diff --git a/Cargo.lock b/Cargo.lock
index de93390b2..8f4f2b13c 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -3237,9 +3237,9 @@ dependencies = [
 
 [[package]]
 name = "zlob"
-version = "1.6.3"
+version = "1.6.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f0f010d9b10495cc0037cb47cedc143c39c7fa2d59f752dca19036fd0234a210"
+checksum = "7ed9cfbcfaf165adcf07d994f85a92effaef03029b53afc0872152671efcfbda"
 dependencies = [
  "bindgen",
  "bitflags 2.11.0",
diff --git a/Cargo.toml b/Cargo.toml
index 99854c592..261e9b5ac 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -36,7 +36,7 @@ ignore = "0.4.22"
 memmap2 = "0.9"
 mimalloc = "0.1.47"
 signal-hook-registry = "1.4"
-zlob = { version = "=1.6.3" }
+zlob = { version = "=1.6.5" }
 
 mlua = { version = "0.11.1", features = ["module", "luajit"] }
 neo_frizbee = { version = "0.11.0", features = ["match_end_col"] }
diff --git a/crates/fff-core/src/file_picker.rs b/crates/fff-core/src/file_picker.rs
index bbb62096b..346504074 100644
--- a/crates/fff-core/src/file_picker.rs
+++ b/crates/fff-core/src/file_picker.rs
@@ -506,9 +506,19 @@ impl FileItem {
         let is_binary = is_known_binary_extension(path);
         // SAFETY-ish: paths on macOS/Linux are bytes; lossy conversion mirrors
         // the existing `to_string_lossy()` behavior on non-UTF8 names.
-        let rel_str = String::from_utf8_lossy(relative_path).into_owned();
+        let decoded = String::from_utf8_lossy(relative_path).into_owned();
+        // The caller's offset indexes the raw bytes; re-measure it against the
+        // decoded string so it never lands inside a U+FFFD (#799).
+        let dir_bytes = relative_path
+            .get(..basename_offset as usize)
+            .unwrap_or(relative_path);
+        let basename_offset = if std::str::from_utf8(dir_bytes).is_ok() {
+            basename_offset
+        } else {
+            String::from_utf8_lossy(dir_bytes).len() as u16
+        };
         let item = Self::new_raw(basename_offset, size, modified, git_status, is_binary);
-        (item, rel_str)
+        (item, decoded)
     }
 
     pub(crate) fn update_frecency_scores(
diff --git a/crates/fff-core/src/walk/zlob.rs b/crates/fff-core/src/walk/zlob.rs
index 9dc2f3504..98eddf1fc 100644
--- a/crates/fff-core/src/walk/zlob.rs
+++ b/crates/fff-core/src/walk/zlob.rs
@@ -54,19 +54,17 @@ pub(crate) fn walk_collect_files(
         if !entry.is_file() {
             // unlike ripgrep walker zlob doesnt show .git files
             if entry.is_dir() {
-                let rel_bytes = entry.relative_path_bytes();
-                if !rel_bytes.is_empty() {
-                    let mut rel = String::from_utf8_lossy(rel_bytes).into_owned();
-                    rel.push('/');
-                    collected.lock().1.push(rel);
+                let relative_path = entry.relative_path_lossy();
+                if !relative_path.is_empty() {
+                    let mut relative_path = relative_path.into_owned();
+                    relative_path.push('/');
+                    collected.lock().1.push(relative_path);
                 }
             }
 
             return WalkState::Continue;
         }
 
-        let rel_bytes = entry.relative_path_bytes();
-
         // `basename()` returns `&str` for files only.
         let basename = entry.basename().unwrap_or("");
         let is_binary = is_known_binary_extension_basename(basename);
@@ -78,14 +76,16 @@ pub(crate) fn walk_collect_files(
             .map(|ns| (ns / 1_000_000_000).max(0) as u64)
             .unwrap_or(0);
 
-        let basename_offset = entry.basename_offset_in_relative();
+        // Lossy pair: the offset must index the decoded string, not the raw
+        // bytes, or it lands inside a U+FFFD on invalid-UTF-8 names (#799).
+        let basename_offset = entry.basename_offset_in_relative_lossy() as u16;
         // zlob emits '/'-separated relative paths, which is fff's canonical
         // internal form on every platform — store them verbatim.
-        let rel_str = String::from_utf8_lossy(rel_bytes).into_owned();
+        let relative_path = entry.relative_path_lossy().into_owned();
         let item = FileItem::new_raw(basename_offset, size, modified, None, is_binary);
 
         let mut guard = collected.lock();
-        guard.0.push((item, rel_str));
+        guard.0.push((item, relative_path));
         let n = guard.0.len();
         drop(guard);
 
diff --git a/crates/fff-core/tests/invalid_utf8_paths.rs b/crates/fff-core/tests/invalid_utf8_paths.rs
new file mode 100644
index 000000000..d758cc6d3
--- /dev/null
+++ b/crates/fff-core/tests/invalid_utf8_paths.rs
@@ -0,0 +1,261 @@
+// Regression test for https://github.com/dmtrKovalenko/fff/issues/799
+// Unix-only: Windows filenames are UTF-16 and cannot carry these bytes.
+#![cfg(unix)]
+
+use std::path::{Path, PathBuf};
+use tempfile::TempDir;
+
+use fff_search::file_picker::FilePicker;
+use fff_search::{FilePickerOptions, FuzzySearchOptions, PaginationArgs, QueryParser};
+
+/// Byte-exact corpus from the reporter's tarball (fff-invalid-utf8-repro.tar.gz,
+/// sha256 fc0011038d27bbd63a3ca67a929d4f342b4895f922062fe86bd958712c67de08):
+/// four nested dirs plus one empty file, names carrying invalid UTF-8,
+/// control bytes and truncated multibyte sequences.
+const REPRO_COMPONENTS: &[(&str, bool)] = &[
+    // (hex-encoded name, is_dir)
+    ("64617461322e6361622e657874726163746564", true),
+    (
+        "2366711cf02a03386b398b0b95824b120cd90135e777c6356ab77ead67ed0be428e3c5a7c49084cd62468cef4f780b46a97ff5e2be8afefd0962a9cfae50f861a2f637fcfe62f44c32aa96392ce1873f392ce18785905893a6c8f7b9b80b39f9cacc576a22f42758c9bd1e3071bbf041ee629cf0bbd09771b2",
+        true,
+    ),
+    (
+        "33516e5b0214f6bdf7056a22dd6886262defcb33811eb9d4bf84b321422cd4b2beb87e5db8c4aedf7fae5e45a5c5fae7e9ba39b1547f8cc3778738f9a94391b3875f20c4cb0a1b4b163938f39558e66ab0e81d0bcacd4139fd62cbf3bcfd9142a29ab511ad9c7766390abad8143428261982ba754133515066a9edf0d34c13e33e35858f65eee4adb6144fdb11546fc442574121adb24191fb95ae6ee3c40a9d787cde8cf1941c8ca7e4f903f2a5b5eaee4b5fe5b18ad62b3a48dc67431d94e0d6503b8e7bf48cd7c3be9c8569598e6f23db3222",
+        true,
+    ),
+    (
+        "6248697c3afe892703961a9d668d9ac88f557d357768c5e665ac9abc505920a44c06a56fff",
+        true,
+    ),
+    (
+        "07fa04794121fad9f17a52a373a636bec7d08d0fa68e1e7a61ce8751bbf68d06ab74fe07206dde4cadb44730f93b49f2987148fe1f28093980f5df7fcbac6b46d9837e31e97e1bae9f581a665209e43f783f06a65666d1853a9f36dea3b56aa3dcfdf7c08ccdbd7423b460de92c7157f506c1f6057fab76abb2bdae6056ccadc49dac6f5ee1e8af2c499f60dc350e3df6a5b9f38135b5dbd60c182e099af1d1f6d3c9ba26dda282e77ec59bde5afcf3e1204f7f8fd8d67bb6bdb37b65cee783f888f8454cf5d967b239ee8cd956dc3e616",
+        false,
+    ),
+];
+
+/// Hand-made corpus covering the shapes the tarball doesn't: invalid bytes in
+/// the leading component, an invalid byte right before the separator, a name
+/// that is invalid only in its basename, and a valid-UTF-8 multibyte dir whose
+/// offset must stay untouched. Entries are '/'-joined hex components.
+const EXTRA_CASES: &[(&str, bool)] = &[
+    // "\xff\xffx/file.txt" — the minimal case from the triage report
+    ("ffff78/66696c652e747874", false),
+    // "a\xffb\xff/plain.md" — invalid byte immediately before the separator
+    ("61ff62ff/706c61696e2e6d64", false),
+    // valid dir, invalid basename: "ok/\xc3(bad"
+    ("6f6b/c328626164", false),
+    // "каталог/main.rs" — valid multibyte, offset must equal the raw one
+    ("d0bad0b0d182d0b0d0bbd0bed0b3/6d61696e2e7273", false),
+];
+
+#[test]
+fn scans_invalid_utf8_corpus_without_panicking() {
+    let tmp = TempDir::new().unwrap();
+    let Some(expected) = build_chain(tmp.path(), REPRO_COMPONENTS) else {
+        eprintln!("filesystem rejects invalid UTF-8 names; skipping");
+        return;
+    };
+
+    // Pre-fix this aborts the process from the background scan thread.
+    let picker = scan(tmp.path());
+    let indexed = all_paths(&picker);
+
+    assert!(
+        indexed.contains(&expected),
+        "the reporter's file must be indexed under its lossy path"
+    );
+    assert_offsets_are_char_boundaries(&picker);
+    assert_dirs_indexed(&picker, &expected);
+}
+
+#[test]
+fn scans_mixed_encoding_corpus_without_panicking() {
+    let tmp = TempDir::new().unwrap();
+    let Some(()) = build_paths(tmp.path(), EXTRA_CASES) else {
+        eprintln!("filesystem rejects invalid UTF-8 names; skipping");
+        return;
+    };
+
+    let picker = scan(tmp.path());
+    let indexed = all_paths(&picker);
+
+    assert!(
+        indexed
+            .iter()
+            .any(|p| p.ends_with("\u{FFFD}\u{FFFD}x/file.txt")),
+        "expected the \\xff\\xffx/file.txt case, got {indexed:?}"
+    );
+    // A valid-UTF-8 multibyte dir must keep the raw offset verbatim.
+    let cyrillic = indexed
+        .iter()
+        .find(|p| p.contains("каталог"))
+        .expect("cyrillic path must be indexed");
+    assert!(cyrillic.ends_with("каталог/main.rs"), "got {cyrillic:?}");
+    assert_offsets_are_char_boundaries(&picker);
+}
+
+/// Searching by a plain ASCII component of a path whose other components are
+/// invalid UTF-8 must still work end to end.
+#[test]
+fn searches_across_invalid_utf8_parents() {
+    let tmp = TempDir::new().unwrap();
+    if build_chain(tmp.path(), REPRO_COMPONENTS).is_none() {
+        eprintln!("filesystem rejects invalid UTF-8 names; skipping");
+        return;
+    }
+
+    let picker = scan(tmp.path());
+    let parser = QueryParser::default();
+    let parsed = parser.parse("data2.cab.extracted");
+    let result = picker.fuzzy_search(
+        &parsed,
+        None,
+        FuzzySearchOptions {
+            max_threads: 1,
+            pagination: PaginationArgs {
+                offset: 0,
+                limit: 50,
+            },
+            ..Default::default()
+        },
+    );
+
+    assert_eq!(result.items.len(), 1, "the single corpus file must match");
+    assert!(
+        result.items[0]
+            .relative_path(&picker)
+            .starts_with("data2.cab.extracted/")
+    );
+}
+
+fn scan(base: &Path) -> FilePicker {
+    let mut picker = FilePicker::new(FilePickerOptions {
+        base_path: base.to_string_lossy().into_owned(),
+        enable_mmap_cache: false,
+        watch: false,
+        ..Default::default()
+    })
+    .expect("FilePicker::new");
+    picker.collect_files().expect("collect_files");
+    picker
+}
+
+fn all_paths(picker: &FilePicker) -> Vec<String> {
+    picker
+        .get_files()
+        .iter()
+        .map(|f| f.relative_path(picker))
+        .collect()
+}
+
+/// The exact invariant #799 violated: the stored basename offset must be a
+/// char boundary of the decoded path and must point right past a separator.
+fn assert_offsets_are_char_boundaries(picker: &FilePicker) {
+    for file in picker.get_files().iter() {
+        let path = file.relative_path(picker);
+        let offset = file.filename_offset_in_relative_path();
+        assert!(
+            path.is_char_boundary(offset),
+            "offset {offset} is not a char boundary of {path:?}"
+        );
+        assert!(offset <= path.len(), "offset {offset} past end of {path:?}");
+        if offset > 0 {
+            assert_eq!(
+                &path[offset - 1..offset],
+                "/",
+                "offset must follow a separator in {path:?}"
+            );
+        }
+        assert!(
+            !path[offset..].contains('/'),
+            "basename split wrong: {path:?}"
+        );
+    }
+}
+
+/// Every ancestor of the corpus file must land in the searchable dir table.
+fn assert_dirs_indexed(picker: &FilePicker, file_path: &str) {
+    let dirs = picker
+        .get_dirs()
+        .iter()
+        .map(|d| d.relative_path(picker))
+        .collect::<Vec<_>>();
+
+    let mut prefix = String::new();
+    let components: Vec<&str> = file_path.split('/').collect();
+    for component in &components[..components.len() - 1] {
+        prefix.push_str(component);
+        prefix.push('/');
+        assert!(
+            dirs.iter()
+                .any(|d| d.trim_end_matches('/') == prefix.trim_end_matches('/')),
+            "missing dir {prefix:?} in {dirs:?}"
+        );
+    }
+}
+
+/// Materializes the tarball chain under `root`: each entry nests inside the
+/// previous one. Returns the '/'-joined lossy path of the leaf file, or `None`
+/// when the filesystem refuses invalid UTF-8 names (APFS returns EILSEQ, ZFS
+/// with utf8only likewise) so the test skips instead of failing.
+fn build_chain(root: &Path, components: &[(&str, bool)]) -> Option<String> {
+    let mut path = PathBuf::from(root);
+    let mut decoded = String::new();
+
+    for (hex, is_dir) in components {
+        let bytes = unhex(hex);
+        path.push(os_str(&bytes));
+        if !decoded.is_empty() {
+            decoded.push('/');
+        }
+        decoded.push_str(&String::from_utf8_lossy(&bytes));
+
+        if !create(&path, *is_dir) {
+            return None;
+        }
+    }
+
+    Some(decoded)
+}
+
+/// Materializes independent '/'-joined hex paths under `root`.
+fn build_paths(root: &Path, specs: &[(&str, bool)]) -> Option<()> {
+    for (hex, is_dir) in specs {
+        let mut path = PathBuf::from(root);
+        let components: Vec<Vec<u8>> = hex.split('/').map(unhex).collect();
+        let (last, parents) = components.split_last().unwrap();
+
+        for parent in parents {
+            path.push(os_str(parent));
+            if !create(&path, true) {
+                return None;
+            }
+        }
+
+        path.push(os_str(last));
+        if !create(&path, *is_dir) {
+            return None;
+        }
+    }
+    Some(())
+}
+
+fn os_str(bytes: &[u8]) -> &std::ffi::OsStr {
+    use std::os::unix::ffi::OsStrExt;
+    std::ffi::OsStr::from_bytes(bytes)
+}
+
+fn create(path: &Path, is_dir: bool) -> bool {
+    if is_dir {
+        std::fs::create_dir_all(path).is_ok()
+    } else {
+        std::fs::write(path, b"").is_ok()
+    }
+}
+
+fn unhex(s: &str) -> Vec<u8> {
+    (0..s.len())
+        .step_by(2)
+        .map(|i| u8::from_str_radix(&s[i..i + 2], 16).unwrap())
+        .collect()
+}