You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
docs site: production has not deployed since dpl_2nfW… — 19 merges to main, five epic PRs invisible #12333
https://objectstack.ai is serving deployment dpl_2nfWjGjSwZjakVEmUG1kBWD6697r. That id is unchanged across 19 merges to main, four of which touched apps/docs, over roughly two hours.
dpl: dpl_2nfWjGjSwZjakVEmUG1kBWD6697r | age: 6510 | x-vercel-cache: HIT
The live generation can be dated precisely by what it does and does not contain:
So production is pinned to a build made between 16:52 and 16:57 on 2026-08-25.
Control: this is not CDN staleness that a request can defeat. Cache-Control: no-cache / Pragma: no-cache request headers and a cache-busting query string both still return x-vercel-cache: HIT with an age of 6510s, and the data-dpl-id embedded in the HTML — which is a property of the build, not of the cache entry — has not moved.
Why it matters
Five merged PRs from epic #12243 are invisible to users and to crawlers, and the epic's definition of done is written against the live site, so it cannot be verified while this holds. Three of its five boxes were ticked against production earlier today; the rest cannot be.
What is needed, and why it is not a PM action
Everything visible from GitHub is green: all five PRs merged through the merge queue with full CI. Whatever is stuck is on the Vercel side — a failing production build, a paused project, a promotion that is not happening — and reading that requires the Vercel dashboard or API token for the docs project. No agent in this repo has it.
Maintainer action: check the docs project's deployment list for a failed or queued production build since ~16:55 on 2026-08-25.
If it turns out a build is failing, that failure is the real card and this one closes into it. If deployments simply resumed on their own, close this and the epic's verification continues.
Corroboration: a probe polled the live homepage <title> every 90 seconds for 50 attempts (~75 minutes) waiting for it to change to the value merged in #12284. It never changed. Final readings:
homepage title: ObjectStack — Apps small enough for AI to hold whole. age: 8311
homepage canonical: (absent)
docs h1 count: 2 age: 7859
docs canonical: (absent)
mdx x-robots-tag: (absent) age: 7863
robots.txt: (still the three-line version, no Allow: /llms.txt) age: 8240
Ages are now ~2.3 hours and rising monotonically, which is the signature of a generation nobody has replaced rather than of a cache that keeps being refilled.
Still pinned. dpl_2nfWjGjSwZjakVEmUG1kBWD6697r unchanged.
The count of merged-but-invisible work from epic #12243 is now eight PRs: #12262, #12284, #12303, #12305, #12325, #12351, #12355, #12364. Everything in that list passed the merge queue with full CI and is on main; none of it is reachable by a user or a crawler.
Root cause found — the maintainer performed exactly the action this card requested (read the Vercel deployment list) and pasted the failing production build log in the PM session, 2026-08-27, verbatim:
15:45:30.171 @objectstack/spec:build: ELIFECYCLE Command failed with exit code 137.
15:45:30.206 @objectstack/spec#build: ERROR command (/vercel/path0/packages/spec) /vercel/.local/share/pnpm/.tools/pnpm/10.31.0/bin/pnpm run build exited (137)
15:45:30.234 ERROR run failed: command exited (137)
15:45:30.307 Error: Command "cd ../.. && pnpm turbo run build --filter=@objectstack/docs" exited with 137
Exit 137 = 128 + SIGKILL: the build container's OOM killer. Every production build since the ~16:55 2026-08-25 window has been dying this way, which is why data-dpl-id never moved — the mystery this card measured is fully explained.
Per this card's own protocol ("If it turns out a build is failing, that failure is the real card and this one closes into it"): the failure card is #12677, dispatched to a dev seat. The data-dpl-id re-check command migrates there as the post-merge verification, and epic #12243's site-level verification resumes once the id moves.
Activity
os-zhuang commentedon Aug 25, 2026
Corroboration: a probe polled the live homepage
<title>every 90 seconds for 50 attempts (~75 minutes) waiting for it to change to the value merged in #12284. It never changed. Final readings:Ages are now ~2.3 hours and rising monotonically, which is the signature of a generation nobody has replaced rather than of a cache that keeps being refilled.
os-zhuang commentedon Aug 25, 2026
Still pinned.
dpl_2nfWjGjSwZjakVEmUG1kBWD6697runchanged.The count of merged-but-invisible work from epic #12243 is now eight PRs: #12262, #12284, #12303, #12305, #12325, #12351, #12355, #12364. Everything in that list passed the merge queue with full CI and is on
main; none of it is reachable by a user or a crawler.os-zhuang commentedon Aug 27, 2026
Root cause found — the maintainer performed exactly the action this card requested (read the Vercel deployment list) and pasted the failing production build log in the PM session, 2026-08-27, verbatim:
Exit 137 = 128 + SIGKILL: the build container's OOM killer. Every production build since the ~16:55 2026-08-25 window has been dying this way, which is why
data-dpl-idnever moved — the mystery this card measured is fully explained.Per this card's own protocol ("If it turns out a build is failing, that failure is the real card and this one closes into it"): the failure card is #12677, dispatched to a dev seat. The
data-dpl-idre-check command migrates there as the post-merge verification, and epic #12243's site-level verification resumes once the id moves.Closing into #12677.
Generated by Claude Code