[loopx-agent]

Goal And Delivered Outcome

2.2.0 已发布，Context delivery 稿仍把它称为候选；EdgeBench 把早期 Hook 的“完整分页后保持连续性”写成了当前机制。三篇文章还需要讲清同一个区别：磁盘能恢复完整结构，不等于模型已经读到决定所需的事实。

本次更新三篇中英文章，以实际发布的机制串起渐进读取、恢复语义和研究证据：

• Context delivery：补充默认开启路径的文件树，结合现有语音任务、JSON Pointer、批量选读、目录和差分数据解释机制；新增“隐私约束没变，但 Agent 漏读了”的反例。当前 recovery 可直接选读必要事实，无须先通读旧基线或重放全部差分。
• TS 控制面：从历史回执不能续租，接到保存的基线不能证明模型已读；补充 compaction 中恢复原任务、保留原结算身份的语义。保留早中期事务、效果身份、协调、字段更新等核心案例。
• EdgeBench：明确早期 Hook 路径的历史条件，区分合成字节测量、发布安装/恢复验证、真实研究轨迹三类证据；保留全部历史表格、失败、比较窗口和未完成资格。

Outcome basis: maintainer-requested post-release blog refinement. Base: main at 8ac384cdcfcf3d6c683ffab43db1a5c14dba0343. No issue is closed by this editorial change.

Author Declaration

• Written by: model_agent — OpenAI Codex.
• Specification and revision: maintainer-requested editorial scope above; the released required-work-context contract at v2.2.0 supplies the mechanism. The historical teaching measurements retain their original source.

Criterion	Disposition	Path / evidence
Explain progressive reads and the enabled native file path with a concrete example	implemented	Both blog/context-delivery/ articles; file tree and voice privacy example
Separate complete file reconstruction from model reading/retention	implemented	#file-and-model and TS #context-continuity; version-pinned owner source checked
Preserve historical data and its experimental scope	implemented	All six pages’ original table HTML remains byte-identical; no experiment rerun or savings inference
Deliver a complete bilingual proposal for review, without merging	implemented	Six static HTML pages; this PR remains unmerged

Self-check: read the modified bilingual text and relevant released owners; build, SEO, static invariants and actual desktop/mobile browser checks below. No new runtime/API/configuration, scoring or fixture behavior.

Scope And Continuation

Complete within this editorial slice. Review only: leave this PR unmerged pending maintainer content review. The released product and historical measurements remain independent of this proposal. Broader long-horizon effectiveness, complete-cost evidence and editorial acceptance are not certified by this PR.

A same-domain readability refinement replaces the long default-path list with a compact file tree and keeps its labels inside the 390px viewport. No new abstraction or tracked capture pack is introduced.

Validation

• Tested revision: 4cc1cb1bd56b5d999f5ac1c322555d45cbd70220; the six tested pages are byte-identical before/after rebase onto current main.
• Run state: finished. Input classes: public_fixture, synthetic.

Check kind	Result	Evidence / limitation
static	passed	npm run build --prefix apps/presentation/site — TypeScript and Vite client/SSR build
static	passed	npm run smoke:seo --prefix apps/presentation/site — five prerendered routes, 28 canonical targets, assets and metadata
regression_parity	passed	Original table HTML unchanged in all six pages; unique fragment IDs and the changed cross-article fragment links resolve
static	passed	Six-file public-boundary scan: zero errors/warnings; git diff --check clean; DCO sign-off present
manual	passed	Actual built-site desktop 1716×995 and narrow 390×844 views in Ego Lite; Chinese/English file-tree layout, missed-read example, cross-article click and browser Back
integration	not_applicable	Static article changes; no product runtime/backend, native host or benchmark was changed or requalified

Source checks distinguish the current v2.2.0 implementation from the older measurement checkout. Byte counts remain byte counts; no token, financial cost or task-outcome improvement is inferred. CI status is independently visible below and was not substituted for these checks.

Frontend / Visual Evidence

• UI impact: changed (article content and the default-path diagram).
• Source data: public_fixture / synthetic article material only.
• Boundary: actual locally built static site, before 07b555494550c12ff6e3dc4b139cdf7cd1bdae32 and after the tested head above; these captures do not claim a deployed Pages update or installed App qualification. The intervening base commits do not change these pages. No persistent public preview was created.
• Attention review: retain the existing article hierarchy and on-demand table of contents; use one file diagram, then the concrete privacy decision. Existing long tables remain in their scroll containers. Both language file trees fit the narrow viewport without horizontal overflow. No loading/error/permission state or interaction behavior changed.
• Craft review: existing article styling and tokens retained; removed the nested preformatted frame inside the diagram, with no new grouping-only cards, motion, status glyphs or color values.

First viewport — after, desktop

Before — desktop first viewport

First viewport — 390px before and after

Before:

After:

Changed mechanism sections — 390px

The enabled file path:

Saved files versus the facts the Agent has read:

Type of Change

• Documentation update

LoopX Area

• Public docs or presentation surface

Shared-authority RFC fixture impact

N/A — this explains existing semantics and historical work; it makes no migration, provider promotion or shared-authority implementation claim.

Boundary Checklist

• Diff, body and attached screenshots contain no private operating context, credentials, raw research traces or local paths.
• No benchmark run, scoring change or duplicate research work.
• Scope remains the six bilingual article pages.
• No paired RFC changed.
• Actual before/after desktop and narrow screenshots are embedded above; captures remain outside the tracked tree.
• The single commit has a DCO sign-off.

[loopx-agent]

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

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

这份六页文章更新可以批准。判断覆盖版本事实、完整 diff、历史证据和实际阅读体验；作者保留的 maintainer 内容审阅与未合并边界继续有效。

动机 / Motivation

阅读三篇中英技术文章、想判断当前读法与历史数据含义的读者。原稿仍把 2.2 称为候选，并容易把早期 Hook 的完整分页条件当成当前文件差分前提；本次区分已发布机制、实际必要读取与历史研究，补上隐私约束未变但漏读的具体决定。当前静态构建和六页实际浏览支持可阅读、可追溯的说明；34 个原表格及原测量文件不变，没有把发版当成模型收益或旧实验补验。本 PR 只交付文章提案；部署、安装、真实模型理解／采用和完整研究成果／费用资格均独立，作者要求内容审阅前保持未合并。

改动思路 / Approach

复用现有静态文章、样式和锚点，把版本与证据范围说清；不新增 runtime、配置、权限或第二份交付决策 owner。本 PR 边界是六个 HTML 文件的公开讲解与文件树呈现，保留原测量和历史失败。只换发版标签不足以修正“文件可还原”与“模型已读”的混淆，另开一篇文章或增加功能入口也没有必要。当前修订把具体语音决定、恢复身份与历史测量串回原文章，并用现有 diagram 容器中的紧凑树替代长列表；图明确标为布局示意，实际文件和命令仍以当前调用返回为准。

本轮将经验提示用于检查真正交付给读者的内容和整页可读性，不把 build 成功当作解释成立，也不为静态讲解强加完整模型实验。历史 Hook 的限定没有消失，而是放回它原来的研究窗口。

具体改动 / Changes

完整审阅六个 HTML 文件，+40/-14，head 4cc1cb1bd56b5d999f5ac1c322555d45cbd70220、merge base 8ac384cdcfcf3d6c683ffab43db1a5c14dba0343。规格为 released required-work-context，固定 revision 2c3524857e8a3dd553335b79444333786f492c43。已用官方 release API 与远端 peeled tag 核实 v2.2.0 于 2026-11-01 18:25:40Z 正式发布。

本次映射行 FILE-READING、ORIGINAL-TURN、SCOPED-DEFAULT、COST-BOUNDARY 均 implemented：分别对应保存文件不要求模型通读旧基线、未结束工作保留原 Task／capture／结算身份、受支持 native 范围内的默认与 off 优先级，以及字节机制不能替代模型成果／费用证据。它们依据发布版已有合同，不由本 PR 的改写自证。

关键代码讲解（静态 HTML 内容与锚点）

1. Context delivery 的 #native-path 使用原 diagram 与样式，说明 source、必要视图、周期交付和 reader；没有新增脚本或真正的文件生成流程。
2. #file-and-model 给出决定性的反例：即使隐私字段未变、文件已保存，漏读约束的 Agent 在外发音频前仍要补回必要事实并遵守授权。选读 EOF、磁盘基线和模型掌握的内容没有被互相替代。
3. TS 文的 #context-continuity 从旧 receipt 不能续租，接到文件不能证明理解；原工作中断后保留 Task 和结算身份。新跨文章链接能到达同语言的具体例子，浏览器返回也回到原段落。
4. EdgeBench 的 历史 Hook 段落 与末段区分合成字节、安装／恢复检查、真实研究轨迹，不把不同版本、窗口和指标拼成节费结果。两版所有 34 个原表格的 HTML 字节、全部 script blocks、原 example.py 和 measurements.json 均与 base 一致；原测量三项 core owner 在其两个声明 revision 的 blob 也已独立核对相同。

对主干的风险 / Risks

最强风险是文案把“正式发布”或“能重建文件”升级成模型保留与真实研究收益。已逐条比较发布版合同和六页新旧措辞，并实际读了隐私反例、历史负面成本表及三个证据层的底部说明。新文案保留原研究未完成状态；没有运行新模型、音频外发或 benchmark 来补写旧结果。

PASS：生产 site build、SEO、六页 fragment／ID 检查、原表格和脚本字节对照、公开边界扫描、DCO；quick premerge 的 4 项直接和 3 项选中检查通过。实际 built site 在 1280×900 与 390×844 核对修改区域，文件树在窄屏内完整显示；中文首屏／展开目录、历史长表、两版底部和跨文章返回路径已查看。正文维持 16–17px；可读性来自实际画面，不只来自无 overflow 的数值。

HOLD：当前 base 的 CQR 已 stale，aggregate premerge 退出 1，合并 gate 仍未通过；作者还明确保留 maintainer 内容审阅。CI 未查询、轮询或等待，状态／原因保持未知。新 ID 是本地 HTML 锚点，公开文章不自动加载为 Agent/system instruction，亦不产生默认、权限或运行状态变化；因此没有把一个 runtime 关闭开关当作静态讲解的展示条件。

我的整体评价 / Verdict

文章这一有界成果已达到请求：读者能区分当前文件机制、决定需要的读取和历史数据能证明的范围。long_horizon 为 preserved，历史失败／原时钟／未完成研究不被发版覆盖；user_experience 为 improved，有具体负例、清楚版本入口和可用的跨文章阅读路径。已在实际 built site 查看六页的桌面／窄屏修改区、历史长表与底部说明，另核中文首屏、展开目录和 before/after 文件树；正文保持16–17px，窄屏图与段落可读，未见新增页面横向溢出。未验证站外部署、安装版 App、真实模型理解／采用或研究收益；本轮没有启动研究或模型运行。

复用现有静态文章、样式和锚点，把版本与证据范围说清；不新增 runtime、配置、权限或第二份交付决策 owner。本 PR 边界是六个 HTML 文件的公开讲解与文件树呈现，保留原测量和历史失败。代码量与内容相称，现有样式和原资料被复用，文件树的局部简化可以保留。批准只代表此精确 head 的文章提案可接受；保留内容审阅、CQR 和独立维护者门禁，本轮不合并、不部署。

English verdict: APPROVE at 4cc1cb1bd56b5d999f5ac1c322555d45cbd70220. The bilingual articles match the released mechanism while preserving historical evidence and readable navigation; CI, deployment and maintainer content acceptance remain separate.