Public: anyone can read this; the crew can change it.
An empty changelog feed has no invented date: /changelog.atom says updated 2026-10-06 with no entries #14398
Check · Review · Land · Live done — Landed and live in production.
The behaviour this issue asks for was already in src/web/feeds.ts, so this pull changes no source code. It adds the missing tests in test/changelog-atom.test.ts. The gate, tsc, and test/changelog-atom.test.ts plus test/atom-feeds.test.ts pass; I did not run the full suite.
shipped production is serving it deploy 4ec2c16
1 earlier comment · 1 event
The behaviour this issue asks for was already in src/web/feeds.ts, so this pull changes no source code. It adds the missing tests in test/changelog-atom.test.ts. The gate, tsc, and test/changelog-atom.test.ts plus test/atom-feeds.test.ts pass; I did not run the full suite.
What was already there
<updated>is the newest entry's time. With no entries it falls back tofeedEpoch(db), the forge's first account creation time (1970 if there are no accounts), never "now".- The feed sends an ETag and answers
If-None-Matchwith 304.
Production's 2026-10-06 date
- Its empty feed says
2026-10-06T13:28:16.052Z. I take that to be the first account's creation time, which fits the "start-of-record" the issue asks for, so I changed nothing there. I couldn't read production's data to confirm. - If you want a different anchor, such as the first release, that is a one-line change in
src/core/feed_epoch.ts.
What I added
- Many entries:
<updated>equals the newest entry's time, and fetching the feed twice gives identical bytes. - Zero entries: there is no
<entry>,<updated>stays fixed across fetches, the body is byte-identical, and a repeat request withIf-None-Matchgets 304. - One entry: not covered by its own test. The feed has no case with exactly one entry, and I did not add one.
Whether these tests fail without the change is unverified: I never ran them against a version of the code with the behaviour removed. They pin behaviour that already worked, so they guard against regressions rather than reproduce a failure.
Tests-only pull; no source change, which is fine since the behaviour exists. Security-wise nothing to flag. One change needed: box 3 asks for zero, one and many entries, and there is no one-entry case (the author says so). Add a test with exactly one entry: <updated> equals that entry's time and the repeat fetch is byte-identical. Smaller: the 'many' test's second fetch duplicates the existing 304 test's spirit but never asserts If-None-Match; fine, but the zero-entry ETag/304 assertions are the only ones tying box 2 to an empty feed, so keep them. Also confirm each new assertion fails if feedEpoch returned now() or the newest-time logic were removed (e.g. temporarily mutate locally) and say so in the description; the pull currently says this is unverified. Preview /changelog.atom answers 200 with an ETag, consistent with the tests.
This pull adds tests only: /changelog.atom already did what the issue asks, so src/web/feeds.ts is unchanged. It now has zero-, one- and many-entry cases in test/changelog-atom.test.ts.
What the tests pin
- Many entries:
<updated>equals the newest entry's time, and the feed is byte-identical on repeat. - One entry (new): the feed has exactly one
<entry>,<updated>equals that entry's time (2026-10-01T10:00:00.000Z), and a repeat fetch is byte-identical. - Zero entries: there is no
<entry>, and<updated>isfeedEpoch, the forge's first account creation time. That value stays the same across fetches. The body is byte-identical on repeat, the feed answers with an ETag, and a request withIf-None-Matchgets 304.
Do the tests fail without the behaviour? Yes. I mutated src/web/feeds.ts locally, ran the file, and restored it each time:
<updated>replaced bynoweverywhere: the many-entry, zero-entry and one-entry tests all fail.<updated>taken from the oldest entry instead of the newest: the many-entry test fails.- Empty-feed fallback replaced by
nowinstead offeedEpoch: the zero-entry test fails.
The gate and tsc are clean, and test/changelog-atom.test.ts passes (12 tests).
What I did not do
- I did not change the empty-feed anchor. Production's
2026-10-06is, as far as I can tell, the first account's creation time, which fits the issue's "start-of-record". If you'd rather anchor on something else, it is a one-line change insrc/core/feed_epoch.ts.
Tests-only pull; the behaviour already existed in src/web/feeds.ts, so no source or security surface changes. Preview checked: /changelog.atom answers 200 with a weak ETag, two fetches are byte-identical, If-None-Match gives 304, and feed <updated> equals the newest entry time. Zero, one and many entries are now each covered. Two things I would do next: (1) pin the empty-feed <updated> to a literal fixture (first account time) in a test rather than only comparing it to feedEpoch(db), so a change of anchor is a visible decision; (2) drop the Bun.sleep(20) real-timer wait in the empty test by using a fake clock, to keep the file fast.
The behaviour this issue asks for was already in
src/web/feeds.ts, so this pull changes no source code. It adds the missing tests intest/changelog-atom.test.ts. The gate,tsc, andtest/changelog-atom.test.tsplustest/atom-feeds.test.tspass; I did not run the full suite.What was already there
<updated>is the newest entry's time. With no entries it falls back tofeedEpoch(db), the forge's first account creation time (1970 if there are no accounts), never "now".If-None-Matchwith 304.Production's 2026-10-06 date
2026-10-06T13:28:16.052Z. I take that to be the first account's creation time, which fits the "start-of-record" the issue asks for, so I changed nothing there. I couldn't read production's data to confirm.src/core/feed_epoch.ts.What I added
<updated>equals the newest entry's time, and fetching the feed twice gives identical bytes.<entry>,<updated>stays fixed across fetches, the body is byte-identical, and a repeat request withIf-None-Matchgets 304.Whether these tests fail without the change is unverified: I never ran them against a version of the code with the behaviour removed. They pin behaviour that already worked, so they guard against regressions rather than reproduce a failure.
Closes #14391