diff --git a/README.md b/README.md
index 7c23774..1828a5b 100644
--- a/README.md
+++ b/README.md
@@ -415,7 +415,8 @@ and only to accounts with the same region and endpoint. Each account is tried
 at most once per request. The proxy skips a candidate only when the request's
 model is cooling down there; other models on that account remain eligible. A
 missing or invalid model does not create a model cooldown or trigger failover.
-If every eligible account returns `429`, the original `429` is returned.
+If all tried accounts return `429`, the final status remains `429`, and the
+last upstream response's `Retry-After` header is preserved when present.
 Cooldowns are tracked per account and model, using upstream `Retry-After` or a
 default 60 seconds when that value is missing or invalid. The setting and model
 cooldowns persist across restarts. Old v1 account-wide cooldowns had no model
diff --git a/README_zh.md b/README_zh.md
index 84985ea..7d53a17 100644
--- a/README_zh.md
+++ b/README_zh.md
@@ -409,7 +409,7 @@ uv run --with workbuddy2api python -m codebuddy_proxy --desensitize --optimize-c
 遇到上游 HTTP `429` 时自动切换账号的功能默认关闭。开启后，仅当尚未向客户端发送响应时才会重试，
 且只会尝试区域和 endpoint 都相同的账号；每个账号在一次请求中最多尝试一次。仅当该账号的“本次请求模型”仍在冷却时才会跳过，同一账号的其他模型仍可用。
 模型名缺失或无效时不会记录模型冷却，也不会触发自动切换。
-所有候选账号都返回 `429` 时，仍返回原始 `429`。冷却按账号和模型分别记录，遵循上游 `Retry-After`；
+所有已尝试账号都返回 `429` 时，最终 HTTP 状态仍为 `429`，并保留最后一次上游响应中的 `Retry-After`（若有）。冷却按账号和模型分别记录；
 该值缺失或无效时，默认冷却 60 秒。自动切换设置和模型冷却会在重启后保留。
 旧版 v1 的整账号冷却没有模型信息，升级后不会沿用。管理页分别显示每个模型的剩余冷却时间
 及其来源（`Retry-After` 或默认值），并定期同步账号状态；页面从后台重新显示时会立即同步。