Ben Kunkle
d224ea86e8
Closes EP-193
Edit prediction context collection previously reused the editor
goto-definition / goto-type-definition path, which resolves every LSP
result into a `LocationLink` — opening a buffer per target (worktree
creation, buffer registration, anchor conversion) before edit prediction
gets a chance to filter. On remote projects the host additionally
created a peer-visible buffer per link and the client waited on each
one.
This adds a bespoke request path that returns raw results first and only
does the expensive work for results that survive filtering:
- New `EditPredictionDefinition { path: ProjectPath, range:
Range<Unclipped<PointUtf16>> }` boundary type. LSP wire types never
leave `lsp_command.rs`; the workspace-only filter and URI→`ProjectPath`
resolution happen while normalizing the LSP response, before any buffer
exists, so the type encodes the workspace-only invariant.
- New `GetEditPredictionDefinitions` /
`GetEditPredictionTypeDefinitions` LSP commands and proto messages.
Responses carry only path + UTF-16 range — the host never calls
`create_buffer_for_peer` and the client never waits on remote buffers.
- One public API: `Project::edit_prediction_definitions(buffer,
position, include_type_definitions, cx)` fires both LSP requests
concurrently and returns a merged, deduped list ("not capable" = empty).
The definition/type-definition split was never used downstream, so
`CacheEntry` now holds a single list and the fetch pipeline runs one
task per identifier.
- `edit_prediction_context` dedupes raw results before opening buffers;
survivors open via `project.open_buffer(ProjectPath)` (skipping the
invisible-worktree/yarn machinery of `open_local_buffer_via_lsp`, and
working identically on remote projects), then clip → anchor →
`MAX_TARGET_LEN`.
- Removes the `workspace_only` flag from `GetDefinitions` /
`GetTypeDefinitions`: edit prediction was its only user. The editor path
always sent `false`, which proto3 doesn't encode, so normal
goto-definition requests are wire-identical; old peers still sending
`true` get unfiltered results (graceful degradation). The proto field
numbers are `reserved`.
Tested with a local test proving filtering precedes buffer opening (an
out-of-workspace target never spawns a worktree, an oversized target is
excluded from related files) and a collab test proving the remote
round-trip returns correct paths/ranges without opening buffers on the
client.
Release Notes:
- N/A
Zed Server
This crate is what we run at https://collab.zed.dev.
It contains our back-end logic for collaboration, to which we connect from the Zed client via a websocket after authenticating via https://zed.dev, which is a separate repo running on Vercel.
Local Development
Database setup
Before you can run the collab server locally, you'll need to set up a zed Postgres database. Follow the steps sequentially:
- Ensure you have postgres installed. If not, install with
brew install postgresql@15. - Follow the steps on Brew's formula and verify your
$PATHcontains/opt/homebrew/opt/postgresql@15/bin. - If you hadn't done it before, create the
postgresuser withcreateuser -s postgres. - You are now ready to run the
bootstrapscript:
script/bootstrap
This script will set up the zed Postgres database, and populate it with some users. It requires internet access, because it fetches some users from the GitHub API.
The script will create several admin users, who you'll sign in as by default when developing locally. The GitHub logins for the default users are specified in the seed.default.json file.
To use a different set of admin users, create crates/collab/seed.json.
{
"admins": ["yourgithubhere"],
"channels": ["zed"]
}
Testing collaborative features locally
In one terminal, run Zed's collaboration server and the livekit dev server:
foreman start
In a second terminal, run two or more instances of Zed.
script/zed-local -2
This script starts one to four instances of Zed, depending on the -2, -3 or -4 flags. Each instance will be connected to the local collab server, signed in as a different user from seed.json or seed.default.json.
Deployment
We run two instances of collab:
- Staging (https://staging-collab.zed.dev)
- Production (https://collab.zed.dev)
Both of these run on the Kubernetes cluster hosted in Digital Ocean.
Deployment is triggered by pushing to the collab-staging (or collab-production) tag in GitHub. The best way to do this is:
./script/deploy-collab staging./script/deploy-collab production
You can tell what is currently deployed with ./script/what-is-deployed.