diff --git a/CHANGELOG.md b/CHANGELOG.md
index 6b153b2..d37b2cf 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -2,6 +2,48 @@
 
 ## Unreleased
 
+### Python bindings: in-place editing (2026-09-27)
+- **`clawhdf5.File(path, 'r+')`** (and `'a'` on an existing file) opens a
+  file for editing through `clawhdf5::FileEditor`, holding its exclusive
+  lock until `close()`:
+  - `ds[key] = value` with h5py's keys (integers, slices with steps, `...`,
+    one increasing index list) and broadcasting (numpy's rules for slices
+    and integers, allowing extra leading length-1 axes; the exact shape for
+    an index list, a scalar only where h5py expands it). A numpy array is
+    converted to the dataset's dtype as libhdf5 converts it in native byte
+    order (integers saturate; floats are truncated toward zero and clipped;
+    integers go into h5py's bool enum by value, as libhdf5 stores them);
+    other values through `numpy.asarray(value, dtype=ds.dtype)`, as h5py
+    does. NaN into an integer dataset is a `ValueError`.
+  - `ds.resize(shape)` / `ds.resize(n, axis=k)` with h5py's argument rules
+    and errors (`TypeError` for a dataset that is not chunked).
+  - `obj.attrs[name] = value`, `attrs.create(name, data, shape, dtype)`,
+    `attrs.modify`: numeric, bool, complex, bytes and `str` data of any
+    shape, stored with the HDF5 types h5py uses (bool as the `FALSE`/`TRUE`
+    enum, complex as the `r`/`i` compound), except that `str` becomes
+    fixed-length UTF-8.
+  - Every edit is written and synced before it returns, then the file is
+    reopened: datasets and attrs objects taken earlier see new shapes and
+    attributes, and reads on other threads wait while an edit is written.
+  - What the editor cannot do raises `NotImplementedError` and writes
+    nothing (deleting attributes or objects, creating datasets or groups,
+    compound fields by name, variable-length data, ...;
+    `docs/known-issues.md`).
+- `File.mode`, `File.flush()` (a no-op), `Dataset.chunks`.
+- Tests (`tests/test_edit.py`): each edit applied to two copies of a file,
+  by h5py and by clawhdf5, and both read back through h5py after every
+  edit, on files h5py writes with `libver` earliest, v114 and latest and on
+  a clawhdf5-written one: a fixed sequence over every chunk index kind,
+  compact/contiguous/gzip layouts and numeric, bool, enum, complex, string
+  and compound types, and 16 random sequences of 40 edits (writes,
+  resizes, attributes); where h5py refuses an edit clawhdf5 must refuse it
+  and leave its file unchanged. A matrix of every numeric source dtype into
+  every numeric dataset dtype at the edge values, dense attribute storage,
+  locking, objects seeing each other's edits, readers racing a writer
+  (never a partly written dataset). Every edited file must pass `h5dump`
+  and, in `ci-test.sh`, `h5rs check`. The read-vs-h5py suite also runs on
+  a file opened `'r+'`.
+
 ### Python bindings: remote files (2026-09-27)
 - **`clawhdf5.File(url)`** opens `http://` URLs (and `https://`, `s3://`,
   `gs://`, `az://` in a wheel built with the `https`, `s3`, `gcs`,
diff --git a/README.md b/README.md
index dccee48..6d8f2a6 100644
--- a/README.md
+++ b/README.md
@@ -542,6 +542,22 @@ f = clawhdf5.File.open_url("http://data.example.org/run42.h5", block_size=256 *
                            headers={"Authorization": "Bearer ..."})
 ```
 
+An existing file opened with `"r+"` is edited in place (through
+`clawhdf5::FileEditor`), with h5py's indexing, broadcasting and numeric
+conversion; each edit is on disk when the statement returns:
+
+```python
+with clawhdf5.File("data.h5", "r+") as f:
+    f["group/temperatures"][100:200, ::4] = 0.0
+    f["series"].resize(5000, axis=0)   # chunked datasets, within maxshape
+    f["series"][4000:] = new_values
+    f["group"].attrs["calibrated"] = True
+```
+
+Creating or deleting datasets, groups and attributes in an existing file is
+not supported (`NotImplementedError`); limits are in
+[known issues](docs/known-issues.md).
+
 The default build reads `http://` URLs only; build with
 `maturin develop --release --features https` (rustls with ring, which
 compiles C) for `https://`, and `--features s3` (or `gcs`, `azure`) for
@@ -561,7 +577,8 @@ non-default fill value (`docs/known-issues.md`). An index list is read one
 group of neighbouring chunks at a time.
 Writing (`File(path, "w")`, `create_dataset`, `create_group`, `attrs[...] =`)
 covers `float64`, `float32`, `int64`, `int32` and `uint8` arrays. The tests
-in `crates/clawhdf5-py/tests` compare every read with h5py; run them with
+in `crates/clawhdf5-py/tests` compare every read and every in-place edit
+with h5py; run them with
 `pip install pytest h5py && pytest crates/clawhdf5-py/tests`.
 
 ### Agent Memory
diff --git a/crates/clawhdf5-py/README.md b/crates/clawhdf5-py/README.md
index 16e4169..50b0d17 100644
--- a/crates/clawhdf5-py/README.md
+++ b/crates/clawhdf5-py/README.md
@@ -98,6 +98,41 @@ chunks=..., compression="gzip")`, `create_group` and `attrs[...] = ...`
 writes `float64`, `float32`, `int64`, `int32` and `uint8` arrays; the file is
 written on `close()`.
 
+## Editing a file in place
+
+`clawhdf5.File(path, "r+")` (or `"a"` on an existing file) edits the file
+where it is, through clawhdf5's `FileEditor`; the file is locked until
+`close()`, and every edit is written and synced before the statement
+returns.
+
+```python
+with clawhdf5.File("data.h5", "r+") as f:
+    ds = f["grid"]
+    ds[10:20, ::2] = 0                  # h5py keys and broadcasting
+    ds[[1, 4, 7], 3] = [1.5, 2.5, 3.5]  # one index list: exact shape
+    f["series"].resize((5000, 3))       # or .resize(5000, axis=0)
+    f["series"].attrs["units"] = "K"
+    f.attrs.create("version", 2, dtype="u1")
+```
+
+- Values: a numpy array is converted to the dataset's dtype as libhdf5
+  converts it (integers saturate at the target's limits; floats are
+  truncated toward zero and clipped); anything else goes through
+  `numpy.asarray(value, dtype=ds.dtype)`, as in h5py. Writing NaN into an
+  integer dataset raises `ValueError` (libhdf5 would store an arbitrary
+  value). A few libhdf5 edge cases differ on purpose; see
+  `docs/known-issues.md`.
+- Shapes: `ds.resize` grows or shrinks chunked datasets within their
+  `maxshape`, as h5py; datasets and `attrs` objects taken before an edit
+  see its result.
+- Attributes: numeric, bool, complex, bytes and `str` data of any shape.
+  `str` is stored as a fixed-length UTF-8 string (h5py stores a
+  variable-length one), so h5py reads it back as `bytes`.
+- Not supported (`NotImplementedError`, nothing written): creating or
+  deleting datasets, groups and attributes, writing compound fields by
+  name, variable-length data, HDF5 array types, and whatever
+  `FileEditor` refuses (listed in `docs/known-issues.md`).
+
 ## Tests
 
 ```bash
@@ -108,8 +143,10 @@ pytest crates/clawhdf5-py/tests
 `tests/test_read_vs_h5py.py` compares every read with h5py on a file h5py
 writes, opened locally and over HTTP (an in-process range server,
 `tests/conftest.py`); `tests/test_remote.py` checks remote reads (requests,
-failures, the GIL). `scripts/ci-test.sh` builds the wheel and runs these
-in CI.
+failures, the GIL); `tests/test_edit.py` applies every edit through h5py and
+clawhdf5 to copies of a file and compares them through h5py (and `h5dump`,
+and `h5rs check` when `CLAWHDF5_H5RS` names it). `scripts/ci-test.sh` builds
+the wheel and runs these in CI.
 
 ## License
 
diff --git a/crates/clawhdf5-py/src/attrs.rs b/crates/clawhdf5-py/src/attrs.rs
index 615214b..5b3f365 100644
--- a/crates/clawhdf5-py/src/attrs.rs
+++ b/crates/clawhdf5-py/src/attrs.rs
@@ -1,23 +1,62 @@
 //! PyAttrs — dict-like access to HDF5 attributes.
 
-use std::sync::{Arc, Mutex};
+use std::sync::{Arc, Mutex, PoisonError};
 
 use clawhdf5_format::attribute::AttributeMessage;
-use pyo3::exceptions::{PyKeyError, PyTypeError, PyValueError};
+use pyo3::exceptions::{PyKeyError, PyNotImplementedError, PyTypeError, PyValueError};
 use pyo3::prelude::*;
 use pyo3::types::{PyList, PyTuple};
 
 use crate::convert::{Converter, Elements, resolve_vl};
 use crate::handle::Handle;
-use crate::{OwnedAttrValue, PyEmpty, attr_value_to_py, node, py_to_attr_value};
+use crate::{OwnedAttrValue, PyEmpty, attr_value_to_py, edit, node, py_to_attr_value};
+
+/// The attributes of an object in a file opened for reading (or editing).
+struct ReadAttrs {
+    handle: Arc<Handle>,
+    addr: u64,
+    path: String,
+    /// Sorted by name, with the file generation they were read at: an edit
+    /// (`attrs[name] = value`, here or through another handle on the same
+    /// object) makes them re-read.
+    cache: Mutex<(u64, Arc<Vec<AttributeMessage>>)>,
+}
+
+impl ReadAttrs {
+    fn current(&self, py: Python<'_>) -> PyResult<Arc<Vec<AttributeMessage>>> {
+        let generation = self.handle.generation();
+        {
+            let cached = self.cache.lock().unwrap_or_else(PoisonError::into_inner);
+            if cached.0 == generation {
+                return Ok(Arc::clone(&cached.1));
+            }
+        }
+        let (addr, path) = (self.addr, &self.path);
+        let attrs = Arc::new(self.handle.with(py, |f| node::attributes(f, addr, path))?);
+        *self.cache.lock().unwrap_or_else(PoisonError::into_inner) =
+            (generation, Arc::clone(&attrs));
+        Ok(attrs)
+    }
+
+    fn check_writable(&self) -> PyResult<()> {
+        if self.handle.is_writable() {
+            return Ok(());
+        }
+        Err(PyErr::new::<pyo3::exceptions::PyIOError, _>(
+            "cannot set attributes on a read-only file (open it with mode 'r+')",
+        ))
+    }
+
+    fn set(&self, py: Python<'_>, name: &str, value: clawhdf5_rs::AttrValue) -> PyResult<()> {
+        self.check_writable()?;
+        let path = node::name(&self.path);
+        self.handle.edit(py, |ed| ed.set_attr(&path, name, &value))
+    }
+}
 
 /// Backing storage for attributes.
 enum AttrsInner {
-    /// Attributes of an object in a file opened for reading, sorted by name.
-    Read {
-        handle: Arc<Handle>,
-        attrs: Vec<AttributeMessage>,
-    },
+    Read(ReadAttrs),
     /// Writable attribute list shared with a parent (PyFile or PyGroup).
     Write(Arc<Mutex<Vec<(String, OwnedAttrValue)>>>),
 }
@@ -27,8 +66,11 @@ enum AttrsInner {
 /// In read mode, values are what h5py returns: numpy scalars for scalar
 /// attributes, numpy arrays otherwise, `str` for variable-length strings,
 /// `numpy.bytes_` for fixed-length ones, and `Empty` for a null dataspace.
-/// In write mode, attributes set here are accumulated and written when
-/// the parent file is closed.
+/// In a file opened with `'r+'`, `attrs[name] = value` adds or replaces an
+/// attribute in the file at once (as h5py stores it, except that `str`
+/// values become fixed-length UTF-8 strings). In write mode (`'w'`),
+/// attributes set here are accumulated and written when the parent file is
+/// closed.
 #[pyclass(name = "Attrs")]
 pub struct PyAttrs {
     inner: AttrsInner,
@@ -43,9 +85,15 @@ impl PyAttrs {
         addr: u64,
         path: &str,
     ) -> PyResult<Self> {
-        let attrs = handle.with(py, |f| node::attributes(f, addr, path))?;
+        let generation = handle.generation();
+        let attrs = Arc::new(handle.with(py, |f| node::attributes(f, addr, path))?);
         Ok(Self {
-            inner: AttrsInner::Read { handle, attrs },
+            inner: AttrsInner::Read(ReadAttrs {
+                handle,
+                addr,
+                path: path.to_string(),
+                cache: Mutex::new((generation, attrs)),
+            }),
         })
     }
 
@@ -55,14 +103,48 @@ impl PyAttrs {
             inner: AttrsInner::Write(store),
         }
     }
+
+    fn set_value(
+        &self,
+        py: Python<'_>,
+        key: &str,
+        value: &Bound<'_, PyAny>,
+        dtype: Option<&Bound<'_, PyAny>>,
+        shape: Option<&Bound<'_, PyAny>>,
+    ) -> PyResult<()> {
+        match &self.inner {
+            AttrsInner::Read(r) => {
+                r.check_writable()?;
+                let value = edit::attr_value(py, value, dtype, shape)?;
+                r.set(py, key, value)
+            }
+            AttrsInner::Write(store) => {
+                if dtype.is_some() || shape.is_some() {
+                    return Err(PyNotImplementedError::new_err(
+                        "attrs.create with a dtype or shape is only supported in a file opened \
+                         with 'r+'",
+                    ));
+                }
+                let owned = py_to_attr_value(value)?;
+                let mut guard = store.lock().unwrap();
+                // Replace existing key if present.
+                if let Some(entry) = guard.iter_mut().find(|(k, _)| k == key) {
+                    entry.1 = owned;
+                } else {
+                    guard.push((key.to_string(), owned));
+                }
+                Ok(())
+            }
+        }
+    }
 }
 
 #[pymethods]
 impl PyAttrs {
     fn __getitem__(&self, py: Python<'_>, key: &str) -> PyResult<Py<PyAny>> {
         match &self.inner {
-            AttrsInner::Read { handle, attrs } => match attrs.iter().find(|a| a.name == key) {
-                Some(attr) => Ok(attr_to_py(py, handle, attr)?.unbind()),
+            AttrsInner::Read(r) => match r.current(py)?.iter().find(|a| a.name == key) {
+                Some(attr) => Ok(attr_to_py(py, &r.handle, attr)?.unbind()),
                 None => Err(PyKeyError::new_err(format!(
                     "Can't open attribute (can't locate attribute: '{key}')"
                 ))),
@@ -80,36 +162,63 @@ impl PyAttrs {
         }
     }
 
-    fn __setitem__(&self, key: &str, value: &Bound<'_, PyAny>) -> PyResult<()> {
+    /// `attrs[name] = value`. In a file opened with `'r+'` this writes the
+    /// attribute (numeric, bool, complex, bytes and str data, any shape)
+    /// into the file before returning; see the class docs.
+    fn __setitem__(&self, py: Python<'_>, key: &str, value: &Bound<'_, PyAny>) -> PyResult<()> {
+        self.set_value(py, key, value, None, None)
+    }
+
+    /// Deleting attributes is not supported in a file (the in-place editor
+    /// cannot remove them); in write mode it removes a pending attribute.
+    fn __delitem__(&self, key: &str) -> PyResult<()> {
         match &self.inner {
-            AttrsInner::Read { .. } => Err(PyErr::new::<pyo3::exceptions::PyIOError, _>(
-                "cannot set attributes on a read-only file",
-            )),
+            AttrsInner::Read(_) => Err(PyNotImplementedError::new_err(format!(
+                "cannot delete attribute '{key}': deleting attributes is not supported by \
+                 clawhdf5's in-place editor"
+            ))),
             AttrsInner::Write(store) => {
-                let owned = py_to_attr_value(value)?;
                 let mut guard = store.lock().unwrap();
-                // Replace existing key if present.
-                if let Some(entry) = guard.iter_mut().find(|(k, _)| k == key) {
-                    entry.1 = owned;
-                } else {
-                    guard.push((key.to_string(), owned));
+                let before = guard.len();
+                guard.retain(|(k, _)| k != key);
+                if guard.len() == before {
+                    return Err(PyKeyError::new_err(key.to_string()));
                 }
                 Ok(())
             }
         }
     }
 
-    fn __len__(&self) -> usize {
+    /// h5py's `attrs.create(name, data, shape=None, dtype=None)`: `data`
+    /// converted to `dtype` and reshaped to `shape` first.
+    #[pyo3(signature = (name, data, shape=None, dtype=None))]
+    fn create(
+        &self,
+        py: Python<'_>,
+        name: &str,
+        data: &Bound<'_, PyAny>,
+        shape: Option<&Bound<'_, PyAny>>,
+        dtype: Option<&Bound<'_, PyAny>>,
+    ) -> PyResult<()> {
+        self.set_value(py, name, data, dtype, shape)
+    }
+
+    /// h5py's `attrs.modify(name, value)`: same as `attrs[name] = value`.
+    fn modify(&self, py: Python<'_>, name: &str, value: &Bound<'_, PyAny>) -> PyResult<()> {
+        self.set_value(py, name, value, None, None)
+    }
+
+    fn __len__(&self, py: Python<'_>) -> PyResult<usize> {
         match &self.inner {
-            AttrsInner::Read { attrs, .. } => attrs.len(),
-            AttrsInner::Write(store) => store.lock().unwrap().len(),
+            AttrsInner::Read(r) => Ok(r.current(py)?.len()),
+            AttrsInner::Write(store) => Ok(store.lock().unwrap().len()),
         }
     }
 
-    fn __contains__(&self, key: &str) -> bool {
+    fn __contains__(&self, py: Python<'_>, key: &str) -> PyResult<bool> {
         match &self.inner {
-            AttrsInner::Read { attrs, .. } => attrs.iter().any(|a| a.name == key),
-            AttrsInner::Write(store) => store.lock().unwrap().iter().any(|(k, _)| k == key),
+            AttrsInner::Read(r) => Ok(r.current(py)?.iter().any(|a| a.name == key)),
+            AttrsInner::Write(store) => Ok(store.lock().unwrap().iter().any(|(k, _)| k == key)),
         }
     }
 
@@ -119,15 +228,17 @@ impl PyAttrs {
         Ok(iter)
     }
 
-    fn __repr__(&self) -> String {
-        let n = self.__len__();
-        format!("<HDF5 Attrs ({n} members)>")
+    fn __repr__(&self, py: Python<'_>) -> String {
+        match self.__len__(py) {
+            Ok(n) => format!("<HDF5 Attrs ({n} members)>"),
+            Err(_) => "<HDF5 Attrs>".to_string(),
+        }
     }
 
     /// The value of `key`, or `default` if there is no such attribute.
     #[pyo3(signature = (key, default=None))]
     fn get(&self, py: Python<'_>, key: &str, default: Option<Py<PyAny>>) -> PyResult<Py<PyAny>> {
-        if self.__contains__(key) {
+        if self.__contains__(py, key)? {
             self.__getitem__(py, key)
         } else {
             Ok(default.unwrap_or_else(|| py.None()))
@@ -137,7 +248,7 @@ impl PyAttrs {
     /// Return attribute names as a list.
     fn keys(&self, py: Python<'_>) -> PyResult<Py<PyAny>> {
         let names: Vec<String> = match &self.inner {
-            AttrsInner::Read { attrs, .. } => attrs.iter().map(|a| a.name.clone()).collect(),
+            AttrsInner::Read(r) => r.current(py)?.iter().map(|a| a.name.clone()).collect(),
             AttrsInner::Write(store) => store
                 .lock()
                 .unwrap()
@@ -152,9 +263,10 @@ impl PyAttrs {
     /// Return attribute values as a list.
     fn values(&self, py: Python<'_>) -> PyResult<Py<PyAny>> {
         let vals: Vec<Py<PyAny>> = match &self.inner {
-            AttrsInner::Read { handle, attrs } => attrs
+            AttrsInner::Read(r) => r
+                .current(py)?
                 .iter()
-                .map(|a| attr_to_py(py, handle, a).map(Bound::unbind))
+                .map(|a| attr_to_py(py, &r.handle, a).map(Bound::unbind))
                 .collect::<PyResult<_>>()?,
             AttrsInner::Write(store) => store
                 .lock()
@@ -173,9 +285,10 @@ impl PyAttrs {
     /// Return attribute (key, value) pairs as a list of tuples.
     fn items(&self, py: Python<'_>) -> PyResult<Py<PyAny>> {
         let pairs: Vec<(String, Py<PyAny>)> = match &self.inner {
-            AttrsInner::Read { handle, attrs } => attrs
+            AttrsInner::Read(r) => r
+                .current(py)?
                 .iter()
-                .map(|a| Ok((a.name.clone(), attr_to_py(py, handle, a)?.unbind())))
+                .map(|a| Ok((a.name.clone(), attr_to_py(py, &r.handle, a)?.unbind())))
                 .collect::<PyResult<_>>()?,
             AttrsInner::Write(store) => store
                 .lock()
@@ -248,19 +361,3 @@ fn prefix_err(py: Python<'_>, name: &str, e: PyErr) -> PyErr {
         PyValueError::new_err(msg)
     }
 }
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn write_attrs_len() {
-        let store = Arc::new(Mutex::new(Vec::new()));
-        store
-            .lock()
-            .unwrap()
-            .push(("key".into(), OwnedAttrValue::I64(99)));
-        let attrs = PyAttrs::from_write(store);
-        assert_eq!(attrs.__len__(), 1);
-    }
-}
diff --git a/crates/clawhdf5-py/src/dataset.rs b/crates/clawhdf5-py/src/dataset.rs
index 24c9b51..d837f73 100644
--- a/crates/clawhdf5-py/src/dataset.rs
+++ b/crates/clawhdf5-py/src/dataset.rs
@@ -9,12 +9,12 @@
 //! Python threads reading the same or different datasets run in parallel,
 //! and a remote file's network reads never hold the GIL.
 
-use std::sync::Arc;
+use std::sync::{Arc, Mutex, PoisonError};
 
 use clawhdf5_format::datatype::Datatype;
 use clawhdf5_format::object_header::ObjectHeader;
 use clawhdf5_rs::File;
-use pyo3::exceptions::{PyTypeError, PyValueError};
+use pyo3::exceptions::{PyNotImplementedError, PyOSError, PyTypeError, PyValueError};
 use pyo3::prelude::*;
 use pyo3::types::{PyList, PyTuple};
 
@@ -22,7 +22,7 @@ use crate::attrs::PyAttrs;
 use crate::convert::{Converter, Elements, VlError, resolve_vl};
 use crate::handle::Handle;
 use crate::select::{self, Plan};
-use crate::{PyEmpty, node, to_py_err};
+use crate::{PyEmpty, edit, node, to_py_err};
 
 /// What opening a dataset reads from the file (without the GIL).
 pub(crate) struct DatasetMeta {
@@ -67,8 +67,9 @@ pub struct PyDataset {
     /// Where the dataset's object header is: reads open it from here rather
     /// than resolve `path` again.
     addr: u64,
-    /// `None` for a dataset with a null dataspace (h5py's `Empty`).
-    shape: Option<Vec<u64>>,
+    /// The shape (`None` for a null dataspace, h5py's `Empty`), with the
+    /// file generation it was read at: an edit (a resize) may change it.
+    shape: Mutex<(u64, Option<Vec<u64>>)>,
     /// The chunk shape, for a chunked dataset.
     chunks: Option<Vec<u64>>,
     datatype: Datatype,
@@ -84,13 +85,14 @@ impl PyDataset {
         addr: u64,
         meta: DatasetMeta,
     ) -> Self {
+        let generation = handle.generation();
         let conv = crate::no_panic(|| Converter::new(py, &meta.datatype, handle.offset_size))
             .map_err(|e| e.value(py).to_string());
         Self {
             handle,
             path,
             addr,
-            shape: meta.shape,
+            shape: Mutex::new((generation, meta.shape)),
             chunks: meta.chunks,
             datatype: meta.datatype,
             conv,
@@ -103,10 +105,54 @@ impl PyDataset {
             .map_err(|msg| PyTypeError::new_err(format!("{}: {msg}", node::name(&self.path))))
     }
 
+    /// The current shape: the one read at open, or re-read after an edit.
+    fn dims(&self, py: Python<'_>) -> PyResult<Option<Vec<u64>>> {
+        let generation = self.handle.generation();
+        {
+            let cached = self.shape.lock().unwrap_or_else(PoisonError::into_inner);
+            if cached.0 == generation {
+                return Ok(cached.1.clone());
+            }
+        }
+        let addr = self.addr;
+        let null = self
+            .shape
+            .lock()
+            .unwrap_or_else(PoisonError::into_inner)
+            .1
+            .is_none();
+        let shape = if null {
+            None
+        } else {
+            Some(self.handle.with(py, |f| {
+                f.dataset_at(addr)
+                    .and_then(|ds| ds.shape())
+                    .map_err(to_py_err)
+            })?)
+        };
+        *self.shape.lock().unwrap_or_else(PoisonError::into_inner) = (generation, shape.clone());
+        Ok(shape)
+    }
+
+    fn check_writable(&self) -> PyResult<()> {
+        if self.handle.is_writable() {
+            Ok(())
+        } else {
+            Err(PyOSError::new_err(format!(
+                "{}: the file is open read-only; open it with mode 'r+' to change it",
+                node::name(&self.path)
+            )))
+        }
+    }
+
     /// Read the selection described by `plan` into a numpy array.
-    fn read_plan<'py>(&self, py: Python<'py>, plan: &Plan) -> PyResult<Bound<'py, PyAny>> {
+    fn read_plan<'py>(
+        &self,
+        py: Python<'py>,
+        plan: &Plan,
+        dims: &[u64],
+    ) -> PyResult<Bound<'py, PyAny>> {
         let conv = self.converter()?;
-        let dims = self.shape.as_deref().unwrap_or(&[]);
         let out_shape = plan.out_shape();
 
         let arr = if plan.is_empty() {
@@ -272,7 +318,7 @@ impl PyDataset {
     /// The shape of the dataset (`None` for an empty/null dataspace).
     #[getter]
     fn shape<'py>(&self, py: Python<'py>) -> PyResult<Bound<'py, PyAny>> {
-        match &self.shape {
+        match self.dims(py)? {
             Some(s) => Ok(PyTuple::new(py, s)?.into_any()),
             None => Ok(py.None().into_bound(py)),
         }
@@ -281,7 +327,7 @@ impl PyDataset {
     /// The maximum shape (`None` per unlimited dimension), like h5py.
     #[getter]
     fn maxshape<'py>(&self, py: Python<'py>) -> PyResult<Bound<'py, PyAny>> {
-        let Some(shape) = &self.shape else {
+        let Some(shape) = self.dims(py)? else {
             return Ok(py.None().into_bound(py));
         };
         let addr = self.addr;
@@ -292,7 +338,7 @@ impl PyDataset {
                     .and_then(|ds| ds.max_dimensions())
                     .map_err(to_py_err)
             })?
-            .unwrap_or_else(|| shape.clone());
+            .unwrap_or(shape);
         let items: Vec<Option<u64>> = max
             .into_iter()
             .map(|d| (d != u64::MAX).then_some(d))
@@ -300,6 +346,15 @@ impl PyDataset {
         Ok(PyTuple::new(py, items)?.into_any())
     }
 
+    /// The chunk shape, or `None` for a dataset that is not chunked.
+    #[getter]
+    fn chunks<'py>(&self, py: Python<'py>) -> PyResult<Bound<'py, PyAny>> {
+        match &self.chunks {
+            Some(c) => Ok(PyTuple::new(py, c)?.into_any()),
+            None => Ok(py.None().into_bound(py)),
+        }
+    }
+
     /// The dataset's numpy dtype, as h5py reports it.
     #[getter]
     fn dtype<'py>(&self, py: Python<'py>) -> PyResult<Bound<'py, PyAny>> {
@@ -307,14 +362,14 @@ impl PyDataset {
     }
 
     #[getter]
-    fn ndim(&self) -> usize {
-        self.shape.as_ref().map_or(0, Vec::len)
+    fn ndim(&self, py: Python<'_>) -> PyResult<usize> {
+        Ok(self.dims(py)?.map_or(0, |s| s.len()))
     }
 
     /// Number of elements (`None` for an empty/null dataspace, as h5py).
     #[getter]
-    fn size(&self) -> Option<u64> {
-        self.shape.as_ref().map(|s| s.iter().product())
+    fn size(&self, py: Python<'_>) -> PyResult<Option<u64>> {
+        Ok(self.dims(py)?.map(|s| s.iter().product()))
     }
 
     /// The dataset's full name, e.g. `/group/data`.
@@ -323,7 +378,8 @@ impl PyDataset {
         node::name(&self.path)
     }
 
-    /// The dataset's attributes (read-only, dict-like).
+    /// The dataset's attributes (dict-like; writable in a file opened with
+    /// `'r+'`).
     #[getter]
     fn attrs(&self, py: Python<'_>) -> PyResult<PyAttrs> {
         PyAttrs::read(py, Arc::clone(&self.handle), self.addr, &self.path)
@@ -338,7 +394,7 @@ impl PyDataset {
         py: Python<'py>,
         key: &Bound<'py, PyAny>,
     ) -> PyResult<Bound<'py, PyAny>> {
-        let Some(dims) = &self.shape else {
+        let Some(dims) = self.dims(py)? else {
             let is_empty_tuple = key.cast::<PyTuple>().is_ok_and(|t| t.is_empty());
             let is_ellipsis = key.is_instance_of::<pyo3::types::PyEllipsis>();
             if is_empty_tuple || is_ellipsis {
@@ -347,8 +403,109 @@ impl PyDataset {
             }
             return Err(PyValueError::new_err("Empty datasets cannot be sliced"));
         };
-        let plan = select::parse(key, dims)?;
-        self.read_plan(py, &plan)
+        let plan = select::parse(key, &dims)?;
+        self.read_plan(py, &plan, &dims)
+    }
+
+    /// Write with h5py indexing (file opened with `'r+'`): `ds[key] = value`.
+    ///
+    /// The key is what `ds[key]` reads (without compound field names). The
+    /// value is converted to the dataset's dtype as h5py converts it (a
+    /// numpy array as libhdf5 does, clipping out-of-range numbers; anything
+    /// else through `numpy.asarray(value, dtype=ds.dtype)`), and broadcast
+    /// to the selection as h5py broadcasts. The edit is written and synced
+    /// before this returns; what the in-place editor cannot write raises
+    /// `NotImplementedError` and leaves the file as it was.
+    fn __setitem__(
+        &self,
+        py: Python<'_>,
+        key: &Bound<'_, PyAny>,
+        value: &Bound<'_, PyAny>,
+    ) -> PyResult<()> {
+        self.check_writable()?;
+        let Some(dims) = self.dims(py)? else {
+            return Err(PyNotImplementedError::new_err(
+                "writing to an empty (null dataspace) dataset is not supported",
+            ));
+        };
+        let plan = select::parse(key, &dims)?;
+        if !plan.fields.is_empty() {
+            return Err(PyNotImplementedError::new_err(
+                "writing compound fields by name is not supported by clawhdf5's in-place editor; \
+                 write whole elements",
+            ));
+        }
+        let category = edit::category(&self.datatype)?;
+        let conv = self.converter()?;
+        let bytes = edit::dataset_bytes(
+            py,
+            value,
+            conv.dtype.bind(py),
+            category,
+            &plan,
+            self.chunks.as_deref(),
+        )?;
+        if plan.is_empty() {
+            return Ok(());
+        }
+        let sel = edit::selection(&plan, &dims)?;
+        let path = node::name(&self.path);
+        self.handle
+            .edit(py, |ed| ed.write_selection(&path, &sel, &bytes))
+    }
+
+    /// Change the dataset's shape (file opened with `'r+'`), as h5py's
+    /// `Dataset.resize`: `ds.resize((100, 20))`, or `ds.resize(100, axis=0)`.
+    /// Only chunked datasets, within their maximum shape; new elements read
+    /// as the fill value.
+    #[pyo3(signature = (size, axis=None))]
+    fn resize(&self, py: Python<'_>, size: &Bound<'_, PyAny>, axis: Option<isize>) -> PyResult<()> {
+        self.check_writable()?;
+        let Some(dims) = self.dims(py)? else {
+            return Err(PyTypeError::new_err("Empty datasets cannot be resized"));
+        };
+        if self.chunks.is_none() {
+            return Err(PyTypeError::new_err("Only chunked datasets can be resized"));
+        }
+        let shape: Vec<u64> = match axis {
+            Some(axis) => {
+                let rank = dims.len();
+                let a = usize::try_from(axis)
+                    .ok()
+                    .filter(|&a| a < rank)
+                    .ok_or_else(|| {
+                        PyValueError::new_err(format!(
+                            "Invalid axis (0 to {} allowed)",
+                            rank.saturating_sub(1)
+                        ))
+                    })?;
+                let n: u64 = size.extract().map_err(|_| {
+                    PyTypeError::new_err("Argument must be a single int if axis is specified")
+                })?;
+                let mut s = dims.clone();
+                s[a] = n;
+                s
+            }
+            // As h5py: without `axis` the size is a sequence (`tuple(size)`).
+            None => size.extract().map_err(|_| {
+                PyTypeError::new_err(format!(
+                    "'{}' object is not iterable",
+                    size.get_type()
+                        .name()
+                        .map(|n| n.to_string())
+                        .unwrap_or_default()
+                ))
+            })?,
+        };
+        if shape.len() != dims.len() {
+            return Err(PyValueError::new_err(format!(
+                "new shape {shape:?} has {} dimensions, the dataset {}",
+                shape.len(),
+                dims.len()
+            )));
+        }
+        let path = node::name(&self.path);
+        self.handle.edit(py, |ed| ed.resize(&path, &shape))
     }
 
     /// `numpy.asarray(ds)` reads the whole dataset.
@@ -360,20 +517,20 @@ impl PyDataset {
         copy: Option<bool>,
     ) -> PyResult<Bound<'py, PyAny>> {
         let _ = copy; // every read is a fresh array
-        let Some(dims) = &self.shape else {
+        let Some(dims) = self.dims(py)? else {
             return Err(PyValueError::new_err("an empty dataset has no array value"));
         };
         let ellipsis = pyo3::types::PyEllipsis::get(py).to_owned().into_any();
-        let plan = select::parse(&ellipsis, dims)?;
-        let arr = self.read_plan(py, &plan)?;
+        let plan = select::parse(&ellipsis, &dims)?;
+        let arr = self.read_plan(py, &plan, &dims)?;
         match dtype {
             Some(dt) => arr.call_method1("astype", (dt,)),
             None => Ok(arr),
         }
     }
 
-    fn __len__(&self) -> PyResult<usize> {
-        match self.shape.as_deref() {
+    fn __len__(&self, py: Python<'_>) -> PyResult<usize> {
+        match self.dims(py)?.as_deref() {
             Some([first, ..]) => Ok(*first as usize),
             _ => Err(PyTypeError::new_err(
                 "Attempt to take len() of scalar dataset",
@@ -391,9 +548,10 @@ impl PyDataset {
                 .unwrap_or_default(),
             Err(_) => format!("{:?}", self.datatype),
         };
-        let shape = match &self.shape {
-            Some(s) => format!("{s:?}"),
-            None => "None".to_string(),
+        let shape = match self.dims(py) {
+            Ok(Some(s)) => format!("{s:?}"),
+            Ok(None) => "None".to_string(),
+            Err(_) => "?".to_string(),
         };
         format!(
             "<HDF5 dataset \"{}\": shape {shape}, type \"{dtype}\">",
diff --git a/crates/clawhdf5-py/src/edit.rs b/crates/clawhdf5-py/src/edit.rs
new file mode 100644
index 0000000..05dd536
--- /dev/null
+++ b/crates/clawhdf5-py/src/edit.rs
@@ -0,0 +1,387 @@
+//! In-place editing (`clawhdf5.File(path, 'r+')`) through `FileEditor`:
+//! turning what Python assigns into the bytes, selections and attribute
+//! values the editor takes.
+//!
+//! Value conversion follows h5py (see `edit_helpers.py`, run inside the
+//! extension module); what `FileEditor` cannot do is `NotImplementedError`
+//! before anything is written.
+
+use std::ffi::CString;
+
+use clawhdf5_format::datatype::{
+    CharacterSet, CompoundMember, Datatype, DatatypeByteOrder, EnumMember, StringPadding,
+};
+use clawhdf5_format::selection::Selection;
+use clawhdf5_rs::AttrValue;
+use pyo3::exceptions::{PyNotImplementedError, PyTypeError};
+use pyo3::prelude::*;
+use pyo3::sync::PyOnceLock;
+use pyo3::types::{PyBytes, PyModule, PyTuple};
+
+use crate::select::{Axis, Plan};
+
+/// The helper module, compiled once.
+pub(crate) fn helpers(py: Python<'_>) -> PyResult<&Bound<'_, PyModule>> {
+    static HELPERS: PyOnceLock<Py<PyModule>> = PyOnceLock::new();
+    let module = HELPERS.get_or_try_init(py, || -> PyResult<Py<PyModule>> {
+        let code = CString::new(include_str!("edit_helpers.py"))
+            .map_err(|e| PyTypeError::new_err(e.to_string()))?;
+        Ok(PyModule::from_code(
+            py,
+            &code,
+            c"clawhdf5/edit_helpers.py",
+            c"clawhdf5._edit_helpers",
+        )?
+        .unbind())
+    })?;
+    Ok(module.bind(py))
+}
+
+fn not_implemented(what: impl std::fmt::Display) -> PyErr {
+    PyNotImplementedError::new_err(format!(
+        "{what} is not supported by clawhdf5's in-place editor"
+    ))
+}
+
+/// How values for a dataset of type `dt` are converted (a category of
+/// `edit_helpers._convert_array`), or why they cannot be written.
+pub(crate) fn category(dt: &Datatype) -> PyResult<&'static str> {
+    match dt {
+        Datatype::FixedPoint { .. } => Ok("int"),
+        Datatype::FloatingPoint { .. } => Ok("float"),
+        Datatype::Enumeration {
+            base_type, members, ..
+        } => {
+            let is_bool = base_type.type_size() == 1
+                && members.len() == 2
+                && members
+                    .iter()
+                    .any(|m| m.name == "FALSE" && m.value.first() == Some(&0))
+                && members
+                    .iter()
+                    .any(|m| m.name == "TRUE" && m.value.first() == Some(&1));
+            Ok(if is_bool { "bool" } else { "enum" })
+        }
+        Datatype::String {
+            padding: StringPadding::NullPad,
+            ..
+        } => Ok("string"),
+        Datatype::String { padding, .. } => Err(not_implemented(format!(
+            "writing fixed-length strings padded {padding:?} (libhdf5 converts them \
+             differently from numpy)"
+        ))),
+        Datatype::Compound { size, members } => {
+            if is_complex(*size, members) {
+                return Ok("complex");
+            }
+            check_exact(dt)?;
+            Ok("exact")
+        }
+        Datatype::Opaque { .. } => Ok("exact"),
+        Datatype::Array { .. } => Err(not_implemented("writing HDF5 array-type elements")),
+        Datatype::VariableLength { .. } => Err(not_implemented("writing variable-length data")),
+        Datatype::Reference { .. } => Err(not_implemented("writing references")),
+        Datatype::BitField { .. } => Err(not_implemented("writing bitfields")),
+        Datatype::Time { .. } => Err(not_implemented("writing time values")),
+    }
+}
+
+/// h5py's complex numbers: a compound of two identical floats `r`, `i`.
+fn is_complex(size: u32, members: &[CompoundMember]) -> bool {
+    matches!(members, [r, i] if r.name == "r" && i.name == "i"
+        && r.datatype == i.datatype
+        && matches!(r.datatype, Datatype::FloatingPoint { size: fs, .. }
+            if r.byte_offset == 0 && i.byte_offset == u64::from(fs) && size == 2 * fs))
+}
+
+/// Compound members written byte for byte from the same numpy dtype: fine
+/// unless libhdf5 would convert them on the way (strings padded other than
+/// with NULs), or the editor cannot write them at all.
+fn check_exact(dt: &Datatype) -> PyResult<()> {
+    match dt {
+        Datatype::Compound { members, .. } => {
+            members.iter().try_for_each(|m| check_exact(&m.datatype))
+        }
+        Datatype::Array { base_type, .. } => check_exact(base_type),
+        Datatype::String {
+            padding: StringPadding::NullPad,
+            ..
+        }
+        | Datatype::FixedPoint { .. }
+        | Datatype::FloatingPoint { .. }
+        | Datatype::Enumeration { .. }
+        | Datatype::Opaque { .. }
+        | Datatype::BitField { .. } => Ok(()),
+        Datatype::String { .. } => Err(not_implemented(
+            "writing compounds with strings not padded with NULs",
+        )),
+        Datatype::VariableLength { .. } => Err(not_implemented(
+            "writing compounds with variable-length members",
+        )),
+        Datatype::Reference { .. } => Err(not_implemented("writing references")),
+        Datatype::Time { .. } => Err(not_implemented("writing time values")),
+    }
+}
+
+/// Largest point selection an index-list write builds (one coordinate
+/// vector per element).
+const MAX_POINTS: usize = 1 << 22;
+
+/// The selection `plan` writes, whose elements are numbered as the value's
+/// (row-major over the selection's shape).
+pub(crate) fn selection(plan: &Plan, dims: &[u64]) -> PyResult<Selection> {
+    if plan.axes.is_empty() {
+        return Ok(Selection::All);
+    }
+    if plan.list_axis().is_none() {
+        let (reads, _) = plan.reads(dims, None, 1);
+        return match <[_; 1]>::try_from(reads) {
+            Ok([read]) => Ok(read.sel),
+            Err(_) => Err(PyTypeError::new_err("internal error: several hyperslabs")),
+        };
+    }
+    // An index list: the points, in the value's order.
+    let per_axis: Vec<Vec<u64>> = plan
+        .axes
+        .iter()
+        .map(|a| match a {
+            Axis::Index(i) => vec![*i],
+            Axis::Slice { start, step, count } => (0..*count).map(|k| start + k * step).collect(),
+            Axis::List(v) => v.clone(),
+        })
+        .collect();
+    let n = per_axis
+        .iter()
+        .try_fold(1usize, |acc, v| acc.checked_mul(v.len()))
+        .filter(|&n| n <= MAX_POINTS)
+        .ok_or_else(|| {
+            not_implemented(format!(
+                "an index-list write of more than {MAX_POINTS} elements (write it in slices)"
+            ))
+        })?;
+    let mut points = Vec::with_capacity(n);
+    let mut at = vec![0usize; per_axis.len()];
+    for _ in 0..n {
+        points.push(at.iter().zip(&per_axis).map(|(&i, v)| v[i]).collect());
+        for d in (0..at.len()).rev() {
+            at[d] += 1;
+            if at[d] < per_axis[d].len() {
+                break;
+            }
+            at[d] = 0;
+        }
+    }
+    Ok(Selection::Points(points))
+}
+
+/// The bytes to write for `value` under `plan`, in the dataset's dtype.
+pub(crate) fn dataset_bytes(
+    py: Python<'_>,
+    value: &Bound<'_, PyAny>,
+    dtype: &Bound<'_, PyAny>,
+    category: &str,
+    plan: &Plan,
+    chunks: Option<&[u64]>,
+) -> PyResult<Vec<u8>> {
+    let shape = PyTuple::new(py, plan.out_shape())?;
+    let fancy = plan.list_axis().is_some();
+    let chunk_elems = chunks.map_or(0, |c| c.iter().fold(1u64, |a, &d| a.saturating_mul(d)));
+    let bytes = helpers(py)?.call_method1(
+        "dataset_values",
+        (value, dtype, category, shape, fancy, chunk_elems),
+    )?;
+    Ok(bytes.cast::<PyBytes>()?.as_bytes().to_vec())
+}
+
+fn ieee_float(size: u32, byte_order: DatatypeByteOrder) -> Option<Datatype> {
+    let (exponent_location, exponent_size, mantissa_size, exponent_bias) = match size {
+        2 => (10, 5, 10, 15),
+        4 => (23, 8, 23, 127),
+        8 => (52, 11, 52, 1023),
+        _ => return None,
+    };
+    Some(Datatype::FloatingPoint {
+        size,
+        byte_order,
+        bit_offset: 0,
+        bit_precision: (size * 8) as u16,
+        exponent_location,
+        exponent_size,
+        mantissa_location: 0,
+        mantissa_size,
+        exponent_bias,
+    })
+}
+
+/// The HDF5 datatype h5py writes for a numpy dtype string (`'<i4'`,
+/// `'|b1'`, `'>f8'`, `'<c16'`, `'|S5'`).
+fn datatype_of(dtype: &str) -> Option<Datatype> {
+    let order = match dtype.as_bytes().first()? {
+        b'<' | b'|' | b'=' => DatatypeByteOrder::LittleEndian,
+        b'>' => DatatypeByteOrder::BigEndian,
+        _ => return None,
+    };
+    let kind = dtype.as_bytes().get(1)?;
+    let size: u32 = dtype.get(2..)?.parse().ok()?;
+    match kind {
+        b'b' if size == 1 => Some(Datatype::Enumeration {
+            size: 1,
+            base_type: Box::new(Datatype::FixedPoint {
+                size: 1,
+                byte_order: DatatypeByteOrder::LittleEndian,
+                signed: true,
+                bit_offset: 0,
+                bit_precision: 8,
+            }),
+            members: vec![
+                EnumMember {
+                    name: "FALSE".into(),
+                    value: vec![0],
+                },
+                EnumMember {
+                    name: "TRUE".into(),
+                    value: vec![1],
+                },
+            ],
+        }),
+        b'i' | b'u' if matches!(size, 1 | 2 | 4 | 8) => Some(Datatype::FixedPoint {
+            size,
+            byte_order: order,
+            signed: *kind == b'i',
+            bit_offset: 0,
+            bit_precision: (size * 8) as u16,
+        }),
+        b'f' => ieee_float(size, order),
+        b'c' => {
+            let part = ieee_float(size / 2, order)?;
+            Some(Datatype::Compound {
+                size,
+                members: vec![
+                    CompoundMember {
+                        name: "r".into(),
+                        byte_offset: 0,
+                        datatype: part.clone(),
+                    },
+                    CompoundMember {
+                        name: "i".into(),
+                        byte_offset: u64::from(size / 2),
+                        datatype: part,
+                    },
+                ],
+            })
+        }
+        b'S' if size > 0 => Some(Datatype::String {
+            size,
+            padding: StringPadding::NullPad,
+            charset: CharacterSet::Ascii,
+        }),
+        _ => None,
+    }
+}
+
+/// An attribute value as h5py would store it (`attrs[name] = value`, or
+/// `attrs.create(name, data, shape, dtype)`), except that `str` data is
+/// stored as fixed-length UTF-8 strings (h5py stores variable-length ones,
+/// which the editor cannot write).
+pub(crate) fn attr_value(
+    py: Python<'_>,
+    value: &Bound<'_, PyAny>,
+    dtype: Option<&Bound<'_, PyAny>>,
+    shape: Option<&Bound<'_, PyAny>>,
+) -> PyResult<AttrValue> {
+    if value.is_instance_of::<crate::PyEmpty>() {
+        return Err(not_implemented(
+            "writing an empty (null dataspace) attribute",
+        ));
+    }
+    let (kind, dt, dims, data): (String, String, Vec<u64>, Vec<u8>) = helpers(py)?
+        .call_method1("attr_value", (value, dtype, shape))?
+        .extract()?;
+    let datatype = if kind == "str" {
+        let size: u32 = dt
+            .parse()
+            .map_err(|_| PyTypeError::new_err("bad string size"))?;
+        Datatype::String {
+            size,
+            padding: StringPadding::NullPad,
+            charset: CharacterSet::Utf8,
+        }
+    } else {
+        datatype_of(&dt).ok_or_else(|| not_implemented(format!("an attribute of dtype {dt}")))?
+    };
+    Ok(AttrValue::Raw {
+        datatype,
+        shape: dims,
+        data,
+    })
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn numpy_dtypes_map_to_h5py_types() {
+        assert!(matches!(
+            datatype_of("<i4"),
+            Some(Datatype::FixedPoint {
+                size: 4,
+                signed: true,
+                byte_order: DatatypeByteOrder::LittleEndian,
+                ..
+            })
+        ));
+        assert!(matches!(
+            datatype_of(">u2"),
+            Some(Datatype::FixedPoint {
+                size: 2,
+                signed: false,
+                byte_order: DatatypeByteOrder::BigEndian,
+                ..
+            })
+        ));
+        assert!(matches!(
+            datatype_of("<f2"),
+            Some(Datatype::FloatingPoint { size: 2, .. })
+        ));
+        let c = datatype_of("<c16").unwrap();
+        assert!(matches!(&c, Datatype::Compound { size: 16, members } if is_complex(16, members)));
+        assert_eq!(category(&c).unwrap(), "complex");
+        let b = datatype_of("|b1").unwrap();
+        assert_eq!(category(&b).unwrap(), "bool");
+        assert!(matches!(
+            datatype_of("|S5"),
+            Some(Datatype::String { size: 5, .. })
+        ));
+        assert!(datatype_of("<f16").is_none());
+        assert!(datatype_of("<M8").is_none());
+        assert!(datatype_of("|S0").is_none());
+    }
+
+    #[test]
+    fn list_writes_become_points_in_value_order() {
+        let plan = Plan {
+            axes: vec![
+                Axis::List(vec![1, 4]),
+                Axis::Slice {
+                    start: 0,
+                    step: 2,
+                    count: 2,
+                },
+                Axis::Index(3),
+            ],
+            fields: vec![],
+            scalar: false,
+        };
+        let sel = selection(&plan, &[5, 4, 4]).unwrap();
+        assert_eq!(
+            sel,
+            Selection::Points(vec![
+                vec![1, 0, 3],
+                vec![1, 2, 3],
+                vec![4, 0, 3],
+                vec![4, 2, 3]
+            ])
+        );
+    }
+}
diff --git a/crates/clawhdf5-py/src/edit_helpers.py b/crates/clawhdf5-py/src/edit_helpers.py
new file mode 100644
index 0000000..0210d4e
--- /dev/null
+++ b/crates/clawhdf5-py/src/edit_helpers.py
@@ -0,0 +1,175 @@
+"""Values for in-place writes (clawhdf5.File(path, 'r+')), prepared the way
+h5py prepares them, so `ds[key] = value` stores what h5py would store.
+
+Loaded by the extension module (src/edit.rs); not a public API.
+
+h5py converts in two ways, and so does this module:
+
+- a value that is not a numpy array (a list, a Python or numpy scalar) is
+  converted by numpy straight to the dataset's dtype
+  (`numpy.asarray(value, dtype=ds.dtype)`), with numpy's rules and errors;
+- a numpy array is converted by libhdf5, whose numeric conversions clip to
+  the target's range instead of wrapping: integers saturate, floats are
+  truncated toward zero and clipped, a double too large for a float becomes
+  infinity. That is what `_convert_array` reproduces. Where libhdf5 has no
+  meaningful answer — NaN into an integer, for which it writes a different
+  arbitrary value per type — this raises ValueError instead of guessing.
+"""
+
+import numpy as np
+
+
+def _no_path(src, dst):
+    return TypeError(f"No conversion path for dtype: {src!r} -> {dst!r}")
+
+
+def _to_int(arr, dtype):
+    """Integer target: libhdf5's saturating conversion."""
+    info = np.iinfo(dtype)
+    kind = arr.dtype.kind
+    if kind == "b":
+        return arr.astype(dtype)
+    if kind in "iu":
+        src = np.iinfo(arr.dtype)
+        lo = max(info.min, src.min)
+        hi = min(info.max, src.max)
+        clipped = np.clip(arr, np.array(lo, arr.dtype), np.array(hi, arr.dtype))
+        return clipped.astype(dtype)
+    if kind == "f":
+        if np.isnan(arr).any():
+            raise ValueError(
+                "cannot write NaN to an integer dataset (libhdf5 would store an arbitrary value)"
+            )
+        t = np.trunc(arr.astype(np.float64))
+        # info.max + 1 and info.min are powers of two: exact as floats.
+        over = t >= float(info.max + 1)
+        under = t < float(info.min)
+        out = np.where(over | under, 0.0, t).astype(dtype)
+        out[over] = info.max
+        out[under] = info.min
+        return out
+    raise _no_path(arr.dtype, dtype)
+
+
+def _convert_array(arr, dtype, category):
+    kind = arr.dtype.kind
+    if category in ("int", "enum"):
+        if arr.dtype == dtype and kind in "iu":
+            return arr
+        if category == "enum" and kind not in "iu":
+            raise _no_path(arr.dtype, dtype)
+        return _to_int(arr, dtype)
+    if category == "bool":
+        if kind == "b":
+            return arr.astype(dtype)
+        if kind in "iu":
+            # h5py's bool is an enum over int8; libhdf5 converts integers
+            # into it by value (saturating), not to FALSE/TRUE, so 3 is
+            # stored as 3. Keep those bytes: a view, not a cast.
+            return _to_int(arr, np.dtype("i1")).view(dtype)
+        raise _no_path(arr.dtype, dtype)
+    if category == "float":
+        if kind not in "biuf":
+            raise _no_path(arr.dtype, dtype)
+        with np.errstate(over="ignore", invalid="ignore"):
+            return arr.astype(dtype)
+    if category == "complex":
+        if kind != "c":
+            raise _no_path(arr.dtype, dtype)
+        with np.errstate(over="ignore", invalid="ignore"):
+            return arr.astype(dtype)
+    if category == "string":
+        if kind != "S":
+            raise _no_path(arr.dtype, dtype)
+        return arr.astype(dtype)
+    # "exact": compound and opaque types, written only from the same dtype.
+    if arr.dtype == dtype:
+        return arr
+    raise _no_path(arr.dtype, dtype)
+
+
+def _convert_other(value, dtype, category):
+    if category == "string":
+        items = np.asarray(value, dtype=object)
+        if any(isinstance(x, str) for x in items.flat):
+            meta = dtype.metadata or {}
+            if meta.get("h5py_encoding") == "utf-8":
+                enc = [x.encode("utf-8") if isinstance(x, str) else x for x in items.flat]
+                return np.array(enc, dtype=dtype).reshape(items.shape)
+    return np.asarray(value, dtype=dtype)
+
+
+def _broadcast(arr, shape, fancy, chunk_elems):
+    """h5py's broadcasting: numpy's rules against the selection's shape
+    (extra leading length-1 axes allowed) for slices and integers. For an
+    index list, the exact shape; a scalar only where h5py expands it to the
+    whole selection (a chunked dataset whose chunk holds at least as many
+    elements as the selection)."""
+    if arr.shape == shape:
+        return arr
+    if fancy:
+        size = int(np.prod(shape))
+        if arr.ndim == 0 and ((chunk_elems > 0 and size <= chunk_elems) or len(shape) == 1):
+            return np.broadcast_to(arr, shape)
+        raise TypeError("Broadcasting is not supported for complex selections")
+    if arr.ndim == 0:
+        return np.broadcast_to(arr, shape)
+    err = TypeError(f"Can't broadcast {arr.shape} -> {shape}")
+    src = arr.shape
+    while len(src) > len(shape) and src[0] == 1:
+        src = src[1:]
+    if len(src) > len(shape):
+        raise err
+    try:
+        return np.broadcast_to(arr.reshape(src), shape)
+    except ValueError:
+        raise err from None
+
+
+def dataset_values(value, dtype, category, shape, fancy, chunk_elems):
+    """The bytes to write for `value` under a selection of `shape`, as a
+    C-ordered array of the dataset's dtype."""
+    if isinstance(value, np.ndarray):
+        arr = _convert_array(value, dtype, category)
+    else:
+        arr = _convert_other(value, dtype, category)
+    arr = _broadcast(arr, tuple(shape), fancy, chunk_elems)
+    return np.ascontiguousarray(arr, dtype=dtype).tobytes()
+
+
+def attr_value(value, dtype=None, shape=None):
+    """(kind, dtype string, shape, bytes) for an attribute value, h5py's
+    `attrs[name] = value` / `attrs.create(name, data, shape, dtype)`:
+
+    - "str": `str` data (h5py would store a variable-length string; this
+      stores a fixed-length UTF-8 string, which clawhdf5 can write);
+      the dtype string is the byte length of the longest element;
+    - "raw": a numeric, bool or bytes array, as numpy lays it out.
+    """
+    if dtype is not None:
+        arr = np.asarray(value, dtype=dtype, order="C")
+    else:
+        arr = np.asarray(value, order="C")
+    if shape is not None:
+        arr = arr.reshape(shape)
+    kind = arr.dtype.kind
+    if kind == "O":
+        if arr.size and all(isinstance(x, str) for x in arr.flat):
+            kind = "U"
+        elif arr.size and all(isinstance(x, bytes) for x in arr.flat):
+            arr = arr.astype(bytes)
+            kind = "S"
+        else:
+            raise TypeError(
+                f"clawhdf5 cannot write an attribute of Python objects ({value!r:.60})"
+            )
+    if kind == "U":
+        enc = [str(x).encode("utf-8") for x in arr.flat]
+        size = max([len(b) for b in enc] + [1])
+        data = np.array(enc, dtype=f"S{size}").reshape(arr.shape)
+        return ("str", str(size), arr.shape, data.tobytes())
+    if kind in "biufcS":
+        return ("raw", arr.dtype.str, arr.shape, np.ascontiguousarray(arr).tobytes())
+    raise NotImplementedError(
+        f"clawhdf5 cannot write an attribute of dtype {arr.dtype} in place"
+    )
diff --git a/crates/clawhdf5-py/src/file.rs b/crates/clawhdf5-py/src/file.rs
index 48d94c0..2764e61 100644
--- a/crates/clawhdf5-py/src/file.rs
+++ b/crates/clawhdf5-py/src/file.rs
@@ -5,7 +5,7 @@ use std::path::PathBuf;
 use std::sync::{Arc, Mutex};
 use std::time::Duration;
 
-use pyo3::exceptions::PyValueError;
+use pyo3::exceptions::{PyNotImplementedError, PyValueError};
 use pyo3::prelude::*;
 use pyo3::types::{PyDict, PyList};
 
@@ -95,6 +95,18 @@ impl PyFile {
         }
         match mode {
             "r" => Ok(Self::from_handle(Handle::open_local(py, path)?, filename)),
+            "r+" => Ok(Self::from_handle(
+                Handle::open_editable(py, path)?,
+                filename,
+            )),
+            "a" if std::path::Path::new(path).exists() => Ok(Self::from_handle(
+                Handle::open_editable(py, path)?,
+                filename,
+            )),
+            "a" => Err(PyNotImplementedError::new_err(format!(
+                "mode 'a' on {path}, which does not exist: clawhdf5 can only edit an existing \
+                 file in place; create a new one with mode 'w'"
+            ))),
             "w" => Ok(Self {
                 filename,
                 inner: Some(FileInner::Write(WriteState {
@@ -105,7 +117,7 @@ impl PyFile {
                 })),
             }),
             other => Err(PyValueError::new_err(format!(
-                "unsupported mode '{other}'; expected 'r' or 'w'"
+                "unsupported mode '{other}'; expected 'r', 'r+', 'a' or 'w'"
             ))),
         }
     }
@@ -226,7 +238,10 @@ impl PyFile {
             PyErr::new::<pyo3::exceptions::PyIOError, _>("file is already closed")
         })?;
         match inner {
-            FileInner::Read(_) => Ok(()),
+            FileInner::Read(root) => {
+                root.handle.close();
+                Ok(())
+            }
             FileInner::Write(state) => finalize_write(state),
         }
     }
@@ -289,6 +304,30 @@ impl PyFile {
         "/"
     }
 
+    /// `'r'` for a file opened read-only (a local file or a URL), `'r+'`
+    /// for one open for editing or writing, as h5py reports it.
+    #[getter]
+    fn mode(&self) -> PyResult<&'static str> {
+        match &self.inner {
+            Some(FileInner::Read(root)) if !root.handle.is_writable() => Ok("r"),
+            Some(_) => Ok("r+"),
+            None => Err(PyErr::new::<pyo3::exceptions::PyIOError, _>(
+                "file is closed",
+            )),
+        }
+    }
+
+    /// Nothing to do: every edit is written and synced when it is made, and
+    /// a file opened with 'w' is written on `close()`.
+    fn flush(&self) {}
+
+    /// Deleting objects is not supported (h5py's `del f[name]`).
+    fn __delitem__(&self, key: &str) -> PyResult<()> {
+        Err(PyNotImplementedError::new_err(format!(
+            "cannot delete '{key}': deleting objects is not supported by clawhdf5"
+        )))
+    }
+
     /// The path (or URL) the file was opened with.
     #[getter]
     fn filename(&self) -> &str {
@@ -389,6 +428,13 @@ impl PyFile {
     fn write_state_mut(&mut self) -> PyResult<&mut WriteState> {
         match &mut self.inner {
             Some(FileInner::Write(s)) => Ok(s),
+            Some(FileInner::Read(root)) if root.handle.is_writable() => {
+                Err(PyNotImplementedError::new_err(
+                    "creating datasets or groups in an existing file is not supported by \
+                     clawhdf5's in-place editor (mode 'r+' changes values, shapes and \
+                     attributes)",
+                ))
+            }
             Some(FileInner::Read(_)) => Err(PyErr::new::<pyo3::exceptions::PyIOError, _>(
                 "cannot write to a file opened for reading",
             )),
diff --git a/crates/clawhdf5-py/src/group.rs b/crates/clawhdf5-py/src/group.rs
index 36506b9..4020f16 100644
--- a/crates/clawhdf5-py/src/group.rs
+++ b/crates/clawhdf5-py/src/group.rs
@@ -3,7 +3,7 @@
 use std::collections::HashMap;
 use std::sync::{Arc, Mutex, OnceLock};
 
-use pyo3::exceptions::{PyIOError, PyKeyError, PyOSError, PyValueError};
+use pyo3::exceptions::{PyIOError, PyKeyError, PyNotImplementedError, PyOSError, PyValueError};
 use pyo3::prelude::*;
 use pyo3::types::PyList;
 
@@ -301,12 +301,23 @@ impl PyGroup {
                 state.lock().unwrap().datasets.push(spec);
                 Ok(())
             }
-            GroupInner::Read { .. } => Err(PyIOError::new_err(
+            GroupInner::Read(g) if g.handle.is_writable() => Err(PyNotImplementedError::new_err(
+                "creating datasets or groups in an existing file is not supported by \
+                 clawhdf5's in-place editor (mode 'r+' changes values, shapes and attributes)",
+            )),
+            GroupInner::Read(_) => Err(PyIOError::new_err(
                 "cannot create datasets on a read-only group",
             )),
         }
     }
 
+    /// Deleting objects is not supported (h5py's `del group[name]`).
+    fn __delitem__(&self, key: &str) -> PyResult<()> {
+        Err(PyNotImplementedError::new_err(format!(
+            "cannot delete '{key}': deleting objects is not supported by clawhdf5"
+        )))
+    }
+
     /// Attribute access.
     #[getter]
     fn attrs(&self, py: Python<'_>) -> PyResult<PyAttrs> {
diff --git a/crates/clawhdf5-py/src/handle.rs b/crates/clawhdf5-py/src/handle.rs
index 8649f74..754de6f 100644
--- a/crates/clawhdf5-py/src/handle.rs
+++ b/crates/clawhdf5-py/src/handle.rs
@@ -2,25 +2,34 @@
 //!
 //! Every read goes through [`Handle::with`], which releases the GIL and
 //! parses through `File::storage()`, so the same code serves a local file
-//! (memory-mapped) and a remote one (`clawhdf5-remote`: range requests
-//! through a block cache, so a network read never holds the GIL).
+//! (memory-mapped), a remote one (`clawhdf5-remote`: range requests through
+//! a block cache, so a network read never holds the GIL) and a file open
+//! for editing.
+//!
+//! A file opened with `'r+'` also holds a [`FileEditor`]. An edit takes the
+//! file's write lock, so no read runs while the file changes underneath it,
+//! and reopens the file afterwards: reads after an edit see the new bytes
+//! (a grown file, a new dataspace), never a stale mapping or chunk cache.
+//! Objects that cache something an edit can change compare
+//! [`Handle::generation`] with the value they cached it at.
 //!
 //! Lock discipline (no deadlock with the GIL): the file lock is only taken
 //! with the GIL released, and code that holds it never touches Python.
 
 use std::path::PathBuf;
-use std::sync::{Arc, PoisonError, RwLock};
+use std::sync::atomic::{AtomicU64, Ordering};
+use std::sync::{Arc, Mutex, PoisonError, RwLock};
 
-use clawhdf5_rs::File;
+use clawhdf5_rs::{File, FileEditor};
 use pyo3::exceptions::PyOSError;
 use pyo3::prelude::*;
 
-use crate::to_py_err;
+use crate::{panic_text, to_py_err};
 
 /// Where the file's bytes come from.
 pub(crate) enum Source {
     /// A local path (memory-mapped).
-    Local(#[allow(dead_code)] PathBuf),
+    Local(PathBuf),
     /// A URL, read through `clawhdf5-remote`'s block cache.
     Remote {
         url: String,
@@ -29,8 +38,14 @@ pub(crate) enum Source {
 }
 
 pub(crate) struct Handle {
+    /// The file as last opened; `None` if reopening it after an edit failed
+    /// (every read is then an error rather than a read of stale bytes).
     file: RwLock<Option<File>>,
+    /// For `'r+'`: the editor, until the file is closed.
+    editor: Option<Mutex<Option<FileEditor>>>,
     source: Source,
+    /// Bumped by every edit.
+    generation: AtomicU64,
     pub offset_size: u8,
     pub length_size: u8,
     pub root: u64,
@@ -41,13 +56,15 @@ fn closed_after_failed_reopen() -> PyErr {
 }
 
 impl Handle {
-    fn new(file: File, source: Source) -> Arc<Self> {
+    fn new(file: File, source: Source, editor: Option<FileEditor>) -> Arc<Self> {
         let sb = file.superblock();
         let (offset_size, length_size, root) =
             (sb.offset_size, sb.length_size, sb.root_group_address);
         Arc::new(Self {
             file: RwLock::new(Some(file)),
+            editor: editor.map(|e| Mutex::new(Some(e))),
             source,
+            generation: AtomicU64::new(0),
             offset_size,
             length_size,
             root,
@@ -57,7 +74,25 @@ impl Handle {
     /// A local file, read-only.
     pub(crate) fn open_local(py: Python<'_>, path: &str) -> PyResult<Arc<Self>> {
         let file = py.detach(|| crate::no_panic(|| File::open(path).map_err(to_py_err)))?;
-        Ok(Self::new(file, Source::Local(PathBuf::from(path))))
+        Ok(Self::new(file, Source::Local(PathBuf::from(path)), None))
+    }
+
+    /// A local file, open for in-place editing (`'r+'`): the editor takes
+    /// the file's exclusive lock and checks that it can edit the file, then
+    /// the file is opened for reading.
+    pub(crate) fn open_editable(py: Python<'_>, path: &str) -> PyResult<Arc<Self>> {
+        let (file, editor) = py.detach(|| {
+            crate::no_panic(|| {
+                let editor = FileEditor::open(path).map_err(to_py_err)?;
+                let file = File::open(path).map_err(to_py_err)?;
+                Ok((file, editor))
+            })
+        })?;
+        Ok(Self::new(
+            file,
+            Source::Local(PathBuf::from(path)),
+            Some(editor),
+        ))
     }
 
     /// A remote file (`http(s)://`, `s3://`, ...).
@@ -79,6 +114,7 @@ impl Handle {
                 url: url.to_string(),
                 storage,
             },
+            None,
         ))
     }
 
@@ -102,6 +138,17 @@ impl Handle {
         })
     }
 
+    /// Edits so far: objects that cache something an edit can change (a
+    /// dataset's shape, an object's attributes) re-read it when this moved.
+    pub(crate) fn generation(&self) -> u64 {
+        self.generation.load(Ordering::Acquire)
+    }
+
+    /// Whether the file was opened for editing (`'r+'`), even if closed since.
+    pub(crate) fn is_writable(&self) -> bool {
+        self.editor.is_some()
+    }
+
     /// The remote file's block cache.
     pub(crate) fn remote_storage(&self) -> Option<&clawhdf5_remote::RemoteStorage> {
         match &self.source {
@@ -117,6 +164,61 @@ impl Handle {
             Source::Local(_) => None,
         }
     }
+
+    /// Release the editor, and with it the file's lock. Objects still
+    /// open keep reading the file as it was last written; an edit through
+    /// them is an error.
+    pub(crate) fn close(&self) {
+        if let Some(ed) = &self.editor {
+            ed.lock().unwrap_or_else(PoisonError::into_inner).take();
+        }
+    }
+
+    /// Apply one edit with the GIL released. No read runs while it writes,
+    /// and the file is reopened afterwards — also after a failed edit, since
+    /// a commit that failed part-way may have changed the file.
+    pub(crate) fn edit<R: Send>(
+        &self,
+        py: Python<'_>,
+        f: impl FnOnce(&mut FileEditor) -> Result<R, clawhdf5_rs::Error> + Send,
+    ) -> PyResult<R> {
+        let Some(editor) = &self.editor else {
+            return Err(PyOSError::new_err(match self.source {
+                Source::Remote { .. } => "remote files are read-only",
+                Source::Local(_) => {
+                    "the file is open read-only; open it with mode 'r+' to change it"
+                }
+            }));
+        };
+        let Source::Local(path) = &self.source else {
+            return Err(PyOSError::new_err("remote files are read-only"));
+        };
+        py.detach(|| {
+            let mut ed = editor.lock().unwrap_or_else(PoisonError::into_inner);
+            let ed = ed
+                .as_mut()
+                .ok_or_else(|| PyOSError::new_err("the file is closed"))?;
+            let mut file = self.file.write().unwrap_or_else(PoisonError::into_inner);
+            let result = std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| f(ed)));
+            // Drop the old mapping and its chunk cache before reopening.
+            *file = None;
+            let reopened = std::panic::catch_unwind(|| File::open(path));
+            self.generation.fetch_add(1, Ordering::AcqRel);
+            match reopened {
+                Ok(Ok(f)) => *file = Some(f),
+                Ok(Err(e)) => return Err(to_py_err(e)),
+                Err(_) => return Err(closed_after_failed_reopen()),
+            }
+            drop(file);
+            match result {
+                Ok(r) => r.map_err(to_py_err),
+                Err(p) => Err(crate::InternalError::new_err(format!(
+                    "clawhdf5 internal error (please report it): {}",
+                    panic_text(&*p)
+                ))),
+            }
+        })
+    }
 }
 
 /// A `clawhdf5_remote::Error` as a Python exception: the network side
diff --git a/crates/clawhdf5-py/src/lib.rs b/crates/clawhdf5-py/src/lib.rs
index db89d8d..4ae81f7 100644
--- a/crates/clawhdf5-py/src/lib.rs
+++ b/crates/clawhdf5-py/src/lib.rs
@@ -7,11 +7,19 @@
 //!
 //! with clawhdf5.File('data.h5', 'r') as f:
 //!     data = f['dataset_name'][:]
+//!
+//! with clawhdf5.File('http://host/data.h5') as f:   # range requests
+//!     block = f['dataset_name'][10:20]
+//!
+//! with clawhdf5.File('data.h5', 'r+') as f:          # in-place edits
+//!     f['dataset_name'][0] = 1.5
+//!     f.attrs['note'] = 'edited'
 //! ```
 
 mod attrs;
 mod convert;
 mod dataset;
+mod edit;
 mod file;
 mod group;
 mod handle;
diff --git a/crates/clawhdf5-py/tests/test_edit.py b/crates/clawhdf5-py/tests/test_edit.py
new file mode 100644
index 0000000..99090e2
--- /dev/null
+++ b/crates/clawhdf5-py/tests/test_edit.py
@@ -0,0 +1,742 @@
+"""In-place editing: clawhdf5.File(path, 'r+') against h5py.
+
+Every edit is applied twice, to two copies of the same file: once through
+h5py (libhdf5) and once through clawhdf5 (FileEditor). After every edit both
+files are read back with h5py and must hold the same shapes, values and
+attributes; clawhdf5's own view must agree; when h5py refuses an edit,
+clawhdf5 must refuse it too and leave its file as it was. Files are written
+by h5py (libver earliest and latest, so every chunk index kind) and by
+clawhdf5; `h5dump` must read every result."""
+
+import io
+import os
+import shutil
+import subprocess
+import threading
+
+import numpy as np
+import pytest
+
+import clawhdf5
+
+
+# ---------------------------------------------------------------------------
+# Files
+# ---------------------------------------------------------------------------
+
+ENUM = {"RED": 0, "GREEN": 1, "BLUE": 7}
+
+
+def _h5py_file(h5py, path, libver):
+    rng = np.random.default_rng(1)
+    with h5py.File(path, "w", libver=libver) as f:
+        f.create_dataset("i4", data=np.arange(60, dtype="<i4").reshape(6, 10))
+        f.create_dataset("be_i2", data=np.arange(24, dtype=">i2").reshape(4, 6))
+        f.create_dataset("u1", data=np.arange(16, dtype="u1"))
+        f.create_dataset("f2", data=rng.standard_normal(12).astype("<f2"))
+        f.create_dataset("f4_2d", data=rng.standard_normal((8, 9)).astype("<f4"))
+        f.create_dataset("f8_3d", data=rng.standard_normal((4, 5, 6)))
+        f.create_dataset("u8", data=np.arange(10, dtype="<u8"))
+        f.create_dataset("c8", data=(np.arange(6) + 1j * np.arange(6)).astype("<c8"))
+        f.create_dataset("bool", data=np.array([True, False, True, True]))
+        f.create_dataset("enum", data=np.array([0, 1, 7, 0], dtype="i1"),
+                         dtype=h5py.enum_dtype(ENUM, basetype="i1"))
+        f.create_dataset("s5", data=np.array([b"ab", b"cdefg", b""], dtype="S5"))
+        cmp_dt = np.dtype([("id", "<i4"), ("x", "<f8"), ("tag", "S3")])
+        f.create_dataset("cmp", data=np.array([(i, i / 2, b"t%d" % i) for i in range(5)], dtype=cmp_dt))
+        f.create_dataset("scalar", data=np.float64(3.5))
+        # Chunked: fixed maxshape (v4 fixed array under latest), one
+        # unlimited dimension (extensible array), two (v2 B-tree), one chunk.
+        f.create_dataset("chunk_fixed", data=np.arange(100, dtype="<i8").reshape(10, 10), chunks=(3, 4))
+        f.create_dataset("chunk_ext", data=rng.standard_normal((12, 7)), chunks=(5, 7), maxshape=(None, 7))
+        f.create_dataset("chunk_bt2", data=np.arange(30, dtype="<i4").reshape(5, 6), chunks=(2, 2),
+                         maxshape=(None, None))
+        f.create_dataset("chunk_gzip", data=np.arange(400, dtype="<f4").reshape(20, 20), chunks=(6, 6),
+                         compression="gzip", maxshape=(40, 40))
+        f.create_dataset("chunk_single", data=np.arange(12, dtype="<u2").reshape(3, 4), chunks=(3, 4),
+                         maxshape=(3, 4))
+        f.create_dataset("chunk_fill", shape=(8,), dtype="<i4", chunks=(3,), maxshape=(20,), fillvalue=-1)
+        f.create_dataset("vlen", data=["a", "bb"], dtype=h5py.string_dtype())
+        # Compact layout (low-level API).
+        dcpl = h5py.h5p.create(h5py.h5p.DATASET_CREATE)
+        dcpl.set_layout(h5py.h5d.COMPACT)
+        space = h5py.h5s.create_simple((7,))
+        dsid = h5py.h5d.create(f.id, b"compact", h5py.h5t.STD_I32LE, space, dcpl=dcpl)
+        dsid.write(h5py.h5s.ALL, h5py.h5s.ALL, np.arange(7, dtype="<i4"))
+        g = f.create_group("grp")
+        g.create_dataset("leaf", data=np.arange(5.0))
+        g.attrs["units"] = "m"
+        f.attrs["version"] = np.int32(1)
+
+
+def _clawhdf5_file(path):
+    with clawhdf5.File(str(path), "w") as f:
+        f.create_dataset("i4", data=np.arange(60, dtype="<i4").reshape(6, 10))
+        f.create_dataset("f8", data=np.linspace(0, 1, 30).reshape(5, 6))
+        f.create_dataset("chunk_gzip", data=np.arange(400, dtype="<f4").reshape(20, 20),
+                         chunks=[6, 6], compression="gzip")
+        f.create_dataset("u1", data=np.arange(16, dtype="u1"))
+        g = f.create_group("grp")
+        g.create_dataset("leaf", data=np.arange(5.0))
+        g.attrs["units"] = "m"
+        f.attrs["version"] = 1
+
+
+# h5py's libver: "earliest" (v1 B-tree chunk indexes), "v114" (the 1.10+
+# indexes: fixed and extensible arrays, v2 B-trees, single chunk) and
+# "latest" (HDF5 2.0's newest format, which h5dump 1.14 cannot read).
+SOURCES = ["h5py-earliest", "h5py-v114", "h5py-latest", "clawhdf5"]
+
+
+def _make(h5py, tmp_path, source):
+    base = tmp_path / f"base-{source}.h5"
+    if source == "clawhdf5":
+        _clawhdf5_file(base)
+    else:
+        _h5py_file(h5py, str(base), source.split("-")[1])
+    theirs = tmp_path / f"theirs-{source}.h5"
+    ours = tmp_path / f"ours-{source}.h5"
+    shutil.copy(base, theirs)
+    shutil.copy(base, ours)
+    return str(theirs), str(ours), str(base)
+
+
+# ---------------------------------------------------------------------------
+# Comparing files through h5py
+# ---------------------------------------------------------------------------
+
+
+def _norm_attr(v):
+    """Attribute values comparable across the two writers: clawhdf5 stores
+    `str` as fixed-length UTF-8 (h5py reads bytes), h5py as variable-length
+    (h5py reads str)."""
+    if isinstance(v, bytes):
+        return ("str", v.decode("utf-8"))
+    if isinstance(v, str):
+        return ("str", v)
+    arr = np.asarray(v)
+    if arr.dtype.kind in "SO":
+        return ("strs", [x.decode() if isinstance(x, bytes) else x for x in arr.ravel().tolist()], arr.shape)
+    return (arr.dtype.str, arr.shape, arr.tobytes())
+
+
+def snapshot(h5py, path):
+    """What h5py sees in the file: every dataset's shape, dtype, bytes and
+    attributes (read without locking: clawhdf5 may hold the file open)."""
+    out = {}
+    with h5py.File(path, "r", locking=False) as f:
+        def visit(name, obj):
+            attrs = {k: _norm_attr(obj.attrs[k]) for k in obj.attrs}
+            if isinstance(obj, h5py.Dataset):
+                if obj.dtype.kind == "O":
+                    data = [x for x in obj[...].ravel().tolist()]
+                else:
+                    data = obj[()].tobytes() if obj.shape is not None else None
+                out[name] = (obj.shape, obj.dtype.str, obj.maxshape, data, attrs)
+            else:
+                out[name] = ("group", attrs)
+        visit("/", f)
+        f.visititems(visit)
+    return out
+
+
+def assert_same_files(h5py, theirs, ours, what):
+    a, b = snapshot(h5py, theirs), snapshot(h5py, ours)
+    assert a.keys() == b.keys(), what
+    for k in a:
+        assert a[k] == b[k], f"{what}: {k} differs\n h5py:     {a[k]}\n clawhdf5: {b[k]}"
+
+
+def assert_ours_reads_like_h5py(h5py, f, path, what):
+    """clawhdf5's own view of the file it is editing matches h5py's."""
+    with h5py.File(path, "r", locking=False) as t:
+        for name in ["i4", "chunk_ext", "chunk_bt2", "chunk_gzip", "f8_3d", "cmp", "bool", "enum", "scalar"]:
+            if name not in t:
+                continue
+            o = f[name]
+            assert o.shape == t[name].shape, f"{what}: {name} shape"
+            assert o.maxshape == t[name].maxshape, f"{what}: {name} maxshape"
+            np.testing.assert_array_equal(o[()], t[name][()], err_msg=f"{what}: {name}")
+        for obj in ["/", "grp"]:
+            assert sorted(f[obj].attrs.keys()) == sorted(t[obj].attrs.keys()), what
+            for k in t[obj].attrs:
+                assert _norm_attr(f[obj].attrs[k]) == _norm_attr(t[obj].attrs[k]), f"{what}: {obj}.attrs[{k}]"
+
+
+def h5dump_reads(path, base=None):
+    """h5dump (libhdf5 1.14) reads every object and value of `path` — when it
+    reads the unedited `base` (it cannot read HDF5 2.0's newest format)."""
+    exe = shutil.which("h5dump")
+    if exe is None:
+        if os.environ.get("CLAWHDF5_REQUIRE_INTEROP") == "1":
+            pytest.fail("h5dump is required (CLAWHDF5_REQUIRE_INTEROP=1)")
+        return
+    h5rs = os.environ.get("CLAWHDF5_H5RS")
+    if h5rs:
+        # clawhdf5's structural and checksum validator (scripts/ci-test.sh
+        # points this at the h5rs it built).
+        r = subprocess.run([h5rs, "check", path], capture_output=True, text=True)
+        assert r.returncode == 0, (r.stdout + r.stderr)[-2000:]
+    if base is not None and subprocess.run([exe, "-H", base], capture_output=True).returncode != 0:
+        return
+    r = subprocess.run([exe, path], capture_output=True, text=True)
+    assert r.returncode == 0, r.stderr[-2000:]
+
+
+# ---------------------------------------------------------------------------
+# Applying one edit both ways
+# ---------------------------------------------------------------------------
+
+
+def _native_conversion(h5py, value, ds_dtype):
+    """`value` as libhdf5 converts it to `ds_dtype` in native byte order.
+
+    libhdf5 2.0 (h5py 3.16) converts numbers differently when either side is
+    not in native byte order (its "soft" conversions): a float in (-1, 0)
+    becomes the integer type's minimum instead of 0, and an unsigned integer
+    too large for the signed type of the same size wraps instead of
+    saturating. clawhdf5 applies the native-order results to every byte
+    order, so the reference is h5py converting into a native dataset of the
+    same kind; the result then reaches the real dataset by a plain byte
+    swap."""
+    if not isinstance(value, np.ndarray):
+        return value
+    if value.dtype.kind not in "biuf" or ds_dtype.kind not in "biuf":
+        return value
+    if value.dtype.isnative and ds_dtype.isnative:
+        return value
+    with h5py.File(io.BytesIO(), "w") as tmp:
+        d = tmp.create_dataset("t", shape=value.shape, dtype=ds_dtype.newbyteorder("="))
+        d[...] = value.astype(value.dtype.newbyteorder("="))
+        return np.asarray(d[()])
+
+
+def _apply(f, op, h5py=None):
+    """Apply `op` to `f`; with `h5py`, `f` is an h5py file and a numpy array
+    value is first converted as libhdf5 converts in native byte order (see
+    `_native_conversion`)."""
+    kind = op[0]
+    if kind == "set":
+        _, name, key, value = op
+        if h5py is not None:
+            value = _native_conversion(h5py, value, f[name].dtype)
+        f[name][key] = value
+    elif kind == "resize":
+        _, name, size, axis = op
+        if axis is None:
+            f[name].resize(size)
+        else:
+            f[name].resize(size, axis=axis)
+    elif kind == "attr":
+        _, obj, name, value = op
+        f[obj].attrs[name] = value
+    else:
+        raise AssertionError(op)
+
+
+def edit_both(h5py, theirs, ours_path, ours, op):
+    """Apply `op` with h5py and with clawhdf5 (`ours`, open 'r+'); the two
+    files must then read the same through h5py. If h5py refuses, clawhdf5
+    must refuse and its file must be unchanged. Returns h5py's error."""
+    before = snapshot(h5py, ours_path)
+    try:
+        with h5py.File(theirs, "r+") as t:
+            _apply(t, op, h5py)
+    except Exception as e:  # noqa: BLE001 - h5py refuses: so must we
+        try:
+            _apply(ours, op)
+        except Exception:  # noqa: BLE001
+            pass
+        else:
+            pytest.fail(f"{op!r:.300}: h5py refused ({type(e).__name__}: {e}), clawhdf5 did not")
+        assert snapshot(h5py, ours_path) == before, f"{op!r}: clawhdf5 changed the file while failing"
+        return e
+    _apply(ours, op)
+    assert_same_files(h5py, theirs, ours_path, repr(op)[:200])
+    return None
+
+
+# ---------------------------------------------------------------------------
+# Tests
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.parametrize("source", SOURCES)
+def test_edit_sequence_matches_h5py(h5py, tmp_path, source):
+    theirs, ours_path, base = _make(h5py, tmp_path, source)
+    ops = [
+        ("set", "i4", 0, 99),
+        ("set", "i4", (slice(1, 5, 2), slice(None, None, 3)), np.array([[1.5, -2.5, 1e12, -1e12]])),
+        ("set", "i4", (slice(None), 2), np.arange(6, dtype="<i8") * 1000),
+        ("set", "i4", ([0, 2, 5], slice(4, 6)), np.array([[1, 2], [3, 4], [5, 6]])),
+        ("set", "i4", (Ellipsis, -1), np.int8(-7)),
+        ("set", "i4", (3, 3), 12345.9),
+        ("set", "u1", slice(2, 8), np.array([-5, 0, 300, 255, 256, 1], dtype="<i4")),
+        ("set", "u1", slice(0, 3), [1, 2, 3]),
+        ("set", "chunk_gzip", (slice(0, 20, 7), slice(3, 17)), 42.25),
+        ("set", "chunk_gzip", (slice(5, 11), slice(5, 11)), np.ones((6, 6), dtype="<f8") * np.pi),
+        ("set", "grp/leaf", slice(None), np.array([1, 2, 3, 4, 5], dtype="<u2")),
+        ("attr", "/", "version", np.int32(2)),
+        ("attr", "/", "count", 7),
+        ("attr", "grp", "scale", np.array([0.5, 0.25], dtype="<f4")),
+        ("attr", "grp", "matrix", np.arange(6, dtype=">i8").reshape(2, 3)),
+        ("attr", "i4", "flag", np.bool_(True)),
+        ("attr", "i4", "z", np.complex64(1 - 2j)),
+        ("attr", "i4", "raw", np.bytes_(b"abc")),
+        ("set", "i4", slice(0, 2), np.zeros((3, 10))),  # shape mismatch: refused by both
+    ]
+    if source != "clawhdf5":
+        ops += [
+            ("set", "be_i2", (slice(None), slice(1, 3)), np.array([70000, -70000], dtype="<i8")),
+            ("set", "f2", slice(None, None, 4), np.array([1e6, -3.25, 0.1])),
+            ("set", "f4_2d", (2, slice(None)), np.linspace(-1, 1, 9)),
+            ("set", "f8_3d", (slice(1, 3), 2, slice(None, None, 2)), np.arange(3, dtype="<i2")),
+            ("set", "u8", slice(None), np.array([-1, 0, 2**63, 1e30, -1e30, 5.5, 2, 3, 4, 5])),
+            ("set", "c8", slice(1, 3), np.array([1 + 1j, 2 - 2j], dtype="<c16")),
+            ("set", "c8", 0, np.float64(1.0)),  # h5py: no conversion path
+            ("set", "bool", slice(None), np.array([0, 3, 0, -1], dtype="<i4")),
+            ("set", "bool", 1, np.array(True)),
+            ("set", "enum", slice(0, 2), np.array([7, 1], dtype="<i4")),
+            ("set", "s5", 0, np.bytes_(b"xyzuvw")),
+            ("set", "s5", slice(1, 3), [b"q", b"rs"]),
+            ("set", "s5", 2, np.array("uni")),  # h5py: no conversion from 'U'
+            ("set", "cmp", 2, np.array((9, 9.5, b"zz"), dtype=[("id", "<i4"), ("x", "<f8"), ("tag", "S3")])),
+            ("set", "cmp", slice(3, 5), [(1, 0.5, b"a"), (2, 1.5, b"b")]),
+            ("set", "scalar", (), 7.25),
+            ("set", "scalar", Ellipsis, np.float32(-1.5)),
+            ("set", "compact", slice(1, 6, 2), np.array([10, 20, 30])),
+            ("set", "chunk_fixed", (slice(2, 9), slice(1, 10, 4)), np.arange(21).reshape(7, 3)),
+            ("set", "chunk_single", (1, slice(None)), np.array([9, 8, 7, 6])),
+            ("resize", "chunk_ext", (20, 7), None),
+            ("set", "chunk_ext", slice(12, 20), np.full((8, 7), 2.5)),
+            ("resize", "chunk_ext", 9, 0),
+            ("resize", "chunk_ext", 16, 0),
+            ("resize", "chunk_bt2", (9, 11), None),
+            ("set", "chunk_bt2", (slice(4, 9), slice(5, 11)), np.arange(30).reshape(5, 6)),
+            ("resize", "chunk_bt2", (3, 3), None),
+            ("resize", "chunk_bt2", (7, 8), None),
+            ("resize", "chunk_gzip", (40, 25), None),
+            ("resize", "chunk_gzip", (41, 25), None),  # beyond maxshape: refused
+            ("resize", "chunk_fill", 15, None),  # h5py: a size without axis must be a tuple
+            ("resize", "chunk_fill", (15,), None),
+            ("set", "chunk_fill", slice(10, 12), [1, 2]),
+            ("resize", "chunk_fill", (4,), None),
+            ("resize", "chunk_fill", (20,), None),
+            ("resize", "i4", (7, 10), None),  # not chunked: refused
+        ]
+    with clawhdf5.File(ours_path, "r+") as ours:
+        assert ours.mode == "r+"
+        for i, op in enumerate(ops):
+            edit_both(h5py, theirs, ours_path, ours, op)
+            if i % 5 == 0:
+                assert_ours_reads_like_h5py(h5py, ours, ours_path, repr(op))
+        assert_ours_reads_like_h5py(h5py, ours, ours_path, "end")
+    h5dump_reads(ours_path, base)
+    # Reopened, clawhdf5 reads what h5py reads.
+    with clawhdf5.File(ours_path, "r") as f:
+        assert f.mode == "r"
+        assert_ours_reads_like_h5py(h5py, f, ours_path, "reopened")
+
+
+def _random_key(rng, shape):
+    key = []
+    for n in shape:
+        r = rng.random()
+        if n == 0 or r < 0.3:
+            start = int(rng.integers(0, n + 1)) if n else 0
+            stop = int(rng.integers(start, n + 1)) if n else 0
+            step = int(rng.integers(1, 4))
+            key.append(slice(start, stop, step))
+        elif r < 0.55:
+            key.append(int(rng.integers(-n, n)))
+        elif r < 0.7:
+            k = int(rng.integers(1, min(n, 4) + 1))
+            key.append(sorted(rng.choice(n, size=k, replace=False).tolist()))
+        else:
+            key.append(slice(None))
+    # Only one index list per key.
+    lists = [i for i, k in enumerate(key) if isinstance(k, list)]
+    for i in lists[1:]:
+        key[i] = slice(None)
+    return tuple(key)
+
+
+def _selection_shape(key, shape):
+    out = []
+    fancy = False
+    for k, n in zip(key, shape):
+        if isinstance(k, slice):
+            out.append(len(range(*k.indices(n))))
+        elif isinstance(k, list):
+            out.append(len(k))
+            fancy = True
+    return tuple(out), fancy
+
+
+def _random_value(rng, sel_shape, fancy, dtype):
+    r = rng.random()
+    if r < 0.2 or not sel_shape:
+        v = rng.standard_normal() * 1000
+        return np.float64(v) if rng.random() < 0.5 else int(v)
+    shape = list(sel_shape)
+    if not fancy and r < 0.35 and shape:
+        shape[0] = 1  # broadcast along the first axis
+    kind = rng.choice(["same", "f8", "i8", "u1", "f4"])
+    if kind == "same" and np.dtype(dtype).kind in "iuf":
+        dt = np.dtype(dtype)
+    else:
+        dt = np.dtype(str(kind) if kind != "same" else "f8")
+    base = rng.standard_normal(size=shape) * (10 ** rng.integers(0, 6))
+    with np.errstate(all="ignore"):
+        return base.astype(dt)
+
+
+@pytest.mark.parametrize("source", SOURCES)
+@pytest.mark.parametrize("seed", [0, 1, 2, 3])
+def test_random_edits_match_h5py(h5py, tmp_path, source, seed):
+    """Random writes (slices, steps, integers, index lists, broadcasts,
+    other dtypes and out-of-range values), resizes and attributes, each
+    compared with h5py applying the same edit."""
+    rng = np.random.default_rng(1000 * seed + SOURCES.index(source))
+    theirs, ours_path, base = _make(h5py, tmp_path, source)
+    with h5py.File(theirs, "r") as t:
+        names = [n for n in ["i4", "u1", "f8", "f4_2d", "f8_3d", "be_i2", "chunk_fixed", "chunk_ext",
+                             "chunk_bt2", "chunk_gzip", "chunk_fill", "compact", "grp/leaf"] if n in t]
+    resizable = [n for n in ["chunk_ext", "chunk_bt2", "chunk_gzip", "chunk_fill"] if n in names]
+    refused = 0
+    with clawhdf5.File(ours_path, "r+") as ours:
+        for step in range(40):
+            r = rng.random()
+            if r < 0.15 and resizable:
+                name = str(rng.choice(resizable))
+                with h5py.File(theirs, "r") as t:
+                    maxshape = t[name].maxshape
+                    shape = t[name].shape
+                new = tuple(int(rng.integers(0, (m if m is not None else s + 10) + 1)) for m, s in zip(maxshape, shape))
+                op = ("resize", name, new, None)
+            elif r < 0.25:
+                obj = str(rng.choice(["/", "grp", names[0]]))
+                choices = [np.int16(rng.integers(-100, 100)), rng.standard_normal(3),
+                           np.arange(int(rng.integers(1, 5)), dtype=">u4"), np.float32(0.5)]
+                value = choices[int(rng.integers(0, len(choices)))]
+                op = ("attr", obj, f"a{int(rng.integers(0, 4))}", value)
+            else:
+                name = str(rng.choice(names))
+                with h5py.File(theirs, "r") as t:
+                    shape, dtype = t[name].shape, t[name].dtype
+                key = _random_key(rng, shape)
+                sel_shape, fancy = _selection_shape(key, shape)
+                op = ("set", name, key, _random_value(rng, sel_shape, fancy, dtype))
+            if edit_both(h5py, theirs, ours_path, ours, op) is not None:
+                refused += 1
+        assert_ours_reads_like_h5py(h5py, ours, ours_path, "end")
+    assert refused < 30
+    h5dump_reads(ours_path, base)
+
+
+NUMERIC = ["<i1", "<u1", "<i2", ">u2", "<i4", "<u4", ">i8", "<u8", "<f2", "<f4", ">f8"]
+
+
+def _quiet(f):
+    with np.errstate(all="ignore"):
+        return f()
+
+
+def _libhdf5_undefined(vals, target):
+    """The values whose conversion to `target` libhdf5 2.0 (h5py 3.16) gets
+    wrong even in native byte order, where its C casts are undefined
+    behaviour; clawhdf5 saturates them as libhdf5's range handling
+    intends (docs/known-issues.md):
+
+    - half floats into unsigned integers: negatives wrap (-1 -> 65535) and
+      +inf becomes 0; into signed integers, +-inf becomes the minimum;
+    - a float equal to the integer maximum rounded up in the float's
+      precision (float32(2**31 - 1) == 2**31 -> int32, float64(2**64 - 1)
+      -> uint64) becomes the minimum (or 0);
+    - a double between 65504 and 65520 into a half float becomes infinity
+      (IEEE rounds it down to 65504, as numpy does).
+    """
+    bad = np.zeros(vals.shape, dtype=bool)
+    if vals.dtype.kind != "f":
+        return bad
+    if target.kind in "iu":
+        if vals.dtype.itemsize == 2:
+            bad |= np.isinf(vals)
+            if target.kind == "u":
+                bad |= vals <= -1
+        top = _quiet(lambda: np.array(np.iinfo(target).max).astype(vals.dtype))
+        if float(top) > np.iinfo(target).max:
+            bad |= vals == top
+    if target.kind == "f" and target.itemsize == 2 and vals.dtype.itemsize > 2:
+        bad |= (np.abs(vals) > 65504) & (np.abs(vals) < 65520)
+    return bad
+
+
+def test_numeric_conversions_match_h5py(h5py, tmp_path):
+    """Every numeric source dtype into every numeric dataset dtype, with
+    values at and beyond the targets' limits, as libhdf5 converts them."""
+    edge = np.array([0, 1, -1, -0.3, 2.5, -2.5, 3.7, -3.7, 127.9, -128.9, 200.5, 255.5, 256, -129,
+                     32767.5, 40000, 65504, 70000, -70000, 2**31 - 1, 2**31, -2**31 - 1,
+                     4e9, 1e15, -1e15, 1e19, 1e300, -1e300, np.inf, -np.inf])
+    sources = {
+        "f8": edge,
+        "f4": _quiet(lambda: edge.astype("<f4")),
+        "f2": np.array([0, 1, -1, 2.5, -3.5, 65504, -65504, np.inf, -np.inf, 100.5], dtype="<f2"),
+        "i8": np.array([0, 1, -1, 127, 128, -129, 255, 256, 32768, -32769, 65536, 2**31, -2**31 - 1,
+                        2**32, 2**62, -2**63, 2**63 - 1], dtype="<i8"),
+        "u8": np.array([0, 1, 127, 128, 255, 256, 65535, 65536, 2**31, 2**32, 2**63, 2**64 - 1], dtype="<u8"),
+        "i1": np.array([-128, -1, 0, 1, 127], dtype="i1"),
+        "u2": np.array([0, 255, 256, 65535], dtype=">u2"),
+        "b": np.array([True, False, True]),
+    }
+    for target in NUMERIC:
+        for sname, src in sources.items():
+            vals = src[~_libhdf5_undefined(src, np.dtype(target))]
+            path_t = str(tmp_path / f"t_{target[1:]}_{sname}.h5")
+            path_o = str(tmp_path / f"o_{target[1:]}_{sname}.h5")
+            with h5py.File(path_t, "w") as f:
+                f.create_dataset("d", shape=vals.shape, dtype=target)
+            shutil.copy(path_t, path_o)
+            what = f"{vals.dtype} -> {target}"
+            try:
+                with h5py.File(path_t, "r+") as f:
+                    f["d"][...] = _native_conversion(h5py, vals, f["d"].dtype)
+            except Exception:  # noqa: BLE001
+                with clawhdf5.File(path_o, "r+") as f, pytest.raises(Exception):
+                    f["d"][...] = vals
+                continue
+            with clawhdf5.File(path_o, "r+") as f:
+                f["d"][...] = vals
+            with h5py.File(path_t, "r") as a, h5py.File(path_o, "r") as b:
+                assert a["d"][...].tobytes() == b["d"][...].tobytes(), (
+                    f"{what}: h5py {a['d'][...].tolist()} clawhdf5 {b['d'][...].tolist()}"
+                )
+
+
+def test_nan_into_an_integer_dataset_is_refused(h5py, tmp_path):
+    """libhdf5 stores NaN as an arbitrary integer (0, the minimum or 2**63,
+    depending on the type); clawhdf5 refuses and writes nothing."""
+    path = str(tmp_path / "nan.h5")
+    with h5py.File(path, "w") as f:
+        f.create_dataset("d", data=np.arange(4, dtype="<i4"))
+    with clawhdf5.File(path, "r+") as f:
+        with pytest.raises(ValueError, match="NaN"):
+            f["d"][...] = np.array([1.0, np.nan, 2.0, 3.0])
+        # A Python list goes through numpy, which refuses NaN too.
+        with pytest.raises(ValueError):
+            f["d"][0:2] = [np.nan, 1.0]
+    with h5py.File(path, "r") as f:
+        np.testing.assert_array_equal(f["d"][...], np.arange(4))
+
+
+def test_unsupported_edits_are_clear_errors(h5py, tmp_path):
+    path = str(tmp_path / "u.h5")
+    _h5py_file(h5py, path, "earliest")
+    before = snapshot(h5py, path)
+    with clawhdf5.File(path, "r+") as f:
+        with pytest.raises(NotImplementedError, match="delet"):
+            del f.attrs["version"]
+        with pytest.raises(NotImplementedError, match="delet"):
+            del f["i4"]
+        with pytest.raises(NotImplementedError, match="delet"):
+            del f["grp"]["leaf"]
+        with pytest.raises(NotImplementedError):
+            f.create_dataset("new", data=np.arange(3.0))
+        with pytest.raises(NotImplementedError):
+            f.create_group("newgrp")
+        with pytest.raises(NotImplementedError):
+            f["grp"].create_dataset("new", data=np.arange(3.0))
+        with pytest.raises(NotImplementedError, match="variable-length"):
+            f["vlen"][0] = "x"
+        with pytest.raises(NotImplementedError, match="field"):
+            f["cmp"]["id"] = np.arange(5)
+        with pytest.raises(NotImplementedError):
+            f.attrs["empty"] = clawhdf5.Empty("f8")
+        with pytest.raises(TypeError, match="chunked"):
+            f["i4"].resize((7, 10))
+        with pytest.raises(ValueError):
+            f["chunk_gzip"].resize((41, 20))
+        with pytest.raises(ValueError, match="axis"):
+            f["chunk_ext"].resize(3, axis=2)
+        with pytest.raises(TypeError):
+            f["i4"][0] = np.array(["a"] * 10)
+    assert snapshot(h5py, path) == before
+    h5dump_reads(path)
+
+
+def test_read_only_files_and_modes(h5py, tmp_path):
+    path = str(tmp_path / "m.h5")
+    with h5py.File(path, "w") as f:
+        f.create_dataset("d", data=np.arange(4, dtype="<i4"), chunks=(2,), maxshape=(None,))
+    with clawhdf5.File(path, "r") as f:
+        with pytest.raises(OSError, match="r\\+"):
+            f["d"][0] = 1
+        with pytest.raises(OSError):
+            f["d"].resize((8,))
+        with pytest.raises(OSError):
+            f.attrs["x"] = 1
+    with pytest.raises(NotImplementedError, match="does not exist"):
+        clawhdf5.File(str(tmp_path / "missing.h5"), "a")
+    with pytest.raises(ValueError, match="mode"):
+        clawhdf5.File(path, "rw")
+    with clawhdf5.File(path, "a") as f:
+        assert f.mode == "r+"
+        f["d"][1] = 10
+    with h5py.File(path, "r") as f:
+        assert f["d"][1] == 10
+
+
+def test_the_file_is_locked_while_open_for_editing(h5py, tmp_path):
+    path = str(tmp_path / "lock.h5")
+    with h5py.File(path, "w") as f:
+        f.create_dataset("d", data=np.arange(4, dtype="<i4"))
+    f = clawhdf5.File(path, "r+")
+    with pytest.raises(OSError):
+        clawhdf5.File(path, "r+")
+    with pytest.raises(OSError):
+        h5py.File(path, "r+")
+    f.close()
+    with h5py.File(path, "r+") as g:
+        g["d"][0] = 5
+    with clawhdf5.File(path, "r+") as g:
+        g["d"][1] = 6
+    with h5py.File(path, "r") as g:
+        np.testing.assert_array_equal(g["d"][...], [5, 6, 2, 3])
+
+
+def test_objects_see_edits_made_through_others(h5py, tmp_path):
+    """A dataset or attrs object taken before an edit reports the file as
+    it is after it: the new shape, the new attribute."""
+    path = str(tmp_path / "live.h5")
+    with h5py.File(path, "w") as f:
+        f.create_dataset("d", data=np.arange(6.0), chunks=(4,), maxshape=(None,))
+    with clawhdf5.File(path, "r+") as f:
+        d1 = f["d"]
+        d2 = f["d"]
+        attrs = d1.attrs
+        assert len(attrs) == 0 and "u" not in attrs
+        d2.resize((10,))
+        assert d1.shape == (10,) and d1.size == 10 and len(d1) == 10
+        np.testing.assert_array_equal(d1[6:], np.zeros(4))
+        d2.attrs["u"] = "m/s"
+        assert "u" in attrs and attrs["u"] == b"m/s" and len(attrs) == 1
+        f.attrs.create("shaped", np.arange(6), shape=(2, 3), dtype="<i2")
+        f.attrs.modify("shaped2", [1.5, 2.5])
+    with h5py.File(path, "r") as f:
+        assert f["d"].shape == (10,)
+        assert f["d"].attrs["u"] == b"m/s"
+        assert f.attrs["shaped"].dtype == np.dtype("<i2") and f.attrs["shaped"].shape == (2, 3)
+        np.testing.assert_array_equal(f.attrs["shaped2"], [1.5, 2.5])
+
+
+def test_attribute_types_as_h5py_reads_them(h5py, tmp_path):
+    path = str(tmp_path / "attrs.h5")
+    with h5py.File(path, "w") as f:
+        f.create_group("g")
+    values = {
+        "i8": 5,
+        "f8": 2.5,
+        "i1": np.int8(-3),
+        "u8": np.uint64(2**64 - 1),
+        ">f4": np.array([1.5, 2.5], dtype=">f4"),
+        "f2": np.float16(0.5),
+        "b": True,
+        "barr": np.array([True, False]),
+        "c16": np.complex128(1 + 2j),
+        "bytes": b"raw",
+        "sarr": np.array([b"a", b"bcd"]),
+        "str": "héllo",
+        "strs": ["x", "yz"],
+        "2d": np.arange(12, dtype="<u2").reshape(3, 4),
+        "empty": np.zeros((0,), dtype="<i4"),
+    }
+    with clawhdf5.File(path, "r+") as f:
+        for k, v in values.items():
+            f["g"].attrs[k] = v
+        # Replace one, with another type and size.
+        f["g"].attrs["i8"] = np.arange(100.0)
+    with h5py.File(path, "r") as f:
+        a = f["g"].attrs
+        np.testing.assert_array_equal(a["i8"], np.arange(100.0))
+        assert a["f8"] == 2.5 and a["f8"].dtype == np.float64
+        assert a["i1"] == -3 and a["i1"].dtype == np.int8
+        assert a["u8"] == 2**64 - 1 and a["u8"].dtype == np.uint64
+        assert a[">f4"].dtype == np.dtype(">f4")
+        assert a["f2"].dtype == np.float16
+        assert a["b"] is np.True_ or a["b"] == True  # noqa: E712
+        assert a["barr"].dtype == np.bool_
+        assert a["c16"] == 1 + 2j
+        assert a["bytes"] == b"raw"
+        assert list(a["sarr"]) == [b"a", b"bcd"]
+        # str is stored as fixed-length UTF-8: h5py reads bytes.
+        assert a["str"].decode("utf-8") == "héllo"
+        assert [x.decode() for x in a["strs"]] == ["x", "yz"]
+        assert a["2d"].shape == (3, 4) and a["2d"].dtype == np.dtype("<u2")
+        assert a["empty"].shape == (0,)
+    with clawhdf5.File(path, "r") as f:
+        assert f["g"].attrs["c16"] == 1 + 2j
+        assert f["g"].attrs["str"].decode("utf-8") == "héllo"
+    h5dump_reads(path)
+
+
+def test_many_attributes_move_to_dense_storage(h5py, tmp_path):
+    """Past the compact limit (8 attributes under libver v114) the object's
+    attributes move to dense storage; h5py reads all of them."""
+    path = str(tmp_path / "dense.h5")
+    with h5py.File(path, "w", libver="v114") as f:
+        f.create_dataset("d", data=np.arange(3))
+    with clawhdf5.File(path, "r+") as f:
+        for i in range(20):
+            f["d"].attrs[f"a{i:02d}"] = np.full(i + 1, i, dtype="<i2")
+    with h5py.File(path, "r") as f:
+        assert sorted(f["d"].attrs.keys()) == [f"a{i:02d}" for i in range(20)]
+        for i in range(20):
+            np.testing.assert_array_equal(f["d"].attrs[f"a{i:02d}"], np.full(i + 1, i))
+    h5dump_reads(path)
+
+
+def test_reads_never_see_a_half_written_edit(h5py, tmp_path):
+    """Readers on other threads while one thread rewrites a dataset: every
+    read returns one whole version (all elements equal), never a mix."""
+    path = str(tmp_path / "race.h5")
+    with h5py.File(path, "w") as f:
+        f.create_dataset("d", data=np.zeros((64, 64)), chunks=(16, 16), compression="gzip")
+    errors = []
+    with clawhdf5.File(path, "r+") as f:
+        stop = threading.Event()
+
+        def read():
+            ds = f["d"]
+            while not stop.is_set():
+                a = ds[...]
+                if not (a == a.flat[0]).all():
+                    errors.append(a)
+                    return
+
+        readers = [threading.Thread(target=read) for _ in range(3)]
+        for t in readers:
+            t.start()
+        try:
+            for k in range(1, 25):
+                f["d"][...] = float(k)
+        finally:
+            stop.set()
+            for t in readers:
+                t.join()
+        np.testing.assert_array_equal(f["d"][...], np.full((64, 64), 24.0))
+    assert not errors, "a read saw a partly written dataset"
+
+
+def test_close_releases_the_file(h5py, tmp_path):
+    path = str(tmp_path / "close.h5")
+    with h5py.File(path, "w") as f:
+        f.create_dataset("d", data=np.arange(4, dtype="<i4"))
+    f = clawhdf5.File(path, "r+")
+    ds = f["d"]
+    f.close()
+    # The handle still reads the file as last written, but cannot edit it.
+    np.testing.assert_array_equal(ds[...], np.arange(4))
+    with pytest.raises(OSError, match="closed"):
+        ds[0] = 1
+    with h5py.File(path, "r+") as g:
+        g["d"][0] = 9
diff --git a/crates/clawhdf5-py/tests/test_read_vs_h5py.py b/crates/clawhdf5-py/tests/test_read_vs_h5py.py
index f3132a7..4fab9d2 100644
--- a/crates/clawhdf5-py/tests/test_read_vs_h5py.py
+++ b/crates/clawhdf5-py/tests/test_read_vs_h5py.py
@@ -178,12 +178,15 @@ def _write_fixture(h5py, path):
         g.attrs["depth"] = np.int8(3)
 
 
-@pytest.fixture(scope="module", params=["local", "http", "http-1k-blocks"])
+@pytest.fixture(scope="module", params=["local", "r+", "http", "http-1k-blocks"])
 def pair(request, h5py, tmp_path_factory):
-    """The fixture file through h5py and through clawhdf5: opened locally,
-    and over HTTP range requests (a local server in this process) with the
+    """The fixture file through h5py and through clawhdf5: opened locally
+    (read-only, and for editing: a copy, since editing locks the file), and
+    over HTTP range requests (a local server in this process) with the
     default 1 MiB blocks and with 1 KiB blocks, so every structure is read
     through many small ranges."""
+    import shutil
+
     from conftest import RangeServer
 
     root = tmp_path_factory.mktemp("h5")
@@ -193,6 +196,10 @@ def pair(request, h5py, tmp_path_factory):
     server = None
     if request.param == "local":
         ours = clawhdf5.File(path, "r")
+    elif request.param == "r+":
+        copy = str(root / "editable.h5")
+        shutil.copy(path, copy)
+        ours = clawhdf5.File(copy, "r+")
     else:
         server = RangeServer(root)
         if request.param == "http":
diff --git a/docs/known-issues.md b/docs/known-issues.md
index bf4ae74..e33b80c 100644
--- a/docs/known-issues.md
+++ b/docs/known-issues.md
@@ -116,6 +116,50 @@ space it leaves is too small for its next, larger version.
 **No journal.** A crash while an edit patches existing structures can leave
 the file inconsistent; see the `FileEditor` documentation.
 
+## Python in-place editing (`clawhdf5.File(path, 'r+')`) limits
+
+**Status:** open (added 2026-09-27). The Python bindings edit through
+`FileEditor`, so its limits above apply, raised as `NotImplementedError`
+before anything is written. On top of them:
+
+- **No new or deleted objects:** `create_dataset`/`create_group` in an
+  `'r+'` file, `del f[name]` and `del obj.attrs[name]` raise
+  `NotImplementedError` (the editor changes values, shapes and
+  attributes only). Mode `'a'` works on an existing file only.
+- **Not writable from Python:** compound fields by name (`ds['x'] = …`;
+  whole elements of the same structured dtype are), HDF5 array-type
+  elements, variable-length data, strings padded with spaces or
+  NUL-terminated (libhdf5 converts those differently from numpy; NUL-padded
+  ones, h5py's, are writable), compounds containing such strings, null
+  dataspaces, and index-list writes of more than 2²² elements (write them
+  in slices).
+- **`str` attributes are fixed-length UTF-8**, where h5py writes
+  variable-length strings: h5py reads them back as `bytes`
+  (`numpy.bytes_`), not `str`.
+- **Numeric conversion follows libhdf5's native-order results, not its
+  bugs.** Arrays are converted as libhdf5 converts them (integers
+  saturate, floats are truncated toward zero and clipped), checked value by
+  value against h5py 3.16 / HDF5 2.0 in
+  `crates/clawhdf5-py/tests/test_edit.py::test_numeric_conversions_match_h5py`
+  (2026-09-27, tank). Where libhdf5 itself is inconsistent, clawhdf5
+  differs from h5py on purpose:
+  - NaN into an integer dataset is a `ValueError` (libhdf5 stores 0, the
+    minimum or 2⁶³ depending on the type);
+  - when the dataset or the array is not in native byte order, libhdf5's
+    "soft" conversions store a float in (-1, 0) as the integer minimum and
+    wrap an unsigned value too large for the signed type of the same size
+    (65535 → -1); clawhdf5 gives 0 and the maximum, as libhdf5 does in
+    native order;
+  - libhdf5's native casts that are undefined in C: half floats into
+    unsigned integers (-1 → 65535, +inf → 0), half-float ±inf into signed
+    integers (→ minimum), a float equal to the integer maximum rounded up
+    in its precision (`float32(2**31 - 1)` into `int32`, `float64(2**64 -
+    1)` into `uint64`: → minimum or 0); clawhdf5 saturates;
+  - a double in (65504, 65520) into a half float: libhdf5 stores infinity,
+    clawhdf5 (numpy) rounds to 65504 as IEEE 754 does.
+- **Each edit reopens the file** (a new memory map) so that reads see it;
+  reads from other threads wait while an edit is written.
+
 ## Selection reads that decode more than the selection
 
 **Status:** open (documented 2026-09-26). `Dataset::read_selection` (and so
diff --git a/scripts/ci-test.sh b/scripts/ci-test.sh
index 5731358..7ba09a2 100755
--- a/scripts/ci-test.sh
+++ b/scripts/ci-test.sh
@@ -272,7 +272,10 @@ python_package() {
         -i "$PYTHON" \
         --out "$out/wheel" || return 1
     "$PYTHON" -m pip install --quiet --no-deps --target "$out/site" "$out"/wheel/*.whl || return 1
-    PYTHONPATH="$out/site" "$PYTHON" -m pytest -q -p no:cacheprovider \
+    # The editing tests run `h5rs check` on every file they edit.
+    cargo build -q -p clawhdf5-tools || return 1
+    CLAWHDF5_H5RS="${CARGO_TARGET_DIR:-$root/target}/debug/h5rs" \
+        PYTHONPATH="$out/site" "$PYTHON" -m pytest -q -p no:cacheprovider \
         "$root/crates/clawhdf5-py/tests"
 }
 if "$PYTHON" -m maturin --version >/dev/null 2>&1 && "$PYTHON" -c "import pytest" >/dev/null 2>&1; then