Agent disposal does not propagate to continuable subagents — two independent lifecycle managers lack a dispose handoff contract, leaving orphaned children running indefinitely （Agent dispose 不传播到可续传子 agent——两个独立生命周期管理器缺少 dispose 交接契约，孤儿子 agent 持续运行）

[yamingmou]

Summary

Disposing an agent does not dispose its continuable subagents — the agent/disposed event listener in SubagentContinuationManager only cleans up tracking references (closingScopes), never calls this.dispose(activation) on the disposed agent's children. Orphaned subagents continue running, consuming API quota, producing PARENT_UNAVAILABLE errors, and blocking the continuation manager's drain barrier.

Suggested fix: Extend the agent/disposed listener to call this.dispose(activation) for the disposed agent's children — reuses existing machinery, no new code paths.

Trigger Condition

1. Agent A creates a continuable subagent B via ctx.subagents.startContinuable()
2. Agent A is disposed (session close / process shutdown / manual dispose)
3. Subagent B is not disposed — becomes an orphan

Minimal Repro

// 1. Agent A starts a continuable subagent B
const sub = await ctx.subagents.startContinuable({ /* ... */ });
// 2. Agent A is disposed (session close or manual)
await agentA.dispose();
// 3. Observe: subagent B is still alive — activations Map retains its entry,
//    B continues running, produces PARENT_UNAVAILABLE when trying to report,
//    and blocks the drain barrier if B has its own children.

Related

• [#3984]([Bug] mcp-client: activeServerNames 生命周期泄漏导致无法创建新会话 #3984) — serverName lifecycle leak (WeakMap key scope mismatch). Single-module registry scope issue; this report is a cross-module dispose handoff gap.
• [#3994]([Bug] Terminal background-job records are never reclaimed in a live session: unbounded store growth, O(n) list/start scans, and a silent memory footprint that grows with session age #3994) — terminal background-job records never reclaimed. Single-module job table leak; this report is a cross-module resource leak.
• [#4017](工具调用失败留下悬挂 tool_calls,会话被 400 INVALID_REQUEST 永久卡死 #4017) — dangling tool/call → model API 400. Async cleanup gap; this report is a lifecycle handoff gap.
• [#4555]([Bug] Subagents always fail with 400 "Reasoning is mandatory" when parent uses a thinking-required model (Ox Alpha via OpenRouter) #4555) — reasoningEffort not propagated to child agents. In our audit baseline (rc.2) AgentOptions has no such field, so the child never receives it; the reporter (rc.6) describes the same gap as resolveChildAgentOptions only forwarding provider/model/maxTokens ("silently dropped"). Same architectural pattern either way: incomplete parent→child option propagation in dsh-subagent. (Independently reproduced for spawn-backed subagents in [bug] Spawn-backed subagents drop reasoningEffort; endpoints that require thinking reject every delegation #4666.)
• [#1105]([Bug Report] subagent 显式指定 `model` 被选择层覆盖：workflow `agent(prompt, {model})` 的模型不生效 #1105) — model overwritten by selectionFor. Same pattern in the model selection layer: explicit AgentOptions ignored by a downstream layer.

---

Environment

• @deepseek-ai/dsh-agent-loop@0.1.1-rc.2
• @deepseek-ai/dsh-subagent@0.1.1-rc.2
• @deepseek-ai/dsh-agent@0.1.1-rc.2
• Also verified on audit profile audit20260822
• Date: 2026-08-27

---

Detailed evidence chain and fix analysis follow in the comments below.

---

About this report

EN: Reported by the OfferKuai (Offer快) Team — an AI startup building full-lifecycle job-application services, guided by the belief that "users need results, not repeated conversations." Founder: Zhaofeng (Yaming). We use DeepSeek Harness as part of our daily development workflow; this report is our way of contributing back to the ecosystem. Website: https://www.offerkuai.com/ | Contact: <contact@offerkuai.com>

---

概述

Dispose 一个 agent 时，其可续传子 agent（continuable subagent）不会被自动清理——SubagentContinuationManager 中的 agent/disposed 事件监听器只清理跟踪引用（closingScopes），从不对被 dispose agent 的子代调用 this.dispose(activation)。孤儿子 agent 持续运行、消耗 API 额度、产生 PARENT_UNAVAILABLE 错误，并阻塞 continuation manager 的 drain 屏障。

建议修复：扩展 agent/disposed 监听器，对被 dispose agent 的子代调用 this.dispose(activation)——复用现有机制，无需新增代码路径。

触发条件

1. Agent A 通过 ctx.subagents.startContinuable() 创建可续传子 agent B
2. Agent A 被 dispose（会话关闭/进程关闭/手动）
3. 子 agent B 未被 dispose → 变成孤儿

最小复现

// 1. Agent A 启动可续传子 agent B
const sub = await ctx.subagents.startContinuable({ /* ... */ });
// 2. Agent A 被 dispose（会话关闭或手动）
await agentA.dispose();
// 3. 观察：子 agent B 仍然存活——activations Map 保留其条目，
//    B 继续运行，汇报时产生 PARENT_UNAVAILABLE，如果 B 有子代还会阻塞 drain 屏障。

相关 Issue

• [#3984]([Bug] mcp-client: activeServerNames 生命周期泄漏导致无法创建新会话 #3984) — serverName 生命周期泄漏（WeakMap key scope 不匹配）。单模块注册表 scope 问题；本报告是跨模块 dispose 交接缺口。
• [#3994]([Bug] Terminal background-job records are never reclaimed in a live session: unbounded store growth, O(n) list/start scans, and a silent memory footprint that grows with session age #3994) — 任务终态不回收。单模块任务表泄漏；本报告是跨模块资源泄漏。
• [#4017](工具调用失败留下悬挂 tool_calls,会话被 400 INVALID_REQUEST 永久卡死 #4017) — 悬挂 tool/call → 模型 API 400。异步收尾缺口；本报告是生命周期交接缺口。
• [#4555]([Bug] Subagents always fail with 400 "Reasoning is mandatory" when parent uses a thinking-required model (Ox Alpha via OpenRouter) #4555) — reasoningEffort 未继承到子代理。在我们的审计基线（rc.2）中 AgentOptions 无该字段，子代理自始收不到；原帖（rc.6）将同一缺口描述为 resolveChildAgentOptions 仅转发 provider/model/maxTokens（"被静默丢弃"）。无论哪个版本，同一架构模式：dsh-subagent 中父→子选项传播不完整（另见 [bug] Spawn-backed subagents drop reasoningEffort; endpoints that require thinking reject every delegation #4666 独立复现 spawn-backed 子代理丢 effort）。
• [#1105]([Bug Report] subagent 显式指定 `model` 被选择层覆盖：workflow `agent(prompt, {model})` 的模型不生效 #1105) — model 被 selectionFor 覆盖。同一模式在模型选择层：显式 AgentOptions 被下游层忽略。

（详细证据链与修复分析见下方跟帖。）

环境

• @deepseek-ai/dsh-agent-loop@0.1.1-rc.2
• @deepseek-ai/dsh-subagent@0.1.1-rc.2
• @deepseek-ai/dsh-agent@0.1.1-rc.2
• 审计 profile audit20260822 同样验证
• 日期：2026-08-27

署名

ZH: 本报告由 OfferKuai（Offer快）团队提交 —— 一家专注 AI 全流程求职托管的创业团队，核心理念是「用户要的是结果，不是重复对话」。创始人：Zhaofeng（Yaming）。我们将 DeepSeek Harness 用于日常开发工作流，这份报告是我们对社区的回馈。官网：https://www.offerkuai.com/ | 联系：<contact@offerkuai.com>

[Jstn-1g]

Thanks for independently surfacing this. I reproduced the orphaned-child lifecycle failure against the current official v0.1.2-alpha.1 head (cd5ef8148158c3a752a658978873241fdf8e2bbc) and ported the earlier #4239 reference implementation to that exact baseline.

One lifecycle correction to the suggested hook: agent/disposed is emitted only after the exact Agent has left the registry, and returned listener promises are observed but not awaited. A drain started there cannot make parent disposal wait for descendants, cannot reliably traverse a disposed intermediate parent, and does not fence an admitted materialization already crossing Agent publication. The ownership boundary needs to be the parent's own Agent scope, before registry removal.

Current-alpha reference:

• exact commit: Jstn-1g@b2b75f7
• branch: https://github.com/Jstn-1g/deepseek-harness/tree/reference/discussion-4793-parent-scope-drain-alpha1
• prior design/reproduction thread: subagent: a continuable child of a non-continuation parent is owned by nothing, and outlives the parent unheard #4239 (comment)

The patch installs one exact-Agent, WeakSet-guarded cleanup in each non-continuation parent's own scope. That cleanup publishes the descendant admission cutoff before its first await, waits admitted materializations through publication or rollback, cancels the selected forest top-down, releases handles child-first, and finally re-clears any settlement notice that crossed after AgentLoop's initial disposal clear. Explicit host drains, ACP ownership, and continuation-managed parent recursion remain separate and unchanged.

The final re-clear is causal, not defensive decoration: removing it makes a dispose -> persist -> same-session resume regression fail with exactly one stale subagent-settled notice. Restored, persistence records the teardown notice insert followed by its removal, and resume sees none. The failure path and simultaneous drain/re-clear failure are also covered.

Verification on the exact public commit:

• 146/146 focused tests pass (112 continuation, 27 report, 7 ACP disposal);
• affected continuation.ts: 100% statements, branches, functions, and lines;
• 789 deterministic relevant repository tests pass, 2 skipped;
• full typecheck, lint, 32/32 doc-sync gates, translation pairing, and duplication checks pass;
• independent exact-commit review found no remaining blocker.

This is a tested reference implementation, not an upstream merge claim.

中文摘要：已在当前官方 v0.1.2-alpha.1 精确基线复现。agent/disposed 在 Agent 离开注册表后才触发，异步监听器也不会被生命周期等待，因此不能作为可靠的父子销毁交接点。参考修复把一次性 cleanup 安装到非 continuation 父 Agent 自身作用域：先关闭后代准入，等待已进入的物化事务，再自顶向下取消、child-first 释放，最后清除父销毁期间晚到的 settlement notice，避免它在同一 Session 恢复后重新出现。146 项聚焦测试和受影响源码 100% 覆盖通过；这是参考实现，尚未声称已合入上游。

[argszero]

Independent source verification of both the report and the reference fix, against the same alpha.1 commit (cd5ef81481) in the local tree:

Reported bug — confirmed. SubagentContinuationManager's agent/disposed listener does exactly one thing: this.closingScopes.delete(agent) (packages/subagent/subagent/src/continuation.ts:386-387). It never touches activations children, so a continuable child survives its parent's disposal — the orphan is real.

Lifecycle correction — confirmed at source, and it's the right call. Two facts back Jstn-1g's move off the agent/disposed hook:

1. dispose() in the agent registry does this.store.delete(entry.id) before emitDisposed(entry) (packages/core/agent/src/index.ts), so the event fires only after the agent has already left the registry — a handler there can no longer reach the parent's scope.
2. emitDisposed wraps each callback as void Promise.resolve(returned).catch(...) — listener promises are observed, never awaited, so a drain started there cannot gate parent disposal on descendants.

Both facts independently confirm that the ownership boundary has to be the parent's own scope before registry removal, which is exactly where the reference patch installs its cleanup.

Also worth noting for anyone reading this thread in the future: this is the dispose half of the same "weak cross-module lifecycle contract" pattern the maintainers are consolidating — the sibling symptom (#4666, option propagation on spawn) landed natively in alpha.1, and the browser-UUID sibling (#4756) landed via packages/util/crypto in the same release. The child-first release + post-dispose settlement re-clear detail here (stale subagent-settled notice on dispose → persist → resume) matches the SubagentResult.diagnostic observability gap I raised in #4666 — good to see the family converging on tested reference implementations.

[yamingmou]

Your independent check against the same alpha.1 commit lines up with what we saw when we were poking at why the agent/disposed hook can't act as the parent→child handoff point — specifically your two points: registry delete before emitDisposed, and listener promises observed-not-awaited. Same conclusion on where the ownership boundary should sit: the parent's own scope, before registry removal.

On the pattern: you're right that this is the dispose half of the weak cross-module lifecycle contract. The spawn/option half (#4666) did land in alpha.1, but the dispose/orphan half still isn't merged upstream — so far only Jstn-1g's tested reference impl exists (146/146 focused tests, no upstream merge yet). That gap is what we'd been digging into and left some notes on at #4793.

The part where you tie the post-dispose settlement re-clear to the SubagentResult.diagnostic gap from #4666 is what made our ears perk up — orphan + diagnostic-swallow looking like two faces of one root cause is exactly the thread we'd been pulling on. The lifecycle-handoff angle is the next one we're taking on our side: same two faces, plus the broader question of where cross-module ownership boundaries should actually sit. Still rough notes, but that's the direction. Same family from a different angle — good to have it confirmed independently.

---

你用同一个 alpha.1 提交做的独立核验，和我们当时扒「为什么 agent/disposed 钩子扛不起父子销毁交接」时看到的完全对得上——尤其你那两点：注册表先 delete 再 emitDisposed；监听器 promise 只被观察、不被 await。所有权边界该落在哪，结论一致：父 Agent 自身作用域、且在离开注册表之前。

关于模式定性：你说得对，这是「弱跨模块生命周期契约」的 dispose 这一半。spawn/选项传播那一半（#4666）确实进了 alpha.1，但 dispose/孤儿这一半至今没合入上游——目前只有 Jstn-1g 那个经过测试的参考实现（146/146 聚焦测试，尚无上游合入）。这个缺口正是我们之前一直在抠、并在 #4793 留了些笔记的地方。

你把「dispose 后 settlement 二次清理」和 #4666 里那个 SubagentResult.diagnostic 缺口连起来，这一点最让我们眼前一亮——孤儿与 diagnostic 被吞像是同一根因的两面，也正是我们当时在拽的那根线。生命周期交接这个角度，正是我们这边接下来要推进的：同一根因的两面，外加「跨模块所有权边界究竟该落在哪」这个更大的问题。还是粗糙笔记，但这是我们打算推进的方向。同一族从另一角度被点出来，有独立佐证挺好。