Skip to main content
← Back to Blog

2026.10.03 The model registry stops dropping models, plus a PR reviewer fix

Three model-registry PRs so new model families and their prices land automatically, the PR reviewer re-reviews pushed fixes, and admin flags failed signups and redacts tokens in bounces.

Behind the scenes stuff. Five PRs merged this week, all in the ChameleonLabs repo, and three of them are about the same problem: the model registry quietly losing models.

The model registry

Quick context. The registry is the set of feeds everything else reads when it needs a current model id or a price at runtime. If a model falls out of the feed, it falls out of everything downstream. So the feed must never silently drop a model.

Leland found models missing. The cause (#314) was on the pricing side. OpenAI's pricing page only lists most models after you click "All models", and the flagship table had moved into a different page structure than our parser expected. We were reading a truncated page. The rule now is simple: we never drop a model unless its retirement is actually verified. Missing from a page doesn't count.

Then #316. Leland's comment on an earlier issue was "we should recognize all models," and he's right. Before this, a new model family sat in an unmatched bucket until somebody edited a regex. Fable and Mythos both sat there before we hand-added them (a little embarrassing, given what we build on). Now any claude-<word>-<major>[-<minor>] is a family automatically, and on the OpenAI side any single-word tier becomes gpt-<tier>. A new family should be usable the day it ships, not the day someone notices.

Naturally the first run after that flagged something. gemini-omni-1.1-flash showed up as unpriced. Google's pricing page calls it "Gemini Omni Flash." The API display name is "Gemini Omni 1.1 Flash." Same model, two names, no match. #319 teaches the shared price lookup (used by both the pricing guard and the feed verifier) to try the exact id, then the date-stripped id, then the provider display name, and as a last resort a version-less match. Three small PRs, but together they mean the registry should keep up with the providers without us babysitting it.

Use our research free in your projects with the model-registry-client

The PR reviewer finally re-reviews

Our automated PR reviewer had a gap. Push fixes after a "request changes" verdict and... nothing. The webhook only queued reviews on opened, reopened and ready_for_review. The README promised that /review or a review-me marker would force one, but comments and labels never queued an intake, so those markers only got read when one of those three events happened to arrive. #312 (below) sat on its stale request-changes verdict because of exactly this. #313 fixes it. Push to a PR, get a fresh verdict.

Admin: failed signups and tokens in bounces

On 2026-09-26 we got two email bounces from a bot-pattern signup (an invalid, dot-padded Gmail address with a random display name). One was a verification email, one a password reset. Neither account was ever real: unverified, never signed in, no Stripe, no subscriptions, no app access. But /admin/members listed them like any other member, and the bounces we forward to ourselves carried account tokens. #312 flags failed registrations in the admin list so they're obviously not members, and redacts those tokens before a bounce gets forwarded. Account tokens don't belong in a forwarded email, period.

That's the week. If you've hit a model id our registry resolves wrong, tell us — that's exactly the kind of report that turned into three of these PRs.

Maestro AI

Hello, there