# 0023: The GPUI dependency moves to a pinned Zed revision, and the accessibility adapter is written before it moves

Status: accepted. Supersedes the dependency choice in decision 0017; that
record's disposition of Floem, its account of the licence closure, and its
reasoning about accent colour all stand.

Decision 0017 chose published `gpui 0.2.2` from crates.io over a Zed git
revision, accepted a known-unmet accessibility gate as the price, and described
the way back: "Reversing this decision means changing one dependency line and
one adapter implementation." Both halves of that sentence have since been
falsified. The published crate is not a release line that can be migrated from
later, and the adapter that was supposed to make the migration cheap was never
written.

## The published crate is not a channel

Measured against the crates.io API and Zed's `main` on 2026-09-01:

- `gpui`'s newest version is `0.2.2`, published **2025-10-22**. Every non-yanked
  release was published in a seventeen-day window in October 2025, and nothing
  has been published in the ten months since.
- Zed's `main` still declares `version = "0.2.2"` in `crates/gpui/Cargo.toml`.
  Ten months of development, no version bump: nobody upstream is preparing a
  publish.
- The platform layer has since been split into `gpui_platform`, `gpui_linux`,
  `gpui_macos`, `gpui_windows`, `gpui_wgpu`, `gpui_web` and `gpui_apple`. All of
  them inherit `publish = false` from Zed's workspace; only `crates/gpui` sets
  `publish = true`. On `main` the `x11` and `wayland` features of `gpui` itself
  are now empty markers, because the backends they used to select live in
  `gpui_linux`.
- Published 0.2.2 has no `accesskit` dependency at all, and
  `docs.rs/gpui/0.2.2/gpui/enum.Role.html` is a 404. At `ce48461e`,
  `crates/gpui/Cargo.toml:50` carries `accesskit.workspace = true`.
- Published 0.2.2 renders through `blade-graphics`. Upstream rendering has moved
  to `gpui_wgpu`.

So the choice was never "0.2.2 now, a newer published version later". Even if
Zed published tomorrow, the publishable crate no longer contains a Linux
platform backend. Staying means freezing permanently on an October-2025
artifact, and permanently failing the hard gate in `UX_VERTICAL_SLICE.md` that
says the accessibility tree must name each pane, its selected state, its
execution target, and its waiting state.

## A pinned revision is not an unpinned upstream

Decision 0017's objection was "a git dependency on an unpinned upstream". That
is not what is adopted here. With a full-SHA `rev`, a committed `Cargo.lock`,
and `deny.toml` naming the permitted Git sources, a git dependency is *more*
pinned than `gpui = "0.2.2"`, which is a caret requirement that would silently
accept a `0.2.3` nobody reviewed. The real cost is that bumping the revision
becomes a deliberate scheduled act. That cost is accepted and named here rather
than avoided.

The target revision is `ce48461eaadd16c65c31f835511ab96bd3b6e746`
(2026-08-31), which is the revision `spikes/gpui-accessibility-probe` already
pins and already proves compiles with stable accessibility IDs, workspace/pane
roles, selected state, and real `EntityInputHandler`/`ElementInputHandler`
implementations. Its toolchain is Rust 1.97.1, moving the workspace from 1.94.1.

## The adapter comes first, and it does not exist yet

Decision 0017 said the mitigation was "a narrow `a11y` adapter trait, introduced
now while there are few call sites", so that the migration would be "a swap
rather than a rewrite". No such trait, module, or file exists anywhere in the
workspace. The only trace is `apps/lumbridge/src/sidebar/model.rs`, where
`describe()` produces the screen-reader sentence, is tested, is called by
nothing, and carries an `#[allow(dead_code, reason = "wired up with the
accessibility adapter, decision 0017")]`.

Meanwhile the call sites stopped being few: `apps/lumbridge/src` now builds 147
elements and gives only 20 of them a stable identity.

Writing the adapter is therefore the first step and happens **before** the
dependency changes, on 0.2.2, where its bodies are no-ops. That keeps the
dependency swap a swap, as 0017 intended, instead of paying for the swap and
the wiring in one unreviewable change. The `#[allow(dead_code)]` is deleted by
that step rather than carried.

## Order, and why each step is alone

Each stage lands separately because each fails differently, and a combined
failure is unattributable:

1. **The `a11y` adapter, still on 0.2.2.** No-op bodies; wire the elements that
   carry identity; give `describe()` a caller.
2. **The toolchain, still on 0.2.2.** Move `rust-toolchain.toml`, the workspace
   `rust-version`, and the toolchain named in `.gitea/workflows/ci.yml` — which
   hardcodes it twice rather than reading the file — to 1.97.1. Three minor
   versions of new `clippy::pedantic` lints arrive as hard errors under
   `-D warnings`. Isolating this stage is what stops a lint avalanche being
   mistaken for a GPUI API break.
3. **`deny.toml`.** Add the Zed Git sources to `allow-git`. This record
   originally said there were two, `zed-industries/zed` and
   `zed-industries/font-kit`. **That was wrong: there are five.** The graph also
   carries `xim-rs` (X11 input methods, feeding `gpui_linux` -- the IME path,
   so load-bearing for a hard gate rather than incidental), `wasm_thread`
   (reached twice, through `gpui_web` and through `scheduler` <- `gpui`), and
   `scap`, which resolves in the lockfile without being reachable in this
   target's graph and so is named before some later platform's build finds it
   first.

   The error is worth keeping visible rather than quietly corrected. It was made
   by reading the manifests of the two crates being added, which is exactly the
   reasoning that misses a transitive Git source; only resolving the graph
   answers the question. `cargo deny check sources` is what caught it, on the
   commit after the allowance was written -- which is the argument for running
   that check at all.
4. **The dependency swap.** `gpui` and `gpui_platform` at the pinned rev;
   `wayland`/`x11` requested on `gpui_platform`, where they still mean
   something; and `Application::new()` — which does not exist at that revision —
   becomes `gpui_platform::application()`.
5. **Real semantics in the adapter**, and the probe's AccessKit tree test ported
   into `apps/lumbridge`.
6. **The probe changes job**: it stops proving AccessKit exists and starts
   proving the next revision still compiles.

The renderer swap from blade to wgpu in stage 4 was called the least
characterised risk here, and at build time it cost nothing: the whole closure
compiled on the existing `libxkbcommon-x11` shim, with no Vulkan or GL loader
problem and no new native dependency. That is not the same as the risk being
disproven. Nothing in `cargo check`, `clippy`, or the test suite opens a window,
so the renderer has still never run. The risk moved from build time to first
launch; it did not go away.

The rest of stage 4 was smaller than expected. The entire API delta across the
ten-month jump was three signatures over nineteen sites: `Window::focus` taking
the `App` it used to reach through the window, `flex_shrink` taking the factor
it used to assume, and `Application::new` becoming
`gpui_platform::application()`. No crate outside `apps/lumbridge` needed any
change, because none of them imports a gpui type -- the narrow boundary
`ARCHITECTURE.md` insisted on is the reason this was an afternoon.

## What is knowingly still unmet

A passing unit test over an AccessKit `TreeUpdate` is not an assistive-technology
claim. The gate is not met until AT-SPI/Orca on Linux and VoiceOver on macOS
have been driven by hand. Stage 5 makes the semantics real; it does not close
the gate, and nothing in this record should be read as closing it.

## Enforcement

This record is not self-enforcing, and the previous one was not obeyed. Three
mechanical checks carry it instead:

- `scripts/workspace-guard.sh` asserts the gpui source and version in
  `Cargo.lock` — the lockfile, because a manifest states an intent while the
  lockfile states what would actually compile. It runs in the fast headless job,
  which is not `continue-on-error`, so it gates every push.
- `deny.toml`'s `unknown-git = "deny"` plus an explicit `allow-git` list makes a
  fourth Git source a CI failure. `scripts/ci.sh` now actually runs
  `cargo deny check sources`; it previously ran only `check licenses`, and the
  source policy had never once been executed.
- `AGENTS.md` names the gpui dependency, `rust-toolchain.toml`, the workspace
  `rust-version`, and `deny.toml` as things a contributor or agent may not
  change without a decision record.