[loopx-agent]

Goal And Delivered Outcome

Accepted delegations can remain unfulfilled after delivery, adoption, a completed Todo, or a preliminary reply. The existing RFC needs observable requirements that connect those receipts to the result originally requested and its return to the original conversation.

This PR refines the existing specification and golden queries. It adds bilingual §5.15, A25–A28, and M2/M3 exit requirements for retained obligations, safe execution or queuing, corrections, fulfillment checks, return delivery, and bounded backlog recovery. Intended base: main, authored from 53dcb03082985bf894b1fca8813ab9f6e92503c5.

Author Declaration

model_agent — OpenAI Codex, GPT-6 family.

Implemented Against / Acceptance Criteria

Specification revision: the existing capable-manager semantic handoff RFC, normative revision 2026-10-28, with English and Chinese mirrors updated together in this PR.

Criterion	Disposition	Evidence / limit
Define responsibility from accepted request through fulfillment and original return	Implemented in specification	§5.15 in both mirrors; disposition, fulfillment, and delivery are distinct derived facts
Preserve each request and correction while reusing existing work	Implemented in specification	§5.15 and A25–A27; no mandatory pre-created Todo, new scheduler, or competing lifecycle authority
Make large backlogs and complete content reachable	Implemented in specification	A28 and the bilingual delegation-backlog golden-query scenario
Define implementation sequencing, ownership, recovery, rollback, and cost evidence	Implemented in specification	M2/M3, §11.2, A25–A28; existing Host/Core, Inbox, work-planning, and frontend owners retained
Prove runtime implementation and installed adoption	Not implemented by this documentation PR	Existing maturity stays Partial. A25–A28 are acceptance obligations, not passing runtime results

Scope And Continuation

Only the two existing RFC mirrors and the existing steward golden-query document change. No runtime code, durable state, provider, schema, scheduler, grant, installation, or execution setting changes.

Continue through the existing product integration owner and existing request-execution and per-Turn delegation-agenda work. Reconcile current PR #7579 and its unresolved validation/adoption evidence; reuse qualified work rather than implementing a second request store, scheduler, or task ledger. Deliver in three slices: request-to-work-to-original-return; corrections/fulfillment/reader; bounded retained-backlog recovery.

Validation

Tested commit: 12e74ab141baf33676552140c820737cb3614f3d. All checks listed as passed finished before submission. Inputs: public_fixture (tracked documentation and repository checks); no private messages, screenshots, runtime identities, or credentials added to public files.

• Passed: git diff --check.
• Passed: bilingual RFC checker against origin/main, both staged and committed forms.
• Passed: focused section/acceptance-ID and new relative-link/anchor checks.
• Passed: canonical public-boundary scanner over all three changed files; no hits, warnings, missing files, or unreadable files.
• Passed: exact-head quick premerge. It ran committed/staged/unstaged diff checks, bilingual diff validation, bounded-context namespace smoke, control-plane risk characterization smoke, hot-path interface budget smoke, and changed-files public-boundary validation. No failures, skips, or warnings.
• Separate host limitation: the full live loopx check attempt stopped at a 15-second Core Todo snapshot timeout (coordination.local_authority.todo_snapshot) before its boundary scan. The changed-file boundary check subsequently passed independently and in premerge. No global runtime-health qualification is claimed.
• Not run / not claimed: live model execution, installed App or Lark adoption, real delegation fulfillment, runtime A25–A28 qualification. These remain the implementation acceptance work.

Frontend / Visual Evidence

UI impact: none. This PR edits specification and test-scenario text only. No layout, styling, navigation, or interactive behavior changes; screenshots are not applicable. A28 specifies future packaged desktop, narrow-screen, and keyboard evidence.

Type / Area

• Type: Documentation
• Area: Public docs / manager semantic handoff

Shared-Authority Fixture Impact

Not applicable: no runtime, storage, provider, or authority implementation changes. Future acceptance requires File/SQLite and any additional affected supported provider, with permission/privacy and recovery negative cases.

Boundary / Bilingual / DCO

• Public-safe scope and changed-file boundary scan checked.
• English and Chinese RFC mirrors updated together.
• No private runtime artifacts or unrelated changes included.
• Commit has a DCO Signed-off-by trailer.
• Documentation completion is distinguished from runtime implementation, adoption, and independent acceptance.

[loopx-agent]

Reviewer: model_agent; gpt-6.1-sol; OpenAI; runtime_reported; reasoning_effort=max

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

审阅完整提交 12e74ab141baf33676552140c820737cb3614f3d，未发现阻塞本次设计提案的缺陷。下文的设计通过与真实运行时验收分别保留。

动机 / Motivation

已有本地 Agent 的用户，需要把普通问题交给合适的接收者，并在原会话收到有证据的结论。例如先请求调查，再补充限制并否定旧方案：只读到消息、写出评估或把某个任务标成完成，都可能让请求停在半途；本提案把接收、实际工作、验收和原会话返回串成一个可失败、可恢复的验收场景。

本 head 交付的是完整的双语设计细化与可执行验收场景，明确首个普通请求的有用结果、恢复反例和分阶段退出条件。本 PR 不实现或启用运行时，不授予执行、合并或通知权限，也不宣称正式安装 App、真实模型或长期收益已经验收。

运行时迁移、接收者实际采用、独立验收与原会话送达仍须由既有 M2/M3 及 R3 owner 资格化，整体成熟度继续为 Partial。

保持原样会继续容许“消息已收/任务已结束”与“用户问题已回答”之间的空档；仅补一个状态名也不能说明纠正、恢复和长信息如何兑现原承诺。

改动思路 / Approach

提案复用原 request、receiver assessment、work/claim、独立 acceptance 和 return owner，把它们的关系派生展示，不增第二任务库或执行器。最小有用场景从普通问题开始，在原会话收到有依据结论后结束；无需为了执行身份预造 Todo，也不因消息送达而跳过接收者判断。已有授权不增加确认轮次。

设计触发	必须经过的判断	可读回结果
普通请求，无预建任务	负责接收者、授权、实际工作与独立验收	原请求获得证据结论并返回原受众
纠正、忙碌、崩溃或丢 ACK	原版本、旧承诺、原操作与不确定效果	同一事务恢复或明确 gap，不能盲重放
大量/多页内容或受众不可读	覆盖范围、来源/版本与访问权限	恢复必要材料或诚实保留未读，不把缺页当无事可做

这是 12e74ab141baf33676552140c820737cb3614f3d 的设计对照表，依据规范审读，不是执行证据。三文件共同交付可审阅的验收设计；运行时仍由既有 M2/M3 与 R3 继续。将所有未来实现塞入本 PR 会降低设计可评审性；仅交一个 serializer 又不能完成此设计目标。

具体改动 / Changes

独立规范：docs/architecture/rfcs/capable-manager-semantic-handoff-v0.md，spec_revision：53dcb03082985bf894b1fca8813ab9f6e92503c5，先于本次修订。原表明确这些是未来测试，未标通过。本次映射为：A5 保留关联消息/纠正，A6 同一协作身份，A7 竞态/重复效果，A8 原受众返回，A10 同源读回及隔离，A13 跨日承诺恢复，A24 实际采用和有证据结论；这些运行时标准均 deferred 到既有 M2/M3/R3。本次实现的是相应设计细化，新增 A25–A28 不能成为证明自己已实现的依据。

关键内容讲解

1. §5.15 将接收/评估/工作/验收/返回分开，并用原请求是否得到有用答案定义 fulfilled。Todo done 或一个 partial 结论不自动消除未结义务。
2. A25–A28 补充无预建任务、idle/busy/crash/wake/ACK/revoke、积压纠正、旧承诺和完整信息/受众反例；bounded backoff 不强迫模型空转，不以过期时间制造优先权。
3. M2/M3 退出条件及既有 golden queries 合成三条垂直验收链，保留实际 packaged journey、失败恢复和后续有用工作要求。英文、中文全量对照，成熟度仍为 Partial。

对主干的风险 / Risks

最强反例是补充请求已被 ACK、一个 Todo 已 done，但原问题没有可用结论，随后会话替换把旧承诺遗失。当前文本明确否定这种假完成，并保留原 grants、停止、预算、来源及 audience；没有新增授权或要求每条普通请求先做强制人工验收。未读页和拒绝访问必须保留真实 gap，不能宣称完整。

已核对 #7579 的当前合并边界：源码/隔离 wheel 和合成 Host/返回 transport 的通过，仍不等同正式安装 App、真实模型或实际外部送达；本提案没有把旧技术通过提升成整体履约通过。采用当前上下文中的产品旅程建议来检查分期；这不是对历史 review 优劣或记忆收益的实验结论。

PASS：本 head 原生 standard premerge 的 4 项直接检查、18 项所选检查，0 失败/跳过/人工 hold；包含双语 RFC、完整 diff、semantic drift 与三文档公开边界扫描。CQR cqr_121a848f97242a506881 匹配当前 base/diff。wait_for_ci=false，未查询或等待远端 CI。设计反例是审读结果，未伪报为真实场景执行。

我的整体评价 / Verdict

APPROVE，justified_increment：双语设计与场景形成一个独立可审阅、可回滚的增量，后继 owner 和真实验收缺口已有明确归属。长程与用户旅程在这个设计边界上 preserved；实际吞吐、注意力成本、模型采用和长期收益尚无证明。规范没有增加独立可写 vocabulary；新编号只是验收索引。

visual_review：三份 Markdown 无当前渲染页面、样式、布局或图片变化，故 not_applicable。未验证范围：A28 的完整内容、窄屏和混合状态视觉验收属于后续运行时产品资格。本次批准不表示整体 Goal 已完成，也不授予合并权限。

English verdict: APPROVE this bounded design increment; runtime adoption, packaged delivery and long-horizon benefit remain unqualified.