{
 "site": "javis.im",
 "kind": "analysis-index",
 "generated_at": "2026-10-27T20:11:25+08:00",
 "count": 19,
 "articles": [
  {
   "slug": "hermes-agent",
   "published": "2026-10-27",
   "title_zh": "7 个月只有 137★，再 8 个月冲到 257,319★：Hermes Agent 与 OpenClaw，两条都成了，钱的走法相反",
   "title_en": "137★ for 7 months, then 257,319★ in 8 more: Hermes Agent and OpenClaw — both made it, with the money flowing in opposite directions",
   "summary_zh": "对比横评拆解：取证日 2026-10-27，同一个爆款赛道里两条都跑通的路径。Hermes Agent 建仓 2025-07-22，前 7 个月只有 137★（2026-02-25 快照），官网 03-05 首次被 Wayback 收录 + 桌面版 + 一键安装把代码存量兑换成关注，7 个多月冲到 250,484★，取证日 API 257,319★ / 页面 257,321★、59,466 commits、401 贡献者；对照方 OpenClaw（第 18 篇读数）393,064★、9 周 +237,633★。两条曲线最终都收敛到日均约 260★ 的存量速度。八格总表每格双方都有出处：星标双口径、爆发曲线、提交与集中度（top1 47.7% vs 58.9%）、issue 关闭率 53.7% vs 91.7%、渠道 7 vs 29、资金（$90M B 轮 @$1.5B、累计 $158M，vs 501(c)(3) 基金会零收入）、风险事件（CVE×2 CVSS 7.3 vs CVE-2026-33579 的 9.4 提权）、HN 声量结构。三条差异轴：钱的走法相反（公司靠 Portal 订阅 $0/$20/$100/$200，基金会三处明写无订阅、无托管、无 token）；HN 高分的叙事主体不同（项目被讨论 vs 用户被第三方计费）；治理动作不同（issue #10232 28 秒内 4 次封禁 + 标题改成句点 vs 把事故写成 10 人招聘、LTS 与第三方审计）。假官方域名战两边都有冒名站，Hermes 侧的冒名站还在收钱（FlyHermes Pro $49/mo）。口径三档分标：可第三方复核 / 官方自估无审计（2,400 万克隆、约 2.5% 全球 token 用量、$36M→$100M 年化）标【存疑】/ 粉丝站与指控方营销博客不作官方口径。结论：尖峰决定不了成败，平台期的性质才决定；拆任何一条成功路径先问三句——钱从哪来、头条在讲事件还是在讲产品、事故之后组织变了没有。",
   "summary_en": "Comparative head-to-head teardown: evidence day 2026-10-27, two breakout paths in the same track that both made it. Hermes Agent was created 2025-07-22 and sat at 137★ after its first 7 months (2026-02-25 snapshot); the official site first archived by Wayback on 03-05, plus a desktop build and a one-line installer, converted that code stock into attention — 250,484★ within 7-plus months, reading 257,319★ via API / 257,321★ on the page with 59,466 commits and 401 contributors on evidence day. The comparison side, OpenClaw (readings from piece #18), is at 393,064★ with +237,633★ in 9 weeks. Both curves end up converging on the same stock-growth speed of roughly 260 stars a day. The eight-cell table gives both sides a source in every cell: dual star readings, breakout curve, commits and concentration (top author 47.7% vs 58.9%), issue close rate 53.7% vs 91.7%, channels 7 vs 29, money ($90M Series B at $1.5B and $158M raised vs a zero-revenue 501(c)(3) foundation), risk events (two CVEs at CVSS 7.3 vs CVE-2026-33579 at 9.4), and HN score structure. Three difference axes: money flows in opposite directions (a company monetizing Portal subscriptions at $0/$20/$100/$200 vs a foundation stating no subscription, no hosted tier, no token in three places); HN's top stories have different narrative subjects (the project being discussed vs users being billed by third parties); and governance moves differ (issue #10232 getting 4 blocks in 28 seconds and a title renamed to a period vs turning an incident into 10 hires, LTS and a third-party audit). Both sides face impersonation domains, and the Hermes-side fake one still charges money (FlyHermes Pro at $49/mo). Figures are tiered three ways: third-party reproducible / official self-report with no audit (24 million clones, about 2.5% of global AI token usage, $36M to $100M annualized) tagged [Unverified] / fan and accuser marketing sites never treated as official. Conclusion: the spike does not decide success, the nature of the plateau does — and any success path is worth three questions: where does the money come from, is the headline about an event or the product, and did the organization change after the incident.",
   "url_zh": "https://javis.im/analysis/hermes-agent/",
   "url_en": "https://javis.im/analysis/hermes-agent/index-en.html"
  },
  {
   "slug": "openclaw",
   "published": "2026-10-26",
   "title_zh": "零收入的 393,064★：OpenClaw 把钱挡在门外，六大金主反而排队打钱",
   "title_en": "A $0-Revenue 393,064★: OpenClaw Shut the Money Out — and Six Deep-Pocketed Donors Queued Up Anyway",
   "summary_zh": "商业画布拆解：取证日 2026-10-26，OpenClaw 393,064★ / 119,647 commits / 271 release / 383 贡献者，建仓 11 个月——9 周冲 +237,633★（日均约 1,585 星）随后 6 个月只 +47,347★（日均约 259 星）。第四格收入是空的：官网、README、页脚三处明写无付费版、无托管、无 token，钱反而从六个方向进来——501(c)(3) 基金会六大机构捐助方（Microsoft、NVIDIA、OpenAI、Red Hat、密歇根大学、Offline Holdings）+ GitHub Sponsors 329 位；成本格=10 人全职 + 8 个在招岗位拖着全部吞吐（top1 独占 67,083 提交=56.1%、top10 占 89%、7 个 bot 账号）；生态壁垒=29 个渠道、65 个周边项目、56,178 已关闭 issue、117,531 已关闭 PR；关键活动=44 天 30 版 + LTS + Trail of Bits 第三方审计 + Enterprise 接驳；代价线=CVE-2026-33579 的 CVSS 9.4 提权与「太创始人驱动」复盘直接换来 10 人全职团队，Clawd→Moltbot→OpenClaw 三连改名沉淀为品牌资产。口径三档分标（可第三方复核 / 官方自报无审计 / 冒名站数字冲突不采信）。结论：零收入是最强差异化也是最大单点风险，壁垒不在代码里——MIT 人人可 fork，fork 不走生态、issue 历史与捐助方背书。",
   "summary_en": "Business model canvas teardown: on evidence day 2026-10-26 OpenClaw reads 393,064★ / 119,647 commits / 271 releases / 383 contributors at 11 months old — +237,633★ in the first 9 weeks (about 1,585 stars a day), then only +47,347★ over the next 6 months (about 259 a day). Cell four, revenue, is empty: the homepage, the README and the footer all state no paid tier, no hosted service, no token — and money queues in from six directions anyway: six institutional donors to the 501(c)(3) foundation (Microsoft, NVIDIA, OpenAI, Red Hat, University of Michigan, Offline Holdings) plus 329 GitHub Sponsors. Cost and organization = 10 full-time staff and 8 open roles carrying the whole throughput (top author 67,083 commits = 56.1%, top 10 = 89%, 7 bot accounts); ecosystem moat = 29 channels, 65 satellite projects, 56,178 closed issues, 117,531 closed PRs; key activities = 30 releases in 44 days, LTS, a Trail of Bits third-party audit and the Enterprise track; the cost line = CVE-2026-33579 (CVSS 9.4 privilege escalation) and the \"too founder-driven\" post-mortem directly producing the 10-person team, while the Clawd to Moltbot to OpenClaw renaming sequence was written up as brand asset. Figures are sorted into three tiers (third-party reproducible / official self-report with no audit / impersonation-site numbers rejected as conflicting). Conclusion: zero revenue is both the strongest differentiation and the biggest single point of failure, and the moat is not in the code — MIT lets anyone fork OpenClaw, but a fork does not inherit the ecosystem, the issue history or the donors' endorsement.",
   "url_zh": "https://javis.im/analysis/openclaw/",
   "url_en": "https://javis.im/analysis/openclaw/index-en.html"
  },
  {
   "slug": "opencode",
   "published": "2026-10-25",
   "title_zh": "9 个月 40 倍、星标只涨 5 倍：opencode 221,369★ 背后的三本账，只有两本能对上",
   "title_en": "40× in 9 Months, Stars Only 5×: The Three Ledgers Behind opencode's 221,369★, Only Two of Which Reconcile",
   "summary_zh": "增长拆解：opencode 18 个月从 206★ 到 221,369★（取证日仍挂周榜、单周 +6,857★），官网宣称月活 9 个月 40 倍（400K→16M）而同期星标宣称只有 5.07 倍；四个大数字里 stars 221,369 与 20,893 commits 可第三方复现，950 contributors 卡在实测两套口径 455（默认分支注册）与 1,069（含匿名）之间、官网 2026-04-01 页内载荷写的是 453，16M monthly devs 无任何第三方口径（npm 月窗口 CLI 合计 11.7M、落后 @openai/codex 7.2 倍，且同一周官方还在用 millions of daily users 讲日活）；渠道拆解显示 HN 八条头条里三条是安全事故或法务冲突（泄露披露 975 分、Anthropic 法务 PR 19 分钟合并、CVE-2026-22812 的 3 处静默修复 + 8 项未修），一条是要求别再用它的 420 分长文；供给侧 173 天发 100 版、约每 1.7 天一版，贡献榜第 2 名是 opencode-agent[bot]、前 10 人占前 100 名提交的 83.1%；主体 2026-01-02 前后从 sst 迁到 anomalyco，钱线三张图并排：Zen 自称 not for profit（充 $20 收 $1.23 手续费）、Enterprise 自带 LLM 网关不收 token 费走 SOC 2 Type 2、GitHub 全免费 MIT 获客。结论三本账：star 账可复现、npm 账换了尺子（位次可信、绝对值不可信）、950 与 16M 账不可复现——标存疑并把两套口径摆出来，不替它补算法。",
   "summary_en": "Growth teardown: opencode went from 206★ to 221,369★ in 18 months (still on the weekly trending board on evidence day, +6,857★ in one week), while the site claims monthly active devs grew 40× in nine months (400K to 16M) against only 5.07× on stars in the same window. Of the four headline figures, 221,369 stars and 20,893 commits reproduce against third parties, but 950 contributors sits between two measured bases — 455 registered identities on the default branch and 1,069 including anonymous — while the site's own 2026-04-01 page payload reads 453; 16M monthly devs has no third-party ruler at all (npm CLI total 11.7M for the month window, 7.2× behind @openai/codex, and the official 2.0 note was still talking about \"millions of daily users\" the same week). The channel breakdown shows three of eight front-page HN stories are security incidents or legal conflicts — the breach disclosure at 975 points, Anthropic's legal PR merged in 19 minutes, CVE-2026-22812 with 3 silent fixes and 8 unfixed — plus a 420-point \"stop using it\" essay. On the supply side: 100 releases in 173 days (about one every 1.7 days), opencode-agent[bot] second on the contributor board, and the top 10 authors holding 83.1% of the top-100 commit count. The entity moved from sst to anomalyco around 2026-01-02, and three money maps sit side by side: Zen calls itself not for profit ($1.23 fee on a $20 top-up), Enterprise charges nothing for tokens behind your own LLM gateway and points at SOC 2 Type 2, and GitHub is free under MIT. The conclusion reads three ledgers: the star ledger reconciles, the npm ledger uses a different ruler (rank and order of magnitude credible, absolute value not), and the 950 / 16M ledger cannot be reproduced — marked doubtful with both bases laid out instead of an invented formula.",
   "url_zh": "https://javis.im/analysis/opencode/",
   "url_en": "https://javis.im/analysis/opencode/index-en.html"
  },
  {
   "slug": "context-mode",
   "published": "2026-10-24",
   "title_zh": "Issues 关门那一周，main 已经静默四个月：把 context-mode 的 599 个文件、4 篇 ADR 和「707.5k+ users」徽章摊开",
   "title_en": "Issues Closed for a Week, main Quiet for Four Months: Laying Out context-mode's 599 Files, 4 ADRs and the \"707.5k+ users\" Badge",
   "summary_zh": "技术栈考古拆解：context-mode 8 个月从 0 到 32,008 星（5 周 +8,981 / +39%，10-10→10-20 十天 +5,085），热度最高的同一个月治理收口——Issues 与 PR 在 10-10 至 10-17 之间双双关停（issue #482/#981 走 API 直接 410、open_issues_count 停在 356、文档死链遍布），main 的人类提交停在 2026-06-29、此后 184 个提交全是机器人改安装量徽章；机制三层全公开可核：子进程沙箱（README 自认 not a full OS sandbox、堵 #852 逃逸向量）、SQLite FTS5 + BM25 + 双路 Reciprocal Rank Fusion、5 类 hook + ≤2 KB 分层快照与 19 条 Session Guide（流程图写 15 categories）；口径三处对账 = 98% 是 ctx_execute_file 子项小计、整份基准 96%、21 个场景最差 13%，「707.5k+ users」是 npm 643.6k 与 marketplace 63.9k 两个累计下载数相加，而月下载已从 122,969 退到 87,392，ADR-0004 自认状态栏 56% 到 95.4% 差 49 个百分点；许可与商业同步转向 = 2026-03-03 MIT→ELv2 封死托管 SaaS 路线、同期 Free / $20 / $49 / Team / Enterprise 四档价目推进；质疑面（注入判罚 #911、Windows 挂起、孤儿进程等六条）原页取证日 404，全按评测站转引标存疑，带广告位的评测站只当二手引源。",
   "summary_en": "Technical archaeology teardown: context-mode went from 0 to 32,008 stars in 8 months (+8,981 / +39% in five weeks, +5,085 in the ten days from 10-10 to 10-20), and governance closed in that same month - Issues and PRs were both shut down between 10-10 and 10-17 (issues #482/#981 return API 410, open_issues_count frozen at 356, dead doc links everywhere), the last human commit on main sits at 2026-06-29 and all 184 commits after it are bots refreshing install badges; three layers of mechanism are public and checkable: a subprocess sandbox (the README admits it is 'not a full OS sandbox' and closes the #852 escape vector), SQLite FTS5 + BM25 with two-path Reciprocal Rank Fusion, and 5 hook types with a <=2 KB tiered snapshot plus a 19-item Session Guide (the flow chart says 15 categories); three numbers audited - 98% is the subtotal of the ctx_execute_file line while the full benchmark reads 96% and the worst of 21 scenarios saves 13%, '707.5k+ users' adds npm 643.6k to marketplace 63.9k even though monthly downloads fell from 122,969 to 87,392, and ADR-0004 admits the status bar drifted from 56% to 95.4%, a 49-point swing; licence and business moved together - 2026-03-03 MIT to ELv2 closed the hosted-SaaS route while Free / $20 / $49 / Team / Enterprise tiers advanced; the challenge side (the #911 injection verdict, Windows hangs, orphaned processes - six items) has 404 pages at capture date, all treated as second-hand citations marked doubtful, and the ad-bearing review sites are used only as second-hand sources.",
   "url_zh": "https://javis.im/analysis/context-mode/",
   "url_en": "https://javis.im/analysis/context-mode/index-en.html"
  },
  {
   "slug": "linear",
   "published": "2026-10-23",
   "title_zh": "一年估值翻倍、零新钱入账：Linear 的商业画布上，写着一次没公告的 25% 涨价和一句「Issue tracking is dead」",
   "title_en": "Valuation Doubled in a Year with Zero New Money: Linear's Business Model Canvas Holds an Unannounced 25% Price Hike and a Sentence Saying \"Issue tracking is dead\"",
   "summary_zh": "商业画布拆解：Linear 一年估值 $1.25B → $2.5B 翻倍而公司账上零新钱——$99M 员工 tender 全部流向员工与前员工（Dealroom：money went to employees rather than balance sheet）；官方口径 $100M ARR、40,000 家付费公司、NRR 177%、付费用户 127,000；静默涨价被 Wayback 双快照抓出：2025-10-07 Basic $8 / Business $14 → 11-13 $10 / $16（+25% / +14.3%，无公告、无生效日期，比价站判定窗口与本机快照差一天）；CEO《Issue tracking is dead》与 5 个月后 now 页 95% 付费工作区接 agent 并列陈列、不连成一条曲线；渠道格近乎空白（二手转述 $36,000 营销费、1 名销售）；Atlassian FY26 订阅 ARR $6,606M ÷ $100M ≈ 66×，同品类一年内 Height 关店、Linear 翻倍；3-24 variable shadowing 权限事故约 1 小时、约 48 小时后才发邮件；员工数四个读数互斥、不采单值。",
   "summary_en": "Business model canvas teardown: Linear doubled its valuation from $1.25B to $2.5B in a year with zero new money on its books — the $99M employee tender went entirely to employees and former employees (Dealroom: money went to employees rather than balance sheet); official figures claim $100M ARR, 40,000 paying companies, NRR 177% and 127,000 paying users; a silent price hike was caught by two Wayback snapshots: 2025-10-07 Basic $8 / Business $14 → 11-13 $10 / $16 (+25% / +14.3%, no announcement, no effective date, comparison-site window one day off our snapshot); the CEO's \"Issue tracking is dead\" sits beside the now page's 95% of paying workspaces on agents five months later — listed side by side, never merged into one curve; the channel block is nearly empty (second-hand reports of a $36,000 marketing spend and a single salesperson); Atlassian FY26 subscription ARR $6,606M ÷ $100M ≈ 66×, with Height shutting down and Linear doubling inside the same year; the Mar 24 variable-shadowing permission incident lasted about an hour and the email followed roughly 48 hours later; four mutually exclusive employee-count readings, no single value adopted.",
   "url_zh": "https://javis.im/analysis/linear/",
   "url_en": "https://javis.im/analysis/linear/index-en.html"
  },
  {
   "slug": "claude-plugins",
   "published": "2026-10-22",
   "title_zh": "11 个月 41 → 43,746 星：Anthropic 官方插件仓的 4,630 次提交里，73.8% 是机器人写的",
   "title_en": "41 to 43,746 Stars in 11 Months: 73.8% of the 4,630 Commits in Anthropic's Official Plugin Repo Were Written by Bots",
   "summary_zh": "增长拆解：Anthropic 官方插件仓 11 个月从 41 星涨到 43,746 星、最近一周 +5,703；4,630 次提交里 github-actions[bot] 占 3,417 次 = 73.8%，5,516 个 PR 里 3,721 个 = 67.5% 是 bump，releases 与 tags 双 0；marketplace.json 315 条中 262 条指向外部仓 = 83% 是指针不是代码；同一件事有 315 / 341 / 2,000+ 三个对外数字，媒体写 launched May 22 而建仓是 2025-11-20；仓本身零收入，母厂 Series H $65B、$965B post-money，Claude Code run-rate $2.5B；提交门设在付费计划之后，1,045 个 open issue 里「已发布却不见」三条同型复发。",
   "summary_en": "Growth teardown: Anthropic's official plugin repository went from 41 to 43,746 stars in 11 months, +5,703 in the last week; of 4,630 commits, github-actions[bot] accounts for 3,417 = 73.8%, and 3,721 of 5,516 PRs = 67.5% are bumps, while releases and tags are both empty; 262 of the 315 entries in marketplace.json point at external repositories, so 83% are pointers rather than code; the same thing carries three public numbers, 315 / 341 / 2,000+, and the press says launched May 22 while the repository was created on 2025-11-20; the repository itself earns nothing, the parent raised $65B in Series H at $965B post-money, and Claude Code's run-rate is $2.5B; the submission gate sits behind a paid plan, and 1,045 open issues include three recurrences of published but not visible.",
   "url_zh": "https://javis.im/analysis/claude-plugins/",
   "url_en": "https://javis.im/analysis/claude-plugins/index-en.html"
  },
  {
   "slug": "voicestudio",
   "published": "2026-10-22",
   "title_zh": "6 个月 80,473 星、HN 只有 6 分：这个「免费 ElevenLabs」的默认引擎，权重写着 CC-BY-NC",
   "title_en": "80,473 Stars in Six Months, Six Points on HN: The Default Engine of This \"Free ElevenLabs\" Ships With CC-BY-NC Weights",
   "summary_zh": "对比横评：官网自建四列比价表逐行对账（646 语言我逐行数成立，ElevenLabs「74」官方只写 70+）；价格 /bin/bash/9/99 对 →90；646 语言里 265 个训练材料不足 10 小时；三层许可里应用 AGPL-3.0 商用免授权、默认引擎 k2-fsa/OmniVoice 权重却是 CC-BY-NC，HF API 连许可徽章都不出；6 个月 80,473 星、HN 只有 6 分 0 评论；Pro 9/年闭源、Cloud 明写 Nothing is live yet。",
   "summary_en": "Head-to-head comparison: the site's own four-column table checked row by row (646 languages verified, ElevenLabs' 74 officially only 70+); /bin/bash/9/99 against  to 90; 265 of the 646 languages have under ten hours of training material; three licence layers where the app is AGPL-3.0 free for commercial use but the default k2-fsa/OmniVoice weights are CC-BY-NC and the HF API shows no badge; 80,473 stars in six months against 6 points on HN; Pro is closed-source at 9 a year and Cloud says Nothing is live yet.",
   "url_zh": "https://javis.im/analysis/voicestudio/",
   "url_en": "https://javis.im/analysis/voicestudio/index-en.html"
  },
  {
   "slug": "braintrust",
   "published": "2026-10-21",
   "title_zh": "主仓 29 星、月下载 689 万次：$80M B 轮、第三方记作 $800M 的 Braintrust，开源外壳里包着自研 Rust 数据库，旗舰 proxy 还被自己标了 deprecated",
   "title_en": "29 Stars on the Main Repo, 6.89M Monthly Downloads: Braintrust (a $80M Series B, third parties say $800M) Wraps an Open-Source Shell Around a Homegrown Rust Database — and Just Flagged Its Own Flagship Proxy as Deprecated",
   "summary_zh": "npm+PyPI 月下载 1,346 万次而主 SDK 仓只有 29 星：108 个开源仓外圈、控制面/数据面分离与自研 Brainstore 内核、proxy 弃用迁移考古，附 2026-05 AWS 泄密与四段资本口径对账。",
   "summary_en": "13.46M monthly npm+PyPI downloads against 29 stars on the main SDK repo: an archaeology of the 108-repo open outer ring, the control/data-plane split and homegrown Brainstore core, and the proxy deprecation — plus the May 2026 AWS breach and four funding calibers reconciled.",
   "url_zh": "https://javis.im/analysis/braintrust/",
   "url_en": "https://javis.im/analysis/braintrust/index-en.html"
  },
  {
   "slug": "calcom",
   "published": "2026-10-21",
   "title_zh": "五年开源、49,185 星：Cal.com 把生产代码搬进私库，公仓只剩自认「别上生产」的 Cal.diy",
   "title_en": "Five Years of Open Source, 49,185 Stars: Cal.com Moved Production Code to a Private Repo, and the Public One Is Cal.diy — Which Says Do Not Put Me in Production",
   "summary_zh": "49,185 星的开源主仓 2026-04 改名 Cal.diy 并删掉 12 类企业功能（含整个 API v1）：AGPLv3+ee 商业许可到 MIT 社区版 + 闭源商业版的许可考古、722 篇博客的程序化 SEO、四档价目与 Calendly 征战页、融资停在 2022-04 的 $32.4M 与收入 $5.1M 口径对账。",
   "summary_en": "The 49,185-star open repo was renamed Cal.diy in April 2026 with 12 enterprise feature categories removed (including the whole API v1): an archaeology from AGPLv3 plus a commercial ee/ license to an MIT community edition plus a closed commercial edition, 722 blog posts of programmatic SEO, four price tiers and a Calendly conquest page, funding frozen at $32.4M since 2022-04, and a $5.1M revenue basis reconciled.",
   "url_zh": "https://javis.im/analysis/calcom/",
   "url_en": "https://javis.im/analysis/calcom/index-en.html"
  },
  {
   "slug": "deepinfer",
   "published": "2026-10-21",
   "title_zh": "开仓 5 天、4 个 commit：DeepSeek 把「单集群每天 155 万亿 token」的推理引擎开源了，缺件也写在同一份 README 里",
   "title_en": "Five Days Old, Four Commits: DeepSeek Open-Sourced the Inference Engine Behind 155 Trillion Tokens a Day — And Listed the Missing Parts in the Same README",
   "summary_zh": "技术栈考古：内核层 6 个自研算子库与 93 个 TileLang 文件、引擎层 DSpark/SWA bounded replay/batch-invariant、系统层 VE·P·D 分离与四类负载均衡只在文档里，生产口径单集群 155 万亿 token/天、8/GPU·时；血统 README 自认 built primarily on vLLM，开仓形态 4 commit / 3 贡献者 / 0 release / issues 关闭，星数三时点 14,754→15,121、HN 2,200 分。",
   "summary_en": "Tech-stack archaeology: six self-built operator libraries and 93 TileLang files at the kernel layer, DSpark/SWA bounded replay/batch-invariant at the engine layer, VE·P·D disaggregation and four load balancers that exist only in the docs, production basis of 155 trillion tokens a day at US$38 per GPU-hour; the README admits it was built primarily on vLLM, the repo shape is 4 commits / 3 contributors / 0 releases / issues off, stars moved 14,754 to 15,121 in 46 minutes, HN 2,200 points.",
   "url_zh": "https://javis.im/analysis/deepinfer/",
   "url_en": "https://javis.im/analysis/deepinfer/index-en.html"
  },
  {
   "slug": "n8n",
   "published": "2026-10-20",
   "title_zh": "自动化老树开新花：n8n 手握 208,124 星、7 个月从 US$2.5B 翻到 US$5.2B，官方文档却写着「我们不称自己为开源」",
   "title_en": "Old Tree, New Blossoms: n8n Holds 208,124 Stars, Doubled the Company to US$5.2B in Seven Months — While Its Own Docs Say We Do Not Call Ourselves Open Source",
   "summary_zh": "208,124 星的 fair-code 自动化平台：一次执行=整条流程的计费、170 万月活搭建者与 1,400 家企业、SAP 战投把公司翻到 US$5.2B，而官方文档写明「我们不称自己为开源」——四个集成数口径同日并存的商业画布拆解。",
   "summary_en": "A fair-code automation platform with 208,124 stars: billing on one full workflow run, 1.7M monthly builders and 1,400 enterprises, an SAP investment that doubled the company to US$5.2B — and official docs stating 'we do not call ourselves open source'. A business-canvas teardown where four integration counts coexist on the same day.",
   "url_zh": "https://javis.im/analysis/n8n/",
   "url_en": "https://javis.im/analysis/n8n/index-en.html"
  },
  {
   "slug": "zed",
   "published": "2026-10-20",
   "title_zh": "被 GitHub 关停的 Atom 团队回来了：Zed 91,923 星、Stack Overflow 2026 官方口径 11.4%，而它刚把自己仓库的 PR 关掉",
   "title_en": "The Team Behind Atom, Shut Down by GitHub, Is Back: Zed at 91,923 Stars and 11.4% in the Stack Overflow 2026 Data — And It Just Turned Off PRs on Its Own Repository",
   "summary_zh": "对比横评两组数：星/issue 口径对账（91,923 vs VS Code 193,972，3,104=2,316+788）+ Stack Overflow 2026 官方常用率 11.4% 第 11 位；Rust 97.69% 的重写动机、性能表系当事人自写、2M B 轮与 /bin/bash/0/0、关掉自家 PR 的 Delta，风险清单 7 条，34 条来源四色标注。",
   "summary_en": "A two-set head-to-head: star/issue basis reconciliation (91,923 vs VS Code 193,972; 3,104 = 2,316 + 788) plus Stack Overflow 2026 official usage of 11.4% in 11th place; the 97.69% Rust rewrite motive, a performance table written by the contestant itself, 2M Series B and /bin/bash/0/0 tiers, Delta turning off its own PRs; 7-item risk list, 34 sources in four confidence colors.",
   "url_zh": "https://javis.im/analysis/zed/",
   "url_en": "https://javis.im/analysis/zed/index-en.html"
  },
  {
   "slug": "langchain",
   "published": "2026-10-19",
   "title_zh": "148,045 星、月下载 1.70 亿：LangChain 三年把自己写的高层抽象废掉两轮，1.0 留下的那个还搬了两次家",
   "title_en": "148,045 Stars, 170M Downloads a Month: LangChain Scrapped Its Own High-Level Abstraction Twice in Three Years — and the One It Kept Has Moved House Twice",
   "summary_zh": "技术栈考古拆 Chain→LCEL→LangGraph→1.0 四代迁移：官方带日期家谱、CEO 在 480 分 HN 帖认账「当初抽象过了头」、同一 agent 抽象在 langchain/langgraph 之间搬了两次家；星数/issue/下载/发布日四组口径逐一对账，25 条来源四色标注。",
   "summary_en": "Tech-stack archaeology of the Chain to LCEL to LangGraph to 1.0 migration: the official dated family tree, the CEO's confession inside a 480-point HN thread, one agent abstraction that moved house twice between langchain and langgraph; four sets of numbers reconciled (stars, issues, downloads, release dates) — 25 sources, four-color tags.",
   "url_zh": "https://javis.im/analysis/langchain/",
   "url_en": "https://javis.im/analysis/langchain/index-en.html"
  },
  {
   "slug": "plausible",
   "published": "2026-10-19",
   "title_zh": "773 字节脚本、24 小时销毁的盐、10 个人服务 21,000 家付费：审计 Plausible 的「免横幅」承诺",
   "title_en": "A 773-Byte Script, a Salt Deleted Every 24 Hours, 10 People Serving 21,000 Paying Customers: Auditing Plausible's \"No Banner Needed\" Promise",
   "summary_zh": "风险审计四层：脚本层实测（gzip 773 字节、源码零 Cookie API）、管道层对账（salt 24 小时销毁、六年公式未变）、法律层（奥地利两裁 GA 违法 vs 欧盟 Art 88a 仍是提案）、商业层（零外部投资、10 人、21,000 家付费、star 排第二档）；33 条来源四色标注。",
   "summary_en": "Four-pass risk audit: script layer measured (773 bytes gzipped, zero Cookie APIs in source), pipeline reconciled (salt deleted every 24h, formula unchanged for six years), legal layer (two Austrian rulings against GA vs EU Art 88a still a proposal), business layer (zero outside investment, 10 people, 21,000 paying customers, mid-pack stars); 33 sources in four confidence colors.",
   "url_zh": "https://javis.im/analysis/plausible/",
   "url_en": "https://javis.im/analysis/plausible/index-en.html"
  },
  {
   "slug": "supabase",
   "published": "2026-10-19",
   "title_zh": "111,751 星只是副产品：YC 最新批次 55% 用 Supabase、周下载 1,990 万，官方口径 70% 的新数据库由 AI agent 创建",
   "title_en": "111,751 Stars Are Just a Side Effect: 55% of the Latest YC Batch Builds on Supabase, 19.9M Downloads a Week — and by the Company's Own Count, 70% of New Databases Come from AI Agents",
   "summary_zh": "增长拆解四条曲线：星与下载（npm 17 个月约 33 倍）、YC 采用率 40%→55%、七轮 95M 再加 2026-10 的 50M、需求侧 70% 换挡到 AI agent；五组口径对账 + 风险清单，34 条来源四色标注。",
   "summary_en": "Growth teardown across four curves: stars and downloads (npm roughly 33x in 17 months), YC adoption 40% to 55%, seven rounds totaling 95M plus 50M in 2026-10, and 70% of demand switching to AI agents; five reconciliations plus a risk list, 34 sources in four confidence colors.",
   "url_zh": "https://javis.im/analysis/supabase/",
   "url_en": "https://javis.im/analysis/supabase/index-en.html"
  },
  {
   "slug": "bun",
   "published": "2026-10-16",
   "title_zh": "48,243 vs 25,181 req/s：96.2k 星的 Bun 横评 Node 与 Deno，$0 收入时被 Anthropic 买走",
   "title_en": "48,243 vs 25,181 req/s: The 96.2k-Star Bun Benchmarks Against Node and Deno — and Sold to Anthropic at $0 Revenue",
   "summary_zh": "对比横评框架拆 Bun：官方四面板 vs 第三方对齐口径（DB 上场三家收敛 ~12,000 req/s）、$0 收入被 Anthropic 收购的账本、Zig→Rust 重写导火索与无 LTS/兼容长尾扣分项，14 条来源四色标注。",
   "summary_en": "Bun head-to-head: four official panels vs aligned third-party measurements (three runtimes converge near 12,000 req/s behind a database), the /bin/bash-revenue acquisition ledger, the Zig-to-Rust rewrite trigger, and the no-LTS/compatibility-tail deductions — 14 sources, four-color tags.",
   "url_zh": "https://javis.im/analysis/bun/",
   "url_en": "https://javis.im/analysis/bun/index-en.html"
  },
  {
   "slug": "litellm",
   "published": "2026-10-14",
   "title_zh": "不抽 token 一分钱：61.2k 星的 LiteLLM，怎么把 $0 开源货卖成企业年单",
   "title_en": "Zero Per-Token Cut: How LiteLLM Turns a 61.2k-Star Open-Source Gateway Into Annual Enterprise Contracts",
   "summary_zh": "商业画布九格拆 LiteLLM：$0 永久免费的开源核心 + never per token 的企业年单，61.2k 星就是它的货架；附 CVE 全记录、ARR 冲突数据与同页数字口径差。",
   "summary_en": "LiteLLM's nine-block business canvas: a free-forever open-source core plus never-per-token enterprise annual contracts, with 61.2k GitHub stars as the shelf — plus the full CVE record, conflicting ARR figures, and on-page number discrepancies.",
   "url_zh": "https://javis.im/analysis/litellm/",
   "url_en": "https://javis.im/analysis/litellm/index-en.html"
  },
  {
   "slug": "openrouter",
   "published": "2026-10-14",
   "title_zh": "600 万亿 token，5 个月翻 6 倍：OpenRouter 如何把「中间商」做成 $1.3B 生意",
   "title_en": "600 Trillion Tokens, 6x in 5 Months: How OpenRouter Turned 'The Middleman' Into a $1.3B Business",
   "summary_zh": "拆解 OpenRouter：不拥有任何模型，靠兼容性、免费层、开放数据飞轮和 agent 浪潮四台发动机，把过路费生意做到 13 亿美元估值。",
   "summary_en": "An OpenRouter teardown: owning zero models, it built a .3B toll business on four engines — compatibility, a free tier, an open-data flywheel, and the agent wave.",
   "url_zh": "https://javis.im/analysis/openrouter/",
   "url_en": "https://javis.im/analysis/openrouter/index-en.html"
  },
  {
   "slug": "javis-im",
   "published": "2026-10-13",
   "title_zh": "零服务器、零数据库、零追踪：一个私人 AI 站点的 4 层防线",
   "title_en": "Zero Servers, Zero Databases, Zero Trackers: The 4-Layer Defense of a Personal AI Site",
   "summary_zh": "拆解 javis.im：三个 Cloudflare 组件、零服务器、零追踪，把安全声明全部做成可当场验证的行为。",
   "summary_en": "Tearing down javis.im: three Cloudflare components, zero servers, zero trackers — every security claim executed as verifiable behavior.",
   "url_zh": "https://javis.im/analysis/javis-im/",
   "url_en": "https://javis.im/analysis/javis-im/index-en.html"
  }
 ]
}