diff --git a/changes/+roadmap.doc.md b/changes/+roadmap.doc.md
new file mode 100644
index 0000000000..8a473acac9
--- /dev/null
+++ b/changes/+roadmap.doc.md
@@ -0,0 +1 @@
+Added a Roadmap page to the documentation outlining future plans and intended changes to the library.
diff --git a/docs/roadmap.md b/docs/roadmap.md
new file mode 100644
index 0000000000..cbe7ce79d6
--- /dev/null
+++ b/docs/roadmap.md
@@ -0,0 +1,303 @@
+# Roadmap
+
+This page describes where Zarr-Python is headed: the goals for the next major
+cycle of work, the changes we intend to make, and how those changes will be
+released. It is a living document; discussion and counter-proposals are welcome
+on the
+[zarr-python issue tracker](https://github.com/zarr-developers/zarr-python/issues).
+
+*The history of this roadmap, including the detailed technical proposals it
+was distilled from, can be traced in the
+[zarr-python-planning](https://github.com/d-v-b/zarr-python-planning)
+repository.*
+
+!!! note
+
+    This roadmap reflects the current thinking of the core developers. It is a
+    statement of direction, not a schedule: the work ships when it is ready,
+    and individual items may change shape as the proposals are discussed and
+    refined.
+
+## Where we are
+
+The [3.0 release](https://github.com/zarr-developers/zarr-python/releases/tag/v3.0.0)
+was a total redesign of the library's internals, with three goals: full support
+for the Zarr V2 and V3 storage formats, storage APIs ergonomic for high-latency
+(cloud) storage, and backwards compatibility with Zarr-Python 2.x where
+possible. Those goals were largely achieved, and more than a year on, the
+2.x → 3.x transition is effectively resolved.
+
+The 3.x redesign was carried out under hard backwards-compatibility
+constraints, and it inherited many structural patterns from the 2.x
+implementation it replaced. The library has never had a release cycle whose
+primary goal was the *shape* of the internals. The next body of work — which we
+call **"v4"** — is that overdue investment.
+
+## Goals
+
+If the 3.0 goals could be sloganized as "migrate to Zarr V3, and improve cloud
+storage support", the slogan for the v4 goals is:
+**"a frictionless Zarr-based Python ecosystem for chunked arrays"**. Zarr-Python
+should be *foundational* for the growing number of Python packages that work
+with data in the Zarr format. Concretely, that means pushing in these
+directions:
+
+- Give Zarr-Python users excellent performance, out of the box.
+- Make Zarr-Python APIs ergonomic and useful for developers.
+- Expand our scope to cover vital quality-of-life routines like data copying,
+  rechunking, and the like.
+- Ease the growth of Python tools across all levels of the Zarr stack.
+- Accelerate the implementation of new codecs, chunk grids, chunk key
+  encodings, etc.
+
+An important design input: [zarrs](https://github.com/zarrs/zarrs) (Rust) and
+[TensorStore](https://github.com/google/tensorstore) (C++) are two independent
+Zarr implementations that have converged on the same architectural patterns —
+sync-first codec APIs, per-codec concurrency budgets, adaptive sharded-read
+strategies, request deduplication, conditional reads. We treat them as
+complementary rather than competitive: Zarr-Python aims to be the best
+pure-Python Zarr implementation *and* the best wrapper around the
+compiled-language implementations, so that users who need native throughput can
+get it without leaving the Zarr-Python API surface.
+
+## The Zarr stack
+
+Different applications need different levels of Zarr support: a convention
+validator only needs to read metadata documents; a visualization tool may only
+need read-only array access; other tools need everything. We think of this as a
+"Zarr stack", from most abstract to most concrete:
+
+1. **Conventions** — application and/or domain-specific schemas built on top of Zarr (OME-NGFF,
+   GeoZarr, anndata-zarr, multiscales).
+2. **Groups** — Zarr hierarchies, traversal, group-level attributes.
+3. **Arrays** — the user-facing array object, plus indexing and slicing.
+4. **Chunk decoding** — the codec pipeline.
+5. **Chunk addressing** — chunk grids and key encodings that map array
+   coordinates to store keys.
+6. **Stores** — the key-value layer.
+7. **Metadata** — pure data documents describing arrays and groups.
+
+Today, Zarr-Python is a monolith that serves every level: a consumer who only
+needs metadata handling has to install the full dependency footprint of the
+whole library, and a faster chunk-decoding implementation cannot plug in
+without re-implementing the layers above it. The v4 direction is to re-shape
+Zarr-Python around the stack, so that each level is something you can depend
+on, conform to, or replace, without buying every other level:
+
+- **A focused package per level** — `zarr-metadata`, `zarr-store`,
+  `zarr-codec`, `zarr-dtype`, with `zarr` as the facade that composes them.
+  The first of these,
+  [`zarr-metadata`](https://pypi.org/project/zarr-metadata/), is already
+  published.
+- **A documented interface per level** — capability protocols for stores, a
+  small stateless codec API, pure-data dtypes.
+- **A conformance suite per level** — so that alternative implementations of a
+  level can verify they behave correctly.
+- **Engine pluggability at the chunk-decoding level** — alternative engines
+  (zarrs, TensorStore) can take over IO without re-implementing hierarchy
+  traversal, indexing, or metadata handling.
+
+## What we intend to change
+
+Each theme below is backed by a detailed technical proposal; the summaries
+here describe the intended end state.
+
+### Foundation: a functional core
+
+Refactor the internals around a *functional core* — pure data structures and
+pure functions for the algebra of Zarr (metadata, chunk layouts, slice
+planning, codec walking) — with the side-effecting protocols (stores, codecs)
+at the edges. This is an internal change that makes the per-level package split
+implementable and provides a clean substrate for engine pluggability.
+
+### Foundation: a formal hierarchy layer
+
+Name and specify the layer that sits between the store API (key-agnostic
+bytes) and the user-facing `Array` / `Group` facade, as a small set of typed
+verbs (`read_array_metadata`, `write_chunk`, `list_children`,
+`read_selection`, ...). Alternative engines implement the verbs end-to-end;
+hierarchy-aware caching wraps them; chunk-introspection APIs expose them.
+
+### Codecs
+
+The current codec API wraps every codec in an unnecessary async layer (a
+profiling hotspot), bakes batching into every signature, and forces output
+allocation even when the caller has a buffer ready. Rewrite the codec API as a
+small, stateless capability bundle — sync-first encode/decode, single-element
+signatures, optional `decode_into`, capability flags — decoupled from the rest
+of the library, with a compatibility shim for existing codecs and clear paths
+for migrating Zarr V2 codecs that still have no V3 equivalent.
+
+### Stores
+
+The store abstraction conflates lifecycle, path handling, sync/async,
+capability advertisement, and read-only semantics into one inheritance
+hierarchy, and the resulting friction has produced a recurring stream of
+regressions. Redesign stores as composable capability protocols (`Get`, `Put`,
+`List`, ...) with composable wrappers (caching, range coalescing, retries),
+transactional semantics, and a shared conformance suite that backends and
+wrappers parameterize.
+
+### Performance
+
+A cross-cutting theme that ties the codec, store, and functional-core work
+into one performance story: typed, library-owned concurrency resources with
+dask-safe defaults; synchronous codec encode/decode on the default read path;
+range coalescing; pre-allocated decode buffers; in-flight request
+deduplication; ETag-style conditional reads; a unified caching substrate with
+sensible defaults; an adaptive whole-shard-vs-coalesced read strategy; and
+pluggable high-performance backends (zarrs, TensorStore) selectable with a
+keyword argument, so the same `Array` and `Group` — and the same Xarray, Dask,
+and napari integrations — work at native throughput. A benchmark suite for the
+target access patterns lands first, so every performance lever ships with
+before/after numbers.
+
+### Lazy indexing
+
+`Array.__getitem__` performs IO eagerly and returns NumPy arrays, which makes Zarr
+arrays the odd one out among modern array libraries and blocks compliance with
+the [Python Array API](https://data-apis.org/array-api/) standard. Add an
+opt-in `array.lazy[...]` accessor backed by a stable coordinate-mapping algebra
+(the `IndexTransform` work in
+[#3906](https://github.com/zarr-developers/zarr-python/pull/3906)), plus a
+small query planner that turns chained selections into a single IO plan before
+any chunks are fetched. No new array type is introduced. Whether the *default*
+of bare `array[...]` ever flips from eager to lazy is an explicit, separate
+decision — see [decision points](#decision-points) below.
+
+### Data types
+
+First-class support for ML-specific dtypes — `bfloat16`, the `float8`
+variants, packed `int4`/`uint4` — via
+[`ml_dtypes`](https://github.com/jax-ml/ml_dtypes), using the exact identifiers
+registered in `zarr-extensions` so the data stays readable by other
+implementations. Ragged arrays, variable-length strings, and an investigation
+of Apache Arrow as a substrate for the dtypes the Array API cannot express are
+follow-on work on the same substrate.
+
+### Device-agnostic IO
+
+Make Zarr-Python's IO surfaces device-agnostic rather than adding GPU support
+as a bolted-on feature: stores and codecs grow APIs for writing into a
+caller-provided buffer (`read_into`, `decode_into`), and the `Array` facade
+returns array-like objects in the user's chosen Array API namespace. GPU
+support falls out once the assumption of CPU destinations is removed, and CPU
+paths get faster too, because pre-allocated output buffers eliminate per-chunk
+allocation.
+
+### Observability
+
+Two pillars: **performance metrics and tracing** (a small library-owned
+`Metrics` object plus OpenTelemetry auto-instrumentation across stores, codecs,
+caches, and the engine boundary) and **stored-state introspection** (public
+APIs for asking about chunk-level structure, materialization, byte ranges, and
+storage footprint without reading the chunks — the surface projects like
+VirtualiZarr and Kerchunk have been asking for).
+
+### Configuration, registries, and plugins
+
+Move configuration from "global mutable state read implicitly" to "typed data
+passed explicitly": a typed config object replacing the untyped global `donfig`
+dict, array-scoped runtime config passed at open time, a registry redesign that
+addresses implementations by stable identity and resolves plugin name-conflicts
+deliberately, and named profiles replacing global mutators. This substrate is
+where the performance-lever defaults (concurrency, caching, engine selection)
+will live, so it lands early.
+
+
+### Coordinated and distributed writes
+
+Give the two patterns that actually produce large Zarr archives — parallel
+disjoint-region writes and append-along-axis growth — a design home: disjoint
+chunk-aligned region writes with alignment *checked* rather than assumed, a
+create-then-hand-out-regions primitive, and single-writer resize/append, all on
+plain Zarr V3. Stronger guarantees (atomicity, reader isolation, concurrent
+appenders) are enabled through the seam a transactional engine such as
+[Icechunk](https://icechunk.io/) builds on, rather than implemented in
+Zarr-Python itself.
+
+### Missing APIs
+
+User-facing conveniences that users have been asking for, in some cases for
+years: hierarchy navigation helpers, chunk introspection, explicit constructors
+replacing `mode=`, a typed exception hierarchy, rich reprs, context-manager
+support, data copying, and an in-library rechunking primitive.
+
+## How the work will be released
+
+**"v4" names this whole body of work, delivered across many releases — it is
+not a single "4.0" feature release.** The work is organized into three streams
+that run in parallel:
+
+| Stream | Release vehicle | Scope |
+|---|---|---|
+| **Additive value** | 3.x minor releases, shipping continuously | The overwhelming majority of the plan, including the entire foundation. No migration required. |
+| **Deprecation accumulation** | Warnings across the 3.x line | Each surface is deprecated only *after* its additive replacement has shipped, so users always have a migration target before they see a warning. |
+| **Breaking removals** | One minimal, late major release (4.0.0) | Removal of the deprecated surfaces, and *only* those, after deprecation windows have elapsed and downstream libraries have had release windows to adapt. |
+
+The additive stream is itself roughly ordered:
+
+1. **Ship-now wins** — dependency-free improvements that land first: the
+   benchmark suite, store-layer range coalescing, in-flight request
+   deduplication, the sync codec path on default reads, ML dtype support,
+   constructor and display UX.
+2. **Foundation** — the functional-core refactor, the per-level package split,
+   the new stores API, the hierarchy verbs, the typed configuration substrate,
+   the full concurrency and caching rework, and the codec API rewrite. Mostly
+   invisible to users, all additive.
+3. **User-facing surface** — opt-in lazy indexing and the query planner,
+   device-agnostic IO, observability, chunk introspection, and the zarrs and
+   TensorStore engine wrappers, built on the foundation.
+
+The eventual 4.0.0 release contains only removals whose replacements shipped
+earlier: the legacy `Store` ABC and the `Buffer`/`prototype` read contract, the
+`mode=` constructors, the internal `sync()` bridge, and — conditionally — the
+eager `array[...]` path. Nothing new is delivered there; it is the only release
+downstream maintainers must treat as breaking, and it arrives after the value
+has already been delivered additively.
+
+### Backwards-compatibility commitments
+
+The v4 work changes the public API: methods will be renamed, signatures will
+change, deprecated patterns will be removed, and the codec and store APIs will
+be rewritten. We believe the changes are worth the cost, and we commit to the
+following:
+
+- **Conformance with community standards.** Where a relevant cross-language
+  standard exists, we conform to it: the Python Array API at the array surface,
+  the Zarr V3 spec and its extensions at the storage layer, OpenTelemetry for
+  tracing, and standard buffer-protocol and device-interop conventions for
+  device-agnostic IO.
+- **Functional coverage.** Anything you can do in Zarr-Python 3.x you will
+  still be able to do once the v4 work has landed — sometimes through a renamed
+  API, but the capability is preserved. We will not remove the ability to read
+  or write any Zarr-format data that 3.x supports.
+- **A deprecation window for every change.** Renames and removals land through
+  deprecation cycles, and downstream libraries (Xarray, Dask, napari) get
+  release windows to absorb each change before the next one lands.
+- **Generous legacy support** If necessary, we can keep old code around in a `legacy` module. Pydantic used a similar strategy to manage their 2.0 release: see https://pydantic.dev/docs/validation/dev/get-started/migration/#using-pydantic-v1-features-in-a-v1v2-environment.
+
+### Decision points
+
+Flipping the default of bare `array[...]` from eager to lazy is the single
+highest-migration-cost item in the plan, so it is handled as an explicit
+decision, not bundled into the additive work. The opt-in `array.lazy[...]`
+accessor ships first, with no default change. Whether the default ever flips
+hinges on whether Array API conformance at the bare-`__getitem__` surface turns
+out to be a hard requirement; if it does, the flip happens as a long-window
+deprecation with an explicit eager escape hatch and downstream coordination —
+never as a reason to adopt a major version.
+
+## How to get involved
+
+- **Discuss the plans.** Comments and counter-proposals on any of the themes
+  above are welcome on the
+  [issue tracker](https://github.com/zarr-developers/zarr-python/issues) and in
+  the [developer chat](https://ossci.zulipchat.com/).
+- **Review in-flight work.** The `IndexTransform` algebra that lazy indexing is
+  built on is in review at
+  [#3906](https://github.com/zarr-developers/zarr-python/pull/3906).
+- **Weigh in as a downstream maintainer.** If your project's use of
+  Zarr-Python would be affected by the codec API rewrite, the stores rewrite,
+  or the lazy-indexing work, the planning phase is the time to surface
+  workloads or patterns that don't fit.
\ No newline at end of file
diff --git a/mkdocs.yml b/mkdocs.yml
index 46bfc1764c..25516a56dd 100644
--- a/mkdocs.yml
+++ b/mkdocs.yml
@@ -93,6 +93,7 @@ nav:
       - '<code class="doc-symbol doc-symbol-toc doc-symbol-function"></code> <code>zarr.zeros</code>': api/zarr/functions/zeros.md
       - '<code class="doc-symbol doc-symbol-toc doc-symbol-function"></code> <code>zarr.zeros_like</code>': api/zarr/functions/zeros_like.md
   - release-notes.md
+  - roadmap.md
   - contributing.md
 hooks:
   - mkdocs_hooks.py