diff --git a/lib/fmt.c b/lib/fmt.c
index a3c6ff6..3a727cb 100644
--- a/lib/fmt.c
+++ b/lib/fmt.c
@@ -1867,7 +1867,7 @@ static int print_bcs (hcl_fmtout_t* fmtout, const hcl_bch_t* ptr, hcl_oow_t len)
 
 	if (HCL_UNLIKELY(!hcl->io.udo_wrtr))
 	{
-		hcl_seterrmsg (hcl, HCL_EINVAL, "no user-defined output handler");
+		hcl_seterrbmsg (hcl, HCL_EINVAL, "no user-defined output handler");
 		return -1;
 	}
 
@@ -1895,7 +1895,7 @@ static int print_ucs (hcl_fmtout_t* fmtout, const hcl_uch_t* ptr, hcl_oow_t len)
 
 	if (HCL_UNLIKELY(!hcl->io.udo_wrtr))
 	{
-		hcl_seterrmsg (hcl, HCL_EINVAL, "no user-defined output handler");
+		hcl_seterrbmsg (hcl, HCL_EINVAL, "no user-defined output handler");
 		return -1;
 	}
 
diff --git a/lib/hcl.h b/lib/hcl.h
index 5d08aab..e42fba4 100644
--- a/lib/hcl.h
+++ b/lib/hcl.h
@@ -1208,6 +1208,12 @@ struct hcl_lxc_t
 };
 typedef struct hcl_lxc_t hcl_lxc_t;
 
+
+/*#define HCL_CCI_BUF_LEN (1)*/
+#if !defined(HCL_CCI_BUF_LEN)
+#define HCL_CCI_BUF_LEN (2048)
+#endif
+
 typedef struct hcl_io_cciarg_t hcl_io_cciarg_t;
 struct hcl_io_cciarg_t
 {
@@ -1232,8 +1238,8 @@ struct hcl_io_cciarg_t
 	int is_bytes; /* set this to non-zero if the handler fills the buffer with bytes */
 	union
 	{
-		hcl_ooch_t c[2048]; /* TODO: resize this if necessary */
-		hcl_uint8_t b[2048 * HCL_SIZEOF(hcl_ooch_t)]; /* TODO: resize this if necessary */
+		hcl_ooch_t c[HCL_CCI_BUF_LEN];
+		hcl_uint8_t b[HCL_CCI_BUF_LEN * HCL_SIZEOF(hcl_ooch_t)];
 	} buf;
 
 	/**
diff --git a/lib/read.c b/lib/read.c
index b4f8882..6be17e5 100644
--- a/lib/read.c
+++ b/lib/read.c
@@ -2968,12 +2968,16 @@ static int feed_from_includee (hcl_t* hcl)
 
 				curinp->b.pos = 0;
 				curinp->b.len = curinp->xlen;
+//printf ("curinp->xlen = %d\n", (int)curinp->xlen);
 			}
 			avail = curinp->b.len - curinp->b.pos; /* available in the read buffer */
 			saved_rsd_len = curinp->rsd.len;
 
-			if (curinp->rsd.len > 0)
+//printf ("saved_rsd_len = %d avail=%d\n", (int)saved_rsd_len, (int)avail);
+			if (saved_rsd_len > 0)
 			{
+				/* there is data in the residue buffer. use the residue buffer to
+				 * locate a proper multi-byte sequence */
 				hcl_oow_t cpl; /* number of bytes to copy to the residue buffer */
 				HCL_ASSERT (hcl, curinp->b.pos == 0);
 				cpl = HCL_COUNTOF(curinp->rsd.buf) - curinp->rsd.len;
@@ -2985,7 +2989,9 @@ static int feed_from_includee (hcl_t* hcl)
 					curinp->b.pos += cpl; /* advance this because the bytes moved to the residue buffer */
 				}
 				inplen = curinp->rsd.len;
-				n = cmgr->bctouc(curinp->rsd.buf, inplen, &c);
+				n = cmgr->bctouc(&curinp->rsd.buf[0], inplen, &c);
+//printf ("residue -> inplen = %d cpl = %d avail=%d\n", (int)inplen, (int)cpl, (int)avail);
+				if (n > 0 && n <= inplen) curinp->b.pos -= curinp->rsd.len - saved_rsd_len;
 			}
 			else
 			{
@@ -2998,19 +3004,20 @@ static int feed_from_includee (hcl_t* hcl)
 				hcl_seterrbfmt (hcl, HCL_EECERR, "invalid byte sequence in %js", curinp->name);
 				return -1;
 			}
-			if (n > inplen)
+			if (n > inplen) /* incomplete sequence */
 			{
-				/* incomplete sequence */
-				HCL_ASSERT (hcl, avail < HCL_COUNTOF(curinp->rsd.buf));
-
-				/* TODO: wrong */
-				HCL_MEMCPY (curinp->rsd.buf, &curinp->buf.b[curinp->b.pos], avail);
-				curinp->rsd.len = avail;
-				curinp->b.pos = curinp->b.len;
+				hcl_oow_t cpl;
+				HCL_ASSERT (hcl, curinp->rsd.len < HCL_COUNTOF(curinp->rsd.buf));
+				cpl = HCL_COUNTOF(curinp->rsd.buf) - curinp->rsd.len;
+				if (cpl > avail) cpl = avail;
+				HCL_MEMCPY(&curinp->rsd.buf[curinp->rsd.len], &curinp->buf.b[curinp->b.pos], cpl);
+				curinp->rsd.len += cpl;
+				curinp->b.pos += cpl;
 				goto start_over;
 			}
 
 			/* how much taken from the read buffer as input */
+			HCL_ASSERT (hcl, n >= saved_rsd_len);
 			taken = n - saved_rsd_len;
 		}
 		else
@@ -3044,6 +3051,7 @@ static int feed_from_includee (hcl_t* hcl)
 		}
 	#endif
 
+//hcl_logbfmt(hcl, HCL_LOG_STDERR, "[%jc]\n", c);
 		x = feed_char(hcl, c);
 		if (x <= -1) return -1;
 		if (x >= 1)
diff --git a/pas/hcl.pas b/pas/hcl.pas
index a88d76f..dbdff5f 100644
--- a/pas/hcl.pas
+++ b/pas/hcl.pas
@@ -47,7 +47,7 @@ type
 		name: pwidechar;
 		handle: pointer;
 		is_bytes: integer;
-		buf: array[0..2047] of widechar;
+		buf: array[0..(2048 - 1)] of widechar;
 		xlen: System.SizeUint;
 		includer: CciArgPtr;
 	end;