@zhushanwen/pi-subagent-workflow 8.13.0 → 8.14.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@zhushanwen/pi-subagent-workflow",
-  "version": "8.13.0",
+  "version": "8.14.0",
   "type": "module",
   "main": "index.ts",
   "description": "Unified subagent execution and multi-agent workflow orchestration for Pi — spawned-process agent runtime with sync/background modes, stateful workflow management with persistence, state machine, and execution tracing.",
@@ -54,21 +54,21 @@
     ]
   },
   "dependencies": {
-    "@xyz-agent/extension-protocol": "0.10.0",
-    "@xyz-agent/session-delivery": "0.3.1",
-    "@zhushanwen/pi-ext-guards": "0.3.0",
-    "@zhushanwen/pi-extension-logger": "0.6.0",
-    "@zhushanwen/subagent-core": "0.10.0",
-    "@zhushanwen/pi-subagent-cli": "0.3.0",
-    "@zhushanwen/zcode-subagent-cli": "0.3.0"
+    "@xyz-agent/extension-protocol": "0.11.0",
+    "@xyz-agent/session-delivery": "0.4.1",
+    "@zhushanwen/pi-ext-guards": "0.4.0",
+    "@zhushanwen/subagent-core": "0.10.1",
+    "@zhushanwen/pi-subagent-cli": "0.4.0",
+    "@zhushanwen/zcode-subagent-cli": "0.3.1",
+    "@zhushanwen/pi-extension-logger": "0.6.0"
   },
   "peerDependencies": {
     "@earendil-works/pi-ai": "^0.84.4",
     "@earendil-works/pi-coding-agent": "^0.84.4",
     "@earendil-works/pi-tui": "^0.84.4",
     "typebox": "*",
-    "@zhushanwen/pi-pending-notifications": "0.7.1",
-    "@zhushanwen/pi-structured-output": "5.1.5"
+    "@zhushanwen/pi-pending-notifications": "0.7.2",
+    "@zhushanwen/pi-structured-output": "5.1.7"
   },
   "peerDependenciesMeta": {
     "@earendil-works/pi-coding-agent": {

package/src/host/inflight-reporter.ts

@@ -37,9 +37,10 @@
 // （30min 有界），不丢 errs-safe 兜底。
 
 import type { ExtensionContext } from "@earendil-works/pi-coding-agent";
-import { SUBAGENT_INFLIGHT_MARKER, isInFlightReportAck } from "@xyz-agent/extension-protocol";
+import { SUBAGENT_INFLIGHT_MARKER, callMarkerRpc, isInFlightReportAck } from "@xyz-agent/extension-protocol";
 import { getInFlightSnapshot } from "@zhushanwen/subagent-core";
 import { getLogger } from "@zhushanwen/pi-extension-logger";
+import { toErrorMessage } from "@zhushanwen/pi-ext-guards";
 
 /** select 通道级超时（控制面单请求，秒级校准——超时默认原则规则 19）。取值对齐
  *  plugin-bridge 启动 sync 的 2s 自愈闸：session_start 首帧可能早于 runtime adapter
@@ -142,25 +143,36 @@ export function createInFlightReporter(opts: InFlightReporterOpts = {}): InFligh
       sessionId: getSessionId(active),
       emittedAt: Date.now(),
     });
-    let value: unknown;
-    try {
-      value = await active.ui.select(SUBAGENT_INFLIGHT_MARKER, [payload], { timeout: selectTimeoutMs });
-    } catch (err) {
-      // 通道异常折叠（plugin-bridge callBridge 同款：不静默吞，但只首败 warn）。
-      value = undefined;
-      logFailure("select channel threw", err);
-    }
+    // 发送+折叠半边走 protocol 的 callMarkerRpc 原语（D8，fire-and-forget：void 发起
+    // 不变）：ok:false 四态（cancelled/timeout/channel-error/non-json）统一折叠进下方
+    // 延迟重试路径；送达判据 = ack 全等匹配（不是 JSON 消费），留在本侧。原语的失败
+    // 留痕经注入的 log 承载本侧「首败 warn / 后续 debug」防刷屏策略。
+    // guiCtx = ExtensionContext 的 GuiContext 最小子集（ask-user runRpcInteraction 同款
+    // 先例：ui.custom 泛型签名静态不兼容，callMarkerRpc 只读 ui.select）。
+    const guiCtx = {
+      mode: active.mode,
+      hasUI: active.hasUI,
+      ui: { select: active.ui.select.bind(active.ui) },
+    };
+    const result = await callMarkerRpc(guiCtx, SUBAGENT_INFLIGHT_MARKER, payload, {
+      timeout: selectTimeoutMs,
+      log: primitiveLog,
+    });
     attemptInFlight = false;
-    if (typeof value === "string" && isInFlightReportAck(value)) {
+    if (result.ok && isInFlightReportAck(result.value)) {
       // 送达确认：清重试与失败计数，补推积压脏帧。
       failureCount = 0;
       clearRetryTimer();
       if (dirty && ctx !== null) kick();
       return;
     }
-    // 失败折叠（resolve undefined = 超时/取消/无路由）→ 延迟重试，累计到顶放弃
-    //（放弃后镜像按 absent-report 走 errs 推迟，30min 有界，errs-safe 兜底不丢）。
-    logFailure("no ack (timeout, cancelled, or runtime without marker routing)", value);
+    // 失败折叠（resolve undefined = 超时/取消/无路由 / 回包非 ack / 通道异常）→ 延迟
+    // 重试，累计到顶放弃（放弃后镜像按 absent-report 走 errs 推迟，30min 有界，
+    // errs-safe 兜底不丢）。
+    logFailure(
+      result.ok ? "no ack (non-ack response)" : `no ack (${result.reason})`,
+      result.ok ? result.value : undefined,
+    );
     failureCount += 1;
     if (failureCount >= maxAttempts) {
       givenUp = true;
@@ -181,11 +193,17 @@ export function createInFlightReporter(opts: InFlightReporterOpts = {}): InFligh
     }
   }
 
+  /** 原语留痕注入（D8）：msg/detail 由 callMarkerRpc 产出；防刷屏策略（首败 warn /
+   * 后续 debug）留本侧 logFailure。detail 是原语侧小对象，序列化保信息。 */
+  function primitiveLog(msg: string, detail?: object): void {
+    logFailure(msg, detail === undefined ? undefined : JSON.stringify(detail));
+  }
+
   function logFailure(reason: string, detail: unknown): void {
     if (!firstFailureLogged) {
       firstFailureLogged = true;
       logger.warn(`[subagent-inflight] in-flight report failed (${reason}); retrying every ${retryDelayMs}ms (bounded at ${maxAttempts} attempts)`, {
-        detail: detail instanceof Error ? detail.message : String(detail),
+        detail: toErrorMessage(detail),
       });
       return;
     }

package/src/injectors/model-list-injector.ts

@@ -42,6 +42,7 @@ import { getLogger } from "@zhushanwen/pi-extension-logger";
 
 // C5①/C5⑦：渲染统一走 core barrel（formatModelList + ModelEntry 类型为 barrel 导出面）
 import { formatModelList, type ModelEntry } from "@zhushanwen/subagent-core";
+import { toErrorMessage } from "@zhushanwen/pi-ext-guards";
 
 const logger = getLogger("injector");
 
@@ -87,7 +88,7 @@ export function setupModelListInjector(pi: ExtensionAPI): void {
 				return { systemPrompt: event.systemPrompt + injection };
 			} catch (err) {
 				logger.error("[model-list-injector] before_agent_start failed", {
-					reason: err instanceof Error ? err.message : String(err),
+					reason: toErrorMessage(err),
 				});
 			}
 		},

package/src/interface/format.ts

@@ -13,6 +13,8 @@ import os from "node:os";
 
 import { truncateToWidth, visibleWidth } from "@earendil-works/pi-tui";
 
+import { firstContentText } from "@xyz-agent/extension-protocol";
+
 import type { AgentEventLogEntry, DisplayItem, ExecutionStatus } from "@zhushanwen/subagent-core";
 import { DEFAULT_AGENT_NAME } from "@zhushanwen/subagent-core";
 import type {
@@ -562,12 +564,13 @@ export function formatTokenStat(
 /**
  * renderResult 的文本兜底：从 result.content[0] 提取纯文本。
  * 多处 tool 的 renderResult 曾各自内联此逻辑，提取后统一调用。
+ * 内核收敛至 extension-protocol firstContentText（ext-simplify-17 D9）；
+ * 本包装保留 content 可选的宽入参形态（tool renderResult 回调契约 content 可缺省）。
  */
 export function renderTextFallback(
   result: { content?: Array<{ type: string; text?: string }> },
 ): string {
-  const first = result.content?.[0];
-  return first?.type === "text" ? (first.text ?? "") : "";
+  return firstContentText({ content: result.content ?? [] });
 }
 
 /** Format a single activity line: ToolName(argsPreview). */

package/src/interface/subagent-tool-schema.ts

@@ -28,7 +28,7 @@ export { SLUG_MAX_LENGTH };
 
 // Params schema（跨包契约测试的真实 typebox 校验入口）。
 //
-// action:"start" 的 17 字段（task/slug/agent/model/...）拍平在顶层，不再用 startParam
+// action:"start" 的 16 字段（task/slug/agent/model/...）拍平在顶层，不再用 startParam
 // 嵌套容器包。原因：弱模型（GLM/DeepSeek）信任 schema 结构信号 > 文本信号，经常省略
 // startParam 嵌套层把 task/slug 直接平铺到顶层导致调用失败。拍平后 schema 结构与模型
 // 的自然倾向一致，消除这层误用。task/slug 必填性由 startHandler runtime 校验（flat
@@ -40,7 +40,7 @@ export { SLUG_MAX_LENGTH };
 // 反映必填性。勿在此基础上继续堆 action 条件逻辑——要加就拆 tool。
 export const SubagentParams = Type.Object({
   action: StringEnum(["start", "list", "cancel", "message", "close", "fork-from"], {
-    description: "Operation: 'start' runs a subagent, 'list' shows subagents, 'cancel' stops a background subagent, 'message' sends a follow-up to any of your subagents (running or idle — an idle one transparently revives and continues on its original session file; one-shot subagents are auto-upgraded to conversation mode on first message), 'close' archives a subagent (immediately when idle; after the current round, or immediately with force:true, when running), 'fork-from' spawns a NEW subagent inheriting an older one's history (recovery for restart-disconnected subagents; the old record is untouched).",
+    description: "Operation: 'start' runs a subagent, 'list' shows subagents, 'cancel' stops a background subagent, 'message' sends a follow-up to any of your subagents (running or idle — an idle one transparently revives and continues on its original session file), 'close' archives a subagent (immediately when idle; after the current round, or immediately with force:true, when running), 'fork-from' spawns a NEW subagent inheriting an older one's history (recovery for restart-disconnected subagents; the old record is untouched).",
   }),
   // ── action:"start" fields (flattened to top level). task/slug REQUIRED for start. ──
   // Missing/empty task or slug throws at runtime (startHandler).
@@ -88,40 +88,34 @@ export const SubagentParams = Type.Object({
     description: 'Override the working directory for the subagent execution. Must be an absolute path (no "~" shorthand, no relative paths); ".." segments are rejected. Defaults to the parent session\'s cwd.',
     pattern: "^/",
   })),
-  conversation: Type.Optional(Type.Boolean({
-    description:
-      "Enable continuous chat with this subagent. When true, the subagent stays available after each reply — you can send follow-up messages (action:'message') and it keeps the full conversation context across rounds, with no need to re-spawn or re-explain. " +
-      "\nUse conversation:true for: multi-round collaboration (iterative review-fix loops, back-and-forth refinement), any task where you expect to send follow-up messages after the initial result. " +
-      "\nOmit (or false) for: one-shot tasks — single exploration, lookup, file read, code generation that needs no follow-up. The subagent runs once, notifies on completion, and is cleaned up automatically (default). " +
-      "\nFor long-interval collaboration (each round spaced >5min apart), set conversation:true AND increase idleTimeoutMs to avoid premature timeout. " +
-      "Cost: a conversation-mode subagent holds resources (memory, and a worktree if enabled) until you explicitly end it with action:'close'. Always close when done.",
-  })),
+  // [modeless 波5] conversation 参数已删除：chatMode 字段消亡后「模式」不存在——
+  // 一切 record 永续可续聊（idle 后 message 即续、fork-from 可继承），无模式开关可表达。
   idleTimeoutMs: Type.Optional(Type.Number({
     description:
-      "Idle timeout in milliseconds for conversation-mode subagents. Controls how long an idle subagent (between rounds) stays alive before automatic cleanup. " +
-      "Default: 300000 (5min). Override for long-interval collaboration where each round is spaced >5min apart. " +
-      "Pass 0 or a negative value to DISABLE idle cleanup entirely (subagent stays alive until explicitly closed). " +
-      "Only meaningful with conversation:true; ignored for one-shot subagents.",
+      "Idle-recycle cadence for ALL subagents (modeless: every subagent stays continuable — this is NOT a mode switch). Controls how long an idle subagent (between rounds, no activity) stays before being automatically archived. " +
+      "Default: 300000 (5min). Raise it for long-interval collaboration where your next message may arrive more than 5min after a round ends. " +
+      "Pass 0 or a negative value to DISABLE idle recycling entirely (subagent stays available until you close it). " +
+      "Priority: this param > env XYZ_SUBAGENT_IDLE_TIMEOUT_MS > default.",
   })),
   engine: Type.Optional(StringEnum(["pi", "zcode"], {
     description:
       "Execution engine for this subagent. Omit to inherit the global config. " +
       "Three-layer priority: this parameter > agent .md frontmatter engine > config.json defaultEngine. " +
-      "Non-pi engines do not support conversation/fork/worktree (rejected before the subagent is created).",
+      "Non-pi engines do not support fork/worktree (rejected before the subagent is created).",
   })),
   collect: Type.Optional(StringEnum(["async", "sync"], {
     description:
-      "Completion-notification collection mode for one-shot subagents (subagent-sync-collect). " +
+      "Completion-notification routing (NOT a record mode — batch membership is routing bookkeeping only). " +
       "Omit to use the config default (currently async). " +
-      "Use 'sync' when you dispatch >=2 independent one-shot subagents whose results you will combine: " +
-      "their completions are held until ALL pending sync subagents finish, then delivered as ONE batch " +
-      "notification (single wake-up, results inline). You may keep dispatching more sync subagents in " +
-      "later turns — they join the same pending batch. " +
-      "Independent means no member's prompt or work depends on another member's output — dependent " +
-      "tasks must be chained across messages (one start after the prior completes), never batched. " +
-      "Use 'async' (or omit) for immediate per-subagent completion notifications. " +
-      "Incompatible with conversation:true — that combination is rejected immediately before start; " +
-      "remove either conversation or collect.",
+      "'async' = each subagent's completion notifies immediately. " +
+      "'sync' = batch wake-up: when you dispatch >=2 independent subagents whose results you will " +
+      "combine, their completions are held until ALL pending sync members finish, then delivered as " +
+      "ONE batch notification (single wake-up, results inline); when the batch closes, its members " +
+      "are automatically archived. Batch members cannot be messaged — use action:'fork-from' to " +
+      "continue from one instead. You may keep dispatching more sync subagents in later turns — " +
+      "they join the same pending batch. Independent means no member's prompt or work depends on " +
+      "another member's output — dependent tasks must be chained across messages (one start after " +
+      "the prior completes), never batched.",
   })),
   // action:"list" → listParam OPTIONAL (all fields optional, defaults apply). Ignored by other actions.
   listParam: Type.Optional(Type.Object({
@@ -147,11 +141,12 @@ export const SubagentParams = Type.Object({
   })),
   // action:"message" → messageParam.subagentId + text REQUIRED. Any reachable subagent works —
   // running joins the in-flight round (D2 打断入队)；idle transparently revives on the same
-  // session file（[U4 §3.2.3] 万物可续——形态枚举 gate 消亡）；one-shot auto-upgrades to
-  // conversation mode on first message (SP-5)。description 与实现锚点见 messageHandler。
+  // session file（[U4 §3.2.3] 万物可续——形态枚举 gate 与 SP-5 升级路径均消亡，message 直接
+  // 续聊任何 idle record）。引擎续聊能力轴 gate 保留（core messageHandler 入口）。
+  // description 与实现锚点见 messageHandler。
   messageParam: Type.Optional(Type.Object({
     subagentId: Type.String({
-      description: "REQUIRED for action:'message'. The subagentId to message. Any subagent reachable in this session tree works, running or idle: an idle subagent transparently revives on the same id and continues writing its original session file (a one-shot is auto-upgraded to conversation mode on first message); a running subagent has your message interrupt-and-join its in-flight round. Rejections: unknown id, session file held by another live process, a record from a different session tree, workflow-origin records (their results belong to the workflow run), and one-shot records on engines that do not support conversation upgrade.",
+      description: "REQUIRED for action:'message'. The subagentId to message. Any subagent reachable in this session tree works, running or idle: an idle subagent transparently revives on the same id and continues writing its original session file; a running subagent has your message interrupt-and-join its in-flight round. Rejections: unknown id, session file held by another live process, a record from a different session tree, workflow-origin records (their results belong to the workflow run), and records on engines that do not support continuation (fork-from or re-dispatch instead).",
     }),
     text: Type.String({
       description: "REQUIRED for action:'message'. The message to send. Whitespace-only throws.",

package/src/interface/subagent-tool.ts

@@ -162,11 +162,11 @@ action:"list" before action:"start" — a reusable subagent may exist; compactio
 
 ## Actions
 
-- action:"start" — run a subagent. Pass task and slug as top-level fields (REQUIRED). Optional: agent, model, thinkingLevel, engine, collect, skillPath, appendSystemPrompt, schema, maxTurns, graceTurns, fork, worktree, cwd, conversation, idleTimeoutMs. Background only: returns a subagentId immediately, notifies on completion.
-- action:"message" — send a follow-up to any of your subagents — running or idle (idle revives in place; one-shots become conversation-mode); full context retained. REQUIRED messageParam: { subagentId, text }. The reply auto-notifies.
+- action:"start" — run a subagent. Pass task and slug as top-level fields (REQUIRED). Optional: agent, model, thinkingLevel, engine, collect, skillPath, appendSystemPrompt, schema, maxTurns, graceTurns, fork, worktree, cwd, idleTimeoutMs. Background only: returns a subagentId immediately, notifies on completion.
+- action:"message" — send a follow-up to any of your subagents — running or idle (idle revives in place; full context retained). REQUIRED messageParam: { subagentId, text }. The reply auto-notifies.
 - action:"close" — archive a subagent (hidden from list, recoverable): idle closes immediately; running finishes the current round first unless force:true (then terminates mid-round). REQUIRED closeParam: { subagentId }.
 - action:"list" — list subagents. listParam: { includeFinished?, includeWorkflow?, limit? } (all optional; includeWorkflow defaults false — workflow-dispatched subagents are hidden unless true). Read an item's sessionFile for full detail.
-- action:"cancel" — stop a background subagent (for conversation-mode use close). REQUIRED cancelParam: { subagentId }.
+- action:"cancel" — stop a background subagent (to archive it instead, use close). REQUIRED cancelParam: { subagentId }.
 - action:"fork-from" — restart-disconnect recovery: spawn a NEW subagent inheriting the old one's history via --fork. REQUIRED forkFromParam: { sourceSubagentId }. Optional: prompt (continuation; default handover frame). Returns { newSubagentId, sourceSessionFile }. Rejects still-running / foreign-live / worktree-bound sources; unparseable history anchors are guided to action:"message" (same-id reopen).
 
 ## Examples
@@ -191,8 +191,8 @@ Completion auto-notifies you (steer wakes the next turn):
 
 ## Batch collection (collect)
 
-- collect:"sync" — >=2 independent one-shot subagents whose results you will combine: completions are held until every pending sync member finishes, then ONE batch notification delivers all results inline (one wake-up). Later sync starts join the same batch; each sync start response reports {"collect":{"mode":"sync","pendingSyncCount":N}}.
-- collect:"async" (default, omit) — immediate per-subagent completion; for conversational work or when each result is needed early.
+- collect:"sync" — >=2 independent subagents whose results you will combine: completions are held until every pending sync member finishes, then ONE batch notification delivers all results inline (one wake-up) and batch members auto-archive. Later sync starts join the same batch; each sync start response reports {"collect":{"mode":"sync","pendingSyncCount":N}}. Batch members cannot be messaged — fork-from continues from one.
+- collect:"async" (default, omit) — immediate per-subagent completion; use when each result is needed early.
 - Subagents in one sync batch must not depend on each other's output — dependent tasks must be chained across messages (see Calling patterns), never batched.
 Items over budget are truncated with a pointer: session_read {"action":"result","session":"<id>"} fetches the full text.
 
@@ -204,16 +204,15 @@ Items over budget are truncated with a pointer: session_read {"action":"result",
 - Treating subagent results as authoritative without verification.
 - Canceling by guessing a subagentId instead of using action:"list" first.
 
-## Continuous chat (conversation mode)
+## Continuing a subagent (modeless)
 
-conversation:true keeps a subagent available across replies — action:"message" continues with full context, action:"close" releases it (always close when done). For review/fix loops and long-interval rounds (>5min apart, raise idleTimeoutMs); omit for one-shot tasks.
-idleTimeoutMs: idle timeout before auto-cleanup (default 300000 / 5min; env XYZ_SUBAGENT_IDLE_TIMEOUT_MS overrides globally, per-call wins).
+Every subagent stays continuable — no mode switch: action:"message" revives an idle record in place (or joins a running one's round), action:"fork-from" branches a new subagent from old history, action:"close" archives it.
+idleTimeoutMs: idle-recycle cadence for ALL subagents — idle records auto-archive on expiry (default 300000 / 5min; 0/negative disables; env XYZ_SUBAGENT_IDLE_TIMEOUT_MS: global default, per-call wins).
 
 ## You cannot
 
 - Get a synchronous/inline result — start always returns a subagentId immediately (background).
 - Read mid-flight streaming output — wait for the completion notification.
-- Combine collect:"sync" with conversation:true — rejected before start; sync is one-shot only (remove one).
 - See intermediate signals while a sync batch waits — nothing arrives until the whole batch closes. Hung member: action:"list" shows what is still running; action:"cancel" it — cancelled members count as terminal and the batch closes.
 
 ## Calling patterns

package/src/jsonl-run-store.ts

@@ -90,7 +90,7 @@ import {
   pruneStateFilesBeyondCap,
   type RunSnapshot,
 } from "@zhushanwen/subagent-core";
-import { toErrorMessage } from "@zhushanwen/pi-ext-guards";
+import { isEnoentError, toErrorMessage } from "@zhushanwen/pi-ext-guards";
 
 // ── Workflow-record self-describing entry (W17, D4) ─────────
 
@@ -226,16 +226,6 @@ async function loadRunFromStateFile(filePath: string): Promise<WorkflowRun | nul
 
 // ── JsonlRunStore ────────────────────────────────────────────
 
-/** Node fs 错误 code 判定（ENOENT = 路径不存在，并发删除场景）。 */
-function isEnoentError(err: unknown): boolean {
-  return (
-    typeof err === "object" &&
-    err !== null &&
-    "code" in err &&
-    (err as { code: unknown }).code === "ENOENT"
-  );
-}
-
 const logger = getLogger("subagents");
 
 /**

package/src/session-lifecycle.ts

@@ -300,7 +300,6 @@ function appendSubagentIdentityEntry(pi: ExtensionAPI): void {
         process.env.PI_SUBAGENT_FORK_DEPTH !== undefined
           ? Number(process.env.PI_SUBAGENT_FORK_DEPTH)
           : undefined,
-      chatMode: process.env.PI_SUBAGENT_CHAT_MODE === "true",
       // [review round2] worktree 隔离标志（session-runner 注入）：跨重启重建路径据此
       // 拒绝续聊（handle 不可序列化，reattach 不可行）。
       worktree: process.env.PI_SUBAGENT_WORKTREE === "true",
@@ -373,7 +372,8 @@ export function bindLedgerHostAndRecover(pi: ExtensionAPI, ctx: ExtensionContext
 /**
  * 随迁块 4 的进程级维护三连（各 try-catch「失败记日志不阻断」，设计 §3.4）：
  * 过期 session 文件清理 / ADR-035 manifest tmp 恢复 / ADR-035 worktree reaper 扫描。
- * 另含 [E1] sync 批崩溃恢复（per-session 域，不用 oncePerProcess——见块内注释）。
+ * （[modeless 波5] 原 [E1] sync 批崩溃恢复接线已摘除——collectMode 记录态消亡后
+ * core 侧 recoverSyncCollectBatch 已是 accepted-no-op，调用点随之退役。）
  */
 async function runProcessLevelMaintenance(
   agentDir: string,
@@ -381,22 +381,6 @@ async function runProcessLevelMaintenance(
   service: SubagentService,
   deps: SessionLifecycleDeps,
 ): Promise<void> {
-  // [E1] sync 批崩溃恢复（subagent-sync-collect 设计 §3.1.5 E1，U5 接线）：扫描主
-  // session 末条 entry 重建批缓冲；全员终态未投递 → notifyBatch 补发 + 统一补
-  // batchFinalized 标记（账本同 hash 幂等拒绝也算已投递）。须晚于 initSession（孤儿
-  // 终态恢复先行收敛 running 成员，createOrReuseServices 内部同步完成）与 ledger
-  // bind（补发走 notifyBatch 写账链，见上方 bindNotifyLedgerHost）。per-session 域
-  // （主 session 文件），不用 oncePerProcess；best-effort 不阻断 session_start。
-  // 时序约束在调用点 setupSessionLifecycle 已满足（bindLedgerHostAndRecover 与
-  // createOrReuseServices 均先于本 helper 调用）。
-  try {
-    service.recoverSyncCollectBatch();
-  } catch (err) {
-    logger.warn("[subagents] sync collect batch recovery failed", {
-      reason: toErrorMessage(err),
-    });
-  }
-
   try {
     // 递归扫描 <agentDir>/subagents + unlink 超 TTL 跨 session 文件属进程级维护
     // ——oncePerProcess 守卫防双跑（u-audit-fix）。