crier
a note·30 August 2026 ·26 weeks · 3,417 commits

The floor
does not
compound


Twenty-six weeks of work, measured. Output peaked in mid-July and has roughly halved since, which sounds like a slowdown and is not one. The number that fell is the wrong number, and the thing that actually changed is visible in a different chart.

the metric that fails

Commits per week measures the wrong thing, three ways

It tracks the model in use. The Co-Authored-By trailers run Opus 4.6 through 4.7, 4.8 and 5, and both peak weeks — 437 and 414 — sit squarely in the 4.8 era. It tracks how the work was chunked: weft’s 256-commit week was a fleet of workers each landing one CSS section against an oracle, while the 88 commits of the week after contain a bare-metal SSH login that took days to earn. And it inverts during integration — the broadest week in the whole record, touching 23 repositories, is also one of the quietest by count.

Split by where the commits landed, though, it says something real.

modus — the language and the kernel every other repository
Commits per week, modus versus every other repository Language bring-up Library explosion Integration 100 200 300 400 1 Mar 25 5 Apr 26 10 May 26 15 Jun 26 20 Jul 26 25 Aug 26 11 months 437
Commits per week by author date, week 1 to 26. modus never speeds up and never stops — 37 to 186 commits a week from April onward, straight through the explosion around it. Everything else goes from nothing to 371 in a single week and back down. The dashed rule marks the eleven months between the first commit and the second.
Show the data
#week endingmodusothertotalrepos
12025-03-151011
22026-02-284041
32026-03-073031
42026-03-143031
52026-04-11500501
62026-04-1810701071
72026-04-25820821
82026-05-0215901591
92026-05-09620621
102026-05-1611101111
112026-05-23870871
122026-05-3011101111
132026-06-0617301731
142026-06-1311301131
152026-06-20186412273
162026-06-2772521244
172026-07-046637143714
182026-07-116117223317
192026-07-184324829127
202026-07-251639841429
212026-08-0110419029434
222026-08-0811210621838
232026-08-151809327338
242026-08-22379913642
252026-08-2941478842
262026-09-05011011042
what the shape says

Three projects, in sequence

Weeks 1–14 are one repository. A language being built, with nothing to reuse because nothing else exists yet. Fifty to a hundred and seventy commits a week, all of it modus: a garbage collector, a calling convention, a numeric tower, an object system, and a test harness that had to be rebuilt after it was caught losing 17,587 tests without saying so.

Weeks 15–22 are the explosion. The workspace goes from one repository to thirty-eight. Ten appear in a single week, twice. This is where the eye-catching numbers live, and where the compounding starts: each new thing is written on top of things written days earlier.

Weeks 23–26 are integration. Zero new repositories in three of the four. Forty-two exist and the work is making them behave as one machine — which is where kiln, glass-sdl and operandi-gui came from. Fewer commits, more repositories touched per commit, bigger consequences. It reads as a slowdown on the chart and is the opposite.

the measure that survives

Hours, from nothing to working

If commits are a bad unit, elapsed time to a verifiable capability is a good one — especially here, where almost every library is graded against somebody else’s implementation, so “working” has a date attached rather than an opinion.

Hours from a repository's first commit to a working capability 0h 24h 48h 72h 96h 120h weft 120.9h URL parser → Acid3 100/100 scribe 84.2h scaffold → shaping + hinting shuttle 53.6h skeleton → 88% of test262 cairn 23.4h read a repo → serve one over SSH webrtc-media 22.9h entropy coder → Safari decodes it natrium 8.2h hashing floor → the NaCl set reed 5.8h MP3 → the whole of Opus folio 2.2h PDF parser → graded against pdfium
Hours between a repository’s first commit and the commit that demonstrates the capability, from the git log. A PDF renderer reached a pdfium-graded harness in two hours and twelve minutes; a VP8 encoder written from RFC 6386 reached “Safari decodes our frames” in twenty-three. The two longest are the two that had no reference to lean on at the start.

These are not the durations of finished software — folio kept going for another day on JBIG2, fax decoding and CJK character maps. They are the durations of the gap between nothing and a thing that passes somebody else’s test, which is the interval where a project either has momentum or does not.

why it got faster

There was something to build on

The acceleration is not typing speed. It is that by July there was a substrate. Repositories born before 15 July depend on a mean of 1.47 other workspace libraries; those born after, 2.04. The tails carry more than the mean: weft draws on nine of them, glass on eight, cairn and loom on five each.

The clearest single case is a chain nobody planned in advance. pagetree was written in June as a crash-safe B-tree for a Bitcoin node’s UTXO set. In July cabinet became a filesystem on top of it, and cairn — a from-scratch git — learned to keep an entire repository inside a cabinet. The first link in that chain took nine hours: pagetree's first commit landed at 23:39 on a Sunday and the Bitcoin node was storing its UTXO set in it by 08:46 the next morning. Which is what lets a git implementation run on a machine with no filesystem, and why kiln can clone thirty-three repositories into a container image that contains no git binary at all.

and why it did not

The floor does not compound

Look at the copper band in the first chart again. It does not bend. modus takes 37 to 186 commits every week from April to now, through the explosion, through the integration phase, at a rate essentially independent of everything happening above it.

That is the honest limit on all of this. Breadth work got roughly an order of magnitude from oracles, fleets and a growing substrate. Depth work — a collector, a calling convention, a tag scheme, a symbol table — got about one. You cannot fan eight workers at a garbage collector, because the bug is never in eight places; it is in the interaction between two, and finding it is serial no matter how many hands are available.

The best evidence is how long wrong beliefs survived:

The shallow mistakes were caught by tooling in days. The deep one needed somebody to load real software and watch what broke, and no amount of parallelism shortens that.

what to watch

The bottleneck is the part that cannot be parallelised

Forty-one repositories compound. One does not, and it is the one underneath all of them. That is not a criticism of how the work is being done — it is the shape of the problem. But it means the useful question for the next twenty-six weeks is not “how many commits”; it is whether the floor can be made to compound too. Every build check, every oracle, every census before a deletion is an attempt at exactly that: turning a thing somebody has to remember into a thing the machine refuses to let you get wrong.