feat(session,llm): restore session runtime on resume; align Responses API details

README.md

@@ -74,9 +74,10 @@ responses OpenAI Responses API /v1/responses
 ```
 
 The Responses API client converts the conversation history to Responses input
-items (`user`/`assistant`/`system` messages, `function_call`, and
+items (`user`/`assistant` messages, `function_call`, and
-`function_call_output`), streams `response.output_text.delta` events, and
+`function_call_output`), sends the system prompt as the top-level
-supports function calling through `response.output_item.added`,
+`instructions` field, streams `response.output_text.delta` events, and supports
+function calling through `response.output_item.added`,
 `response.function_call_arguments.delta`, and `response.output_item.done`.
 
 `models` is required. `/model model <name>` is restricted to that list. Each
@@ -98,10 +99,11 @@ none, middle, high, xhigh, max
 ```
 
 When configured or changed with `/model thinking ...`, the value is sent as a
-top-level request field named `thinking`. For OpenAI reasoning models on the
+top-level request field. The default field depends on `api_type`: `chat` uses
-Responses API, set `thinking_param: reasoning` so the level is sent as the
+`thinking`, while `responses` uses the official `reasoning` field, sent as
-official `reasoning: {"effort": ...}` field; agentu maps its `middle` level to
+`{"effort": ...}` with agentu's `middle` level mapped to the official `medium`
-the official `medium` value and passes `none`, `high`, `xhigh`, `max` through.
+value (`none`, `high`, `xhigh`, `max` pass through). Override with
+`thinking_param` if a provider expects a different field name.
 
 ## CLI Flags
 

internal/session/manager.go

@@ -68,11 +68,29 @@ func (m *Manager) Resume(id string) error {
 		return err
 	}
 	m.currentSession = sess
+	m.restoreRuntime(sess)
 	m.agent.SetMessages(sess.Messages)
 	m.agent.SetLastUsage(sess.LastUsage)
 	return nil
 }
 
+// restoreRuntime switches the manager and agent back to the provider, model,
+// and thinking level recorded on a resumed session. Values that are no longer
+// valid for the current config are skipped and the defaults remain active.
+func (m *Manager) restoreRuntime(sess *Session) {
+	if sess.Provider != "" && sess.Provider != m.providerName {
+		if _, ok := m.cfg.Providers[sess.Provider]; ok {
+			_, _ = m.SetProvider(sess.Provider)
+		}
+	}
+	if sess.Model != "" && sess.Model != m.provider.Model {
+		_, _ = m.SetModel(sess.Model)
+	}
+	if sess.Thinking != "" && sess.Thinking != m.provider.Thinking {
+		_, _ = m.SetThinking(sess.Thinking)
+	}
+}
+
 // Save persists the current session to disk.
 func (m *Manager) Save() error {
 	if m.store == nil || m.currentSession == nil {
@@ -82,6 +100,8 @@ func (m *Manager) Save() error {
 	m.currentSession.LastUsage = m.agent.LastUsage()
 	m.currentSession.Provider = m.providerName
 	m.currentSession.Model = m.provider.Model
+	m.currentSession.APIType = m.provider.APIType
+	m.currentSession.Thinking = m.provider.Thinking
 	// Auto-name if empty
 	if m.currentSession.Name == "" {
 		m.currentSession.Name = DefaultName(m.currentSession.Messages)
@@ -150,6 +170,8 @@ func (m *Manager) NewSession() (string, error) {
 			m.currentSession.Messages = msgs
 			m.currentSession.Provider = m.providerName
 			m.currentSession.Model = m.provider.Model
+			m.currentSession.APIType = m.provider.APIType
+			m.currentSession.Thinking = m.provider.Thinking
 			if m.currentSession.Name == "" {
 				m.currentSession.Name = DefaultName(msgs)
 			}
@@ -257,7 +279,7 @@ func (m *Manager) SetThinking(level string) (string, error) {
 	m.provider.Thinking = level
 	m.provider.ThinkingConfigured = true
 	if m.provider.ThinkingParam == "" {
-		m.provider.ThinkingParam = "thinking"
+		m.provider.ThinkingParam = config.DefaultThinkingParam(m.provider.APIType)
 	}
 	m.cfg.Providers[m.providerName] = m.provider
 	m.syncAgent(nil)

internal/session/manager_test.go

@@ -403,3 +403,132 @@ func TestManagerUsesModelContextForAgentCompaction(t *testing.T) {
 		t.Fatalf("answerCalls = %d, want 1", provider.answerCalls)
 	}
 }
+
+func TestManagerResumeRestoresRuntime(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "text/event-stream")
+		_, _ = w.Write([]byte(`data: {"type":"response.output_text.delta","item_id":"msg_1","output_index":0,"content_index":0,"delta":"ok"}` + "\n\n"))
+		_, _ = w.Write([]byte(`data: {"type":"response.completed","response":{"id":"resp_1","status":"completed"}}` + "\n\n"))
+	}))
+	defer server.Close()
+
+	cfg := &config.Config{
+		Providers: map[string]config.ProviderConfig{
+			"chat": {
+				Name:               "chat",
+				APIType:            config.APITypeChat,
+				BaseURL:            "https://chat.example.com",
+				APIKey:             "sk-test",
+				Model:              "model-a",
+				Models:             []string{"model-a"},
+				Thinking:           "none",
+				ThinkingParam:      "thinking",
+				ThinkingConfigured: true,
+			},
+			"resp": {
+				Name:               "resp",
+				APIType:            config.APITypeResponses,
+				BaseURL:            server.URL,
+				APIKey:             "sk-test",
+				Model:              "m1",
+				Models:             []string{"m1", "m2"},
+				Thinking:           "none",
+				ThinkingParam:      "reasoning",
+				ThinkingConfigured: true,
+			},
+		},
+	}
+	assistant := agent.New(agent.Options{
+		Provider:        &recordingProvider{},
+		Model:           "model-a",
+		Thinking:        "none",
+		ThinkingParam:   "thinking",
+		ThinkingEnabled: true,
+	})
+	store, err := NewStore(filepath.Join(t.TempDir(), "sessions"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	manager, err := NewManager(cfg, assistant, server.Client(), store)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	// Build a session on the responses provider with a custom model and thinking.
+	manager.newSession()
+	if _, err := manager.SetProvider("resp"); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := manager.SetModel("m2"); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := manager.SetThinking("xhigh"); err != nil {
+		t.Fatal(err)
+	}
+	if err := manager.Save(); err != nil {
+		t.Fatal(err)
+	}
+	id := manager.CurrentSessionID()
+
+	// Drift away from the saved runtime.
+	if _, err := manager.SetProvider("chat"); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := manager.SetModel("model-a"); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := manager.SetThinking("none"); err != nil {
+		t.Fatal(err)
+	}
+
+	if err := manager.Resume(id); err != nil {
+		t.Fatal(err)
+	}
+	if got := manager.CurrentProvider(); got != "resp" {
+		t.Fatalf("provider = %q, want resp", got)
+	}
+	if got := manager.CurrentModel(); got != "m2" {
+		t.Fatalf("model = %q, want m2", got)
+	}
+	if got := manager.CurrentThinking(); got != "xhigh" {
+		t.Fatalf("thinking = %q, want xhigh", got)
+	}
+	if got := manager.CurrentAPIType(); got != config.APITypeResponses {
+		t.Fatalf("api type = %q, want responses", got)
+	}
+	runtime := assistant.RuntimeInfo()
+	if runtime.ProviderName != "resp" || runtime.Model != "m2" || runtime.Thinking != "xhigh" {
+		t.Fatalf("agent runtime = %#v", runtime)
+	}
+}
+
+func TestManagerResumeFallsBackWhenRuntimeMissing(t *testing.T) {
+	manager, assistant, _ := testManager(t)
+
+	sess := &Session{
+		ID:       GenerateID(),
+		Provider: "ghost",
+		Model:    "ghost-model",
+		Thinking: "xhigh",
+		Messages: []llm.Message{{Role: llm.RoleUser, Content: "hi"}},
+	}
+	if err := manager.store.Save(sess); err != nil {
+		t.Fatal(err)
+	}
+	if err := manager.Resume(sess.ID); err != nil {
+		t.Fatal(err)
+	}
+	if got := manager.CurrentProvider(); got != "one" {
+		t.Fatalf("provider = %q, want default one", got)
+	}
+	if got := manager.CurrentModel(); got != "model-a" {
+		t.Fatalf("model = %q, want default model-a", got)
+	}
+	if got := manager.CurrentThinking(); got != "xhigh" {
+		t.Fatalf("thinking = %q, want restored xhigh", got)
+	}
+	runtime := assistant.RuntimeInfo()
+	if runtime.ProviderName != "one" || runtime.Model != "model-a" {
+		t.Fatalf("agent runtime = %#v", runtime)
+	}
+}

internal/session/store.go

@@ -22,6 +22,8 @@ type Session struct {
 	UpdatedAt time.Time     `json:"updated_at"`
 	Provider  string        `json:"provider"`
 	Model     string        `json:"model"`
+	APIType   string        `json:"api_type,omitempty"`
+	Thinking  string        `json:"thinking,omitempty"`
 	Messages  []llm.Message `json:"messages"`
 	LastUsage *llm.Usage    `json:"last_usage,omitempty"`
 }

pkg/config/config.go

@@ -241,7 +241,7 @@ func normalizeProvider(name string, raw rawProviderConfig) ProviderConfig {
 	}
 	applyModelDefaults(&provider)
 	if provider.ThinkingConfigured && provider.ThinkingParam == "" {
-		provider.ThinkingParam = "thinking"
+		provider.ThinkingParam = DefaultThinkingParam(provider.APIType)
 	}
 	return provider
 }
@@ -268,6 +268,17 @@ func IsAPIType(value string) bool {
 	return false
 }
 
+// DefaultThinkingParam returns the default request field used for the
+// thinking level. Chat-completions providers historically use a top-level
+// "thinking" field; the OpenAI Responses API uses "reasoning" with an
+// {"effort": ...} value.
+func DefaultThinkingParam(apiType string) string {
+	if apiType == APITypeResponses {
+		return "reasoning"
+	}
+	return "thinking"
+}
+
 func normalizeModelConfigs(rawModels []rawModelConfig) ([]string, map[string]ModelConfig) {
 	models := make([]string, 0, len(rawModels))
 	configs := make(map[string]ModelConfig, len(rawModels))

pkg/config/config_test.go

@@ -441,3 +441,39 @@ providers:
 		t.Fatalf("err = %v", err)
 	}
 }
+
+func TestLoadDefaultsThinkingParamByAPIType(t *testing.T) {
+	dir := t.TempDir()
+	t.Setenv("AGENTU_TEST_KEY", "sk-test")
+	path := filepath.Join(dir, "agentu.yaml")
+	if err := os.WriteFile(path, []byte(`
+providers:
+  chat:
+    api_type: chat
+    base_url: https://chat.example.com
+    api_key: ${AGENTU_TEST_KEY}
+    models:
+      - id: chat-model
+        thinking: none
+  responses:
+    api_type: responses
+    base_url: https://responses.example.com
+    api_key: ${AGENTU_TEST_KEY}
+    models:
+      - id: responses-model
+        thinking: high
+`), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	cfg, err := Load(path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if got := cfg.Providers["chat"].ThinkingParam; got != "thinking" {
+		t.Fatalf("chat thinking param = %q, want %q", got, "thinking")
+	}
+	if got := cfg.Providers["responses"].ThinkingParam; got != "reasoning" {
+		t.Fatalf("responses thinking param = %q, want %q", got, "reasoning")
+	}
+}

pkg/llm/responses.go

@@ -124,6 +124,9 @@ func buildResponsesRequest(req ChatRequest) map[string]any {
 		"input":  buildResponsesInput(req.Messages),
 		"stream": true,
 	}
+	if instructions := responsesInstructions(req.Messages); instructions != "" {
+		payload["instructions"] = instructions
+	}
 	if len(req.Tools) > 0 {
 		payload["tools"] = buildResponsesTools(req.Tools)
 	}
@@ -153,7 +156,9 @@ func buildResponsesInput(messages []Message) []any {
 	for _, msg := range messages {
 		switch msg.Role {
 		case RoleSystem:
-			items = append(items, responsesMessageInput{Role: "system", Content: msg.Content})
+			// System instructions are sent as the top-level "instructions"
+			// field and must not be duplicated inside the input array.
+			continue
 		case RoleUser:
 			items = append(items, responsesMessageInput{Role: "user", Content: msg.Content})
 		case RoleAssistant:
@@ -179,6 +184,22 @@ func buildResponsesInput(messages []Message) []any {
 	return items
 }
 
+// responsesInstructions extracts system messages into a single instructions
+// string, which is the OpenAI Responses API way to pass developer/system
+// guidance (and the most prompt-cache friendly).
+func responsesInstructions(messages []Message) string {
+	var instructions []string
+	for _, msg := range messages {
+		if msg.Role != RoleSystem {
+			continue
+		}
+		if text := strings.TrimSpace(msg.Content); text != "" {
+			instructions = append(instructions, text)
+		}
+	}
+	return strings.Join(instructions, "\n\n")
+}
+
 func buildResponsesTools(tools []Tool) []responsesTool {
 	out := make([]responsesTool, 0, len(tools))
 	for _, tool := range tools {

pkg/llm/responses_test.go

@@ -59,12 +59,15 @@ func TestBuildResponsesRequest(t *testing.T) {
 	if raw["model"] != "gpt-test" || raw["stream"] != true {
 		t.Fatalf("model/stream = %#v / %#v", raw["model"], raw["stream"])
 	}
+	if raw["instructions"] != "be concise" {
+		t.Fatalf("instructions = %#v", raw["instructions"])
+	}
 	if raw["tool_choice"] != "auto" {
 		t.Fatalf("tool_choice = %#v", raw["tool_choice"])
 	}
 
 	input, ok := raw["input"].([]any)
-	if !ok || len(input) != 6 {
+	if !ok || len(input) != 5 {
 		t.Fatalf("input = %#v", raw["input"])
 	}
 	checkInputItem := func(index int, wantType, wantKey, wantValue string) {
@@ -80,16 +83,15 @@ func TestBuildResponsesRequest(t *testing.T) {
 			t.Fatalf("input[%d].type = %#v, want %q", index, item["type"], wantType)
 		}
 	}
-	checkInputItem(0, "", "role", "system")
+	checkInputItem(0, "", "role", "user")
-	checkInputItem(1, "", "role", "user")
+	checkInputItem(1, "", "role", "assistant")
-	checkInputItem(2, "", "role", "assistant")
+	checkInputItem(2, "function_call", "call_id", "call_1")
-	checkInputItem(3, "function_call", "call_id", "call_1")
+	checkInputItem(3, "function_call_output", "call_id", "call_1")
-	checkInputItem(4, "function_call_output", "call_id", "call_1")
+	checkInputItem(4, "", "role", "assistant")
-	checkInputItem(5, "", "role", "assistant")
+	if item := input[2].(map[string]any); item["name"] != "file_read" || item["arguments"] != `{"path":"README.md"}` {
-	if item := input[3].(map[string]any); item["name"] != "file_read" || item["arguments"] != `{"path":"README.md"}` {
 		t.Fatalf("function_call item = %#v", item)
 	}
-	if item := input[4].(map[string]any); item["output"] != "contents" {
+	if item := input[3].(map[string]any); item["output"] != "contents" {
 		t.Fatalf("function_call_output item = %#v", item)
 	}
 
@@ -117,6 +119,36 @@ func TestBuildResponsesRequest(t *testing.T) {
 	}
 }
 
+func TestResponsesInstructionsCombinesSystemMessages(t *testing.T) {
+	payload := buildResponsesRequest(ChatRequest{
+		Model: "gpt-test",
+		Messages: []Message{
+			{Role: RoleSystem, Content: "first instruction"},
+			{Role: RoleUser, Content: "hi"},
+			{Role: RoleSystem, Content: "second instruction"},
+		},
+	})
+	data, err := json.Marshal(payload)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var raw map[string]any
+	if err := json.Unmarshal(data, &raw); err != nil {
+		t.Fatal(err)
+	}
+	if raw["instructions"] != "first instruction\n\nsecond instruction" {
+		t.Fatalf("instructions = %#v", raw["instructions"])
+	}
+	input, ok := raw["input"].([]any)
+	if !ok || len(input) != 1 {
+		t.Fatalf("input = %#v", raw["input"])
+	}
+	item, ok := input[0].(map[string]any)
+	if !ok || item["role"] != "user" {
+		t.Fatalf("input[0] = %#v", input[0])
+	}
+}
+
 func TestOpenAIResponsesClientStreamsContentAndToolCalls(t *testing.T) {
 	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 		if r.URL.Path != "/v1/responses" {