diff --git a/CHANGELOG.md b/CHANGELOG.md
index 655318d..c240648 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -4,6 +4,9 @@ All notable changes to LowQuery are documented here. The format follows [Keep a
 
 ## [Unreleased]
 
+### Fixed
+- Oracle: running a query that returns a CLOB or BLOB again in the same session could fail with ORA-01002 (fetch out of sequence) ([#17](https://github.com/LowBloat/LowQuery/issues/17)). It happened when the server parsed the query again, which on Oracle 23ai often came by the third run of a query such as `dbms_metadata.get_ddl`, and on every run after one that had failed, until reconnecting. Both causes were in the Oracle driver LowQuery ships and are fixed there; Show DDL no longer needs to bypass the statement cache.
+
 ## [0.5.1] - 2026-10-22
 
 ### Security
diff --git a/src-tauri/src/drivers/integration_tests.rs b/src-tauri/src/drivers/integration_tests.rs
index efb4457..d1d7462 100644
--- a/src-tauri/src/drivers/integration_tests.rs
+++ b/src-tauri/src/drivers/integration_tests.rs
@@ -415,6 +415,142 @@ async fn oracle_wide_rows_with_repeated_values() {
     }
 }
 
+/// Queries returning CLOBs run again and again in one session, through the
+/// driver's statement cache. LOBs are fetched as text, which takes a define
+/// round trip on the first run. ORA-01002 (fetch out of sequence) followed
+/// when that run failed, and when the server parsed and described the query
+/// again on a later run.
+#[tokio::test]
+#[ignore]
+async fn oracle_clob_queries_run_repeatedly() {
+    let Some((profile, pw)) = ora_profile() else {
+        return;
+    };
+    let s = Session::open(&profile, pw.as_deref())
+        .await
+        .expect("connection");
+    let schema = profile.user.to_uppercase();
+    let _ = s.execute(&req("drop table lq_clob purge", 0)).await;
+    s.execute(&req(
+        "create table lq_clob (id number primary key, doc clob)",
+        0,
+    ))
+    .await
+    .unwrap();
+    // Values longer than 4000 characters, which no VARCHAR2 define can hold.
+    s.execute(&req(
+        "insert into lq_clob
+         select level, to_clob(rpad('a', 4000, 'a')) || rpad('b', 4000, 'b') || level
+         from dual connect by level <= 3",
+        0,
+    ))
+    .await
+    .unwrap();
+    s.execute(&req("commit", 0)).await.unwrap();
+
+    let long_doc = |id: usize| format!("{}{}{id}", "a".repeat(4000), "b".repeat(4000));
+    type Check = Box<dyn Fn(&[Vec<Option<String>>])>;
+    let queries: [(&str, Check); 4] = [
+        (
+            "select dbms_metadata.get_ddl('TABLE', 'LQ_CLOB') from dual",
+            Box::new(|rows| {
+                assert_eq!(rows.len(), 1);
+                let ddl = rows[0][0].as_deref().unwrap_or_default();
+                assert!(
+                    ddl.contains("CREATE TABLE") && ddl.contains("LQ_CLOB"),
+                    "{ddl}"
+                );
+            }),
+        ),
+        (
+            "select to_clob('LowQuery ' || level) from dual connect by level <= 3",
+            Box::new(|rows| {
+                let got: Vec<_> = rows.iter().map(|r| r[0].clone()).collect();
+                let want: Vec<_> = (1..=3).map(|i| Some(format!("LowQuery {i}"))).collect();
+                assert_eq!(got, want);
+            }),
+        ),
+        (
+            "select id, doc from lq_clob order by id",
+            Box::new(move |rows| {
+                assert_eq!(rows.len(), 3);
+                for (i, row) in rows.iter().enumerate() {
+                    assert_eq!(row[0].as_deref(), Some((i + 1).to_string().as_str()));
+                    assert_eq!(row[1].as_deref(), Some(long_doc(i + 1).as_str()));
+                }
+            }),
+        ),
+        // The first row only: the cursor is left open with rows to fetch.
+        (
+            "select doc from lq_clob order by id",
+            Box::new(move |rows| {
+                assert_eq!(rows.len(), 1);
+                assert_eq!(rows[0][0].as_deref(), Some(long_doc(1).as_str()));
+            }),
+        ),
+    ];
+    // New statistics invalidate the cursors on LQ_CLOB, as statistics
+    // feedback does on its own: the next run parses the query again and the
+    // server describes its columns again, as CLOB.
+    let invalidate =
+        "begin dbms_stats.gather_table_stats(user, 'LQ_CLOB', no_invalidate => false); end;";
+    for (n, (sql, check)) in queries.iter().enumerate() {
+        let limit = if n == 3 { 1 } else { 100 };
+        for run in 1..=5 {
+            if run == 3 {
+                s.execute(&req(invalidate, 0)).await.unwrap();
+            }
+            let r = s
+                .execute(&req(sql, limit))
+                .await
+                .unwrap_or_else(|e| panic!("{sql}: run {run}: {e}"));
+            check(&r.rows);
+        }
+    }
+    // A query that fails while its LOB column is defined (ORA-01476 here),
+    // then succeeds once the data changes.
+    let _ = s.execute(&req("drop table lq_clob_flag purge", 0)).await;
+    s.execute(&req("create table lq_clob_flag (n number)", 0))
+        .await
+        .unwrap();
+    s.execute(&req("insert into lq_clob_flag values (0)", 0))
+        .await
+        .unwrap();
+    s.execute(&req("commit", 0)).await.unwrap();
+    let flagged = "select to_clob('n=' || (select 1 / n from lq_clob_flag)) from dual";
+    for run in 1..=2 {
+        let err = s.execute(&req(flagged, 100)).await.unwrap_err();
+        assert!(err.to_string().contains("ORA-01476"), "run {run}: {err}");
+    }
+    s.execute(&req("update lq_clob_flag set n = 1", 0))
+        .await
+        .unwrap();
+    s.execute(&req("commit", 0)).await.unwrap();
+    for _ in 3..=5 {
+        assert_eq!(one_row(&s, flagged).await[0].as_deref(), Some("n=1"));
+    }
+    s.execute(&req("drop table lq_clob_flag purge", 0))
+        .await
+        .unwrap();
+    // Show DDL reads DBMS_METADATA with binds, five times in a row; the
+    // table has no comments, so one of its queries fails every time.
+    for run in 1..=5 {
+        let ddl = s
+            .object_ddl(&schema, "LQ_CLOB", "table")
+            .await
+            .unwrap_or_else(|e| panic!("Show DDL run {run}: {e}"));
+        assert!(ddl.contains("LQ_CLOB"), "{ddl}");
+    }
+    // The session is still usable afterwards.
+    assert_eq!(
+        one_row(&s, "select count(*) from lq_clob").await[0].as_deref(),
+        Some("3")
+    );
+    s.execute(&req("drop table lq_clob purge", 0))
+        .await
+        .unwrap();
+}
+
 #[tokio::test]
 #[ignore]
 async fn oracle_connects_through_tns_alias() {
diff --git a/src-tauri/src/drivers/oracle.rs b/src-tauri/src/drivers/oracle.rs
index 8656439..c9fcfc2 100644
--- a/src-tauri/src/drivers/oracle.rs
+++ b/src-tauri/src/drivers/oracle.rs
@@ -580,9 +580,7 @@ impl OraSession {
     }
 }
 
-/// First column of every row, as text. The statement stays out of the
-/// driver's cache: on 23ai, running a cached DBMS_METADATA query a third time
-/// in a session fails with ORA-01002 (fetch out of sequence).
+/// First column of every row, as text.
 fn first_texts(conn: &Connection, sql: &str, params: &[String]) -> AppResult<Vec<String>> {
     let binds: Vec<&dyn oracledb::ToDbValue> = params
         .iter()
@@ -591,7 +589,6 @@ fn first_texts(conn: &Connection, sql: &str, params: &[String]) -> AppResult<Vec
     let cursor = conn
         .statement(sql)
         .map_err(map_error)?
-        .exclude_from_cache()
         .build()
         .map_err(map_error)?
         .query(&binds)
diff --git a/src-tauri/vendor/oracledb/LOWQUERY.md b/src-tauri/vendor/oracledb/LOWQUERY.md
index cfdbeba..7c65706 100644
--- a/src-tauri/vendor/oracledb/LOWQUERY.md
+++ b/src-tauri/vendor/oracledb/LOWQUERY.md
@@ -1,7 +1,7 @@
 # Patched oracledb
 
 This is `oracledb` 26.0.0-beta.4 from crates.io (Oracle's Rust thin driver,
-UPL-1.0 OR Apache-2.0) with two fixes; everything else is unchanged.
+UPL-1.0 OR Apache-2.0) with four fixes; everything else is unchanged.
 
 1. `src/messages/execute.rs`: the execute response now knows the number of
    columns of a statement that was already described. When a query runs again
@@ -13,7 +13,29 @@ UPL-1.0 OR Apache-2.0) with two fixes; everything else is unchanged.
 2. `src/response/mod.rs`: a bit vector is cleared after the row it belongs to,
    as python-oracledb does, so a later row without one is not read with stale
    duplicate flags.
+3. `src/client/mod.rs`, `src/statement/cache.rs`: a query with LOB columns
+   is executed, then sent again to define those columns (as text, or as LOB
+   locators). When either step failed, for example with an error raised while
+   fetching the first rows, the statement went back to the statement cache
+   still marked as needing a define. Its next execution then only sent the
+   define and fetched from a cursor that was never executed, failing with
+   ORA-01002 (fetch out of sequence), and so did every later execution of the
+   same SQL in the session. The cursor of such a statement is now closed and
+   the next execution parses it again, as python-oracledb does after an
+   error.
+4. `src/metadata.rs`, `src/statement/mod.rs`: when a query that ran before is
+   parsed again by the server (statistics feedback, new statistics, other
+   invalidations), the server describes its columns again on the next
+   execution, LOB columns as CLOB/BLOB, although the cursor keeps its define
+   and returns them as LONG/LONG RAW along with the rows. The driver took that
+   description as a new LOB column and sent a define after the rows had been
+   fetched, failing with ORA-01002 (fetch out of sequence). On Oracle 23ai
+   this showed on about the third run of a cached `DBMS_METADATA` query. Like
+   python-oracledb's `_adjust_metadata`, a column described as CLOB/BLOB
+   whose previous type on the same cursor was a string/LONG (or RAW/LONG
+   RAW) keeps the LONG type and needs no define.
 
 `drivers::integration_tests::oracle_wide_rows_with_repeated_values` covers
-both on Oracle 21c and 23ai. Remove the `[patch.crates-io]` entry in
-`Cargo.toml` and this folder once an upstream release includes the fixes.
+the first two and `oracle_clob_queries_run_repeatedly` the last two, on
+Oracle 21c and 23ai. Remove the `[patch.crates-io]` entry in `Cargo.toml` and this
+folder once an upstream release includes the fixes.
diff --git a/src-tauri/vendor/oracledb/src/client/mod.rs b/src-tauri/vendor/oracledb/src/client/mod.rs
index 35d409b..c247249 100644
--- a/src-tauri/vendor/oracledb/src/client/mod.rs
+++ b/src-tauri/vendor/oracledb/src/client/mod.rs
@@ -743,11 +743,21 @@ impl Client {
         } else {
             let mut message =
                 ExecuteMessage::new(statement, params, parse_only);
-            let mut response = self.process_message(&mut message)?;
-            response.finalize_rows(client_ref, statement.out_metadata());
+            let result = self.process_message(&mut message);
+            // LowQuery patch: when the execution fails before or during the
+            // define of LOB columns, the cursor is left in an unknown state
+            // (not executed, or without its define); it is closed so the next
+            // execution parses the statement again. Otherwise that execution
+            // would perform a define only, then fetch from a cursor that was
+            // never executed (ORA-01002: fetch out of sequence).
             if statement.requires_define() {
+                if result.is_err() {
+                    self.statement_cache.clear_cursor(statement);
+                }
                 statement.clear_requires_define();
             }
+            let mut response = result?;
+            response.finalize_rows(client_ref, statement.out_metadata());
             Ok(response)
         }
     }
diff --git a/src-tauri/vendor/oracledb/src/metadata.rs b/src-tauri/vendor/oracledb/src/metadata.rs
index 340c76d..c7bc82b 100644
--- a/src-tauri/vendor/oracledb/src/metadata.rs
+++ b/src-tauri/vendor/oracledb/src/metadata.rs
@@ -84,6 +84,42 @@ impl Metadata {
         }
     }
 
+    /// LowQuery patch, from python-oracledb's `_adjust_metadata`: when a
+    /// query runs again and the server describes it again (after parsing it
+    /// anew), a CLOB or BLOB column whose cursor already has a define as
+    /// string/bytes is still returned as LONG/LONG RAW, so it gets the same
+    /// metadata as `define_metadata` and needs no new define.
+    pub(crate) fn adjust_for_previous(&mut self, previous: &Metadata) {
+        let previous_type = previous.db_type.ora_type_num;
+        match self.db_type.ora_type_num {
+            constants::ORA_TYPE_NUM_CLOB
+                if matches!(
+                    previous_type,
+                    constants::ORA_TYPE_NUM_CHAR
+                        | constants::ORA_TYPE_NUM_LONG
+                        | constants::ORA_TYPE_NUM_VARCHAR
+                ) =>
+            {
+                self.db_type = DbType::from_ora_type_and_csfrm(
+                    constants::ORA_TYPE_NUM_LONG,
+                    previous.db_type.csfrm,
+                );
+                self.max_size = 0;
+            }
+            constants::ORA_TYPE_NUM_BLOB
+                if matches!(
+                    previous_type,
+                    constants::ORA_TYPE_NUM_RAW
+                        | constants::ORA_TYPE_NUM_LONG_RAW
+                ) =>
+            {
+                self.db_type = crate::DB_TYPE_LONG_RAW;
+                self.max_size = 0;
+            }
+            _ => {}
+        }
+    }
+
     /// Returns a new structure suitable for defining CLOB/BLOB as
     /// string/bytes.
     pub(crate) fn define_metadata(&self) -> Metadata {
diff --git a/src-tauri/vendor/oracledb/src/statement/cache.rs b/src-tauri/vendor/oracledb/src/statement/cache.rs
index 2f5daf3..13bc3a3 100644
--- a/src-tauri/vendor/oracledb/src/statement/cache.rs
+++ b/src-tauri/vendor/oracledb/src/statement/cache.rs
@@ -131,6 +131,13 @@ impl StatementCache {
         !self.empty_slots.is_empty() || !self.free_used_slots.is_empty()
     }
 
+    /// Forgets the database cursor of a statement and closes it on the next
+    /// round trip (LowQuery patch).
+    pub(crate) fn clear_cursor(&mut self, info: &mut CachedStatement) {
+        self.close_cursor(info);
+        info.clear_cursor();
+    }
+
     /// Returns whether the statement cache has cursors that need to be closed.
     pub(crate) fn has_cursors_to_close(&self) -> bool {
         !self.cursors_to_close.is_empty()
diff --git a/src-tauri/vendor/oracledb/src/statement/mod.rs b/src-tauri/vendor/oracledb/src/statement/mod.rs
index 05eded8..35b4972 100644
--- a/src-tauri/vendor/oracledb/src/statement/mod.rs
+++ b/src-tauri/vendor/oracledb/src/statement/mod.rs
@@ -374,9 +374,17 @@ impl CachedStatement {
         if num_columns > 0 {
             resp.read_u8()?;
         }
-        self.out_metadata.clear();
-        for _ in 0..num_columns {
+        // LowQuery patch: on a cursor that already ran, the columns described
+        // before tell whether it has a define for its LOB columns.
+        let mut previous = std::mem::take(&mut self.out_metadata);
+        if !self.has_cursor() {
+            previous.clear();
+        }
+        for i in 0..num_columns {
             let mut metadata = Metadata::from_response(resp, client)?;
+            if let Some(previous) = previous.get(i as usize) {
+                metadata.adjust_for_previous(previous);
+            }
             if metadata.requires_define() {
                 self.requires_define = true;
                 self.no_prefetch = true;