[Feature]: config-level retainModels allow-list so operators keep configured models in the authoritative live catalog

Problem to solve

When a provider does live model discovery, the authoritative catalog drops configured models that the upstream does not (yet) return — even when the operator explicitly configured them and the wire mapping exists in code. Concretely:

google-antigravity (Cloud Code Assist) does not return gemini-3.7-flash from :fetchAvailableModels for accounts where the model is not yet provisioned, so the picker loses the model even though the operator configured it (see [Bug]: v2.15.1 drops gemini-3.7-flash from Antigravity live discovery on macOS #1683 for the user-facing symptom).
Today the only retention paths are: the hardcoded CALLABLE_CONFIGURED_COMPATIBILITY_MODELS allow-list (kimi, xai), combo targets, and the Vertex default seed. There is no operator-facing, config-level way to say "keep this configured model in the catalog even if discovery omits it."

Operators who run proxies for their own account know their provider's roadmap (e.g. "gemini-3.7-flash is coming to my region; keep the row so it appears the moment the upstream enables it") and should not need a code change per model.

Proposed solution

Add a config-level retention list per provider: providers.<name>.retainModels: string[].

{
  "providers": {
    "google-antigravity": {
      "retainModels": ["gemini-3.7-flash"],
      "modelCosts": {
        "gemini-3.7-flash": { "input": 0.75, "output": 3.75, "cacheRead": 0.15, "cacheWrite": 0 }
      }
    }
  }
}

In mergeConfiguredModelsIntoLiveCatalog (src/codex/catalog/provider-fetch.ts), extend the retention condition:

if (
  seedVertexDefault === true
  || shouldRetainConfiguredProviderModel(name, candidate.id)
+ || (retainConfiguredModelIds?.has(candidate.id) === true)
+ || (providerRetainModels?.has(candidate.id) === true)
  || (retainComboTargets && retainConfiguredModelIds?.has(candidate.id) === true)
) {
  out.push(candidate);
  present.add(candidate.id);
  continue;
}

When a retained model is invoked but the upstream returns 404/not-found, surface a clear diagnostic (log + warnDroppedConfiguredIdsOnce-style one-shot warning) so a model that "exists in the picker but is not callable yet" is not silently hidden — the operator sees exactly why.
Seed CALLABLE_CONFIGURED_COMPATIBILITY_MODELS with google-antigravity: ["gemini-3.7-flash"] as the first entry, so the picker row exists out-of-the-box for antigravity users while the upstream catches up (mirrors the existing kimi/xai pattern).

Alternatives considered

Keep the hardcoded allow-list only: works but requires a code change + release per model/per provider; does not scale to operator-specific account states.
Make discovery "never drop configured models": would resurrect stale/broken models for everyone (the reason the authoritative catalog exists); the retention must be explicit and opt-in.
Document "wait for upstream": not actionable for operators who know their model is coming; the row should exist with a clear "not callable yet" state.

Additional context

Related open issue: [Bug]: v2.15.1 drops gemini-3.7-flash from Antigravity live discovery on macOS #1683 (gemini-3.7-flash dropped from Antigravity live discovery).
Precedent for static + live coexistence: other gateways keep a static model list alongside live discovery (e.g. OmniRoute QODER_STATIC_MODELS + live --list-models fallback in open-sse/services/qoderCli.ts).
The existing warnDroppedConfiguredIdsOnce infrastructure already logs dropped configured ids once per signature — the retention path should reuse it for the inverse case (retained-but-404) so logs stay quiet in steady state.

Acceptance criteria:

retainModels entry absent from discovery → row present in catalog (visibility list).
Invoking a retained-but-unavailable model → 4xx from upstream is surfaced with a one-shot diagnostic log; the catalog row stays.
Without retainModels → current behavior unchanged (no behavior change for existing users).

⚠️ Possible Duplicate Issue(s)

[Bug]: v2.15.1 drops gemini-3.7-flash from Antigravity live discovery on macOS #1683

🔗 Related PRs

#354 - fix(models): surface discovery failure status when /v1/models returns 401 (#329) [open]
#652 - feat(providers): add bounded model discovery contract [merged]
#963 - fix(catalog): preserve provider capabilities on custom models [open]
#1331 - fix(catalog): keep failover combos when live discovery omits members [merged]
#1658 - fix(antigravity): discover live tiered models from CCA [merged]

📝 Issue Planner

_{Check the box below or use the @coderabbitai plan command to generate an implementation plan and prompts that you can use with your favorite coding assistant.}

Create Plan

🧪 Issue enrichment is currently in open beta.

To disable automatic issue enrichment, add the following to your .coderabbit.yaml:

issue_enrichment:
  auto_enrich:
    enabled: false

💬 Have feedback or questions? Drop into our discord!

리뷰 · 우선순위 42 / 80

라이브 디스커버리가 아직 안 돌려주는 설정 모델을 운영자가 카탈로그에 남겨 두고 싶다는 요청입니다. google-antigravity의 gemini-3.7-flash처럼 계정에 아직 프로비저닝되지 않은 모델이 피커에서 사라지는 증상(#1683)이 배경입니다. 모든 설정 모델을 무조건 남기면 깨진 줄이 다시 살아나서, 명시적 허용 목록이 필요합니다.

현재 dev의 src/codex/catalog/provider-fetch.ts를 보면 CALLABLE_CONFIGURED_COMPATIBILITY_MODELS는 kimi와 xai만 들고 있습니다. shouldRetainConfiguredProviderModel()도 그 하드코드와 opencode-free 규칙만 봅니다. mergeConfiguredModelsIntoLiveCatalog()는 Vertex 시드, 위 함수, 콤보 타깃만 남기고 나머지는 droppedConfiguredIds로 보냅니다. providers.*.retainModels 필드는 이 조건에 없습니다.

이슈가 제안한 설정 형태는 운영자 허용 목록을 코드 변경 없이 넣는 것입니다. 유지된 모델을 호출했는데 업스트림이 404면, 피커에서 지우지 말고 warnDroppedConfiguredIdsOnce 같은 한 번만 찍는 진단이 필요합니다. 설정이 없으면 지금 동작을 바꾸면 안 됩니다.

해결방안: mergeConfiguredModelsIntoLiveCatalog 유지 조건에 설정 retainModels를 더하고, 404는 한 줄 진단으로 남기면 됩니다. antigravity 시드를 하드코드에 넣는 일은 제품 판단이지만, 설정 경로가 있으면 모델마다 릴리스할 필요는 없습니다. 디스커버리에서 무조건 설정 모델을 남기는 패치는 이 이슈와 맞지 않습니다.

이 댓글은 grok-bot이 작성했습니다

리뷰 · 우선순위 42 / 80 · 유지

라이브 디스커버리가 모델을 안 주면 피커에서 빠짐. 지금 남는 길은 하드코딩 CALLABLE_CONFIGURED_COMPATIBILITY_MODELS(kimi/xai), 콤보 타겟, Vertex 시드뿐임. 설정으로 retainModels 주고 남겨 달라는 요청. 증상은 #1683.

유지함. 전부 남기면 죽은 모델이 다시 살아나서, opt-in만. 호출 404는 한 번만 경고하고 행은 유지.

해결방안: providers.<name>.retainModels + mergeConfiguredModelsIntoLiveCatalog 조건 추가. 기본 동작은 그대로.

이 댓글은 grok-bot이 작성했습니다

fix(review): retained-model 404 diagnostic, cursor retention, whitesp…

feat(catalog): config-level retainModels allowlist for authoritative …

fix(review): retained-model 404 diagnostic, cursor retention, whitesp…

feat(catalog): config-level retainModels allowlist for authoritative …

fix(review): retained-model 404 diagnostic, cursor retention, whitesp…

fix(catalog): clear retained model diagnostic memos across generation…

fix(catalog): preserve retained diagnostics across multi-pass gathers…

feat(catalog): config-level retainModels allowlist for authoritative …

fix(review): retained-model 404 diagnostic, cursor retention, whitesp…

fix(catalog): initialize cache generation before observed callback an…

fix(catalog): address review feedback on retainModels state preservat…

docs(types): align retainModels JSDoc with independent seeding contra…

feat(catalog): honor per-provider retainModels opt-in (closes lidge-j…

fix(catalog): recognize YYMMDD/MMDD/YYMM dated model id suffixes (lid…

fix(catalog): recognize YYMMDD/MMDD/YYMM dated model id suffixes (#3024)

test(catalog): guard the dated fold's direction at the merge loop, no…

fix(catalog): widen the dated-variant fold and guard its direction at…

feat(catalog): per-provider retainModels opt-in (#1690, carries #2860) (

Landed on dev via #3206 (squash 6a6efa928), carrying #2860's retention predicate and #2122's config/migration design.

What shipped:

providers.<name>.retainModels: string[] keeps listed ids in the routed catalog when live discovery omits them. Ids do not need to be repeated in models; they get the same context/effort hints and survive liveModels: false.
Validated at config load and through the management PATCH (null/empty clears); echoed in the provider DTO.
CLI: ocx provider edit <name> --retain-models <id,id|->.
Rename migration covers the new list.
Default behavior is unchanged when the key is absent.

Not included (deliberately): seeding gemini-3.7-flash as an antigravity product default (with the config key available this no longer needs a release), a dashboard field (the PATCH contract is ready for one), and a 404-time warning. selectedModels still narrows what is visible, so an id must be in both lists when an allowlist is active.

fix(catalog): widen the dated-variant fold and guard its direction at…

feat(catalog): per-provider retainModels opt-in (lidge-jun#1690, carr…

fix(catalog): widen the dated-variant fold and guard its direction at…

feat(catalog): per-provider retainModels opt-in (lidge-jun#1690, carr…

Navigation Menu

Repository navigation

[Feature]: config-level retainModels allow-list so operators keep configured models in the authoritative live catalog #1690

Problem to solve

Proposed solution

Alternatives considered

Additional context

리뷰 · 우선순위 42 / 80

리뷰 · 우선순위 42 / 80 · 유지

68 remaining items

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

[Feature]: config-level retainModels allow-list so operators keep configured models in the authoritative live catalog #1690

Description

Problem to solve

Proposed solution

Alternatives considered

Additional context

Activity

coderabbitai commented on Aug 14, 2026

lidge-jun commented on Aug 19, 2026

리뷰 · 우선순위 42 / 80

lidge-jun commented on Aug 19, 2026

리뷰 · 우선순위 42 / 80 · 유지

68 remaining items

lidge-jun commented on Sep 1, 2026

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions