fix(delegate-task): add subagent turn limit and model routing transparency

原因：
- subagent 无最大步数限制，陷入 tool-call 死循环时可无限运行，造成巨额 API 费用
- category 路由将 subagent 静默切换到与父 session 不同的模型，用户完全无感知

改动：
- sync-session-poller: 新增 maxAssistantTurns 参数（默认 300），每检测到新 assistant 消息
  计数一次，超限后调用 abortSyncSession 并返回明确错误信息
- sync-task: task 完成时在返回字符串中显示实际使用的模型；若与父 session 模型不同，
  加 ⚠️ 警告提示用户发生了静默路由

影响：
- 现有行为不变，maxAssistantTurns 为可选参数，默认值 300 远高于正常任务所需轮次
- 修复 #2571：用户一个下午因 Sisyphus-Junior 死循环 + 静默路由到 Gemini 3.1 Pro
  烧掉 $350+，且 OpenCode 显示费用仅为实际的一半

src/tools/delegate-task/sync-session-poller.ts

@@ -39,6 +39,8 @@ export function isSessionComplete(messages: SessionMessage[]): boolean {
   return lastUser.info.id < lastAssistant.info.id
 }
 
+const DEFAULT_MAX_ASSISTANT_TURNS = 300
+
 export async function pollSyncSession(
   ctx: ToolContextWithMetadata,
   client: OpencodeClient,
@@ -48,16 +50,20 @@ export async function pollSyncSession(
     toastManager: { removeTask: (id: string) => void } | null | undefined
     taskId: string | undefined
     anchorMessageCount?: number
+    maxAssistantTurns?: number
   },
   timeoutMs?: number
 ): Promise<string | null> {
   const syncTiming = getTimingConfig()
   const maxPollTimeMs = Math.max(timeoutMs ?? getDefaultSyncPollTimeoutMs(), 50)
+  const maxTurns = input.maxAssistantTurns ?? DEFAULT_MAX_ASSISTANT_TURNS
   const pollStart = Date.now()
   let pollCount = 0
   let timedOut = false
+  let assistantTurnCount = 0
+  let lastSeenAssistantId: string | undefined
 
-  log("[task] Starting poll loop", { sessionID: input.sessionID, agentToUse: input.agentToUse })
+  log("[task] Starting poll loop", { sessionID: input.sessionID, agentToUse: input.agentToUse, maxTurns })
 
   while (Date.now() - pollStart < maxPollTimeMs) {
     if (ctx.abort?.aborted) {
@@ -112,7 +118,23 @@ export async function pollSyncSession(
       break
     }
 
+    // 计数新出现的 assistant 轮次，用于熔断无限循环
     const lastAssistant = [...msgs].reverse().find((m) => m.info?.role === "assistant")
+    if (lastAssistant?.info?.id && lastAssistant.info.id !== lastSeenAssistantId) {
+      lastSeenAssistantId = lastAssistant.info.id
+      assistantTurnCount++
+      if (assistantTurnCount >= maxTurns) {
+        log("[task] Max assistant turns reached, aborting to prevent infinite loop", {
+          sessionID: input.sessionID,
+          assistantTurnCount,
+          maxTurns,
+        })
+        abortSyncSession(client, input.sessionID, "max_turns_exceeded")
+        if (input.toastManager && input.taskId) input.toastManager.removeTask(input.taskId)
+        return `Task aborted: subagent exceeded ${maxTurns} assistant turns without completing. This usually indicates an infinite tool-call loop. Session ID: ${input.sessionID}`
+      }
+    }
+
     const hasAssistantText = msgs.some((m) => {
       if (m.info?.role !== "assistant") return false
       const parts = m.parts ?? []

src/tools/delegate-task/sync-task.ts

@@ -149,9 +149,23 @@ export async function executeSyncTask(
 
       const duration = formatDuration(startTime)
 
+      // 检测模型路由是否与父 session 不同，给用户可见的提示
+      const actualModelStr = categoryModel
+        ? `${categoryModel.providerID}/${categoryModel.modelID}`
+        : undefined
+      const parentModelStr = parentContext.model
+        ? `${parentContext.model.providerID}/${parentContext.model.modelID}`
+        : undefined
+      const modelRoutingNote =
+        actualModelStr && parentModelStr && actualModelStr !== parentModelStr
+          ? `\n⚠️  Model routing: parent used ${parentModelStr}, this subagent used ${actualModelStr} (via category: ${args.category ?? "unknown"})`
+          : actualModelStr
+            ? `\nModel: ${actualModelStr}${args.category ? ` (category: ${args.category})` : ""}`
+            : ""
+
       return `Task completed in ${duration}.
 
-Agent: ${agentToUse}${args.category ? ` (category: ${args.category})` : ""}
+Agent: ${agentToUse}${args.category ? ` (category: ${args.category})` : ""}${modelRoutingNote}
 
 ---