Duplicate settle notice for background subagent: closing message delivered twice ("finished" then "reported" again)

[omnisurge]

Observed on @deepseek-ai/dsh@0.1.0-rc.6.

What happened

1. Parent starts a background subagent (subagent tool, run_in_background: true).
2. While the child is still working, parent sends a follow-up via send_message (message parks, becomes the child's next turn).
3. The child's second turn completes and the agent settles.

Expected: exactly one settle notice containing the final assistant message.

Actual: the identical closing message was delivered to the parent twice — first as the finished ... closing message: notice, then again minutes later as a Background subagent <id> reported: notice with the same content.

Impact

• Token waste: the parent's context ingests the same long final message twice.
• Context noise / double-processing risk: the parent can mistake the duplicate for new work (we caught it only by content recognition).
• In long orchestration sessions (workflows, multi-agent), duplicated settle notices scale the waste linearly.

Guess at the seam

Looks like two delivery paths both fire for the same settled turn — the "finished" notice on settle and a queued "reported" delivery that wasn't deduplicated after a mid-flight send_message turn. Happy to provide session logs if useful.

[DRAG0NM]

已定位根因：continuable 后台子代理在 turn 结束时若调用了注入的 report 工具（系统提示明确鼓励这么做），父会话会收到两条内容相同的投递：settle notice（"finished … Its closing message: X"）与 report relay（"Background subagent <id> reporte… 完整源码定位与修复方案见 #2192 。

[omnisurge]

Confirmed — the root cause matches our repro exactly: our parent session did call send_message mid-flight (the child then ran a second turn), and the duplicate arrived as "minutes later" because it queued behind the parent's remaining turn. The steer (next-step) vs followup (next-turn) asymmetry in #2192 explains both the duplication and the order inversion we saw.

Fix proposal A (mark relayed output on the Activation, omit closing output in notifySettlement when already relayed) sounds right to us. Thanks for the thorough source-level analysis!

[DRAG0NM]

Thanks for confirming against your original repro — the mid-flight send_message + queued-parent-turn timeline fits the steer/followup asymmetry exactly. Glad proposal A (mark relayed output, omit duplicate closing output) matches your expectation; we'll keep the report updated if a patch lands.

[TTTPOB]

以下GPT生成并且人类大概审阅

补充一下当前新版 DSH 的情况。

现在原来的单向 report 工具已经被双向的 send_message 机制取代了，官方 release note 也明确写了：

Let parent Agents and continuable child Agents exchange follow-up messages through send_message, replacing the one-way report tool.

但原先这个重复交付问题在新的实现里似乎仍然以另一种形式存在。

当前 continuable subagent 启动时，会被额外注入类似下面的提示：

Your parent agent id is "...".
Before you finish, send your result to that agent with
send_message({ agent_id: "...", message: "<self-contained result>" }).
...

与此同时，subagent settle 时，runtime 本身又会自动把 child 的 final assistant message 作为 settlement notice 交给 parent。

因此如果 child 按这条提示在结束前调用一次 send_message，然后再正常输出 final message，就会出现语义上的重复交付：

send_message(final result)
        ↓
parent 收到一次

child final / settle
        ↓
settlement notice 再自动携带 final result
        ↓
parent 又收到一次

除了 token/context 浪费之外，send_message 和 settle 都可能触发 parent 的后续处理，因此实际使用时还会导致 parent 被重复唤醒，甚至把第二次收到的内容误判成新的 subagent 更新。

我也尝试过在 AGENTS.md 中加入相反规则，例如要求：

Do not send the final result with send_message immediately before settling.
The final assistant message will be delivered automatically by settlement.
Use send_message only for interim findings.

但实际测试效果不好。

subagent 确实同时看到了两段互相冲突的提示：一段来自 AGENTS.md，另一段是创建 continuable subagent 时直接追加到当前 task 后面的 withContinuableReturnGuidance()。

模型自己也明确表示两条指令存在冲突，并倾向于遵守后面这个更加 task-local、措辞更加直接的硬编码提示。也就是说，单纯通过 AGENTS.md 覆盖并不能可靠解决问题，反而会给模型制造不必要的 instruction conflict。

目前我使用的临时 workaround 是直接用 pnpm patch 修改 @deepseek-ai/dsh-subagent 中这句 guidance。

大致方法：

pnpm patch @deepseek-ai/dsh-subagent@<实际安装版本> \
  --edit-dir ./dsh-subagent-patch

然后找到编译后的文件中对应字符串：

rg -n 'Before you finish|send your result' ./dsh-subagent-patch

通常修改发布包里的 lib/index.js，把语义改成类似：

Your final assistant message is automatically delivered to the parent when
you finish. Do not send the same final result with send_message immediately
before finishing. Use send_message for interim findings when they may change
what the parent should do next.

最后：

pnpm patch-commit ./dsh-subagent-patch

这样 pnpm 会生成 patch，并在以后安装该版本依赖时自动应用。

现在的 send_message + settlement 语义，更清晰的指令是：

send_message = 中途进展 / 需要及时通知 parent 的信息
final assistant message = 最终结果，由 settlement 自动交付

而不是要求 child 在结束前用 send_message 再主动发送一次最终结果。

[TTTPOB]

https://gist.github.com/TTTPOB/2a56776d9b2200018693bad33dcdce19