[chrarnoldus]

Summary

OpenCode Go serves its Gemini models only over Google's native Gemini API. The gateway can only proxy chat_completions, messages and responses (GatewayChatApiKind), so we listed opencode-go/gemini-4-argon-contributor in the catalog even though every request to it failed.

The models.dev sync for direct BYOK providers now drops any model whose per-model provider.npm is a native Gemini API SDK (@ai-sdk/google or @ai-sdk/google-vertex). All other models are kept.

Evidence

models.dev (https://models.dev/api.json, provider opencode-go, fetched 2026-10-16):

• Provider default: npm: "@ai-sdk/openai-compatible", api: "https://opencode.ai/inference/go/openai/v1".
• gemini-4-argon-contributor: provider: { npm: "@ai-sdk/google", api: "https://opencode.ai/inference/go/google/v1beta" }.
• Every other per-model override uses @ai-sdk/anthropic (minimax, qwen) or @ai-sdk/openai (gpt, grok, muse), or overrides only api.
• https://opencode.ai/zen/go/v1/models lists gemini-4-argon-contributor, so the available-models check doesn't filter it out. The SDK field is the only thing in the data that marks it as unusable.

Production: the Axiom vercel dataset shows ModelProtocolUnsupported ("Model does not support this protocol.") errors from OpenCode Go for opencode-go/gemini-4-argon-contributor on /api/v1/chat/completions.

Why a Gemini SDK denylist, not a model-ID filter or an SDK allowlist

• Not a model-ID filter: the SDK field describes the real constraint, a protocol we can't proxy. It covers future Gemini models whatever their name. It also keeps a Gemini model that is later served over an OpenAI-compatible endpoint.
• Not an allowlist: the first version of this PR kept only @ai-sdk/openai-compatible, @ai-sdk/openai and @ai-sdk/anthropic. That would wrongly hide working models, because several SDK names still speak Chat Completions. Per-model overrides found across all of models.dev:
  • opencode (OpenCode Zen) has a model with npm: "@ai-sdk/mistral" on https://opencode.ai/inference/openai/v1, its OpenAI-compatible endpoint.
  • standardcompute has a model with npm: "@openrouter/ai-sdk-provider".
  • Provider defaults include more Chat Completions-compatible SDKs, such as @ai-sdk/xai, @ai-sdk/groq, @ai-sdk/deepinfra, @ai-sdk/togetherai and @ai-sdk/cerebras.
• Failure mode: an unknown native SDK still gets listed, which is how things work today. A Chat Completions-compatible model is never silently hidden.
• Scope: the filter applies to every models.dev-backed provider, which is safe. None of nvidia, alibaba-token-plan, edenai, zai-coding-plan, ollama-cloud, xiaomi-token-plan-ams or xiaomi-token-plan-sgp has per-model provider overrides. In the current catalog the filter removes exactly one model: opencode-go/gemini-4-argon-contributor.
• Malformed data: the schema field uses .catch(undefined). If the field is malformed, the model is kept.

Already-synced data

No schema change is needed. syncDirectByokModels runs from /api/cron/sync-providers every 5 minutes, and each successful sync fully replaces the provider's direct_byok_model_lists row with onConflictDoUpdate. The first sync after deploy removes the stale Gemini entry. After that, gateway instances stop serving it once their 10-minute cachedEnhancedDirectByokModelList cache expires. A new DB test seeds a stale opencode-go row that contains the Gemini model and checks that a re-sync removes it.

Request behaviour after the change

getDirectByokModel no longer resolves opencode-go/gemini-4-argon-contributor, so the gateway doesn't send it to OpenCode Go. A user with credits gets the gateway's own 404 model_not_found ("The requested model '…' does not exist…") with no upstream call. A personal account or organization with a zero balance hits the credits check before that 404. This is existing behaviour for any model that drops out of a direct BYOK list, and this PR doesn't change it.

Verification

• jest src/lib/providers/direct-byok/sync-direct-byok.test.ts ../../packages/web-shared/src/lib/ai-gateway/providers/direct-byok (from apps/web, NODE_ENV=test): 8 suites, 87 tests passed.
• Against main's sync-direct-byok.ts, the 2 new tests fail. Against the earlier allowlist version, the parse test fails because the @ai-sdk/mistral and @openrouter/ai-sdk-provider models are dropped. With this change, both tests pass.
• pnpm --filter web typecheck: passed.
• pnpm --filter @kilocode/web-shared typecheck: passed. The package isn't changed; this was run as a sanity check.
• pnpm exec oxlint --config .oxlintrc.json on both changed files: 0 warnings, 0 errors.
• pnpm exec oxfmt on both changed files, then git diff --check: clean.

[kilo-code-bot]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

Executive Summary

The incremental change replaces the SDK allowlist in parseModelsDevProviderModels with a native-Gemini denylist (@ai-sdk/google, @ai-sdk/google-vertex); the updated filter, the .catch(undefined) schema hardening, and the revised test assertions are correct and consistent with the changed code, and I found no new issues.

Overview

Incremental review at commit 57ec27850f9f319ba4f19de772bb5de6ea8fdc1e (since e270c5ff). The denylist keeps Chat Completions-compatible SDKs such as @ai-sdk/mistral and @openrouter/ai-sdk-provider (verified by the updated test) while excluding the per-model Gemini overrides that the gateway cannot proxy. The new test inputs cover both a model that must be excluded (vertex-gemini) and aliases that must be kept.

Notes (non-blocking, not inline findings):

• Switching from an allowlist to a denylist means native SDKs other than Gemini (e.g. @ai-sdk/amazon-bedrock) are no longer excluded. No current modelsDevFetcher provider has such a per-model override in the models.dev catalog, so this does not affect live behavior today; a future non-Gemini native-API provider would need a matching entry in GEMINI_API_MODELS_DEV_SDKS.
• The filter inspects only the per-model provider.npm override, not the provider-level default. This matches the current FETCHERS set and is documented in the PR.

Files Reviewed (2 files)

• apps/web/src/lib/providers/direct-byok/sync-direct-byok.ts
• apps/web/src/lib/providers/direct-byok/sync-direct-byok.test.ts

Previous Review Summary (commit e270c5f)

Current summary above is authoritative. Previous snapshots are kept for context only.

Previous review (commit e270c5f)

Status: No Issues Found | Recommendation: Merge

Executive Summary

The change adds a provider.npm SDK filter to parseModelsDevProviderModels, excluding models the gateway cannot proxy (e.g. OpenCode Go's Gemini); the logic, schema hardening via .catch(undefined), and added tests are correct and I found no issues in the changed code.

Files Reviewed (2 files)

• apps/web/src/lib/providers/direct-byok/sync-direct-byok.ts
• apps/web/src/lib/providers/direct-byok/sync-direct-byok.test.ts

Notes (non-blocking, not inline findings):

• The new filter only inspects the per-model provider.npm override, not the provider-level default. This is safe for the current FETCHERS set as documented in the PR, but any future modelsDevFetcher whose provider default is a non-proxyable SDK (e.g. @ai-sdk/google, @ai-sdk/amazon-bedrock) would keep its models. Consider a provider-level default check if such a provider is added.
• Allowed set (@ai-sdk/openai-compatible, @ai-sdk/openai, @ai-sdk/anthropic) matches the gateway's GatewayChatApiKind protocols and opencode-go's declared supported_chat_apis; @ai-sdk/amazon-bedrock is correctly excluded.

---

_{Reviewed by deepseek-v4.1-flash · Input: 36.8K · Output: 5.4K · Cached: 276.5K}

_{Review guidance: REVIEW.md from base branch main}