diff --git a/internal/probe/capability.go b/internal/probe/capability.go
new file mode 100644
index 00000000..7fc9aad5
--- /dev/null
+++ b/internal/probe/capability.go
@@ -0,0 +1,157 @@
+package probe
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"strings"
+	"time"
+)
+
+type TransportProfile struct {
+	SupportsOpenAIModels          bool     `json:"supports_openai_models"`
+	SupportsOpenAIChatCompletions bool     `json:"supports_openai_chat_completions"`
+	SupportsOpenAIResponses       bool     `json:"supports_openai_responses"`
+	SupportsAnthropicMessages     bool     `json:"supports_anthropic_messages"`
+	AuthStyle                     string   `json:"auth_style"`
+	ModelIDStyle                  string   `json:"model_id_style"`
+	KnownAdvisories               []string `json:"known_advisories"`
+}
+
+type ModelCapabilityProfile struct {
+	RawModelID              string `json:"raw_model_id"`
+	NormalizedModelID       string `json:"normalized_model_id"`
+	CanonicalModelFamily    string `json:"canonical_model_family"`
+	SupportsStream          string `json:"supports_stream"`
+	SupportsTools           string `json:"supports_tools"`
+	SupportsReasoningFields string `json:"supports_reasoning_fields"`
+	SmokeChatOK             bool   `json:"smoke_chat_ok"`
+}
+
+type CapabilityProfile struct {
+	TransportProfile TransportProfile         `json:"transport_profile"`
+	ModelProfiles    []ModelCapabilityProfile `json:"model_profiles"`
+}
+
+func ProbeCapabilities(ctx context.Context, baseURL, apiKey string, rawModels []string) (*CapabilityProfile, error) {
+	profile := &CapabilityProfile{
+		TransportProfile: TransportProfile{
+			SupportsOpenAIModels: len(rawModels) > 0,
+			AuthStyle:            "bearer",
+			ModelIDStyle:         detectModelIDStyle(rawModels),
+			KnownAdvisories:      []string{},
+		},
+		ModelProfiles: make([]ModelCapabilityProfile, 0, len(rawModels)),
+	}
+
+	responsesStatus, err := probeJSONEndpoint(ctx, baseURL, apiKey, "/v1/responses", map[string]any{
+		"model": firstNonEmptyModel(rawModels),
+		"input": "ping",
+	})
+	if err != nil {
+		return nil, err
+	}
+	profile.TransportProfile.SupportsOpenAIResponses = responsesStatus >= http.StatusOK && responsesStatus < http.StatusMultipleChoices
+
+	for _, rawModel := range rawModels {
+		modelProfile := ModelCapabilityProfile{
+			RawModelID:              strings.TrimSpace(rawModel),
+			NormalizedModelID:       NormalizeModelID(rawModel),
+			CanonicalModelFamily:    CanonicalModelFamily(rawModel),
+			SupportsStream:          "unknown",
+			SupportsTools:           "unknown",
+			SupportsReasoningFields: "unknown",
+		}
+
+		chatStatus, err := probeJSONEndpoint(ctx, baseURL, apiKey, "/v1/chat/completions", map[string]any{
+			"model": modelProfile.RawModelID,
+			"messages": []map[string]string{
+				{"role": "user", "content": "ping"},
+			},
+			"max_tokens":  8,
+			"temperature": 0,
+		})
+		if err != nil {
+			return nil, err
+		}
+
+		modelProfile.SmokeChatOK = chatStatus >= http.StatusOK && chatStatus < http.StatusMultipleChoices
+		if modelProfile.SmokeChatOK {
+			profile.TransportProfile.SupportsOpenAIChatCompletions = true
+		}
+		if chatStatus == http.StatusForbidden {
+			appendAdvisory(&profile.TransportProfile.KnownAdvisories, "initial_probe_race_expected")
+		}
+
+		profile.ModelProfiles = append(profile.ModelProfiles, modelProfile)
+	}
+
+	if !profile.TransportProfile.SupportsOpenAIResponses && profile.TransportProfile.SupportsOpenAIChatCompletions {
+		appendAdvisory(&profile.TransportProfile.KnownAdvisories, "responses_unsupported_but_chat_ok")
+	}
+
+	return profile, nil
+}
+
+func probeJSONEndpoint(ctx context.Context, baseURL, apiKey, path string, payload any) (int, error) {
+	requestURL, err := joinGatewayPath(baseURL, path)
+	if err != nil {
+		return 0, fmt.Errorf("resolve %s endpoint: %w", path, err)
+	}
+
+	var body bytes.Buffer
+	if payload != nil {
+		if err := json.NewEncoder(&body).Encode(payload); err != nil {
+			return 0, fmt.Errorf("encode %s probe payload: %w", path, err)
+		}
+	}
+
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, requestURL, &body)
+	if err != nil {
+		return 0, fmt.Errorf("build %s request: %w", path, err)
+	}
+	req.Header.Set("Content-Type", "application/json")
+	if token := strings.TrimSpace(apiKey); token != "" {
+		req.Header.Set("Authorization", "Bearer "+token)
+	}
+
+	resp, err := (&http.Client{Timeout: 15 * time.Second}).Do(req)
+	if err != nil {
+		return 0, fmt.Errorf("request %s: %w", path, err)
+	}
+	defer resp.Body.Close()
+
+	return resp.StatusCode, nil
+}
+
+func detectModelIDStyle(rawModels []string) string {
+	for _, rawModel := range rawModels {
+		if strings.Contains(strings.TrimSpace(rawModel), "/") {
+			return "vendor_prefixed"
+		}
+	}
+	return "canonical"
+}
+
+func appendAdvisory(values *[]string, advisory string) {
+	if advisory == "" {
+		return
+	}
+	for _, existing := range *values {
+		if existing == advisory {
+			return
+		}
+	}
+	*values = append(*values, advisory)
+}
+
+func firstNonEmptyModel(rawModels []string) string {
+	for _, rawModel := range rawModels {
+		if trimmed := strings.TrimSpace(rawModel); trimmed != "" {
+			return trimmed
+		}
+	}
+	return "ping"
+}
diff --git a/internal/probe/capability_test.go b/internal/probe/capability_test.go
new file mode 100644
index 00000000..8d33f08b
--- /dev/null
+++ b/internal/probe/capability_test.go
@@ -0,0 +1,119 @@
+package probe
+
+import (
+	"context"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+)
+
+func TestProbeCapabilities(t *testing.T) {
+	t.Parallel()
+
+	t.Run("responses unsupported but chat works", func(t *testing.T) {
+		t.Parallel()
+
+		var responseCalls int
+		server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+			switch r.URL.Path {
+			case "/v1/responses":
+				responseCalls++
+				http.Error(w, `{"error":"unsupported"}`, http.StatusForbidden)
+			case "/v1/chat/completions":
+				w.Header().Set("Content-Type", "application/json")
+				_, _ = w.Write([]byte(`{"id":"chatcmpl-1","choices":[{"message":{"content":"pong"}}]}`))
+			default:
+				t.Fatalf("unexpected path %q", r.URL.Path)
+			}
+		}))
+		defer server.Close()
+
+		profile, err := ProbeCapabilities(context.Background(), server.URL, "sk-test", []string{"kimi-k2.6"})
+		if err != nil {
+			t.Fatalf("ProbeCapabilities() error = %v", err)
+		}
+		if !profile.TransportProfile.SupportsOpenAIChatCompletions {
+			t.Fatal("SupportsOpenAIChatCompletions = false, want true")
+		}
+		if profile.TransportProfile.SupportsOpenAIResponses {
+			t.Fatal("SupportsOpenAIResponses = true, want false")
+		}
+		if responseCalls == 0 {
+			t.Fatal("responses endpoint was not probed")
+		}
+		if !containsString(profile.TransportProfile.KnownAdvisories, "responses_unsupported_but_chat_ok") {
+			t.Fatalf("KnownAdvisories = %#v, want responses advisory", profile.TransportProfile.KnownAdvisories)
+		}
+	})
+
+	t.Run("records per model capability profile", func(t *testing.T) {
+		t.Parallel()
+
+		server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+			switch r.URL.Path {
+			case "/v1/responses":
+				w.WriteHeader(http.StatusOK)
+				_, _ = w.Write([]byte(`{"id":"resp_1"}`))
+			case "/v1/chat/completions":
+				body := make([]byte, r.ContentLength)
+				_, _ = r.Body.Read(body)
+				w.Header().Set("Content-Type", "application/json")
+				_, _ = w.Write([]byte(`{"id":"chatcmpl-1","choices":[{"message":{"content":"ok"}}]}`))
+			default:
+				t.Fatalf("unexpected path %q", r.URL.Path)
+			}
+		}))
+		defer server.Close()
+
+		profile, err := ProbeCapabilities(context.Background(), server.URL, "sk-test", []string{"deepseek-ai/DeepSeek-V4-Pro", "kimi-k2.6"})
+		if err != nil {
+			t.Fatalf("ProbeCapabilities() error = %v", err)
+		}
+		if len(profile.ModelProfiles) != 2 {
+			t.Fatalf("len(ModelProfiles) = %d, want 2", len(profile.ModelProfiles))
+		}
+		if profile.ModelProfiles[0].NormalizedModelID != "deepseek-v4-pro" {
+			t.Fatalf("NormalizedModelID = %q, want %q", profile.ModelProfiles[0].NormalizedModelID, "deepseek-v4-pro")
+		}
+		if profile.ModelProfiles[0].CanonicalModelFamily != "deepseek-v4-pro" {
+			t.Fatalf("CanonicalModelFamily = %q, want %q", profile.ModelProfiles[0].CanonicalModelFamily, "deepseek-v4-pro")
+		}
+		if !profile.ModelProfiles[0].SmokeChatOK {
+			t.Fatal("SmokeChatOK = false, want true")
+		}
+	})
+
+	t.Run("records initial probe advisory on transient auth race", func(t *testing.T) {
+		t.Parallel()
+
+		server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+			switch r.URL.Path {
+			case "/v1/responses":
+				http.Error(w, `{"error":"forbidden"}`, http.StatusForbidden)
+			case "/v1/chat/completions":
+				http.Error(w, `{"error":"warmup"}`, http.StatusForbidden)
+			default:
+				t.Fatalf("unexpected path %q", r.URL.Path)
+			}
+		}))
+		defer server.Close()
+
+		profile, err := ProbeCapabilities(context.Background(), server.URL, "sk-test", []string{"kimi-k2.6"})
+		if err != nil {
+			t.Fatalf("ProbeCapabilities() error = %v", err)
+		}
+		if !containsString(profile.TransportProfile.KnownAdvisories, "initial_probe_race_expected") {
+			t.Fatalf("KnownAdvisories = %#v, want initial probe advisory", profile.TransportProfile.KnownAdvisories)
+		}
+	})
+}
+
+func containsString(values []string, want string) bool {
+	for _, value := range values {
+		if strings.TrimSpace(value) == want {
+			return true
+		}
+	}
+	return false
+}
diff --git a/internal/probe/completion.go b/internal/probe/completion.go
new file mode 100644
index 00000000..a78c034e
--- /dev/null
+++ b/internal/probe/completion.go
@@ -0,0 +1,123 @@
+package probe
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"strings"
+	"time"
+)
+
+type CompletionResult struct {
+	Model          string
+	HTTPStatus     int
+	LatencyMs      int64
+	Classification string
+	Error          string
+}
+
+func ResolveSmokeModel(requested []string, rawModels []string, profile *CapabilityProfile) (string, []string, error) {
+	recommended := RecommendModels(requested, rawModels)
+	for _, candidate := range recommended {
+		if profileAllowsSmoke(profile, candidate) {
+			return candidate, recommended, nil
+		}
+	}
+
+	for _, rawModel := range rawModels {
+		if strings.TrimSpace(rawModel) == "" {
+			continue
+		}
+		if profileAllowsSmoke(profile, rawModel) {
+			return rawModel, recommended, nil
+		}
+	}
+
+	if len(rawModels) > 0 && strings.TrimSpace(rawModels[0]) != "" {
+		return rawModels[0], recommended, nil
+	}
+
+	return "", recommended, fmt.Errorf("no smoke model available")
+}
+
+func SmokeCompletion(ctx context.Context, baseURL, apiKey, model string, profile *CapabilityProfile) (*CompletionResult, error) {
+	model = strings.TrimSpace(model)
+	if model == "" {
+		return nil, fmt.Errorf("model is required")
+	}
+
+	path := "/v1/chat/completions"
+	classification := "chat_completions"
+	payload := map[string]any{
+		"model": model,
+		"messages": []map[string]string{
+			{"role": "user", "content": "ping"},
+		},
+		"max_tokens":  8,
+		"temperature": 0,
+	}
+
+	if profile != nil && profile.TransportProfile.SupportsOpenAIResponses {
+		path = "/v1/responses"
+		classification = "responses"
+		payload = map[string]any{
+			"model": model,
+			"input": "ping",
+		}
+	}
+
+	requestURL, err := joinGatewayPath(baseURL, path)
+	if err != nil {
+		return nil, fmt.Errorf("resolve smoke endpoint: %w", err)
+	}
+
+	var body bytes.Buffer
+	if err := json.NewEncoder(&body).Encode(payload); err != nil {
+		return nil, fmt.Errorf("encode smoke payload: %w", err)
+	}
+
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, requestURL, &body)
+	if err != nil {
+		return nil, fmt.Errorf("build smoke request: %w", err)
+	}
+	req.Header.Set("Content-Type", "application/json")
+	if token := strings.TrimSpace(apiKey); token != "" {
+		req.Header.Set("Authorization", "Bearer "+token)
+	}
+
+	startedAt := time.Now()
+	resp, err := (&http.Client{Timeout: 15 * time.Second}).Do(req)
+	if err != nil {
+		return nil, fmt.Errorf("request smoke completion: %w", err)
+	}
+	defer resp.Body.Close()
+
+	result := &CompletionResult{
+		Model:          model,
+		HTTPStatus:     resp.StatusCode,
+		LatencyMs:      time.Since(startedAt).Milliseconds(),
+		Classification: classification,
+	}
+	if resp.StatusCode < http.StatusOK || resp.StatusCode >= http.StatusMultipleChoices {
+		result.Error = fmt.Sprintf("unexpected_status_%d", resp.StatusCode)
+	}
+	return result, nil
+}
+
+func profileAllowsSmoke(profile *CapabilityProfile, rawModel string) bool {
+	if profile == nil || len(profile.ModelProfiles) == 0 {
+		return true
+	}
+
+	targetRaw := strings.TrimSpace(rawModel)
+	targetCanonical := CanonicalModelFamily(rawModel)
+	for _, modelProfile := range profile.ModelProfiles {
+		if strings.TrimSpace(modelProfile.RawModelID) == targetRaw || modelProfile.CanonicalModelFamily == targetCanonical {
+			return modelProfile.SmokeChatOK
+		}
+	}
+
+	return false
+}
diff --git a/internal/probe/completion_test.go b/internal/probe/completion_test.go
new file mode 100644
index 00000000..691aea0f
--- /dev/null
+++ b/internal/probe/completion_test.go
@@ -0,0 +1,94 @@
+package probe
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+)
+
+func TestResolveSmokeModel(t *testing.T) {
+	t.Parallel()
+
+	t.Run("uses requested alias when matched", func(t *testing.T) {
+		t.Parallel()
+
+		profile := &CapabilityProfile{
+			ModelProfiles: []ModelCapabilityProfile{
+				{RawModelID: "kimi-k2.6", CanonicalModelFamily: "kimi-2.6", SmokeChatOK: true},
+			},
+		}
+
+		model, recommended, err := ResolveSmokeModel([]string{"kimi 2.6"}, []string{"kimi-k2.6"}, profile)
+		if err != nil {
+			t.Fatalf("ResolveSmokeModel() error = %v", err)
+		}
+		if model != "kimi-k2.6" {
+			t.Fatalf("ResolveSmokeModel() model = %q, want %q", model, "kimi-k2.6")
+		}
+		if len(recommended) != 1 || recommended[0] != "kimi-k2.6" {
+			t.Fatalf("recommended = %#v, want discovered alias", recommended)
+		}
+	})
+
+	t.Run("falls back to discovered model with smoke support", func(t *testing.T) {
+		t.Parallel()
+
+		profile := &CapabilityProfile{
+			ModelProfiles: []ModelCapabilityProfile{
+				{RawModelID: "deepseek-ai/DeepSeek-V4-Pro", CanonicalModelFamily: "deepseek-v4-pro", SmokeChatOK: true},
+			},
+		}
+
+		model, recommended, err := ResolveSmokeModel([]string{"unknown"}, []string{"deepseek-ai/DeepSeek-V4-Pro"}, profile)
+		if err != nil {
+			t.Fatalf("ResolveSmokeModel() error = %v", err)
+		}
+		if model != "deepseek-ai/DeepSeek-V4-Pro" {
+			t.Fatalf("ResolveSmokeModel() model = %q, want discovered model", model)
+		}
+		if len(recommended) != 0 {
+			t.Fatalf("recommended = %#v, want empty for unknown request", recommended)
+		}
+	})
+}
+
+func TestSmokeCompletion(t *testing.T) {
+	t.Parallel()
+
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/v1/chat/completions" {
+			t.Fatalf("path = %q, want chat completions fallback", r.URL.Path)
+		}
+		var payload map[string]any
+		if err := json.NewDecoder(r.Body).Decode(&payload); err != nil {
+			t.Fatalf("decode request body: %v", err)
+		}
+		if payload["model"] != "kimi-k2.6" {
+			t.Fatalf("payload model = %v, want kimi-k2.6", payload["model"])
+		}
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"id":"chatcmpl-1","choices":[{"message":{"content":"pong"}}]}`))
+	}))
+	defer server.Close()
+
+	profile := &CapabilityProfile{
+		TransportProfile: TransportProfile{
+			SupportsOpenAIChatCompletions: true,
+			SupportsOpenAIResponses:       false,
+			KnownAdvisories:               []string{"responses_unsupported_but_chat_ok"},
+		},
+	}
+
+	result, err := SmokeCompletion(context.Background(), server.URL, "sk-test", "kimi-k2.6", profile)
+	if err != nil {
+		t.Fatalf("SmokeCompletion() error = %v", err)
+	}
+	if result.HTTPStatus != http.StatusOK {
+		t.Fatalf("HTTPStatus = %d, want %d", result.HTTPStatus, http.StatusOK)
+	}
+	if result.Classification != "chat_completions" {
+		t.Fatalf("Classification = %q, want %q", result.Classification, "chat_completions")
+	}
+}