[adivekar-utexas]

Summary

• The targetFormat field (added in [BUG] opencode-go: custom models added via UI use wrong targetFormat (oa-compat) — custom model save appears to succeed but model still routes as OpenAI-compatible #2905) was fully wired through the Zod schema, the /api/provider-models POST/PUT routes, the DB layer (addCustomModel / updateCustomModel), and the routing engine (getModelTargetFormat in src/sse/handlers/chatHelpers.ts) — but had no UI control anywhere in the providers dashboard.
• This made the documented use case (a custom opencode-go model that must use the Anthropic Messages shape) unreachable from the UI. Users had to call the API directly or patch the DB.
• Add a "Target Format" select dropdown next to the existing "API Format" selector in both the add-custom-model form and the per-model edit form. Includes a "Default (auto)" option (no override) plus the six targetFormat enum values: openai, openai-responses, claude, gemini, gemini-cli, antigravity. A small badge on each model row shows the configured target format when set.

Why this matters

The Compatibility popover on each model row already shows a "Client Request Protocol" dropdown with values OpenAI Chat Completions / OpenAI Responses API / Anthropic Messages. That dropdown is a view selector for the per-protocol compat toggles underneath it — it is not a persisted setting, and selecting "Anthropic Messages" there does not route the model through /v1/messages. That routing override lives in the separate targetFormat field, which had no UI exposure until this PR.

After this PR, a user who wants opencode-go/glm-5.2-max (or any other custom model on any provider) to use the Anthropic Messages wire format can:

1. Dashboard → Providers → opencode-go → Add Custom Model (or edit existing)
2. Set Target Format = Anthropic Messages
3. Save

The model then routes through /v1/messages with thinking.type: "enabled" and effort fields passed through, instead of being rejected by the OpenAI chat completions endpoint with Extra inputs are not permitted, field: 'effort'.

Files changed

• src/app/(dashboard)/dashboard/providers/[id]/components/CustomModelsSection.tsx — add editingTargetFormat / newTargetFormat state; wire into handleAdd, beginEdit, cancelEdit, saveEdit; render select dropdown in both add form and edit form; render badge on model row when targetFormat is set
• src/app/(dashboard)/dashboard/providers/[id]/providerPageHelpers.ts — add targetFormat?: string to CompatModelRow type
• src/i18n/messages/en.json — add 6 new keys: targetFormatLabel, targetFormatHint, targetFormatAuto, targetFormatGemini, targetFormatGeminiCli, targetFormatAntigravity. (Reuses existing compatProtocolOpenAI / compatProtocolOpenAIResponses / compatProtocolClaude labels for the OpenAI/Claude options.) Other locales fall back to English via deepMergeFallback in src/i18n/request.ts — no per-locale changes required.

Test plan

• Existing targetFormat + provider-models + i18n test suites all pass (153/153):
  • tests/unit/custom-model-target-format.test.ts — backend persistence of targetFormat (unchanged, still passes)
  • tests/unit/provider-models-route.test.ts — POST/PUT route handling
  • tests/unit/provider-models-config.test.ts, provider-models-management-route.test.ts, provider-models-token-limits.test.ts, glm-provider-model-import-route.test.ts, provider-models-v1-route.test.ts
  • tests/unit/i18n-fallback.test.ts, i18n-cli-namespaces.test.ts, i18n-nest-dotted-keys.test.ts, i18n-pt-br.test.ts, i18n-quota-share-english.test.ts, audit-eventtype-i18n.test.ts
• Pre-commit hooks pass (prettier, eslint, doc-sync, any-budget, tracked-artifacts)
• No new lint errors

Manual verification

1. Open Dashboard → Providers → opencode-go → Add Custom Model
2. Set Model ID = glm-5.2-max, Target Format = Anthropic Messages, Save
3. Send a request to opencode-go/glm-5.2-max
4. Verify in the dashboard logs that the provider request URL is https://opencode.ai/zen/go/v1/messages (not /v1/chat/completions)
5. Verify the request body contains thinking.type: "enabled" and effort fields
6. Verify the response is 200 (not 400 with Extra inputs are not permitted, field: 'effort')

Notes

• The non-EN locale files are not updated in this PR. They will display English text for the new labels until the maintainers run npm run i18n:sync-ui to propagate placeholders. The runtime fallback in src/i18n/request.ts:122 (deepMergeFallback) ensures no missing-key errors.

Made with Cursor

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!