), or from a terminal:
magpie provider fallback <id> <provider/model>…(noneclears it).magpie providersshows each chain. - When it kicks in: 402, 408, 429 and 5xx replies, and 400/401/403 replies that say "quota", "insufficient balance", "rate limit", "额度", "余额", "限流"… — not ordinary request errors.
- Only before the reply starts: your agent always gets one whole answer from one provider, never half from each.
- A provider that just failed rests for a minute at the back of the line, so the next requests go straight to the one that works.
- The request log shows which provider answered and which were skipped, and why.
备用：一个供应商额度见底，下一个接上
供应商现在可以设置「备用」模型：额度用完、被限流、过载或故障时，请求自动转给下一个。
- 在供应商编辑器的 备用 里挑选其他供应商的模型并排序，或在终端运行
magpie provider fallback <id> <provider/model>…（none清空）。 - 触发条件：402、408、429、5xx，以及写明额度/余额/限流等字样的 400/401/403；普通请求错误不会切换。
- 只在回复开始前切换，agent 收到的永远是一个供应商给出的完整回复。
- 刚失败的供应商会暂停一分钟再用。