fix(agent): 明确 pending 消息消费决策

packages/core/src/llm-core/agent/legacy-executor.ts

@@ -234,7 +234,7 @@ export async function* runAgent(
 
             yield {
                 type: 'round-decision',
-                canContinue: false
+                willConsumePendingMessages: false
             }
 
             const pending = queue?.drain() ?? []
@@ -262,7 +262,7 @@ export async function* runAgent(
 
             yield {
                 type: 'round-decision',
-                canContinue: !tool?.returnDirect
+                willConsumePendingMessages: !tool?.returnDirect
             }
 
             yield {
@@ -293,6 +293,31 @@ export async function* runAgent(
         const last = newSteps[newSteps.length - 1]
         const tool = last ? toolMap[last.action.tool?.toLowerCase()] : undefined
 
+        if (last?.observation === '__character_reply_final__') {
+            yield {
+                type: 'round-decision',
+                willConsumePendingMessages: false
+            }
+
+            const pending = queue?.drain() ?? []
+            if (pending.length > 0) {
+                yield {
+                    type: 'human-update',
+                    messages: pending
+                }
+            }
+
+            yield {
+                type: 'done',
+                output: '',
+                log: last.action.log,
+                steps,
+                replyEmitted: true
+            }
+
+            return
+        }
+
         if (tool?.returnDirect && last != null) {
             const pending = queue?.drain() ?? []
             if (pending.length > 0) {
@@ -317,7 +342,7 @@ export async function* runAgent(
 
     yield {
         type: 'round-decision',
-        canContinue: false
+        willConsumePendingMessages: false
     }
 
     yield {

packages/core/src/llm-core/agent/sub-agent.ts

@@ -409,7 +409,7 @@
         'Delegate a focused task to a specialized agent when parallel work, deeper investigation, or a narrower prompt will help.',
         'Use the exact agent name from the injected catalog.',
         'If delegated work may take a while, set background=true so it can continue beyond the normal tool timeout.',
         'Use action=list or action=status to inspect background tasks, action=message to send more guidance while they run, and action=run with the same id to continue the same session later.'
     ].join('\n')
 }
 
@@ -464,7 +464,7 @@
                 .enum(['run', 'status', 'list', 'message'])
                 .optional()
                 .describe(
                     'run starts or resumes an agent task, status inspects one task, list shows recent tasks in this conversation, message sends live guidance to a running background task.'
                 ),
             agent: z
                 .string()
@@ -784,6 +784,7 @@
             at: Date.now(),
             title: '最终输出',
             text:
+                (event.replyEmitted ? '最终回复已由工具发送。' : '') ||
                 getMessageContent(event.message?.content ?? '') ||
                 event.output ||
                 event.log
@@ -882,7 +883,7 @@
         `agent: ${task.agentName}`,
         `run_id: ${run.runId}`,
         `state: ${run.state}`,
         `resume_hint: use ${toolName} with {"action":"run","id":"${task.id}","prompt":"next instruction"} to continue this session. Add "background":true when the work may take a while.`,
         '',
         output.trim() || '(empty)'
     ].join('\n')
@@ -975,7 +976,7 @@
 
     if (run?.state !== 'running') {
         lines.push(
             `resume_hint: use ${toolName} with {"action":"run","id":"${task.id}","prompt":"next instruction"} to continue this session. Add "background":true when the work may take a while.`
         )
     }
 

packages/core/src/llm-core/agent/types.ts

@@ -244,14 +244,15 @@ export type AgentEvent =
       }
     | {
           type: 'round-decision'
-          canContinue?: boolean
+          willConsumePendingMessages?: boolean
       }
     | {
           type: 'done'
           output: string
           log: string
           steps: AgentStep[]
           message?: AIMessage
+          replyEmitted?: boolean
       }
 
 export interface AgentRuntimeConfigurable {

packages/core/src/services/chat.ts

@@ -968,7 +968,7 @@ type ActiveRequest = {
     abortController: AbortController
     chatMode: string
     messageQueue: MessageQueue
-    roundDecisionResolvers: ((canContinue: boolean) => void)[]
+    roundDecisionResolvers: ((willConsume: boolean) => void)[]
     lastDecision?: boolean
 }
 
@@ -1096,13 +1096,14 @@ class ChatInterfaceWrapper {
                 toolMask: mask,
                 onAgentEvent: async (agentEvent) => {
                     if (agentEvent.type === 'round-decision') {
-                        activeRequest.lastDecision = agentEvent.canContinue
-                        if (agentEvent.canContinue == null) {
+                        activeRequest.lastDecision =
+                            agentEvent.willConsumePendingMessages
+                        if (agentEvent.willConsumePendingMessages == null) {
                             return
                         }
 
                         for (const resolve of activeRequest.roundDecisionResolvers) {
-                            resolve(agentEvent.canContinue)
+                            resolve(agentEvent.willConsumePendingMessages)
                         }
                         activeRequest.roundDecisionResolvers = []
                     }
@@ -1202,11 +1203,11 @@ class ChatInterfaceWrapper {
         }
 
         return new Promise((resolve) => {
-            activeRequest.roundDecisionResolvers.push((canContinue) => {
-                if (canContinue) {
+            activeRequest.roundDecisionResolvers.push((willConsume) => {
+                if (willConsume) {
                     activeRequest.messageQueue.push(message)
                 }
-                resolve(canContinue)
+                resolve(willConsume)
             })
         })
     }