feat(agentx): publish the AgentX optimizations pages / 发布 AgentX 优化专题页面

[functionstackx]

Summary

AgentX's first months produced 50+ upstream pull requests across the serving stack, and that work only existed as an internal document. This publishes it.

• /agentx/optimizations — the industry-impact index: why agentic traces surface costs a single-turn 8k/1k scenario cannot, a primer on the distributed inference ecosystem, a card per project, the day-zero enablement and correctness section, and what the AgentX matrix activates.
• /agentx/optimizations/<framework> — one page per project: vllm, sglang, tensorrt-llm, atom, aiter, dynamo, lmcache, mooncake. Each opens with its headline measurements, then works down the stack section by section.
• Wiring on /agentx — a callout card with a button into the index plus a direct link per project, so a reader who already knows which engine they run can skip the index.

Also in this PR

• A results CTA on the AgentX card — the card explained how the benchmark is built but offered no route to the numbers. It now carries a primary "View AgentX Performance Results" button to /overview (/zh/overview on the Chinese page).
• Restored footer and landing copy — reverts the footer description and the "Explore InferenceX" lead and platform-coverage lines to their pre-feat(agentx): publish AgentX methodology and canonical routes #755 wording, in both languages. The Chinese footer e2e assertion follows.

Sourcing

Every claim carries the pull request it came from — 123 distinct PRs, extracted from the source document's hyperlinks rather than retyped. Sections store them as { repo, number } instead of prose links, so:

• the reference set is shared by both languages and cannot drift between them,
• a PR chip renders as vLLM #43447 and links to https://github.com/vllm-project/vllm/pull/43447,
• the "50+ upstream PRs" claim in the lead is asserted against the registry in a unit test.

Ten diagrams from the source deck ship under public/images/agentx-optimizations/ (2.5 MB total, against 8.8 MB for the existing methodology figures) and are keyed by asset, so captions are translated but the image is never duplicated per locale.

The editorial notes in the source document — the "TO BE UPDATED" chart annotation and the request for more PRs — are deliberately not published, and a test fails if those markers reappear. The PR-timeline chart they annotate is left out for the same reason; it can land once it is refreshed.

Chinese

/zh/agentx/optimizations and /zh/agentx/optimizations/<framework> are hand-authored siblings for all 16 pages, with zhAlternates / enAlternates hreflang pairing, ZH_OG_LOCALE, and sitemap entries. Product names, PR references, hardware SKUs, and identifiers stay English; the prose around them is translated. /agentx already sits under ZH_MIRRORED_ROUTES as a prefix route, so the header toggle follows automatically.

Tests

• agentx-optimizations.test.ts (13 cases): unique slugs and section ids, known repositories, no duplicate PR inside a section, the 50+ claim, https-only references, every referenced figure present on disk with sane dimensions and real alt text, every declared figure used exactly once, no editorial markers, and EN↔ZH structural parity (section ids, paragraph counts, highlight values, figure keys, PR lists, and CJK content in every translated string).
• agentx-optimizations.cy.ts (6 specs): the /agentx callout and its per-project links, the index with its cards and sourced PRs, a project page with measurements/figures/PR hrefs, a 200 for all 16 pages, a 404 for an unknown slug, and the Chinese index and project page including hreflang.
• Full local run: typecheck, lint, fmt, test:unit (176 files), and bun run build — all 16 pages prerender statically.

Overlay support

Not applicable — these are static content pages with no inference or evaluation chart data.

中文说明

AgentX 上线头几个月在推理栈各层推动了 50 多个上游 PR，此前这些内容只存在于内部文档中。本 PR 将其正式发布。

• /agentx/optimizations —— 行业影响索引页：说明为什么智能体 trace 能暴露单轮 8k/1k 场景测不出的开销，包含分布式推理生态简介、各项目卡片、day-zero 支持与正确性章节，以及 AgentX 测试矩阵激活了什么。
• /agentx/optimizations/<framework> —— 每个项目一页：vllm、sglang、tensorrt-llm、atom、aiter、dynamo、lmcache、mooncake。每页先给出关键测量结果，再按小节自上而下深入推理栈。
• /agentx 页面接入 —— 新增引导卡片，提供进入索引页的按钮和每个项目的直达链接，已确定所用引擎的读者可以跳过索引页。

本 PR 的其他改动

• AgentX 卡片新增结果入口 —— 该卡片此前只说明基准测试如何构建，却没有通往结果的入口，现在新增主按钮"查看 AgentX 性能结果"，指向 /overview（中文页面为 /zh/overview）。
• 恢复页脚与落地页文案 —— 将页脚描述以及"探索 InferenceX"的引导语与平台覆盖说明恢复为 feat(agentx): publish AgentX methodology and canonical routes #755 之前的中英文版本，中文页脚的 e2e 断言同步更新。

引用来源

每条结论都附带对应的 PR，共 123 个，全部从源文档的超链接中提取而非手工重录。各小节以 { repo, number } 结构而非正文链接存储：

• 中英文共用同一份引用集合，不会产生漂移；
• PR 标签渲染为 vLLM #43447，链接到 https://github.com/vllm-project/vllm/pull/43447；
• 导语中"50+ 上游 PR"的说法由单元测试对照注册表校验。

源文档中的十张示意图放在 public/images/agentx-optimizations/ 下（合计 2.5 MB，作为对比，现有方法论图片为 8.8 MB），并按资源 key 引用，因此只翻译图注、不按语言重复存放图片。

源文档中的编辑批注——"TO BE UPDATED"图注与补充 PR 的请求——刻意未予发布，并有测试在这些标记重新出现时报错；该批注所标记的 PR 时间线图表同样暂未收录，待其更新后可再补上。

中文页面

/zh/agentx/optimizations 与 /zh/agentx/optimizations/<framework> 为全部 16 个页面提供人工撰写的中文版本，配套 zhAlternates / enAlternates hreflang 配对、ZH_OG_LOCALE 以及 sitemap 条目。产品名、PR 引用、硬件 SKU 与代码标识符保持英文，其余正文均已翻译。/agentx 已作为前缀路由登记在 ZH_MIRRORED_ROUTES 中，因此顶栏语言切换自动生效。

测试

• agentx-optimizations.test.ts（13 项）：slug 与 section id 唯一、仓库名合法、同一小节内不重复列出 PR、"50+"说法成立、引用链接仅限 https、引用的图片均已随仓库提供且尺寸与 alt 文案合理、每张声明的图片恰好使用一次、不含编辑批注标记，以及中英文结构一致性（section id、段落数量、highlight 数值、图片 key、PR 列表，并检查每条译文包含中文字符）。
• agentx-optimizations.cy.ts（6 项）：/agentx 引导卡片及各项目链接、索引页卡片与 PR 引用、项目页的测量结果/图片/PR 链接、全部 16 个页面返回 200、未知 slug 返回 404，以及中文索引页与项目页（含 hreflang）。
• 本地完整校验：typecheck、lint、fmt、test:unit（176 个测试文件）以及 bun run build——16 个页面全部静态预渲染。

覆盖层支持

不适用——这些是静态内容页面，不涉及推理或评估图表数据。

🤖 Generated with Claude Code

---

Note

Low Risk
Mostly static marketing/content pages and copy changes; no auth, data pipelines, or inference runtime logic. Risk is limited to content accuracy and SEO/metadata correctness.

Overview
Publishes AgentX industry-impact content that was previously internal: a static index at /agentx/optimizations and eight per-project pages (vllm, sglang, tensorrt-llm, atom, aiter, dynamo, lmcache, mooncake), with matching /zh/... routes, sitemap entries, metadata/JSON-LD, and prerendered params. Content lives in agentx-optimizations.ts plus a mirrored Chinese prose layer; sections cite upstream work as structured { repo, number } PRs (123 distinct), with diagrams keyed from public/images/agentx-optimizations/.

/agentx gains an optimizations callout (index CTA + per-framework links) and a primary View AgentX Performance Results button on the methodology card linking to /overview (locale-aware). New UI components render the index, articles, figures, and analytics-tracked links.

Also reverts footer and landing “Explore InferenceX” strings to pre-#755 wording (EN/ZH), updates the Chinese footer e2e expectation, and adjusts nudge banner Cypress setup so the launch modal stays dismissed when testing the banner (backdrop overlap).

Tests: agentx-optimizations.test.ts (registry parity, 50+ PRs, figures on disk) and agentx-optimizations.cy.ts (navigation, PR hrefs, 16 pages, 404).

^{Reviewed by Cursor Bugbot for commit eb56ba3. Bugbot is set up for automated code reviews on this repo. Configure here.}

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
inferencemax-app	Ready	Preview	Aug 19, 2026 3:19am

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 3615d1c. Configure here.}