Add lifecycle-aware NullFormatter placeholder for formatter tool extensions

[Copilot]

Formatter extensions (black, autopep8, isort) register a placeholder DocumentFormattingEditProvider at activation so the extension appears in the picker before the LSP starts — but never dispose it once the LSP reaches Running. VS Code sees two providers with the same selector and shows the extension twice in the formatter picker (vscode-black-formatter#752).

The fix moves the full placeholder lifecycle into this shared package so every formatter extension gets it for free.

New: ToolConfig.isFormatter

Optional boolean flag. When true, createToolContext automatically manages the placeholder lifetime. Linter-only tools (pylint, flake8, mypy, …) leave it unset — zero behavioral change.

New: NullFormatter class (typescript/src/nullFormatter.ts)

Lifecycle-aware wrapper around registerDocumentFormattingEditProvider. Exported as public API for consumers that want manual control.

export class NullFormatter implements Disposable {
    register(): void   // no-op if already registered
    unregister(): void // no-op if not registered
    isRegistered(): boolean
    dispose(): void
}

createToolContext wiring (typescript/src/activation.ts)

When isFormatter: true:

• Calls nullFormatter.register() eagerly at context creation (extension visible in picker from activation)
• After restartServer() resolves: if client is already State.Running (initial transition is missed by the onDidChangeState listener — same pattern as the existing busy-status reset), disposes the placeholder immediately
• onDidChangeState listener: Running → unregister(), Stopped/Starting → register() (covers restart cycles)
• dispose() calls nullFormatter?.dispose()

Tests

Six new tests in activation.test.ts cover: activation-time registration, disposal on State.Running via listener, full Stopped → Starting → Running restart cycle, the already-Running synchronous guard, and isFormatter: false/unset no-ops. Mock client extended with state getter and simulateStateChange().

Version bump

0.6.1 → 0.7.0 (new public API: ToolConfig.isFormatter, NullFormatter export).

[rchiodo]

Core fix looks good. A few non-blocking concerns worth addressing before merge: the re-registration coverage claimed in the CHANGELOG/JSDoc doesn't fully hold for extension-driven restarts, the failure path can leave the extension with no placeholder, and the new shared-mock additions are unused by the actual test suite.

[rchiodo]

Approved via Review Center.

[rchiodo]

Solid fix that moves the formatter placeholder lifecycle into the shared package. One thing worth a look: the eager register() at the top of runServer() (needed to cover failure paths) can transiently re-introduce the duplicate-formatter window during a healthy restart — guarding it on state !== Running would close that gap. Otherwise the change and tests look good.

[rchiodo]

Approved via Review Center.

[rchiodo]

GitHub cannot anchor PR review comments to unchanged lines in the diff. Falling back to a general PR comment for typescript/src/activation.ts:L188.

📍 typescript/src/activation.ts:184
The catch (ex) block calls nullFormatter?.register() with no guard, unlike the top-of-runServer register you correctly gated on state !== State.Running. The Skeptic flagged this and I verified it against the diff: when a healthy Running client exists and getProjectRoot()/getWorkspaceSettings() throws, this registers the placeholder while the old client is still Running and serving textDocument/formatting → two providers on the same selector (#752). Worse than the prior transient — the failed restart leaves ctx.lsClient unchanged and Running, so the duplicate persists until the next successful restart. Apply the same guard: if (!disposed && nullFormatter && (!ctx.lsClient || ctx.lsClient.state !== State.Running)) nullFormatter.register(); (this also prevents resurrecting the placeholder after dispose() during an in-flight restart).

[verified]

[rchiodo]

Solid fix overall, but one blocking issue: the failure-path register() in runServer's catch block lacks the same state !== State.Running guard you correctly applied at the top of the function, which can re-introduce the duplicate-formatter symptom (#752) when a restart fails with a healthy client still running. Please apply the same guard there.

[edvilme]

Addressed in 4a2bfc2. The catch block's
egister()\ now uses the same !ctx.lsClient || ctx.lsClient.state !== State.Running\ guard, preventing the duplicate-formatter symptom when a restart fails while a healthy client is still running.

[rchiodo]

GitHub cannot anchor PR review comments to unchanged lines in the diff. Falling back to a general PR comment for typescript/src/activation.ts:L186.

The empty-interpreter branch (workspaceSetting.interpreter.length === 0) stops the client, nulls ctx.lsClient, disposes serverDisposables, and returns — but unlike the other no-running-client exits it never calls nullFormatter.register(). If a Running formatter has its interpreter cleared, the top-of-runServer guard sees the old client still Running and skips register(), so the placeholder surviving depends on stop() synchronously emitting State.Stopped to the listener. To make this robust, either add the standard guard (if (nullFormatter && (!ctx.lsClient || ctx.lsClient.state !== State.Running)) nullFormatter.register();) to this branch, or add a test that clears the interpreter while Running and asserts the placeholder ends registered.

[rchiodo]

Solid fix with good test coverage. Two things before merge: remove the inline #752 reference from the code comment (keep it in the CHANGELOG/PR only), and double-check the empty-interpreter restart branch re-registers the placeholder so a Running formatter that loses its interpreter can't disappear from the picker.

[edvilme]

Addressed in 76e1bad:

• Removed the inline #752\ reference from the code comment (kept in CHANGELOG/PR only).
• Added
  ullFormatter?.register()\ in the empty-interpreter branch so a Running formatter that loses its interpreter stays visible in the picker.

[rchiodo]

Solid, well-tested fix for the duplicate-formatter regression. A couple of non-blocking notes below on the empty-interpreter exit path and test fidelity; safe to merge once you've had a look.

[rchiodo]

Approved via Review Center.