6a0d7ca Route mixed-API models from catalog metadata · zut

M packages/agent/build.go +15 -14


@@ -758,10 +758,11 @@ func (r Resolved) NewClient() provider.Client {
758	758		case "deepseek":
759	759		return wrap(provider.NewDeepSeek(r.Credential, r.BaseURL))
760	760		case "openai":
761		-	if strings.HasPrefix(r.Model, "gpt-5.6-") {
762		-	return wrap(provider.NewOpenAIResponsesNamed(r.Credential, r.BaseURL, "openai"))
763		-	}
764		-	return wrap(provider.NewOpenAI(r.Credential, r.BaseURL))
	761	+	return provider.NewModelRouter("openai",
	762	+	wrap(provider.NewOpenAI(r.Credential, r.BaseURL)),
	763	+	map[string]provider.Client{
	764	+	provider.APIResponses: wrap(provider.NewOpenAIResponsesNamed(r.Credential, r.BaseURL, "openai")),
	765	+	})
765	766		case "openai-codex":
766	767		inner := wrap(provider.NewOpenAICodex(r.Credential, r.AccountID, r.BaseURL))
767	768		return r.wrapWithRefresh(inner)

@@ -777,12 +778,7 @@ func (r Resolved) NewClient() provider.Client {
777	778		case "groq":
778	779		return wrap(provider.NewGroq(r.Credential, r.BaseURL))
779	780		case "xai":
780		-	var inner provider.Client
781		-	if r.Model == "grok-4.5" {
782		-	inner = wrap(provider.NewOpenAIResponsesNamed(r.Credential, r.BaseURL, "xai"))
783		-	} else {
784		-	inner = wrap(provider.NewXAI(r.Credential, r.BaseURL))
785		-	}
	781	+	inner := r.newXAIClient(r.Credential)
786	782		if r.AuthMethod == "oauth" {
787	783		return r.wrapWithRefresh(inner)
788	784		}

@@ -854,6 +850,14 @@ func (r Resolved) withHTTPClient(c provider.Client) provider.Client {
854	850		return provider.WithHTTPClient(c, provider.NewHTTPClient(true))
855	851		}
856	852
	853	+	func (r Resolved) newXAIClient(token string) provider.Client {
	854	+	return provider.NewModelRouter("xai",
	855	+	r.withHTTPClient(provider.NewXAI(token, r.BaseURL)),
	856	+	map[string]provider.Client{
	857	+	provider.APIResponses: r.withHTTPClient(provider.NewOpenAIResponsesNamed(token, r.BaseURL, "xai")),
	858	+	})
	859	+	}
	860	+
857	861		// wrapWithRefresh wraps an OAuth client so the access token is
858	862		// refreshed automatically before each API call. Without this, long
859	863		// sessions (hours) silently fail when the 1-hour token expires.

@@ -882,10 +886,7 @@ func (r Resolved) wrapWithRefresh(inner provider.Client) provider.Client {
882	886		// anthropic-messages on api.kimi.com/coding.
883	887		return r.withHTTPClient(provider.NewKimiCodingWithHeaders(token, baseURL, kimiCodeHeaders()))
884	888		case "xai":
885		-	if r.Model == "grok-4.5" {
886		-	return r.withHTTPClient(provider.NewOpenAIResponsesNamed(token, baseURL, "xai"))
887		-	}
888		-	return r.withHTTPClient(provider.NewXAI(token, baseURL))
	889	+	return r.newXAIClient(token)
889	890		default:
890	891		return r.withHTTPClient(provider.NewAnthropicOAuth(token, baseURL))
891	892		}

 	case "deepseek":
 		return wrap(provider.NewDeepSeek(r.Credential, r.BaseURL))
 	case "openai":
-		if strings.HasPrefix(r.Model, "gpt-5.6-") {
+		return provider.NewModelRouter("openai",
-			return wrap(provider.NewOpenAIResponsesNamed(r.Credential, r.BaseURL, "openai"))
+			wrap(provider.NewOpenAI(r.Credential, r.BaseURL)),
-		}
+			map[string]provider.Client{
-		return wrap(provider.NewOpenAI(r.Credential, r.BaseURL))
+				provider.APIResponses: wrap(provider.NewOpenAIResponsesNamed(r.Credential, r.BaseURL, "openai")),
+			})
 	case "openai-codex":
 		inner := wrap(provider.NewOpenAICodex(r.Credential, r.AccountID, r.BaseURL))
 		return r.wrapWithRefresh(inner)
 	case "groq":
 		return wrap(provider.NewGroq(r.Credential, r.BaseURL))
 	case "xai":
-		var inner provider.Client
+		inner := r.newXAIClient(r.Credential)
-		if r.Model == "grok-4.5" {
-			inner = wrap(provider.NewOpenAIResponsesNamed(r.Credential, r.BaseURL, "xai"))
-		} else {
-			inner = wrap(provider.NewXAI(r.Credential, r.BaseURL))
-		}
 		if r.AuthMethod == "oauth" {
 			return r.wrapWithRefresh(inner)
 		}
 	return provider.WithHTTPClient(c, provider.NewHTTPClient(true))
 }
+func (r Resolved) newXAIClient(token string) provider.Client {
+	return provider.NewModelRouter("xai",
+		r.withHTTPClient(provider.NewXAI(token, r.BaseURL)),
+		map[string]provider.Client{
+			provider.APIResponses: r.withHTTPClient(provider.NewOpenAIResponsesNamed(token, r.BaseURL, "xai")),
+		})
+}
 // wrapWithRefresh wraps an OAuth client so the access token is
 // refreshed automatically before each API call. Without this, long
 // sessions (hours) silently fail when the 1-hour token expires.
 			// anthropic-messages on api.kimi.com/coding.
 			return r.withHTTPClient(provider.NewKimiCodingWithHeaders(token, baseURL, kimiCodeHeaders()))
 		case "xai":
-			if r.Model == "grok-4.5" {
+			return r.newXAIClient(token)
-				return r.withHTTPClient(provider.NewOpenAIResponsesNamed(token, baseURL, "xai"))
-			}
-			return r.withHTTPClient(provider.NewXAI(token, baseURL))
 		default:
 			return r.withHTTPClient(provider.NewAnthropicOAuth(token, baseURL))
 		}

M packages/agent/modes/interactive.go +2 -2


@@ -4654,8 +4654,8 @@ func (i *Interactive) swapModel(prov, model string, builder func(string, string)
4654	4654		return
4655	4655		}
4656	4656		// Same provider AND not a rescue retry: just swap the model on
4657		-	// the existing agent — no rebuild needed because the underlying
4658		-	// client is reusable. Rescue retries always rebuild so a stale
	4657	+	// the existing agent. Mixed-API providers dispatch from model metadata,
	4658	+	// so the client remains reusable. Rescue retries always rebuild so a stale
4659	4659		// auth header / base URL can't carry over.
4660	4660		if !rescue && i.agent != nil && m.Provider == i.cfg.Provider {
4661	4661		i.mu.Lock()


@@ -4654,8 +4654,8 @@ func (i *Interactive) swapModel(prov, model string, builder func(string, string)
4654		return	4654		return
4655		}	4655		}
4656		// Same provider AND not a rescue retry: just swap the model on	4656		// Same provider AND not a rescue retry: just swap the model on
4657	-	// the existing agent — no rebuild needed because the underlying	4657	+	// the existing agent. Mixed-API providers dispatch from model metadata,
4658	-	// client is reusable. Rescue retries always rebuild so a stale	4658	+	// so the client remains reusable. Rescue retries always rebuild so a stale
4659		// auth header / base URL can't carry over.	4659		// auth header / base URL can't carry over.
4660		if !rescue && i.agent != nil && m.Provider == i.cfg.Provider {	4660		if !rescue && i.agent != nil && m.Provider == i.cfg.Provider {
4661		i.mu.Lock()	4661		i.mu.Lock()

M packages/provider/deferred_tools_test.go +1 -1


@@ -87,7 +87,7 @@ func TestKimiK3AndGrok45CatalogMetadata(t *testing.T) {
87	87		if err != nil {
88	88		t.Fatal(err)
89	89		}
90		-	if !m.Reasoning \|\| m.MaxOutput != 30000 {
	90	+	if !m.Reasoning \|\| m.MaxOutput != 30000 \|\| m.API != APIResponses {
91	91		t.Fatalf("xai/grok-4.5 metadata = %+v", m)
92	92		}
93	93		}


@@ -87,7 +87,7 @@ func TestKimiK3AndGrok45CatalogMetadata(t *testing.T) {
87		if err != nil {	87		if err != nil {
88		t.Fatal(err)	88		t.Fatal(err)
89		}	89		}
90	-	if !m.Reasoning \|\| m.MaxOutput != 30000 {	90	+	if !m.Reasoning \|\| m.MaxOutput != 30000 \|\| m.API != APIResponses {
91		t.Fatalf("xai/grok-4.5 metadata = %+v", m)	91		t.Fatalf("xai/grok-4.5 metadata = %+v", m)
92		}	92		}
93		}	93		}

M packages/provider/extra_models.go +1 -1


@@ -42,7 +42,7 @@ var supplementCatalog = []Model{
42	42		BaseURL: "https://ai-gateway.vercel.sh"},
43	43
44	44		// ----- xAI Responses API -----
45		-	{Provider: "xai", ID: "grok-4.5", DisplayName: "Grok 4.5",
	45	+	{Provider: "xai", ID: "grok-4.5", DisplayName: "Grok 4.5", API: APIResponses,
46	46		ContextWindow: 2000000, MaxOutput: 30000, Reasoning: true,
47	47		PriceInput: 1.25, PriceOutput: 2.5, PriceCacheRead: 0.2,
48	48		BaseURL: "https://api.x.ai/v1"},


@@ -42,7 +42,7 @@ var supplementCatalog = []Model{
42		BaseURL: "https://ai-gateway.vercel.sh"},	42		BaseURL: "https://ai-gateway.vercel.sh"},
43			43
44		// ----- xAI Responses API -----	44		// ----- xAI Responses API -----
45	-	{Provider: "xai", ID: "grok-4.5", DisplayName: "Grok 4.5",	45	+	{Provider: "xai", ID: "grok-4.5", DisplayName: "Grok 4.5", API: APIResponses,
46		ContextWindow: 2000000, MaxOutput: 30000, Reasoning: true,	46		ContextWindow: 2000000, MaxOutput: 30000, Reasoning: true,
47		PriceInput: 1.25, PriceOutput: 2.5, PriceCacheRead: 0.2,	47		PriceInput: 1.25, PriceOutput: 2.5, PriceCacheRead: 0.2,
48		BaseURL: "https://api.x.ai/v1"},	48		BaseURL: "https://api.x.ai/v1"},

M packages/provider/httpclient.go +5 -0


@@ -45,6 +45,11 @@ func WithHTTPClient(c Client, httpClient *http.Client) Client {
45	45		v.http = httpClient
46	46		case *renamedClient:
47	47		v.inner = WithHTTPClient(v.inner, httpClient)
	48	+	case *modelRouter:
	49	+	v.fallback = WithHTTPClient(v.fallback, httpClient)
	50	+	for api, inner := range v.byAPI {
	51	+	v.byAPI[api] = WithHTTPClient(inner, httpClient)
	52	+	}
48	53		}
49	54		return c
50	55		}


@@ -45,6 +45,11 @@ func WithHTTPClient(c Client, httpClient *http.Client) Client {
45	v.http = httpClient	45		v.http = httpClient
46	case *renamedClient:	46		case *renamedClient:
47	v.inner = WithHTTPClient(v.inner, httpClient)	47		v.inner = WithHTTPClient(v.inner, httpClient)
		48	+	case *modelRouter:
		49	+	v.fallback = WithHTTPClient(v.fallback, httpClient)
		50	+	for api, inner := range v.byAPI {
		51	+	v.byAPI[api] = WithHTTPClient(inner, httpClient)
		52	+	}
48	}	53		}
49	return c	54		return c
50	}	55		}

A packages/provider/model_router.go +42 -0

@@ -0,0 +1,42 @@
	1	+	package provider
	2	+
	3	+	import (
	4	+	"context"
	5	+	"fmt"
	6	+	)
	7	+
	8	+	const (
	9	+	// APICompletions identifies the OpenAI Chat Completions wire API.
	10	+	APICompletions = "openai-completions"
	11	+	// APIResponses identifies the OpenAI Responses wire API.
	12	+	APIResponses = "openai-responses"
	13	+	)
	14	+
	15	+	// modelRouter dispatches each request according to the selected model's API.
	16	+	// This lets one provider expose models that use different wire protocols while
	17	+	// remaining reusable across model switches.
	18	+	type modelRouter struct {
	19	+	name string
	20	+	fallback Client
	21	+	byAPI map[string]Client
	22	+	}
	23	+
	24	+	// NewModelRouter creates a client that dispatches requests using Model.API.
	25	+	// Models with no API override, and models absent from the catalog, use fallback.
	26	+	func NewModelRouter(name string, fallback Client, byAPI map[string]Client) Client {
	27	+	return &modelRouter{name: name, fallback: fallback, byAPI: byAPI}
	28	+	}
	29	+
	30	+	func (c *modelRouter) Name() string { return c.name }
	31	+
	32	+	func (c *modelRouter) Stream(ctx context.Context, req Request) (<-chan Event, error) {
	33	+	client := c.fallback
	34	+	if model, err := FindModel(c.name, req.Model); err == nil && model.API != "" {
	35	+	routed := c.byAPI[model.API]
	36	+	if routed == nil {
	37	+	return nil, fmt.Errorf("provider %q has no client for model API %q", c.name, model.API)
	38	+	}
	39	+	client = routed
	40	+	}
	41	+	return client.Stream(ctx, req)
	42	+	}

@@ -0,0 +1,42 @@
	1	+	package provider
	2	+
	3	+	import (
	4	+	"context"
	5	+	"fmt"
	6	+	)
	7	+
	8	+	const (
	9	+	// APICompletions identifies the OpenAI Chat Completions wire API.
	10	+	APICompletions = "openai-completions"
	11	+	// APIResponses identifies the OpenAI Responses wire API.
	12	+	APIResponses = "openai-responses"
	13	+	)
	14	+
	15	+	// modelRouter dispatches each request according to the selected model's API.
	16	+	// This lets one provider expose models that use different wire protocols while
	17	+	// remaining reusable across model switches.
	18	+	type modelRouter struct {
	19	+	name string
	20	+	fallback Client
	21	+	byAPI map[string]Client
	22	+	}
	23	+
	24	+	// NewModelRouter creates a client that dispatches requests using Model.API.
	25	+	// Models with no API override, and models absent from the catalog, use fallback.
	26	+	func NewModelRouter(name string, fallback Client, byAPI map[string]Client) Client {
	27	+	return &modelRouter{name: name, fallback: fallback, byAPI: byAPI}
	28	+	}
	29	+
	30	+	func (c *modelRouter) Name() string { return c.name }
	31	+
	32	+	func (c *modelRouter) Stream(ctx context.Context, req Request) (<-chan Event, error) {
	33	+	client := c.fallback
	34	+	if model, err := FindModel(c.name, req.Model); err == nil && model.API != "" {
	35	+	routed := c.byAPI[model.API]
	36	+	if routed == nil {
	37	+	return nil, fmt.Errorf("provider %q has no client for model API %q", c.name, model.API)
	38	+	}
	39	+	client = routed
	40	+	}
	41	+	return client.Stream(ctx, req)
	42	+	}

A packages/provider/model_router_test.go +50 -0

@@ -0,0 +1,50 @@
	1	+	package provider
	2	+
	3	+	import (
	4	+	"context"
	5	+	"testing"
	6	+	)
	7	+
	8	+	type routeCaptureClient struct {
	9	+	name string
	10	+	models []string
	11	+	}
	12	+
	13	+	func (c *routeCaptureClient) Name() string { return c.name }
	14	+	func (c *routeCaptureClient) Stream(_ context.Context, req Request) (<-chan Event, error) {
	15	+	c.models = append(c.models, req.Model)
	16	+	out := make(chan Event)
	17	+	close(out)
	18	+	return out, nil
	19	+	}
	20	+
	21	+	func TestModelRouterDispatchesByModelAPI(t *testing.T) {
	22	+	completions := &routeCaptureClient{name: "xai"}
	23	+	responses := &routeCaptureClient{name: "xai"}
	24	+	router := NewModelRouter("xai", completions, map[string]Client{
	25	+	APIResponses: responses,
	26	+	})
	27	+
	28	+	for _, model := range []string{"grok-4.5", "grok-build-0.1"} {
	29	+	stream, err := router.Stream(context.Background(), Request{Model: model})
	30	+	if err != nil {
	31	+	t.Fatal(err)
	32	+	}
	33	+	for range stream {
	34	+	}
	35	+	}
	36	+
	37	+	if len(responses.models) != 1 \|\| responses.models[0] != "grok-4.5" {
	38	+	t.Fatalf("Responses models = %v", responses.models)
	39	+	}
	40	+	if len(completions.models) != 1 \|\| completions.models[0] != "grok-build-0.1" {
	41	+	t.Fatalf("Completions models = %v", completions.models)
	42	+	}
	43	+	}
	44	+
	45	+	func TestModelRouterRejectsMissingAPIClient(t *testing.T) {
	46	+	router := NewModelRouter("xai", &routeCaptureClient{name: "xai"}, nil)
	47	+	if _, err := router.Stream(context.Background(), Request{Model: "grok-4.5"}); err == nil {
	48	+	t.Fatal("expected missing API client error")
	49	+	}
	50	+	}

@@ -0,0 +1,50 @@
	1	+	package provider
	2	+
	3	+	import (
	4	+	"context"
	5	+	"testing"
	6	+	)
	7	+
	8	+	type routeCaptureClient struct {
	9	+	name string
	10	+	models []string
	11	+	}
	12	+
	13	+	func (c *routeCaptureClient) Name() string { return c.name }
	14	+	func (c *routeCaptureClient) Stream(_ context.Context, req Request) (<-chan Event, error) {
	15	+	c.models = append(c.models, req.Model)
	16	+	out := make(chan Event)
	17	+	close(out)
	18	+	return out, nil
	19	+	}
	20	+
	21	+	func TestModelRouterDispatchesByModelAPI(t *testing.T) {
	22	+	completions := &routeCaptureClient{name: "xai"}
	23	+	responses := &routeCaptureClient{name: "xai"}
	24	+	router := NewModelRouter("xai", completions, map[string]Client{
	25	+	APIResponses: responses,
	26	+	})
	27	+
	28	+	for _, model := range []string{"grok-4.5", "grok-build-0.1"} {
	29	+	stream, err := router.Stream(context.Background(), Request{Model: model})
	30	+	if err != nil {
	31	+	t.Fatal(err)
	32	+	}
	33	+	for range stream {
	34	+	}
	35	+	}
	36	+
	37	+	if len(responses.models) != 1 \|\| responses.models[0] != "grok-4.5" {
	38	+	t.Fatalf("Responses models = %v", responses.models)
	39	+	}
	40	+	if len(completions.models) != 1 \|\| completions.models[0] != "grok-build-0.1" {
	41	+	t.Fatalf("Completions models = %v", completions.models)
	42	+	}
	43	+	}
	44	+
	45	+	func TestModelRouterRejectsMissingAPIClient(t *testing.T) {
	46	+	router := NewModelRouter("xai", &routeCaptureClient{name: "xai"}, nil)
	47	+	if _, err := router.Stream(context.Background(), Request{Model: "grok-4.5"}); err == nil {
	48	+	t.Fatal("expected missing API client error")
	49	+	}
	50	+	}

M packages/provider/models.go +4 -3


@@ -10,6 +10,7 @@ type Model struct {
10	10		Provider string // "anthropic" \| "openai"
11	11		ID string // API id
12	12		DisplayName string
	13	+	API string // wire API override for providers that support multiple protocols
13	14		ContextWindow int
14	15		MaxOutput int
15	16		Reasoning bool // supports reasoning/thinking

@@ -317,19 +318,19 @@ var Catalog = []Model{
317	318		Speculative: true,
318	319		},
319	320		{
320		-	Provider: "openai", ID: "gpt-5.6-luna", DisplayName: "GPT-5.6 Luna",
	321	+	Provider: "openai", ID: "gpt-5.6-luna", DisplayName: "GPT-5.6 Luna", API: APIResponses,
321	322		ContextWindow: 272000, MaxOutput: 128000, Reasoning: true,
322	323		PriceInput: 1, PriceOutput: 6, PriceCacheRead: 0.1, PriceCacheWrite: 1.25,
323	324		Speculative: true,
324	325		},
325	326		{
326		-	Provider: "openai", ID: "gpt-5.6-sol", DisplayName: "GPT-5.6 Sol",
	327	+	Provider: "openai", ID: "gpt-5.6-sol", DisplayName: "GPT-5.6 Sol", API: APIResponses,
327	328		ContextWindow: 272000, MaxOutput: 128000, Reasoning: true,
328	329		PriceInput: 5, PriceOutput: 30, PriceCacheRead: 0.5, PriceCacheWrite: 6.25,
329	330		Speculative: true,
330	331		},
331	332		{
332		-	Provider: "openai", ID: "gpt-5.6-terra", DisplayName: "GPT-5.6 Terra",
	333	+	Provider: "openai", ID: "gpt-5.6-terra", DisplayName: "GPT-5.6 Terra", API: APIResponses,
333	334		ContextWindow: 272000, MaxOutput: 128000, Reasoning: true,
334	335		PriceInput: 2.5, PriceOutput: 15, PriceCacheRead: 0.25, PriceCacheWrite: 3.125,
335	336		Speculative: true,

 	Provider      string // "anthropic" | "openai"
 	ID            string // API id
 	DisplayName   string
+	API           string // wire API override for providers that support multiple protocols
 	ContextWindow int
 	MaxOutput     int
 	Reasoning     bool // supports reasoning/thinking
 		Speculative: true,
 	},
 	{
-		Provider: "openai", ID: "gpt-5.6-luna", DisplayName: "GPT-5.6 Luna",
+		Provider: "openai", ID: "gpt-5.6-luna", DisplayName: "GPT-5.6 Luna", API: APIResponses,
 		ContextWindow: 272000, MaxOutput: 128000, Reasoning: true,
 		PriceInput: 1, PriceOutput: 6, PriceCacheRead: 0.1, PriceCacheWrite: 1.25,
 		Speculative: true,
 	},
 	{
-		Provider: "openai", ID: "gpt-5.6-sol", DisplayName: "GPT-5.6 Sol",
+		Provider: "openai", ID: "gpt-5.6-sol", DisplayName: "GPT-5.6 Sol", API: APIResponses,
 		ContextWindow: 272000, MaxOutput: 128000, Reasoning: true,
 		PriceInput: 5, PriceOutput: 30, PriceCacheRead: 0.5, PriceCacheWrite: 6.25,
 		Speculative: true,
 	},
 	{
-		Provider: "openai", ID: "gpt-5.6-terra", DisplayName: "GPT-5.6 Terra",
+		Provider: "openai", ID: "gpt-5.6-terra", DisplayName: "GPT-5.6 Terra", API: APIResponses,
 		ContextWindow: 272000, MaxOutput: 128000, Reasoning: true,
 		PriceInput: 2.5, PriceOutput: 15, PriceCacheRead: 0.25, PriceCacheWrite: 3.125,
 		Speculative: true,

M packages/provider/openai_codex.go +14 -8


@@ -159,11 +159,20 @@ type codexRequest struct {
159	159
160	160		// ---- Request building ----
161	161
162		-	func (c codexClient) buildRequest(req Request) (codexRequest, error) {
163		-	m, err := FindModel("openai-codex", req.Model)
164		-	if err != nil {
165		-	m, err = FindModel("openai", req.Model)
	162	+	func (c *codexClient) findModel(id string) (Model, error) {
	163	+	if c.providerName != "" && c.providerName != "openai-codex" {
	164	+	if m, err := FindModel(c.providerName, id); err == nil {
	165	+	return m, nil
	166	+	}
	167	+	}
	168	+	if m, err := FindModel("openai-codex", id); err == nil {
	169	+	return m, nil
166	170		}
	171	+	return FindModel("openai", id)
	172	+	}
	173	+
	174	+	func (c codexClient) buildRequest(req Request) (codexRequest, error) {
	175	+	m, err := c.findModel(req.Model)
167	176		if err != nil {
168	177		return nil, err
169	178		}

@@ -405,10 +414,7 @@ func (c codexClient) runStream(ctx context.Context, resp http.Response, req Re
405	414		defer close(out)
406	415		defer resp.Body.Close()
407	416
408		-	model, _ := FindModel("openai-codex", req.Model)
409		-	if model.ID == "" {
410		-	model, _ = FindModel("openai", req.Model)
411		-	}
	417	+	model, _ := c.findModel(req.Model)
412	418		providerName := c.providerName
413	419		if providerName == "" {
414	420		providerName = "openai-codex"

 // ---- Request building ----
-func (c *codexClient) buildRequest(req Request) (*codexRequest, error) {
+func (c *codexClient) findModel(id string) (Model, error) {
-	m, err := FindModel("openai-codex", req.Model)
+	if c.providerName != "" && c.providerName != "openai-codex" {
-	if err != nil {
+		if m, err := FindModel(c.providerName, id); err == nil {
-		m, err = FindModel("openai", req.Model)
+			return m, nil
+		}
+	}
+	if m, err := FindModel("openai-codex", id); err == nil {
+		return m, nil
 	}
+	return FindModel("openai", id)
+}
+func (c *codexClient) buildRequest(req Request) (*codexRequest, error) {
+	m, err := c.findModel(req.Model)
 	if err != nil {
 		return nil, err
 	}
 	defer close(out)
 	defer resp.Body.Close()
-	model, _ := FindModel("openai-codex", req.Model)
+	model, _ := c.findModel(req.Model)
-	if model.ID == "" {
-		model, _ = FindModel("openai", req.Model)
-	}
 	providerName := c.providerName
 	if providerName == "" {
 		providerName = "openai-codex"

M packages/provider/openai_codex_test.go +14 -0


@@ -125,6 +125,20 @@ func TestGPT56UsesNativeMaxReasoningEffort(t *testing.T) {
125	125		}
126	126		}
127	127
	128	+	func TestXAIResponsesUsesNamedProviderCatalogAndEndpoint(t *testing.T) {
	129	+	named := NewOpenAIResponsesNamed("token", "https://api.x.ai/v1", "xai").(*renamedClient)
	130	+	c := named.inner.(*codexClient)
	131	+
	132	+	if c.baseURL != "https://api.x.ai/v1/responses" {
	133	+	t.Fatalf("base URL = %q", c.baseURL)
	134	+	}
	135	+	for _, model := range []string{"grok-4.5", "grok-build-0.1"} {
	136	+	if _, err := c.buildRequest(Request{Model: model}); err != nil {
	137	+	t.Errorf("build request for %s: %v", model, err)
	138	+	}
	139	+	}
	140	+	}
	141	+
128	142		func TestOpenAIGPT56DoesNotUseCodexCLIRouting(t *testing.T) {
129	143		named := NewOpenAIResponsesNamed("token", "https://example.test/v1/responses", "openai").(*renamedClient)
130	144		c := named.inner.(*codexClient)


@@ -125,6 +125,20 @@ func TestGPT56UsesNativeMaxReasoningEffort(t *testing.T) {
125	}	125		}
126	}	126		}
127		127
		128	+	func TestXAIResponsesUsesNamedProviderCatalogAndEndpoint(t *testing.T) {
		129	+	named := NewOpenAIResponsesNamed("token", "https://api.x.ai/v1", "xai").(*renamedClient)
		130	+	c := named.inner.(*codexClient)
		131	+
		132	+	if c.baseURL != "https://api.x.ai/v1/responses" {
		133	+	t.Fatalf("base URL = %q", c.baseURL)
		134	+	}
		135	+	for _, model := range []string{"grok-4.5", "grok-build-0.1"} {
		136	+	if _, err := c.buildRequest(Request{Model: model}); err != nil {
		137	+	t.Errorf("build request for %s: %v", model, err)
		138	+	}
		139	+	}
		140	+	}
		141	+
128	func TestOpenAIGPT56DoesNotUseCodexCLIRouting(t *testing.T) {	142		func TestOpenAIGPT56DoesNotUseCodexCLIRouting(t *testing.T) {
129	named := NewOpenAIResponsesNamed("token", "https://example.test/v1/responses", "openai").(*renamedClient)	143		named := NewOpenAIResponsesNamed("token", "https://example.test/v1/responses", "openai").(*renamedClient)
130	c := named.inner.(*codexClient)	144		c := named.inner.(*codexClient)

M packages/provider/openai_responses.go +18 -3


@@ -55,9 +55,7 @@ func NewOpenAIResponses(apiKey, baseURL string) Client {
55	55		// reporting the supplied provider name. This lets the `openai` provider
56	56		// route Responses-only preview models without changing the visible provider.
57	57		func NewOpenAIResponsesNamed(apiKey, baseURL, name string) Client {
58		-	if baseURL == "" {
59		-	baseURL = openaiResponsesDefaultBaseURL
60		-	}
	58	+	baseURL = responsesURL(baseURL)
61	59		httpClient := &http.Client{
62	60		Transport: &openaiResponsesTransport{inner: http.DefaultTransport},
63	61		Timeout: 0,

@@ -72,3 +70,20 @@ func NewOpenAIResponsesNamed(apiKey, baseURL, name string) Client {
72	70		}
73	71		return &renamedClient{inner: inner, name: name}
74	72		}
	73	+
	74	+	// responsesURL accepts either an API root or a complete Responses endpoint.
	75	+	// Catalog model entries generally carry API roots, while explicit overrides
	76	+	// may already include /responses.
	77	+	func responsesURL(baseURL string) string {
	78	+	baseURL = strings.TrimRight(baseURL, "/")
	79	+	if baseURL == "" {
	80	+	return openaiResponsesDefaultBaseURL
	81	+	}
	82	+	if strings.HasSuffix(baseURL, "/responses") {
	83	+	return baseURL
	84	+	}
	85	+	if versionSegmentSuffix.MatchString(baseURL) {
	86	+	return baseURL + "/responses"
	87	+	}
	88	+	return baseURL + "/v1/responses"
	89	+	}


@@ -55,9 +55,7 @@ func NewOpenAIResponses(apiKey, baseURL string) Client {
55		// reporting the supplied provider name. This lets the `openai` provider	55		// reporting the supplied provider name. This lets the `openai` provider
56		// route Responses-only preview models without changing the visible provider.	56		// route Responses-only preview models without changing the visible provider.
57		func NewOpenAIResponsesNamed(apiKey, baseURL, name string) Client {	57		func NewOpenAIResponsesNamed(apiKey, baseURL, name string) Client {
58	-	if baseURL == "" {	58	+	baseURL = responsesURL(baseURL)
59	-	baseURL = openaiResponsesDefaultBaseURL
60	-	}
61		httpClient := &http.Client{	59		httpClient := &http.Client{
62		Transport: &openaiResponsesTransport{inner: http.DefaultTransport},	60		Transport: &openaiResponsesTransport{inner: http.DefaultTransport},
63		Timeout: 0,	61		Timeout: 0,

@@ -72,3 +70,20 @@ func NewOpenAIResponsesNamed(apiKey, baseURL, name string) Client {
72		}	70		}
73		return &renamedClient{inner: inner, name: name}	71		return &renamedClient{inner: inner, name: name}
74		}	72		}
			73	+
			74	+	// responsesURL accepts either an API root or a complete Responses endpoint.
			75	+	// Catalog model entries generally carry API roots, while explicit overrides
			76	+	// may already include /responses.
			77	+	func responsesURL(baseURL string) string {
			78	+	baseURL = strings.TrimRight(baseURL, "/")
			79	+	if baseURL == "" {
			80	+	return openaiResponsesDefaultBaseURL
			81	+	}
			82	+	if strings.HasSuffix(baseURL, "/responses") {
			83	+	return baseURL
			84	+	}
			85	+	if versionSegmentSuffix.MatchString(baseURL) {
			86	+	return baseURL + "/responses"
			87	+	}
			88	+	return baseURL + "/v1/responses"
			89	+	}