feat(usage): add AI budget usage caps

internal/agent/loop_compact.go

@@ -89,14 +89,15 @@ func (l *Loop) compactMessagesInPlace(ctx context.Context, messages []providers.
 
 	inTokens := l.estimateSummaryInputTokens(toSummarize)
 	slog.Info("compact_budget", "agent", l.id, "in_tokens", inTokens, "out_tokens", dynamicSummaryMax(inTokens))
-	resp, err := l.provider.Chat(sctx, providers.ChatRequest{
+	chatReq := providers.ChatRequest{
 		Messages: []providers.Message{{
 			Role:    "user",
 			Content: compactionSummaryPrompt + sb.String(),
 		}},
 		Model:   l.model,
 		Options: map[string]any{"max_tokens": dynamicSummaryMax(inTokens), "temperature": 0.3},
-	})
+	}
+	resp, err := l.callInternalLLMWithUsage(sctx, chatReq, "mid-loop-compaction")
 	if err != nil {
 		slog.Warn("mid_loop_compaction_failed", "agent", l.id, "error", err)
 		return nil