三人、七台电脑。1 到 3 号机跑编码代理,4、5 号机审查,6 号机做 Windows 验证,7 号机负责合入。11 月 21 日 18:54(Asia/Shanghai),7 号机把当天的文章 squash 进了 master,合入提交是 22091b1。仓库里已经有 source/_posts/order-export-anonymous-can-download.md,也有配图和实验包。对代理来说,发布清单上的「开 PR、合并」已经做完。下一步很容易被写成:确认线上 200。可判断依据会悄悄滑成「master 上已经有这篇,所以它上线了」。
22 日 18:41 再打开预定地址,返回的是 404。首页还挂着 11 月 20 日那篇3 号机把冲突收成己方那一侧,刚合进 main 的验签检查没了。我向这些地址发了带 Cache-Control: no-cache 的请求,响应里 eo-cache-status 是 Cache Miss,Age 是 0。首页的 Last-Modified 仍是 Fri, 20 Nov 2026 11:03:33 GMT,也就是 11 月 20 日 19:03:33;新地址 404 页的时间戳是 19:03:27。ETag 也没有变。这不是浏览器用了旧副本,是源站还在卖上一篇。
从 18:54 合入到这次探测,过了 23 小时 46 分。这篇要核对的是:合进默认分支,和读者已经能打开新地址,是不是同一件事;空提交能不能当叫号铃;仓库里有没有「已部署」的信号,可以替代源站探测。它和6 号机贴了 windows-ok,验证过的那个提交已经不是 HEAD谈的是验证标签贴在过期 SHA 上,和 11 月 18 日修过的尾斜杠路由 404 也不是同一件事——那次是目录首页没有被正确解析,旧文章的带斜杠地址会 404,不带斜杠反而能打开。这次是带斜杠和不带斜杠的新地址都是 404,上一篇的地址仍是 200,源站时间戳没有往后走。今天只谈:merge 完成之后,源站有没有真的换版。
实验在这台机器上对线上站点实测,脚本在本地 git 2.47.3 和 Python 3.13 上跑。我没有在 EdgeOne Makers 控制台里点过重新部署,也没有调用部署钩子,也没有用 Makers API Token 上传构建产物。Windows 没有参与。下面用到的「7 号机」和「代理」是这个发布流程里的角色,不是另一次夹具演戏。
上一篇只用了大约 56 秒,这一篇过了二十多个小时还没有
11 月 20 日那篇的 squash 合入是 aa1dc71,时间 11:02:37 UTC,也就是 19:02:37。线上首页和那篇文章页的 Last-Modified 是 11:03:33 GMT。从合入到源站换版,大约 56 秒。站点还是同一套 EdgeOne Makers。我用腾讯云 TEO 接口读到,www.iotxing.com 和 apex iotxing.com 的主机状态都是 online,zone 没有暂停,免费版套餐状态是 normal。自定义域的 CNAME 仍指向 www.iotxing.com.pages.dnsoe9.com。
11 月 21 日 18:54 合入之后,如果还按前一天的节奏,18:56 左右就该能打开新地址。大约 14 分钟后,19:08 又往 master 推了一个空提交 1b9dbd6。它和 22091b1 的 tree 完全一样,都是 1635b5d,统计是 0 次增加、0 次删除。推送的意图是再叫一次构建。到 22 日 18:41 闸门实测时,新地址仍是 404,Cache Miss 上的源站时间戳仍停在 20 日 19:03。
所以,问题不是「这类托管本来就要等半天」。同一条流水线,前一天是一分钟级,后一天是将近一天还没换版。

这不是缓存没刷新,也不是文章没进仓库
合入提交 22091b1 相对前一天的 aa1dc71,增加了 293 行,文章、六张图和实验包都在树上。origin/master 现在的头是 1b9dbd6。我跑了 shipped_from_git.py:只要默认分支上存在 source/_posts/order-export-anonymous-can-download.md,它就打印 SHIPPED,退出码 0。按这个信号,文章在 18:54 就已经「上线」了。
源站探测给出相反的结论。live_gate.py 会发一个带时间戳查询参数的请求,避免读到边缘节点上的旧对象,然后同时检查三件事:状态码是不是 200,正文里有没有这篇才有的 slug,以及 Last-Modified 是不是晚于合入时间 2026-11-21T10:54:44Z。新地址带尾斜杠和不带尾斜杠都返回 404,正文里没有 slug,时间戳早于合入。首页是 200,但没有新 slug,时间戳同样早于合入。闸门的结论是 NOT_LIVE,退出码 1。作为对照,同一套脚本对上一篇使用 11 月 20 日的合入时间,得到 LIVE:200,正文里有旧 slug,时间戳 19:03:33 晚于当天的合入时间 19:02:37。闸门本身会放行一次真正换过版的发布,它拒绝的是这一次还没换版的发布。
页面响应头是 cache-control: public,max-age=0,must-revalidate。即便边缘节点上有一份旧对象,按这个策略也应该回源再验证。我看到的 Cache Miss 和 Age: 0,再加上未变的 ETag,说明源站给出的就是这一版。把问题写成「清一下 CDN」不会让 404 变成 200。
这一次还有一个可以在仓库里直接复现的问题。合入提交 22091b1 里,11 月 21 日那篇的 description 没有加引号,句子中出现了 auth_declared: admin 这样的英文冒号加空格。Astro 用 js-yaml 解析 front matter 时,会把它当成新的映射项,npm run build 在这里停下。空提交 1b9dbd6 和 22091b1 是同一棵 tree,构建结果也会一样。我没有读到 Makers 的构建日志,所以不能把「源站没换版」的根因写成已经证实就是这一行 YAML。我能确定的是:现在 master 上的这棵树,在这台机器上无法完成和线上相同的 npm run build。只推空提交,修不了这件事。今天的修正是给那一行加上引号,正文一个字都没有改。
空提交不是可靠的叫号铃
空提交这一招并不是从来都不灵。10 月 18 日,有一次合入之后大约 35 分钟,源站还在卖前一天的版本。当时往 master 推了空提交 6986207,tree 同样没有变化,大约 40 秒后新版本上线。那次它起作用,是因为监听还在,只是那一次构建慢了,或者被漏掉了一次。
11 月 21 日的空提交 1b9dbd6 没有让源站往前走。内容提交 22091b1 已经改了 293 行,本来就应该触发一次主分支构建;空提交如果要补救,前提是构建器还在收 master 上的 push。监听如果已经不在了,带正文的 squash 和空的 chore 是同一类事件:对仓库来说都是默认分支上的一次推送,对已经不收事件的构建器来说都是静音。
EdgeOne Makers 的创建部署指南把 Git 自动部署写得很直接:项目和 Git 仓库关联后,系统会自动监听推送并触发部署,默认只有主分支。另外还可以在控制台点重新部署,或者用一条不用额外认证的「部署钩子」URL 发 POST。这些机制都不在仓库里露出记录。我在这个仓库里看不到 webhook,也没有保存过部署钩子,本机环境变量里也没有 Makers API Token。所以我不会把「再推一次空提交」写进发布流程当修复步骤。它只证明一件事:当监听还健康时,再推一次有时能把已经在排队或漏掉的构建叫起来。当监听可能已经断开时,再推一次只会在 git log 里多一行 chore。
仓库里没有「已上线」这盏灯
我用 GitHub API 读了合入提交和空提交上所有看起来像部署的信号。webhook 列表是空数组。deployments 列表是空数组。两次提交的 check-runs 总数都是 0,commit status 的 state 是 pending,但 total_count 也是 0——这不是绿灯,是没有任何状态被写回来。这个仓库也没有 GitHub Pages 站点记录。唯一的 Actions 工作流是 win-lab.yml,和站点构建无关。
也就是说,7 号机如果只盯这个仓库的 Checks、Deployments 和 webhook 列表,它看不见任何失败,也看不见任何成功。合并页看起来是干净的,因为根本没有必须通过的检查。Makers 用 GitHub App 一类的站外监听来收推送,这种安装状态用普通的仓库令牌列不出来。我能用 TEO 只读接口确认域名还在线、zone 没有暂停,但读不到 Makers 的构建队列,DescribeAccelerationDomains 在这把密钥下也没有权限。因此我不能单独把根因写成「本月构建额度用完」「自动部署被关掉」或「GitHub 连接断了」——这三件事都要到控制台才能确认。我能确定的是:从仓库和源站这两侧看,合入事件没有变成一次新的源站版本;而这份已经合入的树,在这台机器上连构建都过不了。
高频 AI 辅助开发里,发布步骤常常被写成清单:开 PR、合并、宣布。合并是仓库里立刻能读到的事实,线上换版是仓库外面的另一次动作。清单如果把第二步省略成「等一会儿」,或者把「等一会儿」的成功条件写成「git log 已经看到合入提交」,代理就会在源站还没动的时候对外说「已经上线」。读者点开 404,下一次就不信发布通知了。内部工单也会被提前关掉。
只看 master 会得到「已上线」,探源站会拒绝宣布

把对照写清楚:
shipped_from_git.py:文件在origin/master(1b9dbd6)上存在,结论SHIPPED,退出码 0。github_signals.py:对22091b1和1b9dbd6,hooks 0、deployments 0、检查 0。live_gate.py对新地址:NOT_LIVE,404,Cache Miss。live_gate.py对首页:NOT_LIVE,200 但未换版。live_gate.py对上一篇、使用 11 月 20 日合入时间:LIVE。
真正的闸门不是「默认分支有没有这篇文章」,而是下面这几条要一起成立:
- 新地址(带尾斜杠和不带)要 200;
- 正文里要有这篇才有的句子、标题或 slug;
- 在 Cache Miss(或等价的回源验证)之后,源站
Last-Modified要晚于合入时间; - 首页也要列出这篇,避免只修了单页、列表仍是旧构建的幻觉。这一次连单页都没有,首页检查看起来像是多余的;把它留在闸门里,是为了下一次只部分换版时还能拦住。
超时之后,流程应该停在「未上线,需要人去控制台看 Git 连接、自动部署开关、构建队列和本月构建额度」,而不是再推一个空提交。
超时取 10 分钟,是因为同一套站点在 11 月 20 日只用了大约 56 秒。把等待窗口放到一分钟级,会把一次正常但稍慢的构建误判成失败;放到几小时,又会把已经断开的监听当成「还在排队」。10 分钟够覆盖一次普通的静态站点构建和边缘生效,却不会让「未上线」在发布清单上过夜。如果 live-gate 在第 2 分钟已经失败,第 3 分钟再探一次是合理的重试;到了第 10 分钟仍然是同一份 20 日的 Last-Modified,再等下去不会多出新信息,只会让错误的「已上线」通知更晚,或者让该升级的问题继续被空提交挡住。

这套闸门不适合什么情况
如果站点不是静态托管,源站根本不回 Last-Modified,就不要照抄时间戳。改用构建号、约定好的 /__version.json,或页面里写死的发布标识,规则仍然是:宣布上线之前,要读到这次发布才有的标记。
如果 CI 或部署钩子会把构建结果写回 GitHub Deployment,仓库里就会多一盏灯。live-gate 仍然值得留着做最终确认,因为 Deployment 成功也可能指错环境,或者指的是预览域而不是自定义域。
如果团队每次都是人在控制台点部署、从不依赖 Git 自动构建,问题会变成「有没有人点」。闸门仍然适用,只是升级对象从「Git 连接」变成「去点一下」。
我没有在 Makers 控制台里点过重新部署,也没有对部署钩子发过 POST,也没有用 CLI 上传 dist。因此,这篇不把「控制台点一下一定能修好」写成已经验证的结果。它是官方文档给出的路径,还需要人去操作。我也没有把构建失败的日志读出来,所以不排除队列卡住或额度用尽,只是这两件事目前都还不能从仓库和源站这两侧证实或排除。
你可以明天抄走的
在 7 号机的发布说明里写明:合入之后不得按 SHA 宣布上线。先跑源站探测:
python3 tools/live_gate.py \
--merged-after 2026-11-21T10:54:44Z \
--needle order-export-anonymous-can-download \
--url https://www.iotxing.com/order-export-anonymous-can-download/ \
--url https://www.iotxing.com/ \
--json-out /tmp/live-gate.json
退出码不是 0,就不要发「已上线」。给一个明确的超时,比如 10 分钟;超时后升级到控制台,去看 Git 仓库连接、生产环境的自动部署开关、构建队列,以及免费版按分钟计的构建额度。不要把再推一个空提交当成修复。
如果希望不打开控制台也能重试,单独建一条 Makers 部署钩子,把它放进密钥库,而不是写进仓库。发布流程在 live-gate 失败、并且已经排除「还在构建」之后,可以对这条 URL 发 POST,然后再探一次源站。钩子本身不再认证,泄露等于把部署权交给任何人,所以不要把它提交到 master。
脚本、AGENTS.md 片段和一份工作流样例在 live-origin-gate-kit.tar.gz。工作流 YAML 只做过结构检查,没有在 GitHub Actions 上跑过。
最小的一步:合入之后,先在本地跑一遍宿主也会跑的 npm run build,再向新地址发一个会打到源站的请求。构建失败,或者线上 404,或者 Last-Modified 仍早于合入时间,就不要宣布。
实际跑过什么
探测时间 2026-11-22T10:41:27Z(18:41:27 Asia/Shanghai)。运行时是 git 2.47.3、Python 3.13、gh,站点响应头里的 Server 是 edgeone makers。
22091b1于 2026-11-21T10:54:44Z 合入,相对aa1dc71增加 293 行。1b9dbd6于 2026-11-21T11:08:48Z 推送,tree 与22091b1同为1635b5d,统计 0/0。shipped_from_git.py对origin/master=1b9dbd6,文件存在,结论SHIPPED,退出码 0。live_gate.py对新地址两个 URL:状态 404,eo-cache-status: Cache Miss,Age: 0,Last-Modified2026-11-20T11:03:27Z,结论NOT_LIVE,退出码 1。首页 200,同一类时间戳 11:03:33Z,没有新 slug,结论NOT_LIVE。上一篇对 2026-11-20T11:02:37Z 的合入时间:200,正文含 slug,时间戳 11:03:33Z,结论LIVE,退出码 0。github_signals.py对两个 SHA:hooks 0,deployments 0,check-runs 0,statuses 总数 0,GitHub Pages 接口 404。- TEO:zone
default-pages-zone未暂停,www.iotxing.com状态 online。没有读到 Makers 构建队列。 - 10 月 18 日的空提交
6986207已在仓库里核对:2026-10-18T11:40:25Z,统计 0/0。它当时大约 40 秒后换版,来自这次站点自己的发布记录;这次没有再复现那一次成功。
对已合入的 22091b1 树运行 npm run build 时,js-yaml 在 11 月 21 日那篇未加引号的 description 上失败,触发点是 auth_declared: admin。给该行加上引号后,正文未改。
没有跑:Makers 控制台操作、部署钩子 POST、使用 API Token 的 CLI 上传、GitHub Actions(工作流只做过结构检查)、分支保护或 ruleset、Windows 或 windows-latest。根因里「额度」「开关」「Git 连接」三项,未在控制台验证。
Three people, seven machines. Machines 1 to 3 run coding agents, 4 and 5 review, 6 does Windows validation, and 7 merges. At 18:54 Asia/Shanghai on 21 November, machine 7 squash-merged that day's post to master as 22091b1. The tree already had source/_posts/order-export-anonymous-can-download.md, plus the diagrams and the lab kit. For the agent, the publish checklist items "open a PR" and "merge" were done. The next item is easy to write as "confirm a live 200". The test then quietly becomes: the file is on master, so the post is live.
At 18:41 on the 22nd, the intended URL still returned 404. The homepage was still headed by the 20 November post, Machine 3 kept our side of the conflict, and the signature check that had just landed on main was gone. I sent those URLs with Cache-Control: no-cache. The responses had eo-cache-status: Cache Miss and Age: 0. The homepage Last-Modified was still Fri, 20 Nov 2026 11:03:33 GMT, which is 19:03:33 on 20 November; the 404 for the new path was timestamped 19:03:27. The ETag had not changed either. This was not a stale browser copy. The origin was still serving yesterday.
From the 18:54 merge to this probe, 23 hours and 46 minutes had passed. This post asks whether a merge to the default branch is the same event as readers being able to open the new URL; whether an empty commit can be used as a pager; and whether the repository exposes any "deployed" signal that can replace an origin probe. It is not the same problem as machine 6 pasted windows-ok, and the commit that was validated is no longer HEAD, which is a validation label stuck to a stale SHA. It is also not the trailing-slash routing 404 fixed on 18 November, when directory indexes were not resolved: older posts 404ed with a slash and opened without one. This time both spellings of the new URL are 404, the previous post is still 200, and the origin timestamp has not moved forward. Today is only about this: after the merge, did the origin actually change over?
The probes were run against the live site from this machine. The scripts ran on local git 2.47.3 and Python 3.13. I did not click Redeploy in the EdgeOne Makers console, I did not POST a deploy hook, and I did not upload a build with a Makers API Token. Windows was not involved. "Machine 7" and "the agent" below are roles in this publish flow, not a second staged play.
The previous post took about 56 seconds; this one had not changed over after more than twenty hours
The 20 November squash merge is aa1dc71, at 11:02:37 UTC, 19:02:37 local. The live homepage and that post's Last-Modified is 11:03:33 GMT. From merge to origin changeover, about 56 seconds. It is the same EdgeOne Makers site. Through the Tencent Cloud TEO API, both www.iotxing.com and apex iotxing.com are online, the zone is not paused, and the free plan resource is normal. The custom-domain CNAME still points at www.iotxing.com.pages.dnsoe9.com.
After the 18:54 merge on the 21st, the same pace would have opened the new URL around 18:56. About fourteen minutes later, at 19:08, an empty commit 1b9dbd6 was pushed to master. It has the same tree as 22091b1, 1635b5d, with 0 additions and 0 deletions. The intent was to ring for another build. At the 18:41 probe on the 22nd, the new URL was still 404, and the origin timestamp on a Cache Miss was still 19:03 on the 20th.
So this is not "this kind of host just takes half a day". On the same pipeline, the previous day was on the order of a minute, and the next day had not changed over after almost a day.

This is not an unflushed cache, and the post is in the repository
Merge commit 22091b1 added 293 lines on top of the previous day's aa1dc71. The post, the six diagrams, and the kit are in the tree. origin/master now points at 1b9dbd6. I ran shipped_from_git.py: if source/_posts/order-export-anonymous-can-download.md exists on the default branch, it prints SHIPPED and exits 0. By that signal, the post had been "live" since 18:54.
The origin probe says the opposite. live_gate.py sends a request with a cache-busting query string so it is less likely to read a stale edge object, and then checks three things at once: whether the status is 200, whether the body contains a slug unique to this post, and whether Last-Modified is at or after the merge time 2026-11-21T10:54:44Z. Both the trailing-slash and slashless new URLs returned 404, with no slug in the body, and a timestamp before the merge. The homepage was 200, but it had no new slug, and the same too-old timestamp. The gate's verdict is NOT_LIVE, exit code 1. As a control, the same script against the previous post, using the 20 November merge time, returns LIVE: 200, the old slug is in the body, and 19:03:33 is after that day's 19:02:37 merge. The gate will pass a release that actually changed over. It is refusing this one, which has not.
The response header is cache-control: public,max-age=0,must-revalidate. Even if an edge node had an old object, that policy should revalidate at the origin. The Cache Miss and Age: 0 I saw, plus an unchanged ETag, mean this is the version the origin is serving. Calling it "flush the CDN" will not turn a 404 into a 200.
There is also a failure that can be reproduced in the repository itself. In merge commit 22091b1, the 21 November post's description was left unquoted, and the sentence contains an ASCII colon-space, auth_declared: admin. Astro parses front matter with js-yaml, which treats that as a new mapping entry, and npm run build stops there. The empty commit 1b9dbd6 has the same tree as 22091b1, so the build result would be the same. I have not read a Makers build log, so I cannot write "the origin did not change over because of this YAML line" as a confirmed root cause. What I can say is that the tree now on master cannot complete the same npm run build the host uses, on this machine. Pushing an empty commit does not fix that. Today's fix is to quote that one line; the body of the 21 November post is unchanged.
An empty commit is not a reliable pager
The empty-commit trick is not something that never works. On 18 October, about 35 minutes after a merge, the origin was still serving the previous day's site. An empty commit 6986207 was pushed to master, also with no tree change, and the new version was live about 40 seconds later. It worked that time because the listener was still there, and that build was only slow, or had been missed once.
The 21 November empty commit 1b9dbd6 did not move the origin forward. The content commit 22091b1 had already changed 293 lines, and should have triggered a production-branch build on its own. An empty commit can only help if the builder is still receiving pushes to master. If the listener is already gone, a squash with a real diff and an empty chore are the same kind of event: a push to the default branch, as far as the repository is concerned, and silence, as far as a builder that is no longer receiving events is concerned.
The EdgeOne Makers guide to triggering deploys is direct: once a project is linked to a Git repository, the system listens for pushes and triggers a deploy, and by default only the production branch is included. You can also click Redeploy in the console, or POST a deploy-hook URL that needs no extra authentication. None of those mechanisms leave a record in the repository. This repository has no webhooks, I have not stored a deploy hook, and this machine's environment has no Makers API Token. So I will not write "push another empty commit" into the publish flow as a fix. It only shows this: when the listener is still healthy, another push can sometimes start a build that was queued or missed. When the listener may already be gone, another push is just one more chore line in git log.
The repository has no "it is live" light
I read every deploy-shaped signal GitHub exposes on the merge commit and on the empty commit. The webhook list is an empty array. The deployments list is an empty array. Both commits have 0 check runs. The commit status state is pending, but total_count is also 0 — that is not a green light, it is no status being written back. The repository has no GitHub Pages site record either. The only Actions workflow is win-lab.yml, which is unrelated to building the site.
So if machine 7 only watches Checks, Deployments, and the webhook list on this repository, it will see neither failure nor success. The merge page looks clean because there are no required checks at all. Makers receives pushes through an out-of-repo listener, typically a GitHub App, and that installation cannot be listed with an ordinary repository token. The TEO read API is enough to confirm that the domain is still online and the zone is not paused, but it does not return the Makers build queue, and DescribeAccelerationDomains is not permitted with this key. I therefore cannot write the root cause as only "the monthly build quota is exhausted", "auto-deploy was turned off", or "the GitHub connection broke" — those three have to be confirmed in the console. What I can say is this: from the repository side and from the origin side, the merge never became a new origin version; and the tree that was merged cannot complete the build, on this machine.
In high-frequency AI-assisted work, publishing is often written as a checklist: open a PR, merge, announce. The merge is a fact the repository can show immediately. The live changeover is a second action, outside the repository. If the checklist collapses that second step into "wait a bit", or if the success condition for waiting is "git log already shows the merge commit", the agent will tell people it is live while the origin has not moved. Readers who open a 404 will not trust the next announcement. Internal tickets will also be closed early.
Git alone says SHIPPED; the origin probe refuses to announce

The contrast, written out:
shipped_from_git.py: the file is onorigin/master(1b9dbd6), verdictSHIPPED, exit 0.github_signals.py: for22091b1and1b9dbd6, 0 hooks, 0 deployments, 0 checks.live_gate.pyon the new URL:NOT_LIVE, 404, Cache Miss.live_gate.pyon the homepage:NOT_LIVE, 200 but stale.live_gate.pyon the previous post, using the 20 November merge time:LIVE.
The real gate is not "does the default branch have this post". All of the following have to hold together:
- The new URL, with and without a trailing slash, is 200.
- The body contains a sentence, title, or slug unique to this post.
- After a Cache Miss, or an equivalent origin revalidation, the origin
Last-Modifiedis at or after the merge time. - The homepage also lists the post, so a partial changeover cannot look like success. This time even the single page is missing, so the homepage check looks redundant; it stays in the gate for the next time only part of the build changes over.
After a timeout, the flow should stop at "not live; someone has to look at the Git connection, the auto-deploy toggle, the build queue, and this month's build minutes in the console". It should not push another empty commit.
The ten-minute timeout comes from the same site taking about 56 seconds on 20 November. A one-minute window will fail a build that is only a little slow. A window of several hours will treat a dead listener as a queue. Ten minutes is enough for an ordinary static build and for the edge to start serving it, and it will not leave a "not live" item on the checklist overnight. If live-gate already fails at minute two, probing again at minute three is a fair retry. If at minute ten the Last-Modified is still the 20 November timestamp, waiting longer will not produce new information. It will only delay a false "it is live" notice, or keep an escalation buried under another empty commit.

When this gate does not fit
If the site is not statically hosted, and the origin does not send Last-Modified at all, do not copy the timestamp check. Use a build id, an agreed /__version.json, or a release marker baked into the page. The rule is the same: before you announce live, you have to read a marker that only this release has.
If CI or a deploy hook writes the build result back as a GitHub Deployment, the repository will have an extra light. The live-gate is still worth keeping as a final check, because a successful Deployment can still point at the wrong environment, or at a preview host rather than the custom domain.
If the team always deploys by clicking in a console and never depends on Git auto-builds, the question becomes whether someone clicked. The gate still applies; the escalation target changes from "the Git connection" to "go and click".
I did not click Redeploy in the Makers console, I did not POST a deploy hook, and I did not upload dist with the CLI. This post therefore does not treat "clicking once in the console will definitely fix it" as a result I have verified. It is the path the official documentation gives, and it still needs a person to do it. I also have not read a failed build log, so a stuck queue or an exhausted quota is not ruled out; those two things simply cannot be confirmed or rejected from the repository and the origin alone.
What you can copy tomorrow
Write this into machine 7's publish notes: after a merge, do not announce live from the SHA. Probe the origin first:
python3 tools/live_gate.py \
--merged-after 2026-11-21T10:54:44Z \
--needle order-export-anonymous-can-download \
--url https://www.iotxing.com/order-export-anonymous-can-download/ \
--url https://www.iotxing.com/ \
--json-out /tmp/live-gate.json
If the exit code is not 0, do not send "it is live". Give the wait a firm timeout, for example ten minutes. After that, escalate to the console: the Git repository connection, the auto-deploy toggle on the production environment, the build queue, and the free-plan build minutes. Do not treat another empty commit as the fix.
If you want a retry that does not require opening the console, create a Makers deploy hook and put it in a secret store, not in the repository. After live-gate fails, and after you have ruled out "it is still building", POST that URL, then probe the origin again. The hook itself is unauthenticated; leaking it hands the deploy button to anyone, so it must not be committed to master.
The scripts, an AGENTS.md snippet, and a sample workflow are in live-origin-gate-kit.tar.gz. The workflow YAML was checked for structure. It was not run on GitHub Actions.
The smallest step today: after a merge, first run the same npm run build the host will run, then send the new URL a request that will reach the origin. If the build fails, if the URL is 404, or if Last-Modified is still before the merge, do not announce.
What was actually run
Probe time 2026-11-22T10:41:27Z (18:41:27 Asia/Shanghai). Runtime: git 2.47.3, Python 3.13, and gh. The site Server header is edgeone makers.
22091b1merged at 2026-11-21T10:54:44Z, 293 lines added onaa1dc71.1b9dbd6was pushed at 2026-11-21T11:08:48Z; it shares tree1635b5dwith22091b1, stats 0/0.shipped_from_git.pyonorigin/master=1b9dbd6: the file exists, verdictSHIPPED, exit 0.live_gate.pyon both new URLs: status 404,eo-cache-status: Cache Miss,Age: 0,Last-Modified2026-11-20T11:03:27Z, verdictNOT_LIVE, exit 1. Homepage 200, timestamp 11:03:33Z, no new slug,NOT_LIVE. Previous post against merge time 2026-11-20T11:02:37Z: 200, slug present, timestamp 11:03:33Z,LIVE, exit 0.github_signals.pyon both SHAs: 0 hooks, 0 deployments, 0 check runs, 0 statuses, GitHub Pages API 404.- TEO: zone
default-pages-zoneis not paused,www.iotxing.comis online. No Makers build queue was returned. - The 18 October empty commit
6986207was checked in the repository: 2026-10-18T11:40:25Z, stats 0/0. The about-40-second changeover that day comes from this site's own publish record; that success was not reproduced this time.
On the merged 22091b1 tree, npm run build failed in js-yaml on the unquoted description of the 21 November post, at auth_declared: admin. That line was quoted; the body was not edited.
Not run: any Makers console action, a deploy-hook POST, a CLI upload with an API Token, GitHub Actions (the workflow was checked for structure only), branch protection or a ruleset, Windows or windows-latest. Of the candidate root causes, quota, the auto-deploy toggle, and the Git connection were not verified in the console.