fix: auto-compact logic

internal/llm/agent/agent.go

@@ -296,36 +296,19 @@ func (a *agent) createUserMessage(ctx context.Context, sessionID, content string
 	})
 }
 
-// estimateTokens provides a rough estimate of token count based on character count
-// using a simple heuristic of ~4 characters per token
-func estimateTokens(messages []message.Message) int64 {
-	totalChars := 0
-	for _, msg := range messages {
-		// Get text content from all parts
-		for _, part := range msg.Parts {
-			if textContent, ok := part.(message.TextContent); ok {
-				totalChars += len(textContent.Text)
-			} else {
-				// For non-text parts, add a conservative estimate
-				totalChars += 100
-			}
-		}
-		// Add chars for role (conservative estimate)
-		totalChars += 10
-	}
-	// Heuristic: ~4 chars per token
-	return int64(totalChars / 4)
-}
-
 func (a *agent) streamAndHandleEvents(ctx context.Context, sessionID string, msgHistory []message.Message) (message.Message, *message.Message, error) {
 	// Check if we need to auto-compact based on token count
 	contextWindow := a.provider.Model().ContextWindow
+	maxTokens := a.provider.MaxTokens()
 	threshold := int64(float64(contextWindow) * 0.80)
-	estimatedTokens := estimateTokens(msgHistory)
+	usage, err := a.GetUsage(ctx, sessionID)
+	if err != nil || usage == nil {
+		return message.Message{}, nil, fmt.Errorf("failed to get usage: %w", err)
+	}
 
 	// If we're approaching the context window limit, trigger auto-compaction
-	if estimatedTokens >= threshold {
-		logging.InfoPersist(fmt.Sprintf("Auto-compaction triggered for session %s. Estimated tokens: %d, Threshold: %d", sessionID, estimatedTokens, threshold))
+	if (*usage + maxTokens) >= threshold {
+		logging.InfoPersist(fmt.Sprintf("Auto-compaction triggered for session %s. Estimated tokens: %d, Threshold: %d", sessionID, usage, threshold))
 
 		// Perform compaction with pause/resume to ensure safety
 		if err := a.CompactSession(ctx, sessionID); err != nil {
@@ -357,11 +340,6 @@ func (a *agent) streamAndHandleEvents(ctx context.Context, sessionID string, msg
 
 				// Replace msgHistory with the new compacted version
 				msgHistory = append([]message.Message{summaryMessage}, sessionMessages...)
-
-				// Log the new token estimate after compaction
-				newEstimate := estimateTokens(msgHistory)
-				logging.InfoPersist(fmt.Sprintf("After compaction: Estimated tokens: %d (reduced by %d)",
-					newEstimate, estimatedTokens-newEstimate))
 			}
 		}
 	}
@@ -534,6 +512,16 @@ func (a *agent) processEvent(ctx context.Context, sessionID string, assistantMsg
 	return nil
 }
 
+func (a *agent) GetUsage(ctx context.Context, sessionID string) (*int64, error) {
+	session, err := a.sessions.Get(ctx, sessionID)
+	if err != nil {
+		return nil, fmt.Errorf("failed to get session: %w", err)
+	}
+
+	usage := session.PromptTokens + session.CompletionTokens
+	return &usage, nil
+}
+
 func (a *agent) TrackUsage(ctx context.Context, sessionID string, model models.Model, usage provider.TokenUsage) error {
 	sess, err := a.sessions.Get(ctx, sessionID)
 	if err != nil {