fix: give glue models awareness of agent conversation history

- Pass recent agent context (last 3 exchanges) to ChatReply
- Skip LLM classification when there's active history — follow-ups
  after a build go straight to agent
- Only greetings/thanks route to chat when conversation is active

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: halfaipg

chat.go

@@ -124,6 +124,28 @@ func newChatModel(cfg *Config) chatModel {
 	}
 }
 
+// recentContext extracts the last few user/assistant exchanges from the agent
+// history so that glue models (classify, chat) know what was said.
+func (m *chatModel) recentContext() []ChatMessage {
+	if m.agent == nil {
+		return nil
+	}
+	// Pull user + assistant messages from agent history (skip system, tool)
+	var msgs []ChatMessage
+	for _, msg := range m.agent.history {
+		if msg.Role == "user" || msg.Role == "assistant" {
+			if msg.Content != nil && *msg.Content != "" {
+				msgs = append(msgs, ChatMessage{Role: msg.Role, Content: msg.Content})
+			}
+		}
+	}
+	// Keep last 6 messages max (3 exchanges) to stay cheap
+	if len(msgs) > 6 {
+		msgs = msgs[len(msgs)-6:]
+	}
+	return msgs
+}
+
 func (m chatModel) Init() tea.Cmd {
 	return tea.Batch(
 		textinput.Blink,
@@ -204,6 +226,7 @@ func (m chatModel) Update(msg tea.Msg) (chatModel, tea.Cmd) {
 
 				// Route through glue intent classification
 				hasHistory := m.agent != nil
+				ctx := m.recentContext()
 				intent := m.glue.ClassifyIntent(prompt, hasHistory)
 
 				switch intent {
@@ -214,7 +237,7 @@ func (m chatModel) Update(msg tea.Msg) (chatModel, tea.Cmd) {
 					})
 					glue := m.glue
 					cmds = append(cmds, func() tea.Msg {
-						reply := glue.ChatReply(prompt, nil)
+						reply := glue.ChatReply(prompt, ctx)
 						return glueResultMsg{kind: "chat", text: reply}
 					})
 

glue.go

@@ -80,6 +80,20 @@ When in doubt, pick "agent". Respond with ONLY one word: agent, plan, chat, or c
 
 // ClassifyIntent determines what kind of response a user message needs.
 func (g *GlueClient) ClassifyIntent(userMsg string, hasHistory bool) Intent {
+	// If there's an active conversation, follow-ups almost always go to agent.
+	// Only classify from scratch on the first message or after a greeting.
+	if hasHistory {
+		// Quick check: is this clearly just "thanks" / "hi" / "ok"?
+		lower := strings.ToLower(strings.TrimSpace(userMsg))
+		greetings := []string{"hi", "hey", "hello", "thanks", "thank you", "ok", "cool", "nice", "bye"}
+		for _, g := range greetings {
+			if lower == g {
+				return IntentChat
+			}
+		}
+		return IntentAgent
+	}
+
 	messages := []ChatMessage{
 		{Role: "system", Content: strPtr(classifyPrompt)},
 		{Role: "user", Content: strPtr(userMsg)},
@@ -91,17 +105,12 @@ func (g *GlueClient) ClassifyIntent(userMsg string, hasHistory bool) Intent {
 	}
 
 	result = strings.TrimSpace(strings.ToLower(result))
-	// Extract just the keyword — check plan before agent since "agent" is a substring match risk
 	for _, intent := range []Intent{IntentPlan, IntentChat, IntentClarify, IntentAgent} {
 		if strings.Contains(result, string(intent)) {
-			// If they have conversation history, "clarify" becomes less useful
-			if intent == IntentClarify && hasHistory {
-				return IntentAgent
-			}
 			return intent
 		}
 	}
-	return IntentAgent // default: let the agent handle it
+	return IntentAgent
 }
 
 // ──────────────────────────────────────────────────────────────