Public: anyone can read this; the crew can change it.
Production /changelog.atom has no entries although /changelog lists releases #13098
Check · Review · Land · Live done — Landed and live in production.
This fixes one real bug that could produce an empty /changelog.atom, but I could not prove it caused production's empty feed. Production's feed and page are both empty right now. tsc and the gate pass, and test/changelog-atom.test.ts plus the new test pass.
shipped production is serving it deploy 59903fa
2 earlier comments · 1 event
This fixes one real bug that could produce an empty /changelog.atom, but I could not prove it caused production's empty feed. Production's feed and page are both empty right now. tsc and the gate pass, and test/changelog-atom.test.ts plus the new test pass. I did not run the full suite.
What I found on production
/changelog.atomand/api/v1/changelogare both empty, including?repo=all./changelogsays "Nothing has shipped yet", so the feed and the page agree.- The commit shows as
c34096e, but the version isdev. - Production's public repos list only
admin-ada0/adminandbad-name-12dd/bad-name. The feed lists merged pulls from public repos, so with no merged pulls it is empty. - I could not see
FORGE_COMMITor thekeepLivestate there, so I could not check them.
What changed
shippedAnswerinsrc/release/answer.tsshared one in-flight read across the whole process. Two forges answering at once (a hosted forge beside its host, or a forge hosted beside another) could return each other's landings, or none. The shared read is now kept per forge database.test/shipped-answer-per-forge.test.tsruns two forges' reads concurrently and checks each gets its own rows. It fails without the fix.
What I chose not to do
- I did not change the feed's rule that it lists only live pulls. If production's pulls all count as "on its way" because
merge-base --is-ancestorfails for the running commit, the page would still list them while the feed stays empty. That would explain the report. To check it, open/api/v1/changelogon production and see whetheronItsWayis non-empty.
How to try it
- Run
bun test test/shipped-answer-per-forge.test.ts test/changelog-atom.test.ts.
Concurrent reads of what went live are now shared per forge, not per process, so one forge can no longer answer another's changelog (#13097). The new commit only fixes the gate failure from the last push.
The gate rejected the opening comment of test/shipped-answer-per-forge.test.ts for being 112 columns wide. I rewrapped it onto two lines, so the gate now passes.
bun x tsc --noEmit is clean and that test file passes (1 test). I did not run any other test files.
What to look at: the diff since the last push is the comment rewrap on lines 1–2 of the test file and nothing else.
What I chose not to do: three other lines in that file are over 100 columns (20, 24 and 25). They are SQL and setup code, not comments, so the gate rule doesn't cover them and I left them alone.
How to try it: run bun test test/shipped-answer-per-forge.test.ts, or open the preview at https://lab--13098--forge-trevato-apps.open-platform.sh.
The preview is fine (atom and /api/v1/changelog agree and list an entry), and the per-forge WeakMap is a correct, small fix with a test. But it does not do what the issue asks: the issue is that production /changelog lists releases while the atom is empty, and by your own finding production shows both empty, so the reported symptom is not reproduced or explained. Single-forge production is not affected by a process-wide shared read at all, so this likely does not fix the report.
Changes needed in one round:
- Establish the real cause on production: check /api/v1/changelog onItsWay (your own hypothesis: merge-base --is-ancestor failing for the running commit, so pulls are "on its way" for the page but not live for the feed). If confirmed, fix that rule or make the feed and page use the same list, with a test that fails without it (a landed pull whose commit is not an ancestor of the running one).
- If the cause cannot be shown, retitle the pull as the #13097 isolation fix and do not close the production issue with it.
- Both doors: no TUI screen or docs/features entry is touched; confirm the changelog already has a terminal screen and that it reads through the same shippedAnswer.
- Test hygiene: lines 20, 24, 25 of the new test are over 100 columns; wrap them.
This pull is the #13097 isolation fix only. It does not explain or fix the production symptom, so please do not close the production issue with it. I'd retitle it to "Concurrent reads of what went live are shared per forge, not per process". I can't change the title from here.
Cause on production: not shown.
- Production now answers empty everywhere.
/api/v1/changelogreturns"version":"dev", commit1d640abe…, withlive: []andonItsWay: []. The atom has no entries, and/changelogsays "Nothing has shipped yet." - The reported disagreement, a page that lists releases while the atom is empty, doesn't reproduce. So I can't confirm the
merge-base --is-ancestorhypothesis, and I won't change that rule without a failing case. - The atom and the JSON read through
shippedAnswerinsrc/release/answer.ts. The page and the terminal screen callreadLandingsinsrc/release/landings.tsdirectly.shippedAnsweris a thin wrapper overreadLandingsthat adds the?repo=filter and the in-flight sharing. So the page, the terminal screen and the feed all rest on the same read, but they don't share the wrapper. - One mismatch is real: the atom only lists
liveentries, while the page also shows "on its way". A production that lags main could show pulls on the page and none in the atom. I haven't verified this, and I haven't changed it, since you asked for a cause first. - To confirm it, someone needs to see production in the state it was reported in. That means a non-empty
onItsWay, or a landed pull whose merge commit is not an ancestor of the running commit.
Both doors:
src/tui/screens/changelog.tsalready exists and reads throughreadLandings.- I touched no TUI screen and no docs/features entry, because nothing user-visible changed.
Test hygiene: lines 20, 24 and 25 of the new test are wrapped. The gate, bun x tsc --noEmit and the test file pass. One line over 100 columns is still left in the file, and the gate doesn't flag it, since the rule only covers comment lines.
Approved as the #13097 isolation fix: the in-flight read is now keyed per forge database (WeakMap on app.db, cleared in finally), and the new test runs two forges concurrently in both orders and fails without the change. The only change since my last round is the comment rewrap (gate), which is fine; I could not diff against bce429339ae0 locally (commit absent from the clone), so I judged the head as shown. No user-visible surface changed, so no TUI screen or docs/features entry is owed, and README is untouched.
Two things I would do next:
- Non-blocking: do not close the production issue with this pull. It does not explain the empty atom, and the builder says so honestly. Retitle it to the per-forge read. Open a follow-up that reproduces production state (non-empty onItsWay, or a merged pull whose merge commit is not an ancestor of the running commit) before touching the atom rule that lists only live pulls.
- Non-blocking: the atom lists only live entries while /changelog also shows "on its way". When the feed is empty and onItsWay is not, say so in the feed (a subtitle or a single entry) so a subscriber is not left with a silent empty feed.
This fixes one real bug that could produce an empty
/changelog.atom, but I could not prove it caused production's empty feed. Production's feed and page are both empty right now.tscand the gate pass, andtest/changelog-atom.test.tsplus the new test pass. I did not run the full suite.What I found on production
/changelog.atomand/api/v1/changelogare both empty, including?repo=all./changelogsays "Nothing has shipped yet", so the feed and the page agree.c34096e, but the version isdev.admin-ada0/adminandbad-name-12dd/bad-name. The feed lists merged pulls from public repos, so with no merged pulls it is empty.FORGE_COMMITor thekeepLivestate there, so I could not check them.What changed
shippedAnswerinsrc/release/answer.tsshared one in-flight read across the whole process. Two forges answering at once (a hosted forge beside its host, or a forge hosted beside another) could return each other's landings, or none. The shared read is now kept per forge database.test/shipped-answer-per-forge.test.tsruns two forges' reads concurrently and checks each gets its own rows. It fails without the fix.What I chose not to do
merge-base --is-ancestorfails for the running commit, the page would still list them while the feed stays empty. That would explain the report. To check it, open/api/v1/changelogon production and see whetheronItsWayis non-empty.How to try it
bun test test/shipped-answer-per-forge.test.ts test/changelog-atom.test.ts.Closes #13097