Skip to content

231_lsp-workspace-singleton.md

Latest commit

 

History

History
81 lines (74 loc) · 3.09 KB

231_lsp-workspace-singleton.md

File metadata and controls

81 lines (74 loc) · 3.09 KB
Raw
id 231
title LSP newest-wins workspace singleton
status ✅
summary A leaked editor extension host can outlive its window while staying alive and holding the LSP server's stdin open, so neither the stdin-EOF exit nor the processId watchdog fires and an orphaned `mdsmith lsp` keeps running beside the one a reload spawned. Add a newest-wins workspace singleton: each server records its workspace root in a shared registry under a per-process id; once a newer server claims the same root, the older one sends `mdsmith/superseded` and exits, and the VS Code extension suppresses the restart so exactly one server stays live per workspace.
model
depends-on

LSP newest-wins workspace singleton

Goal

The processId watchdog (internal/lsp/parentwatch.go) reaps an mdsmith lsp server whose editor host has died. It cannot reap one whose host is alive but orphaned — a VS Code / code-server extension host that survives its own window after an abrupt reload. That host keeps the server's stdin pipe open (so no EOF arrives) and still reports as alive (so the watchdog stays quiet), and its language client even respawns the server if it is killed. The result is two servers for one workspace where the user wants one.

Add a newest-wins workspace singleton so the older server steps aside on its own, leaving exactly one live server per workspace root.

Tasks

  1. Add the registry and watcher in internal/lsp/singleton.go: a per-workspace owner file under the OS cache dir (atomic claim, plain read), a watchSingleton poll loop mirroring watchParentProcess, and a startSingletonWatch that claims the workspace and steps aside when a newer owner appears.
  2. On step-aside, send the mdsmith/superseded server-to-client notification, then exit — so the client knows the close is intentional.
  3. Gate it behind Options.EnableWorkspaceSingleton in internal/lsp/server.go; call startSingletonWatch from handleInitialize; enable it in cmd/mdsmith/lsp.go.
  4. In the VS Code extension, suppress the restart on mdsmith/superseded: a decideClose policy in editors/vscode/src/wiring.ts and a markSuperseded hook on the error handler in editors/vscode/src/extension.ts.
  5. Document the behavior and the mdsmith/superseded contract in the troubleshooting section of the VS Code guide.

Acceptance Criteria

  • A second server started for the same workspace root makes the older one send mdsmith/superseded and exit, leaving one live server.
  • The feature is off by default so unit tests neither write to the real cache dir nor leak a watcher goroutine; the singleton tests drive the seams directly.
  • The VS Code error handler does not restart the server after mdsmith/superseded.
  • go test ./... and go tool golangci-lint run pass.
  • go run ./cmd/mdsmith check . passes.