feat(providers/openai): allow reasoning-model overrides and recognize gpt-6+ (#57)

Cherry-picks upstream charmbracelet/fantasy cfb0530 (fix: gpt series 6+, charmbracelet#354) unchanged, extends the gpt-5+ generation match into getResponsesModelConfig so the Responses client sends reasoning parameters (and drops temperature/top_p) for gpt-6 and later, and adds openai.WithReasoningModelFunc / WithLanguageModelReasoningModelFunc so callers can override the classification for both the Responses and Chat Completions clients.

> Xum acted on behalf of @ibetitsmike for this merge.

Author: ibetitsmike

providers/openai/language_model.go

@@ -25,6 +25,7 @@ type languageModel struct {
 	modelID                    string
 	client                     openai.Client
 	objectMode                 fantasy.ObjectMode
+	reasoningModelFunc         func(modelID string) bool
 	prepareCallFunc            LanguageModelPrepareCallFunc
 	mapFinishReasonFunc        LanguageModelMapFinishReasonFunc
 	extraContentFunc           LanguageModelExtraContentFunc
@@ -87,6 +88,13 @@ func WithLanguageModelToPromptFunc(fn LanguageModelToPromptFunc) LanguageModelOp
 	}
 }
 
+// WithLanguageModelReasoningModelFunc overrides reasoning-model detection for Chat Completions.
+func WithLanguageModelReasoningModelFunc(fn func(modelID string) bool) LanguageModelOption {
+	return func(l *languageModel) {
+		l.reasoningModelFunc = fn
+	}
+}
+
 // WithLanguageModelObjectMode sets the object generation mode.
 func WithLanguageModelObjectMode(om fantasy.ObjectMode) LanguageModelOption {
 	return func(l *languageModel) {
@@ -161,7 +169,7 @@ func (o languageModel) prepareParams(call fantasy.Call) (*openai.ChatCompletionN
 		params.PresencePenalty = param.NewOpt(*call.PresencePenalty)
 	}
 
-	if isReasoningModel(o.modelID) {
+	if o.isReasoningModel() {
 		// remove unsupported settings for reasoning models
 		// see https://platform.openai.com/docs/guides/reasoning#limitations
 		if call.Temperature != nil {
@@ -589,6 +597,13 @@ func (o languageModel) Stream(ctx context.Context, call fantasy.Call) (fantasy.S
 	}, nil
 }
 
+func (o languageModel) isReasoningModel() bool {
+	if o.reasoningModelFunc != nil {
+		return o.reasoningModelFunc(o.modelID)
+	}
+	return isReasoningModel(o.modelID)
+}
+
 func isReasoningModel(modelID string) bool {
 	return strings.HasPrefix(modelID, "o1") || strings.Contains(modelID, "-o1") ||
 		strings.HasPrefix(modelID, "o3") || strings.Contains(modelID, "-o3") ||

providers/openai/language_model_hooks.go

@@ -130,7 +130,11 @@ func DefaultPrepareCallFunc(model fantasy.LanguageModel, params *openai.ChatComp
 		}
 	}
 
-	if isReasoningModel(model.Model()) {
+	reasoning := isReasoningModel(model.Model())
+	if lm, ok := model.(interface{ isReasoningModel() bool }); ok {
+		reasoning = lm.isReasoningModel()
+	}
+	if reasoning {
 		if providerOptions.LogitBias != nil {
 			params.LogitBias = nil
 			warnings = append(warnings, fantasy.CallWarning{

providers/openai/language_model_params_test.go

@@ -0,0 +1,88 @@
+package openai
+
+import (
+	"context"
+	"testing"
+
+	"charm.land/fantasy"
+	"github.com/stretchr/testify/require"
+)
+
+func TestPrepareParams_ChatReasoningModelOverride(t *testing.T) {
+	t.Parallel()
+
+	tests := []struct {
+		name          string
+		modelID       string
+		opts          []Option
+		wantReasoning bool
+	}{
+		{name: "unknown default", modelID: "totally-new-model"},
+		{name: "gpt-5 default", modelID: "gpt-5", wantReasoning: true},
+		{
+			name: "force reasoning", modelID: "totally-new-model", wantReasoning: true,
+			opts: []Option{WithReasoningModelFunc(func(modelID string) bool { return modelID == "totally-new-model" })},
+		},
+		{
+			name: "force non-reasoning", modelID: "gpt-5",
+			opts: []Option{WithReasoningModelFunc(func(modelID string) bool { return modelID != "gpt-5" })},
+		},
+		{
+			name: "language model option", modelID: "totally-new-model", wantReasoning: true,
+			opts: []Option{WithLanguageModelOptions(WithLanguageModelReasoningModelFunc(func(string) bool { return true }))},
+		},
+		{
+			name: "provider option takes precedence", modelID: "gpt-5",
+			opts: []Option{
+				WithLanguageModelOptions(WithLanguageModelReasoningModelFunc(func(string) bool { return true })),
+				WithReasoningModelFunc(func(string) bool { return false }),
+			},
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+
+			provider, err := New(tt.opts...)
+			require.NoError(t, err)
+			model, err := provider.LanguageModel(context.Background(), tt.modelID)
+			require.NoError(t, err)
+			lm, ok := model.(languageModel)
+			require.True(t, ok)
+
+			params, warnings, err := lm.prepareParams(fantasy.Call{
+				Prompt:          fantasy.Prompt{testTextMessage(fantasy.MessageRoleUser, "hello")},
+				Temperature:     new(0.7),
+				MaxOutputTokens: new(int64(512)),
+				ProviderOptions: fantasy.ProviderOptions{
+					Name: &ProviderOptions{LogProbs: new(true)},
+				},
+			})
+			require.NoError(t, err)
+
+			if tt.wantReasoning {
+				require.False(t, params.Temperature.Valid())
+				require.False(t, params.MaxTokens.Valid())
+				require.True(t, params.MaxCompletionTokens.Valid())
+				require.Equal(t, int64(512), params.MaxCompletionTokens.Value)
+				require.False(t, params.Logprobs.Valid())
+				var unsupported []string
+				for _, warning := range warnings {
+					require.Equal(t, fantasy.CallWarningTypeUnsupportedSetting, warning.Type)
+					unsupported = append(unsupported, warning.Setting)
+				}
+				require.ElementsMatch(t, []string{"temperature", "Logprobs"}, unsupported)
+			} else {
+				require.True(t, params.Temperature.Valid())
+				require.Equal(t, 0.7, params.Temperature.Value)
+				require.True(t, params.MaxTokens.Valid())
+				require.Equal(t, int64(512), params.MaxTokens.Value)
+				require.False(t, params.MaxCompletionTokens.Valid())
+				require.True(t, params.Logprobs.Valid())
+				require.True(t, params.Logprobs.Value)
+				require.Empty(t, warnings)
+			}
+		})
+	}
+}

providers/openai/openai.go

@@ -31,6 +31,7 @@ type options struct {
 	name                 string
 	useResponsesAPI      bool
 	responsesAPIFunc     func(modelID string) bool
+	reasoningModelFunc   func(modelID string) bool
 	headers              map[string]string
 	userAgent            string
 	client               option.HTTPClient
@@ -143,6 +144,15 @@ func WithResponsesAPIFunc(fn func(modelID string) bool) Option {
 	}
 }
 
+// WithReasoningModelFunc sets a custom classifier for which models are reasoning models.
+// When set, it replaces the built-in model-name heuristics for both the Responses
+// and Chat Completions clients.
+func WithReasoningModelFunc(fn func(modelID string) bool) Option {
+	return func(o *options) {
+		o.reasoningModelFunc = fn
+	}
+}
+
 // WithUserAgent sets an explicit User-Agent header, overriding the default and any
 // value set via WithHeaders.
 func WithUserAgent(ua string) Option {
@@ -194,11 +204,14 @@ func (o *provider) LanguageModel(_ context.Context, modelID string) (fantasy.Lan
 		if objectMode == fantasy.ObjectModeJSON {
 			objectMode = fantasy.ObjectModeAuto
 		}
-		return newResponsesLanguageModel(modelID, o.options.name, client, objectMode), nil
+		return newResponsesLanguageModel(modelID, o.options.name, client, objectMode, o.options.reasoningModelFunc), nil
 	}
 
 	languageModelOptions := append([]LanguageModelOption{}, o.options.languageModelOptions...)
 	languageModelOptions = append(languageModelOptions, WithLanguageModelObjectMode(o.options.objectMode))
+	if o.options.reasoningModelFunc != nil {
+		languageModelOptions = append(languageModelOptions, WithLanguageModelReasoningModelFunc(o.options.reasoningModelFunc))
+	}
 
 	return newLanguageModel(
 		modelID,

providers/openai/responses_language_model.go

@@ -24,19 +24,21 @@ import (
 const topLogprobsMax = 20
 
 type responsesLanguageModel struct {
-	provider   string
-	modelID    string
-	client     openai.Client
-	objectMode fantasy.ObjectMode
+	provider           string
+	modelID            string
+	client             openai.Client
+	objectMode         fantasy.ObjectMode
+	reasoningModelFunc func(modelID string) bool
 }
 
 // newResponsesLanguageModel implements a responses api model.
-func newResponsesLanguageModel(modelID string, provider string, client openai.Client, objectMode fantasy.ObjectMode) responsesLanguageModel {
+func newResponsesLanguageModel(modelID string, provider string, client openai.Client, objectMode fantasy.ObjectMode, reasoningModelFunc func(modelID string) bool) responsesLanguageModel {
 	return responsesLanguageModel{
-		modelID:    modelID,
-		provider:   provider,
-		client:     client,
-		objectMode: objectMode,
+		modelID:            modelID,
+		provider:           provider,
+		client:             client,
+		objectMode:         objectMode,
+		reasoningModelFunc: reasoningModelFunc,
 	}
 }
 
@@ -77,7 +79,8 @@ func getResponsesModelConfig(modelID string) responsesModelConfig {
 		supportsPriorityProcessing: supportsPriorityProcessing,
 	}
 
-	if strings.Contains(strings.ToLower(modelID), "gpt-5-chat") {
+	reasoningGeneration := reasoningGenerationPattern.MatchString(strings.ToLower(modelID))
+	if reasoningGeneration && strings.Contains(strings.ToLower(modelID), "-chat") {
 		return responsesModelConfig{
 			isReasoningModel:           false,
 			systemMessageMode:          defaults.systemMessageMode,
@@ -91,7 +94,7 @@ func getResponsesModelConfig(modelID string) responsesModelConfig {
 		strings.HasPrefix(modelID, "o3") || strings.Contains(modelID, "-o3") ||
 		strings.HasPrefix(modelID, "o4") || strings.Contains(modelID, "-o4") ||
 		strings.HasPrefix(modelID, "oss") || strings.Contains(modelID, "-oss") ||
-		strings.Contains(strings.ToLower(modelID), "gpt-5") ||
+		reasoningGeneration ||
 		strings.Contains(modelID, "codex-") || strings.Contains(modelID, "computer-use") {
 		if strings.Contains(modelID, "o1-mini") || strings.Contains(modelID, "o1-preview") {
 			return responsesModelConfig{
@@ -131,6 +134,15 @@ func (o responsesLanguageModel) prepareParams(call fantasy.Call) (*responses.Res
 	params := &responses.ResponseNewParams{}
 
 	modelConfig := getResponsesModelConfig(o.modelID)
+	if o.reasoningModelFunc != nil {
+		if reasoning := o.reasoningModelFunc(o.modelID); reasoning != modelConfig.isReasoningModel {
+			modelConfig.isReasoningModel = reasoning
+			modelConfig.systemMessageMode = "system"
+			if reasoning {
+				modelConfig.systemMessageMode = "developer"
+			}
+		}
+	}
 
 	if call.TopK != nil {
 		warnings = append(warnings, fantasy.CallWarning{

providers/openai/responses_options.go

@@ -3,6 +3,7 @@ package openai
 
 import (
 	"encoding/json"
+	"regexp"
 	"slices"
 	"strings"
 
@@ -296,18 +297,26 @@ func ParseResponsesOptions(data map[string]any) (*ResponsesProviderOptions, erro
 	return &options, nil
 }
 
+// responsesGenerationPattern matches the model generations that only
+// speak the Responses API: gpt-4 and gpt-5 today, the newer generations
+// as they ship (gpt-6, gpt-10, ...), and never the legacy gpt-3 family
+// that predates it.
+var responsesGenerationPattern = regexp.MustCompile(`gpt-(?:[4-9]|[1-9]\d)`)
+
+// reasoningGenerationPattern is the subset of those generations that reason:
+// gpt-5 and everything after it, never gpt-4.
+var reasoningGenerationPattern = regexp.MustCompile(`gpt-(?:[5-9]|[1-9]\d)`)
+
 // IsResponsesModel checks if a model ID is a Responses API model for OpenAI.
 func IsResponsesModel(modelID string) bool {
 	return slices.Contains(responsesModelIDs, modelID) ||
-		strings.Contains(strings.ToLower(modelID), "gpt-4") ||
-		strings.Contains(strings.ToLower(modelID), "gpt-5")
+		responsesGenerationPattern.MatchString(strings.ToLower(modelID))
 }
 
 // IsResponsesReasoningModel checks if a model ID is a Responses API reasoning model for OpenAI.
 func IsResponsesReasoningModel(modelID string) bool {
 	return slices.Contains(responsesReasoningModelIDs, modelID) ||
-		strings.Contains(strings.ToLower(modelID), "gpt-4") ||
-		strings.Contains(strings.ToLower(modelID), "gpt-5")
+		responsesGenerationPattern.MatchString(strings.ToLower(modelID))
 }
 
 // SearchContextSize controls how much context window space the

providers/openai/responses_options_test.go

@@ -0,0 +1,60 @@
+package openai
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+)
+
+func TestIsResponsesModel(t *testing.T) {
+	tests := []struct {
+		modelID string
+		want    bool
+	}{
+		// Explicitly listed models.
+		{"gpt-4.1", true},
+		{"gpt-4o-mini", true},
+		{"chatgpt-4o-latest", true},
+		{"o3", true},
+		{"gpt-oss-120b", true},
+
+		// Generations caught by the pattern, listed or not.
+		{"gpt-5", true},
+		{"gpt-5.1-codex", true},
+		{"gpt-6-astra", true},
+		{"GPT-6-ASTRA", true},
+		{"gpt-10-turbo", true},
+
+		// Everything predating the Responses API stays on chat
+		// completions.
+		{"gpt-3.5-turbo-1106", true}, // in the explicit list
+		{"gpt-3-turbo-instruct", false},
+		{"babbage-002", false},
+		{"davinci-002", false},
+		{"some-custom-model", false},
+	}
+
+	for _, tt := range tests {
+		assert.Equal(t, tt.want, IsResponsesModel(tt.modelID), tt.modelID)
+	}
+}
+
+func TestIsResponsesReasoningModel(t *testing.T) {
+	tests := []struct {
+		modelID string
+		want    bool
+	}{
+		{"gpt-5.1-codex", true},
+		{"gpt-6-astra", true},
+		{"o4-mini", true},
+		{"gpt-oss-120b", true},
+
+		{"gpt-4.1-mini", true}, // gpt-4 matches, as before
+		{"gpt-3-turbo-instruct", false},
+		{"some-custom-model", false},
+	}
+
+	for _, tt := range tests {
+		assert.Equal(t, tt.want, IsResponsesReasoningModel(tt.modelID), tt.modelID)
+	}
+}