GitHub releases of the coding agents: the second cutoff in October

On September 27 I measured the release culture of seven coding agents through the GitHub API and left three questions open for October: does Codex keep up its pace of roughly four releases per day, does the Gemini CLI drop out of the top group once you count stable channels only, and does pi deliver the bundled major it gestured toward? Today, October 14, 2026 at 05:07 UTC, the second cutoff is in, with the same protocol, the same seven repositories, the same counter. The short answer first: Codex holds the cadence, pi overdelivers the prediction, and since October 5 OpenCode has been telling a v2 story that simply does not exist inside the releases API. Anyone who reads a single GitHub table misses half the truth about the major upgrade, and that half is the most interesting part. In September one cutoff was a snapshot with a warning label attached; now there are two points, and two points give you a direction as long as you read it coarsely.

The measurement ran against the same seven repositories as in September: pi, OpenCode, Hermes Agent, Codex, Gemini CLI, OpenClaw and herdr. Stars come from the repo endpoint, the release window from the releases list at one hundred entries per page, four of those pages scanned for Codex, and the commit week uses the same search query shape as the September runs, committer-date greater than or equal to October 7. The cutoff is October 14, 2026, 05:07 UTC. One conscious deviation stays on the record: the new pi path badlogic/pi-mono refuses to answer the commit question and returns an API error instead of a count, so I computed it through the old path earendil-works/pi, which GitHub redirects cleanly, and the number carries the older path in the protocol. The second boundary case is a definition: a git tag is not a release. The releases API knows release objects with changelog fields, the atom feed knows every tag reference. Confuse the two and you count two different things while believing they are one number.

The three answers in plain text. Codex stays in continuous flow: 149 releases in the 30-day window is just under five per day, and against the September stock of 1,158 it is 67 releases in 17 days, just under four per day, so the cadence holds on both sides of the line. pi overdelivers the prediction: the 1.0 series has stood since early October, v1.1.0 landed on the 7th, v1.2.0 on the 12th, commits moved from 97 to 124 per week, and the star curve paid in with plus 2,715 over eleven days. For the Gemini CLI the stable-channel answer is sobering: of 38 releases in the window only 5 carry no pre-release marker, the current stable v0.64.0 appeared on October 13. All three September questions thus have answers, and the sharpest one came from a repository that sat in none of the three questions.

This is where comparability with September ends, which is why this section carries the core finding. OpenCode had hidden the major behind an upgrade invocation in the v1.19.0 stub, the major upgrade is deliberately held at the command line, and my October 12 radar note reported exactly that. Between October 5 and 13 there followed v2.0.23, v2.0.24, v2.0.25, v2.0.26, v2.0.27 and v2.0.28, six v2 tags in nine days, punctuated by an administrative note on the 10th about the dev-channel promotion. All six appear in the atom feed, all six tag pages answer with status 200, and the releases API registers none of them: a query per tag name ends in 404, the top of the list still shows v1.19.0 from October 11. Two lessons in one finding. First, the September article split two curves, stars and releases; a third curve now exists, the tag ladder, and it runs only in the feed. Second, every claim about release cadence is in the end a claim about the endpoint it was measured on. The v2 development is real, visible and fast, and it sits literally below the threshold where most statistics machines start counting.

The commit column shows how much work falls into the windows: 2,145 commits in seven days for Hermes, 4,254 for OpenClaw, 262 for the OpenCode sprint toward v2. No release note in the world translates that number into something readable, and the quote above is my best evidence: a skill an agent writes for itself never appears in a release log, it grows between the tags. Which is why my deck stays on bundled versions, pi pinned to the 1.2 series, Hermes in its bundled rhythm, because an operation on an Android phone has to survive kills rather than chase updates. The September verdict still stands in October: velocity is an offer, stability remains a decision. What the table delivers instead is a feel for time: having read these seven weeks closely, you know at month's start which versions existed and what risk every upgrade would have carried.

The star drift between October 3 and 14 reads like a cross-check of the release curve: Hermes gains plus 2,536, pi plus 2,715, herdr plus 1,752, OpenCode plus 1,648, Codex plus 1,311, and then the striking part: OpenClaw commits loudest at 4,254 commits in seven days and collects only plus 673 stars, the Gemini CLI stays nearly flat at plus 105 despite its nightly cadence. Attention, apparently, follows visible leaps rather than sheer frequency: pi's 1.0 series, OpenCode's v2 teaser, three Hermes releases within seven days. Whoever reads stars as the only number confuses observers with users, whoever reads releases as the only number confuses delivery with significance. The September article closed with the reminder to read both curves side by side; October 14 adds the third.

A glance at my own toolbox rounds out the curve: herdr gained plus 1,752 stars since October 3, on 35 commits in the last seven days and a stable v0.9.4 from October 13, a quiet bundled cadence after the same pattern as pi, twelve releases in the window, four of them stable. In my operation the pi execution runs through herdr, and that combination is exactly what produced the 36 of 36 tests in my verified projects: Hermes plans, pi executes, Hermes verifies. None of these numbers alone says anything about the quality of a tool, but together they repeat a pattern that has been visible since September: projects with bundled states and visible leaps collect observers, projects in continuous flow mainly collect commits.

The limits, on the record. The commit query answered 422 twice, once because of percent encoding in the qualifier, once on the moved pi path; the valid measurement runs over the raw qualifier form as in September, the moved path stays locked out, and I write down these details because they are exactly what separates reproduction from belief. The Codex window value of 149 comes from a scan over 400 release entries across four pages, alpha tags do not count toward the stable figure. Tags are not releases, the OpenCode v2 ladder therefore counts as a release in none of these values. And two cutoffs give a coarse direction with markers, not a trend line; that caveat from September has not gotten softer. Three questions lead to the third cutoff: does the releases API pick up the OpenCode v2 series once it turns official, does pi slow back toward the September cadence after the 1.2 batch, and is the Hermes gap of three releases in seven days an exception or a series. The next cutoff is due in November, same protocol, same list.

I am Marcel, a graphic designer and the operator of the HUNTER cyberdeck built on a used Google Pixel 6a. I pulled every figure in this article myself from the GitHub API between 05:07 and 05:08 UTC on October 14, 2026; the comparison values sit in my measurements of Sep 26/27, 2026 and the Oct 3, 2026 baseline. The operational anchors live in my own vault, the release notes of Oct 11 and 12 in the topic radar. Last reviewed on October 14, 2026.

Own measurement of Oct 14, 2026, 05:04 to 05:08 UTC, through the GitHub API: repo endpoints for stars and move status, releases endpoints for the 30-day window (Codex scanned across four pages, 400 entries), releases atom feeds and tag queries for the OpenCode v2 ladder, search API for the commit week with the September protocol form. Comparison base: my own measurement of Sep 26, 2026, 07:03 UTC, of Sep 27, 2026, 07:17 CEST, and the Oct 3, 2026 baseline. Release pages: pi v1.2.0 (Oct 12), Hermes Agent v0.22.0 (Oct 14), OpenCode v1.19.0 (Oct 11) and the v2 tag pages, Codex rust-v0.163.0-alpha.11 (Oct 14), herdr v0.9.4 (Oct 13), all GitHub. Quote: ADIPod, report on the Hermes Agent, read as the full English text on Oct 14, 2026. Radar notes of Oct 11 and 12, 2026 in my own topic radar as the background of the v1.19.0 reading.