Engineering impact

160.5 s to 39.5 s median; p90 236 s to 77.5 s. One late-night hour after merge; re-measure on a weekday. Job start and compile start also moved (282 s to 60 s, 430 s to 242 s), but pool routing changes merged in the same hour confound those.

0% to 49%, over one 3.5-hour window of US night. Owned-placed admissions waited 2 s median and 3 s p90 for a runner; the other 27 overflowed to Blacksmith. Of the 26 owned attempt-1 jobs, 5 succeeded, 11 failed and 10 were cancelled (mostly superseded pushes). The failures mix real build failures with owned-runner refusals, which the retry path (cmux#14312, cmux#14325) re-runs as attempt 2; this measurement keeps attempt 1 only, so a refused job counts as owned and its refusal plus re-queue time is not in the wait. Runner speed after start is not compared here.

22.5 s to 3 s median; p90 323 s to 229 s; mean 138 s to 70 s. The baseline is conservative: it starts after cmux#14205 (least-queued Blacksmith pool, 12:13Z) and had about half the macOS job demand of the after window (31 to 41 vs 55 to 89 jobs per hour). At matched demand before cmux#14205 the median was 312 s (n=159), but most of that gap belongs to cmux#14205 and a single capped 6vcpu pool, so it is not claimed here. The after window is 3.5 hours of US night, and owned slots were raised again at 04:32Z. Reruns are excluded because their job created_at can follow started_at. Attempt 1 only: for an owned job that was refused and re-run, the extra refusal and re-queue time is not counted, which flatters the after median.

173.6 s median minus the larger after value (7.3 s). Only two post-merge runs existed at recording time, so this stays observed. Shard 1's job dropped more than the suite did (581 s median to 319 s and 360 s), so part of that drop comes from other merges and is not claimed. Shard 4 was already the longest app-host job, so this saves shard 1 runner time rather than shortening the app-host lane.

90.4 s to 2.4 s median; every before run fell in 90.2 to 90.6 s and every after run in 2.35 to 2.54 s. Six runs that ran the test in under 1 s (a different selection), three on each side, are excluded. The test runs inside the parallel 'Run unit tests' step, so its time is not all wall time.

Upper bound on the lane effect, recorded as runner time. Shard 4 was usually the longest app-host job (median 653 s before, 590 s after), but shard 1 was close (job median 581 s) and ran longer than shard 4 in some after runs (run 36085886212: 653 s against 628 s), so the app-host lane's wall time fell by less than this step until cmux#14333 shortened shard 1. Other merges landed in the same windows and the after sample is five runs, so not all of the step drop is attributable; the per-test drop is. The PR body placed the test in the 'Run CLI no-socket regressions' step, but it runs in 'Run unit tests'; the no-socket step stayed at 218 to 252 s.

82% (9 of 11) to 47% (7 of 15). The task counts are bimodal, so the median (2615 to 1236 tasks) sits on the boundary job of the lower cluster and one more whole-module job would move it back near 2480; it is not claimed. Jobs 7 to 52 changed inputs from the kept build compiled 1 to 177 tasks; jobs 71 to 187 inputs away still rebuilt the module, because the kept build is the previous PR's, not the merge base. The PR's stamp bump made each Mac's first job after merge cold (18 of 35 owned admissions cold after, 13 of 25 before), a one-time cost not netted here.

Null result for time: 384 s median before, 380 s after. p90 rose from 396 s to 836 s on two contended cmux13s jobs (836 s, 1639 s), unrelated to replay. Jobs that avoided the whole-module rebuild compiled in 91 to 241 s. The PR's expected drop to 50 to 100 s did not hold at the median.

One canary run, one edit per mode. Uses 139 s minus 31 s; the second cache-on series gives 98 s. Only pays off when an admission rebuilds on adopted DerivedData. Adding a top-level declaration in the app still recompiles every cmuxTests file with or without the cache (117 s in the same run). Of the remaining 31 s, 25.7 s was cmuxTests' own emit-module job, which cmux#14364 then removed.

Xcode Build Timing Summary, summed task time across the 12 vCPU runner, not wall time. Same caveats as the wall-time claim.

Two edits each way, one runner each. Only pays off in PR admission when it adopts DerivedData to rebuild on: the macOS 15 seeds (cmux#14315, cmux#14358) merged the same morning, and seeds built before this fix used the cache for cmux, so the first adoption after a stale seed can still rebuild the app fully. Xcode 26.6 pools were already incremental and are unchanged.

Uses the smaller before value (448 s minus 41 s). The fix as merged measured 72 s and 44 s in run 36090019330. Same caveat as the task claim: it needs adopted DerivedData to rebuild on.

Median lane job 197 s, gross, on runs that routed to the CLI lane. The mean lane job was 382 s, so the median is the conservative figure. Averaged over all 137 PR runs in the window this is about 59 s per run. This removal also ends the cli-pipe-resolve-time claim in cmux-14257.

9 s to 11 s median: the cost of running the two moved tests in admission, paid by every admission, including runs that never used the CLI lane. Admission is on the critical path, so this is a small increase there.

Step totals: CLI build 88 min, Restore Swift packages 85 min, resolve 32 min, tests about 23 min, the rest setup and checkout. Critical-path effect is small: the lane was the last job to finish in only 3 of 41 runs, so no time claim is made.

Uses the smaller before (31.6 s) minus the larger after (4.2 s). The whole cmux-unit rebuild went from 47 to 55 s to 19 s in the same series, which contains this saving, so it is not claimed separately. A standalone-driver series that still emitted the module took 36.4 s, so the gain comes from dropping the module. Only pays off when an admission rebuilds on adopted DerivedData, and the first admissions after the merge rebuild cmuxTests once because its arguments changed. Too few post-merge admissions existed at recording time (29 minutes after the merge) to measure it there. cmux#14378 (merged 04:40Z) applies the same settings to the local scripts/test-unit.sh.

Backlog clearance, capped at 64 per run, and eligible branches remained, so the total keeps rising. Before, these were only removed by a manual git branch -D. Big Red runs the same job; its receipts were not read here.

751 to 0. Wall time went 684 s to 2,643 s, likely from a 66 GB VM image pull on the same host (unconfirmed); this is a reliability win, not a speed win.

Before, the deployed (launchd) node could not reach the store at all, so deployed cross-machine reads served nothing; lab reads in glaeda-1162 and glaeda-1180 ran from ssh, which the block does not affect. The 326 s build time is not claimed as a reduction: there is no same-host, same-Xcode cold baseline for it (756 s in glaeda-1162 was Xcode 26.3).

One-time reclaim. The steady-state effect is capping growth at about one day of per-commit archives, not measured here.

The PR's stated growth rate (143.4 GiB held after about five days; the PR puts steady growth at about 35 GiB a day, mostly compilation archives).

451 to 129 median on the same pool. Later runs reached a median of 28, but those also had #14188's faster seed cadence and #14190. Build time on the same pool moved only 845 s to 737 s.

148 s to 110 s median. From job start the gain is 23 s, because checkout got about 12 s slower in the same period.

100 s to 79 s median shortly after merge; over the whole day 89 s, and nothing measurable from run start because change detection itself got slower. #14176 accounts for about 5 s of it.

A reporting correction, not a compute saving: queue-only minutes were about 14 times the real runner minutes (4,732 vs 345), so the old report showed about 15 times the real cost.

1,342 s to 16 s median. After 14:00Z the 12 vCPU pool queued too (621 s and 663 s), once PR admissions began preferring it.

2,628 s to 740 s median, including the faster build and #14174 and #14190.

14.6 s to 0.43 s, measured locally on a loaded host with a full clone; a CI shallow checkout will differ and was not measured.

120 s to 40 s compile on the same PR. The broader 388 s to 36 s medians (n=4 each) compare first admissions against reruns and are not like for like. No true CLI-only PR has run yet.

888 s to 435 s median. The before side includes 7 builds that had already adopted main's DerivedData or an early admission seed.

553 s to 518.5 s median over all builds. A seed hit builds in 367.5 s, but 8 of 22 builds missed with no-seed, cause unknown. macOS 15 has no seed until #14315.

The 82 sampled cmuxTests-only runs had 12 to 27 minute build jobs; the duplicate refs' own compile times were not listed. No after-count was taken.

583 s to 371 s. One run each, so indicative only.

Each would have cost 12 to 27 minutes of macOS compile. No after-count was taken.

29 of 29 to 0 of 3. The after seeds also carried cmux-14262's checksum-only mode, so the 0 of 3 is both changes combined. The cmux-unit phase did not get faster: 12 vCPU 129 s median to 112 s (n=1); 6 vCPU 199 s median to 207 s and 242 s. cmuxTests still recompiles fully after adoption (cmux#14320).

10.0 s median to 2.54 s. The before runs span three pools; the after run was 6vcpu macOS 26.

48 s median to 12 s (4 s checkout plus 8 s restore). One after-sample per lane.

35 + 21 + 17 + 15 minutes of checkout in the window.

63 s to 17 s median.

48 s to 15.5 s median.

8 of 22 to 0 of 22 by replaying the new rule on the same skips, not by live runs after merge. One extra, harmless build.

129 to 0. One run each: the before run was 12 vCPU with a same-width seed, the after run 6 vCPU with a j12 seed, and the after seed also carried cmux-14241's directory-time replay. The Swift module re-emits remained until cmux-14275.

62 to 0, one run each. Holds when a seed of the runner's width exists. Owned 14-CPU minis have none yet (cmuxterm-hq#590). cmuxTests still recompiles fully after adoption (cmux#14320).

145 s to 59 s. One run each, so indicative only.

52 to 56 s before, 13 to 16 s after; 36 s is the conservative gap (52 minus 16). cli-pipe-regressions on a mini still missed (41 s, n=1) because its workspace path differs.

15.4 to 10.7 min. It comes from skipping the seed and cache restore: the kept DerivedData still recompiled about 2,600 tasks, because the kept tree is whatever PR the mini built last (cmuxterm-hq#590). Warm owned is still 0.4 min behind the Blacksmith median. One warm job (107903167882) failed after compiling.

0 to 1, job-verified on the fork. By glaeda#1175, 13 runners were registered on manaflow-ai/cmux but not yet taking jobs; glaeda#1200 plans 36. Neither is job-verified here.

756 s to 132 s. Lab measurement, not CI. For the full app the store and reader were on the same mini; cross-machine was shown on the CmuxSettingsUI chain, and for the full app in glaeda-1180.

756 s to 127 s. Members share entries only on the same Xcode build. The network adds about 20 to 30 s.

Parallel execution exceeded 12 minutes without finishing; serial execution completed in 65 seconds. The stored 655-second reduction is therefore a lower bound.

The parallel run produced about 161 issues while the serial run reported 17 real failures; the difference is approximately 144 contention artifacts.

Observed queue p90 was 83.3 minutes on macOS 15 and 1.0 minute on macOS 26. Execution median was about four minutes slower on macOS 26, and the test mixes differed, so the comparison is operational rather than causal.

Execution median was 18.7 minutes on macOS 15 and 22.5 minutes on macOS 26 while total elapsed median still favored macOS 26 after queueing.

The baseline was about eight minutes per consumer; the PR reports the transfer step as roughly 98 percent transport time and moves the path to concurrent ranges with verification and fallback.

The PR projects roughly 450–600 macOS minutes/day avoided at the two-hour default. This record stores the lower edge of that stated range; the PR also gives a roughly 280-minute/day lower-bound scenario.

The seven-day sample recorded 7,197 push-to-main macOS minutes, about 1,030 minutes/day including three hung builds.

The report found 9,763 runs beyond the first for a head that did not change. This is an observation to investigate, not a claim that every repeated run is removable.

The previous window's corresponding p90 was 0.9 minutes, which made the new 119.8-minute queue visible as a capacity regression.

The PR states about 16–17 macOS minutes saved per upload. This record stores the lower edge of that range.

The PR gives an upper ceiling of about 6.5 macOS-hours/day when all 24 hourly opportunities upload. Actual savings depend on upload frequency.

The export contains 1,966,401 self-hosted macOS minutes. ci.yml alone accounts for 1,345,657 minutes, or about 68.4 percent of this total.

This 90-day average is about 16.8 percent above the separate 18,704-minute one-day accounting from September 22.

Floor valuation only: 655,467 equivalent minutes per 30 days at $0.08/min. No 12-vCPU premium is applied. This is replacement-cost exposure, not a claim that every minute was actually billed to Warp or saved by one contributor.

Annualized replacement-cost floor at the 90-day average. Any 12-vCPU Warp usage would raise the equivalent value.

This independent issue-level accounting remains a useful corroborating snapshot but is no longer the primary economic baseline.

Fresh dispatches do not claim lower total runner minutes; this figure is the measured compile portion removed from a rerun by the new job split.

The fix changes diagnostics, not the fail-closed verdict: incomplete runs remain failed while recorded failures are now named.

Exact count from receipts; each reclaim is recoverable from the recorded commit and branch. Free space rose by 11.4 GiB (sum of per-repository before and after readings, so other host activity can move it).

Gross bytes deleted, not sustained free space: DerivedData and scratch regrow when the next build or session needs them. 35.5 GiB of it came after merge; the other 17.3 GiB was deleted by the PR's own build before merge. The minis run the same pressure job but have not been under pressure (194 to 270 GiB free against a 69.1 GiB threshold), so there are no mini receipts.

The clean seed is about 3.5 GiB instead of roughly 5.0 GiB, so every pull request restores about 1.5 GiB less.

Replay estimate at the default 180-second delay; the same analysis measured a 1.4-minute median increase to the green signal.

This moves workload onto sponsored capacity; it does not claim the compute disappeared.

The measured baseline was about 6,864 router runs/day; the once-per-CI-run trigger was expected to produce about 1,253/day, an 82% reduction.

The PR explicitly left warm compile time unmeasured until the next scheduled seed.

The measured build fell from 38.9 seconds to 13.3 seconds and compiled zero app-target Swift files after the move.

Fresh DerivedData took 953 seconds; a new tag using the already-warm shared DerivedData took 35.7 seconds.

The oldest macOS entry was only five minutes old and the shared Debug compilation seed was already missing. Pull-request jobs had 14 save+restore cache steps and wrote 1–3 GB PR-scoped entries.

The PR deliberately did not quote an unmeasured CI savings number before the scheduled seed had run.

Measured guard wall time fell from 385.5 seconds to 124.3 seconds, about a 68% reduction.

Runner consumption rose from 6.43 to 7.49 minutes because parallel jobs duplicate setup work. The change intentionally spends cheap compute to return the macOS admission answer sooner.