feat(agentx): publish AgentX methodology and canonical routes

[cquil11]

Summary

• make /agentx and /zh/agentx the canonical AgentX surfaces, with permanent redirects from /datasets and /zh/datasets, including nested paths and query strings
• surface AgentX in the primary navigation and footer, with a localized NEW badge on desktop and mobile; navigation and launch badges share one 32×16 treatment with labels centered at an exact zero-pixel offset
• keep a concise AgentX overview and add /agentx/methodology plus /zh/agentx/methodology for the full methodology, including 21 figures, captions, alt text, and primary-source links; 15 supplied source figures use their original 4200-pixel files
• explain opt-in trace collection, deterministic payload replacement, v1.0 filtering, AIPerf DAG replay, warmup and profiling controls, concurrency interpretation, speculative-decoding normalization, DRAM-offload controls, and reproducibility limits
• position InferenceX as an agentic inference benchmark across landing, dashboard, overview, About, FAQ, blog, metadata, manifest, share text, and machine-readable discovery surfaces; reserve AgentX for the named long-context, multi-turn coding workload
• add an Agentic inference glossary category covering AgentX, agentic coding workloads, trace replay, closed-loop benchmarks, and subagents
• replace the Kimi K3 launch promotion with an agentic-results banner and modal: the compact banner says Agentic benchmark results are live; the modal names Kimi K3, DeepSeek-V4-Pro, MiniMax-M3, Qwen3.5 397B, and GLM-5.2; both use View results actions and link directly to /inference?i_seq=agentic-traces
• capitalize the visible AgentX Methodology heading and keep all new user-facing content synchronized with natural Simplified Chinese siblings

This PR now includes the semantic-marketing and glossary follow-ups previously reviewed in #756 and #757.

Source and editorial audit

The public methodology is adapted from the methodology section of InferenceXv3/AgentX: How does CUDA Moat Hold up in Agentic Coding?. It summarizes the measured methodology instead of reproducing internal or unfinished prose.

The page omits unresolved cost claims, TODO sections, promotional superlatives, the internal proxy warning, and the unattributed distillation meme. It does not claim that placeholders reconstruct original prompts, code, or tool payloads. New copy was audited against Wikipedia: Signs of AI writing for generic puffery, canned contrast, vague attribution, repetitive conclusions, excessive headings, and related patterns.

Validation

• bun run fmt
• bun run lint
• bun run typecheck
• bun run test:unit — 219 files, 3,957 tests
• focused Cypress component coverage — 24 header tests
• focused Cypress end-to-end coverage — 37 launch, navigation, locale, route, redirect, and methodology tests
• E2E_FIXTURES=1 bun run build — 950 statically generated pages
• desktop and mobile browser verification in English and Simplified Chinese — no horizontal overflow, framework overlay, or page errors

Latest local validation passes all 3,957 unit tests, 27 focused Cypress tests, the 950-page fixture production build, type checking, lint, and formatting. Browser verification confirms all 21 English and Chinese figures load without missing alt text, overflow, or framework errors. Post-push checks were started but not awaited at the request of the reviewer.

Scope

This PR changes static AgentX routes, methodology content, site positioning, glossary content, navigation, and launch promotions. It does not change benchmark measurements, database contents, API response contracts, or inference chart calculations. Unofficial-run overlays are unaffected.

---

Note

Low Risk
Changes are routing, static content, navigation, and marketing copy; benchmark APIs, chart math, and measurement pipelines are untouched. Redirect and link churn is the main regression surface for bookmarks and external references.

Overview
AgentX becomes the public surface for trace datasets and methodology: new /agentx and /zh/agentx pages (overview + dataset list), full /agentx/methodology siblings with 21 figures and sourced copy, while /datasets → /agentx 308 redirects preserve paths and query strings. Old standalone datasets index pages are removed in favor of this tree.

Navigation and discovery move from “Datasets” to AgentX in the header (with a shared NEW badge via NewBadge), footer, scenario tooltip, timeline deep links, sitemap, and llms.txt. Desktop primary nav now uses the xl breakpoint and adds overflow checks around 1009–1024px so the hamburger layout does not clip.

Site messaging reframes InferenceX as an agentic inference benchmark (AgentX as the named long-context coding scenario) across landing, overview titles (“Agentic Inference Costs”), About/FAQ, glossary (new Agentic inference category and terms), blog metadata, manifest, and share text. Several blog MDX captions/links point to /agentx instead of /datasets.

Launch promotions swap Kimi K3 storage keys and copy for agentic benchmark results; banner/modal View results routes to /inference?i_seq=agentic-traces.

Tests update Cypress for new routes, methodology e2e, redirects, and nudge/badge behavior; unit tests for conversation hrefs use /agentx. next.config adds image qualities: [75, 100] for methodology figures.

^{Reviewed by Cursor Bugbot for commit 5aa625f. 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 18, 2026 10:40pm

[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 0fe51da. Configure here.}