feat(tools): stream shell_run output live into the TUI

README.md

@@ -8,6 +8,7 @@ model/provider switching, and local tools.
 
 - TUI-first chat experience with a light theme by default.
 - OpenAI-compatible `/v1/chat/completions` provider support.
+- OpenAI Responses API (`/v1/responses`) provider support.
 - Multiple providers from `~/.agentu/config.yaml`.
 - Session-only provider, model, and thinking-level switching.
 - Read-only project tools enabled by default.
@@ -48,6 +49,7 @@ Use `--config <path>` to load another file.
 ```yaml
 providers:
   loveuer:
+    api_type: chat
     base_url: https://ai.loveuer.com
     api_key: ${AGENTU_API_KEY}
     models:
@@ -64,6 +66,20 @@ Provider names are the keys under `providers`. At startup, agentu selects the
 first provider by name. Use `/model provider <name>` to switch providers for the
 current session.
 
+`api_type` selects which API protocol the provider speaks:
+
+```text
+chat      OpenAI-compatible /v1/chat/completions (default)
+responses OpenAI Responses API /v1/responses
+```
+
+The Responses API client converts the conversation history to Responses input
+items (`user`/`assistant` messages, `function_call`, and
+`function_call_output`), sends the system prompt as the top-level
+`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
 entry is an object with:
 
@@ -83,7 +99,11 @@ none, middle, high, xhigh, max
 ```
 
 When configured or changed with `/model thinking ...`, the value is sent as a
-top-level OpenAI-compatible request field named `thinking`.
+top-level request field. The default field depends on `api_type`: `chat` uses
+`thinking`, while `responses` uses the official `reasoning` field, sent as
+`{"effort": ...}` with agentu's `middle` level mapped to the official `medium`
+value (`none`, `high`, `xhigh`, `max` pass through). Override with
+`thinking_param` if a provider expects a different field name.
 
 ## CLI Flags
 
@@ -147,7 +167,8 @@ Web tools are also available in read-only mode:
 
 - `file_write` for creating or replacing whole files
 - `file_edit` for surgical line-range or pattern edits to existing files
-- `shell_run` for non-interactive shell commands
+- `shell_run` for non-interactive shell commands; output streams live into the
+  TUI while the command runs instead of appearing only after completion
 
 Use `--yolo` for prompts that ask agentu to run or inspect local commands, for
 example:
@@ -174,5 +195,8 @@ Local config and build outputs are intentionally ignored by git.
 
 - Provider/model/thinking changes are session-only; agentu does not write them
   back to `~/.agentu/config.yaml`.
+- `api_type` is per provider and cannot be changed at runtime; use
+  `/model provider <name>` to switch to a provider configured with a different
+  API type.
 - The config schema is intentionally strict during MVP development. Deprecated
   fields such as `provider` or `active_provider` are rejected.

agentu.example.yaml

@@ -1,5 +1,8 @@
 providers:
   loveuer:
+    # API protocol: "chat" (OpenAI-compatible /v1/chat/completions, default)
+    # or "responses" (OpenAI Responses API /v1/responses).
+    api_type: chat
     base_url: https://ai.loveuer.com
     api_key: ${AGENTU_API_KEY}
     models:

internal/agent/agent.go

@@ -24,6 +24,7 @@ const (
 	toolLogArgumentLimit          = 400
 	toolLogOutputPreviewBytes     = 4096
 	toolLogOutputMarker           = "[output]"
+	toolStreamLogLimit            = tools.MaxToolOutputBytes
 
 	summarizerSystemPrompt = "You are a conversation summarizer. Summarize the following conversation history into a concise but detailed summary. Preserve:\n" +
 		"- Key facts, decisions, and conclusions\n" +
@@ -182,6 +183,7 @@ func (a *Agent) SetLastUsage(usage *llm.Usage) {
 	copy := *usage
 	a.lastUsage = &copy
 }
+
 // compactRetainedTokenBudget returns the number of estimated tokens to retain
 // based on compactRecentRetentionPercent, using integer ceiling division.
 func compactRetainedTokenBudget(totalTokens int) int {
@@ -237,7 +239,6 @@ func compactRecentStart(messages []llm.Message, prefixEnd int) int {
 	return compactTurnBoundary(messages, prefixEnd, start)
 }
 
-
 // Compact summarizes older conversation messages while preserving the system
 // prompt and recent messages intact.
 func (a *Agent) Compact(ctx context.Context) error {
@@ -495,32 +496,75 @@ func (a *Agent) toolDefinitions() []llm.Tool {
 	return a.toolRegistry.Definitions()
 }
 
-func (a *Agent) executeTool(ctx context.Context, call llm.ToolCall) string {
+func (a *Agent) executeTool(ctx context.Context, call llm.ToolCall, logs io.Writer) (string, bool) {
 	name := call.Function.Name
 
 	if a.toolRegistry == nil {
-		return tools.FormatError(fmt.Errorf("tool registry is disabled"))
+		return tools.FormatError(fmt.Errorf("tool registry is disabled")), false
 	}
 	tool, ok := a.toolRegistry.Get(name)
 	if !ok {
-		return tools.FormatError(fmt.Errorf("unknown tool: %s", name))
+		return tools.FormatError(fmt.Errorf("unknown tool: %s", name)), false
 	}
 
 	args := strings.TrimSpace(call.Function.Arguments)
 	if args == "" {
 		args = "{}"
 	}
+	raw := json.RawMessage(args)
 
 	toolCtx, cancel := context.WithTimeout(ctx, a.toolTimeout)
 	defer cancel()
-	output, err := tool.Execute(toolCtx, json.RawMessage(args))
+
+	if streaming, ok := tool.(tools.StreamingTool); ok {
+		return a.executeStreamingTool(streaming, toolCtx, raw, name, args, logs)
+	}
+
+	output, err := tool.Execute(toolCtx, raw)
 	if err != nil {
 		if output != "" {
-			return output + "\n" + tools.FormatError(err)
+			return output + "\n" + tools.FormatError(err), false
 		}
-		return tools.FormatError(err)
+		return tools.FormatError(err), false
 	}
-	return output
+	return output, false
+}
+
+func (a *Agent) executeStreamingTool(streaming tools.StreamingTool, ctx context.Context, raw json.RawMessage, name, args string, logs io.Writer) (string, bool) {
+	var streamed int
+	emit := func(chunk string) error {
+		if chunk == "" || logs == nil {
+			return nil
+		}
+		if streamed >= toolStreamLogLimit {
+			return nil
+		}
+		if len(chunk) > toolStreamLogLimit-streamed {
+			chunk = chunk[:toolStreamLogLimit-streamed]
+			streamed = toolStreamLogLimit
+			_, err := fmt.Fprintf(logs, "\n[tool] %s %s\n%s\n%s\n[streaming output truncated after %d bytes]", name, compact(args, toolLogArgumentLimit), toolLogOutputMarker, chunk, toolStreamLogLimit)
+			return err
+		}
+		streamed += len(chunk)
+		_, err := fmt.Fprintf(logs, "\n[tool] %s %s\n%s\n%s", name, compact(args, toolLogArgumentLimit), toolLogOutputMarker, chunk)
+		return err
+	}
+
+	output, err := streaming.ExecuteStream(ctx, raw, emit)
+	if logs != nil {
+		if err != nil {
+			_ = emit("\n" + tools.FormatError(err))
+		} else if streamed == 0 {
+			_ = emit("(no output)")
+		}
+	}
+	if err != nil {
+		if output != "" {
+			return output + "\n" + tools.FormatError(err), true
+		}
+		return tools.FormatError(err), true
+	}
+	return output, true
 }
 
 func (a *Agent) hasMutatingCall(calls []llm.ToolCall) bool {
@@ -550,9 +594,9 @@ func (a *Agent) executeTools(ctx context.Context, calls []llm.ToolCall, logs io.
 	if a.hasMutatingCall(calls) {
 		// Sequential: avoid racing mutating tools.
 		for i, call := range calls {
-			output := a.executeTool(ctx, call)
+			output, streamed := a.executeTool(ctx, call, logs)
 			results[i] = toolExecutionResult{call: call, output: output}
-			if logs != nil {
+			if logs != nil && !streamed {
 				logToolOutput(logs, call, output)
 			}
 		}
@@ -566,9 +610,9 @@ func (a *Agent) executeTools(ctx context.Context, calls []llm.ToolCall, logs io.
 		wg.Add(1)
 		go func(idx int, c llm.ToolCall) {
 			defer wg.Done()
-			output := a.executeTool(ctx, c)
+			output, streamed := a.executeTool(ctx, c, logs)
 			results[idx] = toolExecutionResult{call: c, output: output}
-			if logs != nil {
+			if logs != nil && !streamed {
 				mu.Lock()
 				logToolOutput(logs, c, output)
 				mu.Unlock()

internal/agent/agent_test.go

@@ -87,6 +87,86 @@ func TestAgentRunsToolLoop(t *testing.T) {
 	}
 }
 
+type streamToolProvider struct {
+	calls int
+	t     *testing.T
+}
+
+func (p *streamToolProvider) ChatStream(ctx context.Context, req llm.ChatRequest, emit func(llm.StreamEvent) error) error {
+	p.calls++
+	switch p.calls {
+	case 1:
+		return emit(llm.StreamEvent{ToolCalls: []llm.ToolCallDelta{
+			{Index: 0, ID: "call_1", Type: "function", Name: "test_stream", Arguments: `{}`},
+		}})
+	case 2:
+		last := req.Messages[len(req.Messages)-1]
+		if last.Role != llm.RoleTool || last.Content != "full result" {
+			p.t.Fatalf("last message = %#v", last)
+		}
+		return emit(llm.StreamEvent{Content: "done"})
+	default:
+		p.t.Fatalf("unexpected call count %d", p.calls)
+		return nil
+	}
+}
+
+type streamEchoTool struct{}
+
+func (streamEchoTool) Definition() llm.Tool {
+	return llm.Tool{
+		Type: "function",
+		Function: llm.ToolFunction{
+			Name:        "test_stream",
+			Description: "streaming test",
+			Parameters:  json.RawMessage(`{"type":"object"}`),
+		},
+	}
+}
+
+func (streamEchoTool) Execute(ctx context.Context, raw json.RawMessage) (string, error) {
+	return "full result", nil
+}
+
+func (streamEchoTool) ExecuteStream(ctx context.Context, raw json.RawMessage, emit func(string) error) (string, error) {
+	if err := emit("one\n"); err != nil {
+		return "", err
+	}
+	if err := emit("two\n"); err != nil {
+		return "", err
+	}
+	return "full result", nil
+}
+
+func TestAgentStreamsToolOutputToLogs(t *testing.T) {
+	provider := &streamToolProvider{t: t}
+	a := New(Options{
+		Provider:     provider,
+		Model:        "test",
+		SystemPrompt: "system",
+		ToolRegistry: tools.NewRegistry(streamEchoTool{}),
+	})
+
+	var out strings.Builder
+	var logs strings.Builder
+	if err := a.RunTurn(context.Background(), "stream", &out, &logs); err != nil {
+		t.Fatal(err)
+	}
+	if out.String() != "done" {
+		t.Fatalf("out = %q", out.String())
+	}
+	logText := logs.String()
+	if strings.Count(logText, "[output]") != 2 {
+		t.Fatalf("logs = %q", logText)
+	}
+	if !strings.Contains(logText, "one\n") || !strings.Contains(logText, "two\n") {
+		t.Fatalf("logs missing streamed chunks: %q", logText)
+	}
+	if strings.Contains(logText, "[output]\nfull result") {
+		t.Fatalf("logs contain duplicate full output: %q", logText)
+	}
+}
+
 type contentProvider struct {
 	called bool
 	text   string

internal/app/app.go

@@ -91,10 +91,17 @@ func loadConfig(configPath string) (*config.Config, error) {
 func defaultProvider(cfg *config.Config, httpClient *http.Client) (string, config.ProviderConfig, llm.Provider) {
 	providerName := cfg.DefaultProviderName()
 	providerConfig := cfg.Providers[providerName]
-	provider := llm.NewOpenAICompatibleClient(providerConfig.BaseURL, providerConfig.APIKey, httpClient)
+	provider := newProviderClient(providerConfig, httpClient)
 	return providerName, providerConfig, provider
 }
 
+func newProviderClient(providerConfig config.ProviderConfig, httpClient *http.Client) llm.Provider {
+	if providerConfig.APIType == config.APITypeResponses {
+		return llm.NewOpenAIResponsesClient(providerConfig.BaseURL, providerConfig.APIKey, httpClient)
+	}
+	return llm.NewOpenAICompatibleClient(providerConfig.BaseURL, providerConfig.APIKey, httpClient)
+}
+
 func newAssistant(cfg *config.Config, providerName string, providerConfig config.ProviderConfig, provider llm.Provider, yolo bool, httpClient *http.Client) *agent.Agent {
 	return agent.New(agent.Options{
 		Provider:         provider,

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)
 			}
@@ -204,10 +226,18 @@ func (m *Manager) CurrentThinking() string {
 	return m.provider.Thinking
 }
 
+func (m *Manager) CurrentAPIType() string {
+	if m.provider.APIType == "" {
+		return config.APITypeChat
+	}
+	return m.provider.APIType
+}
+
 func (m *Manager) Info() string {
 	var b strings.Builder
 	fmt.Fprintf(&b, "provider: %s\n", m.CurrentProvider())
 	fmt.Fprintf(&b, "model: %s\n", m.CurrentModel())
+	fmt.Fprintf(&b, "api: %s\n", m.CurrentAPIType())
 	fmt.Fprintf(&b, "thinking: %s\n", m.CurrentThinking())
 	fmt.Fprintf(&b, "providers: %s\n", strings.Join(m.providerNames(), ", "))
 	fmt.Fprintf(&b, "models: %s\n", strings.Join(m.provider.Models, ", "))
@@ -223,7 +253,7 @@ func (m *Manager) SetProvider(name string) (string, error) {
 	}
 	m.providerName = name
 	m.provider = provider
-	m.syncAgent(llm.NewOpenAICompatibleClient(provider.BaseURL, provider.APIKey, m.httpClient))
+	m.syncAgent(providerClient(provider, m.httpClient))
 	return "Switched provider.\n" + m.summary(), nil
 }
 
@@ -249,7 +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)
@@ -257,7 +287,14 @@ func (m *Manager) SetThinking(level string) (string, error) {
 }
 
 func (m *Manager) summary() string {
-	return fmt.Sprintf("provider: %s\nmodel: %s\nthinking: %s", m.CurrentProvider(), m.CurrentModel(), m.CurrentThinking())
+	return fmt.Sprintf("provider: %s\nmodel: %s\napi: %s\nthinking: %s", m.CurrentProvider(), m.CurrentModel(), m.CurrentAPIType(), m.CurrentThinking())
+}
+
+func providerClient(provider config.ProviderConfig, httpClient *http.Client) llm.Provider {
+	if provider.APIType == config.APITypeResponses {
+		return llm.NewOpenAIResponsesClient(provider.BaseURL, provider.APIKey, httpClient)
+	}
+	return llm.NewOpenAICompatibleClient(provider.BaseURL, provider.APIKey, httpClient)
 }
 
 func (m *Manager) maxContextTokens() int {

internal/session/manager_test.go

@@ -2,6 +2,8 @@ package session
 
 import (
 	"context"
+	"net/http"
+	"net/http/httptest"
 	"path/filepath"
 	"strings"
 	"testing"
@@ -109,6 +111,64 @@ func TestManagerRejectsUnknownModelAndThinking(t *testing.T) {
 	}
 }
 
+func TestManagerSwitchesToResponsesProvider(t *testing.T) {
+	var hitPath string
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		hitPath = r.URL.Path
+		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"},
+			},
+			"responses": {
+				Name:    "responses",
+				APIType: config.APITypeResponses,
+				BaseURL: server.URL,
+				APIKey:  "sk-test",
+				Model:   "model-b",
+				Models:  []string{"model-b"},
+			},
+		},
+	}
+	assistant := agent.New(agent.Options{Provider: &recordingProvider{}, Model: "model-a"})
+	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)
+	}
+
+	if _, err := manager.SetProvider("responses"); err != nil {
+		t.Fatal(err)
+	}
+	if manager.CurrentAPIType() != config.APITypeResponses {
+		t.Fatalf("api type = %q", manager.CurrentAPIType())
+	}
+	var out strings.Builder
+	if err := assistant.RunTurn(context.Background(), "hi", &out, nil); err != nil {
+		t.Fatal(err)
+	}
+	if hitPath != "/v1/responses" {
+		t.Fatalf("path = %q, want /v1/responses", hitPath)
+	}
+	if out.String() != "ok" {
+		t.Fatalf("output = %q", out.String())
+	}
+}
+
 func TestManagerExposesModelCommandCandidates(t *testing.T) {
 	manager, _, _ := testManager(t)
 	models := manager.AvailableModels()
@@ -343,3 +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"`
 }

internal/tools/shell.go

@@ -1,6 +1,7 @@
 package tools
 
 import (
+	"bytes"
 	"context"
 	"encoding/json"
 	"errors"
@@ -8,6 +9,7 @@ import (
 	"os"
 	"os/exec"
 	"strings"
+	"sync"
 
 	"agentu/pkg/llm"
 )
@@ -46,6 +48,10 @@ type shellRunArgs struct {
 }
 
 func (t *ShellRunTool) Execute(ctx context.Context, raw json.RawMessage) (string, error) {
+	return t.ExecuteStream(ctx, raw, nil)
+}
+
+func (t *ShellRunTool) ExecuteStream(ctx context.Context, raw json.RawMessage, emit func(string) error) (string, error) {
 	var args shellRunArgs
 	if err := json.Unmarshal(raw, &args); err != nil {
 		return "", fmt.Errorf("parse arguments: %w", err)
@@ -66,16 +72,45 @@ func (t *ShellRunTool) Execute(ctx context.Context, raw json.RawMessage) (string
 	cmd := exec.CommandContext(ctx, shell, "-lc", args.Command)
 	cmd.Dir = workdir
 
-	output, err := cmd.CombinedOutput()
-	text := FormatResult(string(output))
+	var mu sync.Mutex
+	var combined bytes.Buffer
+	writer := &shellStreamWriter{mu: &mu, buf: &combined, emit: emit}
+	cmd.Stdout = writer
+	cmd.Stderr = writer
+
+	runErr := cmd.Run()
+	text := FormatResult(combined.String())
 	if errors.Is(ctx.Err(), context.DeadlineExceeded) {
 		return text, fmt.Errorf("command timed out: %w", ctx.Err())
 	}
-	if err != nil {
+	if runErr != nil {
 		if text != "" {
-			return text, fmt.Errorf("command failed: %w", err)
+			return text, fmt.Errorf("command failed: %w", runErr)
 		}
-		return "", fmt.Errorf("command failed: %w", err)
+		return "", fmt.Errorf("command failed: %w", runErr)
 	}
 	return text, nil
 }
+
+// shellStreamWriter captures command output for the final result while
+// forwarding every write to the streaming callback as it happens.
+type shellStreamWriter struct {
+	mu   *sync.Mutex
+	buf  *bytes.Buffer
+	emit func(string) error
+}
+
+func (w *shellStreamWriter) Write(p []byte) (int, error) {
+	w.mu.Lock()
+	n, writeErr := w.buf.Write(p)
+	w.mu.Unlock()
+	if writeErr != nil {
+		return n, writeErr
+	}
+	if w.emit != nil {
+		if err := w.emit(string(p)); err != nil {
+			return 0, err
+		}
+	}
+	return len(p), nil
+}

internal/tools/shell_test.go

@@ -0,0 +1,71 @@
+package tools
+
+import (
+	"context"
+	"encoding/json"
+	"strings"
+	"sync"
+	"testing"
+)
+
+func TestShellRunStreamsOutputWhileRunning(t *testing.T) {
+	tool := NewShellRunTool(t.TempDir())
+	var mu sync.Mutex
+	var chunks []string
+	output, err := tool.ExecuteStream(context.Background(), json.RawMessage(`{"command":"printf 'a\\n'; sleep 0.05; printf 'b\\n'"}`), func(chunk string) error {
+		mu.Lock()
+		chunks = append(chunks, chunk)
+		mu.Unlock()
+		return nil
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(output, "a") || !strings.Contains(output, "b") {
+		t.Fatalf("output = %q", output)
+	}
+	mu.Lock()
+	joined := strings.Join(chunks, "")
+	mu.Unlock()
+	if len(chunks) == 0 {
+		t.Fatal("expected at least one streamed chunk")
+	}
+	if !strings.Contains(joined, "a") || !strings.Contains(joined, "b") {
+		t.Fatalf("streamed chunks = %q", joined)
+	}
+}
+
+func TestShellRunExecuteCollectsFullOutput(t *testing.T) {
+	tool := NewShellRunTool(t.TempDir())
+	output, err := tool.Execute(context.Background(), json.RawMessage(`{"command":"printf 'hello world'"}`))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if strings.TrimSpace(output) != "hello world" {
+		t.Fatalf("output = %q", output)
+	}
+}
+
+func TestShellRunStreamsStderrAndStdout(t *testing.T) {
+	tool := NewShellRunTool(t.TempDir())
+	var mu sync.Mutex
+	var chunks []string
+	output, err := tool.ExecuteStream(context.Background(), json.RawMessage(`{"command":"echo out; echo err >&2"}`), func(chunk string) error {
+		mu.Lock()
+		chunks = append(chunks, chunk)
+		mu.Unlock()
+		return nil
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(output, "out") || !strings.Contains(output, "err") {
+		t.Fatalf("output = %q", output)
+	}
+	mu.Lock()
+	joined := strings.Join(chunks, "")
+	mu.Unlock()
+	if !strings.Contains(joined, "out") || !strings.Contains(joined, "err") {
+		t.Fatalf("streamed chunks = %q", joined)
+	}
+}

internal/tools/tools.go

@@ -19,6 +19,15 @@ type Tool interface {
 	Execute(ctx context.Context, args json.RawMessage) (string, error)
 }
 
+// StreamingTool is optionally implemented by tools that can report partial
+// output while they execute. emit receives chunks of the tool's output as
+// they are produced; the returned string is still the complete result that
+// gets sent back to the model.
+type StreamingTool interface {
+	Tool
+	ExecuteStream(ctx context.Context, args json.RawMessage, emit func(string) error) (string, error)
+}
+
 // MutatingTool is optionally implemented by tools that modify external state.
 // It is used by the agent to determine whether tool calls can run in parallel.
 type MutatingTool interface {

internal/tui/app.go

@@ -683,15 +683,27 @@ func (m *model) mergeToolLog(text string) bool {
 	if incoming.output == "" {
 		return false
 	}
+	incomingHeader, _, ok := strings.Cut(text, "\n[output]\n")
+	if !ok {
+		return false
+	}
 	for i := len(m.messages) - 1; i >= 0; i-- {
 		if m.messages[i].role != roleTool {
 			continue
 		}
 		existing := parseToolLog(m.messages[i].content)
-		if existing.name == incoming.name && existing.args == incoming.args && existing.output == "" {
-			m.messages[i].content = text
-			return true
+		if existing.name != incoming.name || existing.args != incoming.args {
+			continue
 		}
+		if !strings.Contains(m.messages[i].content, "\n[output]\n") {
+			// Placeholder card from the tool-start log; first chunk fills it.
+			m.messages[i].content = text
+		} else {
+			// Streaming tools send incremental chunks; append the raw delta
+			// so line boundaries are preserved inside the same card.
+			m.messages[i].content += strings.TrimPrefix(text, incomingHeader+"\n[output]\n")
+		}
+		return true
 	}
 	return false
 }

internal/tui/app_test.go

@@ -395,6 +395,28 @@ func TestToolLogOutputUpdatesExistingToolCard(t *testing.T) {
 	}
 }
 
+func TestToolLogStreamingChunksAppendToSameCard(t *testing.T) {
+	m := newModel(context.Background(), nil, Options{ModelName: "test-model"})
+	m.messages = append(m.messages, message{role: roleTool, content: `shell_run echo hi`})
+
+	if !m.mergeToolLog("shell_run echo hi\n[output]\none\n") {
+		t.Fatal("expected first mergeToolLog to return true")
+	}
+	if !m.mergeToolLog("shell_run echo hi\n[output]\ntwo\n") {
+		t.Fatal("expected second mergeToolLog to return true")
+	}
+	if len(m.messages) != 1 {
+		t.Fatalf("messages count = %d, want 1", len(m.messages))
+	}
+	plain := stripANSI(m.renderMessages())
+	if !strings.Contains(plain, "one") || !strings.Contains(plain, "two") {
+		t.Fatalf("rendered output missing streamed lines:\n%s", plain)
+	}
+	if strings.Contains(plain, "onetwo") {
+		t.Fatalf("streamed lines were concatenated:\n%s", plain)
+	}
+}
+
 type tuiCompactProvider struct{}
 
 func (tuiCompactProvider) ChatStream(ctx context.Context, req llm.ChatRequest, emit func(llm.StreamEvent) error) error {

pkg/config/config.go

@@ -14,6 +14,11 @@ import (
 
 const DefaultPath = "~/.agentu/config.yaml"
 
+const (
+	APITypeChat      = "chat"
+	APITypeResponses = "responses"
+)
+
 type Config struct {
 	Providers map[string]ProviderConfig `yaml:"providers"`
 	Agent     AgentConfig               `yaml:"agent"`
@@ -21,6 +26,7 @@ type Config struct {
 
 type ProviderConfig struct {
 	Name               string
+	APIType            string
 	BaseURL            string
 	Model              string
 	Models             []string
@@ -56,6 +62,7 @@ type rawConfig struct {
 }
 
 type rawProviderConfig struct {
+	APIType string           `yaml:"api_type"`
 	BaseURL string           `yaml:"base_url"`
 	Models  []rawModelConfig `yaml:"models"`
 	APIKey  string           `yaml:"api_key"`
@@ -153,6 +160,9 @@ func (c *Config) Validate() error {
 	for name, provider := range c.Providers {
 		var missing []string
 		prefix := "providers." + name
+		if !IsAPIType(provider.APIType) {
+			return fmt.Errorf("%s.api_type must be one of: %s", prefix, strings.Join(APITypes(), ", "))
+		}
 		if strings.TrimSpace(provider.BaseURL) == "" {
 			missing = append(missing, prefix+".base_url")
 		}
@@ -220,6 +230,7 @@ func normalizeProvider(name string, raw rawProviderConfig) ProviderConfig {
 	models, modelConfigs := normalizeModelConfigs(raw.Models)
 	provider := ProviderConfig{
 		Name:         name,
+		APIType:      normalizeAPIType(raw.APIType),
 		BaseURL:      raw.BaseURL,
 		Models:       models,
 		ModelConfigs: modelConfigs,
@@ -230,11 +241,44 @@ func normalizeProvider(name string, raw rawProviderConfig) ProviderConfig {
 	}
 	applyModelDefaults(&provider)
 	if provider.ThinkingConfigured && provider.ThinkingParam == "" {
-		provider.ThinkingParam = "thinking"
+		provider.ThinkingParam = DefaultThinkingParam(provider.APIType)
 	}
 	return provider
 }
 
+func normalizeAPIType(value string) string {
+	value = strings.ToLower(strings.TrimSpace(value))
+	if value == "" {
+		return APITypeChat
+	}
+	return value
+}
+
+func APITypes() []string {
+	return []string{APITypeChat, APITypeResponses}
+}
+
+func IsAPIType(value string) bool {
+	value = strings.ToLower(strings.TrimSpace(value))
+	for _, apiType := range APITypes() {
+		if value == apiType {
+			return true
+		}
+	}
+	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

@@ -370,3 +370,110 @@ providers:
 		t.Fatalf("err = %v", err)
 	}
 }
+
+func TestLoadDefaultsAPITypeToChat(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:
+  test:
+    base_url: https://example.com
+    api_key: ${AGENTU_TEST_KEY}
+    models:
+      - id: test-model
+`), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	cfg, err := Load(path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if got := cfg.Providers["test"].APIType; got != APITypeChat {
+		t.Fatalf("api type = %q, want %q", got, APITypeChat)
+	}
+}
+
+func TestLoadParsesResponsesAPIType(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:
+  openai:
+    api_type: RESPONSES
+    base_url: https://api.openai.com
+    api_key: ${AGENTU_TEST_KEY}
+    models:
+      - id: gpt-test
+`), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	cfg, err := Load(path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if got := cfg.Providers["openai"].APIType; got != APITypeResponses {
+		t.Fatalf("api type = %q, want %q", got, APITypeResponses)
+	}
+}
+
+func TestLoadRejectsInvalidAPIType(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:
+  test:
+    api_type: completions
+    base_url: https://example.com
+    api_key: ${AGENTU_TEST_KEY}
+    models:
+      - id: test-model
+`), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	_, err := Load(path)
+	if err == nil || !strings.Contains(err.Error(), "api_type") {
+		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

@@ -0,0 +1,481 @@
+package llm
+
+import (
+	"bufio"
+	"bytes"
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"net/http"
+	"sort"
+	"strings"
+)
+
+const responsesEndpoint = "/v1/responses"
+
+// OpenAIResponsesClient implements llm.Provider against the OpenAI Responses
+// API (/v1/responses). It maps the agent's chat-style conversation history to
+// Responses input items (messages, function_call, function_call_output) and
+// converts Responses streaming events back into the shared StreamEvent shape.
+type OpenAIResponsesClient struct {
+	baseURL    string
+	apiKey     string
+	httpClient *http.Client
+}
+
+func NewOpenAIResponsesClient(baseURL, apiKey string, httpClient *http.Client) *OpenAIResponsesClient {
+	if httpClient == nil {
+		httpClient = http.DefaultClient
+	}
+	return &OpenAIResponsesClient{
+		baseURL:    strings.TrimRight(baseURL, "/"),
+		apiKey:     apiKey,
+		httpClient: httpClient,
+	}
+}
+
+func (c *OpenAIResponsesClient) ChatStream(ctx context.Context, req ChatRequest, emit func(StreamEvent) error) error {
+	body, err := json.Marshal(buildResponsesRequest(req))
+	if err != nil {
+		return fmt.Errorf("marshal responses request: %w", err)
+	}
+	state := &responsesStreamState{}
+	return postResponsesSSE(ctx, c.baseURL, c.apiKey, c.httpClient, body, func(payload string) error {
+		return state.handle(payload, emit)
+	})
+}
+
+func postResponsesSSE(ctx context.Context, baseURL, apiKey string, httpClient *http.Client, body []byte, handle func(payload string) error) error {
+	var lastErr error
+	for attempt := 1; attempt <= maxChatAttempts; attempt++ {
+		httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, baseURL+responsesEndpoint, bytes.NewReader(body))
+		if err != nil {
+			return fmt.Errorf("create responses request: %w", err)
+		}
+		httpReq.Header.Set("Content-Type", "application/json")
+		httpReq.Header.Set("Accept", "text/event-stream")
+		if apiKey != "" {
+			httpReq.Header.Set("Authorization", "Bearer "+apiKey)
+		}
+
+		resp, err := httpClient.Do(httpReq)
+		if err != nil {
+			lastErr = fmt.Errorf("send responses request: %w", err)
+			if !shouldRetryRequest(ctx, attempt, 0) {
+				return lastErr
+			}
+			if waitErr := waitBeforeRetry(ctx); waitErr != nil {
+				return waitErr
+			}
+			continue
+		}
+
+		if resp.StatusCode < 200 || resp.StatusCode >= 300 {
+			limit := io.LimitReader(resp.Body, 4096)
+			data, _ := io.ReadAll(limit)
+			_ = resp.Body.Close()
+			lastErr = fmt.Errorf("responses request failed: status=%d body=%s", resp.StatusCode, strings.TrimSpace(string(data)))
+			if !shouldRetryRequest(ctx, attempt, resp.StatusCode) {
+				return lastErr
+			}
+			if waitErr := waitBeforeRetry(ctx); waitErr != nil {
+				return waitErr
+			}
+			continue
+		}
+
+		defer resp.Body.Close()
+		return readResponsesSSE(resp.Body, handle)
+	}
+	return lastErr
+}
+
+func readResponsesSSE(r io.Reader, handle func(payload string) error) error {
+	reader := bufio.NewReader(r)
+	for {
+		line, err := reader.ReadString('\n')
+		if err != nil && !errors.Is(err, io.EOF) {
+			return fmt.Errorf("read responses stream: %w", err)
+		}
+
+		line = strings.TrimSpace(line)
+		if strings.HasPrefix(line, "data:") {
+			payload := strings.TrimSpace(strings.TrimPrefix(line, "data:"))
+			if payload == "[DONE]" {
+				return nil
+			}
+			if handleErr := handle(payload); handleErr != nil {
+				return handleErr
+			}
+		}
+
+		if errors.Is(err, io.EOF) {
+			return nil
+		}
+	}
+}
+
+// buildResponsesRequest converts a ChatRequest into the /v1/responses body.
+func buildResponsesRequest(req ChatRequest) map[string]any {
+	payload := map[string]any{
+		"model":  req.Model,
+		"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)
+	}
+	if req.ToolChoice != "" {
+		payload["tool_choice"] = req.ToolChoice
+	}
+	for key, value := range req.Extra {
+		if _, exists := payload[key]; exists {
+			continue
+		}
+		if key == "reasoning" {
+			if level, ok := value.(string); ok {
+				payload[key] = map[string]any{"effort": mapReasoningEffort(level)}
+				continue
+			}
+		}
+		payload[key] = value
+	}
+	return payload
+}
+
+// buildResponsesInput maps chat-style messages to Responses input items.
+// Assistant tool calls become function_call items and tool results become
+// function_call_output items, preserving their position in the conversation.
+func buildResponsesInput(messages []Message) []any {
+	items := make([]any, 0, len(messages)+4)
+	for _, msg := range messages {
+		switch msg.Role {
+		case RoleSystem:
+			// 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:
+			if msg.Content != "" {
+				items = append(items, responsesMessageInput{Role: "assistant", Content: msg.Content})
+			}
+			for _, call := range msg.ToolCalls {
+				items = append(items, responsesFunctionCallInput{
+					Type:      "function_call",
+					CallID:    call.ID,
+					Name:      call.Function.Name,
+					Arguments: call.Function.Arguments,
+				})
+			}
+		case RoleTool:
+			items = append(items, responsesFunctionCallOutput{
+				Type:   "function_call_output",
+				CallID: msg.ToolCallID,
+				Output: msg.Content,
+			})
+		}
+	}
+	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 {
+		out = append(out, responsesTool{
+			Type:        "function",
+			Name:        tool.Function.Name,
+			Description: tool.Function.Description,
+			Parameters:  tool.Function.Parameters,
+		})
+	}
+	return out
+}
+
+func mapReasoningEffort(level string) string {
+	level = strings.ToLower(strings.TrimSpace(level))
+	if level == "middle" {
+		return "medium"
+	}
+	return level
+}
+
+type responsesMessageInput struct {
+	Role    string `json:"role"`
+	Content string `json:"content"`
+}
+
+type responsesFunctionCallInput struct {
+	Type      string `json:"type"`
+	CallID    string `json:"call_id"`
+	Name      string `json:"name"`
+	Arguments string `json:"arguments"`
+}
+
+type responsesFunctionCallOutput struct {
+	Type   string `json:"type"`
+	CallID string `json:"call_id"`
+	Output string `json:"output"`
+}
+
+type responsesTool struct {
+	Type        string          `json:"type"`
+	Name        string          `json:"name"`
+	Description string          `json:"description,omitempty"`
+	Parameters  json.RawMessage `json:"parameters,omitempty"`
+}
+
+// responsesStreamState tracks in-flight output items while parsing SSE events.
+type responsesStreamState struct {
+	calls map[int]*responsesCallBuilder
+	done  bool
+}
+
+type responsesCallBuilder struct {
+	index     int
+	id        string
+	callType  string
+	name      string
+	arguments strings.Builder
+	finalized bool
+}
+
+func (s *responsesStreamState) handle(payload string, emit func(StreamEvent) error) error {
+	if s.done {
+		return nil
+	}
+	var event responsesStreamEvent
+	if err := json.Unmarshal([]byte(payload), &event); err != nil {
+		return fmt.Errorf("parse responses stream payload: %w", err)
+	}
+
+	switch event.Type {
+	case "error":
+		s.done = true
+		return event.err()
+	case "response.failed":
+		s.done = true
+		if event.Response != nil && event.Response.Error != nil {
+			return fmt.Errorf("provider error: %s: %s", event.Response.Error.Code, event.Response.Error.Message)
+		}
+		return fmt.Errorf("provider error: response failed")
+	case "response.incomplete":
+		s.done = true
+		return s.finish(emit, "length", nil)
+	case "response.completed":
+		s.done = true
+		var usage *Usage
+		if event.Response != nil && event.Response.Usage != nil {
+			usage = &Usage{
+				PromptTokens:     event.Response.Usage.InputTokens,
+				CompletionTokens: event.Response.Usage.OutputTokens,
+				TotalTokens:      event.Response.Usage.TotalTokens,
+			}
+		}
+		return s.finish(emit, "stop", usage)
+	case "response.output_text.delta":
+		if event.Delta != "" {
+			return emit(StreamEvent{Content: event.Delta})
+		}
+	case "response.output_item.added":
+		if event.Item != nil && event.Item.Type == "function_call" {
+			s.builder(event.OutputIndex).applyAdded(*event.Item)
+		}
+	case "response.function_call_arguments.delta":
+		if event.Delta != "" {
+			s.builder(event.OutputIndex).arguments.WriteString(event.Delta)
+		}
+	case "response.function_call_arguments.done":
+		builder := s.builder(event.OutputIndex)
+		if event.Name != "" {
+			builder.name = event.Name
+		}
+		if event.Arguments != "" {
+			builder.arguments.Reset()
+			builder.arguments.WriteString(event.Arguments)
+		}
+	case "response.output_item.done":
+		if event.Item != nil && event.Item.Type == "function_call" {
+			builder := s.builder(event.OutputIndex)
+			builder.applyDone(*event.Item)
+			return s.emitCall(builder, emit)
+		}
+	}
+	return nil
+}
+
+// finish flushes any not-yet-finalized function calls and reports the
+// terminal finish reason.
+func (s *responsesStreamState) finish(emit func(StreamEvent) error, finishReason string, usage *Usage) error {
+	if len(s.calls) == 0 {
+		return emit(StreamEvent{FinishReason: finishReason, Usage: usage})
+	}
+	indexes := make([]int, 0, len(s.calls))
+	for index := range s.calls {
+		indexes = append(indexes, index)
+	}
+	sort.Ints(indexes)
+	var deltas []ToolCallDelta
+	for _, index := range indexes {
+		builder := s.calls[index]
+		if builder.finalized {
+			continue
+		}
+		builder.finalized = true
+		deltas = append(deltas, builder.delta())
+	}
+	if len(deltas) > 0 {
+		if err := emit(StreamEvent{ToolCalls: deltas}); err != nil {
+			return err
+		}
+	}
+	return emit(StreamEvent{FinishReason: finishReason, Usage: usage})
+}
+
+func (s *responsesStreamState) emitCall(builder *responsesCallBuilder, emit func(StreamEvent) error) error {
+	if builder.finalized {
+		return nil
+	}
+	builder.finalized = true
+	return emit(StreamEvent{ToolCalls: []ToolCallDelta{builder.delta()}})
+}
+
+func (s *responsesStreamState) builder(index int) *responsesCallBuilder {
+	if s.calls == nil {
+		s.calls = make(map[int]*responsesCallBuilder)
+	}
+	builder := s.calls[index]
+	if builder == nil {
+		builder = &responsesCallBuilder{index: index}
+		s.calls[index] = builder
+	}
+	return builder
+}
+
+func (b *responsesCallBuilder) applyAdded(item responsesStreamItem) {
+	if b.id == "" {
+		b.id = item.CallID
+		if b.id == "" {
+			b.id = item.ID
+		}
+	}
+	if b.name == "" {
+		b.name = item.Name
+	}
+	if item.Arguments != "" {
+		b.arguments.Reset()
+		b.arguments.WriteString(item.Arguments)
+	}
+}
+
+func (b *responsesCallBuilder) applyDone(item responsesStreamItem) {
+	if item.CallID != "" {
+		b.id = item.CallID
+	} else if item.ID != "" {
+		b.id = item.ID
+	}
+	if item.Name != "" {
+		b.name = item.Name
+	}
+	if item.Arguments != "" {
+		b.arguments.Reset()
+		b.arguments.WriteString(item.Arguments)
+	}
+}
+
+func (b *responsesCallBuilder) delta() ToolCallDelta {
+	id := b.id
+	if id == "" {
+		id = fmt.Sprintf("call_%d", b.index)
+	}
+	callType := b.callType
+	if callType == "" {
+		callType = "function"
+	}
+	return ToolCallDelta{
+		Index:     b.index,
+		ID:        id,
+		Type:      callType,
+		Name:      b.name,
+		Arguments: b.arguments.String(),
+	}
+}
+
+type responsesStreamEvent struct {
+	Type        string                   `json:"type"`
+	OutputIndex int                      `json:"output_index"`
+	Delta       string                   `json:"delta"`
+	Arguments   string                   `json:"arguments"`
+	Name        string                   `json:"name"`
+	Code        string                   `json:"code"`
+	Message     string                   `json:"message"`
+	Param       string                   `json:"param"`
+	Item        *responsesStreamItem     `json:"item"`
+	Response    *responsesStreamResponse `json:"response"`
+	Error       *responsesStreamError    `json:"error"`
+}
+
+type responsesStreamItem struct {
+	Type      string `json:"type"`
+	ID        string `json:"id"`
+	CallID    string `json:"call_id"`
+	Name      string `json:"name"`
+	Arguments string `json:"arguments"`
+}
+
+type responsesStreamResponse struct {
+	Status string                `json:"status"`
+	Error  *responsesStreamError `json:"error"`
+	Usage  *responsesStreamUsage `json:"usage"`
+}
+
+type responsesStreamUsage struct {
+	InputTokens  int `json:"input_tokens"`
+	OutputTokens int `json:"output_tokens"`
+	TotalTokens  int `json:"total_tokens"`
+}
+
+type responsesStreamError struct {
+	Type    string `json:"type"`
+	Code    string `json:"code"`
+	Message string `json:"message"`
+	Param   string `json:"param"`
+}
+
+func (e responsesStreamEvent) err() error {
+	code, message, param := e.Code, e.Message, e.Param
+	if code == "" && message == "" && e.Error != nil {
+		code, message, param = e.Error.Code, e.Error.Message, e.Error.Param
+	}
+	if message == "" {
+		if param != "" {
+			return fmt.Errorf("provider error: %s", param)
+		}
+		return fmt.Errorf("provider error")
+	}
+	if code != "" {
+		return fmt.Errorf("provider error: %s: %s", code, message)
+	}
+	return fmt.Errorf("provider error: %s", message)
+}

pkg/llm/responses_test.go

@@ -0,0 +1,337 @@
+package llm
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+)
+
+func TestBuildResponsesRequest(t *testing.T) {
+	req := ChatRequest{
+		Model: "gpt-test",
+		Messages: []Message{
+			{Role: RoleSystem, Content: "be concise"},
+			{Role: RoleUser, Content: "hi"},
+			{
+				Role:    RoleAssistant,
+				Content: "let me check",
+				ToolCalls: []ToolCall{{
+					ID:   "call_1",
+					Type: "function",
+					Function: FunctionCall{
+						Name:      "file_read",
+						Arguments: `{"path":"README.md"}`,
+					},
+				}},
+			},
+			{Role: RoleTool, ToolCallID: "call_1", Content: "contents"},
+			{Role: RoleAssistant, Content: "done"},
+		},
+		Tools: []Tool{{
+			Type: "function",
+			Function: ToolFunction{
+				Name:        "file_read",
+				Description: "Read a local file",
+				Parameters:  json.RawMessage(`{"type":"object","properties":{}}`),
+			},
+		}},
+		ToolChoice: "auto",
+		Extra: map[string]any{
+			"reasoning":   "middle",
+			"thinking":    "high",
+			"temperature": 0.2,
+		},
+	}
+
+	payload := buildResponsesRequest(req)
+	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["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) != 5 {
+		t.Fatalf("input = %#v", raw["input"])
+	}
+	checkInputItem := func(index int, wantType, wantKey, wantValue string) {
+		t.Helper()
+		item, ok := input[index].(map[string]any)
+		if !ok {
+			t.Fatalf("input[%d] = %#v", index, input[index])
+		}
+		if item[wantKey] != wantValue {
+			t.Fatalf("input[%d].%s = %#v, want %q", index, wantKey, item[wantKey], wantValue)
+		}
+		if wantType != "" && item["type"] != wantType {
+			t.Fatalf("input[%d].type = %#v, want %q", index, item["type"], wantType)
+		}
+	}
+	checkInputItem(0, "", "role", "user")
+	checkInputItem(1, "", "role", "assistant")
+	checkInputItem(2, "function_call", "call_id", "call_1")
+	checkInputItem(3, "function_call_output", "call_id", "call_1")
+	checkInputItem(4, "", "role", "assistant")
+	if item := input[2].(map[string]any); item["name"] != "file_read" || item["arguments"] != `{"path":"README.md"}` {
+		t.Fatalf("function_call item = %#v", item)
+	}
+	if item := input[3].(map[string]any); item["output"] != "contents" {
+		t.Fatalf("function_call_output item = %#v", item)
+	}
+
+	tools, ok := raw["tools"].([]any)
+	if !ok || len(tools) != 1 {
+		t.Fatalf("tools = %#v", raw["tools"])
+	}
+	tool := tools[0].(map[string]any)
+	if tool["type"] != "function" || tool["name"] != "file_read" || tool["description"] != "Read a local file" {
+		t.Fatalf("tool = %#v", tool)
+	}
+	if _, ok := tool["parameters"].(map[string]any); !ok {
+		t.Fatalf("tool parameters = %#v", tool["parameters"])
+	}
+
+	reasoning, ok := raw["reasoning"].(map[string]any)
+	if !ok || reasoning["effort"] != "medium" {
+		t.Fatalf("reasoning = %#v", raw["reasoning"])
+	}
+	if raw["thinking"] != "high" {
+		t.Fatalf("thinking = %#v", raw["thinking"])
+	}
+	if raw["temperature"] != 0.2 {
+		t.Fatalf("temperature = %#v", raw["temperature"])
+	}
+}
+
+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" {
+			t.Fatalf("path = %s", r.URL.Path)
+		}
+		if got := r.Header.Get("Authorization"); got != "Bearer sk-test" {
+			t.Fatalf("authorization = %q", got)
+		}
+		body := mustReadBody(t, r)
+		var raw map[string]any
+		if err := json.NewDecoder(strings.NewReader(body)).Decode(&raw); err != nil {
+			t.Fatal(err)
+		}
+		if raw["stream"] != true {
+			t.Fatal("request did not enable stream")
+		}
+		if raw["tool_choice"] != "auto" {
+			t.Fatalf("tool_choice = %#v", raw["tool_choice"])
+		}
+		input, ok := raw["input"].([]any)
+		if !ok || len(input) == 0 {
+			t.Fatalf("input = %#v", raw["input"])
+		}
+
+		w.Header().Set("Content-Type", "text/event-stream")
+		events := []string{
+			`{"type":"response.output_item.added","output_index":0,"item":{"type":"message","id":"msg_1","role":"assistant","content":[]}}`,
+			`{"type":"response.output_text.delta","item_id":"msg_1","output_index":0,"content_index":0,"delta":"Hel"}`,
+			`{"type":"response.output_text.delta","item_id":"msg_1","output_index":0,"content_index":0,"delta":"lo"}`,
+			`{"type":"response.output_item.added","output_index":1,"item":{"type":"function_call","id":"fc_1","call_id":"call_1","name":"file_read","arguments":""}}`,
+			`{"type":"response.function_call_arguments.delta","item_id":"fc_1","output_index":1,"delta":"{\"path\""}`,
+			`{"type":"response.function_call_arguments.delta","item_id":"fc_1","output_index":1,"delta":":\"README.md\"}"}`,
+			`{"type":"response.function_call_arguments.done","item_id":"fc_1","output_index":1,"name":"file_read","arguments":"{\"path\":\"README.md\"}"}`,
+			`{"type":"response.output_item.done","output_index":1,"item":{"type":"function_call","id":"fc_1","call_id":"call_1","name":"file_read","arguments":"{\"path\":\"README.md\"}"}}`,
+			`{"type":"response.output_item.done","output_index":0,"item":{"type":"message","id":"msg_1","role":"assistant","content":[{"type":"output_text","text":"Hello"}]}}`,
+			`{"type":"response.completed","response":{"id":"resp_1","status":"completed","usage":{"input_tokens":10,"output_tokens":20,"total_tokens":30}}}`,
+		}
+		for _, event := range events {
+			_, _ = w.Write([]byte("data: " + event + "\n\n"))
+		}
+	}))
+	defer server.Close()
+
+	client := NewOpenAIResponsesClient(server.URL, "sk-test", server.Client())
+	var content strings.Builder
+	var calls []ToolCallDelta
+	var finishes []string
+	var usage *Usage
+	err := client.ChatStream(context.Background(), ChatRequest{
+		Model:      "gpt-test",
+		Messages:   []Message{{Role: RoleUser, Content: "hi"}},
+		ToolChoice: "auto",
+		Tools: []Tool{{
+			Type: "function",
+			Function: ToolFunction{
+				Name:        "file_read",
+				Description: "Read a local file",
+				Parameters:  json.RawMessage(`{"type":"object"}`),
+			},
+		}},
+	}, func(event StreamEvent) error {
+		content.WriteString(event.Content)
+		calls = append(calls, event.ToolCalls...)
+		if event.Usage != nil {
+			usage = event.Usage
+		}
+		if event.FinishReason != "" {
+			finishes = append(finishes, event.FinishReason)
+		}
+		return nil
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if content.String() != "Hello" {
+		t.Fatalf("content = %q", content.String())
+	}
+	if len(calls) != 1 {
+		t.Fatalf("calls = %#v", calls)
+	}
+	call := calls[0]
+	if call.Index != 1 || call.ID != "call_1" || call.Name != "file_read" {
+		t.Fatalf("call = %#v", call)
+	}
+	if call.Arguments != `{"path":"README.md"}` {
+		t.Fatalf("arguments = %q", call.Arguments)
+	}
+	if len(finishes) != 1 || finishes[0] != "stop" {
+		t.Fatalf("finishes = %#v", finishes)
+	}
+	if usage == nil || usage.PromptTokens != 10 || usage.CompletionTokens != 20 || usage.TotalTokens != 30 {
+		t.Fatalf("usage = %#v", usage)
+	}
+}
+
+func TestOpenAIResponsesClientErrorEvent(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":"error","code":"invalid_request_error","message":"bad request"}` + "\n\n"))
+	}))
+	defer server.Close()
+
+	client := NewOpenAIResponsesClient(server.URL, "sk-test", server.Client())
+	err := client.ChatStream(context.Background(), ChatRequest{Model: "test"}, func(StreamEvent) error {
+		return nil
+	})
+	if err == nil || !strings.Contains(err.Error(), "bad request") {
+		t.Fatalf("err = %v", err)
+	}
+}
+
+func TestOpenAIResponsesClientFailedEvent(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.failed","response":{"id":"resp_1","status":"failed","error":{"code":"server_error","message":"boom"}}}` + "\n\n"))
+	}))
+	defer server.Close()
+
+	client := NewOpenAIResponsesClient(server.URL, "sk-test", server.Client())
+	err := client.ChatStream(context.Background(), ChatRequest{Model: "test"}, func(StreamEvent) error {
+		return nil
+	})
+	if err == nil || !strings.Contains(err.Error(), "boom") {
+		t.Fatalf("err = %v", err)
+	}
+}
+
+func TestOpenAIResponsesClientIncomplete(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":"partial"}` + "\n\n"))
+		_, _ = w.Write([]byte(`data: {"type":"response.incomplete","response":{"id":"resp_1","status":"incomplete","incomplete_details":{"reason":"max_tokens"}}}` + "\n\n"))
+	}))
+	defer server.Close()
+
+	client := NewOpenAIResponsesClient(server.URL, "sk-test", server.Client())
+	var content strings.Builder
+	var finishes []string
+	err := client.ChatStream(context.Background(), ChatRequest{Model: "test"}, func(event StreamEvent) error {
+		content.WriteString(event.Content)
+		if event.FinishReason != "" {
+			finishes = append(finishes, event.FinishReason)
+		}
+		return nil
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if content.String() != "partial" {
+		t.Fatalf("content = %q", content.String())
+	}
+	if len(finishes) != 1 || finishes[0] != "length" {
+		t.Fatalf("finishes = %#v", finishes)
+	}
+}
+
+func TestOpenAIResponsesClientRetriesTransientHTTPError(t *testing.T) {
+	attempts := 0
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		attempts++
+		if attempts == 1 {
+			http.Error(w, "temporary", http.StatusBadGateway)
+			return
+		}
+		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()
+
+	client := NewOpenAIResponsesClient(server.URL, "sk-test", server.Client())
+	var content strings.Builder
+	err := client.ChatStream(context.Background(), ChatRequest{Model: "test"}, func(event StreamEvent) error {
+		content.WriteString(event.Content)
+		return nil
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if attempts != 2 {
+		t.Fatalf("attempts = %d, want 2", attempts)
+	}
+	if content.String() != "ok" {
+		t.Fatalf("content = %q", content.String())
+	}
+}