lmcli/pkg/tui/chat.go

package tui

import (
	"context"
	"encoding/json"
	"fmt"
	"strings"

	cmdutil "git.mlow.ca/mlow/lmcli/pkg/cmd/util"
	models "git.mlow.ca/mlow/lmcli/pkg/lmcli/model"
	tea "github.com/charmbracelet/bubbletea"
	"github.com/charmbracelet/lipgloss"
	"github.com/muesli/reflow/wordwrap"
	"gopkg.in/yaml.v2"
)

type focusState int

const (
	focusInput focusState = iota
	focusMessages
)

type editorTarget int

const (
	input editorTarget = iota
	selectedMessage
)

// custom tea.Msg types
type (
	// sent on each chunk received from LLM
	msgResponseChunk string
	// sent when response is finished being received
	msgResponseEnd string
	// a special case of msgError that stops the response waiting animation
	msgResponseError error
	// sent on each completed reply
	msgAssistantReply models.Message
	// sent when a conversation is (re)loaded
	msgConversationLoaded *models.Conversation
	// sent when a new conversation title is set
	msgConversationTitleChanged string
	// sent when a conversation's messages are laoded
	msgMessagesLoaded []models.Message
)

// styles
var (
	headerStyle = lipgloss.NewStyle().
			PaddingLeft(1).
			PaddingRight(1).
			Background(lipgloss.Color("0"))

	messageHeadingStyle = lipgloss.NewStyle().
				MarginTop(1).
				MarginBottom(1).
				PaddingLeft(1).
				Bold(true)

	userStyle = lipgloss.NewStyle().Faint(true).Foreground(lipgloss.Color("10"))

	assistantStyle = lipgloss.NewStyle().Faint(true).Foreground(lipgloss.Color("12"))

	messageStyle = lipgloss.NewStyle().
			PaddingLeft(2).
			PaddingRight(2)

	inputFocusedStyle = lipgloss.NewStyle().
				Border(lipgloss.RoundedBorder(), true, true, true, false)

	inputBlurredStyle = lipgloss.NewStyle().
				Faint(true).
				Border(lipgloss.RoundedBorder(), true, true, true, false)

	footerStyle = lipgloss.NewStyle()
)

func (m *model) handleChatInput(msg tea.KeyMsg) tea.Cmd {
	switch msg.String() {
	case "esc":
		m.state = stateConversations
		return m.loadConversations()
	case "ctrl+p":
		m.persistence = !m.persistence
	case "ctrl+t":
		m.showToolResults = !m.showToolResults
		m.rebuildMessageCache()
		m.updateContent()
	case "ctrl+w":
		m.wrap = !m.wrap
		m.rebuildMessageCache()
		m.updateContent()
	default:
		switch m.focus {
		case focusInput:
			return m.handleChatInputKey(msg)
		case focusMessages:
			return m.handleChatMessagesKey(msg)
		}
	}
	return nil
}

func (m *model) handleChatUpdate(msg tea.Msg) []tea.Cmd {
	var cmds []tea.Cmd
	switch msg := msg.(type) {
	case msgTempfileEditorClosed:
		contents := string(msg)
		switch m.editorTarget {
		case input:
			m.input.SetValue(contents)
		case selectedMessage:
			m.setMessageContents(m.selectedMessage, contents)
			if m.persistence && m.messages[m.selectedMessage].ID > 0 {
				// update persisted message
				err := m.ctx.Store.UpdateMessage(&m.messages[m.selectedMessage])
				if err != nil {
					cmds = append(cmds, wrapError(fmt.Errorf("Could not save edited message: %v", err)))
				}
			}
			m.updateContent()
		}
	case msgConversationLoaded:
		m.conversation = (*models.Conversation)(msg)
		cmds = append(cmds, m.loadMessages(m.conversation))
	case msgMessagesLoaded:
		m.setMessages(msg)
		m.updateContent()
	case msgResponseChunk:
		chunk := string(msg)
		last := len(m.messages) - 1
		if last >= 0 && m.messages[last].Role.IsAssistant() {
			m.setMessageContents(last, m.messages[last].Content+chunk)
		} else {
			m.addMessage(models.Message{
				Role:    models.MessageRoleAssistant,
				Content: chunk,
			})
		}
		m.updateContent()
		cmds = append(cmds, m.waitForChunk()) // wait for the next chunk
	case msgAssistantReply:
		// the last reply that was being worked on is finished
		reply := models.Message(msg)
		reply.Content = strings.TrimSpace(reply.Content)

		last := len(m.messages) - 1
		if last < 0 {
			panic("Unexpected empty messages handling msgAssistantReply")
		}

		if reply.Role.IsAssistant() && m.messages[last].Role.IsAssistant() {
			// this was a continuation, so replace the previous message with the completed reply
			m.setMessage(last, reply)
		} else {
			m.addMessage(reply)
		}

		if m.persistence {
			var err error
			if m.conversation.ID == 0 {
				err = m.ctx.Store.SaveConversation(m.conversation)
			}
			if err != nil {
				cmds = append(cmds, wrapError(err))
			} else {
				cmds = append(cmds, m.persistConversation())
			}
		}

		if m.conversation.Title == "" {
			cmds = append(cmds, m.generateConversationTitle())
		}

		m.updateContent()
		cmds = append(cmds, m.waitForReply())
	case msgResponseEnd:
		m.waitingForReply = false
		last := len(m.messages) - 1
		if last < 0 {
			panic("Unexpected empty messages handling msgResponseEnd")
		}
		m.setMessageContents(last, strings.TrimSpace(m.messages[last].Content))
		m.updateContent()
		m.status = "Press ctrl+s to send"
	case msgResponseError:
		m.waitingForReply = false
		m.status = "Press ctrl+s to send"
		m.err = error(msg)
	case msgConversationTitleChanged:
		title := string(msg)
		m.conversation.Title = title
		if m.persistence {
			err := m.ctx.Store.SaveConversation(m.conversation)
			if err != nil {
				cmds = append(cmds, wrapError(err))
			}
		}
	}

	var cmd tea.Cmd
	m.spinner, cmd = m.spinner.Update(msg)
	if cmd != nil {
		cmds = append(cmds, cmd)
	}

	prevInputLineCnt := m.input.LineCount()
	inputCaptured := false
	m.input, cmd = m.input.Update(msg)
	if cmd != nil {
		inputCaptured = true
		cmds = append(cmds, cmd)
	}

	if !inputCaptured {
		m.content, cmd = m.content.Update(msg)
		if cmd != nil {
			cmds = append(cmds, cmd)
		}
	}

	// update views once window dimensions are known
	if m.width > 0 {
		m.views.header = m.headerView()
		m.views.footer = m.footerView()
		m.views.error = m.errorView()
		fixedHeight := height(m.views.header) + height(m.views.error) + height(m.views.footer)

		// calculate clamped input height to accomodate input text
		newHeight := max(4, min((m.height-fixedHeight-1)/2, m.input.LineCount()))
		m.input.SetHeight(newHeight)
		m.views.input = m.input.View()

		m.content.Height = m.height - fixedHeight - height(m.views.input)
		m.views.content = m.content.View()
	}

	// this is a pretty nasty hack to ensure the input area viewport doesn't
	// scroll below its content, which can happen when the input viewport
	// height has grown, or previously entered lines have been deleted
	if prevInputLineCnt != m.input.LineCount() {
		// dist is the distance we'd need to scroll up from the current cursor
		// position to position the last input line at the bottom of the
		// viewport. if negative, we're already scrolled above the bottom
		dist := m.input.Line() - (m.input.LineCount() - m.input.Height())
		if dist > 0 {
			for i := 0; i < dist; i++ {
				// move cursor up until content reaches the bottom of the viewport
				m.input.CursorUp()
			}
			m.input, cmd = m.input.Update(nil)
			for i := 0; i < dist; i++ {
				// move cursor back down to its previous position
				m.input.CursorDown()
			}
			m.input, cmd = m.input.Update(nil)
		}
	}

	return cmds
}

func (m *model) handleChatMessagesKey(msg tea.KeyMsg) tea.Cmd {
	switch msg.String() {
	case "tab":
		m.focus = focusInput
		m.updateContent()
		m.input.Focus()
	case "e":
		message := m.messages[m.selectedMessage]
		cmd := openTempfileEditor("message.*.md", message.Content, "# Edit the message below\n")
		m.editorTarget = selectedMessage
		return cmd
	case "ctrl+k":
		if m.selectedMessage > 0 && len(m.messages) == len(m.messageOffsets) {
			m.selectedMessage--
			m.updateContent()
			offset := m.messageOffsets[m.selectedMessage]
			scrollIntoView(&m.content, offset, 0.1)
		}
	case "ctrl+j":
		if m.selectedMessage < len(m.messages)-1 && len(m.messages) == len(m.messageOffsets) {
			m.selectedMessage++
			m.updateContent()
			offset := m.messageOffsets[m.selectedMessage]
			scrollIntoView(&m.content, offset, 0.1)
		}
	case "ctrl+r":
		// resubmit the conversation with all messages up until and including the selected message
		if m.waitingForReply || len(m.messages) == 0 {
			return nil
		}
		m.messages = m.messages[:m.selectedMessage+1]
		m.messageCache = m.messageCache[:m.selectedMessage+1]
		m.updateContent()
		m.content.GotoBottom()
		return m.promptLLM()
	}
	return nil
}

func (m *model) handleChatInputKey(msg tea.KeyMsg) tea.Cmd {
	switch msg.String() {
	case "esc":
		m.focus = focusMessages
		if len(m.messages) > 0 {
			if m.selectedMessage < 0 || m.selectedMessage >= len(m.messages) {
				m.selectedMessage = len(m.messages) - 1
			}
			offset := m.messageOffsets[m.selectedMessage]
			scrollIntoView(&m.content, offset, 0.1)
		}
		m.updateContent()
		m.input.Blur()
	case "ctrl+s":
		userInput := strings.TrimSpace(m.input.Value())
		if strings.TrimSpace(userInput) == "" {
			return nil
		}

		if len(m.messages) > 0 && m.messages[len(m.messages)-1].Role == models.MessageRoleUser {
			return wrapError(fmt.Errorf("Can't reply to a user message"))
		}

		reply := models.Message{
			Role:    models.MessageRoleUser,
			Content: userInput,
		}

		if m.persistence {
			var err error
			if m.conversation.ID == 0 {
				err = m.ctx.Store.SaveConversation(m.conversation)
			}
			if err != nil {
				return wrapError(err)
			}

			// ensure all messages up to the one we're about to add are persisted
			cmd := m.persistConversation()
			if cmd != nil {
				return cmd
			}

			savedReply, err := m.ctx.Store.AddReply(m.conversation, reply)
			if err != nil {
				return wrapError(err)
			}
			reply = *savedReply
		}

		m.input.SetValue("")
		m.addMessage(reply)

		m.updateContent()
		m.content.GotoBottom()
		return m.promptLLM()
	case "ctrl+e":
		cmd := openTempfileEditor("message.*.md", m.input.Value(), "# Edit your input below\n")
		m.editorTarget = input
		return cmd
	}
	return nil
}

func (m *model) renderMessageHeading(i int, message *models.Message) string {
	icon := ""
	friendly := message.Role.FriendlyRole()
	style := lipgloss.NewStyle().Faint(true).Bold(true)

	switch message.Role {
	case models.MessageRoleSystem:
		icon = "⚙️"
	case models.MessageRoleUser:
		style = userStyle
	case models.MessageRoleAssistant:
		style = assistantStyle
	case models.MessageRoleToolCall:
		style = assistantStyle
		friendly = models.MessageRoleAssistant.FriendlyRole()
	case models.MessageRoleToolResult:
		icon = "🔧"
	}

	user := style.Render(icon + friendly)

	var prefix string
	var suffix string

	faint := lipgloss.NewStyle().Faint(true)
	if m.focus == focusMessages {
		if i == m.selectedMessage {
			prefix = "> "
		}
	}

	if message.ID == 0 {
		suffix += faint.Render(" (not saved)")
	}

	return messageHeadingStyle.Render(prefix + user + suffix)
}

func (m *model) renderMessage(msg *models.Message) string {
	sb := &strings.Builder{}
	sb.Grow(len(msg.Content) * 2)
	if msg.Content != "" {
		err := m.ctx.Chroma.Highlight(sb, msg.Content)
		if err != nil {
			sb.Reset()
			sb.WriteString(msg.Content)
		}
	}

	var toolString string
	switch msg.Role {
	case models.MessageRoleToolCall:
		bytes, err := yaml.Marshal(msg.ToolCalls)
		if err != nil {
			toolString = "Could not serialize ToolCalls"
		} else {
			toolString = "tool_calls:\n" + string(bytes)
		}
	case models.MessageRoleToolResult:
		if !m.showToolResults {
			break
		}

		type renderedResult struct {
			ToolName string `yaml:"tool"`
			Result   any
		}

		var toolResults []renderedResult
		for _, result := range msg.ToolResults {
			var jsonResult interface{}
			err := json.Unmarshal([]byte(result.Result), &jsonResult)
			if err != nil {
				// If parsing as JSON fails, treat Result as a plain string
				toolResults = append(toolResults, renderedResult{
					ToolName: result.ToolName,
					Result:   result.Result,
				})
			} else {
				// If parsing as JSON succeeds, marshal the parsed JSON into YAML
				toolResults = append(toolResults, renderedResult{
					ToolName: result.ToolName,
					Result:   &jsonResult,
				})
			}
		}

		bytes, err := yaml.Marshal(toolResults)
		if err != nil {
			toolString = "Could not serialize ToolResults"
		} else {
			toolString = "tool_results:\n" + string(bytes)
		}
	}

	if toolString != "" {
		toolString = strings.TrimRight(toolString, "\n")
		if msg.Content != "" {
			sb.WriteString("\n\n")
		}
		_ = m.ctx.Chroma.HighlightLang(sb, toolString, "yaml")
	}

	content := strings.TrimRight(sb.String(), "\n")

	if m.wrap {
		wrapWidth := m.content.Width - messageStyle.GetHorizontalPadding() - 1
		content = wordwrap.String(content, wrapWidth)
	}

	return messageStyle.Width(0).Render(content)
}

// render the conversation into a string
func (m *model) conversationMessagesView() string {
	sb := strings.Builder{}

	m.messageOffsets = make([]int, len(m.messages))
	lineCnt := 1
	for i, message := range m.messages {
		m.messageOffsets[i] = lineCnt

		switch message.Role {
		case models.MessageRoleToolCall:
			if !m.showToolResults && message.Content == "" {
				continue
			}
		case models.MessageRoleToolResult:
			if !m.showToolResults {
				continue
			}
		}

		heading := m.renderMessageHeading(i, &message)
		sb.WriteString(heading)
		sb.WriteString("\n")
		lineCnt += lipgloss.Height(heading)

		cached := m.messageCache[i]
		sb.WriteString(cached)
		sb.WriteString("\n")
		lineCnt += lipgloss.Height(cached)
	}

	return sb.String()
}

func (m *model) setMessages(messages []models.Message) {
	m.messages = messages
	m.rebuildMessageCache()
}

func (m *model) setMessage(i int, msg models.Message) {
	if i >= len(m.messages) {
		panic("i out of range")
	}
	m.messages[i] = msg
	m.messageCache[i] = m.renderMessage(&msg)
}

func (m *model) addMessage(msg models.Message) {
	m.messages = append(m.messages, msg)
	m.messageCache = append(m.messageCache, m.renderMessage(&msg))
}

func (m *model) setMessageContents(i int, content string) {
	if i >= len(m.messages) {
		panic("i out of range")
	}
	m.messages[i].Content = content
	m.messageCache[i] = m.renderMessage(&m.messages[i])
}

func (m *model) rebuildMessageCache() {
	m.messageCache = make([]string, len(m.messages))
	for i, msg := range m.messages {
		m.messageCache[i] = m.renderMessage(&msg)
	}
}

func (m *model) updateContent() {
	atBottom := m.content.AtBottom()
	m.content.SetContent(m.conversationMessagesView())
	if atBottom {
		// if we were at bottom before the update, scroll with the output
		m.content.GotoBottom()
	}
}

func (m *model) loadConversation(shortname string) tea.Cmd {
	return func() tea.Msg {
		if shortname == "" {
			return nil
		}
		c, err := m.ctx.Store.ConversationByShortName(shortname)
		if err != nil {
			return msgError(fmt.Errorf("Could not lookup conversation: %v", err))
		}
		if c.ID == 0 {
			return msgError(fmt.Errorf("Conversation not found: %s", shortname))
		}
		return msgConversationLoaded(c)
	}
}

func (m *model) loadMessages(c *models.Conversation) tea.Cmd {
	return func() tea.Msg {
		messages, err := m.ctx.Store.Messages(c)
		if err != nil {
			return msgError(fmt.Errorf("Could not load conversation messages: %v\n", err))
		}
		return msgMessagesLoaded(messages)
	}
}

func (m *model) persistConversation() tea.Cmd {
	existingMessages, err := m.ctx.Store.Messages(m.conversation)
	if err != nil {
		return wrapError(fmt.Errorf("Could not retrieve existing conversation messages while trying to save: %v", err))
	}

	existingById := make(map[uint]*models.Message, len(existingMessages))
	for _, msg := range existingMessages {
		existingById[msg.ID] = &msg
	}

	currentById := make(map[uint]*models.Message, len(m.messages))
	for _, msg := range m.messages {
		currentById[msg.ID] = &msg
	}

	for _, msg := range existingMessages {
		_, ok := currentById[msg.ID]
		if !ok {
			err := m.ctx.Store.DeleteMessage(&msg)
			if err != nil {
				return wrapError(fmt.Errorf("Failed to remove messages: %v", err))
			}
		}
	}

	for i, msg := range m.messages {
		if msg.ID > 0 {
			exist, ok := existingById[msg.ID]
			if ok {
				if msg.Content == exist.Content {
					continue
				}
				// update message when contents don't match that of store
				err := m.ctx.Store.UpdateMessage(&msg)
				if err != nil {
					return wrapError(err)
				}
			} else {
				// this would be quite odd... and I'm not sure how to handle
				// it at the time of writing this
			}
		} else {
			newMessage, err := m.ctx.Store.AddReply(m.conversation, msg)
			if err != nil {
				return wrapError(err)
			}
			m.setMessage(i, *newMessage)
		}
	}
	return nil
}

func (m *model) generateConversationTitle() tea.Cmd {
	return func() tea.Msg {
		title, err := cmdutil.GenerateTitle(m.ctx, m.conversation)
		if err != nil {
			return msgError(err)
		}
		return msgConversationTitleChanged(title)
	}
}

func (m *model) waitForReply() tea.Cmd {
	return func() tea.Msg {
		return msgAssistantReply(<-m.replyChan)
	}
}

func (m *model) waitForChunk() tea.Cmd {
	return func() tea.Msg {
		return msgResponseChunk(<-m.replyChunkChan)
	}
}


func (m *model) promptLLM() tea.Cmd {
	m.waitingForReply = true
	m.status = "Press ctrl+c to cancel"

	return func() tea.Msg {
		completionProvider, err := m.ctx.GetCompletionProvider(*m.ctx.Config.Defaults.Model)
		if err != nil {
			return msgError(err)
		}

		requestParams := models.RequestParameters{
			Model:       *m.ctx.Config.Defaults.Model,
			MaxTokens:   *m.ctx.Config.Defaults.MaxTokens,
			Temperature: *m.ctx.Config.Defaults.Temperature,
			ToolBag:     m.ctx.EnabledTools,
		}

		replyHandler := func(msg models.Message) {
			m.replyChan <- msg
		}

		ctx, cancel := context.WithCancel(context.Background())

		canceled := false
		go func() {
			select {
			case <-m.stopSignal:
				canceled = true
				cancel()
			}
		}()

		resp, err := completionProvider.CreateChatCompletionStream(
			ctx, requestParams, m.messages, replyHandler, m.replyChunkChan,
		)

		if err != nil && !canceled {
			return msgResponseError(err)
		}

		return msgResponseEnd(resp)
	}
}