Fix: user messages copied without content

5 files changed, 200 insertions, 10 deletions

diff --git a/Makefile b/Makefile
index 6f0c7bb..b845e6f 100644
--- a/Makefile
+++ b/Makefile
@@ -30,7 +30,7 @@ download-whisper-model: ## Download Whisper model for STT in batteries directory
 		echo "Please run 'make setup-whisper' first to clone the repository."; \
 		exit 1; \
 	fi
-	@cd batteries/whisper.cpp && make large-v3-turbo
+	@cd batteries/whisper.cpp && bash ./models/download-ggml-model.sh large-v3-turbo-q5_0
 	@echo "Whisper model downloaded successfully!"
 
 # Docker targets for STT/TTS services (in batteries directory)
diff --git a/bot.go b/bot.go
index 6c1098e..0001afa 100644
--- a/bot.go
+++ b/bot.go
@@ -74,6 +74,9 @@ func cleanNullMessages(messages []models.RoleMsg) []models.RoleMsg {
 		// Include message if it has content or if it's a tool response (which might have tool_call_id)
 		if msg.HasContent() || msg.ToolCallID != "" {
 			cleaned = append(cleaned, msg)
+		} else {
+			// Log filtered messages for debugging
+			logger.Warn("filtering out message during cleaning", "role", msg.Role, "content", msg.Content, "tool_call_id", msg.ToolCallID, "has_content", msg.HasContent())
 		}
 	}
 	return consolidateConsecutiveAssistantMessages(cleaned)
@@ -103,9 +106,12 @@ func consolidateConsecutiveAssistantMessages(messages []models.RoleMsg) []models
 				if currentAssistantMsg.IsContentParts() || msg.IsContentParts() {
 					// Handle structured content
 					if !currentAssistantMsg.IsContentParts() {
+						// Preserve the original ToolCallID before conversion
+						originalToolCallID := currentAssistantMsg.ToolCallID
 						// Convert existing content to content parts
 						currentAssistantMsg = models.NewMultimodalMsg(currentAssistantMsg.Role, []interface{}{models.TextContentPart{Type: "text", Text: currentAssistantMsg.Content}})
-						currentAssistantMsg.ToolCallID = msg.ToolCallID
+						// Restore the original ToolCallID to preserve tool call linking
+						currentAssistantMsg.ToolCallID = originalToolCallID
 					}
 					if msg.IsContentParts() {
 						currentAssistantMsg.ContentParts = append(currentAssistantMsg.ContentParts, msg.GetContentParts()...)
@@ -119,6 +125,7 @@ func consolidateConsecutiveAssistantMessages(messages []models.RoleMsg) []models
 					} else {
 						currentAssistantMsg.Content = msg.Content
 					}
+					// ToolCallID is already preserved since we're not creating a new message object when just concatenating content
 				}
 			}
 		} else {
@@ -556,9 +563,19 @@ out:
 		})
 	}
 
+	logger.Debug("chatRound: before cleanChatBody", "messages_before_clean", len(chatBody.Messages))
+	for i, msg := range chatBody.Messages {
+		logger.Debug("chatRound: before cleaning", "index", i, "role", msg.Role, "content_len", len(msg.Content), "has_content", msg.HasContent(), "tool_call_id", msg.ToolCallID)
+	}
+
 	// Clean null/empty messages to prevent API issues with endpoints like llama.cpp jinja template
 	cleanChatBody()
 
+	logger.Debug("chatRound: after cleanChatBody", "messages_after_clean", len(chatBody.Messages))
+	for i, msg := range chatBody.Messages {
+		logger.Debug("chatRound: after cleaning", "index", i, "role", msg.Role, "content_len", len(msg.Content), "has_content", msg.HasContent(), "tool_call_id", msg.ToolCallID)
+	}
+
 	colorText()
 	updateStatusLine()
 	// bot msg is done;
@@ -574,8 +591,17 @@ out:
 func cleanChatBody() {
 	if chatBody != nil && chatBody.Messages != nil {
 		originalLen := len(chatBody.Messages)
+		logger.Debug("cleanChatBody: before cleaning", "message_count", originalLen)
+		for i, msg := range chatBody.Messages {
+			logger.Debug("cleanChatBody: before clean", "index", i, "role", msg.Role, "content_len", len(msg.Content), "has_content", msg.HasContent(), "tool_call_id", msg.ToolCallID)
+		}
+
 		chatBody.Messages = cleanNullMessages(chatBody.Messages)
-		logger.Debug("cleaned chat body", "original_len", originalLen, "new_len", len(chatBody.Messages))
+
+		logger.Debug("cleanChatBody: after cleaning", "original_len", originalLen, "new_len", len(chatBody.Messages))
+		for i, msg := range chatBody.Messages {
+			logger.Debug("cleanChatBody: after clean", "index", i, "role", msg.Role, "content_len", len(msg.Content), "has_content", msg.HasContent(), "tool_call_id", msg.ToolCallID)
+		}
 	}
 }
 
@@ -621,6 +647,7 @@ func findCall(msg, toolCall string, tv *tview.TextView) {
 				Content: fmt.Sprintf("Error processing tool call: %v. Please check the JSON format and try again.", err),
 			}
 			chatBody.Messages = append(chatBody.Messages, toolResponseMsg)
+			logger.Debug("findCall: added tool error response", "role", toolResponseMsg.Role, "content_len", len(toolResponseMsg.Content), "message_count_after_add", len(chatBody.Messages))
 			// Trigger the assistant to continue processing with the error message
 			chatRound("", cfg.AssistantRole, tv, false, false)
 			return
@@ -637,6 +664,7 @@ func findCall(msg, toolCall string, tv *tview.TextView) {
 			ToolCallID: lastToolCallID, // Use the stored tool call ID
 		}
 		chatBody.Messages = append(chatBody.Messages, toolResponseMsg)
+		logger.Debug("findCall: added tool not implemented response", "role", toolResponseMsg.Role, "content_len", len(toolResponseMsg.Content), "tool_call_id", toolResponseMsg.ToolCallID, "message_count_after_add", len(chatBody.Messages))
 		// Clear the stored tool call ID after using it
 		lastToolCallID = ""
 
@@ -657,6 +685,7 @@ func findCall(msg, toolCall string, tv *tview.TextView) {
 		ToolCallID: lastToolCallID, // Use the stored tool call ID
 	}
 	chatBody.Messages = append(chatBody.Messages, toolResponseMsg)
+	logger.Debug("findCall: added actual tool response", "role", toolResponseMsg.Role, "content_len", len(toolResponseMsg.Content), "tool_call_id", toolResponseMsg.ToolCallID, "message_count_after_add", len(chatBody.Messages))
 	// Clear the stored tool call ID after using it
 	lastToolCallID = ""
 	// Trigger the assistant to continue processing with the new tool response
diff --git a/bot_test.go b/bot_test.go
new file mode 100644
index 0000000..2d59c3c
--- /dev/null
+++ b/bot_test.go
@@ -0,0 +1,155 @@
+package main
+
+import (
+	"gf-lt/config"
+	"gf-lt/models"
+	"reflect"
+	"testing"
+)
+
+func TestConsolidateConsecutiveAssistantMessages(t *testing.T) {
+	// Mock config for testing
+	testCfg := &config.Config{
+		AssistantRole:                 "assistant",
+		WriteNextMsgAsCompletionAgent: "",
+	}
+	cfg = testCfg
+
+	tests := []struct {
+		name     string
+		input    []models.RoleMsg
+		expected []models.RoleMsg
+	}{
+		{
+			name: "no consecutive assistant messages",
+			input: []models.RoleMsg{
+				{Role: "user", Content: "Hello"},
+				{Role: "assistant", Content: "Hi there"},
+				{Role: "user", Content: "How are you?"},
+			},
+			expected: []models.RoleMsg{
+				{Role: "user", Content: "Hello"},
+				{Role: "assistant", Content: "Hi there"},
+				{Role: "user", Content: "How are you?"},
+			},
+		},
+		{
+			name: "consecutive assistant messages should be consolidated",
+			input: []models.RoleMsg{
+				{Role: "user", Content: "Hello"},
+				{Role: "assistant", Content: "First part"},
+				{Role: "assistant", Content: "Second part"},
+				{Role: "user", Content: "Thanks"},
+			},
+			expected: []models.RoleMsg{
+				{Role: "user", Content: "Hello"},
+				{Role: "assistant", Content: "First part\nSecond part"},
+				{Role: "user", Content: "Thanks"},
+			},
+		},
+		{
+			name: "multiple sets of consecutive assistant messages",
+			input: []models.RoleMsg{
+				{Role: "user", Content: "First question"},
+				{Role: "assistant", Content: "First answer part 1"},
+				{Role: "assistant", Content: "First answer part 2"},
+				{Role: "user", Content: "Second question"},
+				{Role: "assistant", Content: "Second answer part 1"},
+				{Role: "assistant", Content: "Second answer part 2"},
+				{Role: "assistant", Content: "Second answer part 3"},
+			},
+			expected: []models.RoleMsg{
+				{Role: "user", Content: "First question"},
+				{Role: "assistant", Content: "First answer part 1\nFirst answer part 2"},
+				{Role: "user", Content: "Second question"},
+				{Role: "assistant", Content: "Second answer part 1\nSecond answer part 2\nSecond answer part 3"},
+			},
+		},
+		{
+			name: "single assistant message (no consolidation needed)",
+			input: []models.RoleMsg{
+				{Role: "user", Content: "Hello"},
+				{Role: "assistant", Content: "Hi there"},
+			},
+			expected: []models.RoleMsg{
+				{Role: "user", Content: "Hello"},
+				{Role: "assistant", Content: "Hi there"},
+			},
+		},
+		{
+			name: "only assistant messages",
+			input: []models.RoleMsg{
+				{Role: "assistant", Content: "First"},
+				{Role: "assistant", Content: "Second"},
+				{Role: "assistant", Content: "Third"},
+			},
+			expected: []models.RoleMsg{
+				{Role: "assistant", Content: "First\nSecond\nThird"},
+			},
+		},
+		{
+			name: "user messages at the end are preserved",
+			input: []models.RoleMsg{
+				{Role: "assistant", Content: "First"},
+				{Role: "assistant", Content: "Second"},
+				{Role: "user", Content: "Final user message"},
+			},
+			expected: []models.RoleMsg{
+				{Role: "assistant", Content: "First\nSecond"},
+				{Role: "user", Content: "Final user message"},
+			},
+		},
+		{
+			name: "tool call ids preserved in consolidation",
+			input: []models.RoleMsg{
+				{Role: "user", Content: "Hello"},
+				{Role: "assistant", Content: "First part", ToolCallID: "call_123"},
+				{Role: "assistant", Content: "Second part", ToolCallID: "call_123"}, // Same ID
+				{Role: "user", Content: "Thanks"},
+			},
+			expected: []models.RoleMsg{
+				{Role: "user", Content: "Hello"},
+				{Role: "assistant", Content: "First part\nSecond part", ToolCallID: "call_123"},
+				{Role: "user", Content: "Thanks"},
+			},
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			result := consolidateConsecutiveAssistantMessages(tt.input)
+			
+			if len(result) != len(tt.expected) {
+				t.Errorf("Expected %d messages, got %d", len(tt.expected), len(result))
+				t.Logf("Result: %+v", result)
+				t.Logf("Expected: %+v", tt.expected)
+				return
+			}
+			
+			for i, expectedMsg := range tt.expected {
+				if i >= len(result) {
+					t.Errorf("Result has fewer messages than expected at index %d", i)
+					continue
+				}
+				
+				actualMsg := result[i]
+				if actualMsg.Role != expectedMsg.Role {
+					t.Errorf("Message %d: expected role '%s', got '%s'", i, expectedMsg.Role, actualMsg.Role)
+				}
+				
+				if actualMsg.Content != expectedMsg.Content {
+					t.Errorf("Message %d: expected content '%s', got '%s'", i, expectedMsg.Content, actualMsg.Content)
+				}
+				
+				if actualMsg.ToolCallID != expectedMsg.ToolCallID {
+					t.Errorf("Message %d: expected ToolCallID '%s', got '%s'", i, expectedMsg.ToolCallID, actualMsg.ToolCallID)
+				}
+			}
+			
+			// Additional check: ensure no messages were lost
+			if !reflect.DeepEqual(result, tt.expected) {
+				t.Errorf("Result does not match expected:\nResult:   %+v\nExpected: %+v", result, tt.expected)
+			}
+		})
+	}
+}
\ No newline at end of file
diff --git a/extra/tts.go b/extra/tts.go
index 31e6887..c6f373a 100644
--- a/extra/tts.go
+++ b/extra/tts.go
@@ -48,7 +48,7 @@ type KokoroOrator struct {
 
 func (o *KokoroOrator) stoproutine() {
 	<-TTSDoneChan
-	o.logger.Info("orator got done signal")
+	o.logger.Debug("orator got done signal")
 	o.Stop()
 	// drain the channel
 	for len(TTSTextChan) > 0 {
@@ -72,7 +72,7 @@ func (o *KokoroOrator) readroutine() {
 			}
 			text := o.textBuffer.String()
 			sentences := tokenizer.Tokenize(text)
-			o.logger.Info("adding chunk", "chunk", chunk, "text", text, "sen-len", len(sentences))
+			o.logger.Debug("adding chunk", "chunk", chunk, "text", text, "sen-len", len(sentences))
 			for i, sentence := range sentences {
 				if i == len(sentences)-1 { // last sentence
 					o.textBuffer.Reset()
@@ -83,13 +83,13 @@ func (o *KokoroOrator) readroutine() {
 					}
 					continue // if only one (often incomplete) sentence; wait for next chunk
 				}
-				o.logger.Info("calling Speak with sentence", "sent", sentence.Text)
+				o.logger.Debug("calling Speak with sentence", "sent", sentence.Text)
 				if err := o.Speak(sentence.Text); err != nil {
 					o.logger.Error("tts failed", "sentence", sentence.Text, "error", err)
 				}
 			}
 		case <-TTSFlushChan:
-			o.logger.Info("got flushchan signal start")
+			o.logger.Debug("got flushchan signal start")
 			// lln is done get the whole message out
 			if len(TTSTextChan) > 0 { // otherwise might get stuck
 				for chunk := range TTSTextChan {
@@ -110,7 +110,7 @@ func (o *KokoroOrator) readroutine() {
 			remaining := o.textBuffer.String()
 			o.textBuffer.Reset()
 			if remaining != "" {
-				o.logger.Info("calling Speak with remainder", "rem", remaining)
+				o.logger.Debug("calling Speak with remainder", "rem", remaining)
 				if err := o.Speak(remaining); err != nil {
 					o.logger.Error("tts failed", "sentence", remaining, "error", err)
 				}
@@ -171,7 +171,7 @@ func (o *KokoroOrator) requestSound(text string) (io.ReadCloser, error) {
 }
 
 func (o *KokoroOrator) Speak(text string) error {
-	o.logger.Info("fn: Speak is called", "text-len", len(text))
+	o.logger.Debug("fn: Speak is called", "text-len", len(text))
 	body, err := o.requestSound(text)
 	if err != nil {
 		o.logger.Error("request failed", "error", err)
@@ -202,7 +202,7 @@ func (o *KokoroOrator) Speak(text string) error {
 
 func (o *KokoroOrator) Stop() {
 	// speaker.Clear()
-	o.logger.Info("attempted to stop orator", "orator", o)
+	o.logger.Debug("attempted to stop orator", "orator", o)
 	speaker.Lock()
 	defer speaker.Unlock()
 	if o.currentStream != nil {
diff --git a/llm.go b/llm.go
index 38d6c22..599eb4e 100644
--- a/llm.go
+++ b/llm.go
@@ -211,6 +211,8 @@ func (op LCPChat) FormMsg(msg, role string, resume bool) (io.Reader, error) {
 			newMsg = models.NewRoleMsg(role, msg)
 		}
 		chatBody.Messages = append(chatBody.Messages, newMsg)
+		logger.Debug("LCPChat FormMsg: added message to chatBody", "role", newMsg.Role, "content_len", len(newMsg.Content), "message_count_after_add", len(chatBody.Messages))
+
 		// if rag - add as system message to avoid conflicts with tool usage
 		if cfg.RAGEnabled {
 			ragResp, err := chatRagUse(newMsg.Content)
@@ -221,6 +223,7 @@ func (op LCPChat) FormMsg(msg, role string, resume bool) (io.Reader, error) {
 			// Use system role for RAG context to avoid conflicts with tool usage
 			ragMsg := models.RoleMsg{Role: "system", Content: RAGMsg + ragResp}
 			chatBody.Messages = append(chatBody.Messages, ragMsg)
+			logger.Debug("LCPChat FormMsg: added RAG message to chatBody", "role", ragMsg.Role, "rag_content_len", len(ragMsg.Content), "message_count_after_rag", len(chatBody.Messages))
 		}
 	}
 	// openai /v1/chat does not support custom roles; needs to be user, assistant, system
@@ -231,6 +234,7 @@ func (op LCPChat) FormMsg(msg, role string, resume bool) (io.Reader, error) {
 	}
 	for i, msg := range chatBody.Messages {
 		if msg.Role == cfg.UserRole {
+			bodyCopy.Messages[i] = msg
 			bodyCopy.Messages[i].Role = "user"
 		} else {
 			bodyCopy.Messages[i] = msg
@@ -382,6 +386,7 @@ func (ds DeepSeekerChat) FormMsg(msg, role string, resume bool) (io.Reader, erro
 	}
 	for i, msg := range chatBody.Messages {
 		if msg.Role == cfg.UserRole || i == 1 {
+			bodyCopy.Messages[i] = msg
 			bodyCopy.Messages[i].Role = "user"
 		} else {
 			bodyCopy.Messages[i] = msg
@@ -559,6 +564,7 @@ func (or OpenRouterChat) FormMsg(msg, role string, resume bool) (io.Reader, erro
 		bodyCopy.Messages[i] = msg
 		// Standardize role if it's a user role
 		if bodyCopy.Messages[i].Role == cfg.UserRole {
+			bodyCopy.Messages[i] = msg
 			bodyCopy.Messages[i].Role = "user"
 		}
 	}