[huangruiteng]

Goal And Delivered Outcome

LayerFS can preserve complete, branchable engineering environments, but filesystem speed and deduplication do not establish better multi-agent engineering outcomes. This proposal defines a reproducible way to test that connection with LoopX.

• Before → after: no joint design or attribution plan → bilingual RFC with audited source boundaries, a concrete public historical repair demo, independent acceptance, realistic warm-worktree controls, and native/passive/governed comparisons.
• Outcome basis: maintainer request to investigate LayerFS × LoopX for complex multi-agent software-engineering exploration and propose a demo/evaluation RFC and issues in both projects.
• Tracking: LoopX #7406 · LayerFS #461.
• Base: main, starting at 15803cd7dcc2fc517049809bbe58998f04f5acac.
• RFC: English · 中文.

Author Declaration

• Written by: model_agent — OpenAI Codex.
• Specification: this maintainer request, existing Explore ownership, Stride §7.7 research-state branching, and the long-horizon benchmark measurement contract at the base above.

Criterion	Disposition	Path / evidence
Research actual upstream capabilities and limits	implemented	RFC §4 / Appendix A; LayerFS pinned to ad78914e36b258b360daf13aec164b5655814689
Concrete demo and independent acceptance	implemented	§5 / §9; historical #6869 candidate, base reproduction explicitly pending, existing #6875 not treated as oracle
Fair effectiveness/efficiency comparison	implemented	§6 / §9; warm dependencies, total-budget accounting, six policy/provider cells, causal-claim limits
Existing owners and product boundaries	implemented	§3 / §5 / §10; no new runtime schema, capability, scheduler or default
Bilingual discovery	implemented	RFC pair, index, generated status and paired roadmap entry
Live integration and measured benefit	out_of_scope	Proposed M0–M3 work, not represented as delivered

Self-check: inspected public SDK lifecycle/source, Store semantics, runtime/backend docs, existing integration and open upstream issues; reviewed both language versions and source references. A local Docker availability check does not qualify LayerFS.

Scope And Continuation

This is a completed design proposal, with implementation deliberately tracked as future work. It separates:

1. filesystem-only checkpoints from complete conversation/evidence/artifact/opportunity recovery;
2. substrate efficiency from task success and joint causal benefit;
3. a scripted conformance demo from actual model adoption.

The smallest implementation slice is real environment fork → probe → receiver adoption → clean exported patch verification. Existing LoopX and LayerFS owners remain authoritative. The proposal references upstream #411, #456 and #446 rather than creating duplicate GC/worker/roadmap work. LoopX #7406 tracks implementation/evaluation; LayerFS #461 requests upstream API/evaluation feedback. Both issues are published and cross-linked.

Bounded-refactoring review: unnecessary for this documentation-only slice; existing capture/Explore owners are reused in the proposed design, and no speculative provider code was added.

Validation

• Tested revision: 774dcb0e0171cc1dd993d2c2e16494c357a37d17 for final static checks. The 13 RFC-index tests passed on 00a972f34b300f44dbb6aaec24d72e6292ab7fb3; the only subsequent content change adds the two tracking links to both language headers.
• Run state: finished.
• Input classes: public_fixture (public source and documentation).

Check kind	Result	Evidence / limitation
static	passed	RFC status generation and scripts/generate_rfc_status_index.py --check
static	passed	scripts/check_rfc_bilingual.py --base origin/main --staged, then --head HEAD
static	passed	New RFCs' 12 local links each resolve; 17 external references match between languages; whitespace/public-boundary scan
unit	passed	uv run --extra test python -m pytest -q tests/architecture/test_rfc_status_index.py — 13 passed
manual	passed	Source/claim and bilingual semantic review; no promotion of upstream v1.0 measurements into v1.1 or task-level results
real_backend	not_run	No LayerFS build, live integration, model study or performance reproduction; these are the proposed next milestones

Coverage: static/document checks fit the changed Markdown and generated indexes. No product code, authority-store implementation, benchmark runner, scoring rule or UI chrome changes. Runtime suites are not a qualification claim here. The header's Accepted status becomes effective only on merge under the repository RFC lifecycle.

Frontend / Visual Evidence

UI impact: none. Markdown proposal and indexes only; screenshots are not applicable. The RFC defines the later App/CLI/advertised-Lark readiness, launch, error, result and disable journey without claiming it is shipped.

Type of Change

• Documentation update

LoopX Area

• Public docs or presentation surface (RFC content only)

Technical Direction

S4/S8/S11; existing S3/S7 authority and budget boundaries. No portfolio priority change.

Boundary Checklist

• Dedicated branch from verified target main; explicit seven-path staging.
• Public sources only; no credentials, local paths, private Goal state or raw benchmark logs.
• English/Chinese RFC and roadmap pairs updated together; generated status refreshed.
• DCO sign-off included.
• No new runtime vocabulary or shared-authority fixture change; proposed study labels are document-local.
• No merge, installation, paid experiment or production rollout performed.

[loopx-agent]

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

动机

未发现阻塞问题，结论为 APPROVE，对应 774dcb0e0171cc1dd993d2c2e16494c357a37d17 的完整七文件变更。

需要在复杂工程修复中比较多条方案、交接候选环境的 LoopX 与 LayerFS 维护者。
原先的文件系统演示不能回答多 Agent 是否修出更多正确补丁；本提案把真实环境分叉、失败路线、接手者后续工作和干净导出验收放入同一条实验路径。
本次可观察交付是双语设计、可反驳假设、独立验收行和双方已交叉链接的跟踪 issue。
本 PR 不实现 provider，不启动实验，不证明提速或成本收益，也不改变默认行为、权限和预算。
真实 CLI 基线复现、LayerFS 集成、配对实验、完整成本覆盖、模型实际采用和产品旅程仍待 M0–M3 逐项验证。

改动思路

保留预热 Git worktree 作为实用强对照，先检验环境分叉是否正确，再比较同一总预算的正确修复与完整成本。仅 G-W/G-F pilot 只能说明该治理策略下的 workspace 效果；native/passive/governed 六单元及预声明交互对比才支持联合归因。没有收益或出现 I/O 退化也是有效结论。

边界	本提案保留的 owner
问题、发现、预算、采用	现有 Explore / Todo / quota
会话、工具、取消、settled barrier	Host/runtime
文件树和提交谱系	待验证的可选 LayerFS adapter
独立验收与源码交付	现有 evaluator 和 Git/PR

两张图只关联身份，不把文件 Commit 当作研究成果。最小有用路径是 M1 的分叉、负结果、接手者新工作和干净导出补丁，不能停在安装或快照回执。当前边界是设计提案；A6 在产品推广前要求现有入口的发现、授权操作、失败恢复、结果与关闭，不新增一个尚无真实调用方的顶层页面。交付的历史评审经验在这里用于检查有用成果与旅程，未当作自动否决或已证明的记忆收益。

具体改动

关键内容讲解

1. RFC §4 固定 LayerFS ad78914e36b258b360daf13aec164b5655814689。独立核对公开 SDK、exec 实现、Store/runtime/backend 文档和 #411/#456/#446：显式 commit、单 Linux kernel、本地 Store、Disposable/WAL/OFF、无不可达历史 GC、丢失回复和外部副作用均有真实边界。网站的跨机器措辞没有覆盖源码的受支持条件。
2. §5–§8 使用 #6869 的历史版本作为待复现候选，不重复认领 #6875，也不以其补丁当 oracle。controller、凭据与 evaluator 留在 solver 树外；现有 capture 拒绝 symlink，DSH 双插件共存及会话继承明确未验证。回复未知先核原操作，不能盲重放或复制文件恢复权限。
3. §9–§12 把 A1–A6、M0–M3、总预算、预热对照、失败分母、held-out 任务、成本覆盖和停止/回退条件接到双方已发布的 #7406、LayerFS #461。中英文及 README、生成状态、双语路线入口同步。

独立规范依据为 docs/architecture/rfcs/long-horizon-harness-benchmark-research-program-v0.md，版本 15803cd7dcc2fc517049809bbe58998f04f5acac；没有用本 PR 新写的 RFC 自证实现。设计映射如下，均表示设计已覆盖该义务，不表示未来实验已经通过：

原规范 criterion_id	当前处置
6.1 Native outcomes	implemented：§5/§9 A4/A5 以独立正确补丁和最佳有效保留结果衡量
6.2 Efficiency	implemented：§6/§9 同时计准备、全部 worker/coordinator、控制、恢复、验收、清理及已测注意力；缺账单不报完整成本
6.5 Benchmark integrity and anti-cheating	implemented：§5/§7/§9 固定反馈、可见范围、独立 evaluator 与隔离；研究状态另遵循原 Stride §7.7

对主干的风险

最强反例是“快 mount、少存储，却更慢或修得更差”。本文保留冷启动与真实预热 worktree、独立 source/result、副作用隔离、失败/未知分母和准确 durability 边界；没有把上游约 100.1 MB 对 1,994.6 MB 的一次 v1.0 结果外推到本源码或 Agent 总成本。独立阅读上游 dedup/fio 原页也确认了版本和顺序读写/随机写退化的限定。

本次执行生成索引检查、双语检查、13 项 RFC 索引测试、链接与公开边界审阅；标准 diff-selected premerge 的 4 项直接检查及 18 项选中检查全部通过，0 失败、0 人工 hold。两篇 RFC 各 12 个本地链接均存在，17 个外部引用一致。未查询、轮询或等待 CI。没有运行 LayerFS、模型实验、性能复现或安装；这些不是文档静态检查可以代替的证据。

我的整体评价

这是 justified_increment：可用的设计和下一步验收已交付，独立控制、预算与采用链条完整，未扩成第二套控制面。long_horizon 判断为设计层面 improved；user_experience 保持既有产品行为，未来 A6 旅程仍须实际验证。未引入共享 runtime 词汇；本地研究标签没有注册为生产协议。

视觉审阅不适用：完整七路径只有 Markdown 提案、索引和路线内容，没有应用界面或文档 chrome 改动。尚未验证范围为未来打包 Host、LayerFS 实际执行和真实用户理解。最重要的后续观察是 M0 的真实 CLI 基线复现与独立 oracle，而非再增加一个文档或测试计数。批准设计不授予运行、安装、支出或合并权限。

English verdict: APPROVE — 774dcb0e0171cc1dd993d2c2e16494c357a37d17 is a coherent bilingual design proposal. Independent source checks, RFC/index/link validation and scoped premerge passed. It preserves existing ownership, warm controls, full-cost and failure accounting, and a useful receiver-to-clean-patch path. Live integration, causal benefit, native research recovery and packaged adoption remain explicitly unqualified.