feat(collaboration): let existing agent conversations delegate governed work

[huangruiteng]

Problem and result

An existing shell-capable Agent could not use the bound delegation service without adding MCP tools or launching another lead. loopx delegation list/start/read/wait/resume now exposes that same service to the current conversation. Start returns while the detached worker continues; reconnect and retry retain the original operation, and accepted results still require current canonical completion and artifact verification.

Advances the attached-caller part of roadmap R2/R3 (#4574), reusing #4688 and the clarified contracts in #4702. This does not close the continuous-team milestone.

Implementation

• Thin CLI adapter; shared bounded wait and lazy MCP import. No new scheduler, Agent factory, grant store or acceptance authority.
• managed-research-team prepare creates a disposable fixture and bindings without starting a replacement lead. Clarify the required upstream-adoption hash in member tasks/tools after live rejection exposed the missing guidance.
• Update usage and bilingual RFC checkpoints. Classify the new command on command-specific help and repair the existing missing goal-acceptance classification.
• Existing CLI/MCP execution owners are reused. Frontend/Lark settings and session-mode behavior are unchanged; no new configuration surface is introduced.

Validation

• 38 focused tests across CLI processes, shared delegation/MCP and managed research: real File/SQLite authority, client disconnect/retry, single execution, current acceptance, tampered artifacts, invalid inputs, optional MCP absence and prepare-only behavior.
• Real existing-conversation CLI rehearsal: 2 DSH + 2 Ark workers, including cloud-to-local nested delegation; both cloud members adopted exact upstream hashes. The lead consumed four artifacts and independently completed the report: 5 canonical tasks done, Goal active.
• Synthetic conclusions: normalized FCF 40 → 25, revision −15; periods incomparable, one current-period source family, corrected repost stale.
• Preserve failed attempts: stale cloud environment rejected before model work; first model round omitted adoption hashes and failed. Clarifying the output contract produced a successful fresh rehearsal without changing validators. Owned cloud resources were cleaned and absence checked.
• Final premerge: 18 selected checks passed, plus CLI help/docs smokes, focused Ruff, configured mypy and public-boundary scan. The first premerge was invalidated by an edit during execution and rerun on frozen files. Latest-main rebase added only unrelated presentation files; candidate contents were unchanged.

Dynamic provisioning, unattended wakeup, persistent worker-session policy, and complete inbox/queue/steer remain separate RFC work. Local setup, private data, credentials and raw experiment records are excluded. Runtime changes remain for maintainer merge.

[huangruiteng]

Approval conclusion (author-owned PR; GitHub blocks formal self-approval)

Exact head: bbeab4cfd88e4cc3c6fe814a461c080f907a46a9；base: 3406be3d1724c081e8c004d351d78696a852578b。

没有发现阻断本次交付切片的问题。结论为 APPROVE；此评审不授予 runtime 自合并权限。

动机

按 #4574 的 R2/R3、已合并 #4688 的委派服务和 #4702 的边界判断，这是一项有实际调用者的增量：已有 shell 能力的 Agent 可以继续当前对话，调用受治理的委派服务。旧路径要求 MCP 接入，示例 run 还会启动另一个 lead；新路径让原对话完成派发、结果读回及综合。

本 PR 完成的是这个调用入口，不是整个 G1 持续团队目标。动态建队、原会话无人值守唤醒、完整 inbox/queue/steer 和持久云端 worker 会话仍有明确后续 owner。

改动思路

CLI → 现有 Delegations → TS 执行授权/状态转换 → 原 Turn → 当前 pinned 验收 → canonical Todo 完成 → 产物复核。没有第二套调度器、任务账本或验收规则；模型和工具返回“完成”不能跳过这条链。

比较了不做改动、文档指导调用内部 Python、另建调度层三个方案。71 行 CLI 复用现有 owner，补上稳定入口，维护成本合理。前后版本搜索也确认 configure_delegations、Turn 与验收 owner 均被复用。相关重构限于共用 wait、延迟加载可选 MCP，以及 prepare/run 共用初始化。

具体改动

14 个文件，+358/-17。生产变化包括 CLI 注册/dispatch、命令专属帮助分类及共享 wait；127 行新增测试覆盖进程边界。示例增加 prepare-only 和上游哈希字段说明；使用指南、示例 README、session/roadmap 中英文 checkpoint 同步。顺带补齐既有 goal-acceptance 的帮助分类遗漏，没有改变该命令行为。

关键代码讲解

• handle_delegation：检查动作、显式 execute、稳定操作 ID 和 brief 大小，随后调用共享服务。ok 表示调用结果，status 和当前验收产物才表示任务结果。
• Delegations.wait：CLI/MCP 共用有界观察；不启动、不恢复任务，最后再读一次当前状态。
• prepare_execution：复用隔离 fixture 和 operator binding 初始化，返回入口信息，不另起 lead Turn。原 run 复用它。
• write_output：工具说明和 scenario.task 明确要求 adopted_dependencies 的完整哈希。底层验证器保持不变；这是实测拒收促成的契约说明修复。

对主干的风险

最强反例是“调用成功、文件存在，但实际重复执行或没有采用上游证据”。真实 CLI 进程测试覆盖断开重连、相同操作不重复调用宿主、无 execute 不启动、错误身份读回和验收后篡改产物；现有 File/SQLite 测试覆盖接收方未采纳时不能完成。

真实模型第一轮确实遗漏了依赖哈希，另有旧云环境拒绝、超时或非 JSON 结果。这些失败没有被覆盖为成功。明确输出字段后重新建立隔离团队，两个云端成员采用准确上游哈希；其中云端分析员自主派发本地核验员。最终 2 DSH + 2 Ark、四份依赖和总报告共 5 个 canonical Todo 完成，Goal active。云端实验资源已清理并读回不存在。合成结论为 FCF 40→25、差值 −15、期间不可比、一个来源家族。

语义与 CI 对齐

复用现有请求/执行词汇，没有新增状态权威。CLI 是可信本地入口，不是远端身份认证或 Agent 创建协议。未提供 execution config 的 MCP 仍只有原来的五个工具；注册、命令可见、prepare/list/read/wait 都不激活模型执行。领域提示只在显式合成示例中加载，不进入通用控制面。Frontend/Lark 配置和会话模式没有变化。

执行了完全相同的 baseline/head MCP 与 File/SQLite 测试，各 15 passed。另有 6 项新 CLI 测试；本次完整聚焦组合为 38 passed。实际进程对照确认旧版缺少命令、缺 MCP 时共享 worker 无法 import，而新入口与延迟加载生效。最终 premerge 18 项通过；CLI 帮助、docs smoke、Ruff、配置范围内 mypy、隐私扫描通过。首次 premerge 因执行中编辑导致范围漂移而无效，冻结文件后重跑通过。

按当前 Goal 解析出的 wait_for_ci=false，没有查询或等待远端 CI；本地必需检查没有遗留失败或跳过。精确范围 change-quality receipt 已核验有效。

我的整体评价

这是完整、可回滚的原会话入口增量：已有执行与验收能力现在能直接被原协调 Agent 复用，真实混合/嵌套链路也闭合了产物依赖。它不会宣称固定绑定等于动态团队、一次协作等于持续多周期运行。最重要的剩余验证是原请求唤醒、动态工作派生及真实项目连续修订周期，按既有 R2/R3 owner 继续推进。

English verdict: APPROVE - HEAD bbeab4c. Thin CLI reuses governed delegation without replacing the lead conversation. Paired baseline/head compatibility, focused process tests, 18 premerge checks, and a real synthetic nested DSH/Ark team passed. Dynamic provisioning and unattended wakeup remain outside scope; maintainer merge required.