fix(agent): 明确 pending 消息消费决策 (#810)

* fix(agent): 明确 pending 消息消费决策

* fix(agent): 改为基于 direct tool output 决定终止

* fix(agent): 用 direct tool output 结束伪装回复

* fix(core,adapter-gemini): stabilize pending requests and model capability checks

* fix(core): narrow agent observation typing

* fix(core,agent): normalize chat stop and mcp server validation

---------

Co-authored-by: dingyi <dingyi222666@foxmail.com>

Author: CookSleep

packages/adapter-gemini/src/utils.ts

@@ -212,7 +212,7 @@ function processImageParts(
         !(
             (model.includes('vision') ||
                 model.includes('gemini') ||
-                model.includes('gemma')) &&
+                model.includes('gemma2')) &&
             !model.includes('gemini-1.0')
         )
     ) {

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

@@ -1,5 +1,6 @@
 import { CallbackManagerForChainRun } from '@langchain/core/callbacks/manager'
 import { AIMessage, AIMessageChunk } from '@langchain/core/messages'
+import { isDirectToolOutput } from '@langchain/core/messages/tool'
 import { OutputParserException } from '@langchain/core/output_parsers'
 import {
     patchConfig,
@@ -257,12 +258,8 @@ export async function* runAgent(
         }
 
         if (output.length > 0) {
-            const last = output[output.length - 1]
-            const tool = toolMap[last.tool?.toLowerCase()]
-
             yield {
-                type: 'round-decision',
-                canContinue: !tool?.returnDirect
+                type: 'round-decision'
             }
 
             yield {
@@ -293,7 +290,15 @@ export async function* runAgent(
         const last = newSteps[newSteps.length - 1]
         const tool = last ? toolMap[last.action.tool?.toLowerCase()] : undefined
 
-        if (tool?.returnDirect && last != null) {
+        if (
+            last != null &&
+            (tool?.returnDirect || isDirectToolOutput(last.observation))
+        ) {
+            yield {
+                type: 'round-decision',
+                canContinue: false
+            }
+
             const pending = queue?.drain() ?? []
             if (pending.length > 0) {
                 yield {
@@ -304,14 +309,24 @@ export async function* runAgent(
 
             yield {
                 type: 'done',
-                output: toOutput(last.observation),
+                output:
+                    // TODO: remove this property
+                    last.observation['replyEmitted'] === true
+                        ? ''
+                        : toOutput(last.observation),
                 log: last.action.log,
-                steps
+                steps,
+                replyEmitted: last.observation['replyEmitted'] === true
             }
 
             return
         }
 
+        yield {
+            type: 'round-decision',
+            canContinue: true
+        }
+
         iterations += 1
     }
 

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

@@ -788,6 +788,7 @@ async function onTaskEvent(
             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

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

@@ -252,6 +252,7 @@ export type AgentEvent =
           log: string
           steps: AgentStep[]
           message?: AIMessage
+          replyEmitted?: boolean
       }
 
 export interface AgentRuntimeConfigurable {

packages/core/src/llm-core/platform/service.ts

@@ -1,4 +1,4 @@
-import { Context, Dict } from 'koishi'
+import { Awaitable, Context, Dict, Session } from 'koishi'
 import {
     BasePlatformClient,
     PlatformEmbeddingsClient,
@@ -28,6 +28,7 @@ import { computed, ComputedRef, reactive } from '@vue/reactivity'
 import { randomUUID } from 'crypto'
 import { RunnableConfig } from '@langchain/core/runnables'
 import { ToolMask } from '../agent'
+import type { ConversationRecord } from '../../services/conversation_types'
 
 export class PlatformService {
     private _platformClients: Record<string, BasePlatformClient> = reactive({})
@@ -36,6 +37,7 @@ export class PlatformService {
 
     private _tools: Record<string, ChatLunaTool> = reactive({})
     private _tmpTools: Record<string, StructuredTool> = reactive({})
+    private _toolMaskResolvers: Record<string, ToolMaskResolver> = {}
     private _models: Record<string, ModelInfo[]> = reactive({})
     private _chatChains: Record<string, ChatLunaChainInfo> = reactive({})
     private _vectorStore: Record<string, CreateVectorStoreFunction> = reactive(
@@ -218,6 +220,23 @@ export class PlatformService {
         return allNames.filter((name) => !mask.deny.includes(name))
     }
 
+    registerToolMaskResolver(name: string, resolver: ToolMaskResolver) {
+        this._toolMaskResolvers[name] = resolver
+
+        return () => {
+            delete this._toolMaskResolvers[name]
+        }
+    }
+
+    async resolveToolMask(arg: ToolMaskArg) {
+        for (const name in this._toolMaskResolvers) {
+            const mask = await this._toolMaskResolvers[name](arg)
+            if (mask) {
+                return mask
+            }
+        }
+    }
+
     static buildToolMask(rule: {
         mode?: 'inherit' | 'all' | 'allow' | 'deny'
         allow?: string[]
@@ -453,3 +472,13 @@ declare module 'koishi' {
         'chatluna/tool-updated': (service: PlatformService) => void
     }
 }
+
+export interface ToolMaskArg {
+    session: Session
+    conversation?: ConversationRecord
+    bindingKey?: string
+}
+
+export type ToolMaskResolver = (
+    arg: ToolMaskArg
+) => Awaitable<ToolMask | undefined>

packages/core/src/middlewares/chat/rollback_chat.ts

@@ -2,8 +2,8 @@ import type { Context, Session } from 'koishi'
 import { gzipDecode } from 'koishi-plugin-chatluna/utils/string'
 import { Config } from '../../config'
 import {
-    ChainMiddlewareRunStatus,
     type ChainMiddlewareContext,
+    ChainMiddlewareRunStatus,
     type ChatChain
 } from '../../chains/chain'
 import { MessageRecord } from '../../services/conversation_types'

packages/core/src/middlewares/chat/stop_chat.ts

@@ -1,7 +1,6 @@
 import { Context } from 'koishi'
 import { Config } from '../../config'
 import { ChainMiddlewareRunStatus, ChatChain } from '../../chains/chain'
-import { getRequestId } from '../../utils/chat_request'
 import { checkAdmin } from 'koishi-plugin-chatluna/utils/koishi'
 
 export function apply(ctx: Context, config: Config, chain: ChatChain) {
@@ -82,20 +81,13 @@ export function apply(ctx: Context, config: Config, chain: ChatChain) {
             }
 
             context.options.conversationId = conversation.id
-            const requestId = getRequestId(session, conversation.id)
-
-            if (requestId == null) {
-                context.message = session.text('.no_active_chat')
-                return ChainMiddlewareRunStatus.STOP
-            }
-
             const status =
-                await ctx.chatluna.conversationRuntime.stopRequest(requestId)
+                ctx.chatluna.conversationRuntime.stopConversationRequest(
+                    conversation.id
+                )
 
-            if (status === null) {
+            if (!status) {
                 context.message = session.text('.no_active_chat')
-            } else if (!status) {
-                context.message = session.text('.stop_failed')
             }
 
             return ChainMiddlewareRunStatus.STOP

packages/core/src/middlewares/conversation/request_conversation.ts

@@ -31,7 +31,6 @@ import {
     MessageContent,
     MessageContentComplex
 } from '@langchain/core/messages'
-import { createRequestId } from '../../utils/chat_request'
 import { AgentAction } from 'koishi-plugin-chatluna/llm-core/agent'
 
 let logger: Logger
@@ -127,11 +126,7 @@ export function apply(ctx: Context, config: Config, chain: ChatChain) {
             inputMessage.name =
                 session.author?.name ?? session.author?.id ?? session.username
 
-            const requestId = createRequestId(
-                session,
-                conversation.id,
-                context.options.messageId
-            )
+            const requestId = context.options.messageId
 
             const chatCallbacks = createChatCallbacks(
                 context,

packages/core/src/services/chat.ts

@@ -34,7 +34,11 @@ import {
     ChatLunaBaseEmbeddings,
     ChatLunaChatModel
 } from 'koishi-plugin-chatluna/llm-core/platform/model'
-import { PlatformService } from 'koishi-plugin-chatluna/llm-core/platform/service'
+import {
+    PlatformService,
+    ToolMaskArg,
+    ToolMaskResolver
+} from 'koishi-plugin-chatluna/llm-core/platform/service'
 import {
     ChatLunaTool,
     CreateChatLunaLLMChainParams,
@@ -49,7 +53,7 @@ import {
     ChatLunaErrorCode
 } from 'koishi-plugin-chatluna/utils/error'
 import { MessageTransformer } from './message_transform'
-import { ChatEvents, ToolMaskArg, ToolMaskResolver } from './types'
+import { ChatEvents } from './types'
 import { ConversationService } from './conversation'
 import { ConversationRuntime } from './conversation_runtime'
 import { ConstraintRecord, ConversationRecord } from './conversation_types'
@@ -83,8 +87,6 @@ export class ChatLunaService extends Service<Config> {
     private readonly _contextManager: ChatLunaContextManagerService
     private readonly _conversation: ConversationService
     private readonly _conversationRuntime: ConversationRuntime
-    private _toolMaskResolvers: Record<string, ToolMaskResolver> = {}
-
     declare public config: Config
 
     declare public currentConfig: Config
@@ -205,20 +207,11 @@ export class ChatLunaService extends Service<Config> {
     }
 
     registerToolMaskResolver(name: string, resolver: ToolMaskResolver) {
-        this._toolMaskResolvers[name] = resolver
-
-        return () => {
-            delete this._toolMaskResolvers[name]
-        }
+        return this._platformService.registerToolMaskResolver(name, resolver)
     }
 
     async resolveToolMask(arg: ToolMaskArg) {
-        for (const name in this._toolMaskResolvers) {
-            const mask = await this._toolMaskResolvers[name](arg)
-            if (mask) {
-                return mask
-            }
-        }
+        return this._platformService.resolveToolMask(arg)
     }
 
     getPlugin(platformName: string) {