fix: gateway stream idle timeout honors KIN_STREAM_IDLE_TIMEOUT (#485)

CHANGELOG.md

@@ -28,6 +28,7 @@
 ## 未发布
 
 - Anthropic 流式 `event: error` 经 `mapUpstreamError` 重映射，Usage Policy 与 non-stream 一致为 `policy_blocked` / `refusal_guard`。（#503）
+- 修复网关流路径忽略 `KIN_STREAM_IDLE_TIMEOUT`：与面板 `failover.stream_idle_timeout_ms`、`kernel.json` 同步走同一级联（面板 → cfg.limits → 环境变量 → 默认 180 秒），默认仍为 180 秒（#485）。
 
 ## 1.3.182 — 2026-10-17
 

docs/PANEL_API.md

@@ -202,6 +202,8 @@ Claude PATCH `max_sessions` 返回 400 `claude_max_sessions_forbidden`；OpenAI
 
 `PUT` 热更新。`tiers` 必须回传：该块是整体替换而非 patch，缺字段即置空。Claude 分档保存只传播到未手动覆盖的 Claude 槽。OpenAI `codex.quota` 按字段合并，不覆盖其它 Codex 协议/客户端配置，只更新跟随的 OpenAI 限额。
 
+`failover.stream_idle_timeout_ms`（毫秒，30 秒–60 分钟，默认 180 秒）：上游流连续无事件即判超时；网关流路径、各槽 `kernel.json` 的 `idle_timeout_seconds` 与 job 看门狗共用。优先级：面板该字段 → 启动时写入 `cfg.limits.stream_idle_timeout_ms` → 控制面 `.env` 的 `KIN_STREAM_IDLE_TIMEOUT` → 180 秒。保存后热写各槽 `kernel.json`；运行中 kernel 在无任务时自动重启以读取新看门狗值。
+
 `compatibility.persona_preset`（`official` / `official_full` / `zero` / `custom`）和 `compatibility.cache_ttl`（`5m` / `1h`）保存后投影到每个 Claude 槽的 `vms/<id>/run/kernel.json`：`persona_preset`、`system_layout`（`zero`→`zero`，其余→`identity`）、`default_cache_ttl`。响应 `kernel_persona.updated` 是本次字节有变化的槽数。`PATCH /vms/:id` 的 `timezone` / `timezone_follow_proxy` 另把 `timezone` 热写进该槽 `kernel.json`，`timezone_sync.kernel_hot` 表示文件有变化。容器 `TZ` 不在这次写入里。Codex 槽不写。
 
 `compatibility.agent_standing`（字符串，≤2000）与四个按档布尔 map `agent_standing_presets` / `agent_standing_hide_presets` / `persona_env_presets` / `persona_hide_presets` 控制常驻约束、约束遮罩、Environment 和整档 usage 遮罩，不投影到 `kernel.json`，Node 每次请求热读。`GET /api/panel/persona/preview-vars?timezone=<IANA>` 返回 system提示词页预览用的真实模板常量（身份句、官方 agent 全文、按该时区渲染的 Environment），不含 billing；时区非法或缺省按 UTC。

src/lib/core/config.mjs

@@ -25,6 +25,14 @@ export function streamIdleTimeoutMs(routing, fallbackMs = Number(process.env.KIN
   return Math.min(STREAM_IDLE_TIMEOUT_MAX_MS, Math.max(STREAM_IDLE_TIMEOUT_MIN_MS, Math.round(ms)))
 }
 
+/** Gateway stream path: panel routing → cfg.limits → KIN_STREAM_IDLE_TIMEOUT → 180s default. */
+export function resolveStreamIdleTimeoutMs(routing, cfg) {
+  const limitsMs = Number(cfg?.limits?.stream_idle_timeout_ms)
+  const fallback =
+    Number.isFinite(limitsMs) && limitsMs > 0 ? limitsMs : Number(process.env.KIN_STREAM_IDLE_TIMEOUT) || 180_000
+  return streamIdleTimeoutMs(routing, fallback)
+}
+
 /**
  * How long a Claude stream may sit in the queue before the first downstream
  * byte. Cloudflare returns 524 when the origin writes nothing for ~100s.

src/lib/protocol/handle-protocol.mjs

@@ -3,11 +3,7 @@
  * server wiring; this factory owns convert → pool → Go/Rust hop → client.
  */
 import { applyIntercept } from '../core/intercept.mjs'
-import {
-  streamIdleTimeoutMs as idleTimeoutFromRouting,
-  inboundBodyLimit,
-  queueKeepaliveDelayMs,
-} from '../core/config.mjs'
+import { resolveStreamIdleTimeoutMs, inboundBodyLimit, queueKeepaliveDelayMs } from '../core/config.mjs'
 import { applyEagerToolStreaming } from './eager-tool-streaming.mjs'
 import {
   inboundRefusalDeviceId,
@@ -1093,10 +1089,7 @@ export function createHandleProtocol(deps) {
     // First downstream byte during a queue wait. Short misses stay HTTP 429/529;
     // anything still waiting past this writes SSE so Cloudflare does not 524.
     const queueKeepaliveDelay = queueKeepaliveDelayMs(getRouting(), cfg.limits?.queue_keepalive_delay_ms)
-    const streamIdleTimeoutMs = idleTimeoutFromRouting(
-      getRouting(),
-      Number(cfg.limits.stream_idle_timeout_ms) || 180_000,
-    )
+    const streamIdleTimeoutMs = resolveStreamIdleTimeoutMs(getRouting(), cfg)
     const managedKey = req.apiKeyRecord || null
     if (managedKey) {
       const gate = apiKeyStore.acquire(managedKey)

test/unit/routing-file.test.mjs

@@ -11,6 +11,7 @@ import {
   queueKeepaliveDelayMs,
   readRoutingConfigFile,
   routingConfigFile,
+  resolveStreamIdleTimeoutMs,
   streamIdleTimeoutMs,
 } from '../../src/lib/core/config.mjs'
 
@@ -84,6 +85,34 @@ test('streamIdleTimeoutMs prefers the panel value, then the caller fallback', ()
   })
 })
 
+test('resolveStreamIdleTimeoutMs cascades panel, cfg.limits, env, then 180s default', () => {
+  withEnv('KIN_STREAM_IDLE_TIMEOUT', '', () => {
+    assert.equal(resolveStreamIdleTimeoutMs({ failover: {} }, { limits: {} }), 180_000)
+    assert.equal(
+      resolveStreamIdleTimeoutMs(
+        { failover: { stream_idle_timeout_ms: 600_000 } },
+        { limits: { stream_idle_timeout_ms: 2000 } },
+      ),
+      600_000,
+    )
+    assert.equal(resolveStreamIdleTimeoutMs({ failover: {} }, { limits: { stream_idle_timeout_ms: 90_000 } }), 90_000)
+  })
+  withEnv('KIN_STREAM_IDLE_TIMEOUT', '420000', () => {
+    assert.equal(resolveStreamIdleTimeoutMs({ failover: {} }, { limits: {} }), 420_000)
+    assert.equal(
+      resolveStreamIdleTimeoutMs({ failover: {} }, { limits: { stream_idle_timeout_ms: Number.NaN } }),
+      420_000,
+    )
+  })
+})
+
+test('resolveStreamIdleTimeoutMs avoids hardcoded 180k when cfg.limits is unset but env is set', () => {
+  withEnv('KIN_STREAM_IDLE_TIMEOUT', '600000', () => {
+    const cfg = { limits: { max_body_bytes: 1 } }
+    assert.equal(resolveStreamIdleTimeoutMs({ failover: {} }, cfg), 600_000)
+  })
+})
+
 test('queueKeepaliveDelayMs stays under the Cloudflare window', () => {
   withEnv('KIN_QUEUE_KEEPALIVE_DELAY', '', () => {
     assert.equal(queueKeepaliveDelayMs(null), DEFAULT_QUEUE_KEEPALIVE_DELAY_MS)

web/src/features/settings/pool-pane.tsx

@@ -870,7 +870,7 @@ export function PoolPane({
                 <SliderSetting
                   id='failover-stream-idle'
                   label='流空闲超时'
-                  desc='上游流连续无事件超过此时长即断开；大文件 Write 会被上游整段缓冲，过短会误断。保存后各槽 kernel 会在没有请求时自动重启以应用新值。'
+                  desc='上游流连续无事件超过此时长即断开；大文件 Write 会被上游整段缓冲，过短会误断。未在面板设置时可用控制面环境变量 KIN_STREAM_IDLE_TIMEOUT（毫秒）。保存后各槽 kernel 会在没有请求时自动重启以应用新值。'
                   value={numOf(
                     failoverDraft,
                     'stream_idle_timeout_ms',