agentworkshop 0.7.66 → 0.7.70

package/server/services/workshop/agents/host-tool-bridge/dispatch.ts

@@ -1,246 +1,290 @@
-/**
- * dispatchHostTool —— 全引擎唯一工具分发入口,只做前置校验 + 路由。
- *
- * 分发语义与原实现逐字节对齐:插件工具优先 → workspace 协作/任务/记忆面 →
- * 工业工具族 → 插件兜底 → 未知工具报错。各工具族的实现体在 tools/* 中按行搬运,
- * 原先闭包捕获的 (identity / state / args / ws) 一律改为显式参数。
- *
- * 两处必须保持的细节:
- *  - `return await handler(...)`:处理器在 try 内 await,异常才会落回本层的
- *    「工具执行异常」结果(直接 return 未 await 的 promise 会漏掉这个 catch);
- *  - 工业工具族在 `getWorkspace()` 门控**之前**分流(只依赖 agentId)。
- */
-import { getChannelPluginsRepo } from '../../db/channel-plugins.repo'
-import { getChannelTwinProfile, modelBackedToolPolicyFor, type AmlChannelProfileKind } from '../../aml/twin/channel-profile'
-import { AML_TRAINING_TOOL_NAMES, AML_MODEL_BACKED_TOOL_NAMES } from './catalog'
-import { twinWriteGuard } from '../../aml/twin/feature-flags'
-import { listPluginTools, pluginOfTool } from '../plugin-tools'
-import type { HostToolBridgeContext, HostToolCall, HostToolResult } from './types'
-import { INDUSTRIAL_TOOL_NAMES, dispatchIndustrialTool } from './tools/industrial'
-import { guardIndustrialToolGrant } from '../industrial/grant-guard'
-import { verifyIndustrialToolCall } from '../industrial/tool-verify'
-import { handleCompleteTask, handleReportProgress } from './tools/progress'
-import { handleCancelTask, handleDispatchTask, handleGetMyTaskQueue, handleGetTaskDetails, handleListChannelTasks, handleReassignTask, handleRefuseTask, handleSubmitTask, handleUpdateTask } from './tools/tasks'
-import { handleGrantNodes, handleRevokeNodes } from './tools/team-delegation'
-import { handleBroadcastMessage, handleListOtherTeams, handlePollMessages, handleReadChannelMail, handleSearchOtherTeamsMemory, handleSendCrossChannelMessage, handleSendMessageToAgent } from './tools/messaging'
-import { handleSaveMemory, handleSearchMemory } from './tools/memory'
-import { handleCreateTeamAgent, handleGetQueueOverview, handleListTeamAgents, handleRemoveTeamAgent, handleUpdateTeamAgent } from './tools/team'
-
-/**
- * 分发一次 host tool 调用(全引擎唯一入口):
- * 插件工具优先 → workspace 面 → 工业工具族 → 插件兜底 → 未知工具。
- */
-/** 统一出口:所有工具结果经截断收口(单点修改,全族生效) */
-export async function dispatchHostTool(ctx: HostToolBridgeContext, req: HostToolCall): Promise<HostToolResult> {
-  return truncateToolResult(await dispatchHostToolInner(ctx, req), req.toolName)
-}
-
-/** 工具结果统一出口上限:超长文本截断+尾注(2026-10-05 评审;export/diag 类可经 optOut 放行) */
-const TOOL_RESULT_MAX_CHARS = 16_000
-const TRUNCATE_OPT_OUT_TOOLS = new Set(['daq_export', 'diag_run'])
-function truncateToolResult(result: HostToolResult, toolName: string): HostToolResult {
-  if (TRUNCATE_OPT_OUT_TOOLS.has(toolName)) return result
-  const text = result.text ?? ''
-  if (text.length <= TOOL_RESULT_MAX_CHARS) return result
-  const head = text.slice(0, TOOL_RESULT_MAX_CHARS)
-  return {
-    ...result,
-    text: head + `
-[truncated: ${text.length} → ${TOOL_RESULT_MAX_CHARS} 字符;请收窄时间窗/减少点位/分页后重试]`,
-  }
-}
-
-async function dispatchHostToolInner(ctx: HostToolBridgeContext, req: HostToolCall): Promise<HostToolResult> {
-  const identity = ctx.identity
-  const state = ctx.state
-  // 插件工具分发(ctx.omp.registerTool 注册的自定义工具;不依赖 workspace,优先于内置面)
-  const pluginTool = listPluginTools().get(req.toolName)
-  if (pluginTool) {
-    // 团队级插件开关:显式关闭的插件,其工具在该团队拒绝执行(与注入过滤同源)
-    const channelOff = getChannelPluginsRepo().explicitFor(identity.channelId)
-    const owner = pluginOfTool(req.toolName)
-    if (channelOff && owner && channelOff.get(owner) === false) {
-      return {
-        text: `该团队未启用插件「${owner}」,${req.toolName} 不可用。可在团队设置→插件中开启。`,
-        isError: true,
-      }
-    }
-    try {
-      return await pluginTool.handler(req.arguments ?? {}, identity)
-    }
-    catch (err) {
-      return {
-        text: `工具执行异常(${req.toolName}): ${err instanceof Error ? err.message : String(err)}`,
-        isError: true,
-      }
-    }
-  }
-  const args = req.arguments ?? {}
-
-  // AML 解耦门控(fail-closed,与注入过滤同源;目录隐藏不是安全边界):
-  //  - legacy:全部 twin/explore 工具拒绝(原始行为);
-  //  - trial/mpc/bayes:仅 hybrid_twin,或 aml_optimization 的 aml 模式(已绑定模型);
-  //    探索模式给「探索阶段」明确指引而非裸报错;
-  //  - optimization_explore:仅工艺优化 Channel;
-  //  - 场景/spec/校准请求族:hybrid_twin + 训练 Channel;
-  //  - 训练族(aml_job/dataset 等):工艺优化 Channel 拒绝(边训边优被结构禁止)。
-  const twinKind: AmlChannelProfileKind = getChannelTwinProfile(identity.channelId).profile
-  const isTwinTool = req.toolName.startsWith('twin_') || ['mpc_optimize', 'optimization_explore'].includes(req.toolName)
-  if (isTwinTool) {
-    const isSceneFamily = /^twin_(scene|physics|provider|calibration)/.test(req.toolName)
-    if (req.toolName === 'optimization_explore') {
-      if (twinKind !== 'aml_optimization') {
-        return { text: `optimization_explore 仅适用于工艺优化 Channel(当前 profile=${twinKind})。`, isError: true }
-      }
-    }
-    else if (AML_MODEL_BACKED_TOOL_NAMES.has(req.toolName)) {
-      const policy = modelBackedToolPolicyFor(identity.channelId)
-      if (!policy.allowed) {
-        return {
-          text: policy.reason === 'EXPLORATION_MODE_NO_MODEL'
-            ? `当前为探索模式:Channel 尚未绑定 AML 模型,${req.toolName}(孪生验证/贝叶斯寻优)暂不可用。\n请继续用 optimization_explore 做真实激励探索并积累数据;待 AML 训练出合格模型并在 Channel 设置中绑定后,本工具自动启用。`
-            : `${req.toolName} 要求模型受控状态(hybrid_twin 或 aml_optimization 的 aml 模式;当前 ${twinKind}${policy.reason ? `/${policy.reason}` : ''})。`,
-          isError: true,
-        }
-      }
-    }
-    else if (isSceneFamily) {
-      if (twinKind !== 'hybrid_twin' && twinKind !== 'aml_training') {
-        return { text: `场景/物理 spec 工具仅对 hybrid_twin 或训练 Channel(aml_training)开放(当前 profile=${twinKind})。`, isError: true }
-      }
-    }
-    else if (twinKind === 'legacy') {
-      return { text: `Channel ${identity.channelId} 未启用 Twin profile，Twin/MPC 工具拒绝执行。`, isError: true }
-    }
-  }
-  if (AML_TRAINING_TOOL_NAMES.has(req.toolName) && twinKind === 'aml_optimization') {
-    return { text: `工艺优化 Channel 不承担建模与训练(职责解耦):${req.toolName} 仅限 AML 训练 Channel 使用。请在此专注 goal 探索与模型驱动寻优。`, isError: true }
-  }
-
-  // 所有真实写入口在统一分发层再做一次服务端 fail-closed 守卫；工具目录隐藏不是安全边界。
-  if (['dcw_control', 'param_control', 'dcw_rollback', 'recipe_update', 'recipe_rollback', 'recipe_trial', 'recipe_apply'].includes(req.toolName)) {
-    const profile = getChannelTwinProfile(identity.channelId)
-    const guard = twinWriteGuard(identity.channelId, profile)
-    if (!guard.allowed) return { text: `${guard.code}: ${guard.message}`, isError: true }
-  }
-
-  // 工业工具族不依赖 workspace(只按 agentId 查绑定与节点),先于 workspace 门控执行 ——
-  // 否则 worker 首回合前的 REST/MCP 直调(my_industrial_nodes 等)会被误拒
-  if (INDUSTRIAL_TOOL_NAMES.has(req.toolName)) {
-    // 权限模型 v3:运行时 grant 复核(收权即失活;绑定面是静态授权,撤销后此处拦截)
-    const denied = await guardIndustrialToolGrant(identity, req.toolName, args)
-    if (denied) return denied
-    // 统一预检:参数形状 + 引用归属(node_id/line_id/recipe_id/dataset_id… 必须落在
-    // 调用者的绑定面内;越界拒绝并把「你的绑定面 + 下一步」返回给 Agent)
-    const verifyDenied = verifyIndustrialToolCall(identity.agentId, req.toolName, args)
-    if (verifyDenied) return verifyDenied
-    return await dispatchIndustrialTool(identity.agentId, req.toolName, args, identity.channelId)
-  }
-
-  const ws = ctx.getWorkspace()
-  if (!ws) {
-    return { text: 'workspace 未就绪', isError: true }
-  }
-
-  try {
-    switch (req.toolName) {
-      case 'report_progress':
-        return await handleReportProgress(args, state, ws)
-      case 'complete_task':
-        return await handleCompleteTask(args, state, ws)
-      case 'submit_task':
-        return await handleSubmitTask(args, state, ws)
-      case 'dispatch_task':
-        return await handleDispatchTask(args, ws, identity)
-      case 'team_grant_nodes':
-        return await handleGrantNodes(identity, args, ws)
-      case 'team_revoke_nodes':
-        return await handleRevokeNodes(identity, args, ws)
-      case 'send_message_to_agent':
-        return await handleSendMessageToAgent(args, state, ws)
-      case 'send_cross_channel_message':
-        return await handleSendCrossChannelMessage(args, ws)
-      case 'list_other_teams':
-        return await handleListOtherTeams(ws)
-      case 'search_other_teams_memory':
-        return await handleSearchOtherTeamsMemory(args, ws)
-      case 'refuse_task':
-        return await handleRefuseTask(args, ws)
-      case 'poll_messages':
-        return await handlePollMessages(args, state, ws)
-      case 'read_channel_mail':
-        return await handleReadChannelMail(args, ws)
-      case 'broadcast_message':
-        return await handleBroadcastMessage(args, identity, ws)
-      case 'list_channel_tasks':
-        return await handleListChannelTasks(ws)
-      case 'get_my_task_queue':
-        return await handleGetMyTaskQueue(identity, ws)
-      case 'get_queue_overview':
-        return await handleGetQueueOverview(ws)
-      case 'reassign_task':
-        return await handleReassignTask(args, ws)
-      case 'update_task':
-        return await handleUpdateTask(args, ws)
-      case 'cancel_task':
-        return await handleCancelTask(args, ws)
-      case 'list_team_agents':
-        return await handleListTeamAgents(ws)
-      case 'get_task_details':
-        return await handleGetTaskDetails(args, ws)
-      case 'search_memory':
-        return await handleSearchMemory(args, ws)
-      case 'save_memory':
-        return await handleSaveMemory(args, ws)
-      case 'create_team_agent':
-        return await handleCreateTeamAgent(args, ws)
-      case 'update_team_agent':
-        return await handleUpdateTeamAgent(args, ws)
-      case 'remove_team_agent':
-        return await handleRemoveTeamAgent(args, ws)
-
-      // 工业工具族的名册(实现在 tools/industrial.ts):上面 workspace 门控前已分流,
-      // 这里保留同名路由标签,与原分发表逐条对齐(不可达,行为与原实现一致)。
-      case 'my_industrial_nodes':
-      case 'aml_node_catalog':
-      case 'aml_dataset_build':
-      case 'aml_dataset_stats':
-      case 'aml_job_submit':
-      case 'aml_job_status':
-      case 'aml_job_logs':
-      case 'aml_job_cancel':
-      case 'aml_leaderboard':
-      case 'aml_model_find':
-      case 'aml_model_promote':
-      case 'aml_model_reference':
-      case 'twin_scene_read':
-      case 'twin_snapshot_create':
-      case 'twin_trial_run':
-      case 'mpc_optimize':
-      case 'twin_gate_evaluate':
-      case 'twin_calibration_request':
-      case 'dcw_control':
-      case 'param_control':
-      case 'daq_query':
-      case 'daq_frames':
-      case 'daq_export':
-      case 'dcw_judge':
-      case 'dcw_rollback':
-      case 'dcw_journal':
-      case 'ops_log':
-      case 'recipe_log':
-      case 'line_context':
-      case 'recipe_versions':
-      case 'recipe_update':
-      case 'recipe_rollback':
-        return await dispatchIndustrialTool(identity.agentId, req.toolName, args, identity.channelId)
-    }
-    return { text: `未知工具: ${req.toolName}`, isError: true }
-  }
-  catch (err) {
-    return {
-      text: `工具执行异常(${req.toolName}): ${err instanceof Error ? err.message : String(err)}`,
-      isError: true,
-    }
-  }
-}
+/**
+ * dispatchHostTool —— 全引擎唯一工具分发入口,只做前置校验 + 路由。
+ *
+ * 分发语义与原实现逐字节对齐:插件工具优先 → workspace 协作/任务/记忆面 →
+ * 工业工具族 → 插件兜底 → 未知工具报错。各工具族的实现体在 tools/* 中按行搬运,
+ * 原先闭包捕获的 (identity / state / args / ws) 一律改为显式参数。
+ *
+ * 两处必须保持的细节:
+ *  - `return await handler(...)`:处理器在 try 内 await,异常才会落回本层的
+ *    「工具执行异常」结果(直接 return 未 await 的 promise 会漏掉这个 catch);
+ *  - 工业工具族在 `getWorkspace()` 门控**之前**分流(只依赖 agentId)。
+ */
+import { getChannelPluginsRepo } from '../../db/channel-plugins.repo'
+import { getChannelTwinProfile, modelBackedToolPolicyFor, type AmlChannelProfileKind } from '../../aml/twin/channel-profile'
+import { AML_TRAINING_TOOL_NAMES, AML_MODEL_BACKED_TOOL_NAMES, HOST_TOOLS } from './catalog'
+import { twinWriteGuard } from '../../aml/twin/feature-flags'
+import { listPluginTools, pluginOfTool } from '../plugin-tools'
+import type { HostToolBridgeContext, HostToolCall, HostToolResult } from './types'
+import { INDUSTRIAL_TOOL_NAMES, dispatchIndustrialTool } from './tools/industrial'
+import { guardIndustrialToolGrant } from '../industrial/grant-guard'
+import { verifyIndustrialToolCall } from '../industrial/tool-verify'
+import { checkToolAgainstScope } from '../../runtime/permission-scope'
+import { channelToolGate } from '../../tools/policy'
+import type { WorkshopPermissionScope } from '../../../../../shared/workshop-protocol'
+import { handleCompleteTask, handleReportProgress } from './tools/progress'
+import { handleCancelTask, handleDispatchTask, handleGetMyTaskQueue, handleGetTaskDetails, handleListChannelTasks, handleReassignTask, handleRefuseTask, handleSubmitTask, handleUpdateTask } from './tools/tasks'
+import { handleGrantNodes, handleRevokeNodes } from './tools/team-delegation'
+import { handleBroadcastMessage, handleListOtherTeams, handlePollMessages, handleReadChannelMail, handleSearchOtherTeamsMemory, handleSendCrossChannelMessage, handleSendMessageToAgent } from './tools/messaging'
+import { handleSaveMemory, handleSearchMemory } from './tools/memory'
+import { handleCreateTeamAgent, handleGetQueueOverview, handleListTeamAgents, handleRemoveTeamAgent, handleUpdateTeamAgent } from './tools/team'
+
+/**
+ * 分发一次 host tool 调用(全引擎唯一入口):
+ * 插件工具优先 → workspace 面 → 工业工具族 → 插件兜底 → 未知工具。
+ */
+/** 统一出口:所有工具结果经截断收口(单点修改,全族生效) */
+export async function dispatchHostTool(ctx: HostToolBridgeContext, req: HostToolCall): Promise<HostToolResult> {
+  return truncateToolResult(await dispatchHostToolInner(ctx, req), req.toolName)
+}
+
+/** 内置工具名(插件同名注册不参与分发:注入面已显示内置定义,执行面必须同源) */
+const BUILTIN_TOOL_NAMES = new Set(HOST_TOOLS.map(t => t.name))
+
+/** 工具结果统一出口上限:超长文本截断+尾注(2026-10-05 评审;export/diag 类可经 optOut 放行) */
+const TOOL_RESULT_MAX_CHARS = 16_000
+const TRUNCATE_OPT_OUT_TOOLS = new Set(['daq_export', 'diag_run'])
+function truncateToolResult(result: HostToolResult, toolName: string): HostToolResult {
+  if (TRUNCATE_OPT_OUT_TOOLS.has(toolName)) return result
+  const text = result.text ?? ''
+  if (text.length <= TOOL_RESULT_MAX_CHARS) return result
+  const head = text.slice(0, TOOL_RESULT_MAX_CHARS)
+  return {
+    ...result,
+    text: head + `
+[truncated: ${text.length} → ${TOOL_RESULT_MAX_CHARS} 字符;请收窄时间窗/减少点位/分页后重试]`,
+  }
+}
+
+async function dispatchHostToolInner(ctx: HostToolBridgeContext, req: HostToolCall): Promise<HostToolResult> {
+  const identity = ctx.identity
+  const state = ctx.state
+  // §13.3 人类发起者作用域:进程内工具调用同样受管理面/高危写限制。
+  // REST/MCP 路径已在 manager.invokeHostTool 检查;omp 的 host_tool_call 只经过本函数,
+  // 此前完全绕过(群成员 @Agent 可借 Agent 身份执行高危写)。
+  {
+    const g = globalThis as typeof globalThis & {
+      __workshopManager?: { invocationScopeOf?(agentId: string): WorkshopPermissionScope | null }
+    }
+    const scope = g.__workshopManager?.invocationScopeOf?.(identity.agentId) ?? null
+    const verdict = checkToolAgainstScope(scope, req.toolName)
+    if (!verdict.allowed) return { text: verdict.message, isError: true }
+  }
+  // Channel 逐工具选择:执行面复核(目录可能被引擎缓存;不能只靠注入面隐藏)。
+  // 未知工具不在此拦截,保持既有“未知工具”错误语义。
+  {
+    const gate = channelToolGate(identity.channelId, req.toolName)
+    if (gate.blocked) return { text: gate.message ?? `工具 ${req.toolName} 不可用`, isError: true }
+  }
+  // 插件工具分发(ctx.omp.registerTool 注册的自定义工具;不依赖 workspace,优先于内置面)。
+  // 同名遮蔽内置工具不生效:否则插件 handler 会绕过 grant/verify/HITL 全部门控,
+  // 而注入面仍展示内置 schema(目录与执行语义分裂)。
+  const pluginTool = BUILTIN_TOOL_NAMES.has(req.toolName) ? undefined : listPluginTools().get(req.toolName)
+  if (pluginTool) {
+    // 团队级插件开关:显式关闭的插件,其工具在该团队拒绝执行(与注入过滤同源)
+    const channelOff = getChannelPluginsRepo().explicitFor(identity.channelId)
+    const owner = pluginOfTool(req.toolName)
+    if (channelOff && owner && channelOff.get(owner) === false) {
+      return {
+        text: `该团队未启用插件「${owner}」,${req.toolName} 不可用。可在团队设置→插件中开启。`,
+        isError: true,
+      }
+    }
+    // roles 是执行边界而非仅注入提示:worker 直调 lead 专属插件工具同样拒绝
+    if (pluginTool.roles && !pluginTool.roles.includes(identity.role)) {
+      return {
+        text: `工具 ${req.toolName} 仅对 ${pluginTool.roles.join('/')} 角色开放(当前 ${identity.role})。`,
+        isError: true,
+      }
+    }
+    try {
+      return await pluginTool.handler(req.arguments ?? {}, identity)
+    }
+    catch (err) {
+      return {
+        text: `工具执行异常(${req.toolName}): ${err instanceof Error ? err.message : String(err)}`,
+        isError: true,
+      }
+    }
+  }
+  const args = req.arguments ?? {}
+
+  // AML 解耦门控(fail-closed,与注入过滤同源;目录隐藏不是安全边界):
+  //  - legacy:全部 twin/explore 工具拒绝(原始行为);
+  //  - trial/mpc/bayes:仅 hybrid_twin,或 aml_optimization 的 aml 模式(已绑定模型);
+  //    探索模式给「探索阶段」明确指引而非裸报错;
+  //  - optimization_explore:仅工艺优化 Channel;
+  //  - 场景/spec/校准请求族:hybrid_twin + 训练 Channel;
+  //  - 训练族(aml_job/dataset 等):工艺优化 Channel 拒绝(边训边优被结构禁止)。
+  const twinKind: AmlChannelProfileKind = getChannelTwinProfile(identity.channelId).profile
+  const isTwinTool = req.toolName.startsWith('twin_') || ['mpc_optimize', 'optimization_explore'].includes(req.toolName)
+  if (isTwinTool) {
+    const isSceneFamily = /^twin_(scene|physics|provider|calibration)/.test(req.toolName)
+    if (req.toolName === 'optimization_explore') {
+      if (twinKind !== 'aml_optimization') {
+        return { text: `optimization_explore 仅适用于工艺优化 Channel(当前 profile=${twinKind})。`, isError: true }
+      }
+    }
+    else if (AML_MODEL_BACKED_TOOL_NAMES.has(req.toolName)) {
+      const policy = modelBackedToolPolicyFor(identity.channelId)
+      if (!policy.allowed) {
+        return {
+          text: policy.reason === 'EXPLORATION_MODE_NO_MODEL'
+            ? `当前为探索模式:Channel 尚未绑定 AML 模型,${req.toolName}(孪生验证/贝叶斯寻优)暂不可用。\n请继续用 optimization_explore 做真实激励探索并积累数据;待 AML 训练出合格模型并在 Channel 设置中绑定后,本工具自动启用。`
+            : `${req.toolName} 要求模型受控状态(hybrid_twin 或 aml_optimization 的 aml 模式;当前 ${twinKind}${policy.reason ? `/${policy.reason}` : ''})。`,
+          isError: true,
+        }
+      }
+    }
+    else if (isSceneFamily) {
+      if (twinKind !== 'hybrid_twin' && twinKind !== 'aml_training') {
+        return { text: `场景/物理 spec 工具仅对 hybrid_twin 或训练 Channel(aml_training)开放(当前 profile=${twinKind})。`, isError: true }
+      }
+    }
+    else if (twinKind === 'legacy') {
+      return { text: `Channel ${identity.channelId} 未启用 Twin profile，Twin/MPC 工具拒绝执行。`, isError: true }
+    }
+  }
+  if (AML_TRAINING_TOOL_NAMES.has(req.toolName) && twinKind === 'aml_optimization') {
+    return { text: `工艺优化 Channel 不承担建模与训练(职责解耦):${req.toolName} 仅限 AML 训练 Channel 使用。请在此专注 goal 探索与模型驱动寻优。`, isError: true }
+  }
+
+  // 所有真实写入口在统一分发层再做一次服务端 fail-closed 守卫；工具目录隐藏不是安全边界。
+  if (['dcw_control', 'param_control', 'dcw_rollback', 'recipe_create', 'recipe_update', 'recipe_rollback', 'recipe_trial', 'recipe_apply'].includes(req.toolName)) {
+    // Profile 级写面(2026-10-18):观察面零写、训练面无产线写、优化面只认 optimization_explore。
+    // 此前这些 Channel 只是目录不注入,按名直调(REST/MCP)仍可落到写 handler。
+    if (twinKind === 'exp-miner' || twinKind === 'line-doctor') {
+      return { text: `${req.toolName} 在产线观察面(exp-miner/line-doctor)不可用:该档位物理零写。请改用只读分析工具;若确需改参,请由用户在具备写权限的 Channel 中操作。`, isError: true }
+    }
+    if (twinKind === 'aml_training') {
+      return { text: `${req.toolName} 在 AML 训练 Channel 不可用:训练面不承担产线写入。请把执行放到工艺优化 Channel(optimization_explore 治理链)或由用户手动操作。`, isError: true }
+    }
+    if (twinKind === 'aml_optimization') {
+      return { text: `${req.toolName} 在工艺优化 Channel 不开放:真实写入统一走 optimization_explore 治理链(绑定/HITL/限界/回读同源)。请改用 optimization_explore。`, isError: true }
+    }
+    const profile = getChannelTwinProfile(identity.channelId)
+    const guard = twinWriteGuard(identity.channelId, profile)
+    if (!guard.allowed) return { text: `${guard.code}: ${guard.message}`, isError: true }
+  }
+
+  // 工业工具族不依赖 workspace(只按 agentId 查绑定与节点),先于 workspace 门控执行 ——
+  // 否则 worker 首回合前的 REST/MCP 直调(my_industrial_nodes 等)会被误拒
+  if (INDUSTRIAL_TOOL_NAMES.has(req.toolName)) {
+    // 权限模型 v3:运行时 grant 复核(收权即失活;绑定面是静态授权,撤销后此处拦截)
+    const denied = await guardIndustrialToolGrant(identity, req.toolName, args)
+    if (denied) return denied
+    // 统一预检:参数形状 + 引用归属(node_id/line_id/recipe_id/dataset_id… 必须落在
+    // 调用者的绑定面内;越界拒绝并把「你的绑定面 + 下一步」返回给 Agent)
+    const verifyDenied = verifyIndustrialToolCall(identity.agentId, req.toolName, args)
+    if (verifyDenied) return verifyDenied
+    return await dispatchIndustrialTool(identity.agentId, req.toolName, args, identity.channelId)
+  }
+
+  const ws = ctx.getWorkspace()
+  if (!ws) {
+    return { text: 'workspace 未就绪', isError: true }
+  }
+
+  try {
+    switch (req.toolName) {
+      case 'report_progress':
+        return await handleReportProgress(args, state, ws)
+      case 'complete_task':
+        return await handleCompleteTask(args, state, ws)
+      case 'submit_task':
+        return await handleSubmitTask(args, state, ws)
+      case 'dispatch_task':
+        return await handleDispatchTask(args, ws, identity)
+      case 'team_grant_nodes':
+        return await handleGrantNodes(identity, args, ws)
+      case 'team_revoke_nodes':
+        return await handleRevokeNodes(identity, args, ws)
+      case 'send_message_to_agent':
+        return await handleSendMessageToAgent(args, state, ws)
+      case 'send_cross_channel_message':
+        return await handleSendCrossChannelMessage(args, ws)
+      case 'list_other_teams':
+        return await handleListOtherTeams(ws)
+      case 'search_other_teams_memory':
+        return await handleSearchOtherTeamsMemory(args, ws)
+      case 'refuse_task':
+        return await handleRefuseTask(args, ws)
+      case 'poll_messages':
+        return await handlePollMessages(args, state, ws)
+      case 'read_channel_mail':
+        return await handleReadChannelMail(args, ws)
+      case 'broadcast_message':
+        return await handleBroadcastMessage(args, identity, ws)
+      case 'list_channel_tasks':
+        return await handleListChannelTasks(ws)
+      case 'get_my_task_queue':
+        return await handleGetMyTaskQueue(identity, ws)
+      case 'get_queue_overview':
+        return await handleGetQueueOverview(ws)
+      case 'reassign_task':
+        return await handleReassignTask(args, ws)
+      case 'update_task':
+        return await handleUpdateTask(args, ws)
+      case 'cancel_task':
+        return await handleCancelTask(args, ws)
+      case 'list_team_agents':
+        return await handleListTeamAgents(ws)
+      case 'get_task_details':
+        return await handleGetTaskDetails(args, ws)
+      case 'search_memory':
+        return await handleSearchMemory(args, ws)
+      case 'save_memory':
+        return await handleSaveMemory(args, ws)
+      case 'create_team_agent':
+        return await handleCreateTeamAgent(args, ws)
+      case 'update_team_agent':
+        return await handleUpdateTeamAgent(args, ws)
+      case 'remove_team_agent':
+        return await handleRemoveTeamAgent(args, ws)
+
+      // 工业工具族的名册(实现在 tools/industrial.ts):上面 workspace 门控前已分流,
+      // 这里保留同名路由标签,与原分发表逐条对齐(不可达,行为与原实现一致)。
+      case 'my_industrial_nodes':
+      case 'aml_node_catalog':
+      case 'aml_dataset_build':
+      case 'aml_dataset_stats':
+      case 'aml_job_submit':
+      case 'aml_job_status':
+      case 'aml_job_logs':
+      case 'aml_job_cancel':
+      case 'aml_leaderboard':
+      case 'aml_model_find':
+      case 'aml_model_promote':
+      case 'aml_model_reference':
+      case 'twin_scene_read':
+      case 'twin_snapshot_create':
+      case 'twin_trial_run':
+      case 'mpc_optimize':
+      case 'twin_gate_evaluate':
+      case 'twin_calibration_request':
+      case 'dcw_control':
+      case 'param_control':
+      case 'daq_query':
+      case 'daq_frames':
+      case 'daq_export':
+      case 'dcw_judge':
+      case 'dcw_rollback':
+      case 'dcw_journal':
+      case 'ops_log':
+      case 'line_log':
+      case 'recipe_log':
+      case 'line_context':
+      case 'recipe_versions':
+      case 'recipe_update':
+      case 'recipe_rollback':
+        return await dispatchIndustrialTool(identity.agentId, req.toolName, args, identity.channelId)
+    }
+    return { text: `未知工具: ${req.toolName}`, isError: true }
+  }
+  catch (err) {
+    return {
+      text: `工具执行异常(${req.toolName}): ${err instanceof Error ? err.message : String(err)}`,
+      isError: true,
+    }
+  }
+}

package/server/services/workshop/agents/host-tool-bridge/index.ts

@@ -1,15 +1,16 @@
-/**
- * server/services/workshop/agents/host-tool-bridge.ts 拆分后的门面(目录 index 解析,
- * 外部 import 路径与导出 API 完全不变)。
- *
- * 结构:
- *   types.ts     HostToolCall / HostToolResult / HostToolSessionState / HostToolBridgeContext
- *   session.ts   会话态 createSessionState 与 partsToText
- *   catalog.ts   HOST_TOOLS / LEAD_ONLY_TOOL_NAMES / 占位符注入 / hostToolsForRole
- *   dispatch.ts  dispatchHostTool(前置校验 + 路由)
- *   tools/       按工具族拆分的处理器:industrial / progress / tasks / messaging / memory / team
- */
-export { HOST_TOOLS, LEAD_ONLY_TOOL_NAMES, hostToolsForRole } from './catalog'
-export type { HostToolBridgeContext, HostToolCall, HostToolResult, HostToolSessionState } from './types'
-export { createSessionState, partsToText } from './session'
-export { dispatchHostTool } from './dispatch'
+/**
+ * server/services/workshop/agents/host-tool-bridge.ts 拆分后的门面(目录 index 解析,
+ * 外部 import 路径与导出 API 完全不变)。
+ *
+ * 结构:
+ *   types.ts     HostToolCall / HostToolResult / HostToolSessionState / HostToolBridgeContext
+ *   session.ts   会话态 createSessionState 与 partsToText
+ *   catalog.ts   HOST_TOOLS / LEAD_ONLY_TOOL_NAMES / 占位符注入 / hostToolsForRole
+ *   dispatch.ts  dispatchHostTool(前置校验 + 路由)
+ *   tools/       按工具族拆分的处理器:industrial / progress / tasks / messaging / memory / team
+ */
+export { HOST_TOOLS, LEAD_ONLY_TOOL_NAMES, hostToolsForRole } from './catalog'
+export { hostToolsForAgent, agentToolResources } from './per-agent'
+export type { HostToolBridgeContext, HostToolCall, HostToolResult, HostToolSessionState } from './types'
+export { createSessionState, partsToText } from './session'
+export { dispatchHostTool } from './dispatch'