[viztor]

The Go plane's model list was being answered with the Zen catalog. Three places ask which plane a URL belongs to, and they had three answers — because the newer inference endpoint family was added to each of them separately, and each addition was partial.

path	the old substring checks said	is
/inference/go/openai/v1	zen (two of the three knew only /zen/go)	go
/inference/google/v1beta	go (/inference/go is a prefix of it)	zen
/inference/openai/v1	unknown (/inference/v1 is not inside it)	zen

The first row is the reported symptom

The Go plane's base is the inference family now, and the adapter lists its models at ${base}/models — a path the gateway does not serve, so the answer comes from this plugin's enrichment. That enrichment asks which plane the listing is on, read by substring, and it read the modern Go path as the Zen plane.

So the Go plane was answered with the Zen catalog: every Zen model, and none of the Go plane's own — the Gemini contributor and the Muse contributor, both of which the deployment declares.

Measured against the real gateway, through the patched fetch:

listing	before	after
…/zen/go/v1/models	36 Go models	36 Go models
…/inference/go/openai/v1/models	82 Zen models	28 Go models

The other two rows are credentials filed under the wrong account

• The Zen google family's captured key was filed as a Go key, and used as one. A key filed under the wrong account is worse than one filed under none: it is used, and it is the wrong one.
• The Zen openai family's captured key was filed under no account at all, which is the "my key is invalid" report on a key that was fine — the same report this plugin has already been asked about once.

The fix

One predicate answers it now, in src/utils/request-shape.ts, and the three call sites use it: isGoModelsListingUrl, resolveDiscoveryProvider and tierForRequest.

The segment boundary is load-bearing — /inference/go must end there, or the Zen google family is read as the Go plane. A relative path answers too (/zen/go/v1/models), because a caller with no origin passes a path and the plane is a property of the path; a string that is neither a URL nor a path is still not guessed at.

The Go plane's other traffic is unaffected — and measured

endpoint	answer
POST …/inference/go/openai/v1/chat/completions	200, a real completion
POST …/inference/go/google/v1beta/…:generateContent	200, a real completion
GET …/zen/go/v1/usage	200, the quota panel's endpoint

The usage endpoint stays on the older family on purpose: the vendor's inference proxy has no /inference/…/usage counterpart, which is why toGoBaseURL rewrites an inference-family base to https://opencode.ai/zen/go/v1.

Verification

gate	result
vp check	0 errors, 23 pre-existing warnings
suite	1155 passed / 73 files (+5 tests)
coverage (95.9 / 91.3 / 95.9 / 95.9)	96.32 / 91.92 / 96.33 / 96.31
scripts/check.ts	passes

Each of the three old behaviours was planted back and watched fail the new tests — the listing predicate (expected false to be true: the Go-only Gemini model missing), the tier classifier, and the path predicate.

Also in this branch

AGENTS.md records the merge policy: squash, so one pull request is one commit on main, one line in the changelog, and the pull request's title is the release signal.