fix(server): 修复内置 agent 切换时系统提示未注入及 @agent 出现在用户消息中的问题

- set_agent 控制消息处理中，查找对应 AgentDefinition 并将其 getSystemPrompt()
  写入 options.systemPrompt，使 QueryEngine.ask() 能正确使用 agent 系统提示
- 将 "build"（前端默认模式的显示名称）处理为 undefined，切换回默认模式时
  同步清空 options.systemPrompt 和 mainThreadAgentType
- /prompt 和 /prompt_async 接口不再将 agentMentions (@agent-*) 拼入 content，
  该语法仅用于 REPL 命令模式，出现在 server 模式的 prompt 中会被模型误解

Co-Authored-By: Sonnet-4.6-CoStrict <noreply@anthropic.com>

src/cli/print.ts

@@ -3100,9 +3100,28 @@ function runHeadlessStreaming(
           sendControlResponseSuccess(msg)
         } else if (msg.request.subtype === 'set_agent') {
           const agent = typeof msg.request.agent === 'string' ? msg.request.agent : undefined
-          if (agent) {
-            setMainThreadAgentType(agent)
-            saveAgentSetting(agent)
+          // 'build' is the frontend display name for the default (no-agent) mode.
+          // Treat it as clearing the agent, same as undefined.
+          const resolvedAgent = agent === 'build' ? undefined : agent
+          setMainThreadAgentType(resolvedAgent)
+          saveAgentSetting(resolvedAgent ?? '')
+          if (resolvedAgent) {
+            // Apply the agent's system prompt so QueryEngine picks it up on the next turn.
+            // QueryEngine.ask() uses options.systemPrompt (customSystemPrompt) directly,
+            // bypassing buildEffectiveSystemPrompt/mainThreadAgentDefinition. For built-in
+            // agents the getSystemPrompt params are ignored; pass a minimal stub.
+            const agentDef = currentAgents.find(a => a.agentType === resolvedAgent)
+            if (agentDef) {
+              const agentSystemPrompt = isBuiltInAgent(agentDef)
+                ? agentDef.getSystemPrompt({ toolUseContext: { options: {} as never } })
+                : agentDef.getSystemPrompt()
+              if (agentSystemPrompt) {
+                options.systemPrompt = agentSystemPrompt
+              }
+            }
+          } else {
+            // Switching back to default mode: clear any previously applied agent system prompt.
+            options.systemPrompt = undefined
           }
           sendControlResponseSuccess(msg)
         } else if (msg.request.subtype === 'set_max_thinking_tokens') {

src/server/routes/session.ts

@@ -382,8 +382,10 @@ export function createSessionRoutes(
         .join('\n') ?? ''
 
       const agentParts = body.parts?.filter((p) => p.type === 'agent' && p.name) ?? []
-      const agentMentions = agentParts.map((p) => `@agent-${p.name}`).join(' ')
-      const content = [textContent, agentMentions].filter(Boolean).join('\n')
+      // Agent switching is handled via setAgent() control message below.
+      // Do not append @agent-* mentions to content — they are REPL-only syntax
+      // and would be sent verbatim to the model in server mode.
+      const content = textContent
 
       if (!content) throw badRequest('content is required')
       if (handle.prompting) throw conflict('session is already processing a prompt')
@@ -435,8 +437,10 @@ export function createSessionRoutes(
         .join('\n') ?? ''
 
       const agentParts = body.parts?.filter((p) => p.type === 'agent' && p.name) ?? []
-      const agentMentions = agentParts.map((p) => `@agent-${p.name}`).join(' ')
-      const content = [textContent, agentMentions].filter(Boolean).join('\n')
+      // Agent switching is handled via setAgent() control message below.
+      // Do not append @agent-* mentions to content — they are REPL-only syntax
+      // and would be sent verbatim to the model in server mode.
+      const content = textContent
 
       if (!content) {
         throw badRequest('content is required')