Team messages use the target inbox
Team 消息使用目标 inbox
A second Team outbox duplicates Agent inbox ownership with separate identities, acknowledgements, persistence checkpoints, and retry ordering. Callers need an explicit admission result while the continuation owner already provides target locking, lifecycle authorization, and cold recovery.
English
Problem
A second Team outbox duplicates Agent inbox ownership with separate identities, acknowledgements, persistence checkpoints, and retry ordering. Callers need an explicit admission result while the continuation owner already provides target locking, lifecycle authorization, and cold recovery.
Decision
New Team sends return the accepted inbox MessageId or throw. They use live Lead Steer or the host-only subagent Steer adapter, retaining the actual sender through the existing agent-message source and the member name in content. Acceptance follows ordinary Agent persistence; it neither confirms model processing nor adds a synchronous flush guarantee. Team does not retain unaccepted intent, retry it after restart, or deduplicate explicit resends.
Historical mailbox records stay decodable and projected so existing Lead logs replay, but Team never delivers or acknowledges them. The package is experimental with no stability promise, and a second dispatch path with receipt inspection, per-target ordering, and its own disposal wait costs more than the undelivered old messages it would recover. This partially supersedes the mailbox decision in Agent Teams; roster and task decisions remain there. The package reference owns the resulting limitation.
Alternatives considered
Keep the Team outbox for all sends. This preserves a stronger retry promise but duplicates coordination and requires maintaining two acceptance states. Reintroduce it only for a concrete durable send-intent requirement that the Agent inbox cannot serve.
Expand public sibling messaging permissions. Team membership is separate from subagent parent authority. The host adapter authorizes the exact Lead without misattributing the sender or broadening public child controls.
Drain historical pending items after restart. This honors old accepted intent, but it keeps a second delivery owner beside the Agent inbox for data that only an experimental opt-in composition wrote. Removing the decoder as well is also rejected, because Lead logs containing mailbox records would then fail to replay.
Consequences
New attempts have one identity and one inbox owner. A failed attempt is visible to its caller and has no Team retry promise. Historical logs still cost decoder and projection code. A historical message that its target had not recorded before the upgrade is never delivered. Tests cover direct Lead/peer acceptance, cold recovery, sender attribution, failures, cancellation, disposal, and that historical queued mail stays readable and undelivered after restart.
中文
问题
第二套 Team outbox 通过独立身份、确认、持久化检查点和重试排序重复了 Agent inbox 的职责。调用方需要明确的准入结果，而 continuation owner 已提供目标锁、生命周期授权和冷恢复。
决策
新的 Team 发送返回获准 inbox 的 MessageId 或抛出错误。它们使用 live Lead Steer 或 host-only subagent Steer adapter，通过现有 agent-message source 保留真实发送者，并在正文中保留成员名称。接收遵循普通 Agent 的持久化策略，既不确认模型已处理，也不增加同步 flush 保证。Team 不保留尚未接收的发送意图，不在重启后重试，也不对显式重发去重。
历史 mailbox 记录继续可解码并参与投影，使现有 Lead 日志能够回放，但 Team 不再投递或确认它们。该包是实验性的，没有稳定性承诺；另设一条带回执检查、同目标排序和独立 dispose 等待的投递路径，其代价高于它能补回的未投递旧消息。这部分取代了 Agent Teams 的 mailbox 决策；roster 和任务决策仍由原记录负责。包参考定义由此产生的限制。
考虑过的替代方案
为所有发送保留 Team outbox。 这保留了更强的重试承诺，但重复协调职责，并要求维护两个接收状态。只有具体的持久发送意图需求无法由 Agent inbox 满足时，才重新引入它。
扩大公开的 sibling 消息权限。 Team 成员关系与 subagent parent 权限不同。host adapter 授权确切的 Lead，不会错误归属发送者，也不会扩大公开 child control 权限。
重启后补送历史 pending 项。 这能兑现已接收的旧意图，但会在 Agent inbox 之外保留第二个投递 owner，而这些数据只由实验性的可选组合写入。同时删除 decoder 的方案也被否决，因为包含 mailbox 记录的 Lead 日志将无法回放。
后果
新尝试只有一个身份和一个 inbox owner。失败尝试会向调用方报告，且没有 Team 重试承诺。历史日志仍需要 decoder 与投影代码。升级前目标尚未记录的历史消息永远不会被投递。测试覆盖直接 Lead／peer 接收、冷恢复、发送者归属、失败、取消、dispose（资源释放），以及重启后历史排队消息仍可读取且不被投递。