Case study: date-fns#2068

date-fns/date-fns#2068 (open when this was recorded) reports that eachHourOfInterval returns one hour too few when the interval spans the night the clocks go back. It only shows in time zones with daylight saving time, so UTC machines and CI never see it.

File	What it is
new-york.json	capsule from a reporter in America/New_York: the regression test fails
kolkata.json	capsule from a maintainer in Asia/Kolkata: the same test passes
replay.txt	twin replay new-york.json on Solari, recorded 2026-09-29
bisect.txt	twin bisect new-york.json --good kolkata.json on Solari, recorded 2026-09-29
fix.patch	a candidate fix for pkgs/core/src/eachHourOfInterval/index.ts
verify.txt	twin verify new-york.json --patch fix.patch on Solari, recorded 2026-09-29

Both capsules are from date-fns commit 717ce0a (main), Node 26.7.0, pnpm 10.20.0. The regression test is present as an uncommitted change, so it travels inside the capsule’s working-tree diff (fix.patch only touches the source file).

What happened

Capture. Run from pkgs/core, the same command under two time zones, with LANG=C.UTF-8, no LC_* variables and NO_COLOR=1:

pnpm vitest run src/eachHourOfInterval/test.ts

TZ=America/New_York fails (one entry missing, 1604210400000); TZ=Asia/Kolkata passes. twin diff kolkata.json new-york.json shows only the time zone.

Replay (32 s wall). twin rebuilt the environment on a fresh Solari sandbox: Node 26.7.0, the commit fetched by SHA, the working-tree diff applied, pnpm 10.20.0, pnpm install --frozen-lockfile, and TZ=America/New_York. All three attempts failed with the capsule’s signature: REPRODUCED.

Bisect (32 s wall). twin built the Kolkata environment (20 s), confirmed it passes, applied the failing environment’s differences and reported the minimal one: TZ=America/New_York.

Verify (32 s wall). The function advances with date.setHours(date.getHours() + step). On the night of a fall back, one local hour occurs twice and this call resolves to its first occurrence, so the second one is never produced. fix.patch steps by elapsed time instead (date.setTime(+date + step * millisecondsInHour)) and re-aligns to the hour start for zones whose DST shift is not a whole hour. twin applied it on a fresh sandbox at the captured commit: 3 of 3 attempts passed, FIXED.

Wall times are the whole twin command, measured around it. Setup is dominated by pnpm install (12 to 14 s).

Honest notes

• An open pull request, date-fns/date-fns#4140, covers the same function. This example shows the twin workflow, not a claim of a new fix.
• Capture used NO_COLOR=1. Without it, replay printed DIFFERENT FAILURE with an identical assertion: vitest prints the project name as |main| when it sees no color support (the local, piped capture) and as main otherwise (the sandbox run), so the signatures differed.

Run it yourself

export SOLARI_API_KEY=...
twin replay examples/date-fns-2068/new-york.json
twin bisect examples/date-fns-2068/new-york.json --good examples/date-fns-2068/kolkata.json
twin verify examples/date-fns-2068/new-york.json --patch examples/date-fns-2068/fix.patch

Every file referenced here is in examples/date-fns-2068.