[loopx-agent]

Goal And Delivered Outcome

The English and Chinese context-delivery article still described the earlier reading path. This revision follows the same voice task through complete first responses, scope reads before dependent decisions, local transport costs, and exact numeric recovery. Both catalogs and snippets now describe the revised article.

The core 2.2.0 release, the November source snapshot, author-reported process measurements, and archived ANN experiments have separate version labels. Historical scores, byte/token totals, attribution, early-window overhead and the Grid counterexample remain. Causal wording that those observations alone could not support is qualified.

Author Declaration

• Written by OpenAI Codex (GPT-6 family).
• Editorial basis: update the existing bilingual article with the relevant recent optimizations. Existing article and implementation were read at 611019243eeb91a8cf53d5d0f04e562210de0d7e; the branch was rebased onto ecec9f4efd65c84f890af3c58ddb8e2e053b1290 before submission. Public-site files are unchanged between those two base revisions.

Criterion	Disposition	Location	Evidence
Explain complete first reads and body/output budgets	implemented	Paired context-delivery/index.html, first-read-budget	#8300 source and saved-reader regressions
Explain when full scope must be read, preserving authority	implemented	stage-scopes	#8303, File/SQLite selection/claim/ready checks
Separate selected content, CLI stdout and internal transport	implemented	transport-cost	#8298 and #8285 author records; current transport regression
Preserve original values and same-capture recovery	implemented	exact-values and existing recovery section	#8319/#8324, explicit windows, precision and rejection checks
Keep historical measurements and both editions usable	implemented	Article evidence section and both catalogs	Bilingual, link/asset, packaged-site and browser checks

#8304 and the broader #8266 proposal are excluded from delivered-feature claims. No new model inference or benchmark run was started. Process measurements remain attributed author records; the 21 local regressions check mechanisms and correctness, not those performance numbers.

Scope And Continuation

Complete as an editorial source proposal. This changes four HTML files, with no runtime, settings, permissions, figures, example program or original measurement-record changes. Merge and deployment remain with the maintainers. The live site has not been updated by this PR.

The same existing article is the reader journey: choose it in either catalog, switch language, use the contents links, read the code/tables and open the existing figures. No additional user interaction or configuration is introduced.

Validation

• Tested revision: 357f8a8008a720fedbf76bcbd1414d7f417c1e90.
• Run state: finished. Inputs: synthetic and public fixtures.

Check	Result	Coverage and limits
Reader/transport regressions	passed, 21 cases	Bounded attached/linked reads, explicit windows, File/SQLite preparation and ready scopes, one-body transport, exact numbers, off parity, original-capture/hash recovery
node examples/blog-bilingual-index-smoke.mjs	passed	10 paired articles, catalogs, language navigation and retained empirical evidence checks
npm --prefix apps/presentation/site run build -- --base /loopx/	passed	TypeScript, Vite client and SSR public build
Site SEO and base tests	passed	Five prerendered routes, 28 canonical targets, metadata/assets and 2 root/project-base tests
uv run --extra test python examples/static-site-presentation-smoke.py	passed	Existing static presentation contract
Editorial links and preservation	passed	134 local links/assets, all 16 paired heading IDs, source HTML identical to packaged copies, historical figures and measurements unchanged
Public-boundary scan and diff hygiene	passed	Four publication candidates, zero findings, signed commit
Browser reading	passed	Ego Lite, desktop 1440×1000 and narrow 390×844, both languages, catalog → article → language → contents links; code, new process table, recovery and archived charts inspected

The site build does not assemble MkDocs; existing /docs/book/ companion links were outside this local export check. No website deployment, installed-runtime adoption, new matched cost experiment or whole-task/model-token benefit is claimed. The build was made before the rebase; the complete public-site source is unchanged across that rebase, and the 21 mechanisms plus bilingual check were rerun on the final head.

Frontend / Visual Evidence

UI impact: changed editorial content. Source: public fixture. Actual compiled static pages were served under the /loopx/ base; these are packaged local previews, not deployed-site screenshots. The after images match the four HTML files in the tested head.

The existing title, reading column, contents disclosure, typography and figures stay in place. The new three-column table makes the differing cost directions visible. On narrow screens the text and table fit the page; long commands scroll within code blocks. Dense existing figures retain their full-size links and adjacent textual values. Failure, permission and counterexample text remain in the reading flow.

Before: source 611019243e, desktop 1440×1000.

After: 357f8a8008a720fedbf76bcbd1414d7f417c1e90, desktop 1440×1000.

After: Chinese first viewport, 390×844.

After: the process-cost section and its unchanged-content/increased-stdout counterexample, desktop.

No public preview deployment was created; the revised source is available in this PR.

Type of Change

• Documentation update

LoopX Area

• Public docs or presentation surface

Shared-authority RFC fixture impact

N/A: article and catalog text only. No typed authority, backend or RFC changed.

Boundary Checklist

• Public sources and public-safe screenshots only; no private state, local paths, raw trajectories or credentials.
• Existing contributor and archived experiment attribution retained; no duplicate benchmark execution.
• Scope is the paired article and its two catalog entries.
• No paired RFC changed.
• Visual evidence is attached here; no screenshot pack is committed.
• The commit has a DCO Signed-off-by trailer.

[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)

动机 / Motivation

阅读双语文章、想判断上下文读取机制与实验结果分别说明什么的读者。旧文章仍围绕较早的读取路径，并把部分后程结果直接归因于注意力节省；新版沿同一语音任务解释完整首读、范围补读、内部传输与数值恢复，同时分开当前源码和历史实验。编译后的两种目录均可进入文章，语言切换和目录定位可用；新进程成本表与公开作者记录一致，历史图表数值保留，正文不再把这些观察当作模型或跨题收益证明。仅四份文章与目录 HTML 的编辑更新；不改变 runtime、设置、权限、原实验或安装状态，不宣称新的性能复验和模型收益。

评审从实际读者或维护者要完成的工作判断：有可读的事实、原要求与可信回读，才是有用交付。编译成功、元数据一致或通过数量不能单独证明这些结果。这里保留当前证据边界，不要求本 PR 完成无关的整体 runtime 或长期实验。

改动思路 / Approach

在原文章、原双语目录和原 reader 示例上补解释即可；无需新页面、开关、服务或测量框架。 仅四份文章与目录 HTML 的编辑更新；不改变 runtime、设置、权限、原实验或安装状态，不宣称新的性能复验和模型收益。

原静态站点已有文章、语言导航、目录、表格和可下载的教学示例，复用它们便能解释新源码；当前收益首先是避免读者混淆几种成本，而不是宣称缩减字节必然提高成果。目录展开属于自愿阅读，不增加授权或必经确认。

具体改动 / Changes

规范依据是修改前的 docs/guides/context-delivery.md，spec_revision ecec9f4efd65c84f890af3c58ddb8e2e053b1290。逐项判断 Keep three meanings of reuse separate；Read and recover the right object；Qualify default adoption and retirement。对应原文的复用/保留、同源恢复和资格边界都由文章、原reader及公开来源对应；改动本身的说明不能反向认证规范。

关键代码讲解

• h3#first-read-budget：区分正文与整条响应预算，保留同 capture 补读及不重复工作。
• h2#transport-cost：将 CLI、内部 RPC、正文和 stdout 列在同一个对照表，明确数字属于作者记录。
• h3#exact-values：解释整数、浮点与负零的原值恢复，区别文件完整和返回事实准确。
• article.post-listing：让目录摘要与两篇修订后的正文一致，沿用原语言导航和资产。

完整四路径仅 +92/−24，没有 CSS、图像、示例程序或 runtime 修改。两个目录、元数据、导语和更新时间一起更新；四个新增章节沿原阅读路径解释预算、范围、传输和原值。所有原12个 heading ID 保留，两种语言现在同为16个；历史 figure 块逐字不变，历史表格数值序列完全相同。208个 href/src 条目检查没有新增本地失效项；原主站/MkDocs companion 的单独构建边界保留。

六个引用 PR 的真实 merge 均是所标源码 611019243eeb91a8cf53d5d0f04e562210de0d7e 的祖先；#8298 的4→3 CLI、5→4 RPC、326,204→271,741 B、正文8,046 B不变、stdout32,846→33,060 B，及#8285的53,999→38,225 B与完整序列1,021,975→1,006,203 B，均与公开作者正文相符。只确认出处和版本，不把作者数字转成这次独立性能复验。

对主干的风险 / Risks

独立实际 CLI 的同一80条教学历史，在不可变 base 和本 head 均保留57,014 B原文、6,093 B当前必要视图、全部10条要求、精确恢复以及无 live Goal/authority 写入。文章原有5,107 B属于已注明早期版本的教学测量，不将本轮6,093 B倒填成旧实验结果。新增成本表经真实本地 HTTP 读取，与公开记录一致；把私有副本的326,204改成326,205后，独立来源 oracle 失败，正确 head 通过。

相关 reader 21 passed / 3 skipped；另行4个 File/SQLite selection→claim→ready长范围场景验证220条合成 scope、停止/审批、显式off及原Turn保持。编译、10对文章双语检查、静态呈现、SEO五路由/28目标及2个base测试、标准premerge均通过。三项 skip 不算通过；局部本地结果不认证安装或模型行为。

主要风险是把文件存在、某次EOF、局部wire减少或一次较晚峰值读成真实采用/全程收益。新版把这些层次、原Turn恢复、未知成本和早窗/Grid反例放回原文；没有引入新的默认证明。

按本 Goal 策略未查询、轮询或等待 CI。公开图表出处与源码回归分别核验，未新启模型、正式benchmark或性能实验。 线上部署、真实读者理解、模型采用与完整任务净成本未测；MkDocs 伴随站点未构建。

我的整体评价 / Verdict

APPROVE。在原页面补完整解释、限定测量含义，能独立完成这次编辑更新；没有理由为了文章增加新 runtime owner、配置或实验。历史较好成绩与反例同处阅读路径，原范围和恢复义务没有因更短导语被删掉。

在原文章、原双语目录和原 reader 示例上补解释即可；无需新页面、开关、服务或测量框架。 仅四份文章与目录 HTML 的编辑更新；不改变 runtime、设置、权限、原实验或安装状态，不宣称新的性能复验和模型收益。

首屏保留单一文章标题与阅读列，目录默认折叠；成本表把不同方向的量放在一起，正文与限制可沿原目录读到，390px 下未用缩小文字换取容纳。 线上部署、真实读者理解、模型采用与完整任务净成本未测；MkDocs 伴随站点未构建。

English verdict: APPROVE at 357f8a8. The existing bilingual reader now separates source versions, complete reading/recovery, process measurements and archived outcomes; local built-page and CLI checks preserve the scoped contract. Deployment, comprehension and net model benefit remain unverified.