refactor: unify SDK model factories into internal/models

Move CreateModel, BuildReasoningOptions, ReasoningBudgetTokens and
related types from internal/agent to internal/models as NewSDKChatModel,
SDKModelConfig, etc. This eliminates duplicate ClientType constants and
centralises all Twilight AI SDK instance creation in a single package.

NewSDKEmbeddingModel now accepts a clientType parameter and dispatches
to the native Google embedding provider for google-generative-ai,
instead of always using the OpenAI-compatible endpoint.

internal/agent/agent.go

@@ -11,6 +11,7 @@ import (
 	sdk "github.com/memohai/twilight-ai/sdk"
 
 	"github.com/memohai/memoh/internal/agent/tools"
+	"github.com/memohai/memoh/internal/models"
 	"github.com/memohai/memoh/internal/workspace/bridge"
 )
 
@@ -371,9 +372,9 @@ func (*Agent) buildGenerateOptions(cfg RunConfig, tools []sdk.Tool, prepareStep
 	if prepareStep != nil {
 		opts = append(opts, sdk.WithPrepareStep(prepareStep))
 	}
-	opts = append(opts, BuildReasoningOptions(ModelConfig{
-		ClientType: resolveClientType(cfg.Model),
-		ReasoningConfig: &ReasoningConfig{
+	opts = append(opts, models.BuildReasoningOptions(models.SDKModelConfig{
+		ClientType: models.ResolveClientType(cfg.Model),
+		ReasoningConfig: &models.ReasoningConfig{
 			Enabled: cfg.ReasoningEffort != "",
 			Effort:  cfg.ReasoningEffort,
 		},
@@ -381,23 +382,6 @@ func (*Agent) buildGenerateOptions(cfg RunConfig, tools []sdk.Tool, prepareStep
 	return opts
 }
 
-func resolveClientType(model *sdk.Model) string {
-	if model == nil || model.Provider == nil {
-		return ClientTypeOpenAICompletions
-	}
-	name := model.Provider.Name()
-	switch {
-	case strings.Contains(name, "anthropic"):
-		return ClientTypeAnthropicMessages
-	case strings.Contains(name, "google"):
-		return ClientTypeGoogleGenerativeAI
-	case strings.Contains(name, "responses"):
-		return ClientTypeOpenAIResponses
-	default:
-		return ClientTypeOpenAICompletions
-	}
-}
-
 // assembleTools collects tools from all registered ToolProviders.
 func (a *Agent) assembleTools(ctx context.Context, cfg RunConfig) ([]sdk.Tool, error) {
 	if len(a.toolProviders) == 0 {

internal/agent/read_media.go

@@ -7,6 +7,7 @@ import (
 	sdk "github.com/memohai/twilight-ai/sdk"
 
 	agenttools "github.com/memohai/memoh/internal/agent/tools"
+	"github.com/memohai/memoh/internal/models"
 )
 
 func decorateReadMediaTools(model *sdk.Model, tools []sdk.Tool) ([]sdk.Tool, *readMediaDecorationState) {
@@ -14,7 +15,7 @@ func decorateReadMediaTools(model *sdk.Model, tools []sdk.Tool) ([]sdk.Tool, *re
 		return tools, nil
 	}
 
-	clientType := resolveClientType(model)
+	clientType := models.ResolveClientType(model)
 	state := &readMediaDecorationState{
 		pendingImages: make(map[string]sdk.ImagePart),
 	}
@@ -164,7 +165,7 @@ func buildReadMediaImagePart(clientType, imageBase64, mediaType string) sdk.Imag
 	}
 
 	image := imageBase64
-	if clientType != ClientTypeAnthropicMessages {
+	if clientType != string(models.ClientTypeAnthropicMessages) {
 		image = fmt.Sprintf("data:%s;base64,%s", mediaType, imageBase64)
 	}
 	return sdk.ImagePart{

internal/agent/spawn_adapter.go

@@ -6,6 +6,7 @@ import (
 	sdk "github.com/memohai/twilight-ai/sdk"
 
 	"github.com/memohai/memoh/internal/agent/tools"
+	"github.com/memohai/memoh/internal/models"
 )
 
 // SpawnAdapter wraps *Agent to satisfy tools.SpawnAgent without creating
@@ -68,10 +69,10 @@ func SpawnSystemPrompt(sessionType string) string {
 	})
 }
 
-// SpawnModelCreatorFunc returns a tools.ModelCreator that delegates to agent.CreateModel.
+// SpawnModelCreatorFunc returns a tools.ModelCreator that delegates to models.NewSDKChatModel.
 func SpawnModelCreatorFunc() tools.ModelCreator {
 	return func(modelID, clientType, apiKey, baseURL string) *sdk.Model {
-		return CreateModel(ModelConfig{
+		return models.NewSDKChatModel(models.SDKModelConfig{
 			ModelID:    modelID,
 			ClientType: clientType,
 			APIKey:     apiKey,

internal/agent/types.go

@@ -95,21 +95,6 @@ type SystemFile struct {
 	Content  string
 }
 
-// ModelConfig holds provider and model information resolved from DB.
-type ModelConfig struct {
-	ModelID         string
-	ClientType      string
-	APIKey          string //nolint:gosec // carries provider credential material at runtime
-	BaseURL         string
-	ReasoningConfig *ReasoningConfig
-}
-
-// ReasoningConfig controls extended thinking/reasoning behavior.
-type ReasoningConfig struct {
-	Enabled bool
-	Effort  string
-}
-
 func mustMarshal(v any) json.RawMessage {
 	data, err := json.Marshal(v)
 	if err != nil {

internal/compaction/service.go

@@ -11,9 +11,9 @@ import (
 	"github.com/jackc/pgx/v5/pgtype"
 	sdk "github.com/memohai/twilight-ai/sdk"
 
-	"github.com/memohai/memoh/internal/agent"
 	"github.com/memohai/memoh/internal/db"
 	"github.com/memohai/memoh/internal/db/sqlc"
+	"github.com/memohai/memoh/internal/models"
 )
 
 // Service manages context compaction for bot conversations.
@@ -103,7 +103,7 @@ func (s *Service) doCompaction(ctx context.Context, logID pgtype.UUID, sessionUU
 
 	userPrompt := buildUserPrompt(priorSummaries, entries)
 
-	model := agent.CreateModel(agent.ModelConfig{
+	model := models.NewSDKChatModel(models.SDKModelConfig{
 		ClientType: cfg.ClientType,
 		BaseURL:    cfg.BaseURL,
 		APIKey:     cfg.APIKey,

internal/conversation/flow/resolver.go

@@ -270,15 +270,15 @@ func (r *Resolver) resolve(ctx context.Context, req conversation.ChatRequest) (r
 		reasoningEffort = botSettings.ReasoningEffort
 	}
 
-	var reasoningConfig *agentpkg.ReasoningConfig
+	var reasoningConfig *models.ReasoningConfig
 	if reasoningEffort != "" {
-		reasoningConfig = &agentpkg.ReasoningConfig{
+		reasoningConfig = &models.ReasoningConfig{
 			Enabled: true,
 			Effort:  reasoningEffort,
 		}
 	}
 
-	modelCfg := agentpkg.ModelConfig{
+	modelCfg := models.SDKModelConfig{
 		ModelID:         chatModel.ModelID,
 		ClientType:      clientType,
 		APIKey:          provider.ApiKey,
@@ -286,7 +286,7 @@ func (r *Resolver) resolve(ctx context.Context, req conversation.ChatRequest) (r
 		ReasoningConfig: reasoningConfig,
 	}
 
-	sdkModel := agentpkg.CreateModel(modelCfg)
+	sdkModel := models.NewSDKChatModel(modelCfg)
 	sdkMessages := modelMessagesToSDKMessages(nonNilModelMessages(messages))
 
 	runCfg := agentpkg.RunConfig{

internal/conversation/flow/resolver_title.go

@@ -9,7 +9,6 @@ import (
 
 	sdk "github.com/memohai/twilight-ai/sdk"
 
-	agentpkg "github.com/memohai/memoh/internal/agent"
 	"github.com/memohai/memoh/internal/conversation"
 	"github.com/memohai/memoh/internal/db/sqlc"
 	messageevent "github.com/memohai/memoh/internal/message/event"
@@ -105,13 +104,13 @@ func (r *Resolver) generateTitle(ctx context.Context, model models.GetResponse,
 		"Return ONLY the title text, nothing else.\n\n" +
 		"User: " + userSnippet
 
-	modelCfg := agentpkg.ModelConfig{
+	modelCfg := models.SDKModelConfig{
 		ModelID:    model.ModelID,
 		ClientType: provider.ClientType,
 		APIKey:     provider.ApiKey,
 		BaseURL:    provider.BaseUrl,
 	}
-	sdkModel := agentpkg.CreateModel(modelCfg)
+	sdkModel := models.NewSDKChatModel(modelCfg)
 
 	genCtx, cancel := context.WithTimeout(ctx, titleGenerateTimeout)
 	defer cancel()

internal/memory/adapters/builtin/dense_runtime.go

@@ -35,6 +35,7 @@ type denseRuntime struct {
 
 type denseModelSpec struct {
 	modelID    string
+	clientType string
 	baseURL    string
 	apiKey     string
 	dimensions int
@@ -74,7 +75,7 @@ func newDenseRuntime(providerConfig map[string]any, queries *dbsqlc.Queries, cfg
 		return nil, fmt.Errorf("dense runtime: %w", err)
 	}
 
-	embedModel := models.NewSDKEmbeddingModel(spec.baseURL, spec.apiKey, spec.modelID, denseEmbedTimeout)
+	embedModel := models.NewSDKEmbeddingModel(spec.clientType, spec.baseURL, spec.apiKey, spec.modelID, denseEmbedTimeout)
 
 	return &denseRuntime{
 		qdrant:     qClient,
@@ -565,6 +566,7 @@ func resolveDenseEmbeddingModel(ctx context.Context, queries *dbsqlc.Queries, mo
 	}
 	return denseModelSpec{
 		modelID:    strings.TrimSpace(row.ModelID),
+		clientType: strings.TrimSpace(provider.ClientType),
 		baseURL:    strings.TrimSpace(provider.BaseUrl),
 		apiKey:     strings.TrimSpace(provider.ApiKey),
 		dimensions: *cfg.Dimensions,

internal/memory/memllm/client.go

@@ -11,6 +11,7 @@ import (
 
 	"github.com/memohai/memoh/internal/agent"
 	adapters "github.com/memohai/memoh/internal/memory/adapters"
+	"github.com/memohai/memoh/internal/models"
 )
 
 const (
@@ -42,7 +43,7 @@ func New(cfg Config) *Client {
 }
 
 func (c *Client) model() *sdk.Model {
-	return agent.CreateModel(agent.ModelConfig{
+	return models.NewSDKChatModel(models.SDKModelConfig{
 		ModelID:    c.cfg.ModelID,
 		ClientType: c.cfg.ClientType,
 		APIKey:     c.cfg.APIKey,

internal/models/embedding.go

@@ -4,22 +4,34 @@ import (
 	"net/http"
 	"time"
 
+	googleembedding "github.com/memohai/twilight-ai/provider/google/embedding"
 	openaiembedding "github.com/memohai/twilight-ai/provider/openai/embedding"
 	sdk "github.com/memohai/twilight-ai/sdk"
 )
 
 // NewSDKEmbeddingModel creates a Twilight AI SDK EmbeddingModel for the given
-// provider configuration. Currently all embedding providers use the
-// OpenAI-compatible /embeddings endpoint (including Google-hosted models that
-// expose the same wire format), so we route everything through the OpenAI
-// embedding provider. If a future provider requires a different wire protocol,
-// add a branch here.
-func NewSDKEmbeddingModel(baseURL, apiKey, modelID string, timeout time.Duration) *sdk.EmbeddingModel {
+// provider configuration. It dispatches to the native Google embedding provider
+// when clientType is "google-generative-ai", and falls back to the
+// OpenAI-compatible /embeddings endpoint for all other provider types.
+func NewSDKEmbeddingModel(clientType, baseURL, apiKey, modelID string, timeout time.Duration) *sdk.EmbeddingModel {
 	if timeout <= 0 {
 		timeout = 30 * time.Second
 	}
 	httpClient := &http.Client{Timeout: timeout}
 
+	switch ClientType(clientType) {
+	case ClientTypeGoogleGenerativeAI:
+		opts := []googleembedding.Option{
+			googleembedding.WithAPIKey(apiKey),
+			googleembedding.WithHTTPClient(httpClient),
+		}
+		if baseURL != "" {
+			opts = append(opts, googleembedding.WithBaseURL(baseURL))
+		}
+		p := googleembedding.New(opts...)
+		return p.EmbeddingModel(modelID)
+
+	default:
 		opts := []openaiembedding.Option{
 			openaiembedding.WithAPIKey(apiKey),
 			openaiembedding.WithHTTPClient(httpClient),
@@ -27,7 +39,7 @@ func NewSDKEmbeddingModel(baseURL, apiKey, modelID string, timeout time.Duration
 		if baseURL != "" {
 			opts = append(opts, openaiembedding.WithBaseURL(baseURL))
 		}
-
 		p := openaiembedding.New(opts...)
 		return p.EmbeddingModel(modelID)
+	}
 }

internal/models/probe.go

@@ -43,7 +43,7 @@ func (s *Service) Test(ctx context.Context, id string) (TestResponse, error) {
 	// Embedding models don't have a chat Provider in the SDK — probe
 	// the /embeddings endpoint directly.
 	if model.Type == string(ModelTypeEmbedding) {
-		return s.testEmbeddingModel(ctx, baseURL, apiKey, model.ModelID)
+		return s.testEmbeddingModel(ctx, string(clientType), baseURL, apiKey, model.ModelID)
 	}
 
 	sdkProvider := NewSDKProvider(baseURL, apiKey, clientType, probeTimeout)
@@ -100,11 +100,11 @@ func (s *Service) Test(ctx context.Context, id string) (TestResponse, error) {
 // testEmbeddingModel probes an embedding model by performing a minimal
 // embedding request via the Twilight SDK, verifying that the model is
 // reachable and functional rather than merely checking HTTP connectivity.
-func (*Service) testEmbeddingModel(ctx context.Context, baseURL, apiKey, modelID string) (TestResponse, error) {
+func (*Service) testEmbeddingModel(ctx context.Context, clientType, baseURL, apiKey, modelID string) (TestResponse, error) {
 	ctx, cancel := context.WithTimeout(ctx, probeTimeout)
 	defer cancel()
 
-	model := NewSDKEmbeddingModel(baseURL, apiKey, modelID, probeTimeout)
+	model := NewSDKEmbeddingModel(clientType, baseURL, apiKey, modelID, probeTimeout)
 	client := sdk.NewClient()
 
 	start := time.Now()

internal/models/sdk.go

@@ -1,6 +1,8 @@
-package agent
+package models
 
 import (
+	"strings"
+
 	anthropicmessages "github.com/memohai/twilight-ai/provider/anthropic/messages"
 	googlegenerative "github.com/memohai/twilight-ai/provider/google/generativeai"
 	openaicompletions "github.com/memohai/twilight-ai/provider/openai/completions"
@@ -8,23 +10,30 @@ import (
 	sdk "github.com/memohai/twilight-ai/sdk"
 )
 
-// ClientType constants matching the database model configuration.
-const (
-	ClientTypeOpenAICompletions  = "openai-completions"
-	ClientTypeOpenAIResponses    = "openai-responses"
-	ClientTypeAnthropicMessages  = "anthropic-messages"
-	ClientTypeGoogleGenerativeAI = "google-generative-ai"
-)
+// SDKModelConfig holds provider and model information resolved from DB,
+// used to construct a Twilight AI SDK Model instance.
+type SDKModelConfig struct {
+	ModelID         string
+	ClientType      string
+	APIKey          string //nolint:gosec // carries provider credential material at runtime
+	BaseURL         string
+	ReasoningConfig *ReasoningConfig
+}
+
+// ReasoningConfig controls extended thinking/reasoning behavior.
+type ReasoningConfig struct {
+	Enabled bool
+	Effort  string
+}
 
-// Reasoning budget maps per client type.
 var (
 	anthropicBudget = map[string]int{"low": 5000, "medium": 16000, "high": 50000}
 	googleBudget    = map[string]int{"low": 5000, "medium": 16000, "high": 50000}
 )
 
-// CreateModel builds a Twilight AI SDK Model from the resolved model config.
-func CreateModel(cfg ModelConfig) *sdk.Model {
-	switch cfg.ClientType {
+// NewSDKChatModel builds a Twilight AI SDK Model from the resolved model config.
+func NewSDKChatModel(cfg SDKModelConfig) *sdk.Model {
+	switch ClientType(cfg.ClientType) {
 	case ClientTypeOpenAICompletions:
 		opts := []openaicompletions.Option{
 			openaicompletions.WithAPIKey(cfg.APIKey),
@@ -53,7 +62,7 @@ func CreateModel(cfg ModelConfig) *sdk.Model {
 			opts = append(opts, anthropicmessages.WithBaseURL(cfg.BaseURL))
 		}
 		if cfg.ReasoningConfig != nil && cfg.ReasoningConfig.Enabled {
-			budget := ReasoningBudgetTokens(ClientTypeAnthropicMessages, cfg.ReasoningConfig.Effort)
+			budget := ReasoningBudgetTokens(cfg.ClientType, cfg.ReasoningConfig.Effort)
 			opts = append(opts, anthropicmessages.WithThinking(anthropicmessages.ThinkingConfig{
 				Type:         "enabled",
 				BudgetTokens: budget,
@@ -73,7 +82,6 @@ func CreateModel(cfg ModelConfig) *sdk.Model {
 		return p.ChatModel(cfg.ModelID)
 
 	default:
-		// OpenAI-compatible fallback
 		opts := []openaicompletions.Option{
 			openaicompletions.WithAPIKey(cfg.APIKey),
 		}
@@ -86,7 +94,7 @@ func CreateModel(cfg ModelConfig) *sdk.Model {
 }
 
 // BuildReasoningOptions returns SDK generation options for reasoning/thinking.
-func BuildReasoningOptions(cfg ModelConfig) []sdk.GenerateOption {
+func BuildReasoningOptions(cfg SDKModelConfig) []sdk.GenerateOption {
 	if cfg.ReasoningConfig == nil || !cfg.ReasoningConfig.Enabled {
 		return nil
 	}
@@ -95,9 +103,8 @@ func BuildReasoningOptions(cfg ModelConfig) []sdk.GenerateOption {
 		effort = "medium"
 	}
 
-	switch cfg.ClientType {
+	switch ClientType(cfg.ClientType) {
 	case ClientTypeAnthropicMessages:
-		// Anthropic uses thinking budget — no SDK option, handled by provider
 		return nil
 	case ClientTypeOpenAIResponses, ClientTypeOpenAICompletions:
 		return []sdk.GenerateOption{sdk.WithReasoningEffort(effort)}
@@ -113,7 +120,7 @@ func ReasoningBudgetTokens(clientType, effort string) int {
 	if effort == "" {
 		effort = "medium"
 	}
-	switch clientType {
+	switch ClientType(clientType) {
 	case ClientTypeAnthropicMessages:
 		if b, ok := anthropicBudget[effort]; ok {
 			return b
@@ -128,3 +135,21 @@ func ReasoningBudgetTokens(clientType, effort string) int {
 		return 0
 	}
 }
+
+// ResolveClientType infers the client type string from an SDK Model's provider name.
+func ResolveClientType(model *sdk.Model) string {
+	if model == nil || model.Provider == nil {
+		return string(ClientTypeOpenAICompletions)
+	}
+	name := model.Provider.Name()
+	switch {
+	case strings.Contains(name, "anthropic"):
+		return string(ClientTypeAnthropicMessages)
+	case strings.Contains(name, "google"):
+		return string(ClientTypeGoogleGenerativeAI)
+	case strings.Contains(name, "responses"):
+		return string(ClientTypeOpenAIResponses)
+	default:
+		return string(ClientTypeOpenAICompletions)
+	}
+}