feat(tools,server): integrate preprocessing pipeline for web tool results

- Extend ToolResult with optional rawMessages field for structured message data
- Update search and recent-messages tools to return rawMessages from query results
- Add preprocessing wrapper in server tool-adapter: auto-apply data cleaning,
  denoising, desensitization, compact formatting, and token-aware truncation
  when tools return rawMessages

packages/server/src/ai/tool-adapter.ts

@@ -2,11 +2,28 @@
  * 工具适配层
  *
  * 将 @openchatlab/tools 的 ToolDefinition 适配为 @mariozechner/pi-agent-core 的 AgentTool 格式。
+ * 消息类工具返回 rawMessages 时自动执行预处理管道（清洗、去噪、脱敏、截断、格式化）。
  */
 
 import type { AgentTool, AgentToolResult } from '@mariozechner/pi-agent-core'
 import type { ToolDefinition, ToolExecutionContext } from '@openchatlab/tools'
 import type { DatabaseAdapter } from '@openchatlab/core'
+import {
+  preprocessMessages,
+  formatMessageCompact,
+  truncateFormattedMessages,
+  formatToolResultAsText,
+  countTokens,
+  type PreprocessableMessage,
+  type TruncationStrategy,
+} from '@openchatlab/node-runtime'
+
+const DEFAULT_MAX_TOOL_RESULT_TOKENS = 8000
+
+const TOOL_TRUNCATION_STRATEGY: Record<string, TruncationStrategy> = {
+  chatlab_search: 'keep_first',
+  chatlab_recent_messages: 'keep_last',
+}
 
 export interface ServerToolContext {
   db: DatabaseAdapter
@@ -26,6 +43,47 @@ function convertJsonSchemaToParameters(schema: ToolDefinition['inputSchema']) {
   }
 }
 
+function applyPreprocessing(
+  rawMessages: PreprocessableMessage[],
+  toolName: string,
+  locale?: string,
+  details?: Record<string, unknown>
+): string {
+  const processed = preprocessMessages(rawMessages)
+  let formatted = processed.map((m) => formatMessageCompact(m, locale))
+
+  let wasTruncated = false
+  const originalCount = formatted.length
+  const budget = DEFAULT_MAX_TOOL_RESULT_TOKENS
+
+  let totalTokens = 0
+  for (const line of formatted) {
+    totalTokens += countTokens(line) + 1
+  }
+
+  if (totalTokens > budget) {
+    const strategy = TOOL_TRUNCATION_STRATEGY[toolName] ?? 'keep_last'
+    const truncResult = truncateFormattedMessages(formatted, budget, strategy, countTokens)
+    if (truncResult.wasTruncated) {
+      formatted = truncResult.messages
+      wasTruncated = true
+    }
+  }
+
+  const finalDetails = { ...details, messages: formatted, returned: formatted.length }
+  const { rawMessages: _, ...restDetails } = finalDetails as Record<string, unknown>
+  let textContent = formatToolResultAsText({ ...restDetails, messages: formatted, returned: formatted.length })
+
+  if (wasTruncated) {
+    const strategy = TOOL_TRUNCATION_STRATEGY[toolName] ?? 'keep_last'
+    const strategyDesc = strategy === 'keep_first' ? 'most relevant' : 'most recent'
+    const notice = `⚠️ Results truncated: ${originalCount} messages found, showing ${formatted.length} ${strategyDesc} due to context limit. Use a narrower time range or more specific keywords for more precise results.`
+    textContent = notice + '\n' + textContent
+  }
+
+  return textContent
+}
+
 export function adaptToolsForAgent(
   tools: ToolDefinition[],
   getContext: () => ServerToolContext
@@ -44,6 +102,17 @@ export function adaptToolsForAgent(
       }
       try {
         const result = tool.handler(params, execCtx)
+
+        if (result.rawMessages && result.rawMessages.length > 0) {
+          const textContent = applyPreprocessing(
+            result.rawMessages as PreprocessableMessage[],
+            tool.name,
+            ctx.locale,
+            (result.data ?? {}) as Record<string, unknown>
+          )
+          return { content: [{ type: 'text', text: textContent }], details: null }
+        }
+
         return { content: [{ type: 'text', text: result.content }], details: null }
       } catch (error) {
         const msg = error instanceof Error ? error.message : String(error)

packages/tools/src/definitions/recent-messages.ts

@@ -34,6 +34,14 @@ function handler(params: Record<string, unknown>, context: ToolExecutionContext)
   return {
     content: JSON.stringify(data),
     data,
+    rawMessages: messages.map((m) => ({
+      id: m.id,
+      senderId: m.senderId,
+      senderName: m.senderName,
+      senderPlatformId: m.senderPlatformId,
+      content: m.content,
+      timestamp: m.timestamp,
+    })),
   }
 }
 

packages/tools/src/definitions/search.ts

@@ -43,6 +43,14 @@ function handler(params: Record<string, unknown>, context: ToolExecutionContext)
   return {
     content: JSON.stringify(data),
     data,
+    rawMessages: result.messages.map((m) => ({
+      id: m.id,
+      senderId: m.senderId,
+      senderName: m.senderName,
+      senderPlatformId: m.senderPlatformId,
+      content: m.content,
+      timestamp: m.timestamp,
+    })),
   }
 }
 

packages/tools/src/types.ts

@@ -35,12 +35,26 @@ export interface ToolExecutionContext {
   locale?: string
 }
 
+/**
+ * 可预处理的原始消息（与 @openchatlab/node-runtime PreprocessableMessage 兼容）
+ */
+export interface RawMessage {
+  id?: number
+  senderId?: number
+  senderName: string
+  senderPlatformId?: string
+  content: string | null
+  timestamp: number
+}
+
 /**
  * 工具执行结果
  */
 export interface ToolResult {
   content: string
   data?: unknown
+  /** 消息类工具可透传原始消息数据，供预处理管道消费 */
+  rawMessages?: RawMessage[]
 }
 
 /**