agentworkshop 0.7.24 → 0.7.33

package/server/services/workshop/agents/omp-agent.ts

@@ -1,1178 +1,1193 @@
-/**
- * OmpRpcAgentImpl — omp harness 的完整 AgentInterface 实现。
- *
- * 通过 OmpRpcClient 驱动真实 omp 子进程(--mode rpc),将 omp 的原生 agent 能力
- * (LLM 推理 + 内置工具 read/write/edit/bash/grep/glob 等)接入 Workshop 的
- * Channel 编排体系。
- *
- * 核心设计:
- *  - 每个 Agent 实例持有一个 omp 子进程(lazy spawn,跨消息复用)
- *  - AgentWorkspace 方法经 host-tool-bridge 注册为 omp host tools,agent 原生调用
- *    report_progress / complete_task / dispatch_task / send_message 等,无需文本解析
- *  - omp AgentSessionEvent → AgentEvent 五变体映射
- *  - worker run():接收 assign → prompt omp → agent 用原生工具作业 + host tools 管理任务生命周期
- *  - lead supervise():格式化快照 → prompt omp → agent 直接用 host tools 执行调度决策 → 返回 []
- *  - dispose():杀子进程
- *
- * prompt 组装与工具分发已抽至共享层(prompt-builder / host-tool-bridge),
- * 与 codex/dsh/opencode impl 全引擎一致。协议权威:omp://rpc.md。
- */
-import { createLogger } from '../logger'
-import { randomUUID } from 'node:crypto'
-import type {
-  AgentEvent,
-  AgentInterface,
-  AgentInfo,
-  AgentRunContext,
-  AgentRunRequest,
-  SupervisionDecision,
-  SupervisionSnapshot,
-} from './agent-interface'
-import type { A2AMessage } from '../types/a2a'
-import type { AgentContextStats } from '../types/task'
-import {
-  OmpRpcClient,
-  type AgentSessionEvent,
-  type CompactionResult,
-  type HostToolCallRequest,
-} from './adapters/omp-rpc-client'
-import {
-  registerHarnessProcess,
-  bindHarnessProcess,
-  markHarnessProcessExit,
-  killHarnessProcess,
-} from './harness-process'
-import {
-  attachTerminalTap,
-  markTerminalSessionExit,
-} from './harness-terminal'
-import {
-  createSessionState,
-  dispatchHostTool,
-  hostToolsForRole,
-} from './host-tool-bridge'
-import {
-  contextPrefix as buildContextPrefix,
-  createRosterCache,
-  extractJsonArray,
-  parseSteerBanner,
-  peerPrompt,
-  supervisePrompt,
-  systemManual,
-  toolArgsPreview,
-  workerPrompt,
-} from './prompt-builder'
-import { renderPrompt } from '../prompts/loader'
-import { attachOmpPluginBridge, onPluginToolsChange } from './plugin-tools'
-import { ompSettings } from '../settings'
-
-const log = createLogger('workshop.omp')
-
-// 兼容再导出(脚本/测试历史导入面;实现在 host-tool-bridge)
-export { HOST_TOOLS, hostToolsForRole } from './host-tool-bridge'
-
-// ===== 配置 =====
-
-export interface OmpAgentConfig {
-  /** omp 可执行文件路径(默认 'omp',从 PATH 查找) */
-  command?: string
-  /** 额外 CLI 参数 */
-  args?: string[]
-  /** omp 工作目录(默认 process.cwd()) */
-  cwd?: string
-  /** 模型 provider(如 'anthropic'/'openai'/'zhipu');省略则用 omp 默认 */
-  provider?: string
-  /** 模型 ID;省略则用 omp 默认 */
-  model?: string
-  /** thinking level: off/minimal/low/medium/high/xhigh/max */
-  thinkingLevel?: string
-  /** 自定义系统 prompt 前缀(拼接到 agent 系统指令之前) */
-  systemPromptPrefix?: string
-  /** 限制 omp 仅使用指定内置工具(null = 全部) */
-  toolNames?: string[]
-  /** 每轮停滞超时(ms,默认 300000):整轮无任何 omp 事件才中止;工具内阻塞(poll_messages ≤180s)有 tool 事件刷新计时 */
-  promptTimeoutMs?: number
-  /** supervise 轮超时(ms,默认 150000) */
-  superviseTimeoutMs?: number
-  /**
-   * omp 输出模式:'rpc-ui'(默认)= RPC 协议 + UI 上下文 —— ask 工具与
-   * extension_ui_request 对话框可用(监控终端 HITL 通道,强制 noPty);
-   * 'rpc' = 纯协议(无 UI,ask 不可用)。
-   */
-  rpcMode?: 'rpc' | 'rpc-ui'
-  /** channel 级作业场景 prompt(manager 装配时注入;用户场景规范,全员共享) */
-  scenarioPrompt?: string
-  /** agent 身份(由 factory 从 AgentInfo 注入) */
-  agentId?: string
-  name?: string
-  role?: 'lead' | 'worker'
-  channelId?: string
-}
-
-// 插件桥接管(host.mjs ctx.omp 排队注册的工具在此回放;幂等)
-attachOmpPluginBridge()
-
-/** 在跑 agent 实例表(工具注册表变更 → 热重发 set_host_tools,运行时注入无需重spawn) */
-const gLive = globalThis as typeof globalThis & { __ompLiveAgents?: Set<OmpRpcAgentImpl> }
-function liveAgents(): Set<OmpRpcAgentImpl> {
-  return gLive.__ompLiveAgents ??= new Set()
-}
-
-// 工具注册表变更 → 全部在跑会话热更新工具面
-onPluginToolsChange(() => {
-  for (const impl of liveAgents()) {
-    try {
-      impl.refreshPluginTools()
-    }
-    catch { /* 单实例失败不影响其他 */ }
-  }
-})
-
-// ===== OmpRpcAgentImpl =====
-
-export class OmpRpcAgentImpl implements AgentInterface {
-  private readonly config: OmpAgentConfig
-  private client: OmpRpcClient | null = null
-  private workspace: AgentRunContext['workspace'] | null = null
-  private agentInfo: AgentInfo | null = null
-  private hostToolsRegistered = false
-  /** host tool 会话态(当前任务 + 待回执上下文;与共享桥共用) */
-  private readonly toolState = createSessionState()
-  /** 工具桥上下文(分发唯一入口) */
-  private readonly bridgeCtx = {
-    identity: { agentId: '', channelId: '', role: 'worker' as 'lead' | 'worker', name: 'agent' },
-    state: this.toolState,
-    getWorkspace: () => this.workspace,
-  }
-
-  /**
-   * 会话回合状态(steer 可靠注入的依据):
-   * omp 的 steer 仅在回合 streaming 中生效——prompt 已入列但尚未开始输出时,
-   * steer 会"成功"返回但被静默丢弃。因此注入方轮询直到回合输出开始(streamingStarted)
-   * 才发送;若回合在等待期间结束,消息保持 pending 由消费循环处理。
-   */
-  private streaming = false
-  private turnActive = false
-  /** 当前回合产生的 assistant 文本(供诊断) */
-  private turnText = ''
-  /** agent 身份信息(factory 注入;无需等待 init()) */
-  private selfAgentId = ''
-  private agentName = 'agent'
-  private agentRole: 'lead' | 'worker' = 'worker'
-  private channelId = ''
-  /** 团队名册缓存(共享 prompt-builder) */
-  private roster = createRosterCache({
-    selfAgentId: '',
-    listAgents: () => this.workspace!.listAgents(),
-  })
-
-  // ===== 上下文治理(70% 无中断压缩环)=====
-  /** 压缩进行中(平台 gate 与 omp 原生压缩共用互斥位) */
-  private compacting = false
-  private lastCompactAt = 0
-  /** omp 不支持 compact 命令(旧版)→ 永久停用平台压缩;原生 auto-compaction 兜底,harvest 照常 */
-  private compactLegacy = false
-  /** 模型上下文窗口(get_state/get_session_stats 探测;null = 未知,percent 不可算) */
-  private contextWindow: number | null = null
-  private sessionId: string | null = null
-  /** harvest 双路去重(compact 响应与 compaction_end 事件可能双达) */
-  private lastHarvestKey = ''
-  private lastHarvestAt = 0
-
-  /** 总开关(omp.compact_enabled;env AW_OMP_COMPACT_ENABLED=0 显式关闭兼容) */
-  private compactEnabled(): boolean {
-    return ompSettings().compact_enabled
-  }
-
-  private compactThreshold(): number {
-    return ompSettings().compact_threshold
-  }
-
-  private compactMinIntervalMs(): number {
-    return ompSettings().compact_min_interval_ms
-  }
-
-  private compactWaitMs(): number {
-    return ompSettings().compact_wait_ms
-  }
-
-  /** 被动上下文用量快照(无探测 RPC;getStatus 透出用) */
-  getContextStats(): AgentContextStats | null {
-    const usage = this.client?.getContextUsage() ?? null
-    const window = this.contextWindow ?? usage?.contextWindow ?? null
-    if (!usage && window === null) return null
-    const usedTokens = usage?.tokens ?? 0
-    return {
-      usedTokens,
-      contextWindow: window,
-      percent: window && window > 0 ? Math.min(1, usedTokens / window) : null,
-      compacting: this.compacting,
-    }
-  }
-
-  /**
-   * 上下文用量探测:get_session_stats 权威值;不可用退化为被动 usage + 已探测窗口。
-   * percent 归一化到 0-1(omp 可能返回 0-100)。
-   */
-  private async probeUsage(client: OmpRpcClient): Promise<{ tokens: number, contextWindow: number | null, percent: number | null } | null> {
-    try {
-      const resp = await client.send({ type: 'get_session_stats' })
-      const cu = ((resp.data ?? {}) as Record<string, unknown>).contextUsage as
-        | { tokens?: number, contextWindow?: number, percent?: number }
-        | undefined
-      if (cu && typeof cu.tokens === 'number' && cu.tokens > 0) {
-        if (typeof cu.contextWindow === 'number' && cu.contextWindow > 0) {
-          this.contextWindow = cu.contextWindow
-          client.setContextWindow(cu.contextWindow)
-        }
-        const window = this.contextWindow ?? cu.contextWindow ?? null
-        const rawPercent = typeof cu.percent === 'number' ? (cu.percent > 1 ? cu.percent / 100 : cu.percent) : null
-        const percent = rawPercent ?? (window && window > 0 ? Math.min(1, cu.tokens / window) : null)
-        return { tokens: cu.tokens, contextWindow: window, percent }
-      }
-    }
-    catch { /* get_session_stats 不可用 → 被动跟踪 */ }
-    return client.getContextUsage()
-  }
-
-  /**
-   * 上下文治理门控:三条 prompt 路径(worker/peer/supervise)在回合间隙统一调用;
-   * post-settle 经 onTurnSettled 复用同一守卫。任何失败路径都放行——压缩晚一轮
-   * 永远好过中断作业。三重防线:仅回合间隙发起 → get_state 复查 isStreaming/
-   * isCompacting → compacting 互斥位 + 最小间隔防振荡。
-   */
-  private async contextGate(reason: string): Promise<void> {
-    if (!this.compactEnabled() || this.compactLegacy || this.compacting) return
-    const client = this.client
-    if (!client || !client.alive) return
-    try {
-      const usage = await this.probeUsage(client)
-      if (!usage || usage.percent === null || usage.percent < this.compactThreshold()) return
-      if (Date.now() - this.lastCompactAt < this.compactMinIntervalMs()) return
-      // 回合间隙硬校验:流式中/压缩中绝不发起(不中断执行的硬约束)
-      const st = await client.send({ type: 'get_state' })
-      const s = (st.data ?? {}) as Record<string, unknown>
-      if (s.isStreaming === true || s.isCompacting === true) return
-      if (typeof s.sessionId === 'string' && s.sessionId) this.sessionId = s.sessionId
-      this.compacting = true
-      this.lastCompactAt = Date.now()
-      try {
-        const result = await this.runCompaction(client)
-        if (result?.summary) {
-          await this.harvestCompaction(
-            result.summary,
-            result.tokensBefore,
-            result.tokensAfter ?? result.estimatedTokensAfter,
-            reason,
-          )
-        }
-      }
-      finally {
-        this.compacting = false
-      }
-    }
-    catch (err) {
-      const msg = err instanceof Error ? err.message : String(err)
-      if (/compact/i.test(msg) && /unknown|unsupported|invalid|无效|未知|不支持/i.test(msg)) this.compactLegacy = true
-      log.warn(`[OmpRpcAgent:${this.selfAgentId}] contextGate(${reason}) 失败,放行: ${msg}`)
-    }
-  }
-
-  /**
-   * 发起 compact 并等待 compaction_end(响应内含 result 则立即收口;
-   * 超时放行——压缩在 harness 内后台继续,isCompacting 防后续双发)。
-   */
-  private runCompaction(client: OmpRpcClient): Promise<CompactionResult | null> {
-    return new Promise<CompactionResult | null>((resolve) => {
-      let settled = false
-      const finish = (r: CompactionResult | null): void => {
-        if (settled) return
-        settled = true
-        off()
-        clearTimeout(timer)
-        resolve(r)
-      }
-      const off = client.onEvent((event) => {
-        if (event.type === 'compaction_end') {
-          finish((event as { result?: CompactionResult }).result ?? null)
-        }
-        else if (event.type === '__process_exit__') {
-          finish(null)
-        }
-      })
-      const timer = setTimeout(() => {
-        log.warn(`[OmpRpcAgent:${this.selfAgentId}] compact 等待超时(${this.compactWaitMs()}ms)→ 放行(压缩后台继续)`)
-        finish(null)
-      }, this.compactWaitMs())
-      void client.send({ type: 'compact', customInstructions: this.compactionHints() })
-        .then((resp) => {
-          const result = (resp.data as { result?: CompactionResult } | undefined)?.result
-          if (result?.summary) finish(result)
-          // 响应无 result:等 compaction_end 事件(协议权威语义)
-        })
-        .catch((err) => {
-          const msg = err instanceof Error ? err.message : String(err)
-          if (/compact/i.test(msg)) this.compactLegacy = true
-          log.warn(`[OmpRpcAgent:${this.selfAgentId}] compact 发起失败: ${msg}`)
-          finish(null)
-        })
-    })
-  }
-
-  /** 压缩摘要指令(外置 prompts;缺失时省略,omp 用默认策略) */
-  private compactionHints(): string | undefined {
-    try {
-      return renderPrompt('compaction-hints')
-    }
-    catch {
-      return undefined
-    }
-  }
-
-  /**
-   * 压缩摘要 harvest(三路统一:平台主动 compact / omp 阈值自动 / overflow):
-   * 经 workspace.recordSessionMemory 落库为本人 episodic-session 记忆。
-   * 双路去重:同摘要 60s 内只入库一次。
-   */
-  private async harvestCompaction(summary: string, tokensBefore?: number, tokensAfter?: number, reason?: string): Promise<void> {
-    const key = summary.slice(0, 120)
-    if (key === this.lastHarvestKey && Date.now() - this.lastHarvestAt < 60_000) return
-    this.lastHarvestKey = key
-    this.lastHarvestAt = Date.now()
-    try {
-      await this.workspace?.recordSessionMemory?.({ summary, tokensBefore, tokensAfter, reason })
-      log.info(`[OmpRpcAgent:${this.selfAgentId}] 压缩摘要已入记忆(${summary.length} 字,reason=${reason ?? 'auto'})`)
-    }
-    catch (err) {
-      log.warn(`[OmpRpcAgent:${this.selfAgentId}] 压缩摘要入库失败(不影响会话):`, err instanceof Error ? err.message : err)
-    }
-  }
-
-  /** 回合落定钩子(AgentRuntime 在信箱空时调用):后台压缩检查,自守卫不抛错 */
-  async onTurnSettled(): Promise<void> {
-    await this.contextGate('post-settle')
-  }
-
-  constructor(config: Record<string, unknown> = {}) {
-    this.config = config as OmpAgentConfig
-    // effort 统一入口(channel 默认 LLM 注入 config.effort)→ omp thinking 级别
-    if (!this.config.thinkingLevel && typeof config.effort === 'string' && config.effort) {
-      this.config.thinkingLevel = config.effort
-    }
-    // factory 注入的 agent 身份(无需等待 init())
-    this.selfAgentId = this.config.agentId ?? ''
-    this.agentName = this.config.name ?? 'agent'
-    this.agentRole = this.config.role ?? 'worker'
-    this.channelId = this.config.channelId ?? ''
-    this.bridgeCtx.identity = { agentId: this.selfAgentId, channelId: this.channelId, role: this.agentRole, name: this.agentName }
-    this.roster = createRosterCache({
-      selfAgentId: this.selfAgentId,
-      listAgents: () => this.workspace!.listAgents(),
-    })
-    liveAgents().add(this) // 在跑实例表:插件工具热注入的目标集
-  }
-
-  // ===== 生命周期 =====
-
-  async init(input: { agent: AgentInfo, channelId: string }): Promise<void> {
-    this.agentInfo = input.agent
-    this.channelId = input.channelId
-    this.agentName = input.agent.name
-    this.agentRole = input.agent.role
-    this.selfAgentId = input.agent.id
-    this.bridgeCtx.identity = { agentId: this.selfAgentId, channelId: this.channelId, role: this.agentRole, name: this.agentName }
-    this.roster = createRosterCache({
-      selfAgentId: this.selfAgentId,
-      listAgents: () => this.workspace!.listAgents(),
-    })
-  }
-
-  async dispose(): Promise<void> {
-    liveAgents().delete(this)
-    if (this.client) {
-      const pid = this.client.pid
-      await this.client.dispose()
-      this.client = null
-      if (pid) {
-        markHarnessProcessExit(pid, null)
-        markTerminalSessionExit(pid, null)
-      }
-    }
-    this.hostToolsRegistered = false
-  }
-
-  /** 工具注册表变更 → 热重发 set_host_tools(在跑会话立即获得插件新工具,无需重spawn) */
-  refreshPluginTools(): void {
-    const client = this.client
-    if (!client || !client.alive) return
-    void client.send({ type: 'set_host_tools', tools: hostToolsForRole(this.agentRole) }).catch(() => {})
-  }
-
-  /** harness 进程资源信息(运行时资源监控;进程未 spawn/已回收 → null) */
-  getProcessInfo(): { pid: number, alive: boolean, command: string } | null {
-    const client = this.client
-    const pid = client?.pid
-    if (!pid || !client) return null
-    return { pid, alive: client.alive, command: `omp --mode ${this.rpcMode}` }
-  }
-
-  /** 强制终止 harness 进程(进程树;终止后由调用方停止对应 AgentRuntime) */
-  killProcess(): void {
-    const pid = this.client?.pid
-    if (pid) {
-      markTerminalSessionExit(pid, null)
-      killHarnessProcess(pid)
-    }
-    else {
-      this.client?.kill()
-    }
-  }
-
-  /**
-   * harness 进程存活校准(manager sweeper 周期性调用;休眠/强杀后 exit 事件
-   * 可能不达父进程):按 PID 探 OS 实际存在性,进程已死 → 客户端收敛为已退出,
-   * 在途回合经 __process_exit__ 归位,下一回合 ensureClient 重生子进程。
-   */
-  reconcileProcess(): void {
-    this.client?.reconcile()
-  }
-
-  /** omp 输出模式(rpc-ui = 默认,启用 HITL 对话框) */
-  private get rpcMode(): 'rpc' | 'rpc-ui' {
-    return this.config.rpcMode === 'rpc' ? 'rpc' : 'rpc-ui'
-  }
-
-  // ===== 工具桥(共享分发)=====
-
-  /** 引擎无关工具面:REST/MCP 桥直调与 omp 内部 host_tool_call 共用同一实现 */
-  dispatchHostTool(toolName: string, args: Record<string, unknown>): Promise<{ text: string, isError?: boolean }> {
-    return dispatchHostTool(this.bridgeCtx, { toolName, arguments: args })
-  }
-
-  private async handleHostTool(req: HostToolCallRequest): Promise<{ text: string, isError?: boolean }> {
-    return this.dispatchHostTool(req.toolName, req.arguments ?? {})
-  }
-
-  // ===== prompt 组合 =====
-
-  /** 前置上下文(场景 × 身份 × 工业简报 × 名册;共享 prompt-builder) */
-  private async contextPrefix(): Promise<string> {
-    return buildContextPrefix({
-      scenarioPrompt: this.config.scenarioPrompt,
-      systemPromptPrefix: this.config.systemPromptPrefix,
-      agentId: this.selfAgentId,
-      roster: await this.roster.roster(),
-    })
-  }
-
-  private systemManual(): string {
-    return systemManual()
-  }
-
-  /**
-   * 实时消息注入(送达模式协议):
-   *  - 回合 streaming 中 → 立即 steer,返回 'steer'(同轮可见,唯一可标记消费的路径)
-   *  - 回合活跃但尚未 streaming(prompt 排队窗口,上限 20s)→ 等 streaming 开始后 steer
-   *  - 回合被 host 工具阻塞(如 poll_messages 等待)/空闲/发送失败 → 返回 'deferred':
-   *    消息保持 pending,poll_messages 的 Mailbox 到信回调即时取走(毫秒级),或本回合
-   *    结束后由消费循环按 FIFO 起回合处理。
-   */
-  async steer(text: string): Promise<'steer' | 'deferred'> {
-    // 从确定性触发横幅提取回执上下文(AgentRuntime.injectSteer 生成,格式固定):
-    // "[实时消息 from <id>]: ..." + "[系统触发器] 本消息要求回复(reply_to=<messageId>)。"
-    const banner = parseSteerBanner(text)
-    if (banner) {
-      this.toolState.replyContext = banner
-    }
-    const client = this.client
-    if (!client) return 'deferred'
-    try {
-      if (this.streaming) {
-        await client.send({ type: 'steer', message: text })
-        return 'steer'
-      }
-      if (this.turnActive) {
-        // prompt 排队窗口:等 streaming 开始(上限 20s);工具阻塞的回合等不到 → deferred
-        const deadline = Date.now() + 20_000
-        while (Date.now() < deadline && this.turnActive && !this.streaming) {
-          const { promise, resolve } = Promise.withResolvers()
-          setTimeout(resolve, 150)
-          await promise
-        }
-        if (this.streaming && this.turnActive) {
-          await client.send({ type: 'steer', message: text })
-          return 'steer'
-        }
-        return 'deferred'
-      }
-      // 空闲/回合已结束:deferred(消息 pending,消费循环 dequeue 即起回合)
-      return 'deferred'
-    }
-    catch (err) {
-      log.error(`[OmpRpcAgent:${this.selfAgentId}] steer 注入失败(消息保持 pending):`, err instanceof Error ? err.message : err)
-      return 'deferred'
-    }
-  }
-
-  async* run(request: AgentRunRequest, ctx: AgentRunContext): AsyncIterable<AgentEvent> {
-    const kind = request.message.metadata?.['x-aw-task-kind']
-
-    // worker: assign 消息 → 执行任务
-    if (kind === 'assign' && ctx.role === 'worker') {
-      yield* this.workerRun(request, ctx)
-      return
-    }
-
-    // worker/lead: 同事点对点消息(含实时通信触发器)→ 按触发器语义处理并回复;
-    // 人类经 REST 注入的消息(from 空 + x-aw-from-label)同样进入应答流
-    if (!kind && (request.fromAgentId || request.message.metadata?.['x-aw-from-label'])) {
-      yield* this.peerMessageRun(request, ctx)
-      return
-    }
-
-    // 其余消息(lead 的 assign/child-completed 等):no-op(调度由 supervise() 处理)
-  }
-
-  // ===== supervise() =====
-
-  /** supervise 单飞守卫:同一 client 不并发 LLM 回合(残留回合与下一 prompt 混流的根因) */
-  private supervising = false
-
-  async supervise(snapshot: SupervisionSnapshot, ctx: AgentRunContext, opts?: { signal?: AbortSignal }): Promise<SupervisionDecision[]> {
-    await this.ensureClient(ctx)
-    if (!this.client) return []
-    if (this.supervising) return [] // 上一轮 supervise 未收口:跳过本拍(节流即正确)
-
-    // 上下文门控(≥70% 先压缩再调度;回合间隙发起,失败放行)
-    await this.contextGate('supervise')
-
-    const prompt = supervisePrompt({
-      snapshot,
-      agentName: this.agentName,
-      channelId: this.channelId,
-      ctxPrefix: await this.contextPrefix(),
-      manual: this.systemManual(),
-      memory: ctx.memory,
-    })
-    // 150s 上界:supervise 持 lead.execLock 期间信箱消费停顿(更久会拖垮 worker 回执处理);
-    // supervise 是一次真实 LLM 回合:omp 冷启动(插件/MCP 加载 30~90s)+ 慢 provider
-    // 单步可能 >60s,过紧会把正常回合掐成 "Interrupted by user"。默认 150s,
-    // 仅拦真僵死(更紧的预算由调用方 config.superviseTimeoutMs 显式传入);真 abort 已实现。
-    const timeoutMs = this.config.superviseTimeoutMs ?? 150_000
-
-    return new Promise<SupervisionDecision[]>((resolve) => {
-      let assistantText = ''
-      let resolved = false
-
-      const finish = (decisions: SupervisionDecision[]) => {
-        if (resolved) return
-        resolved = true
-        this.supervising = false
-        unsub()
-        clearTimeout(timer)
-        signalUnsub?.()
-        resolve(decisions)
-      }
-
-      const abortTurn = (): void => {
-        // 超时/外部取消:真正中止 omp 当前回合 —— 只 resolve 不 abort 会让残留回合
-        // 与下一个 prompt 在同一 client 混流(决策错位 + token 空烧)
-        log.warn(`[OmpRpcAgent:${this.selfAgentId}] supervise 超时(${timeoutMs}ms)→ abort 当前调度回合`)
-        void this.client?.send({ type: 'abort' }).catch(() => {})
-        finish([])
-      }
-
-      const unsub = this.client!.onEvent((event) => {
-        if (event.type === 'message_update' && event.assistantMessageEvent?.type === 'text_delta') {
-          assistantText += event.assistantMessageEvent.delta ?? ''
-        }
-        if (event.type === 'agent_end' && event.isTerminal !== false) {
-          // 尝试从文本解析 JSON 决策(备用:如果 agent 没用 host tools 而是输出 JSON)
-          const parsed = extractJsonArray(assistantText)
-          if (parsed && parsed.length > 0) {
-            finish(parsed as SupervisionDecision[])
-          }
-          else {
-            // agent 可能已通过 host tools 直接执行了调度,返回空(已执行)
-            finish([])
-          }
-        }
-        if (event.type === '__process_exit__' || event.type === '__error__') {
-          finish([])
-        }
-      })
-
-      const timer = setTimeout(abortTurn, timeoutMs)
-
-      // 外部取消(调度器 cancel 路径)传导:abort 当前 LLM 回合并立即收口
-      let signalUnsub: (() => void) | undefined
-      if (opts?.signal) {
-        if (opts.signal.aborted) {
-          unsub()
-          resolve([])
-          return
-        }
-        const onAbort = (): void => abortTurn()
-        opts.signal.addEventListener('abort', onAbort, { once: true })
-        signalUnsub = () => opts.signal?.removeEventListener('abort', onAbort)
-      }
-
-      this.supervising = true
-      this.client!.send({ type: 'prompt', message: prompt }).catch(() => finish([]))
-    })
-  }
-
-  // ===== 内部:worker 执行 =====
-
-  private async* workerRun(request: AgentRunRequest, ctx: AgentRunContext): AsyncGenerator<AgentEvent, void, unknown> {
-    const taskId = request.taskId
-      ?? (request.message.metadata?.['x-aw-task-id'] as string | undefined)
-    if (!taskId) return
-
-    this.toolState.currentTaskId = taskId
-
-    try {
-      await this.ensureClient(ctx)
-    }
-    catch (err) {
-      yield {
-        kind: 'error',
-        error: {
-          code: 'OMP_SPAWN_FAILED',
-          message: `omp 子进程启动失败: ${err instanceof Error ? err.message : String(err)}`,
-        },
-      }
-      return
-    }
-
-    if (!this.client) {
-      yield { kind: 'error', error: { code: 'OMP_NOT_READY', message: 'omp 客户端未就绪' } }
-      return
-    }
-
-    // 构建 prompt
-    const taskText = request.message.parts.map((p) => {
-      if ('text' in p) return p.text
-      if ('data' in p) return JSON.stringify(p.data)
-      if ('url' in p) return p.url
-      if ('raw' in p) return p.raw
-      return ''
-    }).join('\n')
-    const prompt = workerPrompt({
-      agentName: this.agentName,
-      channelId: this.channelId,
-      taskId,
-      taskText,
-      memory: request.memory,
-      ctxPrefix: await this.contextPrefix(),
-      manual: this.systemManual(),
-    })
-
-    // 流式执行 + 事件映射
-    yield* this.promptAndStream(prompt, taskId, ctx.signal)
-  }
-
-  /**
-   * 点对点消息处理(实时通信驱动;lead 与 worker 通用)。
-   * 触发器语义:metadata['x-aw-require-reply']='true' → 必须经 send_message_to_agent
-   * 回给发送者:执行结果 + 对方所需内容,in_reply_to 关联原消息,并声明是否需再响应。
-   */
-  private async* peerMessageRun(request: AgentRunRequest, ctx: AgentRunContext): AsyncGenerator<AgentEvent, void, unknown> {
-    try {
-      await this.ensureClient(ctx)
-    }
-    catch (err) {
-      // spawn 失败必须显式报错:静默 return 会让 sawRunError=false → 消息被
-      // 直接 markConsumed,实时消息一次性丢失且无任何重试;产出 error 事件
-      // 走消息 requeue(≤2 次),与 workerRun 同口径
-      yield {
-        kind: 'error',
-        error: {
-          code: 'OMP_SPAWN_FAILED',
-          message: `omp 子进程启动失败(peer 消息未消费,将重试): ${err instanceof Error ? err.message : String(err)}`,
-        },
-      }
-      return
-    }
-    if (!this.client) {
-      yield { kind: 'error', error: { code: 'OMP_NOT_READY', message: 'omp 客户端未就绪(peer 消息未消费,将重试)' } }
-      return
-    }
-
-    const msg = request.message
-    // 显示名:agent 同事用 id(名册可解析);人类发送者用 fromLabel;兜底 unknown
-    const fromId = request.fromAgentId
-      ?? (typeof msg.metadata?.['x-aw-from-label'] === 'string' ? msg.metadata['x-aw-from-label'] : undefined)
-      ?? 'unknown'
-    const msgText = msg.parts.map((p) => {
-      if ('text' in p) return p.text
-      if ('data' in p) return JSON.stringify(p.data)
-      if ('url' in p) return p.url
-      if ('raw' in p) return p.raw
-      return ''
-    }).join('\n')
-    const requireReply = msg.metadata?.['x-aw-require-reply'] === 'true'
-    const isReply = typeof msg.metadata?.['x-aw-in-reply-to'] === 'string'
-    // 跨 Channel 来信:发送方是其他 channel 的 lead,回执必须走 send_cross_channel_message
-    const crossChannel = msg.metadata?.['x-aw-cross-channel'] === 'true'
-    const fromChannel = typeof msg.metadata?.['x-aw-from-channel'] === 'string'
-      ? String(msg.metadata['x-aw-from-channel'])
-      : ''
-    // 回执自动关联仅对真实 agent 发送者生效(人类无 agentId 可回投;
-    // 对人类的回复经事件流/时间线可见)
-    this.toolState.replyContext = requireReply && request.fromAgentId && !crossChannel
-      ? { fromId: request.fromAgentId, messageId: msg.messageId }
-      : null
-
-    const prompt = peerPrompt({
-      agentName: this.agentName,
-      role: this.agentRole,
-      channelId: this.channelId,
-      ctxPrefix: await this.contextPrefix(),
-      manual: this.systemManual(),
-      memory: request.memory,
-      fromId,
-      messageId: msg.messageId,
-      requireReply,
-      isReply,
-      crossChannel,
-      fromChannel,
-      msgText,
-    })
-
-    yield* this.promptAndStream(prompt, undefined, ctx.signal)
-  }
-
-  // ===== 内部:prompt 发送 + 事件流桥接 =====
-
-  private async* promptAndStream(
-    prompt: string,
-    taskId: string | undefined,
-    signal: AbortSignal,
-  ): AsyncGenerator<AgentEvent, void, unknown> {
-    const client = this.client
-    if (!client) {
-      throw new Error('omp 客户端未就绪')
-    }
-    // 上下文门控(≥70% 时在回合间隙先压缩再发 prompt;等待有界,失败放行)
-    await this.contextGate('pre-prompt')
-    const queue: AgentEvent[] = []
-    let resolveWait: (() => void) | null = null
-    let isDone = false
-    // 回合生命周期:prompt 已发出 → turnActive;首条 message_update → streaming;message_end/turn_end → 结束
-    this.turnActive = true
-    this.streaming = false
-    this.turnText = ''
-    // 正文流式差额账本:contentIndex → 已流出长度。部分 provider(reasoning 模型/
-    // openai-completions)的 text 块不走 text_delta(只有换行),全文在 text_end.content
-    // 落定 —— 差额兜底把未流出部分补发为 delta,保证前端流式不缺正文。
-    const textSent = new Map<number, number>()
-    const unsubState = client.onEvent((event) => {
-      if (event.type === 'message_update') {
-        const ev = event.assistantMessageEvent as
-          | { type?: string, delta?: string, content?: string, contentIndex?: number }
-          | undefined
-        if (ev?.type === 'text_delta') {
-          this.streaming = true
-          const text = ev.delta ?? ''
-          const ci = typeof ev.contentIndex === 'number' ? ev.contentIndex : 0
-          textSent.set(ci, (textSent.get(ci) ?? 0) + text.length)
-          this.turnText += text
-          // LLM 流式增量透出(AEP agent.delta 事件源):仅 worker/peer 转本走生成器
-          if (text) enqueueDelta(text)
-        }
-        else if (ev?.type === 'text_end' && typeof ev.content === 'string') {
-          const ci = typeof ev.contentIndex === 'number' ? ev.contentIndex : 0
-          const sent = textSent.get(ci) ?? 0
-          if (ev.content.length > sent) {
-            const chunk = ev.content.slice(sent)
-            textSent.set(ci, ev.content.length)
-            this.streaming = true
-            this.turnText += chunk
-            enqueueDelta(chunk)
-          }
-        }
-      }
-      if (event.type === 'message_end' || event.type === 'turn_end') {
-        this.streaming = false
-        flushDelta()
-      }
-      if (event.type === 'agent_end' && event.isTerminal !== false) {
-        this.turnActive = false
-        this.streaming = false
-        flushDelta()
-      }
-      if (event.type === '__process_exit__' || event.type === '__error__') {
-        this.turnActive = false
-        this.streaming = false
-      }
-    })
-
-    const enqueue = (event: AgentEvent): void => {
-      queue.push(event)
-      if (event.kind === 'done' || event.kind === 'error') isDone = true
-      lastActivity = Date.now()
-      resolveWait?.()
-      resolveWait = null
-    }
-
-    // 停滞看门狗:整轮无任何 omp 事件(挂死的 LLM 调用/子进程僵死)时中止回合,
-    // 否则消息永久卡 consuming、队友消息无限堆积。host 工具内阻塞
-    // (poll_messages 最长 180s)与慢 provider(实测单步可 >5min)都有
-    // tool/状态事件刷新计时;默认 600s 覆盖慢回合,只拦真僵死。
-    const idleTimeoutMs = this.config.promptTimeoutMs ?? 600_000
-    let lastActivity = Date.now()
-    let stallTimer: ReturnType<typeof setTimeout> | null = null
-
-    // LLM 流式增量缓冲:50ms 批量合并为一帧 delta(防高频 text_delta 洪泛 WS)
-    let deltaBuf = ''
-    let deltaTimer: ReturnType<typeof setTimeout> | null = null
-    const flushDelta = (): void => {
-      if (deltaTimer) {
-        clearTimeout(deltaTimer)
-        deltaTimer = null
-      }
-      if (!deltaBuf) return
-      const text = deltaBuf
-      deltaBuf = ''
-      enqueue({ kind: 'delta', delta: { text } })
-    }
-    const enqueueDelta = (text: string): void => {
-      deltaBuf += text
-      deltaTimer ??= setTimeout(() => {
-        deltaTimer = null
-        flushDelta()
-      }, 50)
-    }
-
-    // 订阅 omp 事件流
-    const unsub = client.onEvent((event) => {
-      for (const mapped of this.mapOmpEvent(event, taskId)) {
-        enqueue(mapped)
-      }
-    })
-
-    // abort 传导
-    const onAbort = (): void => {
-      log.warn(`[OmpRpcAgent:${this.selfAgentId}] run 被 abort(signal)→ 中止 omp 回合,taskId=${taskId ?? '-'}`)
-      client.send({ type: 'abort' }).catch(() => {})
-      if (!isDone) {
-        enqueue({ kind: 'done', final: taskId ? { taskId } : undefined })
-      }
-    }
-    if (signal.aborted) {
-      onAbort()
-    }
-    else {
-      signal.addEventListener('abort', onAbort, { once: true })
-    }
-
-    // 发送 prompt
-    try {
-      await client.send({ type: 'prompt', message: prompt })
-    }
-    catch (err) {
-      unsub()
-      signal.removeEventListener('abort', onAbort)
-      // 僵死进程回收:prompt send 失败 = stdio 断裂或 60s 命令确认超时 —— 进程
-      // "活着但不干活"(alive-but-wedged)。若仅报错不清理,消息重试(≤2 次)会
-      // 复用同一僵死 stdio 每次空转 60s;必须杀掉并置空,下回合 ensureClient
-      // 全新重生(host tools/模型/终端 tap 随重建)。
-      log.warn(`[OmpRpcAgent:${this.selfAgentId}] prompt 失败 → 回收可疑僵死进程 pid=${client.pid}`)
-      this.killProcess()
-      this.client = null
-      this.hostToolsRegistered = false
-      yield {
-        kind: 'error',
-        error: {
-          code: 'PROMPT_FAILED',
-          message: `omp prompt 失败: ${err instanceof Error ? err.message : String(err)}`,
-        },
-      }
-      return
-    }
-
-    // 产出事件直到 done/error
-    try {
-      while (!isDone || queue.length > 0) {
-        if (queue.length === 0 && !isDone) {
-          const remaining = idleTimeoutMs - (Date.now() - lastActivity)
-          if (remaining <= 0) {
-            client.send({ type: 'abort' }).catch(() => {})
-            // 子进程大概率已僵死:击杀并清引用,下回合 ensureClient 全新重生
-            this.killProcess()
-            this.client = null
-            this.hostToolsRegistered = false
-            enqueue({
-              kind: 'error',
-              error: {
-                code: 'TURN_STALLED',
-                message: `回合停滞 ${Math.round(idleTimeoutMs / 1000)}s 无 omp 事件,已中止重置(消息按已处理落账,后续消息继续)`,
-              },
-            })
-            continue
-          }
-          await new Promise<void>((r) => {
-            resolveWait = r
-            stallTimer = setTimeout(() => {
-              resolveWait = null
-              r()
-            }, remaining + 100)
-          })
-          if (stallTimer) {
-            clearTimeout(stallTimer)
-            stallTimer = null
-          }
-        }
-        while (queue.length > 0) {
-          yield queue.shift()!
-        }
-      }
-    }
-    finally {
-      if (stallTimer) clearTimeout(stallTimer)
-      // 事件监听必须全部解除:本 client 跨回合长期复用,残留的 mapOmpEvent
-      // 闭包会在后续每一帧继续执行并把事件 push 进孤儿队列(内存/CPU 随
-      // 回合数线性劣化,直接违背进程常驻的稳定性目标)
-      unsubState()
-      unsub()
-      signal.removeEventListener('abort', onAbort)
-      this.turnActive = false
-      this.streaming = false
-      this.toolState.currentTaskId = null
-    }
-  }
-
-  // ===== 内部:omp 事件 → AgentEvent 映射 =====
-
-  private mapOmpEvent(event: AgentSessionEvent, taskId: string | undefined): AgentEvent[] {
-    switch (event.type) {
-      case 'agent_start':
-        return [{
-          kind: 'status',
-          status: { state: 'WORKING', timestamp: new Date().toISOString() },
-        }]
-      case 'message_end': {
-        // 消息完成:如果有 message 内容,产出为 status 事件(任务历史追踪)
-        const msg = event.message
-        if (msg && Array.isArray(msg.content)) {
-          const text = (msg.content as Array<{ type?: string, text?: string }>)
-            .filter(c => c.type === 'text')
-            .map(c => c.text ?? '')
-            .join('')
-          if (text) {
-            const a2aMsg: A2AMessage = {
-              messageId: randomUUID(),
-              contextId: this.channelId,
-              role: 'ROLE_AGENT',
-              parts: [{ text }],
-            }
-            return [{
-              kind: 'status',
-              status: { state: 'WORKING', message: a2aMsg, timestamp: new Date().toISOString() },
-            }]
-          }
-        }
-        return []
-      }
-
-      case 'tool_execution_start': {
-        const toolName = event.toolName ?? 'tool'
-        return [{
-          kind: 'status',
-          status: {
-            state: 'WORKING',
-            message: {
-              messageId: randomUUID(),
-              contextId: this.channelId,
-              role: 'ROLE_AGENT',
-              parts: [{ text: `🔧 ${toolName}${toolArgsPreview(event.args)}` }],
-            },
-            timestamp: new Date().toISOString(),
-          },
-        }]
-      }
-
-      case 'agent_end': {
-        if (event.isTerminal === false) return []
-        const events: AgentEvent[] = []
-        // 提取最终 assistant 文本作为 artifact
-        const messages = event.messages ?? []
-        const assistantText = messages
-          .filter(m => m.role === 'assistant')
-          .flatMap(m => m.content.filter(c => c.type === 'text').map(c => c.text ?? ''))
-          .join('')
-        // 回合内的 provider/API 失败(限额 429、鉴权、上游 5xx 等):omp 把错误挂在最后一条
-        // assistant 消息上(stopReason='error' + errorStatus/errorMessage)而非 __error__ 帧。
-        // 不映射出来,上层只能看到「回合结束无产出」,真实原因彻底丢失(且照常重试到耗尽)。
-        const failure = [...messages].reverse().find((m) => {
-          const mm = m as { role?: string, stopReason?: string, errorMessage?: string }
-          return mm.role === 'assistant' && (mm.stopReason === 'error' || typeof mm.errorMessage === 'string')
-        }) as { errorStatus?: number, errorMessage?: string } | undefined
-        if (failure) {
-          const code = failure.errorStatus != null ? `OMP_LLM_${failure.errorStatus}` : 'OMP_LLM_ERROR'
-          const detail = failure.errorMessage ?? 'harness 回合以错误结束(无错误详情)'
-          log.error(`[OmpRpcAgent:${this.selfAgentId}] harness 回合失败 ${code}: ${detail}`)
-          // error 事件即回合终点(队列侧据此收口),不再追加 done
-          return [{ kind: 'error', error: { code, message: detail } }]
-        }
-        if (assistantText) {
-          events.push({
-            kind: 'artifact',
-            artifact: {
-              artifactId: randomUUID(),
-              name: 'output',
-              parts: [{ text: assistantText }],
-            },
-            lastChunk: true,
-            totalChunks: 1,
-          })
-        }
-        events.push({ kind: 'done', final: taskId ? { taskId } : undefined })
-        return events
-      }
-
-      case '__process_exit__':
-        return [{
-          kind: 'error',
-          error: { code: 'OMP_PROCESS_EXIT', message: 'omp 子进程意外退出' },
-        }]
-
-      // ===== 上下文治理事件(回合中到达的原生阈值/overflow 压缩;不污染事件流)=====
-      case 'compaction_start':
-        this.compacting = true
-        return []
-
-      case 'compaction_end': {
-        this.compacting = false
-        const result = (event as { result?: CompactionResult }).result
-        if (result?.summary) {
-          void this.harvestCompaction(
-            result.summary,
-            result.tokensBefore,
-            result.tokensAfter ?? result.estimatedTokensAfter,
-            (event as { reason?: string }).reason,
-          )
-        }
-        return []
-      }
-
-      // 回合完全落定(无自动重试/压缩重试/排队续跑):后续压缩检查由 AgentRuntime 驱动
-      case 'agent_settled':
-        return []
-
-      case '__error__':
-        return [{
-          kind: 'error',
-          error: { code: 'OMP_ERROR', message: (event as { error?: string }).error ?? 'omp 未知错误' },
-        }]
-
-      default:
-        return []
-    }
-  }
-
-  // ===== 内部:omp 客户端管理 =====
-
-  private async ensureClient(ctx: AgentRunContext): Promise<void> {
-    if (!this.workspace) {
-      this.workspace = ctx.workspace
-    }
-    if (!this.agentInfo) {
-      // init() 可能未调用;从 ctx 推断
-      this.channelId = ctx.channelId
-      this.agentRole = ctx.role
-      this.bridgeCtx.identity = { agentId: this.selfAgentId, channelId: this.channelId, role: this.agentRole, name: this.agentName }
-    }
-
-    if (!this.client || !this.client.alive) {
-      // 旧客户端已退出(exit 事件/OS 存活校准 reconcile 触发)——必须丢弃并重生,
-      // 否则后续回合会一直对着死 stdio 报 PROMPT_FAILED 烧重试配额
-      this.client = null
-      this.hostToolsRegistered = false
-      const command = this.config.command ?? 'omp'
-      const client = new OmpRpcClient({
-        command,
-        mode: this.rpcMode,
-        args: [...(this.config.thinkingLevel ? ['--thinking', this.config.thinkingLevel] : [])],
-        args: this.config.args,
-        cwd: this.config.cwd ?? process.cwd(),
-        // 进程退出 → 注册表标记(供运行时资源监控);pid=-1 表示无法取得,忽略
-        onExit: (pid, code) => {
-          if (pid > 0) {
-            markHarnessProcessExit(pid, code)
-            markTerminalSessionExit(pid, code)
-          }
-        },
-      })
-      await client.start()
-
-      // 登记 + 绑定 agent 身份(harness 进程监控的事实源)
-      const pid = client.pid
-      if (pid) {
-        registerHarnessProcess(pid, {
-          harness: 'omp',
-          command,
-          args: ['--mode', this.rpcMode, ...(this.config.args ?? [])],
-        })
-        bindHarnessProcess(pid, {
-          agentId: this.selfAgentId,
-          channelId: this.channelId,
-          name: this.agentName,
-          role: this.agentRole,
-        })
-        // 终端镜像 tap:全部 RPC 帧 → /monitor 终端(实时 TUI 渲染 + HITL)
-        attachTerminalTap(client, {
-          pid,
-          harness: 'omp',
-          agentId: this.selfAgentId,
-          channelId: this.channelId,
-          name: this.agentName,
-          role: this.agentRole,
-        })
-      }
-
-      // 设置模型(如果配置了 provider/model 任一;omp 以当前已设值为缺省补全,
-      // 只配其一时也显式发送,避免 respawn 后静默回落默认模型)
-      if (this.config.provider || this.config.model) {
-        try {
-          // 未配置一侧保持 undefined(JSON 序列化丢键)→ omp 以当前已设值补全
-          await client.send({
-            type: 'set_model',
-            provider: this.config.provider as string,
-            modelId: this.config.model as string,
-          })
-        }
-        catch {
-          // 模型设置失败不致命(用 omp 默认模型)
-        }
-      }
-
-      // 注册 host tools(按角色差异化:lead 全量,worker 剔除调度/团队管理专属工具)
-      client.onHostToolCall(req => this.handleHostTool(req))
-      await client.send({ type: 'set_host_tools', tools: hostToolsForRole(this.agentRole) })
-
-      // 上下文治理探测(feature-detect 一次;失败退化被动 usage 跟踪)+ 原生压缩兜底保持开启
-      void this.probeContext(client).catch(() => {})
-      void client.send({ type: 'set_auto_compaction', enabled: true }).catch(() => {})
-
-      this.client = client
-      this.hostToolsRegistered = true
-    }
-  }
-
-  /** 上下文状态探测:get_state 取 contextWindow/sessionId(一次;失败仅损失 percent 精度) */
-  private async probeContext(client: OmpRpcClient): Promise<void> {
-    try {
-      const resp = await client.send({ type: 'get_state' })
-      const data = (resp.data ?? {}) as Record<string, unknown>
-      const model = data.model as { contextWindow?: number } | undefined
-      if (typeof model?.contextWindow === 'number' && model.contextWindow > 0) {
-        this.contextWindow = model.contextWindow
-        client.setContextWindow(model.contextWindow)
-      }
-      if (typeof data.sessionId === 'string' && data.sessionId) this.sessionId = data.sessionId
-    }
-    catch { /* get_state 不可用:仅被动 usage */ }
-  }
-}
+/**
+ * OmpRpcAgentImpl — omp harness 的完整 AgentInterface 实现。
+ *
+ * 通过 OmpRpcClient 驱动真实 omp 子进程(--mode rpc),将 omp 的原生 agent 能力
+ * (LLM 推理 + 内置工具 read/write/edit/bash/grep/glob 等)接入 Workshop 的
+ * Channel 编排体系。
+ *
+ * 核心设计:
+ *  - 每个 Agent 实例持有一个 omp 子进程(lazy spawn,跨消息复用)
+ *  - AgentWorkspace 方法经 host-tool-bridge 注册为 omp host tools,agent 原生调用
+ *    report_progress / complete_task / dispatch_task / send_message 等,无需文本解析
+ *  - omp AgentSessionEvent → AgentEvent 五变体映射
+ *  - worker run():接收 assign → prompt omp → agent 用原生工具作业 + host tools 管理任务生命周期
+ *  - lead supervise():格式化快照 → prompt omp → agent 直接用 host tools 执行调度决策 → 返回 []
+ *  - dispose():杀子进程
+ *
+ * prompt 组装与工具分发已抽至共享层(prompt-builder / host-tool-bridge),
+ * 与 codex/dsh/opencode impl 全引擎一致。协议权威:omp://rpc.md。
+ */
+import { createLogger } from '../logger'
+import { BaseAgentImpl } from './base-agent'
+import { randomUUID } from 'node:crypto'
+import type {
+  AgentEvent,
+  AgentInterface,
+  AgentInfo,
+  AgentRunContext,
+  AgentRunRequest,
+  SupervisionDecision,
+  SupervisionSnapshot,
+} from './agent-interface'
+import type { A2AMessage } from '../types/a2a'
+import type { AgentContextStats } from '../types/task'
+import {
+  OmpRpcClient,
+  type AgentSessionEvent,
+  type CompactionResult,
+  type HostToolCallRequest,
+} from './adapters/omp-rpc-client'
+import {
+  registerHarnessProcess,
+  bindHarnessProcess,
+  markHarnessProcessExit,
+  killHarnessProcess,
+} from './harness-process'
+import {
+  attachTerminalTap,
+  markTerminalSessionExit,
+} from './harness-terminal'
+import { hostToolsForRole } from './host-tool-bridge'
+import {
+  createRosterCache,
+  extractJsonArray,
+  parseSteerBanner,
+  peerPrompt,
+  supervisePrompt,
+  systemManual,
+  toolArgsPreview,
+  workerPrompt,
+} from './prompt-builder'
+import { renderPrompt } from '../prompts/loader'
+import { attachOmpPluginBridge, onPluginToolsChange } from './plugin-tools'
+import { ompSettings } from '../settings'
+
+const log = createLogger('workshop.omp')
+
+// 兼容再导出(脚本/测试历史导入面;实现在 host-tool-bridge)
+export { HOST_TOOLS, hostToolsForRole } from './host-tool-bridge'
+
+// ===== 配置 =====
+
+export interface OmpAgentConfig {
+  /** omp 可执行文件路径(默认 'omp',从 PATH 查找) */
+  command?: string
+  /** 额外 CLI 参数 */
+  args?: string[]
+  /** omp 工作目录(默认 process.cwd()) */
+  cwd?: string
+  /** 模型 provider(如 'anthropic'/'openai'/'zhipu');省略则用 omp 默认 */
+  provider?: string
+  /** 模型 ID;省略则用 omp 默认 */
+  model?: string
+  /** thinking level: off/minimal/low/medium/high/xhigh/max */
+  thinkingLevel?: string
+  /** 自定义系统 prompt 前缀(拼接到 agent 系统指令之前) */
+  systemPromptPrefix?: string
+  /** 限制 omp 仅使用指定内置工具(null = 全部) */
+  toolNames?: string[]
+  /** 每轮停滞超时(ms,默认 300000):整轮无任何 omp 事件才中止;工具内阻塞(poll_messages ≤180s)有 tool 事件刷新计时 */
+  promptTimeoutMs?: number
+  /** supervise 轮超时(ms,默认 150000) */
+  superviseTimeoutMs?: number
+  /**
+   * omp 输出模式:'rpc-ui'(默认)= RPC 协议 + UI 上下文 —— ask 工具与
+   * extension_ui_request 对话框可用(监控终端 HITL 通道,强制 noPty);
+   * 'rpc' = 纯协议(无 UI,ask 不可用)。
+   */
+  rpcMode?: 'rpc' | 'rpc-ui'
+  /** channel 级作业场景 prompt(manager 装配时注入;用户场景规范,全员共享) */
+  scenarioPrompt?: string
+  /** agent 身份(由 factory 从 AgentInfo 注入) */
+  agentId?: string
+  name?: string
+  role?: 'lead' | 'worker'
+  channelId?: string
+}
+
+// 插件桥接管(host.mjs ctx.omp 排队注册的工具在此回放;幂等)
+attachOmpPluginBridge()
+
+/** 在跑 agent 实例表(工具注册表变更 → 热重发 set_host_tools,运行时注入无需重spawn) */
+const gLive = globalThis as typeof globalThis & { __ompLiveAgents?: Set<OmpRpcAgentImpl> }
+function liveAgents(): Set<OmpRpcAgentImpl> {
+  return gLive.__ompLiveAgents ??= new Set()
+}
+
+// 工具注册表变更 → 全部在跑会话热更新工具面
+onPluginToolsChange(() => {
+  for (const impl of liveAgents()) {
+    try {
+      impl.refreshPluginTools()
+    }
+    catch { /* 单实例失败不影响其他 */ }
+  }
+})
+
+// ===== OmpRpcAgentImpl =====
+
+export class OmpRpcAgentImpl extends BaseAgentImpl implements AgentInterface {
+  private readonly config: OmpAgentConfig
+
+  protected get harnessId(): string {
+    return 'omp'
+  }
+
+  protected configRecord(): Record<string, unknown> {
+    return this.config as unknown as Record<string, unknown>
+  }
+
+  /**
+   * supervise 用:收齐整个回合的事件。omp 的 supervise() 已整体覆盖(直接解析
+   * 决策/执行语义),本实现作为抽象成员的兜底:prompt → text deltas → agent_end,
+   * 文本装进单个 artifact 事件(与基类 extractJsonArray 消费形状一致)。
+   */
+  protected collectTurnEvents(prompt: string, timeoutMs: number, signal?: AbortSignal): Promise<AgentEvent[]> {
+    return new Promise((resolve) => {
+      if (!this.client) {
+        resolve([])
+        return
+      }
+      let text = ''
+      let settled = false
+      const finish = (): void => {
+        if (settled) return
+        settled = true
+        unsub()
+        clearTimeout(timer)
+        signalUnsub?.()
+        resolve(text.trim()
+          ? [{
+              kind: 'artifact',
+              artifact: { artifactId: randomUUID(), name: 'output', parts: [{ text }] },
+              lastChunk: true,
+              totalChunks: 1,
+            } as AgentEvent]
+          : [])
+      }
+      const abortTurn = (): void => {
+        void this.client?.send({ type: 'abort' }).catch(() => {})
+        finish()
+      }
+      const unsub = this.client.onEvent((event) => {
+        if (event.type === 'message_update' && event.assistantMessageEvent?.type === 'text_delta') {
+          text += event.assistantMessageEvent.delta ?? ''
+        }
+        if (event.type === 'agent_end' && event.isTerminal !== false) finish()
+        if (event.type === '__process_exit__' || event.type === '__error__') finish()
+      })
+      const timer = setTimeout(abortTurn, Math.max(timeoutMs, 1_000))
+      let signalUnsub: (() => void) | undefined
+      if (signal) {
+        if (signal.aborted) {
+          finish()
+          return
+        }
+        const onAbort = (): void => abortTurn()
+        signal.addEventListener('abort', onAbort, { once: true })
+        signalUnsub = () => signal.removeEventListener('abort', onAbort)
+      }
+      this.client.send({ type: 'prompt', message: prompt }).catch(() => finish())
+    })
+  }
+
+  private client: OmpRpcClient | null = null
+  private agentInfo: AgentInfo | null = null
+  private hostToolsRegistered = false
+  /**
+   * 会话回合状态(steer 可靠注入的依据):
+   * omp 的 steer 仅在回合 streaming 中生效——prompt 已入列但尚未开始输出时,
+   * steer 会"成功"返回但被静默丢弃。因此注入方轮询直到回合输出开始(streamingStarted)
+   * 才发送;若回合在等待期间结束,消息保持 pending 由消费循环处理。
+   */
+  private streaming = false
+  private turnActive = false
+  /** 当前回合产生的 assistant 文本(供诊断) */
+  private turnText = ''
+  /** agent 身份信息(factory 注入;无需等待 init()) */
+  private agentRole: 'lead' | 'worker' = 'worker'
+
+  // ===== 上下文治理(70% 无中断压缩环)=====
+  /** 压缩进行中(平台 gate 与 omp 原生压缩共用互斥位) */
+  private compacting = false
+  private lastCompactAt = 0
+  /** omp 不支持 compact 命令(旧版)→ 永久停用平台压缩;原生 auto-compaction 兜底,harvest 照常 */
+  private compactLegacy = false
+  /** 模型上下文窗口(get_state/get_session_stats 探测;null = 未知,percent 不可算) */
+  private contextWindow: number | null = null
+  private sessionId: string | null = null
+  /** harvest 双路去重(compact 响应与 compaction_end 事件可能双达) */
+  private lastHarvestKey = ''
+  private lastHarvestAt = 0
+
+  /** 总开关(omp.compact_enabled;env AW_OMP_COMPACT_ENABLED=0 显式关闭兼容) */
+  private compactEnabled(): boolean {
+    return ompSettings().compact_enabled
+  }
+
+  private compactThreshold(): number {
+    return ompSettings().compact_threshold
+  }
+
+  private compactMinIntervalMs(): number {
+    return ompSettings().compact_min_interval_ms
+  }
+
+  private compactWaitMs(): number {
+    return ompSettings().compact_wait_ms
+  }
+
+  /** 被动上下文用量快照(无探测 RPC;getStatus 透出用) */
+  getContextStats(): AgentContextStats | null {
+    const usage = this.client?.getContextUsage() ?? null
+    const window = this.contextWindow ?? usage?.contextWindow ?? null
+    if (!usage && window === null) return null
+    const usedTokens = usage?.tokens ?? 0
+    return {
+      usedTokens,
+      contextWindow: window,
+      percent: window && window > 0 ? Math.min(1, usedTokens / window) : null,
+      compacting: this.compacting,
+    }
+  }
+
+  /**
+   * 上下文用量探测:get_session_stats 权威值;不可用退化为被动 usage + 已探测窗口。
+   * percent 归一化到 0-1(omp 可能返回 0-100)。
+   */
+  private async probeUsage(client: OmpRpcClient): Promise<{ tokens: number, contextWindow: number | null, percent: number | null } | null> {
+    try {
+      const resp = await client.send({ type: 'get_session_stats' })
+      const cu = ((resp.data ?? {}) as Record<string, unknown>).contextUsage as
+        | { tokens?: number, contextWindow?: number, percent?: number }
+        | undefined
+      if (cu && typeof cu.tokens === 'number' && cu.tokens > 0) {
+        if (typeof cu.contextWindow === 'number' && cu.contextWindow > 0) {
+          this.contextWindow = cu.contextWindow
+          client.setContextWindow(cu.contextWindow)
+        }
+        const window = this.contextWindow ?? cu.contextWindow ?? null
+        const rawPercent = typeof cu.percent === 'number' ? (cu.percent > 1 ? cu.percent / 100 : cu.percent) : null
+        const percent = rawPercent ?? (window && window > 0 ? Math.min(1, cu.tokens / window) : null)
+        return { tokens: cu.tokens, contextWindow: window, percent }
+      }
+    }
+    catch { /* get_session_stats 不可用 → 被动跟踪 */ }
+    return client.getContextUsage()
+  }
+
+  /**
+   * 上下文治理门控:三条 prompt 路径(worker/peer/supervise)在回合间隙统一调用;
+   * post-settle 经 onTurnSettled 复用同一守卫。任何失败路径都放行——压缩晚一轮
+   * 永远好过中断作业。三重防线:仅回合间隙发起 → get_state 复查 isStreaming/
+   * isCompacting → compacting 互斥位 + 最小间隔防振荡。
+   */
+  private async contextGate(reason: string): Promise<void> {
+    if (!this.compactEnabled() || this.compactLegacy || this.compacting) return
+    const client = this.client
+    if (!client || !client.alive) return
+    try {
+      const usage = await this.probeUsage(client)
+      if (!usage || usage.percent === null || usage.percent < this.compactThreshold()) return
+      if (Date.now() - this.lastCompactAt < this.compactMinIntervalMs()) return
+      // 回合间隙硬校验:流式中/压缩中绝不发起(不中断执行的硬约束)
+      const st = await client.send({ type: 'get_state' })
+      const s = (st.data ?? {}) as Record<string, unknown>
+      if (s.isStreaming === true || s.isCompacting === true) return
+      if (typeof s.sessionId === 'string' && s.sessionId) this.sessionId = s.sessionId
+      this.compacting = true
+      this.lastCompactAt = Date.now()
+      try {
+        const result = await this.runCompaction(client)
+        if (result?.summary) {
+          await this.harvestCompaction(
+            result.summary,
+            result.tokensBefore,
+            result.tokensAfter ?? result.estimatedTokensAfter,
+            reason,
+          )
+        }
+      }
+      finally {
+        this.compacting = false
+      }
+    }
+    catch (err) {
+      const msg = err instanceof Error ? err.message : String(err)
+      if (/compact/i.test(msg) && /unknown|unsupported|invalid|无效|未知|不支持/i.test(msg)) this.compactLegacy = true
+      log.warn(`[OmpRpcAgent:${this.selfAgentId}] contextGate(${reason}) 失败,放行: ${msg}`)
+    }
+  }
+
+  /**
+   * 发起 compact 并等待 compaction_end(响应内含 result 则立即收口;
+   * 超时放行——压缩在 harness 内后台继续,isCompacting 防后续双发)。
+   */
+  private runCompaction(client: OmpRpcClient): Promise<CompactionResult | null> {
+    return new Promise<CompactionResult | null>((resolve) => {
+      let settled = false
+      const finish = (r: CompactionResult | null): void => {
+        if (settled) return
+        settled = true
+        off()
+        clearTimeout(timer)
+        resolve(r)
+      }
+      const off = client.onEvent((event) => {
+        if (event.type === 'compaction_end') {
+          finish((event as { result?: CompactionResult }).result ?? null)
+        }
+        else if (event.type === '__process_exit__') {
+          finish(null)
+        }
+      })
+      const timer = setTimeout(() => {
+        log.warn(`[OmpRpcAgent:${this.selfAgentId}] compact 等待超时(${this.compactWaitMs()}ms)→ 放行(压缩后台继续)`)
+        finish(null)
+      }, this.compactWaitMs())
+      void client.send({ type: 'compact', customInstructions: this.compactionHints() })
+        .then((resp) => {
+          const result = (resp.data as { result?: CompactionResult } | undefined)?.result
+          if (result?.summary) finish(result)
+          // 响应无 result:等 compaction_end 事件(协议权威语义)
+        })
+        .catch((err) => {
+          const msg = err instanceof Error ? err.message : String(err)
+          if (/compact/i.test(msg)) this.compactLegacy = true
+          log.warn(`[OmpRpcAgent:${this.selfAgentId}] compact 发起失败: ${msg}`)
+          finish(null)
+        })
+    })
+  }
+
+  /** 压缩摘要指令(外置 prompts;缺失时省略,omp 用默认策略) */
+  private compactionHints(): string | undefined {
+    try {
+      return renderPrompt('compaction-hints')
+    }
+    catch {
+      return undefined
+    }
+  }
+
+  /**
+   * 压缩摘要 harvest(三路统一:平台主动 compact / omp 阈值自动 / overflow):
+   * 经 workspace.recordSessionMemory 落库为本人 episodic-session 记忆。
+   * 双路去重:同摘要 60s 内只入库一次。
+   */
+  private async harvestCompaction(summary: string, tokensBefore?: number, tokensAfter?: number, reason?: string): Promise<void> {
+    const key = summary.slice(0, 120)
+    if (key === this.lastHarvestKey && Date.now() - this.lastHarvestAt < 60_000) return
+    this.lastHarvestKey = key
+    this.lastHarvestAt = Date.now()
+    try {
+      await this.workspace?.recordSessionMemory?.({ summary, tokensBefore, tokensAfter, reason })
+      log.info(`[OmpRpcAgent:${this.selfAgentId}] 压缩摘要已入记忆(${summary.length} 字,reason=${reason ?? 'auto'})`)
+    }
+    catch (err) {
+      log.warn(`[OmpRpcAgent:${this.selfAgentId}] 压缩摘要入库失败(不影响会话):`, err instanceof Error ? err.message : err)
+    }
+  }
+
+  /** 回合落定钩子(AgentRuntime 在信箱空时调用):后台压缩检查,自守卫不抛错 */
+  async onTurnSettled(): Promise<void> {
+    await this.contextGate('post-settle')
+  }
+
+  constructor(config: Record<string, unknown> = {}) {
+    super({
+      agentId: typeof config.agentId === 'string' ? config.agentId : '',
+      name: typeof config.name === 'string' ? config.name : undefined,
+      role: config.role === 'lead' ? 'lead' : 'worker',
+      channelId: typeof config.channelId === 'string' ? config.channelId : '',
+    })
+    this.config = config as OmpAgentConfig
+    // effort 统一入口(channel 默认 LLM 注入 config.effort)→ omp thinking 级别
+    if (!this.config.thinkingLevel && typeof config.effort === 'string' && config.effort) {
+      this.config.thinkingLevel = config.effort
+    }
+    // factory 注入的 agent 身份(无需等待 init())
+    this.selfAgentId = this.config.agentId ?? ''
+    this.agentName = this.config.name ?? 'agent'
+    this.agentRole = this.config.role ?? 'worker'
+    this.channelId = this.config.channelId ?? ''
+    this.bridgeCtx.identity = { agentId: this.selfAgentId, channelId: this.channelId, role: this.agentRole, name: this.agentName }
+    this.roster = createRosterCache({
+      selfAgentId: this.selfAgentId,
+      listAgents: () => this.workspace!.listAgents(),
+    })
+    liveAgents().add(this) // 在跑实例表:插件工具热注入的目标集
+  }
+
+  // ===== 生命周期 =====
+
+  async init(input: { agent: AgentInfo, channelId: string }): Promise<void> {
+    this.agentInfo = input.agent
+    this.channelId = input.channelId
+    this.agentName = input.agent.name
+    this.agentRole = input.agent.role
+    this.selfAgentId = input.agent.id
+    this.bridgeCtx.identity = { agentId: this.selfAgentId, channelId: this.channelId, role: this.agentRole, name: this.agentName }
+    this.roster = createRosterCache({
+      selfAgentId: this.selfAgentId,
+      listAgents: () => this.workspace!.listAgents(),
+    })
+  }
+
+  async dispose(): Promise<void> {
+    liveAgents().delete(this)
+    if (this.client) {
+      const pid = this.client.pid
+      await this.client.dispose()
+      this.client = null
+      if (pid) {
+        markHarnessProcessExit(pid, null)
+        markTerminalSessionExit(pid, null)
+      }
+    }
+    this.hostToolsRegistered = false
+  }
+
+  /** 工具注册表变更 → 热重发 set_host_tools(在跑会话立即获得插件新工具,无需重spawn) */
+  refreshPluginTools(): void {
+    const client = this.client
+    if (!client || !client.alive) return
+    void client.send({ type: 'set_host_tools', tools: hostToolsForRole(this.agentRole, this.channelId) }).catch(() => {})
+  }
+
+  /** harness 进程资源信息(运行时资源监控;进程未 spawn/已回收 → null) */
+  getProcessInfo(): { pid: number, alive: boolean, command: string } | null {
+    const client = this.client
+    const pid = client?.pid
+    if (!pid || !client) return null
+    return { pid, alive: client.alive, command: `omp --mode ${this.rpcMode}` }
+  }
+
+  /** 强制终止 harness 进程(进程树;终止后由调用方停止对应 AgentRuntime) */
+  killProcess(): void {
+    const pid = this.client?.pid
+    if (pid) {
+      markTerminalSessionExit(pid, null)
+      killHarnessProcess(pid)
+    }
+    else {
+      this.client?.kill()
+    }
+  }
+
+  /**
+   * harness 进程存活校准(manager sweeper 周期性调用;休眠/强杀后 exit 事件
+   * 可能不达父进程):按 PID 探 OS 实际存在性,进程已死 → 客户端收敛为已退出,
+   * 在途回合经 __process_exit__ 归位,下一回合 ensureClient 重生子进程。
+   */
+  reconcileProcess(): void {
+    this.client?.reconcile()
+  }
+
+  /** omp 输出模式(rpc-ui = 默认,启用 HITL 对话框) */
+  private get rpcMode(): 'rpc' | 'rpc-ui' {
+    return this.config.rpcMode === 'rpc' ? 'rpc' : 'rpc-ui'
+  }
+
+  // ===== 工具桥(共享分发)=====
+
+  /** 引擎无关工具面:REST/MCP 桥直调与 omp 内部 host_tool_call 共用同一实现 */
+  private async handleHostTool(req: HostToolCallRequest): Promise<{ text: string, isError?: boolean }> {
+    return this.dispatchHostTool(req.toolName, req.arguments ?? {})
+  }
+
+  // ===== prompt 组合 =====
+
+  /** 前置上下文(场景 × 身份 × 工业简报 × 名册;共享 prompt-builder) */
+  private systemManual(): string {
+    return systemManual()
+  }
+
+  /**
+   * 实时消息注入(送达模式协议):
+   *  - 回合 streaming 中 → 立即 steer,返回 'steer'(同轮可见,唯一可标记消费的路径)
+   *  - 回合活跃但尚未 streaming(prompt 排队窗口,上限 20s)→ 等 streaming 开始后 steer
+   *  - 回合被 host 工具阻塞(如 poll_messages 等待)/空闲/发送失败 → 返回 'deferred':
+   *    消息保持 pending,poll_messages 的 Mailbox 到信回调即时取走(毫秒级),或本回合
+   *    结束后由消费循环按 FIFO 起回合处理。
+   */
+  async steer(text: string): Promise<'steer' | 'deferred'> {
+    // 从确定性触发横幅提取回执上下文(AgentRuntime.injectSteer 生成,格式固定):
+    // "[实时消息 from <id>]: ..." + "[系统触发器] 本消息要求回复(reply_to=<messageId>)。"
+    const banner = parseSteerBanner(text)
+    if (banner) {
+      this.toolState.replyContext = banner
+    }
+    const client = this.client
+    if (!client) return 'deferred'
+    try {
+      if (this.streaming) {
+        await client.send({ type: 'steer', message: text })
+        return 'steer'
+      }
+      if (this.turnActive) {
+        // prompt 排队窗口:等 streaming 开始(上限 20s);工具阻塞的回合等不到 → deferred
+        const deadline = Date.now() + 20_000
+        while (Date.now() < deadline && this.turnActive && !this.streaming) {
+          const { promise, resolve } = Promise.withResolvers()
+          setTimeout(resolve, 150)
+          await promise
+        }
+        if (this.streaming && this.turnActive) {
+          await client.send({ type: 'steer', message: text })
+          return 'steer'
+        }
+        return 'deferred'
+      }
+      // 空闲/回合已结束:deferred(消息 pending,消费循环 dequeue 即起回合)
+      return 'deferred'
+    }
+    catch (err) {
+      log.error(`[OmpRpcAgent:${this.selfAgentId}] steer 注入失败(消息保持 pending):`, err instanceof Error ? err.message : err)
+      return 'deferred'
+    }
+  }
+
+  // ===== supervise() =====
+
+  /** supervise 单飞守卫:同一 client 不并发 LLM 回合(残留回合与下一 prompt 混流的根因) */
+  private supervising = false
+
+  async supervise(snapshot: SupervisionSnapshot, ctx: AgentRunContext, opts?: { signal?: AbortSignal }): Promise<SupervisionDecision[]> {
+    await this.ensureClient(ctx)
+    if (!this.client) return []
+    if (this.supervising) return [] // 上一轮 supervise 未收口:跳过本拍(节流即正确)
+
+    // 上下文门控(≥70% 先压缩再调度;回合间隙发起,失败放行)
+    await this.contextGate('supervise')
+
+    const prompt = supervisePrompt({
+      snapshot,
+      agentName: this.agentName,
+      channelId: this.channelId,
+      ctxPrefix: await this.contextPrefix(),
+      manual: this.systemManual(),
+      memory: ctx.memory,
+    })
+    // 150s 上界:supervise 持 lead.execLock 期间信箱消费停顿(更久会拖垮 worker 回执处理);
+    // supervise 是一次真实 LLM 回合:omp 冷启动(插件/MCP 加载 30~90s)+ 慢 provider
+    // 单步可能 >60s,过紧会把正常回合掐成 "Interrupted by user"。默认 150s,
+    // 仅拦真僵死(更紧的预算由调用方 config.superviseTimeoutMs 显式传入);真 abort 已实现。
+    const timeoutMs = this.config.superviseTimeoutMs ?? 150_000
+
+    return new Promise<SupervisionDecision[]>((resolve) => {
+      let assistantText = ''
+      let resolved = false
+
+      const finish = (decisions: SupervisionDecision[]) => {
+        if (resolved) return
+        resolved = true
+        this.supervising = false
+        unsub()
+        clearTimeout(timer)
+        signalUnsub?.()
+        resolve(decisions)
+      }
+
+      const abortTurn = (): void => {
+        // 超时/外部取消:真正中止 omp 当前回合 —— 只 resolve 不 abort 会让残留回合
+        // 与下一个 prompt 在同一 client 混流(决策错位 + token 空烧)
+        log.warn(`[OmpRpcAgent:${this.selfAgentId}] supervise 超时(${timeoutMs}ms)→ abort 当前调度回合`)
+        void this.client?.send({ type: 'abort' }).catch(() => {})
+        finish([])
+      }
+
+      const unsub = this.client!.onEvent((event) => {
+        if (event.type === 'message_update' && event.assistantMessageEvent?.type === 'text_delta') {
+          assistantText += event.assistantMessageEvent.delta ?? ''
+        }
+        if (event.type === 'agent_end' && event.isTerminal !== false) {
+          // 尝试从文本解析 JSON 决策(备用:如果 agent 没用 host tools 而是输出 JSON)
+          const parsed = extractJsonArray(assistantText)
+          if (parsed && parsed.length > 0) {
+            finish(parsed as SupervisionDecision[])
+          }
+          else {
+            // agent 可能已通过 host tools 直接执行了调度,返回空(已执行)
+            finish([])
+          }
+        }
+        if (event.type === '__process_exit__' || event.type === '__error__') {
+          finish([])
+        }
+      })
+
+      const timer = setTimeout(abortTurn, timeoutMs)
+
+      // 外部取消(调度器 cancel 路径)传导:abort 当前 LLM 回合并立即收口
+      let signalUnsub: (() => void) | undefined
+      if (opts?.signal) {
+        if (opts.signal.aborted) {
+          unsub()
+          resolve([])
+          return
+        }
+        const onAbort = (): void => abortTurn()
+        opts.signal.addEventListener('abort', onAbort, { once: true })
+        signalUnsub = () => opts.signal?.removeEventListener('abort', onAbort)
+      }
+
+      this.supervising = true
+      this.client!.send({ type: 'prompt', message: prompt }).catch(() => finish([]))
+    })
+  }
+
+  // ===== 内部:worker 执行 =====
+
+  protected async* workerTurn(request: AgentRunRequest, ctx: AgentRunContext): AsyncGenerator<AgentEvent, void, unknown> {
+    const taskId = request.taskId
+      ?? (request.message.metadata?.['x-aw-task-id'] as string | undefined)
+    if (!taskId) return
+
+    this.toolState.currentTaskId = taskId
+
+    try {
+      await this.ensureClient(ctx)
+    }
+    catch (err) {
+      yield {
+        kind: 'error',
+        error: {
+          code: 'OMP_SPAWN_FAILED',
+          message: `omp 子进程启动失败: ${err instanceof Error ? err.message : String(err)}`,
+        },
+      }
+      return
+    }
+
+    if (!this.client) {
+      yield { kind: 'error', error: { code: 'OMP_NOT_READY', message: 'omp 客户端未就绪' } }
+      return
+    }
+
+    // 构建 prompt
+    const taskText = request.message.parts.map((p) => {
+      if ('text' in p) return p.text
+      if ('data' in p) return JSON.stringify(p.data)
+      if ('url' in p) return p.url
+      if ('raw' in p) return p.raw
+      return ''
+    }).join('\n')
+    const prompt = workerPrompt({
+      agentName: this.agentName,
+      channelId: this.channelId,
+      taskId,
+      taskText,
+      memory: request.memory,
+      ctxPrefix: await this.contextPrefix(),
+      manual: this.systemManual(),
+    })
+
+    // 流式执行 + 事件映射
+    yield* this.promptAndStream(prompt, taskId, ctx.signal)
+  }
+
+  /**
+   * 点对点消息处理(实时通信驱动;lead 与 worker 通用)。
+   * 触发器语义:metadata['x-aw-require-reply']='true' → 必须经 send_message_to_agent
+   * 回给发送者:执行结果 + 对方所需内容,in_reply_to 关联原消息,并声明是否需再响应。
+   */
+  protected async* peerTurn(request: AgentRunRequest, ctx: AgentRunContext): AsyncGenerator<AgentEvent, void, unknown> {
+    try {
+      await this.ensureClient(ctx)
+    }
+    catch (err) {
+      // spawn 失败必须显式报错:静默 return 会让 sawRunError=false → 消息被
+      // 直接 markConsumed,实时消息一次性丢失且无任何重试;产出 error 事件
+      // 走消息 requeue(≤2 次),与 workerRun 同口径
+      yield {
+        kind: 'error',
+        error: {
+          code: 'OMP_SPAWN_FAILED',
+          message: `omp 子进程启动失败(peer 消息未消费,将重试): ${err instanceof Error ? err.message : String(err)}`,
+        },
+      }
+      return
+    }
+    if (!this.client) {
+      yield { kind: 'error', error: { code: 'OMP_NOT_READY', message: 'omp 客户端未就绪(peer 消息未消费,将重试)' } }
+      return
+    }
+
+    const msg = request.message
+    // 显示名:agent 同事用 id(名册可解析);人类发送者用 fromLabel;兜底 unknown
+    const fromId = request.fromAgentId
+      ?? (typeof msg.metadata?.['x-aw-from-label'] === 'string' ? msg.metadata['x-aw-from-label'] : undefined)
+      ?? 'unknown'
+    const msgText = msg.parts.map((p) => {
+      if ('text' in p) return p.text
+      if ('data' in p) return JSON.stringify(p.data)
+      if ('url' in p) return p.url
+      if ('raw' in p) return p.raw
+      return ''
+    }).join('\n')
+    const requireReply = msg.metadata?.['x-aw-require-reply'] === 'true'
+    const isReply = typeof msg.metadata?.['x-aw-in-reply-to'] === 'string'
+    // 跨 Channel 来信:发送方是其他 channel 的 lead,回执必须走 send_cross_channel_message
+    const crossChannel = msg.metadata?.['x-aw-cross-channel'] === 'true'
+    const fromChannel = typeof msg.metadata?.['x-aw-from-channel'] === 'string'
+      ? String(msg.metadata['x-aw-from-channel'])
+      : ''
+    // 回执自动关联仅对真实 agent 发送者生效(人类无 agentId 可回投;
+    // 对人类的回复经事件流/时间线可见)
+    this.toolState.replyContext = requireReply && request.fromAgentId && !crossChannel
+      ? { fromId: request.fromAgentId, messageId: msg.messageId }
+      : null
+
+    const prompt = peerPrompt({
+      agentName: this.agentName,
+      role: this.agentRole,
+      channelId: this.channelId,
+      ctxPrefix: await this.contextPrefix(),
+      manual: this.systemManual(),
+      memory: request.memory,
+      fromId,
+      messageId: msg.messageId,
+      requireReply,
+      isReply,
+      crossChannel,
+      fromChannel,
+      msgText,
+    })
+
+    yield* this.promptAndStream(prompt, undefined, ctx.signal)
+  }
+
+  // ===== 内部:prompt 发送 + 事件流桥接 =====
+
+  private async* promptAndStream(
+    prompt: string,
+    taskId: string | undefined,
+    signal: AbortSignal,
+  ): AsyncGenerator<AgentEvent, void, unknown> {
+    const client = this.client
+    if (!client) {
+      throw new Error('omp 客户端未就绪')
+    }
+    // 上下文门控(≥70% 时在回合间隙先压缩再发 prompt;等待有界,失败放行)
+    await this.contextGate('pre-prompt')
+    const queue: AgentEvent[] = []
+    let resolveWait: (() => void) | null = null
+    let isDone = false
+    // 回合生命周期:prompt 已发出 → turnActive;首条 message_update → streaming;message_end/turn_end → 结束
+    this.turnActive = true
+    this.streaming = false
+    this.turnText = ''
+    // 正文流式差额账本:contentIndex → 已流出长度。部分 provider(reasoning 模型/
+    // openai-completions)的 text 块不走 text_delta(只有换行),全文在 text_end.content
+    // 落定 —— 差额兜底把未流出部分补发为 delta,保证前端流式不缺正文。
+    const textSent = new Map<number, number>()
+    const unsubState = client.onEvent((event) => {
+      if (event.type === 'message_update') {
+        const ev = event.assistantMessageEvent as
+          | { type?: string, delta?: string, content?: string, contentIndex?: number }
+          | undefined
+        if (ev?.type === 'text_delta') {
+          this.streaming = true
+          const text = ev.delta ?? ''
+          const ci = typeof ev.contentIndex === 'number' ? ev.contentIndex : 0
+          textSent.set(ci, (textSent.get(ci) ?? 0) + text.length)
+          this.turnText += text
+          // LLM 流式增量透出(AEP agent.delta 事件源):仅 worker/peer 转本走生成器
+          if (text) enqueueDelta(text)
+        }
+        else if (ev?.type === 'text_end' && typeof ev.content === 'string') {
+          const ci = typeof ev.contentIndex === 'number' ? ev.contentIndex : 0
+          const sent = textSent.get(ci) ?? 0
+          if (ev.content.length > sent) {
+            const chunk = ev.content.slice(sent)
+            textSent.set(ci, ev.content.length)
+            this.streaming = true
+            this.turnText += chunk
+            enqueueDelta(chunk)
+          }
+        }
+      }
+      if (event.type === 'message_end' || event.type === 'turn_end') {
+        this.streaming = false
+        flushDelta()
+      }
+      if (event.type === 'agent_end' && event.isTerminal !== false) {
+        this.turnActive = false
+        this.streaming = false
+        flushDelta()
+      }
+      if (event.type === '__process_exit__' || event.type === '__error__') {
+        this.turnActive = false
+        this.streaming = false
+      }
+    })
+
+    const enqueue = (event: AgentEvent): void => {
+      queue.push(event)
+      if (event.kind === 'done' || event.kind === 'error') isDone = true
+      lastActivity = Date.now()
+      resolveWait?.()
+      resolveWait = null
+    }
+
+    // 停滞看门狗:整轮无任何 omp 事件(挂死的 LLM 调用/子进程僵死)时中止回合,
+    // 否则消息永久卡 consuming、队友消息无限堆积。host 工具内阻塞
+    // (poll_messages 最长 180s)与慢 provider(实测单步可 >5min)都有
+    // tool/状态事件刷新计时;默认 600s 覆盖慢回合,只拦真僵死。
+    const idleTimeoutMs = this.config.promptTimeoutMs ?? 600_000
+    let lastActivity = Date.now()
+    let stallTimer: ReturnType<typeof setTimeout> | null = null
+
+    // LLM 流式增量缓冲:50ms 批量合并为一帧 delta(防高频 text_delta 洪泛 WS)
+    let deltaBuf = ''
+    let deltaTimer: ReturnType<typeof setTimeout> | null = null
+    const flushDelta = (): void => {
+      if (deltaTimer) {
+        clearTimeout(deltaTimer)
+        deltaTimer = null
+      }
+      if (!deltaBuf) return
+      const text = deltaBuf
+      deltaBuf = ''
+      enqueue({ kind: 'delta', delta: { text } })
+    }
+    const enqueueDelta = (text: string): void => {
+      deltaBuf += text
+      deltaTimer ??= setTimeout(() => {
+        deltaTimer = null
+        flushDelta()
+      }, 50)
+    }
+
+    // 订阅 omp 事件流
+    const unsub = client.onEvent((event) => {
+      for (const mapped of this.mapOmpEvent(event, taskId)) {
+        enqueue(mapped)
+      }
+    })
+
+    // abort 传导
+    const onAbort = (): void => {
+      log.warn(`[OmpRpcAgent:${this.selfAgentId}] run 被 abort(signal)→ 中止 omp 回合,taskId=${taskId ?? '-'}`)
+      client.send({ type: 'abort' }).catch(() => {})
+      if (!isDone) {
+        enqueue({ kind: 'done', final: taskId ? { taskId } : undefined })
+      }
+    }
+    if (signal.aborted) {
+      onAbort()
+    }
+    else {
+      signal.addEventListener('abort', onAbort, { once: true })
+    }
+
+    // 发送 prompt
+    try {
+      await client.send({ type: 'prompt', message: prompt })
+    }
+    catch (err) {
+      unsub()
+      signal.removeEventListener('abort', onAbort)
+      // 僵死进程回收:prompt send 失败 = stdio 断裂或 60s 命令确认超时 —— 进程
+      // "活着但不干活"(alive-but-wedged)。若仅报错不清理,消息重试(≤2 次)会
+      // 复用同一僵死 stdio 每次空转 60s;必须杀掉并置空,下回合 ensureClient
+      // 全新重生(host tools/模型/终端 tap 随重建)。
+      log.warn(`[OmpRpcAgent:${this.selfAgentId}] prompt 失败 → 回收可疑僵死进程 pid=${client.pid}`)
+      this.killProcess()
+      this.client = null
+      this.hostToolsRegistered = false
+      yield {
+        kind: 'error',
+        error: {
+          code: 'PROMPT_FAILED',
+          message: `omp prompt 失败: ${err instanceof Error ? err.message : String(err)}`,
+        },
+      }
+      return
+    }
+
+    // 产出事件直到 done/error
+    try {
+      while (!isDone || queue.length > 0) {
+        if (queue.length === 0 && !isDone) {
+          const remaining = idleTimeoutMs - (Date.now() - lastActivity)
+          if (remaining <= 0) {
+            client.send({ type: 'abort' }).catch(() => {})
+            // 子进程大概率已僵死:击杀并清引用,下回合 ensureClient 全新重生
+            this.killProcess()
+            this.client = null
+            this.hostToolsRegistered = false
+            enqueue({
+              kind: 'error',
+              error: {
+                code: 'TURN_STALLED',
+                message: `回合停滞 ${Math.round(idleTimeoutMs / 1000)}s 无 omp 事件,已中止重置(消息按已处理落账,后续消息继续)`,
+              },
+            })
+            continue
+          }
+          await new Promise<void>((r) => {
+            resolveWait = r
+            stallTimer = setTimeout(() => {
+              resolveWait = null
+              r()
+            }, remaining + 100)
+          })
+          if (stallTimer) {
+            clearTimeout(stallTimer)
+            stallTimer = null
+          }
+        }
+        while (queue.length > 0) {
+          yield queue.shift()!
+        }
+      }
+    }
+    finally {
+      if (stallTimer) clearTimeout(stallTimer)
+      // 事件监听必须全部解除:本 client 跨回合长期复用,残留的 mapOmpEvent
+      // 闭包会在后续每一帧继续执行并把事件 push 进孤儿队列(内存/CPU 随
+      // 回合数线性劣化,直接违背进程常驻的稳定性目标)
+      unsubState()
+      unsub()
+      signal.removeEventListener('abort', onAbort)
+      this.turnActive = false
+      this.streaming = false
+      this.toolState.currentTaskId = null
+    }
+  }
+
+  // ===== 内部:omp 事件 → AgentEvent 映射 =====
+
+  private mapOmpEvent(event: AgentSessionEvent, taskId: string | undefined): AgentEvent[] {
+    switch (event.type) {
+      case 'agent_start':
+        return [{
+          kind: 'status',
+          status: { state: 'WORKING', timestamp: new Date().toISOString() },
+        }]
+      case 'message_end': {
+        // 消息完成:如果有 message 内容,产出为 status 事件(任务历史追踪)
+        const msg = event.message
+        if (msg && Array.isArray(msg.content)) {
+          const text = (msg.content as Array<{ type?: string, text?: string }>)
+            .filter(c => c.type === 'text')
+            .map(c => c.text ?? '')
+            .join('')
+          if (text) {
+            const a2aMsg: A2AMessage = {
+              messageId: randomUUID(),
+              contextId: this.channelId,
+              role: 'ROLE_AGENT',
+              parts: [{ text }],
+            }
+            return [{
+              kind: 'status',
+              status: { state: 'WORKING', message: a2aMsg, timestamp: new Date().toISOString() },
+            }]
+          }
+        }
+        return []
+      }
+
+      case 'tool_execution_start': {
+        const toolName = event.toolName ?? 'tool'
+        return [{
+          kind: 'status',
+          status: {
+            state: 'WORKING',
+            message: {
+              messageId: randomUUID(),
+              contextId: this.channelId,
+              role: 'ROLE_AGENT',
+              parts: [{ text: `🔧 ${toolName}${toolArgsPreview(event.args)}` }],
+            },
+            timestamp: new Date().toISOString(),
+          },
+        }]
+      }
+
+      case 'agent_end': {
+        if (event.isTerminal === false) return []
+        const events: AgentEvent[] = []
+        // 提取最终 assistant 文本作为 artifact
+        const messages = event.messages ?? []
+        const assistantText = messages
+          .filter(m => m.role === 'assistant')
+          .flatMap(m => m.content.filter(c => c.type === 'text').map(c => c.text ?? ''))
+          .join('')
+        // 回合内的 provider/API 失败(限额 429、鉴权、上游 5xx 等):omp 把错误挂在最后一条
+        // assistant 消息上(stopReason='error' + errorStatus/errorMessage)而非 __error__ 帧。
+        // 不映射出来,上层只能看到「回合结束无产出」,真实原因彻底丢失(且照常重试到耗尽)。
+        const failure = [...messages].reverse().find((m) => {
+          const mm = m as { role?: string, stopReason?: string, errorMessage?: string }
+          return mm.role === 'assistant' && (mm.stopReason === 'error' || typeof mm.errorMessage === 'string')
+        }) as { errorStatus?: number, errorMessage?: string } | undefined
+        if (failure) {
+          const code = failure.errorStatus != null ? `OMP_LLM_${failure.errorStatus}` : 'OMP_LLM_ERROR'
+          const detail = failure.errorMessage ?? 'harness 回合以错误结束(无错误详情)'
+          log.error(`[OmpRpcAgent:${this.selfAgentId}] harness 回合失败 ${code}: ${detail}`)
+          // error 事件即回合终点(队列侧据此收口),不再追加 done
+          return [{ kind: 'error', error: { code, message: detail } }]
+        }
+        if (assistantText) {
+          events.push({
+            kind: 'artifact',
+            artifact: {
+              artifactId: randomUUID(),
+              name: 'output',
+              parts: [{ text: assistantText }],
+            },
+            lastChunk: true,
+            totalChunks: 1,
+          })
+        }
+        events.push({ kind: 'done', final: taskId ? { taskId } : undefined })
+        return events
+      }
+
+      case '__process_exit__':
+        return [{
+          kind: 'error',
+          error: { code: 'OMP_PROCESS_EXIT', message: 'omp 子进程意外退出' },
+        }]
+
+      // ===== 上下文治理事件(回合中到达的原生阈值/overflow 压缩;不污染事件流)=====
+      case 'compaction_start':
+        this.compacting = true
+        return []
+
+      case 'compaction_end': {
+        this.compacting = false
+        const result = (event as { result?: CompactionResult }).result
+        if (result?.summary) {
+          void this.harvestCompaction(
+            result.summary,
+            result.tokensBefore,
+            result.tokensAfter ?? result.estimatedTokensAfter,
+            (event as { reason?: string }).reason,
+          )
+        }
+        return []
+      }
+
+      // 回合完全落定(无自动重试/压缩重试/排队续跑):后续压缩检查由 AgentRuntime 驱动
+      case 'agent_settled':
+        return []
+
+      case '__error__':
+        return [{
+          kind: 'error',
+          error: { code: 'OMP_ERROR', message: (event as { error?: string }).error ?? 'omp 未知错误' },
+        }]
+
+      default:
+        return []
+    }
+  }
+
+  // ===== 内部:omp 客户端管理 =====
+
+  private async ensureClient(ctx: AgentRunContext): Promise<void> {
+    if (!this.workspace) {
+      this.workspace = ctx.workspace
+    }
+    if (!this.agentInfo) {
+      // init() 可能未调用;从 ctx 推断
+      this.channelId = ctx.channelId
+      this.agentRole = ctx.role
+      this.bridgeCtx.identity = { agentId: this.selfAgentId, channelId: this.channelId, role: this.agentRole, name: this.agentName }
+    }
+
+    if (!this.client || !this.client.alive) {
+      // 旧客户端已退出(exit 事件/OS 存活校准 reconcile 触发)——必须丢弃并重生,
+      // 否则后续回合会一直对着死 stdio 报 PROMPT_FAILED 烧重试配额
+      this.client = null
+      this.hostToolsRegistered = false
+      const command = this.config.command ?? 'omp'
+      const client = new OmpRpcClient({
+        command,
+        mode: this.rpcMode,
+        args: [...(this.config.thinkingLevel ? ['--thinking', this.config.thinkingLevel] : [])],
+        args: this.config.args,
+        cwd: this.config.cwd ?? process.cwd(),
+        // 进程退出 → 注册表标记(供运行时资源监控);pid=-1 表示无法取得,忽略
+        onExit: (pid, code) => {
+          if (pid > 0) {
+            markHarnessProcessExit(pid, code)
+            markTerminalSessionExit(pid, code)
+          }
+        },
+      })
+      await client.start()
+
+      // 登记 + 绑定 agent 身份(harness 进程监控的事实源)
+      const pid = client.pid
+      if (pid) {
+        registerHarnessProcess(pid, {
+          harness: 'omp',
+          command,
+          args: ['--mode', this.rpcMode, ...(this.config.args ?? [])],
+        })
+        bindHarnessProcess(pid, {
+          agentId: this.selfAgentId,
+          channelId: this.channelId,
+          name: this.agentName,
+          role: this.agentRole,
+        })
+        // 终端镜像 tap:全部 RPC 帧 → /monitor 终端(实时 TUI 渲染 + HITL)
+        attachTerminalTap(client, {
+          pid,
+          harness: 'omp',
+          agentId: this.selfAgentId,
+          channelId: this.channelId,
+          name: this.agentName,
+          role: this.agentRole,
+        })
+      }
+
+      // 设置模型(如果配置了 provider/model 任一;omp 以当前已设值为缺省补全,
+      // 只配其一时也显式发送,避免 respawn 后静默回落默认模型)
+      if (this.config.provider || this.config.model) {
+        try {
+          // 未配置一侧保持 undefined(JSON 序列化丢键)→ omp 以当前已设值补全
+          await client.send({
+            type: 'set_model',
+            provider: this.config.provider as string,
+            modelId: this.config.model as string,
+          })
+        }
+        catch {
+          // 模型设置失败不致命(用 omp 默认模型)
+        }
+      }
+
+      // 注册 host tools(按角色差异化:lead 全量,worker 剔除调度/团队管理专属工具)
+      client.onHostToolCall(req => this.handleHostTool(req))
+      await client.send({ type: 'set_host_tools', tools: hostToolsForRole(this.agentRole, this.channelId) })
+
+      // 上下文治理探测(feature-detect 一次;失败退化被动 usage 跟踪)+ 原生压缩兜底保持开启
+      void this.probeContext(client).catch(() => {})
+      void client.send({ type: 'set_auto_compaction', enabled: true }).catch(() => {})
+
+      this.client = client
+      this.hostToolsRegistered = true
+    }
+  }
+
+  /** 上下文状态探测:get_state 取 contextWindow/sessionId(一次;失败仅损失 percent 精度) */
+  private async probeContext(client: OmpRpcClient): Promise<void> {
+    try {
+      const resp = await client.send({ type: 'get_state' })
+      const data = (resp.data ?? {}) as Record<string, unknown>
+      const model = data.model as { contextWindow?: number } | undefined
+      if (typeof model?.contextWindow === 'number' && model.contextWindow > 0) {
+        this.contextWindow = model.contextWindow
+        client.setContextWindow(model.contextWindow)
+      }
+      if (typeof data.sessionId === 'string' && data.sessionId) this.sessionId = data.sessionId
+    }
+    catch { /* get_state 不可用:仅被动 usage */ }
+  }
+}