October checkpoint: what three weeks did to coding agent release velocity

Three weeks ago I measured how fast the coding agents ship, and I closed that piece with three unanswered questions. On September 27, 2026 the Codex CLI sat at 134 releases in 30 days, more than four per day, and I promised a second checkpoint with the same method plus the channel-clean count I had openly skipped. That checkpoint is due today: October 19, 2026, the same seven repositories, now with the column I owed. The short version before any number: on paper the Codex pace holds, but the channel math shows that roughly nine tenths of it flows out of the pre-release tap. Counted cleanly, the Gemini CLI drops out of the lead group because its nightly builds inflate the raw figure. And pi shipped the bundled major I asked about. A single checkpoint was never the point, the direction is.

To keep both checkpoints comparable I changed nothing about how the count works. The GitHub API scans the same seven repositories as on September 27: Codex CLI, Gemini CLI, OpenClaw, OpenCode, pi, Hermes Agent and herdr. The window is again a rolling 30 days up to the fetch, and the baseline lives in the September piece dated 27.09.2026, 07:17 CEST. What is new is the channel column: every release gets classified as stable or pre-release, derived from the tag string and GitHub's prerelease flag, so alpha, beta and release-candidate tags no longer slip in as finished versions. Commits over the last seven days again come from the Search API, the same approach as in September, with the same limit: the search spans all branches and gives a sample, not a full ledger. My fetch ran on 19.10.2026 against the public endpoints, the script sits in my vault, and anyone can rerun the same query today. I deliberately did not re-count the lifetime total of all releases, for the window comparisons the first pages of the release list are enough, and that is the honest boundary.

Every value is my own measurement taken on 19.10.2026, set against the baseline from the first checkpoint. Sorted by releases over the last 30 days, with the new column separating the stable channel from the pre-release tap. Read the commit numbers with care, they scan every branch of each repository and are therefore a sample.

The answer is yes, with a footnote. The rolling window of the last 30 days holds 129 Codex releases, the first checkpoint showed 134. The pace therefore stays in the same league at roughly four per day, the rhythm did not flatten out after the September alpha surge. The footnote lives in the channel column: only 15 of the 129 releases come from the stable channel, 114 are alpha and pre-release tags. That was exactly the gap I flagged in September, when I wrote that a channel-clean count would tell whether the daily flood consists of finished versions or of a testing ground. The answer has arrived, and it is the latter. On the stable channel Codex ships a finished tag roughly every second day, in between the alpha band runs without pause. The newest entry in the window reads rust-v0.164.0-alpha.3 from October 17, the stable tag rust-v0.163.0 landed on October 15. Anyone reading this pace as a verdict on quality repeats the September mistake: velocity measures shipping, not value.

It does, and by a wide margin. In the unfiltered list the Gemini CLI still sits near the top with 36 releases in 30 days, barely under the 38 from the first checkpoint. Counted cleanly, 4 of those are stable, the remaining 32 are nightly builds. The September question was whether the nightly channel drives the 30-day figure, and the answer is a clear yes, almost entirely. The star trail barely moves: 277 stars gained since September 27, on a base of 107,164 the smallest gain in the field. Whoever reads only the unfiltered list takes this project for one of the busiest players, whoever reads only the channel column takes it for nearly silent. Both readings are true at once, which is exactly why the unfiltered number is not wrong but hides the release culture.

pi delivered, and on a calendar. Version 1.0.0 landed on October 1, v1.1.0 followed on the 7th, v1.2.0 on the 12th, v1.3.0 on the 15th. Four bundled milestones within fourteen days, all from the stable channel, plus 131 commits in the last seven days against 97 in September. The combination of loud development and few releases did not dissolve, it professionalized: pi now publishes more often, still bundled, and without any pre-release channel. The star trail backs the story: 6,615 stars gained since September 27, the strongest gain in the field, just ahead of OpenCode with 6,551. The September question asked whether the pairing of weekly commits and sparse releases would hold, or whether a bundled major would show up. Both happened, the major shipped, and the commit trail keeps flowing.

OpenCode sticks to its rhythm: 8 releases in 30 days, all stable, newest tag v1.19.3 from October 16. The project sits in a quiet patch mode, and the stars still grew by 6,551. OpenClaw is visibly accelerating: 23 releases in 30 days after 12 in September, 15 of them stable, plus 6,208 commits in seven days per the Search API, the densest activity in the field, with the branch-counting caveat attached. Hermes Agent keeps the third release culture alive: 7 releases in 30 days, all stable, newest tag v0.22.2 from October 15, alongside 3,240 commits in seven days, bundled and curated instead of a daily stream. And herdr shows how a small tool runs two channels: 10 releases in 30 days, 4 stable, 6 pre-release, newest stable tag v0.9.5 from October 14, with 3,766 stars gained in three weeks.

My takeaway from the second checkpoint is an operating decision, not a forecast. The stack on my deck has to survive kills on Android, not prevent them, so the pi worker under herdr only moves to a new version after a review, never on release day. The snapshot from the morning of the measurement shows the operating math: 306 of 306 gateway checks alive over seven days, 2 OOM kills, 1,069 MiB of free memory, 13 jobs, the system was healthy. Velocity is an offer, stability is a decision, and the gates between draft and publication stay part of my stack for exactly that reason. In the table the strongest star gain sits with pi, the second strongest with OpenCode, and the commit trails show continuous work everywhere. For my daily operation a different column matters: how much new stability and how much new risk rides along with every version I pull. No release counter can answer that, only the running operation can.

The limits, explicitly. The channel classification comes from the tag string and the prerelease flag, not from the intent of the project, a nightly build without a flag would wrongly fall into my stable column. The commit numbers from the Search API span all branches, a comparison between two checkpoints can therefore be shaped by a changed branch structure, the values remain a sample. I did not re-fetch the lifetime total of releases today, the window comparisons draw from the first pages of the release list, which is enough for values since September 26. An unauthenticated fetch runs into rate limits, my script therefore paused between requests. Stars over a single day would be a sideshow, which is why these are three-week windows, against the same first-checkpoint bases for all seven repositories. And two checkpoints turn two numbers into a direction, not a line: a third checkpoint would be the next measurement, not the proof of a trend.

I am Marcel, a graphic designer running the HUNTER cyberdeck on a second-hand Google Pixel 6a. The release, commit and star figures in this piece I pulled myself from the GitHub API on 19.10.2026, the baselines come from my measurement of 27.09.2026, 07:17 CEST. The operations snapshot from measurement day, 07:00, lives in my vault under stability-snapshots. Editorial responsibility: Marcel. Last fact-checked on 19.10.2026.