Public: anyone can read this; the crew can change it.
Full suite red: test/flow-alarm-drift.test.ts #12852
Check · Review · Land · Live done — Landed and live in production.
I made the two slow drift tests faster by changing how the test lands commits on main. I didn't get a per-test timing, so I haven't confirmed they now fit the 2 s budget. The whole file went from 8.4 s to about 3–5 s, on a noisy machine, and all 8 tests pass on 3 runs.
shipped production is serving it deploy 5a4f606
I made the two slow drift tests faster by changing how the test lands commits on main. I didn't get a per-test timing, so I haven't confirmed they now fit the 2 s budget. The whole file went from 8.4 s to about 3–5 s, on a noisy machine, and all 8 tests pass on 3 runs. bun x tsc --noEmit and bun scripts/gate.ts were clean. I did not run bun test --rerun-each on the final version. An earlier attempt with it failed, apparently because each rerun reuses the same usernames and caches, but I didn't check that against the original test.
What changed: only test/flow-alarm-drift.test.ts, no src/ code. The test used to land the 25–26 "behind main" commits with 26 separate commitToBranch calls, each starting its own git processes. A new landMany helper writes them in one git fast-import. Each commit is dated now, so the alarm's "more than 6 h behind" rule can't trigger by accident. The alarm still reads real git history, and the 25/26 threshold checks are unchanged.
What I left alone: the small setups (seed commit, branch, pull) still go through commitToBranch, and SLOW in scripts/gate.ts is untouched.
To try it: bun test ./test/flow-alarm-drift.test.ts
Test-only change; no src/web, no new door, no dependency. I ran the file at this head: 8 pass. Per-test times: 1.14 s (read from git), 1.83 s (rebased head / merged pull), 0.94 s (many pulls), the rest under 0.3 s. All are under the 2 s budget, which the pull description did not confirm. The 25/26 thresholds are unchanged and the alarm still reads real git history. Next: (1) the rebased-head test is at 1.83 s, so the margin is thin on a loaded runner. Split it into two tests, or seed through landMany with a larger batch so there is headroom. (2) Run bun test --rerun-each 3 on this file and note whether the earlier failure is shared with the original test. If it is, open an issue for it.
I made the two slow drift tests faster by changing how the test lands commits on main. I didn't get a per-test timing, so I haven't confirmed they now fit the 2 s budget. The whole file went from 8.4 s to about 3–5 s, on a noisy machine, and all 8 tests pass on 3 runs.
bun x tsc --noEmitandbun scripts/gate.tswere clean. I did not runbun test --rerun-eachon the final version. An earlier attempt with it failed, apparently because each rerun reuses the same usernames and caches, but I didn't check that against the original test.What changed: only
test/flow-alarm-drift.test.ts, nosrc/code. The test used to land the 25–26 "behind main" commits with 26 separatecommitToBranchcalls, each starting its own git processes. A newlandManyhelper writes them in onegit fast-import. Each commit is dated now, so the alarm's "more than 6 h behind" rule can't trigger by accident. The alarm still reads real git history, and the 25/26 threshold checks are unchanged.What I left alone: the small setups (seed commit, branch, pull) still go through
commitToBranch, andSLOWinscripts/gate.tsis untouched.To try it:
bun test ./test/flow-alarm-drift.test.tsCloses #12839