refactor(agent): integrate Fantasy SDK, streaming, and MemGPT memory

Core agent loop migration to Fantasy SDK:
- Replace providers.LLMProvider with fantasy.LanguageModel interface
- Add streaming agent loop (runAgentLoopStreaming) with real-time token
  delivery via bus StreamDelta messages
- Inject MemGPT working context into system prompt automatically
- Register memory tool for agent self-managed persistent memory
- Add MemGPT tool implementation (search/write/status operations)
- Exponential backoff retry with jitter for transient LLM failures

Supporting changes:
- bus: add PublishStreamDelta for real-time token streaming
- bus: add StreamDelta message type
- channels: migrate from providers.Message to messages.Message
- config: add ProgressiveDisclosure and memory configuration fields
- heartbeat: update imports for messages package
- cmd/main: wire Fantasy provider and language model creation

Includes integration test scaffolding for the full agent loop.

cmd/picoclaw/main.go

@@ -26,10 +26,10 @@ import (
 	"github.com/sipeed/picoclaw/pkg/config"
 	"github.com/sipeed/picoclaw/pkg/cron"
 	"github.com/sipeed/picoclaw/pkg/devices"
+	picofantasy "github.com/sipeed/picoclaw/pkg/fantasy"
 	"github.com/sipeed/picoclaw/pkg/heartbeat"
 	"github.com/sipeed/picoclaw/pkg/logger"
 	"github.com/sipeed/picoclaw/pkg/migrate"
-	"github.com/sipeed/picoclaw/pkg/providers"
 	"github.com/sipeed/picoclaw/pkg/skills"
 	"github.com/sipeed/picoclaw/pkg/state"
 	"github.com/sipeed/picoclaw/pkg/tools"
@@ -508,14 +508,21 @@ func agentCmd() {
 		os.Exit(1)
 	}
 
-	provider, err := providers.CreateProvider(cfg)
+	fantasyProvider, err := picofantasy.CreateProvider(cfg)
 	if err != nil {
 		fmt.Printf("Error creating provider: %v\n", err)
 		os.Exit(1)
 	}
 
+	ctx := context.Background()
+	languageModel, err := fantasyProvider.LanguageModel(ctx, picofantasy.ModelID(cfg))
+	if err != nil {
+		fmt.Printf("Error creating language model: %v\n", err)
+		os.Exit(1)
+	}
+
 	msgBus := bus.NewMessageBus()
-	agentLoop := agent.NewAgentLoop(cfg, msgBus, provider)
+	agentLoop := agent.NewAgentLoop(cfg, msgBus, languageModel)
 
 	// Print agent startup info (only for interactive mode)
 	startupInfo := agentLoop.GetStartupInfo()
@@ -527,7 +534,6 @@ func agentCmd() {
 		})
 
 	if message != "" {
-		ctx := context.Background()
 		response, err := agentLoop.ProcessDirect(ctx, message, sessionKey)
 		if err != nil {
 			fmt.Printf("Error: %v\n", err)
@@ -643,14 +649,20 @@ func gatewayCmd() {
 		os.Exit(1)
 	}
 
-	provider, err := providers.CreateProvider(cfg)
+	fantasyProvider, err := picofantasy.CreateProvider(cfg)
 	if err != nil {
 		fmt.Printf("Error creating provider: %v\n", err)
 		os.Exit(1)
 	}
 
+	languageModel, err := fantasyProvider.LanguageModel(context.Background(), picofantasy.ModelID(cfg))
+	if err != nil {
+		fmt.Printf("Error creating language model: %v\n", err)
+		os.Exit(1)
+	}
+
 	msgBus := bus.NewMessageBus()
-	agentLoop := agent.NewAgentLoop(cfg, msgBus, provider)
+	agentLoop := agent.NewAgentLoop(cfg, msgBus, languageModel)
 
 	// Print agent startup info
 	fmt.Println("\n📦 Agent Status:")

pkg/agent/context.go

@@ -1,6 +1,7 @@
 package agent
 
 import (
+	"context"
 	"fmt"
 	"os"
 	"path/filepath"
@@ -9,7 +10,8 @@ import (
 	"time"
 
 	"github.com/sipeed/picoclaw/pkg/logger"
-	"github.com/sipeed/picoclaw/pkg/providers"
+	"github.com/sipeed/picoclaw/pkg/memory"
+	"github.com/sipeed/picoclaw/pkg/messages"
 	"github.com/sipeed/picoclaw/pkg/skills"
 	"github.com/sipeed/picoclaw/pkg/tools"
 )
@@ -17,7 +19,8 @@ import (
 type ContextBuilder struct {
 	workspace    string
 	skillsLoader *skills.SkillsLoader
-	memory       *MemoryStore
+	memory       *MemoryStore        // Legacy file-based memory
+	memoryStore  memory.Memory       // New 3-tier MemGPT memory (may be nil)
 	tools        *tools.ToolRegistry // Direct reference to tool registry
 }
 
@@ -48,6 +51,11 @@ func (cb *ContextBuilder) SetToolsRegistry(registry *tools.ToolRegistry) {
 	cb.tools = registry
 }
 
+// SetMemoryStore sets the 3-tier MemGPT memory store for working context injection.
+func (cb *ContextBuilder) SetMemoryStore(ms memory.Memory) {
+	cb.memoryStore = ms
+}
+
 func (cb *ContextBuilder) getIdentity() string {
 	now := time.Now().Format("2006-01-02 15:04 (Monday)")
 	workspacePath, _ := filepath.Abs(filepath.Join(cb.workspace))
@@ -128,12 +136,20 @@ The following skills extend your capabilities. To use a skill, read its SKILL.md
 %s`, skillsSummary))
 	}
 
-	// Memory context
+	// Legacy file-based memory context
 	memoryContext := cb.memory.GetMemoryContext()
 	if memoryContext != "" {
 		parts = append(parts, "# Memory\n\n"+memoryContext)
 	}
 
+	// 3-tier MemGPT working context injection
+	if cb.memoryStore != nil {
+		wcSection := cb.buildWorkingContextSection()
+		if wcSection != "" {
+			parts = append(parts, wcSection)
+		}
+	}
+
 	// Join with "---" separator
 	return strings.Join(parts, "\n\n---\n\n")
 }
@@ -157,8 +173,46 @@ func (cb *ContextBuilder) LoadBootstrapFiles() string {
 	return result
 }
 
-func (cb *ContextBuilder) BuildMessages(history []providers.Message, summary string, currentMessage string, media []string, channel, chatID string) []providers.Message {
+// buildWorkingContextSection returns the working context section for the system prompt.
-	messages := []providers.Message{}
+// It includes the hot-tier working context buffer and memory usage instructions.
+func (cb *ContextBuilder) buildWorkingContextSection() string {
+	if cb.memoryStore == nil {
+		return ""
+	}
+
+	// Use a background context for system prompt building (non-blocking)
+	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Second)
+	defer cancel()
+
+	var parts []string
+
+	// Inject working context (hot tier)
+	wc, err := cb.memoryStore.GetWorkingContext(ctx, "picoclaw", "default")
+	if err == nil && wc != "" {
+		parts = append(parts, "## Working Context\n\n"+wc)
+	}
+
+	// Memory system instructions
+	parts = append(parts, `## Memory System
+
+You have a 3-tier memory system accessible via the **memory** tool:
+
+- **Working Context** (hot): Always loaded. Use memory tool to update it with durable facts.
+- **Recall** (warm): Scored memory items. Search returns the most relevant.
+- **Archival** (cold): Large documents chunked and embedded for retrieval.
+
+Use the memory tool to:
+- **search**: Find relevant memories across all tiers.
+- **write**: Store important facts, preferences, decisions.
+- **status**: Check memory pressure and usage.
+
+Store important user preferences, key decisions, and facts you want to remember long-term.`)
+
+	return strings.Join(parts, "\n\n")
+}
+
+func (cb *ContextBuilder) BuildMessages(history []messages.Message, summary string, currentMessage string, media []string, channel, chatID string) []messages.Message {
+	msgs := []messages.Message{}
 
 	systemPrompt := cb.BuildSystemPrompt()
 
@@ -189,49 +243,41 @@ func (cb *ContextBuilder) BuildMessages(history []providers.Message, summary str
 		systemPrompt += "\n\n## Summary of Previous Conversation\n\n" + summary
 	}
 
-	//This fix prevents the session memory from LLM failure due to elimination of toolu_IDs required from LLM
+	// Note: Orphaned tool messages are now prevented at the source by
-	// --- INICIO DEL FIX ---
+	// tool-call-aware truncation in SessionManager.TruncateHistory().
-	//Diegox-17
-	for len(history) > 0 && (history[0].Role == "tool") {
-		logger.DebugCF("agent", "Removing orphaned tool message from history to prevent LLM error",
-			map[string]interface{}{"role": history[0].Role})
-		history = history[1:]
-	}
-	//Diegox-17
-	// --- FIN DEL FIX ---
 
-	messages = append(messages, providers.Message{
+	msgs = append(msgs, messages.Message{
 		Role:    "system",
 		Content: systemPrompt,
 	})
 
-	messages = append(messages, history...)
+	msgs = append(msgs, history...)
 
-	messages = append(messages, providers.Message{
+	msgs = append(msgs, messages.Message{
 		Role:    "user",
 		Content: currentMessage,
 	})
 
-	return messages
+	return msgs
 }
 
-func (cb *ContextBuilder) AddToolResult(messages []providers.Message, toolCallID, toolName, result string) []providers.Message {
+func (cb *ContextBuilder) AddToolResult(msgs []messages.Message, toolCallID, toolName, result string) []messages.Message {
-	messages = append(messages, providers.Message{
+	msgs = append(msgs, messages.Message{
 		Role:       "tool",
 		Content:    result,
 		ToolCallID: toolCallID,
 	})
-	return messages
+	return msgs
 }
 
-func (cb *ContextBuilder) AddAssistantMessage(messages []providers.Message, content string, toolCalls []map[string]interface{}) []providers.Message {
+func (cb *ContextBuilder) AddAssistantMessage(msgs []messages.Message, content string, toolCalls []map[string]interface{}) []messages.Message {
-	msg := providers.Message{
+	msg := messages.Message{
 		Role:    "assistant",
 		Content: content,
 	}
 	// Always add assistant message, whether or not it has tool calls
-	messages = append(messages, msg)
+	msgs = append(msgs, msg)
-	return messages
+	return msgs
 }
 
 func (cb *ContextBuilder) loadSkills() string {

pkg/agent/integration_test.go

@@ -0,0 +1,560 @@
+package agent
+
+import (
+	"context"
+	"fmt"
+	"os"
+	"strings"
+	"testing"
+	"time"
+
+	fantasy "charm.land/fantasy"
+	"github.com/sipeed/picoclaw/pkg/bus"
+	"github.com/sipeed/picoclaw/pkg/config"
+	"github.com/sipeed/picoclaw/pkg/tools"
+)
+
+// --- Mock language model that simulates tool calls ---
+
+// toolCallingModel simulates an LLM that requests tool calls on first round,
+// then produces a final text response incorporating tool results.
+type toolCallingModel struct {
+	callCount int
+}
+
+func (m *toolCallingModel) Generate(_ context.Context, call fantasy.Call) (*fantasy.Response, error) {
+	m.callCount++
+
+	// First call: check if any tool results already exist in the prompt.
+	// If no tool results found, request a tool call.
+	hasToolResults := false
+	for _, msg := range call.Prompt {
+		for _, part := range msg.Content {
+			if part.GetType() == fantasy.ContentTypeToolResult {
+				hasToolResults = true
+			}
+		}
+	}
+
+	if !hasToolResults && len(call.Tools) > 0 {
+		// Request a tool call
+		return &fantasy.Response{
+			Content: fantasy.ResponseContent{
+				fantasy.ToolCallContent{
+					ToolCallID: "call-1",
+					ToolName:   "echo",
+					Input:      `{"text": "hello from tool"}`,
+				},
+			},
+			FinishReason: fantasy.FinishReasonToolCalls,
+		}, nil
+	}
+
+	// After tool results: produce final response
+	return &fantasy.Response{
+		Content: fantasy.ResponseContent{
+			fantasy.TextContent{Text: "Integration test response with tool output"},
+		},
+		FinishReason: fantasy.FinishReasonStop,
+	}, nil
+}
+
+func (m *toolCallingModel) Stream(_ context.Context, call fantasy.Call) (fantasy.StreamResponse, error) {
+	resp, err := m.Generate(context.Background(), call)
+	if err != nil {
+		return nil, err
+	}
+
+	return func(yield func(fantasy.StreamPart) bool) {
+		// Check if response has tool calls
+		hasToolCalls := false
+		for _, c := range resp.Content {
+			if c.GetType() == fantasy.ContentTypeToolCall {
+				hasToolCalls = true
+			}
+		}
+
+		if hasToolCalls {
+			// Emit tool calls as stream parts
+			for _, c := range resp.Content {
+				if tc, ok := c.(fantasy.ToolCallContent); ok {
+					if !yield(fantasy.StreamPart{
+						Type:          fantasy.StreamPartTypeToolCall,
+						ID:            tc.ToolCallID,
+						ToolCallName:  tc.ToolName,
+						ToolCallInput: tc.Input,
+					}) {
+						return
+					}
+				}
+			}
+			if !yield(fantasy.StreamPart{Type: fantasy.StreamPartTypeFinish, FinishReason: fantasy.FinishReasonToolCalls}) {
+				return
+			}
+		} else {
+			// Emit text as proper stream sequence
+			text := resp.Content.Text()
+			if text != "" {
+				if !yield(fantasy.StreamPart{Type: fantasy.StreamPartTypeTextStart, ID: "text-0"}) {
+					return
+				}
+				// Emit text in chunks to simulate real streaming
+				for i := 0; i < len(text); i += 10 {
+					end := i + 10
+					if end > len(text) {
+						end = len(text)
+					}
+					if !yield(fantasy.StreamPart{Type: fantasy.StreamPartTypeTextDelta, ID: "text-0", Delta: text[i:end]}) {
+						return
+					}
+				}
+				if !yield(fantasy.StreamPart{Type: fantasy.StreamPartTypeTextEnd, ID: "text-0"}) {
+					return
+				}
+			}
+			yield(fantasy.StreamPart{Type: fantasy.StreamPartTypeFinish, FinishReason: fantasy.FinishReasonStop})
+		}
+	}, nil
+}
+
+func (m *toolCallingModel) GenerateObject(_ context.Context, _ fantasy.ObjectCall) (*fantasy.ObjectResponse, error) {
+	return nil, fmt.Errorf("not implemented")
+}
+
+func (m *toolCallingModel) StreamObject(_ context.Context, _ fantasy.ObjectCall) (fantasy.ObjectStreamResponse, error) {
+	return nil, fmt.Errorf("not implemented")
+}
+
+func (m *toolCallingModel) Provider() string { return "mock" }
+func (m *toolCallingModel) Model() string    { return "mock-tool-model" }
+
+// --- Simple echo tool for integration testing ---
+
+type echoTool struct{}
+
+func (t *echoTool) Name() string        { return "echo" }
+func (t *echoTool) Description() string { return "Echo the given text back" }
+func (t *echoTool) Parameters() map[string]interface{} {
+	return map[string]interface{}{
+		"type": "object",
+		"properties": map[string]interface{}{
+			"text": map[string]interface{}{
+				"type":        "string",
+				"description": "Text to echo",
+			},
+		},
+		"required": []string{"text"},
+	}
+}
+
+func (t *echoTool) Execute(_ context.Context, args map[string]interface{}) *tools.ToolResult {
+	text, _ := args["text"].(string)
+	return &tools.ToolResult{
+		ForLLM:  "Echo: " + text,
+		ForUser: "",
+		Silent:  true,
+		IsError: false,
+	}
+}
+
+// --- Integration Tests ---
+
+// TestIntegration_FullAgentLoop_SimpleResponse tests the full agent loop
+// with a simple mock model that returns text directly (no tool calls).
+func TestIntegration_FullAgentLoop_SimpleResponse(t *testing.T) {
+	tmpDir, err := os.MkdirTemp("", "agent-integration-*")
+	if err != nil {
+		t.Fatalf("Failed to create temp dir: %v", err)
+	}
+	defer os.RemoveAll(tmpDir)
+
+	cfg := &config.Config{
+		Agents: config.AgentsConfig{
+			Defaults: config.AgentDefaults{
+				Workspace:         tmpDir,
+				Model:             "test-model",
+				MaxTokens:         4096,
+				MaxToolIterations: 10,
+			},
+		},
+	}
+
+	msgBus := bus.NewMessageBus()
+	model := newMockLanguageModel("Hello from Fantasy agent")
+	al := NewAgentLoop(cfg, msgBus, model)
+
+	ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
+	defer cancel()
+
+	msg := bus.InboundMessage{
+		Channel:    "test",
+		SenderID:   "user1",
+		ChatID:     "chat1",
+		Content:    "Say hello",
+		SessionKey: "test-session-simple",
+	}
+
+	response, err := al.processMessage(ctx, msg)
+	if err != nil {
+		t.Fatalf("processMessage failed: %v", err)
+	}
+
+	if response != "Hello from Fantasy agent" {
+		t.Errorf("Expected 'Hello from Fantasy agent', got: %s", response)
+	}
+
+	// Verify session has messages saved
+	history := al.sessions.GetHistory("test-session-simple")
+	if len(history) == 0 {
+		t.Error("Expected session history to have messages")
+	}
+
+	// First message should be the user's
+	foundUser := false
+	for _, m := range history {
+		if m.Role == "user" && m.Content == "Say hello" {
+			foundUser = true
+			break
+		}
+	}
+	if !foundUser {
+		t.Error("Expected user message in session history")
+	}
+}
+
+// TestIntegration_FullAgentLoop_WithToolCalls tests the full agent loop
+// including tool call execution and response incorporation.
+func TestIntegration_FullAgentLoop_WithToolCalls(t *testing.T) {
+	tmpDir, err := os.MkdirTemp("", "agent-integration-tools-*")
+	if err != nil {
+		t.Fatalf("Failed to create temp dir: %v", err)
+	}
+	defer os.RemoveAll(tmpDir)
+
+	cfg := &config.Config{
+		Agents: config.AgentsConfig{
+			Defaults: config.AgentDefaults{
+				Workspace:         tmpDir,
+				Model:             "mock-tool-model",
+				MaxTokens:         4096,
+				MaxToolIterations: 10,
+			},
+		},
+	}
+
+	msgBus := bus.NewMessageBus()
+	model := &toolCallingModel{}
+	al := NewAgentLoop(cfg, msgBus, model)
+
+	// Register the echo tool
+	al.RegisterTool(&echoTool{})
+
+	ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
+	defer cancel()
+
+	msg := bus.InboundMessage{
+		Channel:    "test",
+		SenderID:   "user1",
+		ChatID:     "chat1",
+		Content:    "Use the echo tool",
+		SessionKey: "test-session-tools",
+	}
+
+	response, err := al.processMessage(ctx, msg)
+	if err != nil {
+		t.Fatalf("processMessage failed: %v", err)
+	}
+
+	// The model returns "Integration test response with tool output" after tool execution
+	if !strings.Contains(response, "Integration test response") {
+		t.Errorf("Expected response to contain 'Integration test response', got: %s", response)
+	}
+
+	// Model should have been called at least twice (tool call + final response)
+	if model.callCount < 2 {
+		t.Errorf("Expected model to be called at least 2 times, got: %d", model.callCount)
+	}
+}
+
+// TestIntegration_ProcessDirect tests the ProcessDirect method
+// which is used by CLI mode for one-shot message processing.
+func TestIntegration_ProcessDirect(t *testing.T) {
+	tmpDir, err := os.MkdirTemp("", "agent-integration-direct-*")
+	if err != nil {
+		t.Fatalf("Failed to create temp dir: %v", err)
+	}
+	defer os.RemoveAll(tmpDir)
+
+	cfg := &config.Config{
+		Agents: config.AgentsConfig{
+			Defaults: config.AgentDefaults{
+				Workspace:         tmpDir,
+				Model:             "test-model",
+				MaxTokens:         4096,
+				MaxToolIterations: 10,
+			},
+		},
+	}
+
+	msgBus := bus.NewMessageBus()
+	model := newMockLanguageModel("Direct CLI response")
+	al := NewAgentLoop(cfg, msgBus, model)
+
+	ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
+	defer cancel()
+
+	response, err := al.ProcessDirect(ctx, "Direct message", "direct-session")
+	if err != nil {
+		t.Fatalf("ProcessDirect failed: %v", err)
+	}
+
+	if response != "Direct CLI response" {
+		t.Errorf("Expected 'Direct CLI response', got: %s", response)
+	}
+}
+
+// --- Streaming-specific mock ---
+
+// streamingModel simulates an LLM with proper streaming token emission.
+// It emits tokens one word at a time via Stream() and also supports Generate()
+// for non-streaming fallback.
+type streamingModel struct {
+	words []string
+}
+
+func newStreamingModel(text string) *streamingModel {
+	return &streamingModel{words: strings.Fields(text)}
+}
+
+func (m *streamingModel) Generate(_ context.Context, _ fantasy.Call) (*fantasy.Response, error) {
+	fullText := strings.Join(m.words, " ")
+	return &fantasy.Response{
+		Content:      fantasy.ResponseContent{fantasy.TextContent{Text: fullText}},
+		FinishReason: fantasy.FinishReasonStop,
+	}, nil
+}
+
+func (m *streamingModel) Stream(_ context.Context, _ fantasy.Call) (fantasy.StreamResponse, error) {
+	words := m.words
+	return func(yield func(fantasy.StreamPart) bool) {
+		if !yield(fantasy.StreamPart{Type: fantasy.StreamPartTypeTextStart, ID: "s-0"}) {
+			return
+		}
+		for i, w := range words {
+			delta := w
+			if i < len(words)-1 {
+				delta += " "
+			}
+			if !yield(fantasy.StreamPart{Type: fantasy.StreamPartTypeTextDelta, ID: "s-0", Delta: delta}) {
+				return
+			}
+		}
+		if !yield(fantasy.StreamPart{Type: fantasy.StreamPartTypeTextEnd, ID: "s-0"}) {
+			return
+		}
+		yield(fantasy.StreamPart{
+			Type:         fantasy.StreamPartTypeFinish,
+			FinishReason: fantasy.FinishReasonStop,
+			Usage:        fantasy.Usage{InputTokens: 10, OutputTokens: int64(len(words)), TotalTokens: 10 + int64(len(words))},
+		})
+	}, nil
+}
+
+func (m *streamingModel) GenerateObject(_ context.Context, _ fantasy.ObjectCall) (*fantasy.ObjectResponse, error) {
+	return nil, fmt.Errorf("not implemented")
+}
+
+func (m *streamingModel) StreamObject(_ context.Context, _ fantasy.ObjectCall) (fantasy.ObjectStreamResponse, error) {
+	return nil, fmt.Errorf("not implemented")
+}
+
+func (m *streamingModel) Provider() string { return "mock" }
+func (m *streamingModel) Model() string    { return "streaming-mock" }
+
+// TestIntegration_Streaming_TextDeltas tests that the streaming agent loop
+// publishes text deltas to the bus and returns the complete text.
+func TestIntegration_Streaming_TextDeltas(t *testing.T) {
+	tmpDir, err := os.MkdirTemp("", "agent-integration-stream-*")
+	if err != nil {
+		t.Fatalf("Failed to create temp dir: %v", err)
+	}
+	defer os.RemoveAll(tmpDir)
+
+	cfg := &config.Config{
+		Agents: config.AgentsConfig{
+			Defaults: config.AgentDefaults{
+				Workspace:         tmpDir,
+				Model:             "streaming-mock",
+				MaxTokens:         4096,
+				MaxToolIterations: 10,
+			},
+		},
+	}
+
+	msgBus := bus.NewMessageBus()
+	model := newStreamingModel("Hello from streaming agent response")
+	al := NewAgentLoop(cfg, msgBus, model)
+
+	ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
+	defer cancel()
+
+	// Collect stream deltas in background
+	var deltas []string
+	var deltaDone = make(chan struct{})
+	go func() {
+		defer close(deltaDone)
+		for {
+			msg, ok := msgBus.SubscribeOutbound(ctx)
+			if !ok {
+				return
+			}
+			if msg.StreamDelta {
+				deltas = append(deltas, msg.Content)
+			}
+		}
+	}()
+
+	// Process with streaming
+	response, err := al.ProcessDirectStreaming(ctx, "Stream me", "stream-session", "test", "chat-1")
+	if err != nil {
+		t.Fatalf("ProcessDirectStreaming failed: %v", err)
+	}
+
+	// Cancel to stop delta collector
+	cancel()
+	<-deltaDone
+
+	// Verify complete response
+	if response != "Hello from streaming agent response" {
+		t.Errorf("Expected 'Hello from streaming agent response', got: %s", response)
+	}
+
+	// Verify stream deltas were published
+	if len(deltas) == 0 {
+		t.Error("Expected stream deltas to be published to bus")
+	}
+
+	// Reconstruct full text from deltas
+	fullFromDeltas := strings.Join(deltas, "")
+	if fullFromDeltas != "Hello from streaming agent response" {
+		t.Errorf("Delta reconstruction mismatch: got '%s'", fullFromDeltas)
+	}
+
+	// Verify session was saved
+	history := al.sessions.GetHistory("stream-session")
+	if len(history) < 2 { // user + assistant
+		t.Errorf("Expected at least 2 messages in session, got %d", len(history))
+	}
+}
+
+// TestIntegration_Streaming_WithToolCalls tests streaming with a model
+// that requests tool calls before producing a final streamed response.
+func TestIntegration_Streaming_WithToolCalls(t *testing.T) {
+	tmpDir, err := os.MkdirTemp("", "agent-integration-stream-tools-*")
+	if err != nil {
+		t.Fatalf("Failed to create temp dir: %v", err)
+	}
+	defer os.RemoveAll(tmpDir)
+
+	cfg := &config.Config{
+		Agents: config.AgentsConfig{
+			Defaults: config.AgentDefaults{
+				Workspace:         tmpDir,
+				Model:             "mock-tool-model",
+				MaxTokens:         4096,
+				MaxToolIterations: 10,
+			},
+		},
+	}
+
+	msgBus := bus.NewMessageBus()
+	model := &toolCallingModel{}
+	al := NewAgentLoop(cfg, msgBus, model)
+	al.RegisterTool(&echoTool{})
+
+	ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
+	defer cancel()
+
+	// Collect deltas
+	var deltas []string
+	var deltaDone = make(chan struct{})
+	go func() {
+		defer close(deltaDone)
+		for {
+			msg, ok := msgBus.SubscribeOutbound(ctx)
+			if !ok {
+				return
+			}
+			if msg.StreamDelta {
+				deltas = append(deltas, msg.Content)
+			}
+		}
+	}()
+
+	response, err := al.ProcessDirectStreaming(ctx, "Use the echo tool (streaming)", "stream-tools-session", "test", "chat-1")
+	cancel()
+	<-deltaDone
+
+	if err != nil {
+		t.Fatalf("ProcessDirectStreaming with tools failed: %v", err)
+	}
+
+	if !strings.Contains(response, "Integration test response") {
+		t.Errorf("Expected response containing 'Integration test response', got: %s", response)
+	}
+
+	// Model should have been called at least twice (tool call step + final response step)
+	if model.callCount < 2 {
+		t.Errorf("Expected at least 2 model calls, got %d", model.callCount)
+	}
+}
+
+// TestIntegration_MultipleMessages tests sequential message processing
+// to verify session history accumulation.
+func TestIntegration_MultipleMessages(t *testing.T) {
+	tmpDir, err := os.MkdirTemp("", "agent-integration-multi-*")
+	if err != nil {
+		t.Fatalf("Failed to create temp dir: %v", err)
+	}
+	defer os.RemoveAll(tmpDir)
+
+	cfg := &config.Config{
+		Agents: config.AgentsConfig{
+			Defaults: config.AgentDefaults{
+				Workspace:         tmpDir,
+				Model:             "test-model",
+				MaxTokens:         4096,
+				MaxToolIterations: 10,
+			},
+		},
+	}
+
+	msgBus := bus.NewMessageBus()
+	model := newMockLanguageModel("Response")
+	al := NewAgentLoop(cfg, msgBus, model)
+
+	sessionKey := "multi-msg-session"
+	ctx := context.Background()
+
+	// Send 3 messages
+	for i := 0; i < 3; i++ {
+		msg := bus.InboundMessage{
+			Channel:    "test",
+			SenderID:   "user1",
+			ChatID:     "chat1",
+			Content:    fmt.Sprintf("Message %d", i+1),
+			SessionKey: sessionKey,
+		}
+
+		_, err := al.processMessage(ctx, msg)
+		if err != nil {
+			t.Fatalf("processMessage #%d failed: %v", i+1, err)
+		}
+	}
+
+	// Verify session history grew
+	history := al.sessions.GetHistory(sessionKey)
+	if len(history) < 6 { // At least 3 user messages + 3 assistant messages
+		t.Errorf("Expected at least 6 messages in history, got: %d", len(history))
+	}
+}

pkg/agent/loop.go

@@ -8,7 +8,6 @@ package agent
 
 import (
 	"context"
-	"encoding/json"
 	"fmt"
 	"os"
 	"path/filepath"
@@ -17,11 +16,15 @@ import (
 	"sync/atomic"
 	"time"
 
+	fantasy "charm.land/fantasy"
 	"github.com/sipeed/picoclaw/pkg/bus"
 	"github.com/sipeed/picoclaw/pkg/config"
 	"github.com/sipeed/picoclaw/pkg/constants"
+	picofantasy "github.com/sipeed/picoclaw/pkg/fantasy"
 	"github.com/sipeed/picoclaw/pkg/logger"
-	"github.com/sipeed/picoclaw/pkg/providers"
+	"github.com/sipeed/picoclaw/pkg/memory/delegate"
+	memstore "github.com/sipeed/picoclaw/pkg/memory/store"
+	"github.com/sipeed/picoclaw/pkg/messages"
 	"github.com/sipeed/picoclaw/pkg/session"
 	"github.com/sipeed/picoclaw/pkg/state"
 	"github.com/sipeed/picoclaw/pkg/tools"
@@ -30,7 +33,7 @@ import (
 
 type AgentLoop struct {
 	bus               *bus.MessageBus
-	provider       providers.LLMProvider
+	languageModel     fantasy.LanguageModel
 	workspace         string
 	model             string
 	contextWindow     int // Maximum context window size in tokens
@@ -39,8 +42,11 @@ type AgentLoop struct {
 	state             *state.Manager
 	contextBuilder    *ContextBuilder
 	tools             *tools.ToolRegistry
+	memoryStore       *memstore.MemoryStore // 3-tier MemGPT memory (nil if init failed)
 	running           atomic.Bool
 	summarizing       sync.Map       // Tracks which sessions are currently being summarized
+	summarizeFailures sync.Map       // Tracks consecutive summarization failures per session (string -> int)
+	cfg               *config.Config // Stored for subagent factory access
 }
 
 // processOptions configures how a message is processed
@@ -53,6 +59,7 @@ type processOptions struct {
 	EnableSummary   bool   // Whether to trigger summarization
 	SendResponse    bool   // Whether to send response via bus
 	NoHistory       bool   // If true, don't load session history (for heartbeat)
+	Streaming       bool   // If true, stream token deltas to bus via OnTextDelta
 }
 
 // createToolRegistry creates a tool registry with common tools.
@@ -101,7 +108,7 @@ func createToolRegistry(workspace string, restrict bool, cfg *config.Config, msg
 	return registry
 }
 
-func NewAgentLoop(cfg *config.Config, msgBus *bus.MessageBus, provider providers.LLMProvider) *AgentLoop {
+func NewAgentLoop(cfg *config.Config, msgBus *bus.MessageBus, model fantasy.LanguageModel) *AgentLoop {
 	workspace := cfg.WorkspacePath()
 	os.MkdirAll(workspace, 0755)
 
@@ -111,7 +118,7 @@ func NewAgentLoop(cfg *config.Config, msgBus *bus.MessageBus, provider providers
 	toolsRegistry := createToolRegistry(workspace, restrict, cfg, msgBus)
 
 	// Create subagent manager with its own tool registry
-	subagentManager := tools.NewSubagentManager(provider, cfg.Agents.Defaults.Model, workspace, msgBus)
+	subagentManager := tools.NewSubagentManager(model, cfg.Agents.Defaults.Model, workspace, msgBus)
 	subagentTools := createToolRegistry(workspace, restrict, cfg, msgBus)
 	// Subagent doesn't need spawn/subagent tools to avoid recursion
 	subagentManager.SetTools(subagentTools)
@@ -133,18 +140,66 @@ func NewAgentLoop(cfg *config.Config, msgBus *bus.MessageBus, provider providers
 	contextBuilder := NewContextBuilder(workspace)
 	contextBuilder.SetToolsRegistry(toolsRegistry)
 
+	// Initialize 3-tier MemGPT memory system
+	var ms *memstore.MemoryStore
+	memDBPath := filepath.Join(workspace, "memory", "picoclaw.db")
+	os.MkdirAll(filepath.Dir(memDBPath), 0755)
+
+	del, err := delegate.NewLibSQLDelegate(memDBPath)
+	if err != nil {
+		logger.WarnCF("agent", "Failed to create memory delegate, memory system disabled",
+			map[string]interface{}{"error": err.Error()})
+	} else {
+		if err := del.Init(context.Background()); err != nil {
+			logger.WarnCF("agent", "Failed to init memory schema, memory system disabled",
+				map[string]interface{}{"error": err.Error()})
+			del.Close()
+		} else {
+			chunker := memstore.NewMarkdownChunker(memstore.DefaultMarkdownChunkerConfig())
+			ms = memstore.New(del, chunker, nil, memstore.Config{
+				ContextWindowTokens:    cfg.Agents.Defaults.MaxTokens,
+				OffloadThresholdTokens: 4000,
+			})
+			contextBuilder.SetMemoryStore(ms)
+
+			// Register the memory tool
+			memTool := NewMemGPTTool(ms, "picoclaw", "default")
+			toolsRegistry.Register(memTool)
+
+			logger.InfoCF("agent", "3-tier memory system initialized",
+				map[string]interface{}{"db_path": memDBPath})
+		}
+	}
+
+	// Register meta-tools for progressive disclosure (tool_search + tool_call)
+	toolsRegistry.RegisterMetaTools()
+
+	// If memory tool is a gateway, mark it visible in progressive mode
+	if ms != nil {
+		toolsRegistry.MarkGateway("memory")
+	}
+
+	// Apply progressive disclosure config
+	if cfg.Tools.ProgressiveDisclosure {
+		toolsRegistry.SetProgressiveDisclosure(true)
+		logger.InfoCF("agent", "Progressive tool disclosure enabled",
+			map[string]interface{}{"gateway_tools": toolsRegistry.ListVisible()})
+	}
+
 	return &AgentLoop{
 		bus:            msgBus,
-		provider:       provider,
+		languageModel:  model,
 		workspace:      workspace,
 		model:          cfg.Agents.Defaults.Model,
-		contextWindow:  cfg.Agents.Defaults.MaxTokens, // Restore context window for summarization
+		contextWindow:  cfg.Agents.Defaults.MaxTokens,
 		maxIterations:  cfg.Agents.Defaults.MaxToolIterations,
 		sessions:       sessionsManager,
 		state:          stateManager,
 		contextBuilder: contextBuilder,
 		tools:          toolsRegistry,
+		memoryStore:    ms,
 		summarizing:    sync.Map{},
+		cfg:            cfg,
 	}
 }
 
@@ -192,6 +247,9 @@ func (al *AgentLoop) Run(ctx context.Context) error {
 
 func (al *AgentLoop) Stop() {
 	al.running.Store(false)
+	if al.memoryStore != nil {
+		al.memoryStore.Close()
+	}
 }
 
 func (al *AgentLoop) RegisterTool(tool tools.Tool) {
@@ -226,6 +284,29 @@ func (al *AgentLoop) ProcessDirectWithChannel(ctx context.Context, content, sess
 	return al.processMessage(ctx, msg)
 }
 
+// ProcessDirectStreaming processes a message with streaming token delivery.
+// Text deltas are published to the bus as StreamDelta messages in real time.
+func (al *AgentLoop) ProcessDirectStreaming(ctx context.Context, content, sessionKey, channel, chatID string) (string, error) {
+	msg := bus.InboundMessage{
+		Channel:    channel,
+		SenderID:   "user",
+		ChatID:     chatID,
+		Content:    content,
+		SessionKey: sessionKey,
+	}
+
+	return al.runAgentLoop(ctx, processOptions{
+		SessionKey:      msg.SessionKey,
+		Channel:         msg.Channel,
+		ChatID:          msg.ChatID,
+		UserMessage:     msg.Content,
+		DefaultResponse: "I've completed processing but have no response to give.",
+		EnableSummary:   true,
+		SendResponse:    false,
+		Streaming:       true,
+	})
+}
+
 // ProcessHeartbeat processes a heartbeat request without session history.
 // Each heartbeat is independent and doesn't accumulate context.
 func (al *AgentLoop) ProcessHeartbeat(ctx context.Context, content, channel, chatID string) (string, error) {
@@ -327,11 +408,14 @@ func (al *AgentLoop) processSystemMessage(ctx context.Context, msg bus.InboundMe
 }
 
 // runAgentLoop is the core message processing logic.
-// It handles context building, LLM calls, tool execution, and response handling.
+// It handles context building, Fantasy agent creation, tool execution, and response handling.
+// When opts.Streaming is true, delegates to runAgentLoopStreaming for real-time token delivery.
 func (al *AgentLoop) runAgentLoop(ctx context.Context, opts processOptions) (string, error) {
+	if opts.Streaming {
+		return al.runAgentLoopStreaming(ctx, opts)
+	}
 	// 0. Record last channel for heartbeat notifications (skip internal channels)
 	if opts.Channel != "" && opts.ChatID != "" {
-		// Don't record internal channels (cli, system, subagent)
 		if !constants.IsInternalChannel(opts.Channel) {
 			channelKey := fmt.Sprintf("%s:%s", opts.Channel, opts.ChatID)
 			if err := al.RecordLastChannel(channelKey); err != nil {
@@ -344,13 +428,13 @@ func (al *AgentLoop) runAgentLoop(ctx context.Context, opts processOptions) (str
 	al.updateToolContexts(opts.Channel, opts.ChatID)
 
 	// 2. Build messages (skip history for heartbeat)
-	var history []providers.Message
+	var history []messages.Message
 	var summary string
 	if !opts.NoHistory {
 		history = al.sessions.GetHistory(opts.SessionKey)
 		summary = al.sessions.GetSummary(opts.SessionKey)
 	}
-	messages := al.contextBuilder.BuildMessages(
+	builtMsgs := al.contextBuilder.BuildMessages(
 		history,
 		summary,
 		opts.UserMessage,
@@ -362,30 +446,89 @@ func (al *AgentLoop) runAgentLoop(ctx context.Context, opts processOptions) (str
 	// 3. Save user message to session
 	al.sessions.AddMessage(opts.SessionKey, "user", opts.UserMessage)
 
-	// 4. Run LLM iteration loop
+	// 4. Split built messages into system prompt, conversation history, and current user prompt.
-	finalContent, iteration, err := al.runLLMIteration(ctx, messages, opts)
+	// BuildMessages returns: [system, ...history, user]
-	if err != nil {
+	systemPrompt := ""
-		return "", err
+	var historyMsgs []messages.Message
+	userPrompt := opts.UserMessage
+
+	if len(builtMsgs) > 0 && builtMsgs[0].Role == "system" {
+		systemPrompt = builtMsgs[0].Content
+		// History is everything between system and last user message.
+		if len(builtMsgs) > 2 {
+			historyMsgs = builtMsgs[1 : len(builtMsgs)-1]
+		}
 	}
 
-	// If last tool had ForUser content and we already sent it, we might not need to send final response
+	// 5. Convert history to Fantasy message format
-	// This is controlled by the tool's Silent flag and ForUser content
+	fantasyHistory := picofantasy.MessagesToFantasy(historyMsgs)
 
-	// 5. Handle empty response
+	// 6. Build adapted tools from PicoClaw registry (with optional offloading)
+	adaptCfg := picofantasy.AdaptedToolsConfig{
+		MemStore:   al.memoryStore,
+		AgentID:    "picoclaw",
+		SessionKey: opts.SessionKey,
+	}
+	adaptedTools := picofantasy.BuildAdaptedTools(al.tools, al.bus, opts.Channel, opts.ChatID, adaptCfg)
+
+	// 7. Create Fantasy agent with tools and configuration
+	agentOpts := []fantasy.AgentOption{
+		fantasy.WithTools(adaptedTools...),
+		fantasy.WithStopConditions(fantasy.StepCountIs(al.maxIterations)),
+	}
+	if systemPrompt != "" {
+		agentOpts = append(agentOpts, fantasy.WithSystemPrompt(systemPrompt))
+	}
+	agent := fantasy.NewAgent(al.languageModel, agentOpts...)
+
+	logger.DebugCF("agent", "Fantasy agent created",
+		map[string]interface{}{
+			"model":          al.model,
+			"tools_count":    len(adaptedTools),
+			"history_count":  len(historyMsgs),
+			"max_iterations": al.maxIterations,
+			"memory_enabled": al.memoryStore != nil,
+		})
+
+	// 8. Call Fantasy agent.Generate()
+	result, err := agent.Generate(ctx, fantasy.AgentCall{
+		Prompt:   userPrompt,
+		Messages: fantasyHistory,
+	})
+	if err != nil {
+		logger.ErrorCF("agent", "Fantasy Generate failed",
+			map[string]interface{}{
+				"error": err.Error(),
+			})
+		return "", fmt.Errorf("agent Generate failed: %w", err)
+	}
+
+	// 9. Save all step messages to session
+	stepCount := len(result.Steps)
+	for _, step := range result.Steps {
+		stepMsgs := picofantasy.StepToMessages(step)
+		for _, m := range stepMsgs {
+			al.sessions.AddFullMessage(opts.SessionKey, m)
+		}
+	}
+
+	// 10. Extract final text
+	finalContent := result.Response.Content.Text()
+
+	// 11. Handle empty response
 	if finalContent == "" {
 		finalContent = opts.DefaultResponse
 	}
 
-	// 6. Save final assistant message to session
+	// 12. Save session
-	al.sessions.AddMessage(opts.SessionKey, "assistant", finalContent)
 	al.sessions.Save(opts.SessionKey)
 
-	// 7. Optional: summarization
+	// 13. Optional: summarization
 	if opts.EnableSummary {
 		al.maybeSummarize(opts.SessionKey)
 	}
 
-	// 8. Optional: send response via bus
+	// 14. Optional: send response via bus
 	if opts.SendResponse {
 		al.bus.PublishOutbound(bus.OutboundMessage{
 			Channel: opts.Channel,
@@ -394,179 +537,165 @@ func (al *AgentLoop) runAgentLoop(ctx context.Context, opts processOptions) (str
 		})
 	}
 
-	// 9. Log response
+	// 15. Log response
 	responsePreview := utils.Truncate(finalContent, 120)
 	logger.InfoCF("agent", fmt.Sprintf("Response: %s", responsePreview),
 		map[string]interface{}{
 			"session_key":  opts.SessionKey,
-			"iterations":   iteration,
+			"steps":        stepCount,
 			"final_length": len(finalContent),
 		})
 
 	return finalContent, nil
 }
 
-// runLLMIteration executes the LLM call loop with tool handling.
+// runAgentLoopStreaming uses Fantasy's agent.Stream() to stream token deltas
-// Returns the final content, iteration count, and any error.
+// to the bus in real time. Structure mirrors runAgentLoop but uses AgentStreamCall
-func (al *AgentLoop) runLLMIteration(ctx context.Context, messages []providers.Message, opts processOptions) (string, int, error) {
+// with OnTextDelta, OnStepFinish, and OnToolCall callbacks.
-	iteration := 0
+func (al *AgentLoop) runAgentLoopStreaming(ctx context.Context, opts processOptions) (string, error) {
-	var finalContent string
+	// 0. Record last channel
+	if opts.Channel != "" && opts.ChatID != "" {
+		if !constants.IsInternalChannel(opts.Channel) {
+			channelKey := fmt.Sprintf("%s:%s", opts.Channel, opts.ChatID)
+			if err := al.RecordLastChannel(channelKey); err != nil {
+				logger.WarnCF("agent", "Failed to record last channel: %v", map[string]interface{}{"error": err.Error()})
+			}
+		}
+	}
 
-	for iteration < al.maxIterations {
+	// 1. Update tool contexts
-		iteration++
+	al.updateToolContexts(opts.Channel, opts.ChatID)
 
-		logger.DebugCF("agent", "LLM iteration",
+	// 2. Build messages
+	var history []messages.Message
+	var summary string
+	if !opts.NoHistory {
+		history = al.sessions.GetHistory(opts.SessionKey)
+		summary = al.sessions.GetSummary(opts.SessionKey)
+	}
+	builtMsgs := al.contextBuilder.BuildMessages(history, summary, opts.UserMessage, nil, opts.Channel, opts.ChatID)
+
+	// 3. Save user message
+	al.sessions.AddMessage(opts.SessionKey, "user", opts.UserMessage)
+
+	// 4. Split into system/history/user
+	systemPrompt := ""
+	var historyMsgs []messages.Message
+	userPrompt := opts.UserMessage
+
+	if len(builtMsgs) > 0 && builtMsgs[0].Role == "system" {
+		systemPrompt = builtMsgs[0].Content
+		if len(builtMsgs) > 2 {
+			historyMsgs = builtMsgs[1 : len(builtMsgs)-1]
+		}
+	}
+
+	// 5. Convert history
+	fantasyHistory := picofantasy.MessagesToFantasy(historyMsgs)
+
+	// 6. Build adapted tools (with optional offloading)
+	streamAdaptCfg := picofantasy.AdaptedToolsConfig{
+		MemStore:   al.memoryStore,
+		AgentID:    "picoclaw",
+		SessionKey: opts.SessionKey,
+	}
+	adaptedTools := picofantasy.BuildAdaptedTools(al.tools, al.bus, opts.Channel, opts.ChatID, streamAdaptCfg)
+
+	// 7. Create Fantasy agent
+	agentOpts := []fantasy.AgentOption{
+		fantasy.WithTools(adaptedTools...),
+		fantasy.WithStopConditions(fantasy.StepCountIs(al.maxIterations)),
+	}
+	if systemPrompt != "" {
+		agentOpts = append(agentOpts, fantasy.WithSystemPrompt(systemPrompt))
+	}
+	fantasyAgent := fantasy.NewAgent(al.languageModel, agentOpts...)
+
+	logger.DebugCF("agent", "Fantasy streaming agent created",
 		map[string]interface{}{
-				"iteration": iteration,
-				"max":       al.maxIterations,
-			})
-
-		// Build tool definitions
-		providerToolDefs := al.tools.ToProviderDefs()
-
-		// Log LLM request details
-		logger.DebugCF("agent", "LLM request",
-			map[string]interface{}{
-				"iteration":         iteration,
 			"model":          al.model,
-				"messages_count":    len(messages),
+			"tools_count":    len(adaptedTools),
-				"tools_count":       len(providerToolDefs),
+			"history_count":  len(historyMsgs),
-				"max_tokens":        8192,
+			"max_iterations": al.maxIterations,
-				"temperature":       0.7,
+			"memory_enabled": al.memoryStore != nil,
-				"system_prompt_len": len(messages[0].Content),
 		})
 
-		// Log full messages (detailed)
+	// 8. Build streaming call with callbacks
-		logger.DebugCF("agent", "Full LLM request",
+	streamCall := fantasy.AgentStreamCall{
-			map[string]interface{}{
+		Prompt:   userPrompt,
-				"iteration":     iteration,
+		Messages: fantasyHistory,
-				"messages_json": formatMessagesForLog(messages),
-				"tools_json":    formatToolsForLog(providerToolDefs),
-			})
 
-		// Call LLM
+		// Stream text deltas to bus in real time
-		response, err := al.provider.Chat(ctx, messages, providerToolDefs, al.model, map[string]interface{}{
+		OnTextDelta: func(id, text string) error {
-			"max_tokens":  8192,
+			if opts.Channel != "" && opts.ChatID != "" {
-			"temperature": 0.7,
-		})
-
-		if err != nil {
-			logger.ErrorCF("agent", "LLM call failed",
-				map[string]interface{}{
-					"iteration": iteration,
-					"error":     err.Error(),
-				})
-			return "", iteration, fmt.Errorf("LLM call failed: %w", err)
-		}
-
-		// Check if no tool calls - we're done
-		if len(response.ToolCalls) == 0 {
-			finalContent = response.Content
-			logger.InfoCF("agent", "LLM response without tool calls (direct answer)",
-				map[string]interface{}{
-					"iteration":     iteration,
-					"content_chars": len(finalContent),
-				})
-			break
-		}
-
-		// Log tool calls
-		toolNames := make([]string, 0, len(response.ToolCalls))
-		for _, tc := range response.ToolCalls {
-			toolNames = append(toolNames, tc.Name)
-		}
-		logger.InfoCF("agent", "LLM requested tool calls",
-			map[string]interface{}{
-				"tools":     toolNames,
-				"count":     len(response.ToolCalls),
-				"iteration": iteration,
-			})
-
-		// Build assistant message with tool calls
-		assistantMsg := providers.Message{
-			Role:    "assistant",
-			Content: response.Content,
-		}
-		for _, tc := range response.ToolCalls {
-			argumentsJSON, _ := json.Marshal(tc.Arguments)
-			assistantMsg.ToolCalls = append(assistantMsg.ToolCalls, providers.ToolCall{
-				ID:   tc.ID,
-				Type: "function",
-				Function: &providers.FunctionCall{
-					Name:      tc.Name,
-					Arguments: string(argumentsJSON),
-				},
-			})
-		}
-		messages = append(messages, assistantMsg)
-
-		// Save assistant message with tool calls to session
-		al.sessions.AddFullMessage(opts.SessionKey, assistantMsg)
-
-		// Execute tool calls
-		for _, tc := range response.ToolCalls {
-			// Log tool call with arguments preview
-			argsJSON, _ := json.Marshal(tc.Arguments)
-			argsPreview := utils.Truncate(string(argsJSON), 200)
-			logger.InfoCF("agent", fmt.Sprintf("Tool call: %s(%s)", tc.Name, argsPreview),
-				map[string]interface{}{
-					"tool":      tc.Name,
-					"iteration": iteration,
-				})
-
-			// Create async callback for tools that implement AsyncTool
-			// NOTE: Following openclaw's design, async tools do NOT send results directly to users.
-			// Instead, they notify the agent via PublishInbound, and the agent decides
-			// whether to forward the result to the user (in processSystemMessage).
-			asyncCallback := func(callbackCtx context.Context, result *tools.ToolResult) {
-				// Log the async completion but don't send directly to user
-				// The agent will handle user notification via processSystemMessage
-				if !result.Silent && result.ForUser != "" {
-					logger.InfoCF("agent", "Async tool completed, agent will handle notification",
-						map[string]interface{}{
-							"tool":        tc.Name,
-							"content_len": len(result.ForUser),
-						})
-				}
-			}
-
-			toolResult := al.tools.ExecuteWithContext(ctx, tc.Name, tc.Arguments, opts.Channel, opts.ChatID, asyncCallback)
-
-			// Send ForUser content to user immediately if not Silent
-			if !toolResult.Silent && toolResult.ForUser != "" && opts.SendResponse {
 				al.bus.PublishOutbound(bus.OutboundMessage{
 					Channel:     opts.Channel,
 					ChatID:      opts.ChatID,
-					Content: toolResult.ForUser,
+					Content:     text,
+					StreamDelta: true,
 				})
-				logger.DebugCF("agent", "Sent tool result to user",
+			}
+			return nil
+		},
+
+		// Save each step's messages to session as they complete
+		OnStepFinish: func(step fantasy.StepResult) error {
+			stepMsgs := picofantasy.StepToMessages(step)
+			for _, m := range stepMsgs {
+				al.sessions.AddFullMessage(opts.SessionKey, m)
+			}
+			return nil
+		},
+
+		// Log tool calls as they happen
+		OnToolCall: func(tc fantasy.ToolCallContent) error {
+			logger.DebugCF("agent", "Streaming tool call",
 				map[string]interface{}{
-						"tool":        tc.Name,
+					"tool": tc.ToolName,
-						"content_len": len(toolResult.ForUser),
+					"id":   tc.ToolCallID,
 				})
+			return nil
+		},
 	}
 
-			// Determine content for LLM based on tool result
+	// 9. Call Fantasy agent.Stream()
-			contentForLLM := toolResult.ForLLM
+	result, err := fantasyAgent.Stream(ctx, streamCall)
-			if contentForLLM == "" && toolResult.Err != nil {
+	if err != nil {
-				contentForLLM = toolResult.Err.Error()
+		logger.ErrorCF("agent", "Fantasy Stream failed",
+			map[string]interface{}{"error": err.Error()})
+		return "", fmt.Errorf("agent Stream failed: %w", err)
 	}
 
-			toolResultMsg := providers.Message{
+	// 10. Extract final text
-				Role:       "tool",
+	finalContent := result.Response.Content.Text()
-				Content:    contentForLLM,
+	if finalContent == "" {
-				ToolCallID: tc.ID,
+		finalContent = opts.DefaultResponse
-			}
-			messages = append(messages, toolResultMsg)
-
-			// Save tool result message to session
-			al.sessions.AddFullMessage(opts.SessionKey, toolResultMsg)
-		}
 	}
 
-	return finalContent, iteration, nil
+	// 11. Save session
+	al.sessions.Save(opts.SessionKey)
+
+	// 12. Summarization
+	if opts.EnableSummary {
+		al.maybeSummarize(opts.SessionKey)
 	}
 
+	// 13. Log response
+	stepCount := len(result.Steps)
+	responsePreview := utils.Truncate(finalContent, 120)
+	logger.InfoCF("agent", fmt.Sprintf("Streaming response: %s", responsePreview),
+		map[string]interface{}{
+			"session_key":  opts.SessionKey,
+			"steps":        stepCount,
+			"final_length": len(finalContent),
+			"total_tokens": result.TotalUsage.TotalTokens,
+		})
+
+	return finalContent, nil
+}
+
+// runLLMIteration — DELETED. Replaced by Fantasy's internal agent loop.
+
 // updateToolContexts updates the context for tools that need channel/chatID info.
 func (al *AgentLoop) updateToolContexts(channel, chatID string) {
 	// Use ContextualTool interface instead of type assertions
@@ -621,16 +750,16 @@ func (al *AgentLoop) GetStartupInfo() map[string]interface{} {
 }
 
 // formatMessagesForLog formats messages for logging
-func formatMessagesForLog(messages []providers.Message) string {
+func formatMessagesForLog(msgs []messages.Message) string {
-	if len(messages) == 0 {
+	if len(msgs) == 0 {
 		return "[]"
 	}
 
 	var result string
 	result += "[\n"
-	for i, msg := range messages {
+	for i, msg := range msgs {
 		result += fmt.Sprintf("  [%d] Role: %s\n", i, msg.Role)
-		if msg.ToolCalls != nil && len(msg.ToolCalls) > 0 {
+		if len(msg.ToolCalls) > 0 {
 			result += "  ToolCalls:\n"
 			for _, tc := range msg.ToolCalls {
 				result += fmt.Sprintf("    - ID: %s, Type: %s, Name: %s\n", tc.ID, tc.Type, tc.Name)
@@ -652,25 +781,6 @@ func formatMessagesForLog(messages []providers.Message) string {
 	return result
 }
 
-// formatToolsForLog formats tool definitions for logging
-func formatToolsForLog(tools []providers.ToolDefinition) string {
-	if len(tools) == 0 {
-		return "[]"
-	}
-
-	var result string
-	result += "[\n"
-	for i, tool := range tools {
-		result += fmt.Sprintf("  [%d] Type: %s, Name: %s\n", i, tool.Type, tool.Function.Name)
-		result += fmt.Sprintf("      Description: %s\n", tool.Function.Description)
-		if len(tool.Function.Parameters) > 0 {
-			result += fmt.Sprintf("      Parameters: %s\n", utils.Truncate(fmt.Sprintf("%v", tool.Function.Parameters), 200))
-		}
-	}
-	result += "]"
-	return result
-}
-
 // summarizeSession summarizes the conversation history for a session.
 func (al *AgentLoop) summarizeSession(sessionKey string) {
 	ctx, cancel := context.WithTimeout(context.Background(), 120*time.Second)
@@ -689,14 +799,13 @@ func (al *AgentLoop) summarizeSession(sessionKey string) {
 	// Oversized Message Guard
 	// Skip messages larger than 50% of context window to prevent summarizer overflow
 	maxMessageTokens := al.contextWindow / 2
-	validMessages := make([]providers.Message, 0)
+	validMessages := make([]messages.Message, 0)
 	omitted := false
 
 	for _, m := range toSummarize {
 		if m.Role != "user" && m.Role != "assistant" {
 			continue
 		}
-		// Estimate tokens for this message
 		msgTokens := len(m.Content) / 4
 		if msgTokens > maxMessageTokens {
 			omitted = true
@@ -710,7 +819,6 @@ func (al *AgentLoop) summarizeSession(sessionKey string) {
 	}
 
 	// Multi-Part Summarization
-	// Split into two parts if history is significant
 	var finalSummary string
 	if len(validMessages) > 10 {
 		mid := len(validMessages) / 2
@@ -722,12 +830,9 @@ func (al *AgentLoop) summarizeSession(sessionKey string) {
 
 		// Merge them
 		mergePrompt := fmt.Sprintf("Merge these two conversation summaries into one cohesive summary:\n\n1: %s\n\n2: %s", s1, s2)
-		resp, err := al.provider.Chat(ctx, []providers.Message{{Role: "user", Content: mergePrompt}}, nil, al.model, map[string]interface{}{
+		resp, err := al.callModel(ctx, mergePrompt)
-			"max_tokens":  1024,
-			"temperature": 0.3,
-		})
 		if err == nil {
-			finalSummary = resp.Content
+			finalSummary = resp
 		} else {
 			finalSummary = s1 + " " + s2
 		}
@@ -743,11 +848,33 @@ func (al *AgentLoop) summarizeSession(sessionKey string) {
 		al.sessions.SetSummary(sessionKey, finalSummary)
 		al.sessions.TruncateHistory(sessionKey, 4)
 		al.sessions.Save(sessionKey)
+		al.summarizeFailures.Delete(sessionKey)
+	} else {
+		var count int
+		if v, ok := al.summarizeFailures.Load(sessionKey); ok {
+			count = v.(int)
+		}
+		count++
+		al.summarizeFailures.Store(sessionKey, count)
+
+		const maxSummarizeFailures = 3
+		const emergencyKeep = 10
+		if count >= maxSummarizeFailures {
+			logger.ErrorCF("agent", "Summarization failed repeatedly, force-truncating session",
+				map[string]interface{}{
+					"session":              sessionKey,
+					"consecutive_failures": count,
+					"keep":                 emergencyKeep,
+				})
+			al.sessions.TruncateHistory(sessionKey, emergencyKeep)
+			al.sessions.Save(sessionKey)
+			al.summarizeFailures.Delete(sessionKey)
+		}
 	}
 }
 
-// summarizeBatch summarizes a batch of messages.
+// summarizeBatch summarizes a batch of messages using the Fantasy LanguageModel directly.
-func (al *AgentLoop) summarizeBatch(ctx context.Context, batch []providers.Message, existingSummary string) (string, error) {
+func (al *AgentLoop) summarizeBatch(ctx context.Context, batch []messages.Message, existingSummary string) (string, error) {
 	prompt := "Provide a concise summary of this conversation segment, preserving core context and key points.\n"
 	if existingSummary != "" {
 		prompt += "Existing context: " + existingSummary + "\n"
@@ -757,20 +884,32 @@ func (al *AgentLoop) summarizeBatch(ctx context.Context, batch []providers.Messa
 		prompt += fmt.Sprintf("%s: %s\n", m.Role, m.Content)
 	}
 
-	response, err := al.provider.Chat(ctx, []providers.Message{{Role: "user", Content: prompt}}, nil, al.model, map[string]interface{}{
+	return al.callModel(ctx, prompt)
-		"max_tokens":  1024,
+}
-		"temperature": 0.3,
+
+// callModel makes a direct call to the Fantasy LanguageModel (no tools, no agent loop).
+// Used for summarization and other simple generation tasks.
+func (al *AgentLoop) callModel(ctx context.Context, prompt string) (string, error) {
+	temp := 0.3
+	maxTokens := int64(1024)
+
+	resp, err := al.languageModel.Generate(ctx, fantasy.Call{
+		Prompt: fantasy.Prompt{
+			fantasy.NewUserMessage(prompt),
+		},
+		Temperature:     &temp,
+		MaxOutputTokens: &maxTokens,
 	})
 	if err != nil {
 		return "", err
 	}
-	return response.Content, nil
+	return resp.Content.Text(), nil
 }
 
 // estimateTokens estimates the number of tokens in a message list.
-func (al *AgentLoop) estimateTokens(messages []providers.Message) int {
+func (al *AgentLoop) estimateTokens(msgs []messages.Message) int {
 	total := 0
-	for _, m := range messages {
+	for _, m := range msgs {
 		total += len(m.Content) / 4 // Simple heuristic: 4 chars per token
 	}
 	return total

pkg/agent/loop_test.go

@@ -2,31 +2,55 @@ package agent
 
 import (
 	"context"
+	"fmt"
 	"os"
 	"path/filepath"
 	"testing"
 	"time"
 
+	fantasy "charm.land/fantasy"
 	"github.com/sipeed/picoclaw/pkg/bus"
 	"github.com/sipeed/picoclaw/pkg/config"
-	"github.com/sipeed/picoclaw/pkg/providers"
 	"github.com/sipeed/picoclaw/pkg/tools"
 )
 
-// mockProvider is a simple mock LLM provider for testing
+// mockLanguageModel is a simple mock fantasy.LanguageModel for testing
-type mockProvider struct{}
+type mockLanguageModel struct {
+	response string
+}
 
-func (m *mockProvider) Chat(ctx context.Context, messages []providers.Message, tools []providers.ToolDefinition, model string, opts map[string]interface{}) (*providers.LLMResponse, error) {
+func newMockLanguageModel(response string) *mockLanguageModel {
-	return &providers.LLMResponse{
+	if response == "" {
-		Content:   "Mock response",
+		response = "Mock response"
-		ToolCalls: []providers.ToolCall{},
+	}
+	return &mockLanguageModel{response: response}
+}
+
+func (m *mockLanguageModel) Generate(_ context.Context, call fantasy.Call) (*fantasy.Response, error) {
+	return &fantasy.Response{
+		Content:      fantasy.ResponseContent{fantasy.TextContent{Text: m.response}},
+		FinishReason: fantasy.FinishReasonStop,
 	}, nil
 }
 
-func (m *mockProvider) GetDefaultModel() string {
+func (m *mockLanguageModel) Stream(_ context.Context, call fantasy.Call) (fantasy.StreamResponse, error) {
-	return "mock-model"
+	return func(yield func(fantasy.StreamPart) bool) {
+		yield(fantasy.StreamPart{Type: fantasy.StreamPartTypeTextDelta, Delta: m.response})
+		yield(fantasy.StreamPart{Type: fantasy.StreamPartTypeFinish, FinishReason: fantasy.FinishReasonStop})
+	}, nil
 }
 
+func (m *mockLanguageModel) GenerateObject(_ context.Context, _ fantasy.ObjectCall) (*fantasy.ObjectResponse, error) {
+	return nil, fmt.Errorf("not implemented")
+}
+
+func (m *mockLanguageModel) StreamObject(_ context.Context, _ fantasy.ObjectCall) (fantasy.ObjectStreamResponse, error) {
+	return nil, fmt.Errorf("not implemented")
+}
+
+func (m *mockLanguageModel) Provider() string { return "mock" }
+func (m *mockLanguageModel) Model() string    { return "mock-model" }
+
 func TestRecordLastChannel(t *testing.T) {
 	// Create temp workspace
 	tmpDir, err := os.MkdirTemp("", "agent-test-*")
@@ -49,8 +73,8 @@ func TestRecordLastChannel(t *testing.T) {
 
 	// Create agent loop
 	msgBus := bus.NewMessageBus()
-	provider := &mockProvider{}
+	model := newMockLanguageModel("")
-	al := NewAgentLoop(cfg, msgBus, provider)
+	al := NewAgentLoop(cfg, msgBus, model)
 
 	// Test RecordLastChannel
 	testChannel := "test-channel"
@@ -66,7 +90,7 @@ func TestRecordLastChannel(t *testing.T) {
 	}
 
 	// Verify persistence by creating a new agent loop
-	al2 := NewAgentLoop(cfg, msgBus, provider)
+	al2 := NewAgentLoop(cfg, msgBus, model)
 	if al2.state.GetLastChannel() != testChannel {
 		t.Errorf("Expected persistent channel '%s', got '%s'", testChannel, al2.state.GetLastChannel())
 	}
@@ -94,8 +118,8 @@ func TestRecordLastChatID(t *testing.T) {
 
 	// Create agent loop
 	msgBus := bus.NewMessageBus()
-	provider := &mockProvider{}
+	model := newMockLanguageModel("")
-	al := NewAgentLoop(cfg, msgBus, provider)
+	al := NewAgentLoop(cfg, msgBus, model)
 
 	// Test RecordLastChatID
 	testChatID := "test-chat-id-123"
@@ -111,7 +135,7 @@ func TestRecordLastChatID(t *testing.T) {
 	}
 
 	// Verify persistence by creating a new agent loop
-	al2 := NewAgentLoop(cfg, msgBus, provider)
+	al2 := NewAgentLoop(cfg, msgBus, model)
 	if al2.state.GetLastChatID() != testChatID {
 		t.Errorf("Expected persistent chat ID '%s', got '%s'", testChatID, al2.state.GetLastChatID())
 	}
@@ -139,8 +163,8 @@ func TestNewAgentLoop_StateInitialized(t *testing.T) {
 
 	// Create agent loop
 	msgBus := bus.NewMessageBus()
-	provider := &mockProvider{}
+	model := newMockLanguageModel("")
-	al := NewAgentLoop(cfg, msgBus, provider)
+	al := NewAgentLoop(cfg, msgBus, model)
 
 	// Verify state manager is initialized
 	if al.state == nil {
@@ -174,8 +198,8 @@ func TestToolRegistry_ToolRegistration(t *testing.T) {
 	}
 
 	msgBus := bus.NewMessageBus()
-	provider := &mockProvider{}
+	model := newMockLanguageModel("")
-	al := NewAgentLoop(cfg, msgBus, provider)
+	al := NewAgentLoop(cfg, msgBus, model)
 
 	// Register a custom tool
 	customTool := &mockCustomTool{}
@@ -220,8 +244,8 @@ func TestToolContext_Updates(t *testing.T) {
 	}
 
 	msgBus := bus.NewMessageBus()
-	provider := &simpleMockProvider{response: "OK"}
+	model := newMockLanguageModel("OK")
-	_ = NewAgentLoop(cfg, msgBus, provider)
+	_ = NewAgentLoop(cfg, msgBus, model)
 
 	// Verify that ContextualTool interface is defined and can be implemented
 	// This test validates the interface contract exists
@@ -251,8 +275,8 @@ func TestToolRegistry_GetDefinitions(t *testing.T) {
 	}
 
 	msgBus := bus.NewMessageBus()
-	provider := &mockProvider{}
+	model := newMockLanguageModel("")
-	al := NewAgentLoop(cfg, msgBus, provider)
+	al := NewAgentLoop(cfg, msgBus, model)
 
 	// Register a test tool and verify it shows up in startup info
 	testTool := &mockCustomTool{}
@@ -295,8 +319,8 @@ func TestAgentLoop_GetStartupInfo(t *testing.T) {
 	}
 
 	msgBus := bus.NewMessageBus()
-	provider := &mockProvider{}
+	model := newMockLanguageModel("")
-	al := NewAgentLoop(cfg, msgBus, provider)
+	al := NewAgentLoop(cfg, msgBus, model)
 
 	info := al.GetStartupInfo()
 
@@ -342,8 +366,8 @@ func TestAgentLoop_Stop(t *testing.T) {
 	}
 
 	msgBus := bus.NewMessageBus()
-	provider := &mockProvider{}
+	model := newMockLanguageModel("")
-	al := NewAgentLoop(cfg, msgBus, provider)
+	al := NewAgentLoop(cfg, msgBus, model)
 
 	// Note: running is only set to true when Run() is called
 	// We can't test that without starting the event loop
@@ -358,21 +382,6 @@ func TestAgentLoop_Stop(t *testing.T) {
 
 // Mock implementations for testing
 
-type simpleMockProvider struct {
-	response string
-}
-
-func (m *simpleMockProvider) Chat(ctx context.Context, messages []providers.Message, tools []providers.ToolDefinition, model string, opts map[string]interface{}) (*providers.LLMResponse, error) {
-	return &providers.LLMResponse{
-		Content:   m.response,
-		ToolCalls: []providers.ToolCall{},
-	}, nil
-}
-
-func (m *simpleMockProvider) GetDefaultModel() string {
-	return "mock-model"
-}
-
 // mockCustomTool is a simple mock tool for registration testing
 type mockCustomTool struct{}
 
@@ -464,8 +473,8 @@ func TestToolResult_SilentToolDoesNotSendUserMessage(t *testing.T) {
 	}
 
 	msgBus := bus.NewMessageBus()
-	provider := &simpleMockProvider{response: "File operation complete"}
+	model := newMockLanguageModel("File operation complete")
-	al := NewAgentLoop(cfg, msgBus, provider)
+	al := NewAgentLoop(cfg, msgBus, model)
 	helper := testHelper{al: al}
 
 	// ReadFileTool returns SilentResult, which should not send user message
@@ -506,8 +515,8 @@ func TestToolResult_UserFacingToolDoesSendMessage(t *testing.T) {
 	}
 
 	msgBus := bus.NewMessageBus()
-	provider := &simpleMockProvider{response: "Command output: hello world"}
+	model := newMockLanguageModel("Command output: hello world")
-	al := NewAgentLoop(cfg, msgBus, provider)
+	al := NewAgentLoop(cfg, msgBus, model)
 	helper := testHelper{al: al}
 
 	// ExecTool returns UserResult, which should send user message

pkg/agent/memgpt_tool.go

@@ -0,0 +1,106 @@
+// PicoClaw - Ultra-lightweight personal AI agent
+// License: MIT
+//
+// Copyright (c) 2026 PicoClaw contributors
+
+package agent
+
+import (
+	"context"
+	"encoding/json"
+
+	memstore "github.com/sipeed/picoclaw/pkg/memory/store"
+	"github.com/sipeed/picoclaw/pkg/tools"
+)
+
+// MemGPTTool wraps store.MemoryTool as a PicoClaw tools.Tool so it can be
+// registered in the ToolRegistry and executed by the Fantasy agent loop.
+type MemGPTTool struct {
+	inner *memstore.MemoryTool
+}
+
+var _ tools.Tool = (*MemGPTTool)(nil)
+
+// NewMemGPTTool creates a PicoClaw tool wrapper around a MemoryTool.
+func NewMemGPTTool(store *memstore.MemoryStore, agentID, session string) *MemGPTTool {
+	return &MemGPTTool{
+		inner: memstore.NewMemoryTool(store, agentID, session),
+	}
+}
+
+func (t *MemGPTTool) Name() string { return "memory" }
+
+func (t *MemGPTTool) Description() string {
+	return "Manage the agent's 3-tier memory system. Actions: search (hybrid keyword+vector), read (by ID), write (to recall or archival), update, delete, status (context pressure). All memory persists across sessions."
+}
+
+func (t *MemGPTTool) Parameters() map[string]interface{} {
+	return map[string]interface{}{
+		"type": "object",
+		"properties": map[string]interface{}{
+			"action": map[string]interface{}{
+				"type":        "string",
+				"description": "The memory operation to perform.",
+				"enum":        []string{"search", "read", "write", "update", "delete", "status"},
+			},
+			"query": map[string]interface{}{
+				"type":        "string",
+				"description": "Search query (for action=search).",
+			},
+			"id": map[string]interface{}{
+				"type":        "string",
+				"description": "Memory ID (for action=read/update/delete).",
+			},
+			"content": map[string]interface{}{
+				"type":        "string",
+				"description": "Content to store or update (for action=write/update).",
+			},
+			"source": map[string]interface{}{
+				"type":        "string",
+				"description": "Source label for archival writes.",
+			},
+			"sector": map[string]interface{}{
+				"type":        "string",
+				"description": "Memory sector: episodic, semantic, procedural, reflective.",
+				"enum":        []string{"episodic", "semantic", "procedural", "reflective"},
+			},
+			"tags": map[string]interface{}{
+				"type":        "string",
+				"description": "Comma-separated tags for the memory entry.",
+			},
+			"tier": map[string]interface{}{
+				"type":        "string",
+				"description": "Storage tier: recall (warm, default) or archival (cold, chunked+embedded).",
+				"enum":        []string{"recall", "archival"},
+			},
+			"limit": map[string]interface{}{
+				"type":        "integer",
+				"description": "Max results for search. Default: 5.",
+			},
+		},
+		"required": []string{"action"},
+	}
+}
+
+func (t *MemGPTTool) Execute(ctx context.Context, args map[string]interface{}) *tools.ToolResult {
+	// Marshal the args back to JSON for the inner MemoryTool.Execute()
+	input, err := json.Marshal(args)
+	if err != nil {
+		return tools.ErrorResult("invalid arguments: " + err.Error())
+	}
+
+	result, err := t.inner.Execute(ctx, string(input))
+	if err != nil {
+		return tools.ErrorResult("memory tool error: " + err.Error())
+	}
+
+	return &tools.ToolResult{
+		ForLLM: result,
+	}
+}
+
+// UpdateSession rebinds the inner MemoryTool to a new session.
+// Called when the agent switches sessions.
+func (t *MemGPTTool) UpdateSession(store *memstore.MemoryStore, agentID, session string) {
+	t.inner = memstore.NewMemoryTool(store, agentID, session)
+}

pkg/bus/bus.go

@@ -3,6 +3,8 @@ package bus
 import (
 	"context"
 	"sync"
+
+	"github.com/sipeed/picoclaw/pkg/logger"
 )
 
 type MessageBus struct {
@@ -24,6 +26,22 @@ func (mb *MessageBus) PublishInbound(msg InboundMessage) {
 	mb.inbound <- msg
 }
 
+// TryPublishInbound attempts a non-blocking send. Returns false if the buffer
+// is full (message dropped). Callers should log and/or notify the user.
+func (mb *MessageBus) TryPublishInbound(msg InboundMessage) bool {
+	select {
+	case mb.inbound <- msg:
+		return true
+	default:
+		logger.InfoCF("bus", "Inbound buffer full, message dropped",
+			map[string]interface{}{
+				"channel":   msg.Channel,
+				"sender_id": msg.SenderID,
+			})
+		return false
+	}
+}
+
 func (mb *MessageBus) ConsumeInbound(ctx context.Context) (InboundMessage, bool) {
 	select {
 	case msg := <-mb.inbound:

pkg/bus/types.go

@@ -14,6 +14,7 @@ type OutboundMessage struct {
 	Channel     string `json:"channel"`
 	ChatID      string `json:"chat_id"`
 	Content     string `json:"content"`
+	StreamDelta bool   `json:"stream_delta,omitempty"` // true = partial token (not a complete message)
 }
 
 type MessageHandler func(InboundMessage) error

pkg/channels/base.go

@@ -4,8 +4,10 @@ import (
 	"context"
 	"fmt"
 	"strings"
+	"sync/atomic"
 
 	"github.com/sipeed/picoclaw/pkg/bus"
+	"github.com/sipeed/picoclaw/pkg/logger"
 )
 
 type Channel interface {
@@ -20,19 +22,20 @@ type Channel interface {
 type BaseChannel struct {
 	config    interface{}
 	bus       *bus.MessageBus
-	running   bool
+	running   atomic.Bool
 	name      string
 	allowList []string
 }
 
 func NewBaseChannel(name string, config interface{}, bus *bus.MessageBus, allowList []string) *BaseChannel {
-	return &BaseChannel{
+	bc := &BaseChannel{
 		config:    config,
 		bus:       bus,
 		name:      name,
 		allowList: allowList,
-		running:   false,
 	}
+	bc.running.Store(false)
+	return bc
 }
 
 func (c *BaseChannel) Name() string {
@@ -40,7 +43,7 @@ func (c *BaseChannel) Name() string {
 }
 
 func (c *BaseChannel) IsRunning() bool {
-	return c.running
+	return c.running.Load()
 }
 
 func (c *BaseChannel) IsAllowed(senderID string) bool {
@@ -100,9 +103,15 @@ func (c *BaseChannel) HandleMessage(senderID, chatID, content string, media []st
 		Metadata:   metadata,
 	}
 
-	c.bus.PublishInbound(msg)
+	if !c.bus.TryPublishInbound(msg) {
+		logger.InfoCF(c.name, "Message dropped (bus full)",
+			map[string]interface{}{
+				"sender_id": msg.SenderID,
+				"channel":   msg.Channel,
+			})
+	}
 }
 
 func (c *BaseChannel) setRunning(running bool) {
-	c.running = running
+	c.running.Store(running)
 }

pkg/channels/maixcam.go

@@ -72,7 +72,7 @@ func (c *MaixCamChannel) acceptConnections(ctx context.Context) {
 		default:
 			conn, err := c.listener.Accept()
 			if err != nil {
-				if c.running {
+				if c.IsRunning() {
 					logger.ErrorCF("maixcam", "Failed to accept connection", map[string]interface{}{
 						"error": err.Error(),
 					})

pkg/config/config.go

@@ -175,6 +175,7 @@ type ProviderConfig struct {
 	APIBase    string `json:"api_base" env:"PICOCLAW_PROVIDERS_{{.Name}}_API_BASE"`
 	Proxy      string `json:"proxy,omitempty" env:"PICOCLAW_PROVIDERS_{{.Name}}_PROXY"`
 	AuthMethod string `json:"auth_method,omitempty" env:"PICOCLAW_PROVIDERS_{{.Name}}_AUTH_METHOD"`
+	Timeout    int    `json:"timeout,omitempty" env:"PICOCLAW_PROVIDERS_{{.Name}}_TIMEOUT"` // seconds, 0 = default (120s)
 }
 
 type GatewayConfig struct {
@@ -200,6 +201,7 @@ type WebToolsConfig struct {
 
 type ToolsConfig struct {
 	Web                   WebToolsConfig `json:"web"`
+	ProgressiveDisclosure bool           `json:"progressive_disclosure" env:"PICOCLAW_TOOLS_PROGRESSIVE_DISCLOSURE"`
 }
 
 func DefaultConfig() *Config {
@@ -332,9 +334,48 @@ func LoadConfig(path string) (*Config, error) {
 		return nil, err
 	}
 
+	if warnings := cfg.Validate(); len(warnings) > 0 {
+		for _, w := range warnings {
+			fmt.Fprintf(os.Stderr, "config warning: %s\n", w)
+		}
+	}
+
 	return cfg, nil
 }
 
+// Validate checks configuration for common issues. Returns a list of
+// warning strings. These are warnings rather than hard errors to maintain
+// backward compatibility, but they indicate values that will likely cause
+// problems at runtime.
+func (c *Config) Validate() []string {
+	c.mu.RLock()
+	defer c.mu.RUnlock()
+
+	var warnings []string
+
+	if c.Agents.Defaults.Model == "" {
+		warnings = append(warnings, "agents.defaults.model is empty: no default LLM model configured")
+	}
+
+	if c.Agents.Defaults.MaxTokens <= 0 {
+		warnings = append(warnings, fmt.Sprintf("agents.defaults.max_tokens=%d: should be > 0", c.Agents.Defaults.MaxTokens))
+	}
+
+	if c.Agents.Defaults.MaxToolIterations <= 0 {
+		warnings = append(warnings, fmt.Sprintf("agents.defaults.max_tool_iterations=%d: should be > 0", c.Agents.Defaults.MaxToolIterations))
+	}
+
+	if c.Gateway.Port < 0 || c.Gateway.Port > 65535 {
+		warnings = append(warnings, fmt.Sprintf("gateway.port=%d: must be in range 1-65535", c.Gateway.Port))
+	}
+
+	if c.Heartbeat.Interval < 0 {
+		warnings = append(warnings, fmt.Sprintf("heartbeat.interval=%d: must be >= 0", c.Heartbeat.Interval))
+	}
+
+	return warnings
+}
+
 func SaveConfig(path string, cfg *Config) error {
 	cfg.mu.RLock()
 	defer cfg.mu.RUnlock()

pkg/heartbeat/service.go

@@ -146,14 +146,11 @@ func (hs *HeartbeatService) runLoop(stopChan chan struct{}) {
 func (hs *HeartbeatService) executeHeartbeat() {
 	hs.mu.RLock()
 	enabled := hs.enabled
+	stopped := hs.stopChan == nil
 	handler := hs.handler
-	if !hs.enabled || hs.stopChan == nil {
-		hs.mu.RUnlock()
-		return
-	}
 	hs.mu.RUnlock()
 
-	if !enabled {
+	if !enabled || stopped {
 		return
 	}