Enha: filter out thinking blocks from chat history, removed {role}:

bot.go

@@ -1087,7 +1087,15 @@ func chatToTextSlice(messages []models.RoleMsg, showSys bool) []string {
 
 func chatToText(messages []models.RoleMsg, showSys bool) string {
 	s := chatToTextSlice(messages, showSys)
-	return strings.Join(s, "\n")
+	text := strings.Join(s, "\n")
+
+	// Collapse thinking blocks if enabled
+	if thinkingCollapsed {
+		placeholder := "[yellow::i][thinking... (press Alt+T to expand)][-:-:-]"
+		text = thinkRE.ReplaceAllString(text, placeholder)
+	}
+
+	return text
 }
 
 func removeThinking(chatBody *models.ChatBody) {

config.example.toml

@@ -48,3 +48,4 @@ EnableMouse = false # Enable mouse support in the UI
 CharSpecificContextEnabled = true
 CharSpecificContextTag = "@"
 AutoTurn = true
+StripThinkingFromAPI = true  # Strip <think> blocks from messages before sending to LLM (keeps them in chat history)

config/config.go

@@ -19,6 +19,7 @@ type Config struct {
 	ToolRole                      string `toml:"ToolRole"`
 	ToolUse                       bool   `toml:"ToolUse"`
 	ThinkUse                      bool   `toml:"ThinkUse"`
+	StripThinkingFromAPI          bool   `toml:"StripThinkingFromAPI"`
 	AssistantRole                 string `toml:"AssistantRole"`
 	SysDir                        string `toml:"SysDir"`
 	ChunkLimit                    uint32 `toml:"ChunkLimit"`

helpfuncs.go

@@ -23,6 +23,25 @@ func isASCII(s string) bool {
 	return true
 }
 
+// stripThinkingFromMsg removes thinking blocks from assistant messages.
+// Skips user, tool, and system messages as they may contain thinking examples.
+func stripThinkingFromMsg(msg models.RoleMsg) *models.RoleMsg {
+	if !cfg.StripThinkingFromAPI {
+		return &msg
+	}
+	// Skip user, tool, and system messages - they might contain thinking examples
+	if msg.Role == cfg.UserRole || msg.Role == cfg.ToolRole || msg.Role == "system" {
+		return &msg
+	}
+	// Strip thinking from assistant messages
+	if thinkRE.MatchString(msg.Content) {
+		msg.Content = thinkRE.ReplaceAllString(msg.Content, "")
+		// Clean up any double newlines that might result
+		msg.Content = strings.TrimSpace(msg.Content)
+	}
+	return &msg
+}
+
 // refreshChatDisplay updates the chat display based on current character view
 // It filters messages for the character the user is currently "writing as"
 // and updates the textView with the filtered conversation

llm.go

@@ -13,28 +13,6 @@ var imageAttachmentPath string // Global variable to track image attachment for
 var lastImg string             // for ctrl+j
 var RAGMsg = "Retrieved context for user's query:\n"
 
-// addPersonaSuffixToLastUserMessage adds the persona suffix to the last user message
-// to indicate to the assistant who it should reply as
-func addPersonaSuffixToLastUserMessage(messages []models.RoleMsg, persona string) []models.RoleMsg {
-	if len(messages) == 0 {
-		return messages
-	}
-	// // Find the last user message to modify
-	// for i := len(messages) - 1; i >= 0; i-- {
-	// 	if messages[i].Role == cfg.UserRole || messages[i].Role == "user" {
-	// 		// Create a copy of the message to avoid modifying the original
-	// 		modifiedMsg := messages[i]
-	// 		modifiedMsg.Content = modifiedMsg.Content + "\n" + persona + ":"
-	// 		messages[i] = modifiedMsg
-	// 		break
-	// 	}
-	// }
-	modifiedMsg := messages[len(messages)-1]
-	modifiedMsg.Content = modifiedMsg.Content + "\n" + persona + ":\n"
-	messages[len(messages)-1] = modifiedMsg
-	return messages
-}
-
 // containsToolSysMsg checks if the toolSysMsg already exists in the chat body
 func containsToolSysMsg() bool {
 	for _, msg := range chatBody.Messages {
@@ -187,17 +165,9 @@ func (lcp LCPCompletion) FormMsg(msg, role string, resume bool) (io.Reader, erro
 	filteredMessages, botPersona := filterMessagesForCurrentCharacter(chatBody.Messages)
 	messages := make([]string, len(filteredMessages))
 	for i, m := range filteredMessages {
-		messages[i] = m.ToPrompt()
+		messages[i] = stripThinkingFromMsg(m).ToPrompt()
 	}
 	prompt := strings.Join(messages, "\n")
-	// strings builder?
-	if !resume {
-		botMsgStart := "\n" + botPersona + ":\n"
-		prompt += botMsgStart
-	}
-	if cfg.ThinkUse && !cfg.ToolUse {
-		prompt += "<think>"
-	}
 	// Add multimodal media markers to the prompt text when multimodal data is present
 	// This is required by llama.cpp multimodal models so they know where to insert media
 	if len(multimodalData) > 0 {
@@ -209,6 +179,14 @@ func (lcp LCPCompletion) FormMsg(msg, role string, resume bool) (io.Reader, erro
 		}
 		prompt = sb.String()
 	}
+	// needs to be after <__media__> if there are images
+	if !resume {
+		botMsgStart := "\n" + botPersona + ":\n"
+		prompt += botMsgStart
+	}
+	if cfg.ThinkUse && !cfg.ToolUse {
+		prompt += "<think>"
+	}
 	logger.Debug("checking prompt for /completion", "tool_use", cfg.ToolUse,
 		"msg", msg, "resume", resume, "prompt", prompt, "multimodal_data_count", len(multimodalData))
 	payload := models.NewLCPReq(prompt, chatBody.Model, multimodalData,
@@ -341,23 +319,21 @@ func (op LCPChat) FormMsg(msg, role string, resume bool) (io.Reader, error) {
 		logger.Debug("LCPChat: RAG message added to chat body", "role", ragMsg.Role,
 			"rag_content_len", len(ragMsg.Content), "message_count_after_rag", len(chatBody.Messages))
 	}
-	filteredMessages, botPersona := filterMessagesForCurrentCharacter(chatBody.Messages)
+	filteredMessages, _ := filterMessagesForCurrentCharacter(chatBody.Messages)
 	// openai /v1/chat does not support custom roles; needs to be user, assistant, system
 	// Add persona suffix to the last user message to indicate who the assistant should reply as
-	if cfg.AutoTurn && !resume {
-		filteredMessages = addPersonaSuffixToLastUserMessage(filteredMessages, botPersona)
-	}
 	bodyCopy := &models.ChatBody{
 		Messages: make([]models.RoleMsg, len(filteredMessages)),
 		Model:    chatBody.Model,
 		Stream:   chatBody.Stream,
 	}
 	for i, msg := range filteredMessages {
-		if msg.Role == cfg.UserRole {
-			bodyCopy.Messages[i] = msg
+		strippedMsg := *stripThinkingFromMsg(msg)
+		if strippedMsg.Role == cfg.UserRole {
+			bodyCopy.Messages[i] = strippedMsg
 			bodyCopy.Messages[i].Role = "user"
 		} else {
-			bodyCopy.Messages[i] = msg
+			bodyCopy.Messages[i] = strippedMsg
 		}
 	}
 	// Clean null/empty messages to prevent API issues
@@ -437,7 +413,7 @@ func (ds DeepSeekerCompletion) FormMsg(msg, role string, resume bool) (io.Reader
 	filteredMessages, botPersona := filterMessagesForCurrentCharacter(chatBody.Messages)
 	messages := make([]string, len(filteredMessages))
 	for i, m := range filteredMessages {
-		messages[i] = m.ToPrompt()
+		messages[i] = stripThinkingFromMsg(m).ToPrompt()
 	}
 	prompt := strings.Join(messages, "\n")
 	// strings builder?
@@ -519,22 +495,20 @@ func (ds DeepSeekerChat) FormMsg(msg, role string, resume bool) (io.Reader, erro
 		logger.Debug("RAG message added to chat body", "message_count", len(chatBody.Messages))
 	}
 	// Create copy of chat body with standardized user role
-	filteredMessages, botPersona := filterMessagesForCurrentCharacter(chatBody.Messages)
+	filteredMessages, _ := filterMessagesForCurrentCharacter(chatBody.Messages)
 	// Add persona suffix to the last user message to indicate who the assistant should reply as
-	if cfg.AutoTurn && !resume {
-		filteredMessages = addPersonaSuffixToLastUserMessage(filteredMessages, botPersona)
-	}
 	bodyCopy := &models.ChatBody{
 		Messages: make([]models.RoleMsg, len(filteredMessages)),
 		Model:    chatBody.Model,
 		Stream:   chatBody.Stream,
 	}
 	for i, msg := range filteredMessages {
-		if msg.Role == cfg.UserRole || i == 1 {
-			bodyCopy.Messages[i] = msg
+		strippedMsg := *stripThinkingFromMsg(msg)
+		if strippedMsg.Role == cfg.UserRole || i == 1 {
+			bodyCopy.Messages[i] = strippedMsg
 			bodyCopy.Messages[i].Role = "user"
 		} else {
-			bodyCopy.Messages[i] = msg
+			bodyCopy.Messages[i] = strippedMsg
 		}
 	}
 	// Clean null/empty messages to prevent API issues
@@ -605,7 +579,7 @@ func (or OpenRouterCompletion) FormMsg(msg, role string, resume bool) (io.Reader
 	filteredMessages, botPersona := filterMessagesForCurrentCharacter(chatBody.Messages)
 	messages := make([]string, len(filteredMessages))
 	for i, m := range filteredMessages {
-		messages[i] = m.ToPrompt()
+		messages[i] = stripThinkingFromMsg(m).ToPrompt()
 	}
 	prompt := strings.Join(messages, "\n")
 	// strings builder?
@@ -718,21 +692,19 @@ func (or OpenRouterChat) FormMsg(msg, role string, resume bool) (io.Reader, erro
 		logger.Debug("RAG message added to chat body", "message_count", len(chatBody.Messages))
 	}
 	// Create copy of chat body with standardized user role
-	filteredMessages, botPersona := filterMessagesForCurrentCharacter(chatBody.Messages)
+	filteredMessages, _ := filterMessagesForCurrentCharacter(chatBody.Messages)
 	// Add persona suffix to the last user message to indicate who the assistant should reply as
-	if cfg.AutoTurn && !resume {
-		filteredMessages = addPersonaSuffixToLastUserMessage(filteredMessages, botPersona)
-	}
 	bodyCopy := &models.ChatBody{
 		Messages: make([]models.RoleMsg, len(filteredMessages)),
 		Model:    chatBody.Model,
 		Stream:   chatBody.Stream,
 	}
 	for i, msg := range filteredMessages {
-		bodyCopy.Messages[i] = msg
+		strippedMsg := *stripThinkingFromMsg(msg)
+		bodyCopy.Messages[i] = strippedMsg
 		// Standardize role if it's a user role
 		if bodyCopy.Messages[i].Role == cfg.UserRole {
-			bodyCopy.Messages[i] = msg
+			bodyCopy.Messages[i] = strippedMsg
 			bodyCopy.Messages[i].Role = "user"
 		}
 	}

main.go

@@ -12,6 +12,7 @@ var (
 	injectRole          = true
 	selectedIndex       = int(-1)
 	shellMode           = false
+	thinkingCollapsed   = false
 	indexLineCompletion = "F12 to show keys help | llm turn: [%s:-:b]%v[-:-:-] (F6) | chat: [orange:-:b]%s[-:-:-] (F1) | toolUseAdviced: [%s:-:b]%v[-:-:-] (ctrl+k) | model: [orange:-:b]%s[-:-:-] (ctrl+l) | skip LLM resp: [%s:-:b]%v[-:-:-] (F10)\nAPI: [orange:-:b]%s[-:-:-] (ctrl+v) | recording: [%s:-:b]%v[-:-:-] (ctrl+r) | writing as: [orange:-:b]%s[-:-:-] (ctrl+q) | bot will write as [orange:-:b]%s[-:-:-] (ctrl+x) | role injection (alt+7) [%s:-:b]%v[-:-:-]"
 	focusSwitcher       = map[tview.Primitive]tview.Primitive{}
 )

tui.go

@@ -96,6 +96,7 @@ var (
 [yellow]Alt+7[white]: toggle role injection (inject role in messages)
 [yellow]Alt+8[white]: show char img or last picked img
 [yellow]Alt+9[white]: warm up (load) selected llama.cpp model
+[yellow]Alt+t[white]: toggle thinking blocks visibility (collapse/expand <think> blocks)
 
 === scrolling chat window (some keys similar to vim) ===
 [yellow]arrows up/down and j/k[white]: scroll up and down
@@ -831,6 +832,20 @@ func init() {
 			injectRole = !injectRole
 			updateStatusLine()
 		}
+		// Handle Alt+T to toggle thinking block visibility
+		if event.Key() == tcell.KeyRune && event.Rune() == 't' && event.Modifiers()&tcell.ModAlt != 0 {
+			thinkingCollapsed = !thinkingCollapsed
+			textView.SetText(chatToText(chatBody.Messages, cfg.ShowSys))
+			colorText()
+			status := "expanded"
+			if thinkingCollapsed {
+				status = "collapsed"
+			}
+			if err := notifyUser("thinking", fmt.Sprintf("Thinking blocks %s", status)); err != nil {
+				logger.Error("failed to send notification", "error", err)
+			}
+			return nil
+		}
 		if event.Key() == tcell.KeyF1 {
 			// chatList, err := loadHistoryChats()
 			chatList, err := store.GetChatByChar(cfg.AssistantRole)