@zhushanwen/pi-subagent-workflow 8.4.0 → 8.6.0

package/src/execution/stream-sink.ts

@@ -16,6 +16,10 @@
  * - dispose 清除 widget + 清 timer
  */
 
+import type { ExtensionMode } from "./host-mode.ts";
+import { resolveHostMode } from "./host-mode.ts";
+import { isRelayActive } from "./relay-env.ts";
+
 /** UI streaming sink 的最小接口（ctx.ui.setWidget 的 duck-typed 子集）。
  *
  * 当前只有一个 adapter（index.ts session_start 包装 ctx.ui.setWidget）。
@@ -98,3 +102,25 @@ export class SubagentStream {
     ]);
   }
 }
+
+/**
+ * background stream 创建策略（widget 私货通道退役步骤 2，E 方案 §6.3）。
+ *
+ * GUI（mode=rpc）+ relay 激活时不创建——drawer 实时改由 runtime tee 供数
+ * （session.subagentEntriesAppended 帧），extension 停发 setWidget 私货，主 pi stdout
+ * 卸载。TUI（widget 行是 TUI 用户的实时预览）与 relay 未激活环境（独立 pi / 无
+ * xyz-agent runtime）原样创建——universal 包自足性不绑架。
+ *
+ * env/mode 显式入参而非内部读 process.env：调用方语义（本 record 的宿主形态）一次
+ * 确定，测试可注入，函数保持纯判定。
+ */
+export function createBackgroundStream(
+  recordId: string,
+  sink: StreamSink | null,
+  mode: ExtensionMode | undefined,
+  env: NodeJS.ProcessEnv | Record<string, string | undefined>,
+): SubagentStream | undefined {
+  if (sink === null) return undefined;
+  if (isRelayActive(env) && resolveHostMode(mode) === "gui") return undefined;
+  return new SubagentStream(recordId, sink);
+}

package/src/execution/subagent-service.ts

@@ -29,6 +29,15 @@ import {
   tryTransition,
 } from "./execution-record.ts";
 import { doFinalizeRecord, doFinalizeRoundToIdle } from "./finalize-record.ts";
+import { getEngineDataDir } from "./engine/common/data-dir.ts";
+import { EngineError } from "./engine/common/errors.ts";
+import { JournalWriter } from "./engine/common/event-journal.ts";
+import { resolveJournalPath } from "./engine/paths.ts";
+import { executeOptionsToEngineTaskSpec } from "./engine/host-task-spec.ts";
+import type { EnginePort, RunContext } from "./engine/port.ts";
+import { DEFAULT_ENGINE_ID, getEngine } from "./engine/registry.ts";
+import { type EngineRouteResult, resolveEngineRouting, routeEngine } from "./engine/routing.ts";
+import type { AgentOutcome } from "./engine/types.ts";
 import { ManifestStore } from "./manifest-store.ts";
 import type { ModelConfigService } from "./model-config-service.ts";
 import type { AgentConfig, ModelInfo, ResolvedModel } from "./model-resolver.ts";
@@ -38,7 +47,7 @@ import { getSubagentRecordsDir, getSubagentSessionDir } from "./path-encoding.ts
 import type { StatusFilter } from "./record-store.ts";
 import { RecordStore } from "./record-store.ts";
 import { MAX_FORK_DEPTH } from "./session-context-resolver.ts";
-import { getChildByRecord, killAllSpawnedChildren, runSpawn, spawnedChildren, type SessionRunnerContext, type SpawnResumeOpts } from "./session-runner.ts";
+import { getChildByRecord, killAllSpawnedChildren, registerSpawnedChildForRecord, runSpawn, spawnedChildren, type SessionRunnerContext, type SpawnResumeOpts } from "./session-runner.ts";
 import { isIdle, isResumable, hasLiveProcessHandle } from "./lifecycle-predicates.ts";
 import { startIdleGc } from "./idle-gc.ts";
 import {
@@ -48,8 +57,8 @@ import {
   resetAllEpipeFailures,
   sendPromptCommand,
 } from "./stdin-writer.ts";
-import type { StreamSink } from "./stream-sink.ts";
-import { SubagentStream } from "./stream-sink.ts";
+import type { StreamSink, SubagentStream } from "./stream-sink.ts";
+import { createBackgroundStream } from "./stream-sink.ts";
 import { writeCancelledTombstone } from "./tombstone-store.ts";
 import type { WorktreeHandle } from "./types.ts";
 import type {
@@ -102,7 +111,7 @@ interface PiLike {
   events: { emit(channel: string, data: unknown): void };
   sendMessage(
     message: { customType: string; content: string; display: boolean; details?: unknown },
-    options?: { triggerTurn?: boolean; deliverAs?: "steer" | "followUp" | "nextTurn" },
+    options?: { triggerTurn?: boolean; deliverAs?: "steer" | "followUp" | "nextTurn" }, // g4-allow: 类型注解——PiLike 接口形状（pi.sendMessage 签面子集），非投递调用
   ): void;
   /** 订阅 pi 事件（D8：notifier 的 settled 边沿订阅用 'agent_settled'）。
    *  pi 0.84.1 的 on 返回 void 且无 off——退订语义由调用侧 disposed 标志包装兑现。
@@ -684,8 +693,47 @@ export class SubagentService {
     // ── 1. IDENTITY 解析（确认 → agentConfig → resolveModel）──
     const identity = await this.resolveIdentity(opts);
 
+    // ── 1.5 引擎路由（D4 chat 入口分叉；U2 升级为 routeEngine 编排）──
+    // 三层解析（调用参数 > agent frontmatter > config.json defaultEngine）仍是同步纯
+    // 函数；解析为非 pi 时升级走 routeEngine（probe 编排 + fallback 三守卫）。时机
+    // 选择：路由（含 probe）在 record 创建前完成——兜底时 record 直接按 pi 语义创建 +
+    // engineFallback 留痕（D5 字节级守护只约束「无 fallback 的纯缺省路径」，兜底路径
+    // 的 entry 允许含 engine/engineFallback 字段）；守卫命中/strict 时 routeEngine 在此
+    // throw，不产生孤儿 record。
+    const routingInput = {
+      callEngine: opts.engine,
+      agentEngine: identity.agentConfig?.engine,
+      globalDefaultEngine: this.modelService.getGlobalConfig().defaultEngine,
+    };
+    const routing = resolveEngineRouting(routingInput);
+    let route: EngineRouteResult | undefined;
+    if (routing.engineId !== DEFAULT_ENGINE_ID) {
+      route = await routeEngine({
+        routing: routingInput,
+        // 守卫 c 判据只看调用方显式指定的 model（resolved model 含 ctxModel 兼底，
+        // 恒非空会把一切兜底误判为 model 绑定命中）
+        taskModel: opts.model,
+        strict: this.modelService.getGlobalConfig().engineRouting?.strict === true,
+        probe: (engineId) => getEngine(engineId).probe(),
+      });
+      if (route.engineId !== DEFAULT_ENGINE_ID) {
+        return this.executeViaEngine(opts, identity, route);
+      }
+      // 兜底成功（典型：默认路由 + probe 失败 + 无守卫命中）→ 落回下方 pi 主路径，
+      // record 创建时按 pi 语义 + engine/engineFallback 留痕（engine = 实际执行引擎）
+    }
+    // D5 字节级守护：无 fallback 的 pi 路由剥掉 opts.engine——createRecordForMode
+    // 不盖章（pi record entry 序列化产物不得新增 engine 键，undefined 经 JSON 省略）。
+    // 兜底路径显式盖 engine='pi' + engineFallback（见上方时机注释）。
+    const piOpts =
+      route?.engineFallback !== undefined
+        ? { ...opts, engine: DEFAULT_ENGINE_ID, engineFallback: route.engineFallback }
+        : opts.engine === undefined
+          ? opts
+          : { ...opts, engine: undefined };
+
     // ── 2. RECORD 创建 + 注册 ──
-    const record = this.createRecordForMode(identity, opts, mode);
+    const record = this.createRecordForMode(identity, piOpts, mode);
     emitPendingRegister(this.pi, record.id, record.agent);
 
     // ── 2.5 worktree 创建（仅 worktree===true 或已传入 handle 时）──
@@ -724,7 +772,7 @@ export class SubagentService {
     // ── 4-7. background 包 detached 立即返回 id ──
     // background detached 运行对 tool 层不可见，完成由 notify 驱动新 turn。
     const bgDetails = project(record);
-    this.kickOffBackground(record, { ...opts, worktree: worktreeHandle }, ctx, identity, signal, priority);
+    this.kickOffBackground(record, { ...piOpts, worktree: worktreeHandle }, ctx, identity, signal, priority);
     return { mode: "background", subagentId: record.id, sessionFile: record.sessionFile, details: bgDetails };
   }
 
@@ -919,6 +967,20 @@ export class SubagentService {
         sendPromptCommand(child, text, { streamingBehavior: interrupt ? "steer" : "followUp" });
         // 热路径成功，清零 EPIPE 连续失败计数（[v4 A-1] 计数器已迁移到 stdin-writer）
         clearEpipeFailure(record.id);
+        // [race-F5] 写后死进程检测：write 同步成功只代表数据进了内核 pipe 缓冲，子进程
+        // 可能在读取前死亡（gate/idle kill 竞速）。exitCode/signalCode 已非 null = 进程已死
+        //（close 事件可能尚未到达），缓冲中的消息将被静默丢弃。只 warn 留证（含 runId 与
+        // 消息类型），不抛错不重试：终态回收已由 kill 路径保证，对死进程重试反而可能二次写。
+        if (child.exitCode !== null || child.signalCode !== null) {
+          logger.warn(
+            `[subagents] deliverMessage: child ${record.id} died around stdin write, message may be lost`,
+            {
+              msgType: interrupt ? "steer" : "followUp",
+              exitCode: child.exitCode,
+              signalCode: child.signalCode,
+            },
+          );
+        }
         // 轮始执行态信号清除 + 迁移上报（residual-fixes U3 补全，与冷路径 resumeRound
         // 对称）：新一轮开跑 = 无轮终信号——清上一轮 result（§5.4 isStreaming 公式要求
         // result undefined 才显示 streaming）与 resumable，appendEntry 让 runtime/W18
@@ -1344,7 +1406,15 @@ export class SubagentService {
     // 直接用 override → 主 agent model。DEFAULT_AGENT_NAME 仅作 record 显示名
     // （TUI 层 extractAgentName 共用，保证显示一致）。
     const agent = opts.agent ?? DEFAULT_AGENT_NAME;
-    const agentConfig = opts.agent ? this.modelService.getAgentConfig(opts.agent) : undefined;
+    // 显式 agent ref（用户点名）失败必须报错，不静默降级：无 require 的 loadByPath
+    // 对相对路径/裸名/文件缺失都返回 undefined → agentConfig undefined → resolveModel
+    // 静默回落 override→主 agent model，用户拿到的 subagent 无 systemPrompt/工具白名单
+    // 且零反馈。require:true 让失败抛出带 <available_subagents> 指引的错误（对齐
+    // workflow name not found 反馈风格）；不传 agent = 默认 general-purpose 语义，
+    // agentConfig 保持 undefined（合法缺省，走 override → ctxModel 兑底）。
+    const agentConfig = opts.agent
+      ? this.modelService.getRequiredAgentConfig(opts.agent)
+      : undefined;
 
     const resolved = this.modelService.resolveModel(
       opts.agent ?? "",
@@ -1389,6 +1459,10 @@ export class SubagentService {
       depth,
       chatMode: opts.conversation === true,
       idleTimeoutMs: opts.idleTimeoutMs,
+      // P4 引擎留痕（D9①）：opts.engine/engineFallback 由引擎适配层写入（PiEngine.run
+      // 从 RunContext 回填；缺省 = pi 投影，存量调用方零感知）
+      engine: opts.engine,
+      engineFallback: opts.engineFallback,
       controller,
     });
 
@@ -1403,6 +1477,194 @@ export class SubagentService {
     return { mode: "background", subagentId: record.id, sessionFile: record.sessionFile, details };
   }
 
+  // ── 引擎分支（D4/D10：非 pi 引擎的 chat 域执行骨架，U0）──────────
+
+  /**
+   * 路由到非 pi 引擎的执行入口：routeEngine（注册表校验 + probe/守卫）已由 execute
+   * 完成——这里只剩 unsupported 预检 → record 创建+盖章 → detached 引擎 run。
+   * 全部同步拒绝发生在 record 创建前（不产生孤儿 record）。
+   */
+  private executeViaEngine(
+    opts: ExecuteOptions,
+    identity: ResolvedIdentity,
+    route: EngineRouteResult,
+  ): ExecutionHandle {
+    const engine = route.engine;
+    this.assertEngineParamSupport(engine, opts);
+    // record 盖章路由结果（D5 仅 pi 缺省不盖章；非 pi 显式留痕，createRecordForMode
+    // 经 opts.engine/engineFallback 读入 record identity——engine 为实际执行引擎，
+    // fallback 路径 from=请求引擎留痕，probe 通过的常态路径恒缺省）
+    const record = this.createRecordForMode(
+      identity,
+      {
+        ...opts,
+        engine: route.engineId,
+        ...(route.engineFallback !== undefined ? { engineFallback: route.engineFallback } : {}),
+      },
+      "background",
+    );
+    emitPendingRegister(this.pi, record.id, record.agent);
+    this.kickOffEngineRun(record, opts, engine);
+    return { mode: "background", subagentId: record.id, sessionFile: record.sessionFile, details: project(record) };
+  }
+
+  /**
+   * 非 pi 引擎的 unsupported 参数预检（D11 处置「调用前拒绝」的判据 = capabilities）。
+   * conversation / fork / worktree 三参数对首期接入的引擎（zcode）均不可用：
+   * conversation 依赖同进程 idle 复用、fork 依赖父 pi session 上下文继承、worktree 依赖
+   * 文件隔离（capabilities.sandbox='none'）。同步 throw，文案含 capabilities 依据与恢复指引。
+   */
+  private assertEngineParamSupport(engine: EnginePort, opts: ExecuteOptions): void {
+    const caps = engine.capabilities();
+    if (opts.conversation === true && caps.conversation === "unsupported") {
+      throw new EngineError(
+        "engine_capability_unsupported",
+        `engine '${engine.id}' 不支持 conversation（capabilities.conversation = 'unsupported'，` +
+          `spawn 单轮模式无同进程 idle 复用，message/close 交互控制面不可用）`,
+        `改用 engine: pi（支持 conversation 续聊），或不传该参数（一次性任务默认形态）`,
+      );
+    }
+    if (opts.fork === true) {
+      throw new EngineError(
+        "engine_capability_unsupported",
+        `engine '${engine.id}' 不支持 fork（fork 依赖父 pi session 上下文继承，` +
+          `capabilities.steer = '${caps.steer}'——非 pi 引擎无父 session 分叉通道）`,
+        `把所需父上下文写进 task 正文后不传 fork，或改用 engine: pi`,
+      );
+    }
+    if ((opts.worktree === true || typeof opts.worktree === "object") && caps.sandbox === "none") {
+      throw new EngineError(
+        "engine_capability_unsupported",
+        `engine '${engine.id}' 不支持 worktree 隔离（capabilities.sandbox = 'none'，` +
+          `引擎未接文件系统隔离层）`,
+        `改用 engine: pi（worktree 隔离可用），或不传该参数（在 parent cwd 执行）`,
+      );
+    }
+  }
+
+  /**
+   * 非 pi 引擎的 detached 执行编排（与 kickOffBackground 同构的 background 语义）：
+   * pool 并发槽（maxConcurrent 对非 pi 引擎同样生效）→ journal 接线（D6 第②级：
+   * taskId=record.id，初始池 key 占位 'shared'，onPoolResolved retarget 到引擎实际
+   * 池 key——路径与 paths.ts 同源推导）→ engine.run（signal 接 record controller，
+   * kill-chain 两级生效）→ engineHandle 回填（终态迁移落 entry 前）→ 终态迁移 →
+   * bg notify（chat 域宿主职责，与 pi 完成通知同语义）。
+   */
+  private kickOffEngineRun(record: ExecutionRecord, opts: ExecuteOptions, engine: EnginePort): void {
+    const signal = record.controller?.signal;
+    void (async () => {
+      try {
+        await this.pool.acquire(PRIORITY_BACKGROUND, this.effectiveMaxConcurrentFor(record), signal);
+      } catch {
+        // S1: 排队中被 abort（signal.aborted）走 cancelled，与已运行被 abort 一致（runAndFinalize 同款）
+        if (signal?.aborted) {
+          await this.finalizeAborted(record);
+        } else {
+          await this.finalizeFailed(record, new Error("aborted"));
+        }
+        return;
+      }
+      // [review MF1] acquire 成功后必须 finally release：不 release 则
+      // DefaultConcurrencyPool._active 永不递减——每次引擎任务泄漏一个并发槽，累计
+      // maxConcurrent 次后全部 background subagent（pi 与引擎共用同一池）在 acquire 队列永久挂起
+      try {
+        await this.runEngineTask(record, opts, engine, signal);
+        // cancel 抢先（closedReason='cancelled'）时 cancelBackground 自己 notify，跳过
+        if (record.closedReason !== "cancelled") {
+          this.notifyComplete(record);
+        }
+      } finally {
+        this.pool.release();
+      }
+    })();
+  }
+
+  /**
+   * kickOffEngineRun 的 acquire 后主体：journal 接线（D6 第②级：taskId=record.id，
+   * 初始池 key 占位 'shared'，onPoolResolved retarget 到引擎实际池 key）→ engine.run
+   * （signal 接 record controller，kill-chain 两级生效）→ engineHandle 回填（终态迁移
+   * 落 entry 前）→ 终态迁移。bg notify 归编排侧（与 kickOffBackground 收尾通知归编排对称）。
+   */
+  private async runEngineTask(
+    record: ExecutionRecord,
+    opts: ExecuteOptions,
+    engine: EnginePort,
+    signal: AbortSignal | undefined,
+  ): Promise<void> {
+    const journal = new JournalWriter({
+      path: resolveJournalPath(getEngineDataDir(), engine.id, "shared", record.id),
+      taskId: record.id,
+      engineId: engine.id,
+    });
+    const retargetJournal = (poolKey: string): void => {
+      journal.retarget(resolveJournalPath(getEngineDataDir(), engine.id, poolKey, record.id));
+    };
+    // 对齐点③：journal 路径权威 = 引擎声明的池 key（writer 初始用占位，retarget 后
+    // 与 handle.poolKey 同源）。模式对齐 SAR 的 journalingOnEvent：先落盘再转发。
+    const runCtx: RunContext = {
+      taskId: record.id,
+      poolKey: "shared",
+      signal,
+      ctxModel: opts.ctxModel,
+      onEvent: (event) => journal.append(event),
+      onPoolResolved: retargetJournal,
+      // D9①：路由层 fallback 留痕投影进 outcome（zcode 无独立 record 通路）
+      ...(record.engineFallback !== undefined ? { engineFallback: record.engineFallback } : {}),
+      // D10 终止链：engine spawn 的子进程注册进 spawnedChildren 记账
+      //（cancelBackground SIGTERM / dispose killAll 收割对非 pi record 生效）
+      onChildSpawned: (child) => registerSpawnedChildForRecord(record.id, child),
+    };
+    try {
+      const { handle, outcome } = await engine.run(executeOptionsToEngineTaskSpec(opts), runCtx);
+      // engineHandle 完整回填（U2：终态迁移落 entry 前）。sessionRef 整体透传——
+      // 失败终态 sessionId 缺失时也回填已有部分（dbPath/poolKey），读侧①级降②级
+      // 的防御形态；journalPath 取 retarget 后的实际落盘路径（writer 是路径权威）。
+      record.engineHandle = {
+        sessionRef: handle.data.sessionRef,
+        poolKey: handle.data.poolKey,
+        journalPath: journal.path,
+      };
+      await journal.close();
+      await this.finalizeEngineOutcome(record, outcome);
+    } catch (err) {
+      // engine.run prepare 期 reject（进程创建前）→ failed 终态（与 runAndFinalize catch 同语义）；
+      // journal 尽力而为收口（②级数据源写失败已由 writer 内部 warn 收敛）
+      await journal.close();
+      await this.finalizeFailed(record, err);
+    }
+  }
+
+  /**
+   * 分层并发配额：depth 越深可用配额越少（下限 1）。fork 深度护栏在池维度的投影，
+   * 公式约定以 concurrency-pool.ts 注释为登记处、此处为唯一代码锚点。
+   */
+  private effectiveMaxConcurrentFor(record: ExecutionRecord): number {
+    return Math.max(1, this.pool.maxConcurrent - record.depth);
+  }
+
+  /**
+   * engine.run resolve 的终态迁移：outcome.error → failed（success=false + error 文案）；
+   * 否则 done（result=content）。CAS 抢锁（tryTransition）防与 cancelBackground 双收尾。
+   */
+  private async finalizeEngineOutcome(record: ExecutionRecord, outcome: AgentOutcome): Promise<void> {
+    if (outcome.sessionFile !== undefined) {
+      record.sessionFile = outcome.sessionFile;
+    }
+    const result: AgentResult = {
+      text: outcome.content,
+      turns: outcome.usage?.turns ?? 0,
+      durationMs: outcome.durationMs ?? Date.now() - record.startedAt,
+      success: outcome.error === undefined,
+      ...(outcome.error !== undefined ? { error: outcome.error } : {}),
+      sessionId: outcome.sessionId ?? record.id,
+      toolCalls: [],
+      ...(outcome.parsedOutput !== undefined ? { parsedOutput: outcome.parsedOutput } : {}),
+    };
+    if (tryTransition(record, "closed", "gc")) {
+      await this.finalizeRecord(record, result, "closed", "gc");
+    }
+  }
+
   // ── 执行内部：run + finalize（sync/bg 共用）──────────────
 
   /** 共享的"干活 + 收尾"——sync 直接 await，background 在 detached 里调。 */
@@ -1421,9 +1683,8 @@ export class SubagentService {
     const pooled = record.mode === "background";
     let acquired = false;
     if (pooled) {
-      const effectiveMaxConcurrent = Math.max(1, this.pool.maxConcurrent - record.depth);
       try {
-        await this.pool.acquire(priority, effectiveMaxConcurrent, signal);
+        await this.pool.acquire(priority, this.effectiveMaxConcurrentFor(record), signal);
         acquired = true;
       } catch {
         // S1: 排队中被 abort（signal.aborted）走 cancelled，与已运行被 abort 一致。
@@ -1575,10 +1836,9 @@ export class SubagentService {
     /** resume 选项（M2-B1）：透传 runAndFinalize→runSpawn。undefined = 新 session。 */
     resume?: SpawnResumeOpts,
   ): void {
-    // 创建 streaming 生命周期对象——streamSink 为 null（session_start 未注入）时降级为 undefined。
-    const stream = this.streamSink
-      ? new SubagentStream(record.id, this.streamSink)
-      : undefined;
+    // 创建 streaming 生命周期对象。策略（含 widget 退役步骤 2：GUI + relay 激活时停发
+    // 私货、TUI/未激活原样创建、sink 未注入降级 undefined）集中在 createBackgroundStream。
+    const stream = createBackgroundStream(record.id, this.streamSink, ctx.mode, process.env);
 
     void this.runAndFinalize(
       record, opts, ctx, identity, signal, priority,