HTML.CAFE — CHINESE/CJK AND AGENT-THEMED SOURCE REVIEW
2026-09-05 UTC. Local files only; no network requests, browser rendering, script execution, or configured API calls.

MEASURED SCREEN
Parent-run html_cafe_language_audit.py reports 9492 matching x########.html files, 3786 with at least one visible-text CJK ideograph, and 271 matching its language/model/test/memory keyword regex. These are file counts, not people, agents or independent episodes. The keyword count is not restricted to CJK files; an English/Japanese page naming Qwen can match.
The script removes script/style/noscript tags with BeautifulSoup, then uses U+4E00–U+9FFF as the CJK signal. One ideograph is enough; Japanese text qualifies. This is not Chinese-language identification. Parsing malformed HTML can expose bundled script strings as visible text: JoySpace is a concrete example. The counts establish an overlooked mixed-language subcorpus, not 3786 Chinese users or 271 swarm candidates.

BOUNDED REVIEW:12 SOURCE FILES
The following classifications use actual saved HTML, script functions, embedded data, metadata and selected code context. Every file was parsed as text. Referenced external scripts were not fetched. Any described client capability remains source-level, not runtime-verified.

1. https://html.cafe/x2e718704
Title: AgentDesk · 本地优先的智能多Agent工作台
Category: marketing/architecture landing-page demo. Talks about task trees, local/cloud models and an upcoming download. Its397-character inline JavaScript only smooth-scrolls anchor links. The file itself is not a running orchestration backend or a completed multi-agent trace.

2. https://html.cafe/x3ccf1a2a
Title: AgentDesk · 本地优先的智能多Agent工作台
Category: expanded variant of the same landing-page concept. Adds benchmark tables, commercial roadmap and future enterprise plans. Inline JavaScript is again 397 characters of anchor scrolling. Benchmark numbers and claimed product features are page claims, not verified measurements. Two similar pages show a design/content relationship, not evidence of autonomous coordination.

3. https://html.cafe/x07e82a7b
Title: Coding Agent
Category: Japanese-language browser coding client. UI asks the user for model-service credentials and GitHub repository connection details. Code contains repository reads, model chat and a GitHub write operation guarded by an approval UI. None was invoked. A credential-input interface plus API code is ordinary application source, not evidence that a model autonomously posted this page or used it as research scratch memory.

4. https://html.cafe/x69226946
Title: NIE Coding Agent — Multi-AI Collaborative Coder
Category: Japanese-language multi-model coding interface/client. Source defines runAgent, deep-think/review/search/image tools, a chat-completions backend, local browser sessions and an interactive activity timeline. The initial visible state has zero logs and awaits user input. This is meaningful implemented client logic, not merely static branding, but contains no observed completed research episode or artifact-publication history. No backend availability or actual collaboration verified.

5. https://html.cafe/x001fcd09
Title: AI 财富管家 · 实验页
Category: elaborate financial-app prototype. Inline script contains preset action steps, delayed transitions and result displays; functions include runAgentSteps and runAgentExec. No fetch calls in the saved file. The so-called agent execution is represented by timed UI behavior rather than evidence of executed financial operations. Personal and financial-looking values are omitted; whether every displayed value is fictional is not independently known.

6. Personal AI-chat archive (public-report URL omitted)
Title withheld here because it identifies a personal conversation archive.
Category: static personal AI-chat archive with Kimi labels, no inline script. Contains sensitive personal discussion, deliberately not summarized or quoted. The model label is part of the pasted conversation and does not identify the publishing actor or prove a lab origin. No research-task scratch-memory behavior observed in the reviewed structure.

7. https://html.cafe/x1fdb9e08
Title: JoySpace 私人空间智能整理
Category: bundled React-like file-organization prototype with extensive local-storage state, preset file/folder fixtures and UI flows. Full-file scan finds one fetch call in the module-preload helper, not a demonstrated research-data retrieval flow. Browser-parser behavior was not tested. BeautifulSoup leaves a large amount of bundled code in its visible-text extraction, making this file an important language-detector caveat. Names, file lists and personal-looking folder details are omitted.

8. https://html.cafe/x30ce82d4
Title: Qwen Coder Studio — Local WebGPU AI
Category: Japanese-language local-model coding client. Source imports Transformers.js, defines model loading/from_pretrained, a tool loop, virtual files and previews. It also defines DuckDuckGo lookup and a URL-fetch proxy. Thus a broad UI claim of no server communication would not describe all optional tools; source behavior is richer than its tagline. No model download or tool call executed. Empty initial conversation and generic client tools do not establish an observed escaped-agent episode. Qwen branding is not author/lab attribution.

9. https://html.cafe/x471b8b78
Title: PDD / TEMU 阶段性研究简报
Category: human-facing research presentation with navigation, editing, export/print and chart-library material. A named author/department and a June2026 date appear as page text; personal attribution and financial details are withheld. No inline fetch calls detected. A formatted report could be human-written or AI-assisted, but is not itself evidence of an autonomous agent's scratch-memory write. Generator identity not established.

10. https://html.cafe/x3ac313a6
Title: G66 任务模块与测试场景Agent应用分析示例
Category: game-QA planning/coverage presentation. Discusses test-agent strategies, game state transitions, QA tools and proposed automation coverage. No inline fetch calls. These are descriptions of testing workflows and claimed percentages, not recorded web-research tasks or measured execution logs.

11. https://html.cafe/xc75c715f
Title: 路演 · 基于“五个发生”的上市公司管理决策 Agent 平台(单文件版)
Category: business pitch plus interactive management-app demo. Local editing/save/export functions and hardcoded scenario/result data underpin the UI. No inline fetch calls. A displayed author byline is a claim within the document; omitted personal/financial details are unnecessary to classification. Agent-role architecture and preset output are not proof of functioning autonomous company analysis.

12. https://html.cafe/xc2075e4e
Title: EverMind AI 面试准备 & 海外推广简报
Category: static interview-preparation/product-marketing brief about persistent agent memory. No inline script. Performance, research and product claims were not fact-checked in this local task. This is a document describing agent-memory products, not an agent using the page to persist its own state.

ATTRIBUTION AND BEHAVIORAL CHECKS
None of the 12 has a generator/author meta tag, and the inspected HTML-comment search found no explicit generated-by/created-by provenance. Some pages have visible author bylines, while the personal archive labels a conversation model; neither establishes who uploaded the HTML. Shared milkymouse.com script references are external includes common to these saved pages, not demonstrated author identity. Those external scripts were not fetched or executed.
A narrow raw-source screen across the 12 found no ZZZROOTTEST, md.succ.ai, datausa.io, ECDC, TESTREF, scratch memory or shared memory strings. This is a limited marker screen, not a comprehensive absence claim. Source APIs, generic fetch tools, demo task lists and quoted agent instructions require actual dated execution/context joins before they can be classified as swarm traces.

DATE AND COVERAGE LIMITATIONS
The handoff describes this 9492-page cache as selected May–July material. This review did not independently reconstruct the gallery-selection procedure or capture history. Content dates, copyright years, report months, future roadmaps and filesystem mtimes cannot prove publication dates. No reviewed page has a newly verified pre-disclosure WARC/Wayback capture from this task.
Only 12 files were closely reviewed. The271-match detector is broad and the other 259 files have not thereby been cleared. Nor do keyword-negative pages exclude English research artifacts, hidden script data, or missed language variants. A useful next step is obtaining actual gallery/crawl timestamps and screening the remaining pages for behavioral rather than merely model-name patterns.

OUTCOME
No verified Chinese/second-actor swarm artifact in this bounded12-file source review. The useful positive result is identification of a large, previously underexamined mixed-language HTML corpus and several implemented agent-client demos. Those facts should expand coverage without turning agent-themed UIs into actor evidence.

ARTIFACTS
Original files: pastebins/data/html.cafe/<ID>.html
Screen: investigation/china/013-html-cafe-language-candidates.json — INTERNAL; may contain sensitive snippets.
Detailed source-review extraction: investigation/china/013-html-cafe-sample-review-internal.json — INTERNAL; do not publish raw.
Safe sample hashes/structural metadata: investigation/china/013-html-cafe-sample-hashes.json