wip

Replace custom normalizePath function with stdlib path.Clean.
Use path.IsAbs and path.Dir for cleaner, more robust code.
Add sibling escape detection to prevent traversal attacks like
"tools/a/b/../../../etc" which normalizes to "etc" (a sibling).

anthropic/anthropic.go

@@ -1,778 +0,0 @@
-package anthropic
-
-import (
-	"crypto/rand"
-	"encoding/base64"
-	"encoding/json"
-	"errors"
-	"fmt"
-	"log/slog"
-	"net/http"
-	"strings"
-	"time"
-
-	"github.com/ollama/ollama/api"
-)
-
-// Error types matching Anthropic API
-type Error struct {
-	Type    string `json:"type"`
-	Message string `json:"message"`
-}
-
-type ErrorResponse struct {
-	Type      string `json:"type"` // always "error"
-	Error     Error  `json:"error"`
-	RequestID string `json:"request_id,omitempty"`
-}
-
-// NewError creates a new ErrorResponse with the appropriate error type based on HTTP status code
-func NewError(code int, message string) ErrorResponse {
-	var etype string
-	switch code {
-	case http.StatusBadRequest:
-		etype = "invalid_request_error"
-	case http.StatusUnauthorized:
-		etype = "authentication_error"
-	case http.StatusForbidden:
-		etype = "permission_error"
-	case http.StatusNotFound:
-		etype = "not_found_error"
-	case http.StatusTooManyRequests:
-		etype = "rate_limit_error"
-	case http.StatusServiceUnavailable, 529:
-		etype = "overloaded_error"
-	default:
-		etype = "api_error"
-	}
-
-	return ErrorResponse{
-		Type:      "error",
-		Error:     Error{Type: etype, Message: message},
-		RequestID: generateID("req"),
-	}
-}
-
-// Request types
-
-// MessagesRequest represents an Anthropic Messages API request
-type MessagesRequest struct {
-	Model         string          `json:"model"`
-	MaxTokens     int             `json:"max_tokens"`
-	Messages      []MessageParam  `json:"messages"`
-	System        any             `json:"system,omitempty"` // string or []ContentBlock
-	Stream        bool            `json:"stream,omitempty"`
-	Temperature   *float64        `json:"temperature,omitempty"`
-	TopP          *float64        `json:"top_p,omitempty"`
-	TopK          *int            `json:"top_k,omitempty"`
-	StopSequences []string        `json:"stop_sequences,omitempty"`
-	Tools         []Tool          `json:"tools,omitempty"`
-	ToolChoice    *ToolChoice     `json:"tool_choice,omitempty"`
-	Thinking      *ThinkingConfig `json:"thinking,omitempty"`
-	Metadata      *Metadata       `json:"metadata,omitempty"`
-}
-
-// MessageParam represents a message in the request
-type MessageParam struct {
-	Role    string `json:"role"`    // "user" or "assistant"
-	Content any    `json:"content"` // string or []ContentBlock
-}
-
-// ContentBlock represents a content block in a message.
-// Text and Thinking use pointers so they serialize as the field being present (even if empty)
-// only when set, which is required for SDK streaming accumulation.
-type ContentBlock struct {
-	Type string `json:"type"` // text, image, tool_use, tool_result, thinking
-
-	// For text blocks - pointer so field only appears when set (SDK requires it for accumulation)
-	Text *string `json:"text,omitempty"`
-
-	// For image blocks
-	Source *ImageSource `json:"source,omitempty"`
-
-	// For tool_use blocks
-	ID    string `json:"id,omitempty"`
-	Name  string `json:"name,omitempty"`
-	Input any    `json:"input,omitempty"`
-
-	// For tool_result blocks
-	ToolUseID string `json:"tool_use_id,omitempty"`
-	Content   any    `json:"content,omitempty"` // string or []ContentBlock
-	IsError   bool   `json:"is_error,omitempty"`
-
-	// For thinking blocks - pointer so field only appears when set (SDK requires it for accumulation)
-	Thinking  *string `json:"thinking,omitempty"`
-	Signature string  `json:"signature,omitempty"`
-}
-
-// ImageSource represents the source of an image
-type ImageSource struct {
-	Type      string `json:"type"` // "base64" or "url"
-	MediaType string `json:"media_type,omitempty"`
-	Data      string `json:"data,omitempty"`
-	URL       string `json:"url,omitempty"`
-}
-
-// Tool represents a tool definition
-type Tool struct {
-	Type        string          `json:"type,omitempty"` // "custom" for user-defined tools
-	Name        string          `json:"name"`
-	Description string          `json:"description,omitempty"`
-	InputSchema json.RawMessage `json:"input_schema,omitempty"`
-}
-
-// ToolChoice controls how the model uses tools
-type ToolChoice struct {
-	Type                   string `json:"type"` // "auto", "any", "tool", "none"
-	Name                   string `json:"name,omitempty"`
-	DisableParallelToolUse bool   `json:"disable_parallel_tool_use,omitempty"`
-}
-
-// ThinkingConfig controls extended thinking
-type ThinkingConfig struct {
-	Type         string `json:"type"` // "enabled" or "disabled"
-	BudgetTokens int    `json:"budget_tokens,omitempty"`
-}
-
-// Metadata for the request
-type Metadata struct {
-	UserID string `json:"user_id,omitempty"`
-}
-
-// Response types
-
-// MessagesResponse represents an Anthropic Messages API response
-type MessagesResponse struct {
-	ID           string         `json:"id"`
-	Type         string         `json:"type"` // "message"
-	Role         string         `json:"role"` // "assistant"
-	Model        string         `json:"model"`
-	Content      []ContentBlock `json:"content"`
-	StopReason   string         `json:"stop_reason,omitempty"`
-	StopSequence string         `json:"stop_sequence,omitempty"`
-	Usage        Usage          `json:"usage"`
-}
-
-// Usage contains token usage information
-type Usage struct {
-	InputTokens  int `json:"input_tokens"`
-	OutputTokens int `json:"output_tokens"`
-}
-
-// Streaming event types
-
-// MessageStartEvent is sent at the start of streaming
-type MessageStartEvent struct {
-	Type    string           `json:"type"` // "message_start"
-	Message MessagesResponse `json:"message"`
-}
-
-// ContentBlockStartEvent signals the start of a content block
-type ContentBlockStartEvent struct {
-	Type         string       `json:"type"` // "content_block_start"
-	Index        int          `json:"index"`
-	ContentBlock ContentBlock `json:"content_block"`
-}
-
-// ContentBlockDeltaEvent contains incremental content updates
-type ContentBlockDeltaEvent struct {
-	Type  string `json:"type"` // "content_block_delta"
-	Index int    `json:"index"`
-	Delta Delta  `json:"delta"`
-}
-
-// Delta represents an incremental update
-type Delta struct {
-	Type        string `json:"type"` // "text_delta", "input_json_delta", "thinking_delta", "signature_delta"
-	Text        string `json:"text,omitempty"`
-	PartialJSON string `json:"partial_json,omitempty"`
-	Thinking    string `json:"thinking,omitempty"`
-	Signature   string `json:"signature,omitempty"`
-}
-
-// ContentBlockStopEvent signals the end of a content block
-type ContentBlockStopEvent struct {
-	Type  string `json:"type"` // "content_block_stop"
-	Index int    `json:"index"`
-}
-
-// MessageDeltaEvent contains updates to the message
-type MessageDeltaEvent struct {
-	Type  string       `json:"type"` // "message_delta"
-	Delta MessageDelta `json:"delta"`
-	Usage DeltaUsage   `json:"usage"`
-}
-
-// MessageDelta contains stop information
-type MessageDelta struct {
-	StopReason   string `json:"stop_reason,omitempty"`
-	StopSequence string `json:"stop_sequence,omitempty"`
-}
-
-// DeltaUsage contains cumulative token usage
-type DeltaUsage struct {
-	OutputTokens int `json:"output_tokens"`
-}
-
-// MessageStopEvent signals the end of the message
-type MessageStopEvent struct {
-	Type string `json:"type"` // "message_stop"
-}
-
-// PingEvent is a keepalive event
-type PingEvent struct {
-	Type string `json:"type"` // "ping"
-}
-
-// StreamErrorEvent is an error during streaming
-type StreamErrorEvent struct {
-	Type  string `json:"type"` // "error"
-	Error Error  `json:"error"`
-}
-
-// FromMessagesRequest converts an Anthropic MessagesRequest to an Ollama api.ChatRequest
-func FromMessagesRequest(r MessagesRequest) (*api.ChatRequest, error) {
-	var messages []api.Message
-
-	if r.System != nil {
-		switch sys := r.System.(type) {
-		case string:
-			if sys != "" {
-				messages = append(messages, api.Message{Role: "system", Content: sys})
-			}
-		case []any:
-			// System can be an array of content blocks
-			var content strings.Builder
-			for _, block := range sys {
-				if blockMap, ok := block.(map[string]any); ok {
-					if blockMap["type"] == "text" {
-						if text, ok := blockMap["text"].(string); ok {
-							content.WriteString(text)
-						}
-					}
-				}
-			}
-			if content.Len() > 0 {
-				messages = append(messages, api.Message{Role: "system", Content: content.String()})
-			}
-		}
-	}
-
-	for _, msg := range r.Messages {
-		converted, err := convertMessage(msg)
-		if err != nil {
-			return nil, err
-		}
-		messages = append(messages, converted...)
-	}
-
-	options := make(map[string]any)
-
-	options["num_predict"] = r.MaxTokens
-
-	if r.Temperature != nil {
-		options["temperature"] = *r.Temperature
-	}
-
-	if r.TopP != nil {
-		options["top_p"] = *r.TopP
-	}
-
-	if r.TopK != nil {
-		options["top_k"] = *r.TopK
-	}
-
-	if len(r.StopSequences) > 0 {
-		options["stop"] = r.StopSequences
-	}
-
-	var tools api.Tools
-	for _, t := range r.Tools {
-		tool, err := convertTool(t)
-		if err != nil {
-			return nil, err
-		}
-		tools = append(tools, tool)
-	}
-
-	var think *api.ThinkValue
-	if r.Thinking != nil && r.Thinking.Type == "enabled" {
-		think = &api.ThinkValue{Value: true}
-	}
-
-	stream := r.Stream
-
-	return &api.ChatRequest{
-		Model:    r.Model,
-		Messages: messages,
-		Options:  options,
-		Stream:   &stream,
-		Tools:    tools,
-		Think:    think,
-	}, nil
-}
-
-// convertMessage converts an Anthropic MessageParam to Ollama api.Message(s)
-func convertMessage(msg MessageParam) ([]api.Message, error) {
-	var messages []api.Message
-	role := strings.ToLower(msg.Role)
-
-	switch content := msg.Content.(type) {
-	case string:
-		messages = append(messages, api.Message{Role: role, Content: content})
-
-	case []any:
-		var textContent strings.Builder
-		var images []api.ImageData
-		var toolCalls []api.ToolCall
-		var thinking string
-		var toolResults []api.Message
-
-		for _, block := range content {
-			blockMap, ok := block.(map[string]any)
-			if !ok {
-				return nil, errors.New("invalid content block format")
-			}
-
-			blockType, _ := blockMap["type"].(string)
-
-			switch blockType {
-			case "text":
-				if text, ok := blockMap["text"].(string); ok {
-					textContent.WriteString(text)
-				}
-
-			case "image":
-				source, ok := blockMap["source"].(map[string]any)
-				if !ok {
-					return nil, errors.New("invalid image source")
-				}
-
-				sourceType, _ := source["type"].(string)
-				if sourceType == "base64" {
-					data, _ := source["data"].(string)
-					decoded, err := base64.StdEncoding.DecodeString(data)
-					if err != nil {
-						return nil, fmt.Errorf("invalid base64 image data: %w", err)
-					}
-					images = append(images, decoded)
-				} else {
-					return nil, fmt.Errorf("invalid image source type: %s. Only base64 images are supported.", sourceType)
-				}
-				// URL images would need to be fetched - skip for now
-
-			case "tool_use":
-				id, ok := blockMap["id"].(string)
-				if !ok {
-					return nil, errors.New("tool_use block missing required 'id' field")
-				}
-				name, ok := blockMap["name"].(string)
-				if !ok {
-					return nil, errors.New("tool_use block missing required 'name' field")
-				}
-				tc := api.ToolCall{
-					ID: id,
-					Function: api.ToolCallFunction{
-						Name: name,
-					},
-				}
-				if input, ok := blockMap["input"].(map[string]any); ok {
-					tc.Function.Arguments = mapToArgs(input)
-				}
-				toolCalls = append(toolCalls, tc)
-
-			case "tool_result":
-				toolUseID, _ := blockMap["tool_use_id"].(string)
-				var resultContent string
-
-				switch c := blockMap["content"].(type) {
-				case string:
-					resultContent = c
-				case []any:
-					for _, cb := range c {
-						if cbMap, ok := cb.(map[string]any); ok {
-							if cbMap["type"] == "text" {
-								if text, ok := cbMap["text"].(string); ok {
-									resultContent += text
-								}
-							}
-						}
-					}
-				}
-
-				toolResults = append(toolResults, api.Message{
-					Role:       "tool",
-					Content:    resultContent,
-					ToolCallID: toolUseID,
-				})
-
-			case "thinking":
-				if t, ok := blockMap["thinking"].(string); ok {
-					thinking = t
-				}
-			}
-		}
-
-		if textContent.Len() > 0 || len(images) > 0 || len(toolCalls) > 0 || thinking != "" {
-			m := api.Message{
-				Role:      role,
-				Content:   textContent.String(),
-				Images:    images,
-				ToolCalls: toolCalls,
-				Thinking:  thinking,
-			}
-			messages = append(messages, m)
-		}
-
-		// Add tool results as separate messages
-		messages = append(messages, toolResults...)
-
-	default:
-		return nil, fmt.Errorf("invalid message content type: %T", content)
-	}
-
-	return messages, nil
-}
-
-// convertTool converts an Anthropic Tool to an Ollama api.Tool
-func convertTool(t Tool) (api.Tool, error) {
-	var params api.ToolFunctionParameters
-	if len(t.InputSchema) > 0 {
-		if err := json.Unmarshal(t.InputSchema, &params); err != nil {
-			return api.Tool{}, fmt.Errorf("invalid input_schema for tool %q: %w", t.Name, err)
-		}
-	}
-
-	return api.Tool{
-		Type: "function",
-		Function: api.ToolFunction{
-			Name:        t.Name,
-			Description: t.Description,
-			Parameters:  params,
-		},
-	}, nil
-}
-
-// ToMessagesResponse converts an Ollama api.ChatResponse to an Anthropic MessagesResponse
-func ToMessagesResponse(id string, r api.ChatResponse) MessagesResponse {
-	var content []ContentBlock
-
-	if r.Message.Thinking != "" {
-		content = append(content, ContentBlock{
-			Type:     "thinking",
-			Thinking: ptr(r.Message.Thinking),
-		})
-	}
-
-	if r.Message.Content != "" {
-		content = append(content, ContentBlock{
-			Type: "text",
-			Text: ptr(r.Message.Content),
-		})
-	}
-
-	for _, tc := range r.Message.ToolCalls {
-		content = append(content, ContentBlock{
-			Type:  "tool_use",
-			ID:    tc.ID,
-			Name:  tc.Function.Name,
-			Input: tc.Function.Arguments,
-		})
-	}
-
-	stopReason := mapStopReason(r.DoneReason, len(r.Message.ToolCalls) > 0)
-
-	return MessagesResponse{
-		ID:         id,
-		Type:       "message",
-		Role:       "assistant",
-		Model:      r.Model,
-		Content:    content,
-		StopReason: stopReason,
-		Usage: Usage{
-			InputTokens:  r.Metrics.PromptEvalCount,
-			OutputTokens: r.Metrics.EvalCount,
-		},
-	}
-}
-
-// mapStopReason converts Ollama done_reason to Anthropic stop_reason
-func mapStopReason(reason string, hasToolCalls bool) string {
-	if hasToolCalls {
-		return "tool_use"
-	}
-
-	switch reason {
-	case "stop":
-		return "end_turn"
-	case "length":
-		return "max_tokens"
-	default:
-		if reason != "" {
-			return "stop_sequence"
-		}
-		return ""
-	}
-}
-
-// StreamConverter manages state for converting Ollama streaming responses to Anthropic format
-type StreamConverter struct {
-	ID              string
-	Model           string
-	firstWrite      bool
-	contentIndex    int
-	inputTokens     int
-	outputTokens    int
-	thinkingStarted bool
-	thinkingDone    bool
-	textStarted     bool
-	toolCallsSent   map[string]bool
-}
-
-func NewStreamConverter(id, model string) *StreamConverter {
-	return &StreamConverter{
-		ID:            id,
-		Model:         model,
-		firstWrite:    true,
-		toolCallsSent: make(map[string]bool),
-	}
-}
-
-// StreamEvent represents a streaming event to be sent to the client
-type StreamEvent struct {
-	Event string
-	Data  any
-}
-
-// Process converts an Ollama ChatResponse to Anthropic streaming events
-func (c *StreamConverter) Process(r api.ChatResponse) []StreamEvent {
-	var events []StreamEvent
-
-	if c.firstWrite {
-		c.firstWrite = false
-		c.inputTokens = r.Metrics.PromptEvalCount
-
-		events = append(events, StreamEvent{
-			Event: "message_start",
-			Data: MessageStartEvent{
-				Type: "message_start",
-				Message: MessagesResponse{
-					ID:      c.ID,
-					Type:    "message",
-					Role:    "assistant",
-					Model:   c.Model,
-					Content: []ContentBlock{},
-					Usage: Usage{
-						InputTokens:  c.inputTokens,
-						OutputTokens: 0,
-					},
-				},
-			},
-		})
-	}
-
-	if r.Message.Thinking != "" && !c.thinkingDone {
-		if !c.thinkingStarted {
-			c.thinkingStarted = true
-			events = append(events, StreamEvent{
-				Event: "content_block_start",
-				Data: ContentBlockStartEvent{
-					Type:  "content_block_start",
-					Index: c.contentIndex,
-					ContentBlock: ContentBlock{
-						Type:     "thinking",
-						Thinking: ptr(""),
-					},
-				},
-			})
-		}
-
-		events = append(events, StreamEvent{
-			Event: "content_block_delta",
-			Data: ContentBlockDeltaEvent{
-				Type:  "content_block_delta",
-				Index: c.contentIndex,
-				Delta: Delta{
-					Type:     "thinking_delta",
-					Thinking: r.Message.Thinking,
-				},
-			},
-		})
-	}
-
-	if r.Message.Content != "" {
-		if c.thinkingStarted && !c.thinkingDone {
-			c.thinkingDone = true
-			events = append(events, StreamEvent{
-				Event: "content_block_stop",
-				Data: ContentBlockStopEvent{
-					Type:  "content_block_stop",
-					Index: c.contentIndex,
-				},
-			})
-			c.contentIndex++
-		}
-
-		if !c.textStarted {
-			c.textStarted = true
-			events = append(events, StreamEvent{
-				Event: "content_block_start",
-				Data: ContentBlockStartEvent{
-					Type:  "content_block_start",
-					Index: c.contentIndex,
-					ContentBlock: ContentBlock{
-						Type: "text",
-						Text: ptr(""),
-					},
-				},
-			})
-		}
-
-		events = append(events, StreamEvent{
-			Event: "content_block_delta",
-			Data: ContentBlockDeltaEvent{
-				Type:  "content_block_delta",
-				Index: c.contentIndex,
-				Delta: Delta{
-					Type: "text_delta",
-					Text: r.Message.Content,
-				},
-			},
-		})
-	}
-
-	for _, tc := range r.Message.ToolCalls {
-		if c.toolCallsSent[tc.ID] {
-			continue
-		}
-
-		if c.textStarted {
-			events = append(events, StreamEvent{
-				Event: "content_block_stop",
-				Data: ContentBlockStopEvent{
-					Type:  "content_block_stop",
-					Index: c.contentIndex,
-				},
-			})
-			c.contentIndex++
-			c.textStarted = false
-		}
-
-		argsJSON, err := json.Marshal(tc.Function.Arguments)
-		if err != nil {
-			slog.Error("failed to marshal tool arguments", "error", err, "tool_id", tc.ID)
-			continue
-		}
-
-		events = append(events, StreamEvent{
-			Event: "content_block_start",
-			Data: ContentBlockStartEvent{
-				Type:  "content_block_start",
-				Index: c.contentIndex,
-				ContentBlock: ContentBlock{
-					Type:  "tool_use",
-					ID:    tc.ID,
-					Name:  tc.Function.Name,
-					Input: map[string]any{},
-				},
-			},
-		})
-
-		events = append(events, StreamEvent{
-			Event: "content_block_delta",
-			Data: ContentBlockDeltaEvent{
-				Type:  "content_block_delta",
-				Index: c.contentIndex,
-				Delta: Delta{
-					Type:        "input_json_delta",
-					PartialJSON: string(argsJSON),
-				},
-			},
-		})
-
-		events = append(events, StreamEvent{
-			Event: "content_block_stop",
-			Data: ContentBlockStopEvent{
-				Type:  "content_block_stop",
-				Index: c.contentIndex,
-			},
-		})
-
-		c.toolCallsSent[tc.ID] = true
-		c.contentIndex++
-	}
-
-	if r.Done {
-		if c.textStarted {
-			events = append(events, StreamEvent{
-				Event: "content_block_stop",
-				Data: ContentBlockStopEvent{
-					Type:  "content_block_stop",
-					Index: c.contentIndex,
-				},
-			})
-		} else if c.thinkingStarted && !c.thinkingDone {
-			events = append(events, StreamEvent{
-				Event: "content_block_stop",
-				Data: ContentBlockStopEvent{
-					Type:  "content_block_stop",
-					Index: c.contentIndex,
-				},
-			})
-		}
-
-		c.outputTokens = r.Metrics.EvalCount
-		stopReason := mapStopReason(r.DoneReason, len(c.toolCallsSent) > 0)
-
-		events = append(events, StreamEvent{
-			Event: "message_delta",
-			Data: MessageDeltaEvent{
-				Type: "message_delta",
-				Delta: MessageDelta{
-					StopReason: stopReason,
-				},
-				Usage: DeltaUsage{
-					OutputTokens: c.outputTokens,
-				},
-			},
-		})
-
-		events = append(events, StreamEvent{
-			Event: "message_stop",
-			Data: MessageStopEvent{
-				Type: "message_stop",
-			},
-		})
-	}
-
-	return events
-}
-
-// generateID generates a unique ID with the given prefix using crypto/rand
-func generateID(prefix string) string {
-	b := make([]byte, 12)
-	if _, err := rand.Read(b); err != nil {
-		// Fallback to time-based ID if crypto/rand fails
-		return fmt.Sprintf("%s_%d", prefix, time.Now().UnixNano())
-	}
-	return fmt.Sprintf("%s_%x", prefix, b)
-}
-
-// GenerateMessageID generates a unique message ID
-func GenerateMessageID() string {
-	return generateID("msg")
-}
-
-// ptr returns a pointer to the given string value
-func ptr(s string) *string {
-	return &s
-}
-
-// mapToArgs converts a map to ToolCallFunctionArguments
-func mapToArgs(m map[string]any) api.ToolCallFunctionArguments {
-	args := api.NewToolCallFunctionArguments()
-	for k, v := range m {
-		args.Set(k, v)
-	}
-	return args
-}

anthropic/anthropic_test.go

@@ -1,953 +0,0 @@
-package anthropic
-
-import (
-	"encoding/base64"
-	"encoding/json"
-	"testing"
-
-	"github.com/google/go-cmp/cmp"
-
-	"github.com/ollama/ollama/api"
-)
-
-const (
-	testImage = `iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR42mNk+A8AAQUBAScY42YAAAAASUVORK5CYII=`
-)
-
-// testArgs creates ToolCallFunctionArguments from a map (convenience function for tests)
-func testArgs(m map[string]any) api.ToolCallFunctionArguments {
-	args := api.NewToolCallFunctionArguments()
-	for k, v := range m {
-		args.Set(k, v)
-	}
-	return args
-}
-
-func TestFromMessagesRequest_Basic(t *testing.T) {
-	req := MessagesRequest{
-		Model:     "test-model",
-		MaxTokens: 1024,
-		Messages: []MessageParam{
-			{Role: "user", Content: "Hello"},
-		},
-	}
-
-	result, err := FromMessagesRequest(req)
-	if err != nil {
-		t.Fatalf("unexpected error: %v", err)
-	}
-
-	if result.Model != "test-model" {
-		t.Errorf("expected model 'test-model', got %q", result.Model)
-	}
-
-	if len(result.Messages) != 1 {
-		t.Fatalf("expected 1 message, got %d", len(result.Messages))
-	}
-
-	if result.Messages[0].Role != "user" || result.Messages[0].Content != "Hello" {
-		t.Errorf("unexpected message: %+v", result.Messages[0])
-	}
-
-	if numPredict, ok := result.Options["num_predict"].(int); !ok || numPredict != 1024 {
-		t.Errorf("expected num_predict 1024, got %v", result.Options["num_predict"])
-	}
-}
-
-func TestFromMessagesRequest_WithSystemPrompt(t *testing.T) {
-	req := MessagesRequest{
-		Model:     "test-model",
-		MaxTokens: 1024,
-		System:    "You are a helpful assistant.",
-		Messages: []MessageParam{
-			{Role: "user", Content: "Hello"},
-		},
-	}
-
-	result, err := FromMessagesRequest(req)
-	if err != nil {
-		t.Fatalf("unexpected error: %v", err)
-	}
-
-	if len(result.Messages) != 2 {
-		t.Fatalf("expected 2 messages, got %d", len(result.Messages))
-	}
-
-	if result.Messages[0].Role != "system" || result.Messages[0].Content != "You are a helpful assistant." {
-		t.Errorf("unexpected system message: %+v", result.Messages[0])
-	}
-}
-
-func TestFromMessagesRequest_WithSystemPromptArray(t *testing.T) {
-	req := MessagesRequest{
-		Model:     "test-model",
-		MaxTokens: 1024,
-		System: []any{
-			map[string]any{"type": "text", "text": "You are helpful."},
-			map[string]any{"type": "text", "text": " Be concise."},
-		},
-		Messages: []MessageParam{
-			{Role: "user", Content: "Hello"},
-		},
-	}
-
-	result, err := FromMessagesRequest(req)
-	if err != nil {
-		t.Fatalf("unexpected error: %v", err)
-	}
-
-	if len(result.Messages) != 2 {
-		t.Fatalf("expected 2 messages, got %d", len(result.Messages))
-	}
-
-	if result.Messages[0].Content != "You are helpful. Be concise." {
-		t.Errorf("unexpected system message content: %q", result.Messages[0].Content)
-	}
-}
-
-func TestFromMessagesRequest_WithOptions(t *testing.T) {
-	temp := 0.7
-	topP := 0.9
-	topK := 40
-	req := MessagesRequest{
-		Model:         "test-model",
-		MaxTokens:     2048,
-		Messages:      []MessageParam{{Role: "user", Content: "Hello"}},
-		Temperature:   &temp,
-		TopP:          &topP,
-		TopK:          &topK,
-		StopSequences: []string{"\n", "END"},
-	}
-
-	result, err := FromMessagesRequest(req)
-	if err != nil {
-		t.Fatalf("unexpected error: %v", err)
-	}
-
-	if result.Options["temperature"] != 0.7 {
-		t.Errorf("expected temperature 0.7, got %v", result.Options["temperature"])
-	}
-	if result.Options["top_p"] != 0.9 {
-		t.Errorf("expected top_p 0.9, got %v", result.Options["top_p"])
-	}
-	if result.Options["top_k"] != 40 {
-		t.Errorf("expected top_k 40, got %v", result.Options["top_k"])
-	}
-	if diff := cmp.Diff([]string{"\n", "END"}, result.Options["stop"]); diff != "" {
-		t.Errorf("stop sequences mismatch: %s", diff)
-	}
-}
-
-func TestFromMessagesRequest_WithImage(t *testing.T) {
-	imgData, _ := base64.StdEncoding.DecodeString(testImage)
-
-	req := MessagesRequest{
-		Model:     "test-model",
-		MaxTokens: 1024,
-		Messages: []MessageParam{
-			{
-				Role: "user",
-				Content: []any{
-					map[string]any{"type": "text", "text": "What's in this image?"},
-					map[string]any{
-						"type": "image",
-						"source": map[string]any{
-							"type":       "base64",
-							"media_type": "image/png",
-							"data":       testImage,
-						},
-					},
-				},
-			},
-		},
-	}
-
-	result, err := FromMessagesRequest(req)
-	if err != nil {
-		t.Fatalf("unexpected error: %v", err)
-	}
-
-	if len(result.Messages) != 1 {
-		t.Fatalf("expected 1 message, got %d", len(result.Messages))
-	}
-
-	if result.Messages[0].Content != "What's in this image?" {
-		t.Errorf("expected content 'What's in this image?', got %q", result.Messages[0].Content)
-	}
-
-	if len(result.Messages[0].Images) != 1 {
-		t.Fatalf("expected 1 image, got %d", len(result.Messages[0].Images))
-	}
-
-	if string(result.Messages[0].Images[0]) != string(imgData) {
-		t.Error("image data mismatch")
-	}
-}
-
-func TestFromMessagesRequest_WithToolUse(t *testing.T) {
-	req := MessagesRequest{
-		Model:     "test-model",
-		MaxTokens: 1024,
-		Messages: []MessageParam{
-			{Role: "user", Content: "What's the weather in Paris?"},
-			{
-				Role: "assistant",
-				Content: []any{
-					map[string]any{
-						"type":  "tool_use",
-						"id":    "call_123",
-						"name":  "get_weather",
-						"input": map[string]any{"location": "Paris"},
-					},
-				},
-			},
-		},
-	}
-
-	result, err := FromMessagesRequest(req)
-	if err != nil {
-		t.Fatalf("unexpected error: %v", err)
-	}
-
-	if len(result.Messages) != 2 {
-		t.Fatalf("expected 2 messages, got %d", len(result.Messages))
-	}
-
-	if len(result.Messages[1].ToolCalls) != 1 {
-		t.Fatalf("expected 1 tool call, got %d", len(result.Messages[1].ToolCalls))
-	}
-
-	tc := result.Messages[1].ToolCalls[0]
-	if tc.ID != "call_123" {
-		t.Errorf("expected tool call ID 'call_123', got %q", tc.ID)
-	}
-	if tc.Function.Name != "get_weather" {
-		t.Errorf("expected tool name 'get_weather', got %q", tc.Function.Name)
-	}
-}
-
-func TestFromMessagesRequest_WithToolResult(t *testing.T) {
-	req := MessagesRequest{
-		Model:     "test-model",
-		MaxTokens: 1024,
-		Messages: []MessageParam{
-			{
-				Role: "user",
-				Content: []any{
-					map[string]any{
-						"type":        "tool_result",
-						"tool_use_id": "call_123",
-						"content":     "The weather in Paris is sunny, 22°C",
-					},
-				},
-			},
-		},
-	}
-
-	result, err := FromMessagesRequest(req)
-	if err != nil {
-		t.Fatalf("unexpected error: %v", err)
-	}
-
-	if len(result.Messages) != 1 {
-		t.Fatalf("expected 1 message, got %d", len(result.Messages))
-	}
-
-	msg := result.Messages[0]
-	if msg.Role != "tool" {
-		t.Errorf("expected role 'tool', got %q", msg.Role)
-	}
-	if msg.ToolCallID != "call_123" {
-		t.Errorf("expected tool_call_id 'call_123', got %q", msg.ToolCallID)
-	}
-	if msg.Content != "The weather in Paris is sunny, 22°C" {
-		t.Errorf("unexpected content: %q", msg.Content)
-	}
-}
-
-func TestFromMessagesRequest_WithTools(t *testing.T) {
-	req := MessagesRequest{
-		Model:     "test-model",
-		MaxTokens: 1024,
-		Messages:  []MessageParam{{Role: "user", Content: "Hello"}},
-		Tools: []Tool{
-			{
-				Name:        "get_weather",
-				Description: "Get current weather",
-				InputSchema: json.RawMessage(`{"type":"object","properties":{"location":{"type":"string"}},"required":["location"]}`),
-			},
-		},
-	}
-
-	result, err := FromMessagesRequest(req)
-	if err != nil {
-		t.Fatalf("unexpected error: %v", err)
-	}
-
-	if len(result.Tools) != 1 {
-		t.Fatalf("expected 1 tool, got %d", len(result.Tools))
-	}
-
-	tool := result.Tools[0]
-	if tool.Type != "function" {
-		t.Errorf("expected type 'function', got %q", tool.Type)
-	}
-	if tool.Function.Name != "get_weather" {
-		t.Errorf("expected name 'get_weather', got %q", tool.Function.Name)
-	}
-	if tool.Function.Description != "Get current weather" {
-		t.Errorf("expected description 'Get current weather', got %q", tool.Function.Description)
-	}
-}
-
-func TestFromMessagesRequest_WithThinking(t *testing.T) {
-	req := MessagesRequest{
-		Model:     "test-model",
-		MaxTokens: 1024,
-		Messages:  []MessageParam{{Role: "user", Content: "Hello"}},
-		Thinking:  &ThinkingConfig{Type: "enabled", BudgetTokens: 1000},
-	}
-
-	result, err := FromMessagesRequest(req)
-	if err != nil {
-		t.Fatalf("unexpected error: %v", err)
-	}
-
-	if result.Think == nil {
-		t.Fatal("expected Think to be set")
-	}
-	if v, ok := result.Think.Value.(bool); !ok || !v {
-		t.Errorf("expected Think.Value to be true, got %v", result.Think.Value)
-	}
-}
-
-// TestFromMessagesRequest_ThinkingOnlyBlock verifies that messages containing only
-// a thinking block (no text, images, or tool calls) are preserved and not dropped.
-func TestFromMessagesRequest_ThinkingOnlyBlock(t *testing.T) {
-	req := MessagesRequest{
-		Model:     "test-model",
-		MaxTokens: 1024,
-		Messages: []MessageParam{
-			{Role: "user", Content: "Hello"},
-			{
-				Role: "assistant",
-				Content: []any{
-					map[string]any{
-						"type":     "thinking",
-						"thinking": "Let me think about this...",
-					},
-				},
-			},
-		},
-	}
-
-	result, err := FromMessagesRequest(req)
-	if err != nil {
-		t.Fatalf("unexpected error: %v", err)
-	}
-
-	if len(result.Messages) != 2 {
-		t.Fatalf("expected 2 messages, got %d", len(result.Messages))
-	}
-
-	assistantMsg := result.Messages[1]
-	if assistantMsg.Thinking != "Let me think about this..." {
-		t.Errorf("expected thinking content, got %q", assistantMsg.Thinking)
-	}
-}
-
-func TestFromMessagesRequest_ToolUseMissingID(t *testing.T) {
-	req := MessagesRequest{
-		Model:     "test-model",
-		MaxTokens: 1024,
-		Messages: []MessageParam{
-			{
-				Role: "assistant",
-				Content: []any{
-					map[string]any{
-						"type": "tool_use",
-						"name": "get_weather",
-					},
-				},
-			},
-		},
-	}
-
-	_, err := FromMessagesRequest(req)
-	if err == nil {
-		t.Fatal("expected error for missing tool_use id")
-	}
-	if err.Error() != "tool_use block missing required 'id' field" {
-		t.Errorf("unexpected error message: %v", err)
-	}
-}
-
-func TestFromMessagesRequest_ToolUseMissingName(t *testing.T) {
-	req := MessagesRequest{
-		Model:     "test-model",
-		MaxTokens: 1024,
-		Messages: []MessageParam{
-			{
-				Role: "assistant",
-				Content: []any{
-					map[string]any{
-						"type": "tool_use",
-						"id":   "call_123",
-					},
-				},
-			},
-		},
-	}
-
-	_, err := FromMessagesRequest(req)
-	if err == nil {
-		t.Fatal("expected error for missing tool_use name")
-	}
-	if err.Error() != "tool_use block missing required 'name' field" {
-		t.Errorf("unexpected error message: %v", err)
-	}
-}
-
-func TestFromMessagesRequest_InvalidToolSchema(t *testing.T) {
-	req := MessagesRequest{
-		Model:     "test-model",
-		MaxTokens: 1024,
-		Messages:  []MessageParam{{Role: "user", Content: "Hello"}},
-		Tools: []Tool{
-			{
-				Name:        "bad_tool",
-				InputSchema: json.RawMessage(`{invalid json`),
-			},
-		},
-	}
-
-	_, err := FromMessagesRequest(req)
-	if err == nil {
-		t.Fatal("expected error for invalid tool schema")
-	}
-}
-
-func TestToMessagesResponse_Basic(t *testing.T) {
-	resp := api.ChatResponse{
-		Model: "test-model",
-		Message: api.Message{
-			Role:    "assistant",
-			Content: "Hello there!",
-		},
-		Done:       true,
-		DoneReason: "stop",
-		Metrics: api.Metrics{
-			PromptEvalCount: 10,
-			EvalCount:       5,
-		},
-	}
-
-	result := ToMessagesResponse("msg_123", resp)
-
-	if result.ID != "msg_123" {
-		t.Errorf("expected ID 'msg_123', got %q", result.ID)
-	}
-	if result.Type != "message" {
-		t.Errorf("expected type 'message', got %q", result.Type)
-	}
-	if result.Role != "assistant" {
-		t.Errorf("expected role 'assistant', got %q", result.Role)
-	}
-	if len(result.Content) != 1 {
-		t.Fatalf("expected 1 content block, got %d", len(result.Content))
-	}
-	if result.Content[0].Type != "text" || result.Content[0].Text == nil || *result.Content[0].Text != "Hello there!" {
-		t.Errorf("unexpected content: %+v", result.Content[0])
-	}
-	if result.StopReason != "end_turn" {
-		t.Errorf("expected stop_reason 'end_turn', got %q", result.StopReason)
-	}
-	if result.Usage.InputTokens != 10 || result.Usage.OutputTokens != 5 {
-		t.Errorf("unexpected usage: %+v", result.Usage)
-	}
-}
-
-func TestToMessagesResponse_WithToolCalls(t *testing.T) {
-	resp := api.ChatResponse{
-		Model: "test-model",
-		Message: api.Message{
-			Role: "assistant",
-			ToolCalls: []api.ToolCall{
-				{
-					ID: "call_123",
-					Function: api.ToolCallFunction{
-						Name:      "get_weather",
-						Arguments: testArgs(map[string]any{"location": "Paris"}),
-					},
-				},
-			},
-		},
-		Done:       true,
-		DoneReason: "stop",
-	}
-
-	result := ToMessagesResponse("msg_123", resp)
-
-	if len(result.Content) != 1 {
-		t.Fatalf("expected 1 content block, got %d", len(result.Content))
-	}
-	if result.Content[0].Type != "tool_use" {
-		t.Errorf("expected type 'tool_use', got %q", result.Content[0].Type)
-	}
-	if result.Content[0].ID != "call_123" {
-		t.Errorf("expected ID 'call_123', got %q", result.Content[0].ID)
-	}
-	if result.Content[0].Name != "get_weather" {
-		t.Errorf("expected name 'get_weather', got %q", result.Content[0].Name)
-	}
-	if result.StopReason != "tool_use" {
-		t.Errorf("expected stop_reason 'tool_use', got %q", result.StopReason)
-	}
-}
-
-func TestToMessagesResponse_WithThinking(t *testing.T) {
-	resp := api.ChatResponse{
-		Model: "test-model",
-		Message: api.Message{
-			Role:     "assistant",
-			Content:  "The answer is 42.",
-			Thinking: "Let me think about this...",
-		},
-		Done:       true,
-		DoneReason: "stop",
-	}
-
-	result := ToMessagesResponse("msg_123", resp)
-
-	if len(result.Content) != 2 {
-		t.Fatalf("expected 2 content blocks, got %d", len(result.Content))
-	}
-	if result.Content[0].Type != "thinking" {
-		t.Errorf("expected first block type 'thinking', got %q", result.Content[0].Type)
-	}
-	if result.Content[0].Thinking == nil || *result.Content[0].Thinking != "Let me think about this..." {
-		t.Errorf("unexpected thinking content: %v", result.Content[0].Thinking)
-	}
-	if result.Content[1].Type != "text" {
-		t.Errorf("expected second block type 'text', got %q", result.Content[1].Type)
-	}
-}
-
-func TestMapStopReason(t *testing.T) {
-	tests := []struct {
-		reason       string
-		hasToolCalls bool
-		want         string
-	}{
-		{"stop", false, "end_turn"},
-		{"length", false, "max_tokens"},
-		{"stop", true, "tool_use"},
-		{"other", false, "stop_sequence"},
-		{"", false, ""},
-	}
-
-	for _, tt := range tests {
-		got := mapStopReason(tt.reason, tt.hasToolCalls)
-		if got != tt.want {
-			t.Errorf("mapStopReason(%q, %v) = %q, want %q", tt.reason, tt.hasToolCalls, got, tt.want)
-		}
-	}
-}
-
-func TestNewError(t *testing.T) {
-	tests := []struct {
-		code int
-		want string
-	}{
-		{400, "invalid_request_error"},
-		{401, "authentication_error"},
-		{403, "permission_error"},
-		{404, "not_found_error"},
-		{429, "rate_limit_error"},
-		{500, "api_error"},
-		{503, "overloaded_error"},
-		{529, "overloaded_error"},
-	}
-
-	for _, tt := range tests {
-		result := NewError(tt.code, "test message")
-		if result.Type != "error" {
-			t.Errorf("NewError(%d) type = %q, want 'error'", tt.code, result.Type)
-		}
-		if result.Error.Type != tt.want {
-			t.Errorf("NewError(%d) error.type = %q, want %q", tt.code, result.Error.Type, tt.want)
-		}
-		if result.Error.Message != "test message" {
-			t.Errorf("NewError(%d) message = %q, want 'test message'", tt.code, result.Error.Message)
-		}
-		if result.RequestID == "" {
-			t.Errorf("NewError(%d) request_id should not be empty", tt.code)
-		}
-	}
-}
-
-func TestGenerateMessageID(t *testing.T) {
-	id1 := GenerateMessageID()
-	id2 := GenerateMessageID()
-
-	if id1 == "" {
-		t.Error("GenerateMessageID returned empty string")
-	}
-	if id1 == id2 {
-		t.Error("GenerateMessageID returned duplicate IDs")
-	}
-	if len(id1) < 10 {
-		t.Errorf("GenerateMessageID returned short ID: %q", id1)
-	}
-	if id1[:4] != "msg_" {
-		t.Errorf("GenerateMessageID should start with 'msg_', got %q", id1[:4])
-	}
-}
-
-func TestStreamConverter_Basic(t *testing.T) {
-	conv := NewStreamConverter("msg_123", "test-model")
-
-	// First chunk
-	resp1 := api.ChatResponse{
-		Model: "test-model",
-		Message: api.Message{
-			Role:    "assistant",
-			Content: "Hello",
-		},
-		Metrics: api.Metrics{PromptEvalCount: 10},
-	}
-
-	events1 := conv.Process(resp1)
-	if len(events1) < 3 {
-		t.Fatalf("expected at least 3 events for first chunk, got %d", len(events1))
-	}
-
-	// Should have message_start, content_block_start, content_block_delta
-	if events1[0].Event != "message_start" {
-		t.Errorf("expected first event 'message_start', got %q", events1[0].Event)
-	}
-	if events1[1].Event != "content_block_start" {
-		t.Errorf("expected second event 'content_block_start', got %q", events1[1].Event)
-	}
-	if events1[2].Event != "content_block_delta" {
-		t.Errorf("expected third event 'content_block_delta', got %q", events1[2].Event)
-	}
-
-	// Final chunk
-	resp2 := api.ChatResponse{
-		Model: "test-model",
-		Message: api.Message{
-			Role:    "assistant",
-			Content: " world!",
-		},
-		Done:       true,
-		DoneReason: "stop",
-		Metrics:    api.Metrics{EvalCount: 5},
-	}
-
-	events2 := conv.Process(resp2)
-
-	// Should have content_block_delta, content_block_stop, message_delta, message_stop
-	hasStop := false
-	for _, e := range events2 {
-		if e.Event == "message_stop" {
-			hasStop = true
-		}
-	}
-	if !hasStop {
-		t.Error("expected message_stop event in final chunk")
-	}
-}
-
-func TestStreamConverter_WithToolCalls(t *testing.T) {
-	conv := NewStreamConverter("msg_123", "test-model")
-
-	resp := api.ChatResponse{
-		Model: "test-model",
-		Message: api.Message{
-			Role: "assistant",
-			ToolCalls: []api.ToolCall{
-				{
-					ID: "call_123",
-					Function: api.ToolCallFunction{
-						Name:      "get_weather",
-						Arguments: testArgs(map[string]any{"location": "Paris"}),
-					},
-				},
-			},
-		},
-		Done:       true,
-		DoneReason: "stop",
-		Metrics:    api.Metrics{PromptEvalCount: 10, EvalCount: 5},
-	}
-
-	events := conv.Process(resp)
-
-	hasToolStart := false
-	hasToolDelta := false
-	for _, e := range events {
-		if e.Event == "content_block_start" {
-			if start, ok := e.Data.(ContentBlockStartEvent); ok {
-				if start.ContentBlock.Type == "tool_use" {
-					hasToolStart = true
-				}
-			}
-		}
-		if e.Event == "content_block_delta" {
-			if delta, ok := e.Data.(ContentBlockDeltaEvent); ok {
-				if delta.Delta.Type == "input_json_delta" {
-					hasToolDelta = true
-				}
-			}
-		}
-	}
-
-	if !hasToolStart {
-		t.Error("expected tool_use content_block_start event")
-	}
-	if !hasToolDelta {
-		t.Error("expected input_json_delta event")
-	}
-}
-
-func TestStreamConverter_ToolCallWithUnmarshalableArgs(t *testing.T) {
-	// Test that unmarshalable arguments (like channels) are handled gracefully
-	// and don't cause a panic or corrupt stream
-	conv := NewStreamConverter("msg_123", "test-model")
-
-	// Create a channel which cannot be JSON marshaled
-	unmarshalable := make(chan int)
-	badArgs := api.NewToolCallFunctionArguments()
-	badArgs.Set("channel", unmarshalable)
-
-	resp := api.ChatResponse{
-		Model: "test-model",
-		Message: api.Message{
-			Role: "assistant",
-			ToolCalls: []api.ToolCall{
-				{
-					ID: "call_bad",
-					Function: api.ToolCallFunction{
-						Name:      "bad_function",
-						Arguments: badArgs,
-					},
-				},
-			},
-		},
-		Done:       true,
-		DoneReason: "stop",
-	}
-
-	// Should not panic and should skip the unmarshalable tool call
-	events := conv.Process(resp)
-
-	// Verify no tool_use block was started (since marshal failed before block start)
-	hasToolStart := false
-	for _, e := range events {
-		if e.Event == "content_block_start" {
-			if start, ok := e.Data.(ContentBlockStartEvent); ok {
-				if start.ContentBlock.Type == "tool_use" {
-					hasToolStart = true
-				}
-			}
-		}
-	}
-
-	if hasToolStart {
-		t.Error("expected no tool_use block when arguments cannot be marshaled")
-	}
-}
-
-func TestStreamConverter_MultipleToolCallsWithMixedValidity(t *testing.T) {
-	// Test that valid tool calls still work when mixed with invalid ones
-	conv := NewStreamConverter("msg_123", "test-model")
-
-	unmarshalable := make(chan int)
-	badArgs := api.NewToolCallFunctionArguments()
-	badArgs.Set("channel", unmarshalable)
-
-	resp := api.ChatResponse{
-		Model: "test-model",
-		Message: api.Message{
-			Role: "assistant",
-			ToolCalls: []api.ToolCall{
-				{
-					ID: "call_good",
-					Function: api.ToolCallFunction{
-						Name:      "good_function",
-						Arguments: testArgs(map[string]any{"location": "Paris"}),
-					},
-				},
-				{
-					ID: "call_bad",
-					Function: api.ToolCallFunction{
-						Name:      "bad_function",
-						Arguments: badArgs,
-					},
-				},
-			},
-		},
-		Done:       true,
-		DoneReason: "stop",
-	}
-
-	events := conv.Process(resp)
-
-	// Count tool_use blocks - should only have 1 (the valid one)
-	toolStartCount := 0
-	toolDeltaCount := 0
-	for _, e := range events {
-		if e.Event == "content_block_start" {
-			if start, ok := e.Data.(ContentBlockStartEvent); ok {
-				if start.ContentBlock.Type == "tool_use" {
-					toolStartCount++
-					if start.ContentBlock.Name != "good_function" {
-						t.Errorf("expected tool name 'good_function', got %q", start.ContentBlock.Name)
-					}
-				}
-			}
-		}
-		if e.Event == "content_block_delta" {
-			if delta, ok := e.Data.(ContentBlockDeltaEvent); ok {
-				if delta.Delta.Type == "input_json_delta" {
-					toolDeltaCount++
-				}
-			}
-		}
-	}
-
-	if toolStartCount != 1 {
-		t.Errorf("expected 1 tool_use block, got %d", toolStartCount)
-	}
-	if toolDeltaCount != 1 {
-		t.Errorf("expected 1 input_json_delta, got %d", toolDeltaCount)
-	}
-}
-
-// TestContentBlockJSON_EmptyFieldsPresent verifies that empty text and thinking fields
-// are serialized in JSON output. The Anthropic SDK requires these fields to be present
-// (even when empty) in content_block_start events to properly accumulate streaming deltas.
-// Without these fields, the SDK throws: "TypeError: unsupported operand type(s) for +=: 'NoneType' and 'str'"
-func TestContentBlockJSON_EmptyFieldsPresent(t *testing.T) {
-	tests := []struct {
-		name     string
-		block    ContentBlock
-		wantKeys []string
-	}{
-		{
-			name: "text block includes empty text field",
-			block: ContentBlock{
-				Type: "text",
-				Text: ptr(""),
-			},
-			wantKeys: []string{"type", "text"},
-		},
-		{
-			name: "thinking block includes empty thinking field",
-			block: ContentBlock{
-				Type:     "thinking",
-				Thinking: ptr(""),
-			},
-			wantKeys: []string{"type", "thinking"},
-		},
-		{
-			name: "text block with content",
-			block: ContentBlock{
-				Type: "text",
-				Text: ptr("hello"),
-			},
-			wantKeys: []string{"type", "text"},
-		},
-	}
-
-	for _, tt := range tests {
-		t.Run(tt.name, func(t *testing.T) {
-			data, err := json.Marshal(tt.block)
-			if err != nil {
-				t.Fatalf("failed to marshal: %v", err)
-			}
-
-			var result map[string]any
-			if err := json.Unmarshal(data, &result); err != nil {
-				t.Fatalf("failed to unmarshal: %v", err)
-			}
-
-			for _, key := range tt.wantKeys {
-				if _, ok := result[key]; !ok {
-					t.Errorf("expected key %q to be present in JSON output, got: %s", key, string(data))
-				}
-			}
-		})
-	}
-}
-
-// TestStreamConverter_ContentBlockStartIncludesEmptyFields verifies that content_block_start
-// events include the required empty fields for SDK compatibility.
-func TestStreamConverter_ContentBlockStartIncludesEmptyFields(t *testing.T) {
-	t.Run("text block start includes empty text", func(t *testing.T) {
-		conv := NewStreamConverter("msg_123", "test-model")
-
-		resp := api.ChatResponse{
-			Model:   "test-model",
-			Message: api.Message{Role: "assistant", Content: "hello"},
-		}
-
-		events := conv.Process(resp)
-
-		var foundTextStart bool
-		for _, e := range events {
-			if e.Event == "content_block_start" {
-				if start, ok := e.Data.(ContentBlockStartEvent); ok {
-					if start.ContentBlock.Type == "text" {
-						foundTextStart = true
-						// Marshal and verify the text field is present
-						data, _ := json.Marshal(start)
-						var result map[string]any
-						json.Unmarshal(data, &result)
-						cb := result["content_block"].(map[string]any)
-						if _, ok := cb["text"]; !ok {
-							t.Error("content_block_start for text should include 'text' field")
-						}
-					}
-				}
-			}
-		}
-
-		if !foundTextStart {
-			t.Error("expected text content_block_start event")
-		}
-	})
-
-	t.Run("thinking block start includes empty thinking", func(t *testing.T) {
-		conv := NewStreamConverter("msg_123", "test-model")
-
-		resp := api.ChatResponse{
-			Model:   "test-model",
-			Message: api.Message{Role: "assistant", Thinking: "let me think..."},
-		}
-
-		events := conv.Process(resp)
-
-		var foundThinkingStart bool
-		for _, e := range events {
-			if e.Event == "content_block_start" {
-				if start, ok := e.Data.(ContentBlockStartEvent); ok {
-					if start.ContentBlock.Type == "thinking" {
-						foundThinkingStart = true
-						data, _ := json.Marshal(start)
-						var result map[string]any
-						json.Unmarshal(data, &result)
-						cb := result["content_block"].(map[string]any)
-						if _, ok := cb["thinking"]; !ok {
-							t.Error("content_block_start for thinking should include 'thinking' field")
-						}
-					}
-				}
-			}
-		}
-
-		if !foundThinkingStart {
-			t.Error("expected thinking content_block_start event")
-		}
-	})
-}

cmd/cmd.go

@@ -520,6 +520,7 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 
 	// Check for experimental flag
 	isExperimental, _ := cmd.Flags().GetBool("experimental")
+	yoloMode, _ := cmd.Flags().GetBool("yolo")
 
 	if interactive {
 		if err := loadOrUnloadModel(cmd, &opts); err != nil {
@@ -547,9 +548,9 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 			}
 		}
 
-		// Use experimental agent loop with
+		// Use experimental agent loop with tools
 		if isExperimental {
-			return xcmd.GenerateInteractive(cmd, opts.Model, opts.WordWrap, opts.Options, opts.Think, opts.HideThinking, opts.KeepAlive)
+			return xcmd.GenerateInteractive(cmd, opts.Model, opts.WordWrap, opts.Options, opts.Think, opts.HideThinking, opts.KeepAlive, yoloMode)
 		}
 
 		return generateInteractive(cmd, opts)
@@ -1764,6 +1765,7 @@ func NewCLI() *cobra.Command {
 	runCmd.Flags().Bool("truncate", false, "For embedding models: truncate inputs exceeding context length (default: true). Set --truncate=false to error instead")
 	runCmd.Flags().Int("dimensions", 0, "Truncate output embeddings to specified dimension (embedding models only)")
 	runCmd.Flags().Bool("experimental", false, "Enable experimental agent loop with tools")
+	runCmd.Flags().BoolP("yolo", "y", false, "Skip all tool approval prompts (use with caution)")
 
 	stopCmd := &cobra.Command{
 		Use:     "stop MODEL",

docs/README.md

@@ -14,7 +14,6 @@
 * [API Reference](https://docs.ollama.com/api)
 * [Modelfile Reference](https://docs.ollama.com/modelfile)
 * [OpenAI Compatibility](https://docs.ollama.com/api/openai-compatibility)
-* [Anthropic Compatibility](./api/anthropic-compatibility.mdx)
 
 ### Resources
 

docs/api/anthropic-compatibility.mdx

@@ -1,406 +0,0 @@
----
-title: Anthropic compatibility
----
-
-Ollama provides compatibility with the [Anthropic Messages API](https://docs.anthropic.com/en/api/messages) to help connect existing applications to Ollama, including tools like Claude Code.
-
-## Recommended models
-
-For coding use cases, models like `glm-4.7:cloud`, `minimax-m2.1:cloud`, and `qwen3-coder` are recommended.
-
-Pull a model before use:
-```shell
-ollama pull qwen3-coder
-ollama pull glm-4.7:cloud
-```
-
-## Usage
-
-### Environment variables
-
-To use Ollama with tools that expect the Anthropic API (like Claude Code), set these environment variables:
-
-```shell
-export ANTHROPIC_BASE_URL=http://localhost:11434
-export ANTHROPIC_API_KEY=ollama  # required but ignored
-```
-
-### Simple `/v1/messages` example
-
-<CodeGroup dropdown>
-
-```python basic.py
-import anthropic
-
-client = anthropic.Anthropic(
-    base_url='http://localhost:11434',
-    api_key='ollama',  # required but ignored
-)
-
-message = client.messages.create(
-    model='qwen3-coder',
-    max_tokens=1024,
-    messages=[
-        {'role': 'user', 'content': 'Hello, how are you?'}
-    ]
-)
-print(message.content[0].text)
-```
-
-```javascript basic.js
-import Anthropic from "@anthropic-ai/sdk";
-
-const anthropic = new Anthropic({
-  baseURL: "http://localhost:11434",
-  apiKey: "ollama", // required but ignored
-});
-
-const message = await anthropic.messages.create({
-  model: "qwen3-coder",
-  max_tokens: 1024,
-  messages: [{ role: "user", content: "Hello, how are you?" }],
-});
-
-console.log(message.content[0].text);
-```
-
-```shell basic.sh
-curl -X POST http://localhost:11434/v1/messages \
--H "Content-Type: application/json" \
--H "x-api-key: ollama" \
--H "anthropic-version: 2023-06-01" \
--d '{
-  "model": "qwen3-coder",
-  "max_tokens": 1024,
-  "messages": [{ "role": "user", "content": "Hello, how are you?" }]
-}'
-```
-
-</CodeGroup>
-
-### Streaming example
-
-<CodeGroup dropdown>
-
-```python streaming.py
-import anthropic
-
-client = anthropic.Anthropic(
-    base_url='http://localhost:11434',
-    api_key='ollama',
-)
-
-with client.messages.stream(
-    model='qwen3-coder',
-    max_tokens=1024,
-    messages=[{'role': 'user', 'content': 'Count from 1 to 10'}]
-) as stream:
-    for text in stream.text_stream:
-        print(text, end='', flush=True)
-```
-
-```javascript streaming.js
-import Anthropic from "@anthropic-ai/sdk";
-
-const anthropic = new Anthropic({
-  baseURL: "http://localhost:11434",
-  apiKey: "ollama",
-});
-
-const stream = await anthropic.messages.stream({
-  model: "qwen3-coder",
-  max_tokens: 1024,
-  messages: [{ role: "user", content: "Count from 1 to 10" }],
-});
-
-for await (const event of stream) {
-  if (
-    event.type === "content_block_delta" &&
-    event.delta.type === "text_delta"
-  ) {
-    process.stdout.write(event.delta.text);
-  }
-}
-```
-
-```shell streaming.sh
-curl -X POST http://localhost:11434/v1/messages \
--H "Content-Type: application/json" \
--d '{
-  "model": "qwen3-coder",
-  "max_tokens": 1024,
-  "stream": true,
-  "messages": [{ "role": "user", "content": "Count from 1 to 10" }]
-}'
-```
-
-</CodeGroup>
-
-### Tool calling example
-
-<CodeGroup dropdown>
-
-```python tools.py
-import anthropic
-
-client = anthropic.Anthropic(
-    base_url='http://localhost:11434',
-    api_key='ollama',
-)
-
-message = client.messages.create(
-    model='qwen3-coder',
-    max_tokens=1024,
-    tools=[
-        {
-            'name': 'get_weather',
-            'description': 'Get the current weather in a location',
-            'input_schema': {
-                'type': 'object',
-                'properties': {
-                    'location': {
-                        'type': 'string',
-                        'description': 'The city and state, e.g. San Francisco, CA'
-                    }
-                },
-                'required': ['location']
-            }
-        }
-    ],
-    messages=[{'role': 'user', 'content': "What's the weather in San Francisco?"}]
-)
-
-for block in message.content:
-    if block.type == 'tool_use':
-        print(f'Tool: {block.name}')
-        print(f'Input: {block.input}')
-```
-
-```javascript tools.js
-import Anthropic from "@anthropic-ai/sdk";
-
-const anthropic = new Anthropic({
-  baseURL: "http://localhost:11434",
-  apiKey: "ollama",
-});
-
-const message = await anthropic.messages.create({
-  model: "qwen3-coder",
-  max_tokens: 1024,
-  tools: [
-    {
-      name: "get_weather",
-      description: "Get the current weather in a location",
-      input_schema: {
-        type: "object",
-        properties: {
-          location: {
-            type: "string",
-            description: "The city and state, e.g. San Francisco, CA",
-          },
-        },
-        required: ["location"],
-      },
-    },
-  ],
-  messages: [{ role: "user", content: "What's the weather in San Francisco?" }],
-});
-
-for (const block of message.content) {
-  if (block.type === "tool_use") {
-    console.log("Tool:", block.name);
-    console.log("Input:", block.input);
-  }
-}
-```
-
-```shell tools.sh
-curl -X POST http://localhost:11434/v1/messages \
--H "Content-Type: application/json" \
--d '{
-  "model": "qwen3-coder",
-  "max_tokens": 1024,
-  "tools": [
-    {
-      "name": "get_weather",
-      "description": "Get the current weather in a location",
-      "input_schema": {
-        "type": "object",
-        "properties": {
-          "location": {
-            "type": "string",
-            "description": "The city and state"
-          }
-        },
-        "required": ["location"]
-      }
-    }
-  ],
-  "messages": [{ "role": "user", "content": "What is the weather in San Francisco?" }]
-}'
-```
-
-</CodeGroup>
-
-## Using with Claude Code
-
-[Claude Code](https://code.claude.com/docs/en/overview) can be configured to use Ollama as its backend:
-
-```shell
-ANTHROPIC_BASE_URL=http://localhost:11434 ANTHROPIC_API_KEY=ollama claude --model qwen3-coder
-```
-
-Or set the environment variables in your shell profile:
-
-```shell
-export ANTHROPIC_BASE_URL=http://localhost:11434
-export ANTHROPIC_API_KEY=ollama
-```
-
-Then run Claude Code with any Ollama model:
-
-```shell
-# Local models
-claude --model qwen3-coder
-claude --model gpt-oss:20b
-
-# Cloud models
-claude --model glm-4.7:cloud
-claude --model minimax-m2.1:cloud
-```
-
-## Endpoints
-
-### `/v1/messages`
-
-#### Supported features
-
-- [x] Messages
-- [x] Streaming
-- [x] System prompts
-- [x] Multi-turn conversations
-- [x] Vision (images)
-- [x] Tools (function calling)
-- [x] Tool results
-- [x] Thinking/extended thinking
-
-#### Supported request fields
-
-- [x] `model`
-- [x] `max_tokens`
-- [x] `messages`
-  - [x] Text `content`
-  - [x] Image `content` (base64)
-  - [x] Array of content blocks
-  - [x] `tool_use` blocks
-  - [x] `tool_result` blocks
-  - [x] `thinking` blocks
-- [x] `system` (string or array)
-- [x] `stream`
-- [x] `temperature`
-- [x] `top_p`
-- [x] `top_k`
-- [x] `stop_sequences`
-- [x] `tools`
-- [x] `thinking`
-- [ ] `tool_choice`
-- [ ] `metadata`
-
-#### Supported response fields
-
-- [x] `id`
-- [x] `type`
-- [x] `role`
-- [x] `model`
-- [x] `content` (text, tool_use, thinking blocks)
-- [x] `stop_reason` (end_turn, max_tokens, tool_use)
-- [x] `usage` (input_tokens, output_tokens)
-
-#### Streaming events
-
-- [x] `message_start`
-- [x] `content_block_start`
-- [x] `content_block_delta` (text_delta, input_json_delta, thinking_delta)
-- [x] `content_block_stop`
-- [x] `message_delta`
-- [x] `message_stop`
-- [x] `ping`
-- [x] `error`
-
-## Models
-
-Ollama supports both local and cloud models.
-
-### Local models
-
-Pull a local model before use:
-
-```shell
-ollama pull qwen3-coder
-```
-
-Recommended local models:
-- `qwen3-coder` - Excellent for coding tasks
-- `gpt-oss:20b` - Strong general-purpose model
-
-### Cloud models
-
-Cloud models are available immediately without pulling:
-
-- `glm-4.7:cloud` - High-performance cloud model
-- `minimax-m2.1:cloud` - Fast cloud model
-
-### Default model names
-
-For tooling that relies on default Anthropic model names such as `claude-3-5-sonnet`, use `ollama cp` to copy an existing model name:
-
-```shell
-ollama cp qwen3-coder claude-3-5-sonnet
-```
-
-Afterwards, this new model name can be specified in the `model` field:
-
-```shell
-curl http://localhost:11434/v1/messages \
-    -H "Content-Type: application/json" \
-    -d '{
-        "model": "claude-3-5-sonnet",
-        "max_tokens": 1024,
-        "messages": [
-            {
-                "role": "user",
-                "content": "Hello!"
-            }
-        ]
-    }'
-```
-
-## Differences from the Anthropic API
-
-### Behavior differences
-
-- API key is accepted but not validated
-- `anthropic-version` header is accepted but not used
-- Token counts are approximations based on the underlying model's tokenizer
-
-### Not supported
-
-The following Anthropic API features are not currently supported:
-
-| Feature | Description |
-|---------|-------------|
-| `/v1/messages/count_tokens` | Token counting endpoint |
-| `tool_choice` | Forcing specific tool use or disabling tools |
-| `metadata` | Request metadata (user_id) |
-| Prompt caching | `cache_control` blocks for caching prefixes |
-| Batches API | `/v1/messages/batches` for async batch processing |
-| Citations | `citations` content blocks |
-| PDF support | `document` content blocks with PDF files |
-| Server-sent errors | `error` events during streaming (errors return HTTP status) |
-
-### Partial support
-
-| Feature | Status |
-|---------|--------|
-| Image content | Base64 images supported; URL images not supported |
-| Extended thinking | Basic support; `budget_tokens` accepted but not enforced |

docs/docs.json

@@ -32,9 +32,7 @@
     "codeblocks": "system"
   },
   "contextual": {
-    "options": [
-      "copy"
-    ]
+    "options": ["copy"]
   },
   "navbar": {
     "links": [
@@ -54,9 +52,7 @@
       "display": "simple"
     },
     "examples": {
-      "languages": [
-        "curl"
-      ]
+      "languages": ["curl"]
     }
   },
   "redirects": [
@@ -101,7 +97,6 @@
           {
             "group": "Integrations",
             "pages": [
-              "/integrations/claude-code",
               "/integrations/vscode",
               "/integrations/jetbrains",
               "/integrations/codex",
@@ -144,8 +139,7 @@
               "/api/streaming",
               "/api/usage",
               "/api/errors",
-              "/api/openai-compatibility",
-              "/api/anthropic-compatibility"
+              "/api/openai-compatibility"
             ]
           },
           {

docs/integrations/claude-code.mdx

@@ -1,69 +0,0 @@
----
-title: Claude Code
----
-
-## Install
-
-Install [Claude Code](https://code.claude.com/docs/en/overview):
-
-<CodeGroup>
-
-```shell macOS / Linux
-curl -fsSL https://claude.ai/install.sh | bash
-```
-
-```powershell Windows
-irm https://claude.ai/install.ps1 | iex
-```
-
-</CodeGroup>
-
-## Usage with Ollama
-
-Claude Code connects to Ollama using the Anthropic-compatible API.
-
-1. Set the environment variables:
-
-```shell
-export ANTHROPIC_BASE_URL=http://localhost:11434
-export ANTHROPIC_API_KEY=ollama
-```
-
-2. Run Claude Code with an Ollama model:
-
-```shell
-claude --model qwen3-coder
-```
-
-Or run with environment variables inline:
-
-```shell
-ANTHROPIC_BASE_URL=http://localhost:11434 ANTHROPIC_API_KEY=ollama claude --model qwen3-coder
-```
-
-## Connecting to ollama.com
-
-1. Create an [API key](https://ollama.com/settings/keys) on ollama.com
-2. Set the environment variables:
-
-```shell
-export ANTHROPIC_BASE_URL=https://ollama.com
-export ANTHROPIC_API_KEY=<your-api-key>
-```
-
-3. Run Claude Code with a cloud model:
-
-```shell
-claude --model glm-4.7:cloud
-```
-
-## Recommended Models
-
-### Cloud models
-- `glm-4.7:cloud` - High-performance cloud model
-- `minimax-m2.1:cloud` - Fast cloud model
-- `qwen3-coder:480b` - Large coding model
-
-### Local models
-- `qwen3-coder` - Excellent for coding tasks
-- `gpt-oss:20b` - Strong general-purpose model

middleware/anthropic.go

@@ -1,149 +0,0 @@
-package middleware
-
-import (
-	"bytes"
-	"encoding/json"
-	"fmt"
-	"io"
-	"net/http"
-
-	"github.com/gin-gonic/gin"
-
-	"github.com/ollama/ollama/anthropic"
-	"github.com/ollama/ollama/api"
-)
-
-// AnthropicWriter wraps the response writer to transform Ollama responses to Anthropic format
-type AnthropicWriter struct {
-	BaseWriter
-	stream    bool
-	id        string
-	model     string
-	converter *anthropic.StreamConverter
-}
-
-func (w *AnthropicWriter) writeError(data []byte) (int, error) {
-	var errData struct {
-		Error string `json:"error"`
-	}
-	if err := json.Unmarshal(data, &errData); err != nil {
-		return 0, err
-	}
-
-	w.ResponseWriter.Header().Set("Content-Type", "application/json")
-	err := json.NewEncoder(w.ResponseWriter).Encode(anthropic.NewError(w.ResponseWriter.Status(), errData.Error))
-	if err != nil {
-		return 0, err
-	}
-
-	return len(data), nil
-}
-
-func (w *AnthropicWriter) writeEvent(eventType string, data any) error {
-	d, err := json.Marshal(data)
-	if err != nil {
-		return err
-	}
-	_, err = w.ResponseWriter.Write([]byte(fmt.Sprintf("event: %s\ndata: %s\n\n", eventType, d)))
-	if err != nil {
-		return err
-	}
-	if f, ok := w.ResponseWriter.(http.Flusher); ok {
-		f.Flush()
-	}
-	return nil
-}
-
-func (w *AnthropicWriter) writeResponse(data []byte) (int, error) {
-	var chatResponse api.ChatResponse
-	err := json.Unmarshal(data, &chatResponse)
-	if err != nil {
-		return 0, err
-	}
-
-	if w.stream {
-		w.ResponseWriter.Header().Set("Content-Type", "text/event-stream")
-
-		events := w.converter.Process(chatResponse)
-		for _, event := range events {
-			if err := w.writeEvent(event.Event, event.Data); err != nil {
-				return 0, err
-			}
-		}
-		return len(data), nil
-	}
-
-	w.ResponseWriter.Header().Set("Content-Type", "application/json")
-	response := anthropic.ToMessagesResponse(w.id, chatResponse)
-	return len(data), json.NewEncoder(w.ResponseWriter).Encode(response)
-}
-
-func (w *AnthropicWriter) Write(data []byte) (int, error) {
-	code := w.ResponseWriter.Status()
-	if code != http.StatusOK {
-		return w.writeError(data)
-	}
-
-	return w.writeResponse(data)
-}
-
-// AnthropicMessagesMiddleware handles Anthropic Messages API requests
-func AnthropicMessagesMiddleware() gin.HandlerFunc {
-	return func(c *gin.Context) {
-		var req anthropic.MessagesRequest
-		err := c.ShouldBindJSON(&req)
-		if err != nil {
-			c.AbortWithStatusJSON(http.StatusBadRequest, anthropic.NewError(http.StatusBadRequest, err.Error()))
-			return
-		}
-
-		if req.Model == "" {
-			c.AbortWithStatusJSON(http.StatusBadRequest, anthropic.NewError(http.StatusBadRequest, "model is required"))
-			return
-		}
-
-		if req.MaxTokens <= 0 {
-			c.AbortWithStatusJSON(http.StatusBadRequest, anthropic.NewError(http.StatusBadRequest, "max_tokens is required and must be positive"))
-			return
-		}
-
-		if len(req.Messages) == 0 {
-			c.AbortWithStatusJSON(http.StatusBadRequest, anthropic.NewError(http.StatusBadRequest, "messages is required"))
-			return
-		}
-
-		chatReq, err := anthropic.FromMessagesRequest(req)
-		if err != nil {
-			c.AbortWithStatusJSON(http.StatusBadRequest, anthropic.NewError(http.StatusBadRequest, err.Error()))
-			return
-		}
-
-		var b bytes.Buffer
-		if err := json.NewEncoder(&b).Encode(chatReq); err != nil {
-			c.AbortWithStatusJSON(http.StatusInternalServerError, anthropic.NewError(http.StatusInternalServerError, err.Error()))
-			return
-		}
-
-		c.Request.Body = io.NopCloser(&b)
-
-		messageID := anthropic.GenerateMessageID()
-
-		w := &AnthropicWriter{
-			BaseWriter: BaseWriter{ResponseWriter: c.Writer},
-			stream:     req.Stream,
-			id:         messageID,
-			model:      req.Model,
-			converter:  anthropic.NewStreamConverter(messageID, req.Model),
-		}
-
-		if req.Stream {
-			c.Writer.Header().Set("Content-Type", "text/event-stream")
-			c.Writer.Header().Set("Cache-Control", "no-cache")
-			c.Writer.Header().Set("Connection", "keep-alive")
-		}
-
-		c.Writer = w
-
-		c.Next()
-	}
-}

middleware/anthropic_test.go

@@ -1,584 +0,0 @@
-package middleware
-
-import (
-	"bytes"
-	"encoding/json"
-	"io"
-	"net/http"
-	"net/http/httptest"
-	"strings"
-	"testing"
-
-	"github.com/gin-gonic/gin"
-	"github.com/google/go-cmp/cmp"
-	"github.com/google/go-cmp/cmp/cmpopts"
-
-	"github.com/ollama/ollama/anthropic"
-	"github.com/ollama/ollama/api"
-)
-
-func captureAnthropicRequest(capturedRequest any) gin.HandlerFunc {
-	return func(c *gin.Context) {
-		bodyBytes, _ := io.ReadAll(c.Request.Body)
-		c.Request.Body = io.NopCloser(bytes.NewReader(bodyBytes))
-		_ = json.Unmarshal(bodyBytes, capturedRequest)
-		c.Next()
-	}
-}
-
-// testProps creates ToolPropertiesMap from a map (convenience function for tests)
-func testProps(m map[string]api.ToolProperty) *api.ToolPropertiesMap {
-	props := api.NewToolPropertiesMap()
-	for k, v := range m {
-		props.Set(k, v)
-	}
-	return props
-}
-
-func TestAnthropicMessagesMiddleware(t *testing.T) {
-	type testCase struct {
-		name string
-		body string
-		req  api.ChatRequest
-		err  anthropic.ErrorResponse
-	}
-
-	var capturedRequest *api.ChatRequest
-	stream := true
-
-	testCases := []testCase{
-		{
-			name: "basic message",
-			body: `{
-				"model": "test-model",
-				"max_tokens": 1024,
-				"messages": [
-					{"role": "user", "content": "Hello"}
-				]
-			}`,
-			req: api.ChatRequest{
-				Model: "test-model",
-				Messages: []api.Message{
-					{Role: "user", Content: "Hello"},
-				},
-				Options: map[string]any{"num_predict": 1024},
-				Stream:  &False,
-			},
-		},
-		{
-			name: "with system prompt",
-			body: `{
-				"model": "test-model",
-				"max_tokens": 1024,
-				"system": "You are helpful.",
-				"messages": [
-					{"role": "user", "content": "Hello"}
-				]
-			}`,
-			req: api.ChatRequest{
-				Model: "test-model",
-				Messages: []api.Message{
-					{Role: "system", Content: "You are helpful."},
-					{Role: "user", Content: "Hello"},
-				},
-				Options: map[string]any{"num_predict": 1024},
-				Stream:  &False,
-			},
-		},
-		{
-			name: "with options",
-			body: `{
-				"model": "test-model",
-				"max_tokens": 2048,
-				"temperature": 0.7,
-				"top_p": 0.9,
-				"top_k": 40,
-				"stop_sequences": ["\n", "END"],
-				"messages": [
-					{"role": "user", "content": "Hello"}
-				]
-			}`,
-			req: api.ChatRequest{
-				Model: "test-model",
-				Messages: []api.Message{
-					{Role: "user", Content: "Hello"},
-				},
-				Options: map[string]any{
-					"num_predict": 2048,
-					"temperature": 0.7,
-					"top_p":       0.9,
-					"top_k":       40,
-					"stop":        []string{"\n", "END"},
-				},
-				Stream: &False,
-			},
-		},
-		{
-			name: "streaming",
-			body: `{
-				"model": "test-model",
-				"max_tokens": 1024,
-				"stream": true,
-				"messages": [
-					{"role": "user", "content": "Hello"}
-				]
-			}`,
-			req: api.ChatRequest{
-				Model: "test-model",
-				Messages: []api.Message{
-					{Role: "user", Content: "Hello"},
-				},
-				Options: map[string]any{"num_predict": 1024},
-				Stream:  &stream,
-			},
-		},
-		{
-			name: "with tools",
-			body: `{
-				"model": "test-model",
-				"max_tokens": 1024,
-				"messages": [
-					{"role": "user", "content": "What's the weather?"}
-				],
-				"tools": [{
-					"name": "get_weather",
-					"description": "Get current weather",
-					"input_schema": {
-						"type": "object",
-						"properties": {
-							"location": {"type": "string"}
-						},
-						"required": ["location"]
-					}
-				}]
-			}`,
-			req: api.ChatRequest{
-				Model: "test-model",
-				Messages: []api.Message{
-					{Role: "user", Content: "What's the weather?"},
-				},
-				Tools: []api.Tool{
-					{
-						Type: "function",
-						Function: api.ToolFunction{
-							Name:        "get_weather",
-							Description: "Get current weather",
-							Parameters: api.ToolFunctionParameters{
-								Type:     "object",
-								Required: []string{"location"},
-								Properties: testProps(map[string]api.ToolProperty{
-									"location": {Type: api.PropertyType{"string"}},
-								}),
-							},
-						},
-					},
-				},
-				Options: map[string]any{"num_predict": 1024},
-				Stream:  &False,
-			},
-		},
-		{
-			name: "with tool result",
-			body: `{
-				"model": "test-model",
-				"max_tokens": 1024,
-				"messages": [
-					{"role": "user", "content": "What's the weather?"},
-					{"role": "assistant", "content": [
-						{"type": "tool_use", "id": "call_123", "name": "get_weather", "input": {"location": "Paris"}}
-					]},
-					{"role": "user", "content": [
-						{"type": "tool_result", "tool_use_id": "call_123", "content": "Sunny, 22°C"}
-					]}
-				]
-			}`,
-			req: api.ChatRequest{
-				Model: "test-model",
-				Messages: []api.Message{
-					{Role: "user", Content: "What's the weather?"},
-					{
-						Role: "assistant",
-						ToolCalls: []api.ToolCall{
-							{
-								ID: "call_123",
-								Function: api.ToolCallFunction{
-									Name:      "get_weather",
-									Arguments: testArgs(map[string]any{"location": "Paris"}),
-								},
-							},
-						},
-					},
-					{Role: "tool", Content: "Sunny, 22°C", ToolCallID: "call_123"},
-				},
-				Options: map[string]any{"num_predict": 1024},
-				Stream:  &False,
-			},
-		},
-		{
-			name: "with thinking enabled",
-			body: `{
-				"model": "test-model",
-				"max_tokens": 1024,
-				"thinking": {"type": "enabled", "budget_tokens": 1000},
-				"messages": [
-					{"role": "user", "content": "Hello"}
-				]
-			}`,
-			req: api.ChatRequest{
-				Model: "test-model",
-				Messages: []api.Message{
-					{Role: "user", Content: "Hello"},
-				},
-				Options: map[string]any{"num_predict": 1024},
-				Stream:  &False,
-				Think:   &api.ThinkValue{Value: true},
-			},
-		},
-		{
-			name: "missing model error",
-			body: `{
-				"max_tokens": 1024,
-				"messages": [
-					{"role": "user", "content": "Hello"}
-				]
-			}`,
-			err: anthropic.ErrorResponse{
-				Type: "error",
-				Error: anthropic.Error{
-					Type:    "invalid_request_error",
-					Message: "model is required",
-				},
-			},
-		},
-		{
-			name: "missing max_tokens error",
-			body: `{
-				"model": "test-model",
-				"messages": [
-					{"role": "user", "content": "Hello"}
-				]
-			}`,
-			err: anthropic.ErrorResponse{
-				Type: "error",
-				Error: anthropic.Error{
-					Type:    "invalid_request_error",
-					Message: "max_tokens is required and must be positive",
-				},
-			},
-		},
-		{
-			name: "missing messages error",
-			body: `{
-				"model": "test-model",
-				"max_tokens": 1024
-			}`,
-			err: anthropic.ErrorResponse{
-				Type: "error",
-				Error: anthropic.Error{
-					Type:    "invalid_request_error",
-					Message: "messages is required",
-				},
-			},
-		},
-		{
-			name: "tool_use missing id error",
-			body: `{
-				"model": "test-model",
-				"max_tokens": 1024,
-				"messages": [
-					{"role": "assistant", "content": [
-						{"type": "tool_use", "name": "test"}
-					]}
-				]
-			}`,
-			err: anthropic.ErrorResponse{
-				Type: "error",
-				Error: anthropic.Error{
-					Type:    "invalid_request_error",
-					Message: "tool_use block missing required 'id' field",
-				},
-			},
-		},
-	}
-
-	endpoint := func(c *gin.Context) {
-		c.Status(http.StatusOK)
-	}
-
-	gin.SetMode(gin.TestMode)
-	router := gin.New()
-	router.Use(AnthropicMessagesMiddleware(), captureAnthropicRequest(&capturedRequest))
-	router.Handle(http.MethodPost, "/v1/messages", endpoint)
-
-	for _, tc := range testCases {
-		t.Run(tc.name, func(t *testing.T) {
-			req, _ := http.NewRequest(http.MethodPost, "/v1/messages", strings.NewReader(tc.body))
-			req.Header.Set("Content-Type", "application/json")
-
-			defer func() { capturedRequest = nil }()
-
-			resp := httptest.NewRecorder()
-			router.ServeHTTP(resp, req)
-
-			if tc.err.Type != "" {
-				// Expect error
-				if resp.Code == http.StatusOK {
-					t.Fatalf("expected error response, got 200 OK")
-				}
-				var errResp anthropic.ErrorResponse
-				if err := json.Unmarshal(resp.Body.Bytes(), &errResp); err != nil {
-					t.Fatalf("failed to unmarshal error: %v", err)
-				}
-				if errResp.Type != tc.err.Type {
-					t.Errorf("expected error type %q, got %q", tc.err.Type, errResp.Type)
-				}
-				if errResp.Error.Type != tc.err.Error.Type {
-					t.Errorf("expected error.type %q, got %q", tc.err.Error.Type, errResp.Error.Type)
-				}
-				if errResp.Error.Message != tc.err.Error.Message {
-					t.Errorf("expected error.message %q, got %q", tc.err.Error.Message, errResp.Error.Message)
-				}
-				return
-			}
-
-			if resp.Code != http.StatusOK {
-				t.Fatalf("unexpected status code: %d, body: %s", resp.Code, resp.Body.String())
-			}
-
-			if capturedRequest == nil {
-				t.Fatal("request was not captured")
-			}
-
-			// Compare relevant fields
-			if capturedRequest.Model != tc.req.Model {
-				t.Errorf("model mismatch: got %q, want %q", capturedRequest.Model, tc.req.Model)
-			}
-
-			if diff := cmp.Diff(tc.req.Messages, capturedRequest.Messages,
-				cmpopts.IgnoreUnexported(api.ToolCallFunctionArguments{}, api.ToolPropertiesMap{})); diff != "" {
-				t.Errorf("messages mismatch (-want +got):\n%s", diff)
-			}
-
-			if tc.req.Stream != nil && capturedRequest.Stream != nil {
-				if *tc.req.Stream != *capturedRequest.Stream {
-					t.Errorf("stream mismatch: got %v, want %v", *capturedRequest.Stream, *tc.req.Stream)
-				}
-			}
-
-			if tc.req.Think != nil {
-				if capturedRequest.Think == nil {
-					t.Error("expected Think to be set")
-				} else if capturedRequest.Think.Value != tc.req.Think.Value {
-					t.Errorf("Think mismatch: got %v, want %v", capturedRequest.Think.Value, tc.req.Think.Value)
-				}
-			}
-		})
-	}
-}
-
-func TestAnthropicMessagesMiddleware_Headers(t *testing.T) {
-	gin.SetMode(gin.TestMode)
-
-	t.Run("streaming sets correct headers", func(t *testing.T) {
-		router := gin.New()
-		router.Use(AnthropicMessagesMiddleware())
-		router.POST("/v1/messages", func(c *gin.Context) {
-			// Check headers were set
-			if c.Writer.Header().Get("Content-Type") != "text/event-stream" {
-				t.Errorf("expected Content-Type text/event-stream, got %q", c.Writer.Header().Get("Content-Type"))
-			}
-			if c.Writer.Header().Get("Cache-Control") != "no-cache" {
-				t.Errorf("expected Cache-Control no-cache, got %q", c.Writer.Header().Get("Cache-Control"))
-			}
-			c.Status(http.StatusOK)
-		})
-
-		body := `{"model": "test", "max_tokens": 100, "stream": true, "messages": [{"role": "user", "content": "Hi"}]}`
-		req, _ := http.NewRequest(http.MethodPost, "/v1/messages", strings.NewReader(body))
-		req.Header.Set("Content-Type", "application/json")
-
-		resp := httptest.NewRecorder()
-		router.ServeHTTP(resp, req)
-	})
-}
-
-func TestAnthropicMessagesMiddleware_InvalidJSON(t *testing.T) {
-	gin.SetMode(gin.TestMode)
-	router := gin.New()
-	router.Use(AnthropicMessagesMiddleware())
-	router.POST("/v1/messages", func(c *gin.Context) {
-		c.Status(http.StatusOK)
-	})
-
-	req, _ := http.NewRequest(http.MethodPost, "/v1/messages", strings.NewReader(`{invalid json`))
-	req.Header.Set("Content-Type", "application/json")
-
-	resp := httptest.NewRecorder()
-	router.ServeHTTP(resp, req)
-
-	if resp.Code != http.StatusBadRequest {
-		t.Errorf("expected status 400, got %d", resp.Code)
-	}
-
-	var errResp anthropic.ErrorResponse
-	if err := json.Unmarshal(resp.Body.Bytes(), &errResp); err != nil {
-		t.Fatalf("failed to unmarshal error: %v", err)
-	}
-
-	if errResp.Type != "error" {
-		t.Errorf("expected type 'error', got %q", errResp.Type)
-	}
-	if errResp.Error.Type != "invalid_request_error" {
-		t.Errorf("expected error type 'invalid_request_error', got %q", errResp.Error.Type)
-	}
-}
-
-func TestAnthropicWriter_NonStreaming(t *testing.T) {
-	gin.SetMode(gin.TestMode)
-
-	router := gin.New()
-	router.Use(AnthropicMessagesMiddleware())
-	router.POST("/v1/messages", func(c *gin.Context) {
-		// Simulate Ollama response
-		resp := api.ChatResponse{
-			Model: "test-model",
-			Message: api.Message{
-				Role:    "assistant",
-				Content: "Hello there!",
-			},
-			Done:       true,
-			DoneReason: "stop",
-			Metrics: api.Metrics{
-				PromptEvalCount: 10,
-				EvalCount:       5,
-			},
-		}
-		data, _ := json.Marshal(resp)
-		c.Writer.WriteHeader(http.StatusOK)
-		_, _ = c.Writer.Write(data)
-	})
-
-	body := `{"model": "test-model", "max_tokens": 100, "messages": [{"role": "user", "content": "Hi"}]}`
-	req, _ := http.NewRequest(http.MethodPost, "/v1/messages", strings.NewReader(body))
-	req.Header.Set("Content-Type", "application/json")
-
-	resp := httptest.NewRecorder()
-	router.ServeHTTP(resp, req)
-
-	if resp.Code != http.StatusOK {
-		t.Fatalf("expected status 200, got %d", resp.Code)
-	}
-
-	var result anthropic.MessagesResponse
-	if err := json.Unmarshal(resp.Body.Bytes(), &result); err != nil {
-		t.Fatalf("failed to unmarshal response: %v", err)
-	}
-
-	if result.Type != "message" {
-		t.Errorf("expected type 'message', got %q", result.Type)
-	}
-	if result.Role != "assistant" {
-		t.Errorf("expected role 'assistant', got %q", result.Role)
-	}
-	if len(result.Content) != 1 {
-		t.Fatalf("expected 1 content block, got %d", len(result.Content))
-	}
-	if result.Content[0].Text == nil || *result.Content[0].Text != "Hello there!" {
-		t.Errorf("expected text 'Hello there!', got %v", result.Content[0].Text)
-	}
-	if result.StopReason != "end_turn" {
-		t.Errorf("expected stop_reason 'end_turn', got %q", result.StopReason)
-	}
-	if result.Usage.InputTokens != 10 {
-		t.Errorf("expected input_tokens 10, got %d", result.Usage.InputTokens)
-	}
-	if result.Usage.OutputTokens != 5 {
-		t.Errorf("expected output_tokens 5, got %d", result.Usage.OutputTokens)
-	}
-}
-
-// TestAnthropicWriter_ErrorFromRoutes tests error handling when routes.go sends
-// gin.H{"error": "message"} without a StatusCode field (which is the common case)
-func TestAnthropicWriter_ErrorFromRoutes(t *testing.T) {
-	gin.SetMode(gin.TestMode)
-
-	tests := []struct {
-		name          string
-		statusCode    int
-		errorPayload  any
-		wantErrorType string
-		wantMessage   string
-	}{
-		// routes.go sends errors without StatusCode in JSON, so we must use HTTP status
-		{
-			name:          "404 with gin.H error (model not found)",
-			statusCode:    http.StatusNotFound,
-			errorPayload:  gin.H{"error": "model 'nonexistent' not found"},
-			wantErrorType: "not_found_error",
-			wantMessage:   "model 'nonexistent' not found",
-		},
-		{
-			name:          "400 with gin.H error (bad request)",
-			statusCode:    http.StatusBadRequest,
-			errorPayload:  gin.H{"error": "model is required"},
-			wantErrorType: "invalid_request_error",
-			wantMessage:   "model is required",
-		},
-		{
-			name:          "500 with gin.H error (internal error)",
-			statusCode:    http.StatusInternalServerError,
-			errorPayload:  gin.H{"error": "something went wrong"},
-			wantErrorType: "api_error",
-			wantMessage:   "something went wrong",
-		},
-		{
-			name:       "404 with api.StatusError",
-			statusCode: http.StatusNotFound,
-			errorPayload: api.StatusError{
-				StatusCode:   http.StatusNotFound,
-				ErrorMessage: "model not found via StatusError",
-			},
-			wantErrorType: "not_found_error",
-			wantMessage:   "model not found via StatusError",
-		},
-	}
-
-	for _, tt := range tests {
-		t.Run(tt.name, func(t *testing.T) {
-			router := gin.New()
-			router.Use(AnthropicMessagesMiddleware())
-			router.POST("/v1/messages", func(c *gin.Context) {
-				// Simulate what routes.go does - set status and write error JSON
-				data, _ := json.Marshal(tt.errorPayload)
-				c.Writer.WriteHeader(tt.statusCode)
-				_, _ = c.Writer.Write(data)
-			})
-
-			body := `{"model": "test-model", "max_tokens": 100, "messages": [{"role": "user", "content": "Hi"}]}`
-			req, _ := http.NewRequest(http.MethodPost, "/v1/messages", strings.NewReader(body))
-			req.Header.Set("Content-Type", "application/json")
-
-			resp := httptest.NewRecorder()
-			router.ServeHTTP(resp, req)
-
-			if resp.Code != tt.statusCode {
-				t.Errorf("expected status %d, got %d", tt.statusCode, resp.Code)
-			}
-
-			var errResp anthropic.ErrorResponse
-			if err := json.Unmarshal(resp.Body.Bytes(), &errResp); err != nil {
-				t.Fatalf("failed to unmarshal error response: %v\nbody: %s", err, resp.Body.String())
-			}
-
-			if errResp.Type != "error" {
-				t.Errorf("expected type 'error', got %q", errResp.Type)
-			}
-			if errResp.Error.Type != tt.wantErrorType {
-				t.Errorf("expected error type %q, got %q", tt.wantErrorType, errResp.Error.Type)
-			}
-			if errResp.Error.Message != tt.wantMessage {
-				t.Errorf("expected message %q, got %q", tt.wantMessage, errResp.Error.Message)
-			}
-		})
-	}
-}

readline/errors.go

@@ -6,6 +6,9 @@ import (
 
 var ErrInterrupt = errors.New("Interrupt")
 
+// ErrExpandOutput is returned when user presses Ctrl+O to expand tool output
+var ErrExpandOutput = errors.New("ExpandOutput")
+
 type InterruptError struct {
 	Line []rune
 }

readline/readline.go

@@ -206,6 +206,9 @@ func (i *Instance) Readline() (string, error) {
 			buf.DeleteBefore()
 		case CharCtrlL:
 			buf.ClearScreen()
+		case CharCtrlO:
+			// Ctrl+O - expand tool output
+			return "", ErrExpandOutput
 		case CharCtrlW:
 			buf.DeleteWord()
 		case CharCtrlZ:

readline/types.go

@@ -18,6 +18,7 @@ const (
 	CharCtrlL     = 12
 	CharEnter     = 13
 	CharNext      = 14
+	CharCtrlO     = 15 // Ctrl+O - used for expanding tool output
 	CharPrev      = 16
 	CharBckSearch = 18
 	CharFwdSearch = 19

server/routes.go

@@ -1544,9 +1544,6 @@ func (s *Server) GenerateRoutes(rc *ollama.Registry) (http.Handler, error) {
 	r.GET("/v1/models/:model", middleware.RetrieveMiddleware(), s.ShowHandler)
 	r.POST("/v1/responses", middleware.ResponsesMiddleware(), s.ChatHandler)
 
-	// Inference (Anthropic compatibility)
-	r.POST("/v1/messages", middleware.AnthropicMessagesMiddleware(), s.ChatHandler)
-
 	if rc != nil {
 		// wrap old with new
 		rs := &registry.Local{

x/agent/approval.go

@@ -4,6 +4,7 @@ package agent
 import (
 	"fmt"
 	"os"
+	"path"
 	"path/filepath"
 	"strings"
 	"sync"
@@ -179,6 +180,7 @@ func FormatDeniedResult(command string, pattern string) string {
 // extractBashPrefix extracts a prefix pattern from a bash command.
 // For commands like "cat tools/tools_test.go | head -200", returns "cat:tools/"
 // For commands without path args, returns empty string.
+// Paths with ".." traversal that escape the base directory return empty string for security.
 func extractBashPrefix(command string) string {
 	// Split command by pipes and get the first part
 	parts := strings.Split(command, "|")
@@ -204,8 +206,8 @@ func extractBashPrefix(command string) string {
 		return ""
 	}
 
-	// Find the first path-like argument (must contain / or start with .)
-	// First pass: look for clear paths (containing / or starting with .)
+	// Find the first path-like argument (must contain / or \ or start with .)
+	// First pass: look for clear paths (containing path separators or starting with .)
 	for _, arg := range fields[1:] {
 		// Skip flags
 		if strings.HasPrefix(arg, "-") {
@@ -215,19 +217,49 @@ func extractBashPrefix(command string) string {
 		if isNumeric(arg) {
 			continue
 		}
-		// Only process if it looks like a path (contains / or starts with .)
-		if !strings.Contains(arg, "/") && !strings.HasPrefix(arg, ".") {
+		// Only process if it looks like a path (contains / or \ or starts with .)
+		if !strings.Contains(arg, "/") && !strings.Contains(arg, "\\") && !strings.HasPrefix(arg, ".") {
 			continue
 		}
-		// If arg ends with /, it's a directory - use it directly
-		if strings.HasSuffix(arg, "/") {
-			return fmt.Sprintf("%s:%s", baseCmd, arg)
+		// Normalize to forward slashes for consistent cross-platform matching
+		arg = strings.ReplaceAll(arg, "\\", "/")
+
+		// Security: reject absolute paths
+		if path.IsAbs(arg) {
+			return "" // Absolute path - don't create prefix
 		}
-		// Get the directory part of a file path
-		dir := filepath.Dir(arg)
+
+		// Normalize the path using stdlib path.Clean (resolves . and ..)
+		cleaned := path.Clean(arg)
+
+		// Security: reject if cleaned path escapes to parent directory
+		if strings.HasPrefix(cleaned, "..") {
+			return "" // Path escapes - don't create prefix
+		}
+
+		// Security: if original had "..", verify cleaned path didn't escape to sibling
+		// e.g., "tools/a/b/../../../etc" -> "etc" (escaped tools/ to sibling)
+		if strings.Contains(arg, "..") {
+			origBase := strings.SplitN(arg, "/", 2)[0]
+			cleanedBase := strings.SplitN(cleaned, "/", 2)[0]
+			if origBase != cleanedBase {
+				return "" // Path escaped to sibling directory
+			}
+		}
+
+		// Check if arg ends with / (explicit directory)
+		isDir := strings.HasSuffix(arg, "/")
+
+		// Get the directory part
+		var dir string
+		if isDir {
+			dir = cleaned
+		} else {
+			dir = path.Dir(cleaned)
+		}
+
 		if dir == "." {
-			// Path is just a directory like "tools" or "src" (no trailing /)
-			return fmt.Sprintf("%s:%s/", baseCmd, arg)
+			return fmt.Sprintf("%s:./", baseCmd)
 		}
 		return fmt.Sprintf("%s:%s/", baseCmd, dir)
 	}
@@ -332,6 +364,8 @@ func AllowlistKey(toolName string, args map[string]any) string {
 }
 
 // IsAllowed checks if a tool/command is allowed (exact match or prefix match).
+// For bash commands, hierarchical path matching is used - if "cat:tools/" is allowed,
+// then "cat:tools/subdir/" is also allowed (subdirectories inherit parent permissions).
 func (a *ApprovalManager) IsAllowed(toolName string, args map[string]any) bool {
 	a.mu.RLock()
 	defer a.mu.RUnlock()
@@ -342,12 +376,20 @@ func (a *ApprovalManager) IsAllowed(toolName string, args map[string]any) bool {
 		return true
 	}
 
-	// For bash commands, check prefix matches
+	// For bash commands, check prefix matches with hierarchical path support
 	if toolName == "bash" {
 		if cmd, ok := args["command"].(string); ok {
 			prefix := extractBashPrefix(cmd)
-			if prefix != "" && a.prefixes[prefix] {
-				return true
+			if prefix != "" {
+				// Check exact prefix match first
+				if a.prefixes[prefix] {
+					return true
+				}
+				// Check hierarchical match: if any stored prefix is a parent of current prefix
+				// e.g., stored "cat:tools/" should match current "cat:tools/subdir/"
+				if a.matchesHierarchicalPrefix(prefix) {
+					return true
+				}
 			}
 		}
 	}
@@ -360,6 +402,40 @@ func (a *ApprovalManager) IsAllowed(toolName string, args map[string]any) bool {
 	return false
 }
 
+// matchesHierarchicalPrefix checks if the given prefix matches any stored prefix hierarchically.
+// For example, if "cat:tools/" is stored, it will match "cat:tools/subdir/" or "cat:tools/a/b/c/".
+func (a *ApprovalManager) matchesHierarchicalPrefix(currentPrefix string) bool {
+	// Split prefix into command and path parts (format: "cmd:path/")
+	colonIdx := strings.Index(currentPrefix, ":")
+	if colonIdx == -1 {
+		return false
+	}
+	currentCmd := currentPrefix[:colonIdx]
+	currentPath := currentPrefix[colonIdx+1:]
+
+	for storedPrefix := range a.prefixes {
+		storedColonIdx := strings.Index(storedPrefix, ":")
+		if storedColonIdx == -1 {
+			continue
+		}
+		storedCmd := storedPrefix[:storedColonIdx]
+		storedPath := storedPrefix[storedColonIdx+1:]
+
+		// Commands must match exactly
+		if currentCmd != storedCmd {
+			continue
+		}
+
+		// Check if current path starts with stored path (hierarchical match)
+		// e.g., "tools/subdir/" starts with "tools/"
+		if strings.HasPrefix(currentPath, storedPath) {
+			return true
+		}
+	}
+
+	return false
+}
+
 // AddToAllowlist adds a tool/command to the session allowlist.
 // For bash commands, it adds the prefix pattern instead of exact command.
 func (a *ApprovalManager) AddToAllowlist(toolName string, args map[string]any) {
@@ -443,11 +519,12 @@ func formatToolDisplay(toolName string, args map[string]any) string {
 		}
 	}
 
-	// For web search, show query
+	// For web search, show query and internet notice
 	if toolName == "web_search" {
 		if query, ok := args["query"].(string); ok {
 			sb.WriteString(fmt.Sprintf("Tool: %s\n", toolName))
-			sb.WriteString(fmt.Sprintf("Query: %s", query))
+			sb.WriteString(fmt.Sprintf("Query: %s\n", query))
+			sb.WriteString("Uses internet via ollama.com")
 			return sb.String()
 		}
 	}
@@ -951,3 +1028,184 @@ func FormatDenyResult(toolName string, reason string) string {
 	}
 	return fmt.Sprintf("User denied execution of %s.", toolName)
 }
+
+// PromptYesNo displays a simple Yes/No prompt and returns the user's choice.
+// Returns true for Yes, false for No.
+func PromptYesNo(question string) (bool, error) {
+	fd := int(os.Stdin.Fd())
+	oldState, err := term.MakeRaw(fd)
+	if err != nil {
+		return false, err
+	}
+	defer term.Restore(fd, oldState)
+
+	selected := 0 // 0 = Yes, 1 = No
+	options := []string{"Yes", "No"}
+
+	// Hide cursor
+	fmt.Fprint(os.Stderr, "\033[?25l")
+	defer fmt.Fprint(os.Stderr, "\033[?25h")
+
+	renderYesNo := func() {
+		// Move to start of line and clear
+		fmt.Fprintf(os.Stderr, "\r\033[K")
+		fmt.Fprintf(os.Stderr, "\033[36m%s\033[0m ", question)
+		for i, opt := range options {
+			if i == selected {
+				fmt.Fprintf(os.Stderr, "\033[1;32m[%s]\033[0m ", opt)
+			} else {
+				fmt.Fprintf(os.Stderr, "\033[90m %s \033[0m ", opt)
+			}
+		}
+		fmt.Fprintf(os.Stderr, "\033[90m(←/→ or y/n, Enter to confirm)\033[0m")
+	}
+
+	renderYesNo()
+
+	buf := make([]byte, 3)
+	for {
+		n, err := os.Stdin.Read(buf)
+		if err != nil {
+			return false, err
+		}
+
+		if n == 1 {
+			switch buf[0] {
+			case 'y', 'Y':
+				selected = 0
+				renderYesNo()
+			case 'n', 'N':
+				selected = 1
+				renderYesNo()
+			case '\r', '\n': // Enter
+				fmt.Fprintf(os.Stderr, "\r\033[K") // Clear line
+				return selected == 0, nil
+			case 3: // Ctrl+C
+				fmt.Fprintf(os.Stderr, "\r\033[K")
+				return false, nil
+			case 27: // Escape - could be arrow key
+				// Read more bytes for arrow keys
+				continue
+			}
+		} else if n == 3 && buf[0] == 27 && buf[1] == 91 {
+			// Arrow keys
+			switch buf[2] {
+			case 'D': // Left
+				if selected > 0 {
+					selected--
+				}
+				renderYesNo()
+			case 'C': // Right
+				if selected < len(options)-1 {
+					selected++
+				}
+				renderYesNo()
+			}
+		}
+	}
+}
+
+// CloudModelOption represents a suggested cloud model for the selection prompt.
+type CloudModelOption struct {
+	Name        string
+	Description string
+}
+
+// PromptModelChoice displays a model selection prompt with multiple options.
+// Returns the selected model name, or empty string if user declined or cancelled.
+func PromptModelChoice(question string, models []CloudModelOption) (string, error) {
+	fd := int(os.Stdin.Fd())
+	oldState, err := term.MakeRaw(fd)
+	if err != nil {
+		return "", err
+	}
+	defer term.Restore(fd, oldState)
+
+	// Build options: models + "No thanks, continue"
+	optionCount := len(models) + 1
+	selected := 0
+
+	// Total lines: question + models + "no thanks" + hint = optionCount + 2
+	totalLines := optionCount + 2
+
+	// Hide cursor
+	fmt.Fprint(os.Stderr, "\033[?25l")
+	defer fmt.Fprint(os.Stderr, "\033[?25h")
+
+	firstRender := true
+
+	render := func() {
+		if !firstRender {
+			fmt.Fprintf(os.Stderr, "\033[%dA\r", totalLines-1)
+		}
+		firstRender = false
+
+		// \r\n needed in raw mode for proper line breaks
+		fmt.Fprintf(os.Stderr, "\033[K\033[36m%s\033[0m\r\n", question)
+
+		for i, model := range models {
+			fmt.Fprintf(os.Stderr, "\033[K")
+			if i == selected {
+				fmt.Fprintf(os.Stderr, "  \033[1;32m> %s\033[0m  \033[90m%s\033[0m\r\n", model.Name, model.Description)
+			} else {
+				fmt.Fprintf(os.Stderr, "    \033[90m%s  %s\033[0m\r\n", model.Name, model.Description)
+			}
+		}
+
+		fmt.Fprintf(os.Stderr, "\033[K")
+		if selected == len(models) {
+			fmt.Fprintf(os.Stderr, "  \033[1;32m> No thanks, continue\033[0m\r\n")
+		} else {
+			fmt.Fprintf(os.Stderr, "    \033[90mNo thanks, continue\033[0m\r\n")
+		}
+
+		fmt.Fprintf(os.Stderr, "\033[K\033[90m(↑/↓ to navigate, Enter to confirm)\033[0m")
+	}
+
+	render()
+
+	buf := make([]byte, 3)
+	for {
+		n, err := os.Stdin.Read(buf)
+		if err != nil {
+			return "", err
+		}
+
+		if n == 1 {
+			switch buf[0] {
+			case 'j', 'J':
+				if selected < optionCount-1 {
+					selected++
+				}
+				render()
+			case 'k', 'K':
+				if selected > 0 {
+					selected--
+				}
+				render()
+			case '\r', '\n':
+				fmt.Fprintf(os.Stderr, "\n")
+				if selected < len(models) {
+					return models[selected].Name, nil
+				}
+				return "", nil
+			case 3: // Ctrl+C
+				fmt.Fprintf(os.Stderr, "\n")
+				return "", nil
+			}
+		} else if n == 3 && buf[0] == 27 && buf[1] == 91 {
+			switch buf[2] {
+			case 'A': // Up
+				if selected > 0 {
+					selected--
+				}
+				render()
+			case 'B': // Down
+				if selected < optionCount-1 {
+					selected++
+				}
+				render()
+			}
+		}
+	}
+}

x/agent/approval_test.go

@@ -151,6 +151,27 @@ func TestExtractBashPrefix(t *testing.T) {
 			command:  "head -n 100",
 			expected: "",
 		},
+		// Path traversal security tests
+		{
+			name:     "path traversal - parent escape",
+			command:  "cat tools/../../etc/passwd",
+			expected: "", // Should NOT create a prefix - path escapes
+		},
+		{
+			name:     "path traversal - deep escape",
+			command:  "cat tools/a/b/../../../etc/passwd",
+			expected: "", // Normalizes to "../etc/passwd" - escapes
+		},
+		{
+			name:     "path traversal - absolute path",
+			command:  "cat /etc/passwd",
+			expected: "", // Absolute paths should not create prefix
+		},
+		{
+			name:     "path with safe dotdot - normalized",
+			command:  "cat tools/subdir/../file.go",
+			expected: "cat:tools/", // Normalizes to tools/file.go - safe, creates prefix
+		},
 	}
 
 	for _, tt := range tests {
@@ -164,6 +185,34 @@ func TestExtractBashPrefix(t *testing.T) {
 	}
 }
 
+func TestApprovalManager_PathTraversalBlocked(t *testing.T) {
+	am := NewApprovalManager()
+
+	// Allow "cat tools/file.go" - creates prefix "cat:tools/"
+	am.AddToAllowlist("bash", map[string]any{"command": "cat tools/file.go"})
+
+	// Path traversal attack: should NOT be allowed
+	if am.IsAllowed("bash", map[string]any{"command": "cat tools/../../etc/passwd"}) {
+		t.Error("SECURITY: path traversal attack should NOT be allowed")
+	}
+
+	// Another traversal variant
+	if am.IsAllowed("bash", map[string]any{"command": "cat tools/../../../etc/shadow"}) {
+		t.Error("SECURITY: deep path traversal should NOT be allowed")
+	}
+
+	// Valid subdirectory access should still work
+	if !am.IsAllowed("bash", map[string]any{"command": "cat tools/subdir/file.go"}) {
+		t.Error("expected cat tools/subdir/file.go to be allowed")
+	}
+
+	// Safe ".." that normalizes to within allowed directory should work
+	// tools/subdir/../other.go normalizes to tools/other.go which is under tools/
+	if !am.IsAllowed("bash", map[string]any{"command": "cat tools/subdir/../other.go"}) {
+		t.Error("expected cat tools/subdir/../other.go to be allowed (normalizes to tools/other.go)")
+	}
+}
+
 func TestApprovalManager_PrefixAllowlist(t *testing.T) {
 	am := NewApprovalManager()
 
@@ -186,6 +235,119 @@ func TestApprovalManager_PrefixAllowlist(t *testing.T) {
 	}
 }
 
+func TestApprovalManager_HierarchicalPrefixAllowlist(t *testing.T) {
+	am := NewApprovalManager()
+
+	// Allow "cat tools/file.go" - this creates prefix "cat:tools/"
+	am.AddToAllowlist("bash", map[string]any{"command": "cat tools/file.go"})
+
+	// Should allow subdirectories (hierarchical matching)
+	if !am.IsAllowed("bash", map[string]any{"command": "cat tools/subdir/file.go"}) {
+		t.Error("expected cat tools/subdir/file.go to be allowed via hierarchical prefix")
+	}
+
+	// Should allow deeply nested subdirectories
+	if !am.IsAllowed("bash", map[string]any{"command": "cat tools/a/b/c/deep.go"}) {
+		t.Error("expected cat tools/a/b/c/deep.go to be allowed via hierarchical prefix")
+	}
+
+	// Should still allow same directory
+	if !am.IsAllowed("bash", map[string]any{"command": "cat tools/another.go"}) {
+		t.Error("expected cat tools/another.go to be allowed")
+	}
+
+	// Should NOT allow different base directory
+	if am.IsAllowed("bash", map[string]any{"command": "cat src/main.go"}) {
+		t.Error("expected cat src/main.go to NOT be allowed")
+	}
+
+	// Should NOT allow different command even in subdirectory
+	if am.IsAllowed("bash", map[string]any{"command": "ls tools/subdir/"}) {
+		t.Error("expected ls tools/subdir/ to NOT be allowed (different command)")
+	}
+
+	// Should NOT allow similar but different directory name
+	if am.IsAllowed("bash", map[string]any{"command": "cat toolsbin/file.go"}) {
+		t.Error("expected cat toolsbin/file.go to NOT be allowed (different directory)")
+	}
+}
+
+func TestApprovalManager_HierarchicalPrefixAllowlist_CrossPlatform(t *testing.T) {
+	am := NewApprovalManager()
+
+	// Allow with forward slashes (Unix-style)
+	am.AddToAllowlist("bash", map[string]any{"command": "cat tools/file.go"})
+
+	// Should work with backslashes too (Windows-style) - normalized internally
+	if !am.IsAllowed("bash", map[string]any{"command": "cat tools\\subdir\\file.go"}) {
+		t.Error("expected cat tools\\subdir\\file.go to be allowed via hierarchical prefix (Windows path)")
+	}
+
+	// Mixed slashes should also work
+	if !am.IsAllowed("bash", map[string]any{"command": "cat tools\\a/b\\c/deep.go"}) {
+		t.Error("expected mixed slash path to be allowed via hierarchical prefix")
+	}
+}
+
+func TestMatchesHierarchicalPrefix(t *testing.T) {
+	am := NewApprovalManager()
+
+	// Add prefix for "cat:tools/"
+	am.prefixes["cat:tools/"] = true
+
+	tests := []struct {
+		name     string
+		prefix   string
+		expected bool
+	}{
+		{
+			name:     "exact match",
+			prefix:   "cat:tools/",
+			expected: true, // exact match also passes HasPrefix - caller handles exact match first
+		},
+		{
+			name:     "subdirectory",
+			prefix:   "cat:tools/subdir/",
+			expected: true,
+		},
+		{
+			name:     "deeply nested",
+			prefix:   "cat:tools/a/b/c/",
+			expected: true,
+		},
+		{
+			name:     "different base directory",
+			prefix:   "cat:src/",
+			expected: false,
+		},
+		{
+			name:     "different command same path",
+			prefix:   "ls:tools/",
+			expected: false,
+		},
+		{
+			name:     "similar directory name",
+			prefix:   "cat:toolsbin/",
+			expected: false,
+		},
+		{
+			name:     "invalid prefix format",
+			prefix:   "cattools",
+			expected: false,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			result := am.matchesHierarchicalPrefix(tt.prefix)
+			if result != tt.expected {
+				t.Errorf("matchesHierarchicalPrefix(%q) = %v, expected %v",
+					tt.prefix, result, tt.expected)
+			}
+		})
+	}
+}
+
 func TestFormatApprovalResult(t *testing.T) {
 	tests := []struct {
 		name     string

x/agent/prompt_test.go

@@ -0,0 +1,25 @@
+package agent
+
+import (
+	"testing"
+)
+
+func TestCloudModelOptionStruct(t *testing.T) {
+	// Test that the struct is defined correctly
+	models := []CloudModelOption{
+		{Name: "glm-4.7:cloud", Description: "GLM 4.7 Cloud"},
+		{Name: "qwen3-coder:480b-cloud", Description: "Qwen3 Coder 480B"},
+	}
+
+	if len(models) != 2 {
+		t.Errorf("expected 2 models, got %d", len(models))
+	}
+
+	if models[0].Name != "glm-4.7:cloud" {
+		t.Errorf("expected glm-4.7:cloud, got %s", models[0].Name)
+	}
+
+	if models[1].Description != "Qwen3 Coder 480B" {
+		t.Errorf("expected 'Qwen3 Coder 480B', got %s", models[1].Description)
+	}
+}

x/cmd/cloudmodel_test.go

@@ -0,0 +1,41 @@
+package cmd
+
+import (
+	"errors"
+	"testing"
+)
+
+func TestCloudModelSwitchRequest(t *testing.T) {
+	// Test the error type
+	req := &CloudModelSwitchRequest{Model: "glm-4.7:cloud"}
+
+	// Test Error() method
+	errMsg := req.Error()
+	expected := "switch to model: glm-4.7:cloud"
+	if errMsg != expected {
+		t.Errorf("expected %q, got %q", expected, errMsg)
+	}
+
+	// Test errors.As
+	var err error = req
+	var switchReq *CloudModelSwitchRequest
+	if !errors.As(err, &switchReq) {
+		t.Error("errors.As should return true for CloudModelSwitchRequest")
+	}
+
+	if switchReq.Model != "glm-4.7:cloud" {
+		t.Errorf("expected model glm-4.7:cloud, got %s", switchReq.Model)
+	}
+}
+
+func TestSuggestedCloudModels(t *testing.T) {
+	// Verify the suggested models are defined
+	if len(suggestedCloudModels) == 0 {
+		t.Error("suggestedCloudModels should not be empty")
+	}
+
+	// Check first model
+	if suggestedCloudModels[0].Name != "glm-4.7:cloud" {
+		t.Errorf("expected first model to be glm-4.7:cloud, got %s", suggestedCloudModels[0].Name)
+	}
+}

x/cmd/run.go

@@ -6,10 +6,12 @@ import (
 	"errors"
 	"fmt"
 	"io"
+	"net/url"
 	"os"
 	"os/signal"
 	"strings"
 	"syscall"
+	"time"
 
 	"github.com/spf13/cobra"
 	"golang.org/x/term"
@@ -22,6 +24,132 @@ import (
 	"github.com/ollama/ollama/x/tools"
 )
 
+// Tool output capping constants
+const (
+	// localModelTokenLimit is the token limit for local models (smaller context).
+	localModelTokenLimit = 4000
+
+	// defaultTokenLimit is the token limit for cloud/remote models.
+	defaultTokenLimit = 10000
+
+	// charsPerToken is a rough estimate of characters per token.
+	// TODO: Estimate tokens more accurately using tokenizer if available
+	charsPerToken = 4
+)
+
+// suggestedCloudModels are the models suggested to users after signing in.
+// TODO(parthsareen): Dynamically recommend models based on user context instead of hardcoding
+var suggestedCloudModels = []agent.CloudModelOption{
+	{Name: "glm-4.7:cloud", Description: "GLM 4.7 Cloud"},
+	{Name: "qwen3-coder:480b-cloud", Description: "Qwen3 Coder 480B"},
+}
+
+// CloudModelSwitchRequest signals that the user wants to switch to a different model.
+type CloudModelSwitchRequest struct {
+	Model string
+}
+
+func (c *CloudModelSwitchRequest) Error() string {
+	return fmt.Sprintf("switch to model: %s", c.Model)
+}
+
+// isLocalModel checks if the model is running locally (not a cloud model).
+// TODO: Improve local/cloud model identification - could check model metadata
+func isLocalModel(modelName string) bool {
+	return !strings.HasSuffix(modelName, "-cloud")
+}
+
+// isLocalServer checks if connecting to a local Ollama server.
+// TODO: Could also check other indicators of local vs cloud server
+func isLocalServer() bool {
+	host := os.Getenv("OLLAMA_HOST")
+	if host == "" {
+		return true // Default is localhost:11434
+	}
+
+	// Parse the URL to check host
+	parsed, err := url.Parse(host)
+	if err != nil {
+		return true // If can't parse, assume local
+	}
+
+	hostname := parsed.Hostname()
+	return hostname == "localhost" || hostname == "127.0.0.1" || strings.Contains(parsed.Host, ":11434")
+}
+
+// truncateToolOutput truncates tool output to prevent context overflow.
+// Uses a smaller limit (4k tokens) for local models, larger (10k) for cloud/remote.
+func truncateToolOutput(output, modelName string) string {
+	var tokenLimit int
+	if isLocalModel(modelName) && isLocalServer() {
+		tokenLimit = localModelTokenLimit
+	} else {
+		tokenLimit = defaultTokenLimit
+	}
+
+	maxChars := tokenLimit * charsPerToken
+	if len(output) > maxChars {
+		return output[:maxChars] + "\n... (output truncated)"
+	}
+	return output
+}
+
+// waitForOllamaSignin shows the signin URL and polls until authentication completes.
+func waitForOllamaSignin(ctx context.Context) error {
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		return err
+	}
+
+	// Get signin URL from initial Whoami call
+	_, err = client.Whoami(ctx)
+	if err != nil {
+		var aErr api.AuthorizationError
+		if errors.As(err, &aErr) && aErr.SigninURL != "" {
+			fmt.Fprintf(os.Stderr, "\n  To sign in, navigate to:\n")
+			fmt.Fprintf(os.Stderr, "      \033[36m%s\033[0m\n\n", aErr.SigninURL)
+			fmt.Fprintf(os.Stderr, "  \033[90mWaiting for sign in to complete...\033[0m")
+
+			// Poll until auth succeeds
+			ticker := time.NewTicker(2 * time.Second)
+			defer ticker.Stop()
+
+			for {
+				select {
+				case <-ctx.Done():
+					fmt.Fprintf(os.Stderr, "\n")
+					return ctx.Err()
+				case <-ticker.C:
+					user, whoamiErr := client.Whoami(ctx)
+					if whoamiErr == nil && user != nil && user.Name != "" {
+						fmt.Fprintf(os.Stderr, "\r\033[K  \033[32mSigned in as %s\033[0m\n", user.Name)
+						return nil
+					}
+					// Still waiting, show dot
+					fmt.Fprintf(os.Stderr, ".")
+				}
+			}
+		}
+		return err
+	}
+	return nil
+}
+
+// promptCloudModelSuggestion shows cloud model suggestions after successful sign-in.
+// Returns the selected model name, or empty string if user declines.
+func promptCloudModelSuggestion() string {
+	fmt.Fprintf(os.Stderr, "\n")
+	fmt.Fprintf(os.Stderr, "\033[1;36mTry cloud models for free!\033[0m\n")
+	fmt.Fprintf(os.Stderr, "\033[90mCloud models offer powerful capabilities without local hardware requirements.\033[0m\n")
+	fmt.Fprintf(os.Stderr, "\n")
+
+	selectedModel, err := agent.PromptModelChoice("Try a cloud model now?", suggestedCloudModels)
+	if err != nil || selectedModel == "" {
+		return ""
+	}
+	return selectedModel
+}
+
 // RunOptions contains options for running an interactive agent session.
 type RunOptions struct {
 	Model        string
@@ -37,6 +165,50 @@ type RunOptions struct {
 	// Agent fields (managed externally for session persistence)
 	Tools    *tools.Registry
 	Approval *agent.ApprovalManager
+
+	// YoloMode skips all tool approval prompts
+	YoloMode bool
+
+	// LastToolOutput stores the full output of the last tool execution
+	// for Ctrl+O expansion. Updated by Chat(), read by caller.
+	LastToolOutput *string
+
+	// LastToolOutputTruncated stores the truncated version shown inline
+	LastToolOutputTruncated *string
+
+	// ActiveModel points to the current model name - can be updated mid-turn
+	// for model switching. If nil, opts.Model is used.
+	ActiveModel *string
+}
+
+// getActiveModel returns the current model name, checking ActiveModel pointer first.
+func getActiveModel(opts *RunOptions) string {
+	if opts.ActiveModel != nil && *opts.ActiveModel != "" {
+		return *opts.ActiveModel
+	}
+	return opts.Model
+}
+
+// showModelConnection displays "Connecting to X on ollama.com" for cloud models.
+func showModelConnection(ctx context.Context, modelName string) error {
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		return err
+	}
+
+	info, err := client.Show(ctx, &api.ShowRequest{Model: modelName})
+	if err != nil {
+		return err
+	}
+
+	if info.RemoteHost != "" {
+		if strings.HasPrefix(info.RemoteHost, "https://ollama.com") {
+			fmt.Fprintf(os.Stderr, "Connecting to '%s' on 'ollama.com' ⚡\n", info.RemoteModel)
+		} else {
+			fmt.Fprintf(os.Stderr, "Connecting to '%s' on '%s'\n", info.RemoteModel, info.RemoteHost)
+		}
+	}
+	return nil
 }
 
 // Chat runs an agent chat loop with tool support.
@@ -77,6 +249,7 @@ func Chat(ctx context.Context, opts RunOptions) (*api.Message, error) {
 	var thinkTagOpened bool = false
 	var thinkTagClosed bool = false
 	var pendingToolCalls []api.ToolCall
+	var consecutiveErrors int // Track consecutive 500 errors for retry limit
 
 	role := "assistant"
 	messages := opts.Messages
@@ -135,7 +308,7 @@ func Chat(ctx context.Context, opts RunOptions) (*api.Message, error) {
 	// Agentic loop: continue until no more tool calls
 	for {
 		req := &api.ChatRequest{
-			Model:    opts.Model,
+			Model:    getActiveModel(&opts),
 			Messages: messages,
 			Format:   json.RawMessage(opts.Format),
 			Options:  opts.Options,
@@ -159,6 +332,61 @@ func Chat(ctx context.Context, opts RunOptions) (*api.Message, error) {
 				return nil, nil
 			}
 
+			var authErr api.AuthorizationError
+			if errors.As(err, &authErr) {
+				p.StopAndClear()
+				fmt.Fprintf(os.Stderr, "\033[33mAuthentication required to use this cloud model.\033[0m\n")
+				result, promptErr := agent.PromptYesNo("Sign in to Ollama?")
+				if promptErr == nil && result {
+					if signinErr := waitForOllamaSignin(ctx); signinErr == nil {
+						suggestedModel := promptCloudModelSuggestion()
+						if suggestedModel != "" {
+							return nil, &CloudModelSwitchRequest{Model: suggestedModel}
+						}
+
+						fmt.Fprintf(os.Stderr, "\033[90mRetrying...\033[0m\n")
+						continue
+					}
+				}
+				return nil, fmt.Errorf("authentication required - run 'ollama signin' to authenticate")
+			}
+
+			// Check for 500 errors (often tool parsing failures) - inform the model
+			var statusErr api.StatusError
+			if errors.As(err, &statusErr) && statusErr.StatusCode >= 500 {
+				consecutiveErrors++
+				p.StopAndClear()
+
+				if consecutiveErrors >= 3 {
+					fmt.Fprintf(os.Stderr, "\033[31m✗ Too many consecutive errors, giving up\033[0m\n")
+					return nil, fmt.Errorf("too many consecutive server errors: %s", statusErr.ErrorMessage)
+				}
+
+				fmt.Fprintf(os.Stderr, "\033[33m⚠ Server error (attempt %d/3): %s\033[0m\n", consecutiveErrors, statusErr.ErrorMessage)
+
+				// Include both the model's response and the error so it can learn
+				assistantContent := fullResponse.String()
+				if assistantContent == "" {
+					assistantContent = "(empty response)"
+				}
+				errorMsg := fmt.Sprintf("Your previous response caused an error: %s\n\nYour response was:\n%s\n\nPlease try again with a valid response.", statusErr.ErrorMessage, assistantContent)
+				messages = append(messages,
+					api.Message{Role: "user", Content: errorMsg},
+				)
+
+				// Reset state and retry
+				fullResponse.Reset()
+				thinkingContent.Reset()
+				thinkTagOpened = false
+				thinkTagClosed = false
+				pendingToolCalls = nil
+				state = &displayResponseState{}
+				p = progress.NewProgress(os.Stderr)
+				spinner = progress.NewSpinner("")
+				p.Add("", spinner)
+				continue
+			}
+
 			if strings.Contains(err.Error(), "upstream error") {
 				p.StopAndClear()
 				fmt.Println("An error occurred while processing your message. Please try again.")
@@ -168,6 +396,9 @@ func Chat(ctx context.Context, opts RunOptions) (*api.Message, error) {
 			return nil, err
 		}
 
+		// Reset consecutive error counter on success
+		consecutiveErrors = 0
+
 		// If no tool calls, we're done
 		if len(pendingToolCalls) == 0 || toolRegistry == nil {
 			break
@@ -216,7 +447,12 @@ func Chat(ctx context.Context, opts RunOptions) (*api.Message, error) {
 			}
 
 			// Check approval (uses prefix matching for bash commands)
-			if !skipApproval && !approval.IsAllowed(toolName, args) {
+			// In yolo mode, skip all approval prompts
+			if opts.YoloMode {
+				if !skipApproval {
+					fmt.Fprintf(os.Stderr, "\033[90m▶ Running: %s\033[0m\n", formatToolShort(toolName, args))
+				}
+			} else if !skipApproval && !approval.IsAllowed(toolName, args) {
 				result, err := approval.RequestApproval(toolName, args)
 				if err != nil {
 					fmt.Fprintf(os.Stderr, "Error requesting approval: %v\n", err)
@@ -247,9 +483,27 @@ func Chat(ctx context.Context, opts RunOptions) (*api.Message, error) {
 				fmt.Fprintf(os.Stderr, "\033[90m▶ Running: %s\033[0m\n", formatToolShort(toolName, args))
 			}
 
-			// Execute the tool
 			toolResult, err := toolRegistry.Execute(call)
 			if err != nil {
+				if errors.Is(err, tools.ErrWebSearchAuthRequired) {
+					fmt.Fprintf(os.Stderr, "\033[33m  Web search requires authentication.\033[0m\n")
+					result, promptErr := agent.PromptYesNo("Sign in to Ollama?")
+					if promptErr == nil && result {
+						if signinErr := waitForOllamaSignin(ctx); signinErr == nil {
+							suggestedModel := promptCloudModelSuggestion()
+							if suggestedModel != "" && opts.ActiveModel != nil {
+								*opts.ActiveModel = suggestedModel
+								showModelConnection(ctx, suggestedModel)
+							}
+
+							fmt.Fprintf(os.Stderr, "\033[90mRetrying web search...\033[0m\n")
+							toolResult, err = toolRegistry.Execute(call)
+							if err == nil {
+								goto toolSuccess
+							}
+						}
+					}
+				}
 				fmt.Fprintf(os.Stderr, "\033[31m  Error: %v\033[0m\n", err)
 				toolResults = append(toolResults, api.Message{
 					Role:       "tool",
@@ -258,20 +512,34 @@ func Chat(ctx context.Context, opts RunOptions) (*api.Message, error) {
 				})
 				continue
 			}
+		toolSuccess:
 
 			// Display tool output (truncated for display)
+			truncatedOutput := ""
 			if toolResult != "" {
 				output := toolResult
 				if len(output) > 300 {
-					output = output[:300] + "... (truncated)"
+					output = output[:300] + "... (truncated, press Ctrl+O to expand)"
 				}
+				truncatedOutput = output
 				// Show result in grey, indented
 				fmt.Fprintf(os.Stderr, "\033[90m  %s\033[0m\n", strings.ReplaceAll(output, "\n", "\n  "))
 			}
 
+			// Store full and truncated output for Ctrl+O toggle
+			if opts.LastToolOutput != nil {
+				*opts.LastToolOutput = toolResult
+			}
+			if opts.LastToolOutputTruncated != nil {
+				*opts.LastToolOutputTruncated = truncatedOutput
+			}
+
+			// Truncate output to prevent context overflow
+			toolResultForLLM := truncateToolOutput(toolResult, getActiveModel(&opts))
+
 			toolResults = append(toolResults, api.Message{
 				Role:       "tool",
-				Content:    toolResult,
+				Content:    toolResultForLLM,
 				ToolCallID: call.ID,
 			})
 		}
@@ -426,30 +694,34 @@ func renderToolCalls(toolCalls []api.ToolCall, plainText bool) string {
 	return out
 }
 
-// checkModelCapabilities checks if the model supports tools.
-func checkModelCapabilities(ctx context.Context, modelName string) (supportsTools bool, err error) {
+// checkModelCapabilities checks if the model supports tools and thinking.
+func checkModelCapabilities(ctx context.Context, modelName string) (supportsTools bool, supportsThinking bool, err error) {
 	client, err := api.ClientFromEnvironment()
 	if err != nil {
-		return false, err
+		return false, false, err
 	}
 
 	resp, err := client.Show(ctx, &api.ShowRequest{Model: modelName})
 	if err != nil {
-		return false, err
+		return false, false, err
 	}
 
 	for _, cap := range resp.Capabilities {
 		if cap == model.CapabilityTools {
-			return true, nil
+			supportsTools = true
+		}
+		if cap == model.CapabilityThinking {
+			supportsThinking = true
 		}
 	}
 
-	return false, nil
+	return supportsTools, supportsThinking, nil
 }
 
 // GenerateInteractive runs an interactive agent session.
 // This is called from cmd.go when --experimental flag is set.
-func GenerateInteractive(cmd *cobra.Command, modelName string, wordWrap bool, options map[string]any, think *api.ThinkValue, hideThinking bool, keepAlive *api.Duration) error {
+// If yoloMode is true, all tool approvals are skipped.
+func GenerateInteractive(cmd *cobra.Command, modelName string, wordWrap bool, options map[string]any, think *api.ThinkValue, hideThinking bool, keepAlive *api.Duration, yoloMode bool) error {
 	scanner, err := readline.New(readline.Prompt{
 		Prompt:         ">>> ",
 		AltPrompt:      "... ",
@@ -463,22 +735,26 @@ func GenerateInteractive(cmd *cobra.Command, modelName string, wordWrap bool, op
 	fmt.Print(readline.StartBracketedPaste)
 	defer fmt.Printf(readline.EndBracketedPaste)
 
-	// Check if model supports tools
-	supportsTools, err := checkModelCapabilities(cmd.Context(), modelName)
+	// Check if model supports tools and thinking
+	supportsTools, supportsThinking, err := checkModelCapabilities(cmd.Context(), modelName)
 	if err != nil {
 		fmt.Fprintf(os.Stderr, "\033[33mWarning: Could not check model capabilities: %v\033[0m\n", err)
 		supportsTools = false
+		supportsThinking = false
 	}
 
+	// Track if session is using thinking mode
+	usingThinking := think != nil && supportsThinking
+
 	// Create tool registry only if model supports tools
 	var toolRegistry *tools.Registry
 	if supportsTools {
 		toolRegistry = tools.DefaultRegistry()
-		fmt.Fprintf(os.Stderr, "Tools available: %s\n", strings.Join(toolRegistry.Names(), ", "))
-
-		// Check for OLLAMA_API_KEY for web search
-		if os.Getenv("OLLAMA_API_KEY") == "" {
-			fmt.Fprintf(os.Stderr, "\033[33mWarning: OLLAMA_API_KEY not set - web search will not work\033[0m\n")
+		if toolRegistry.Count() > 0 {
+			fmt.Fprintf(os.Stderr, "\033[90mTools available: %s\033[0m\n", strings.Join(toolRegistry.Names(), ", "))
+		}
+		if yoloMode {
+			fmt.Fprintf(os.Stderr, "\033[33m⚠ YOLO mode: All tool approvals will be skipped\033[0m\n")
 		}
 	} else {
 		fmt.Fprintf(os.Stderr, "\033[33mNote: Model does not support tools - running in chat-only mode\033[0m\n")
@@ -490,6 +766,11 @@ func GenerateInteractive(cmd *cobra.Command, modelName string, wordWrap bool, op
 	var messages []api.Message
 	var sb strings.Builder
 
+	// Track last tool output for Ctrl+O toggle
+	var lastToolOutput string
+	var lastToolOutputTruncated string
+	var toolOutputExpanded bool
+
 	for {
 		line, err := scanner.Readline()
 		switch {
@@ -502,6 +783,20 @@ func GenerateInteractive(cmd *cobra.Command, modelName string, wordWrap bool, op
 			}
 			sb.Reset()
 			continue
+		case errors.Is(err, readline.ErrExpandOutput):
+			// Ctrl+O pressed - toggle between expanded and collapsed tool output
+			if lastToolOutput == "" {
+				fmt.Fprintf(os.Stderr, "\033[90mNo tool output to expand\033[0m\n")
+			} else if toolOutputExpanded {
+				// Currently expanded, show truncated
+				fmt.Fprintf(os.Stderr, "\033[90m  %s\033[0m\n", strings.ReplaceAll(lastToolOutputTruncated, "\n", "\n  "))
+				toolOutputExpanded = false
+			} else {
+				// Currently collapsed, show full
+				fmt.Fprintf(os.Stderr, "\033[90m  %s\033[0m\n", strings.ReplaceAll(lastToolOutput, "\n", "\n  "))
+				toolOutputExpanded = true
+			}
+			continue
 		case err != nil:
 			return err
 		}
@@ -524,6 +819,9 @@ func GenerateInteractive(cmd *cobra.Command, modelName string, wordWrap bool, op
 			fmt.Fprintln(os.Stderr, "  /bye            Exit")
 			fmt.Fprintln(os.Stderr, "  /?, /help       Help for a command")
 			fmt.Fprintln(os.Stderr, "")
+			fmt.Fprintln(os.Stderr, "Keyboard Shortcuts:")
+			fmt.Fprintln(os.Stderr, "  Ctrl+O          Expand last tool output")
+			fmt.Fprintln(os.Stderr, "")
 			continue
 		case strings.HasPrefix(line, "/"):
 			fmt.Printf("Unknown command '%s'. Type /? for help\n", strings.Fields(line)[0])
@@ -535,25 +833,44 @@ func GenerateInteractive(cmd *cobra.Command, modelName string, wordWrap bool, op
 		if sb.Len() > 0 {
 			newMessage := api.Message{Role: "user", Content: sb.String()}
 			messages = append(messages, newMessage)
+			toolOutputExpanded = false
 
-			opts := RunOptions{
-				Model:        modelName,
-				Messages:     messages,
-				WordWrap:     wordWrap,
-				Options:      options,
-				Think:        think,
-				HideThinking: hideThinking,
-				KeepAlive:    keepAlive,
-				Tools:        toolRegistry,
-				Approval:     approval,
-			}
+		retryChat:
+			for {
+				opts := RunOptions{
+					Model:                   modelName,
+					Messages:                messages,
+					WordWrap:                wordWrap,
+					Options:                 options,
+					Think:                   think,
+					HideThinking:            hideThinking,
+					KeepAlive:               keepAlive,
+					Tools:                   toolRegistry,
+					Approval:                approval,
+					YoloMode:                yoloMode,
+					LastToolOutput:          &lastToolOutput,
+					LastToolOutputTruncated: &lastToolOutputTruncated,
+					ActiveModel:             &modelName,
+				}
 
-			assistant, err := Chat(cmd.Context(), opts)
-			if err != nil {
-				return err
-			}
-			if assistant != nil {
-				messages = append(messages, *assistant)
+				assistant, err := Chat(cmd.Context(), opts)
+				if err != nil {
+					var switchReq *CloudModelSwitchRequest
+					if errors.As(err, &switchReq) {
+						newModel := switchReq.Model
+						if err := switchToModel(cmd.Context(), newModel, &modelName, &supportsTools, &supportsThinking, &toolRegistry, usingThinking); err != nil {
+							fmt.Fprintf(os.Stderr, "\033[33m%v\033[0m\n", err)
+							fmt.Fprintf(os.Stderr, "\033[90mContinuing with %s...\033[0m\n", modelName)
+						}
+						continue retryChat
+					}
+					return err
+				}
+
+				if assistant != nil {
+					messages = append(messages, *assistant)
+				}
+				break retryChat
 			}
 
 			sb.Reset()
@@ -561,6 +878,52 @@ func GenerateInteractive(cmd *cobra.Command, modelName string, wordWrap bool, op
 	}
 }
 
+// switchToModel handles model switching with capability checks and UI updates.
+func switchToModel(ctx context.Context, newModel string, modelName *string, supportsTools, supportsThinking *bool, toolRegistry **tools.Registry, usingThinking bool) error {
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		return fmt.Errorf("could not create client: %w", err)
+	}
+
+	newSupportsTools, newSupportsThinking, capErr := checkModelCapabilities(ctx, newModel)
+	if capErr != nil {
+		return fmt.Errorf("could not check model capabilities: %w", capErr)
+	}
+
+	// TODO(parthsareen): Handle thinking -> non-thinking model switch gracefully
+	if usingThinking && !newSupportsThinking {
+		return fmt.Errorf("%s does not support thinking mode", newModel)
+	}
+
+	// Show "Connecting to X on ollama.com" for cloud models
+	info, err := client.Show(ctx, &api.ShowRequest{Model: newModel})
+	if err == nil && info.RemoteHost != "" {
+		if strings.HasPrefix(info.RemoteHost, "https://ollama.com") {
+			fmt.Fprintf(os.Stderr, "Connecting to '%s' on 'ollama.com' ⚡\n", info.RemoteModel)
+		} else {
+			fmt.Fprintf(os.Stderr, "Connecting to '%s' on '%s'\n", info.RemoteModel, info.RemoteHost)
+		}
+	}
+
+	*modelName = newModel
+	*supportsTools = newSupportsTools
+	*supportsThinking = newSupportsThinking
+
+	if *supportsTools {
+		if *toolRegistry == nil {
+			*toolRegistry = tools.DefaultRegistry()
+		}
+		if (*toolRegistry).Count() > 0 {
+			fmt.Fprintf(os.Stderr, "\033[90mTools available: %s\033[0m\n", strings.Join((*toolRegistry).Names(), ", "))
+		}
+	} else {
+		*toolRegistry = nil
+		fmt.Fprintf(os.Stderr, "\033[33mNote: Model does not support tools - running in chat-only mode\033[0m\n")
+	}
+
+	return nil
+}
+
 // showToolsStatus displays the current tools and approval status.
 func showToolsStatus(registry *tools.Registry, approval *agent.ApprovalManager, supportsTools bool) {
 	if !supportsTools || registry == nil {

x/cmd/run_test.go

@@ -0,0 +1,180 @@
+package cmd
+
+import (
+	"testing"
+)
+
+func TestIsLocalModel(t *testing.T) {
+	tests := []struct {
+		name      string
+		modelName string
+		expected  bool
+	}{
+		{
+			name:      "local model without suffix",
+			modelName: "llama3.2",
+			expected:  true,
+		},
+		{
+			name:      "local model with version",
+			modelName: "qwen2.5:7b",
+			expected:  true,
+		},
+		{
+			name:      "cloud model",
+			modelName: "gpt-4-cloud",
+			expected:  false,
+		},
+		{
+			name:      "cloud model with version",
+			modelName: "claude-3-cloud",
+			expected:  false,
+		},
+		{
+			name:      "empty model name",
+			modelName: "",
+			expected:  true,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			result := isLocalModel(tt.modelName)
+			if result != tt.expected {
+				t.Errorf("isLocalModel(%q) = %v, expected %v", tt.modelName, result, tt.expected)
+			}
+		})
+	}
+}
+
+func TestIsLocalServer(t *testing.T) {
+	tests := []struct {
+		name     string
+		host     string
+		expected bool
+	}{
+		{
+			name:     "empty host (default)",
+			host:     "",
+			expected: true,
+		},
+		{
+			name:     "localhost",
+			host:     "http://localhost:11434",
+			expected: true,
+		},
+		{
+			name:     "127.0.0.1",
+			host:     "http://127.0.0.1:11434",
+			expected: true,
+		},
+		{
+			name:     "custom port on localhost",
+			host:     "http://localhost:8080",
+			expected: true, // localhost is always considered local
+		},
+		{
+			name:     "remote host",
+			host:     "http://ollama.example.com:11434",
+			expected: true, // has :11434
+		},
+		{
+			name:     "remote host different port",
+			host:     "http://ollama.example.com:8080",
+			expected: false,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Setenv("OLLAMA_HOST", tt.host)
+			result := isLocalServer()
+			if result != tt.expected {
+				t.Errorf("isLocalServer() with OLLAMA_HOST=%q = %v, expected %v", tt.host, result, tt.expected)
+			}
+		})
+	}
+}
+
+func TestTruncateToolOutput(t *testing.T) {
+	// Create outputs of different sizes
+	localLimitOutput := make([]byte, 20000)   // > 4k tokens (16k chars)
+	defaultLimitOutput := make([]byte, 50000) // > 10k tokens (40k chars)
+	for i := range localLimitOutput {
+		localLimitOutput[i] = 'a'
+	}
+	for i := range defaultLimitOutput {
+		defaultLimitOutput[i] = 'b'
+	}
+
+	tests := []struct {
+		name          string
+		output        string
+		modelName     string
+		host          string
+		shouldTrim    bool
+		expectedLimit int
+	}{
+		{
+			name:          "short output local model",
+			output:        "hello world",
+			modelName:     "llama3.2",
+			host:          "",
+			shouldTrim:    false,
+			expectedLimit: localModelTokenLimit,
+		},
+		{
+			name:          "long output local model - trimmed at 4k",
+			output:        string(localLimitOutput),
+			modelName:     "llama3.2",
+			host:          "",
+			shouldTrim:    true,
+			expectedLimit: localModelTokenLimit,
+		},
+		{
+			name:          "long output cloud model - uses 10k limit",
+			output:        string(localLimitOutput), // 20k chars, under 10k token limit
+			modelName:     "gpt-4-cloud",
+			host:          "",
+			shouldTrim:    false,
+			expectedLimit: defaultTokenLimit,
+		},
+		{
+			name:          "very long output cloud model - trimmed at 10k",
+			output:        string(defaultLimitOutput),
+			modelName:     "gpt-4-cloud",
+			host:          "",
+			shouldTrim:    true,
+			expectedLimit: defaultTokenLimit,
+		},
+		{
+			name:          "long output remote server - uses 10k limit",
+			output:        string(localLimitOutput),
+			modelName:     "llama3.2",
+			host:          "http://remote.example.com:8080",
+			shouldTrim:    false,
+			expectedLimit: defaultTokenLimit,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Setenv("OLLAMA_HOST", tt.host)
+			result := truncateToolOutput(tt.output, tt.modelName)
+
+			if tt.shouldTrim {
+				maxLen := tt.expectedLimit * charsPerToken
+				if len(result) > maxLen+50 { // +50 for the truncation message
+					t.Errorf("expected output to be truncated to ~%d chars, got %d", maxLen, len(result))
+				}
+				if result == tt.output {
+					t.Error("expected output to be truncated but it wasn't")
+				}
+			} else {
+				if result != tt.output {
+					t.Error("expected output to not be truncated")
+				}
+			}
+		})
+	}
+}

x/tools/registry.go

@@ -3,6 +3,7 @@ package tools
 
 import (
 	"fmt"
+	"os"
 	"sort"
 
 	"github.com/ollama/ollama/api"
@@ -88,9 +89,16 @@ func (r *Registry) Count() int {
 }
 
 // DefaultRegistry creates a registry with all built-in tools.
+// Tools can be disabled via environment variables:
+// - OLLAMA_AGENT_DISABLE_WEBSEARCH=1 disables web_search
+// - OLLAMA_AGENT_DISABLE_BASH=1 disables bash
 func DefaultRegistry() *Registry {
 	r := NewRegistry()
-	r.Register(&WebSearchTool{})
-	r.Register(&BashTool{})
+	if os.Getenv("OLLAMA_AGENT_DISABLE_WEBSEARCH") == "" {
+		r.Register(&WebSearchTool{})
+	}
+	if os.Getenv("OLLAMA_AGENT_DISABLE_BASH") == "" {
+		r.Register(&BashTool{})
+	}
 	return r
 }

x/tools/registry_test.go

@@ -108,6 +108,57 @@ func TestDefaultRegistry(t *testing.T) {
 	}
 }
 
+func TestDefaultRegistry_DisableWebsearch(t *testing.T) {
+	t.Setenv("OLLAMA_AGENT_DISABLE_WEBSEARCH", "1")
+
+	r := DefaultRegistry()
+
+	if r.Count() != 1 {
+		t.Errorf("expected 1 tool with websearch disabled, got %d", r.Count())
+	}
+
+	_, ok := r.Get("bash")
+	if !ok {
+		t.Error("expected bash tool in registry")
+	}
+
+	_, ok = r.Get("web_search")
+	if ok {
+		t.Error("expected web_search to be disabled")
+	}
+}
+
+func TestDefaultRegistry_DisableBash(t *testing.T) {
+	t.Setenv("OLLAMA_AGENT_DISABLE_BASH", "1")
+
+	r := DefaultRegistry()
+
+	if r.Count() != 1 {
+		t.Errorf("expected 1 tool with bash disabled, got %d", r.Count())
+	}
+
+	_, ok := r.Get("web_search")
+	if !ok {
+		t.Error("expected web_search tool in registry")
+	}
+
+	_, ok = r.Get("bash")
+	if ok {
+		t.Error("expected bash to be disabled")
+	}
+}
+
+func TestDefaultRegistry_DisableBoth(t *testing.T) {
+	t.Setenv("OLLAMA_AGENT_DISABLE_WEBSEARCH", "1")
+	t.Setenv("OLLAMA_AGENT_DISABLE_BASH", "1")
+
+	r := DefaultRegistry()
+
+	if r.Count() != 0 {
+		t.Errorf("expected 0 tools with both disabled, got %d", r.Count())
+	}
+}
+
 func TestBashTool_Schema(t *testing.T) {
 	tool := &BashTool{}
 

x/tools/websearch.go

@@ -2,15 +2,19 @@ package tools
 
 import (
 	"bytes"
+	"context"
 	"encoding/json"
+	"errors"
 	"fmt"
 	"io"
 	"net/http"
-	"os"
+	"net/url"
+	"strconv"
 	"strings"
 	"time"
 
 	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/auth"
 )
 
 const (
@@ -18,6 +22,9 @@ const (
 	webSearchTimeout = 15 * time.Second
 )
 
+// ErrWebSearchAuthRequired is returned when web search requires authentication
+var ErrWebSearchAuthRequired = errors.New("web search requires authentication")
+
 // WebSearchTool implements web search using Ollama's hosted API.
 type WebSearchTool struct{}
 
@@ -68,17 +75,13 @@ type webSearchResult struct {
 }
 
 // Execute performs the web search.
+// Uses Ollama key signing for authentication - this makes requests via ollama.com API.
 func (w *WebSearchTool) Execute(args map[string]any) (string, error) {
 	query, ok := args["query"].(string)
 	if !ok || query == "" {
 		return "", fmt.Errorf("query parameter is required")
 	}
 
-	apiKey := os.Getenv("OLLAMA_API_KEY")
-	if apiKey == "" {
-		return "", fmt.Errorf("OLLAMA_API_KEY environment variable is required for web search")
-	}
-
 	// Prepare request
 	reqBody := webSearchRequest{
 		Query:      query,
@@ -90,13 +93,34 @@ func (w *WebSearchTool) Execute(args map[string]any) (string, error) {
 		return "", fmt.Errorf("marshaling request: %w", err)
 	}
 
-	req, err := http.NewRequest("POST", webSearchAPI, bytes.NewBuffer(jsonBody))
+	// Parse URL and add timestamp for signing
+	searchURL, err := url.Parse(webSearchAPI)
+	if err != nil {
+		return "", fmt.Errorf("parsing search URL: %w", err)
+	}
+
+	q := searchURL.Query()
+	q.Add("ts", strconv.FormatInt(time.Now().Unix(), 10))
+	searchURL.RawQuery = q.Encode()
+
+	// Sign the request using Ollama key (~/.ollama/id_ed25519)
+	// This authenticates with ollama.com using the local signing key
+	ctx := context.Background()
+	data := fmt.Appendf(nil, "%s,%s", http.MethodPost, searchURL.RequestURI())
+	signature, err := auth.Sign(ctx, data)
+	if err != nil {
+		return "", fmt.Errorf("signing request: %w", err)
+	}
+
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, searchURL.String(), bytes.NewBuffer(jsonBody))
 	if err != nil {
 		return "", fmt.Errorf("creating request: %w", err)
 	}
 
 	req.Header.Set("Content-Type", "application/json")
-	req.Header.Set("Authorization", "Bearer "+apiKey)
+	if signature != "" {
+		req.Header.Set("Authorization", fmt.Sprintf("Bearer %s", signature))
+	}
 
 	// Send request
 	client := &http.Client{Timeout: webSearchTimeout}
@@ -111,6 +135,9 @@ func (w *WebSearchTool) Execute(args map[string]any) (string, error) {
 		return "", fmt.Errorf("reading response: %w", err)
 	}
 
+	if resp.StatusCode == http.StatusUnauthorized {
+		return "", ErrWebSearchAuthRequired
+	}
 	if resp.StatusCode != http.StatusOK {
 		return "", fmt.Errorf("web search API returned status %d: %s", resp.StatusCode, string(body))
 	}

x/tools/websearch_test.go

@@ -0,0 +1,58 @@
+package tools
+
+import (
+	"errors"
+	"testing"
+)
+
+func TestWebSearchTool_Name(t *testing.T) {
+	tool := &WebSearchTool{}
+	if tool.Name() != "web_search" {
+		t.Errorf("expected name 'web_search', got '%s'", tool.Name())
+	}
+}
+
+func TestWebSearchTool_Description(t *testing.T) {
+	tool := &WebSearchTool{}
+	if tool.Description() == "" {
+		t.Error("expected non-empty description")
+	}
+}
+
+func TestWebSearchTool_Execute_MissingQuery(t *testing.T) {
+	tool := &WebSearchTool{}
+
+	// Test with no query
+	_, err := tool.Execute(map[string]any{})
+	if err == nil {
+		t.Error("expected error for missing query")
+	}
+
+	// Test with empty query
+	_, err = tool.Execute(map[string]any{"query": ""})
+	if err == nil {
+		t.Error("expected error for empty query")
+	}
+}
+
+func TestErrWebSearchAuthRequired(t *testing.T) {
+	// Test that the error type exists and can be checked with errors.Is
+	err := ErrWebSearchAuthRequired
+	if err == nil {
+		t.Fatal("ErrWebSearchAuthRequired should not be nil")
+	}
+
+	if err.Error() != "web search requires authentication" {
+		t.Errorf("unexpected error message: %s", err.Error())
+	}
+
+	// Test that errors.Is works
+	wrappedErr := errors.New("wrapped: " + err.Error())
+	if errors.Is(wrappedErr, ErrWebSearchAuthRequired) {
+		t.Error("wrapped error should not match with errors.Is")
+	}
+
+	if !errors.Is(ErrWebSearchAuthRequired, ErrWebSearchAuthRequired) {
+		t.Error("ErrWebSearchAuthRequired should match itself with errors.Is")
+	}
+}