@hydranium/data-server 1.0.0-next.27 → 1.0.0-next.270

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
@@ -11,13 +11,20 @@ import {
11
11
  createRpcProxy,
12
12
  defineMessage,
13
13
  DisposableCollection,
14
+ HYDRANIUM_ERROR_CODES,
14
15
  isDocumentSource,
15
16
  isElementSource,
17
+ isResponseError,
16
18
  isSyntheticSource,
19
+ messageData,
17
20
  messageError,
18
21
  ReferenceSource,
22
+ SessionClosedError,
23
+ TIMED_OUT,
19
24
  type CloseModelArgs,
25
+ type HydraniumMessageData,
20
26
  type HydraniumResponseError,
27
+ type CanonicalUri,
21
28
  type Disposable,
22
29
  type ElementSource,
23
30
  type FindNextNameArgs,
@@ -29,10 +36,16 @@ import {
29
36
  type ReferenceContext,
30
37
  type ReferenceRequest,
31
38
  type ReferenceTarget,
39
+ type TransferModelSnapshot,
40
+ type ModelVersion,
41
+ type TextState,
32
42
  type Tracer,
33
43
  type TransferDiagnostic,
34
- type TransferDocument,
35
- type TransferElement
44
+ TransferDocument,
45
+ type TransferSavedDocument,
46
+ type TransferElement,
47
+ UNKNOWN_CLIENT_ID,
48
+ textHash
36
49
  } from '@hydranium/protocol';
37
50
  import {
38
51
  DATA_SERVER_DIAGNOSTICS_METHODS,
@@ -45,24 +58,29 @@ import {
45
58
  type StartProfilingArgs,
46
59
  type StopProfilingArgs,
47
60
  type WriteServerHeapSnapshotArgs,
61
+ type CloseSessionArgs,
62
+ type CreateModelDocumentArgs,
63
+ type CreateSessionArgs,
48
64
  type GetModelDocumentArgs,
49
65
  type GetProjectForUriArgs,
66
+ type TransferPersistDocumentArgs,
50
67
  type TransferSaveDocumentArgs,
51
68
  type WatchModelDocumentArgs,
52
69
  type TransferDocumentSavedEvent,
53
70
  type TransferDocumentUpdatedEvent,
54
- type TransferUpdateDocumentArgs
71
+ type TransferUpdateDocumentArgs,
72
+ type TransferUpdateDocumentsArgs
55
73
  } from '@hydranium/protocol/data';
56
- import { REVERT_ON_CLOSE_CLIENT_ID, UNKNOWN_CLIENT_ID } from '@hydranium/core';
74
+ import { DOCUMENT_RELEASE_CLIENT_ID } from '@hydranium/core';
57
75
  import { defaultDataServerDiagnostics } from './default-diagnostics.js';
58
76
 
59
77
  /**
60
78
  * Raised when a profiling command arrives with no capture running.
61
79
  *
62
- * A `ResponseError` rather than a plain `Error` so the identity survives the
63
- * hop: an unwrapped throw inside an RPC handler reaches the client as a generic
64
- * internal error, which carries no code for a caller to switch on and no
65
- * envelope for a translating host to read.
80
+ * A `ResponseError` rather than a plain `Error` so the identity survives to the
81
+ * response boundary, which is where the message is rendered. An unwrapped throw
82
+ * reaches the client as a generic internal error instead: no code for a caller
83
+ * to switch on, and nothing for the boundary to render from.
66
84
  */
67
85
  export const NO_ACTIVE_PROFILE = defineMessage(
68
86
  'hydranium/data-server/no-active-profile',
@@ -73,23 +91,54 @@ export const NO_ACTIVE_PROFILE = defineMessage(
73
91
  * The JSON-RPC code, unrelated to the catalogue code above: this one is numeric,
74
92
  * survives reconstruction and is what a caller switches on.
75
93
  */
76
- export const NO_ACTIVE_PROFILE_CODE = 1002;
94
+ export const NO_ACTIVE_PROFILE_CODE = HYDRANIUM_ERROR_CODES.noActiveProfile;
77
95
 
78
96
  export const noActiveProfileError = (): HydraniumResponseError => messageError(NO_ACTIVE_PROFILE_CODE, NO_ACTIVE_PROFILE);
79
97
 
98
+ /**
99
+ * Raised when a reference question's wait for the build exceeds
100
+ * {@link DataServerOptions.referenceSettleTimeoutMs}.
101
+ *
102
+ * A `ResponseError` because a class name does not cross RPC: without the code a
103
+ * wire client cannot tell a wedged build from a reference that resolves to
104
+ * nothing, which is the distinction the wait exists to keep.
105
+ */
106
+ export const REFERENCE_SETTLE_TIMEOUT = defineMessage(
107
+ 'hydranium/data-server/reference-settle-timeout',
108
+ 'The model is still being processed; try again in a moment.'
109
+ );
110
+
111
+ /** See {@link NO_ACTIVE_PROFILE_CODE} for why this is separate from the catalogue code. */
112
+ export const REFERENCE_SETTLE_TIMEOUT_CODE = HYDRANIUM_ERROR_CODES.referenceSettleTimeout;
113
+
114
+ /** The wait's length rides in `data` as `elapsedMs`, for the code that reacts rather than the reader. */
115
+ export const referenceSettleTimeoutError = (elapsedMs: number): ResponseError<HydraniumMessageData & { readonly elapsedMs: number }> =>
116
+ new ResponseError(REFERENCE_SETTLE_TIMEOUT_CODE, REFERENCE_SETTLE_TIMEOUT.format(), {
117
+ elapsedMs,
118
+ ...messageData(REFERENCE_SETTLE_TIMEOUT)
119
+ });
120
+
80
121
  import type { DataServerDiagnosticsProvider, DataServerProfileCapture } from './diagnostics-provider.js';
81
122
  import type {
82
- ClientTextDocumentChangeEvent,
123
+ ClientSession,
124
+ ClientSessionWriteArgs,
83
125
  HydraniumLanguageServices,
84
126
  LogNameOptions,
127
+ AstDiagnostic,
128
+ AstDocument,
129
+ AstDocumentSavedEvent,
130
+ EncodedTransferDocument,
131
+ ModelDeletedEvent,
132
+ ModelDirtyChangedEvent,
85
133
  ModelService,
134
+ ModelUpdatedEvent,
86
135
  ProjectChangeEvent,
87
136
  ServerSharedServices,
137
+ SessionEndCause,
88
138
  TransferEncoder
89
139
  } from '@hydranium/core';
90
- import type { TextDocument } from 'vscode-languageserver-textdocument';
91
- import { type AstNode, DocumentState, type LangiumDocument, UriUtils, type URI } from '@hydranium/langium';
92
- import type { CancellationToken, MessageConnection } from 'vscode-jsonrpc';
140
+ import { type AstNode, DocumentState, UriUtils, type URI } from '@hydranium/langium';
141
+ import { type MessageConnection, ResponseError } from 'vscode-jsonrpc';
93
142
 
94
143
  /**
95
144
  * Domain separator between text and diagnostics inputs of
@@ -100,34 +149,11 @@ import type { CancellationToken, MessageConnection } from 'vscode-jsonrpc';
100
149
  const FINGERPRINT_SEPARATOR = '\0';
101
150
 
102
151
  /**
103
- * Portable, non-cryptographic 64-bit fingerprint hash (cyrb53), returning a
104
- * 16-char hex digest.
105
- *
106
- * **A cryptographic hash via `node:crypto` is the wrong trade**: the
107
- * fingerprint only needs a stable signal that `(text, diagnostics)` changed
108
- * between emissions, and a `node:*` import would cost
109
- * `@hydranium/data-server` its browser-portability. Parts are fed
110
- * incrementally, char by char, so multi-MB document text is never
111
- * concatenated into one string.
152
+ * `LSPErrorCodes.ServerCancelled`, with which Langium's document wait rejects
153
+ * for a URI that has no document; `vscode-languageserver-protocol` is not a
154
+ * dependency of this package.
112
155
  */
113
- function fingerprintHash(parts: readonly string[]): string {
114
- let h1 = 0xdeadbeef;
115
- let h2 = 0x41c6ce57;
116
- for (const part of parts) {
117
- for (let i = 0; i < part.length; i++) {
118
- const ch = part.charCodeAt(i);
119
- h1 = Math.imul(h1 ^ ch, 2654435761);
120
- h2 = Math.imul(h2 ^ ch, 1597334677);
121
- }
122
- }
123
- h1 = Math.imul(h1 ^ (h1 >>> 16), 2246822507);
124
- h1 ^= Math.imul(h2 ^ (h2 >>> 13), 3266489909);
125
- h2 = Math.imul(h2 ^ (h2 >>> 16), 2246822507);
126
- h2 ^= Math.imul(h1 ^ (h1 >>> 13), 3266489909);
127
- const high = (h2 >>> 0).toString(16).padStart(8, '0');
128
- const low = (h1 >>> 0).toString(16).padStart(8, '0');
129
- return high + low;
130
- }
156
+ const SERVER_CANCELLED = -32802;
131
157
 
132
158
  /**
133
159
  * Which observable state {@link DataServer.computeDocumentFingerprint} hashes to
@@ -140,26 +166,32 @@ function fingerprintHash(parts: readonly string[]): string {
140
166
  * folded in on a cascade rebuild that leaves the document's own text
141
167
  * untouched. Safe for any adopter whose AST extensions fold derived state
142
168
  * into the transfer projection (the framework norm).
143
- * - `'text-diagnostics'` — the cheaper `getText()` + diagnostics hash. An
144
- * opt-DOWN for adopters that fold no derived state and want to avoid the
145
- * transfer-encode per phase event.
169
+ * - `'text-diagnostics'` — the cheaper parsed-text + diagnostics hash. An
170
+ * opt-DOWN for adopters that fold no derived state and want to avoid
171
+ * stringifying the transfer root per phase event.
146
172
  */
147
173
  export type FingerprintStrategy = 'transfer-document' | 'text-diagnostics';
148
174
 
149
- /** Hash a document's raw text + diagnostics — the `'text-diagnostics'` strategy. */
150
- function textDiagnosticsFingerprint(document: LangiumDocument): string {
151
- return fingerprintHash([document.textDocument.getText(), FINGERPRINT_SEPARATOR, JSON.stringify(document.diagnostics ?? [])]);
175
+ /**
176
+ * Hash `text`, the text the model was parsed from, + `model`'s diagnostics — the `'text-diagnostics'` strategy.
177
+ * The text comes from the root's ledger record or CST, as the document's own may have moved on,
178
+ * and a repair in place leaves the CST on the unrepaired text;
179
+ * a root built without a CST hashes as `'transfer-document'`.
180
+ * Absent diagnostics must hash apart from `[]`, as in {@link transferDocumentFingerprint}.
181
+ */
182
+ function textDiagnosticsFingerprint(text: string | undefined, model: EncodedTransferDocument<TransferElement, unknown>['model']): string {
183
+ return text === undefined
184
+ ? transferDocumentFingerprint(model)
185
+ : textHash([text, FINGERPRINT_SEPARATOR, JSON.stringify(model.diagnostics ?? null)]);
152
186
  }
153
187
 
154
- /** Hash an encoded transfer document's root + diagnostics — the `'transfer-document'` strategy. */
155
- function transferDocumentFingerprint(transferDocument: Pick<TransferDocument<TransferElement, unknown>, 'root' | 'diagnostics'>): string {
156
- // null, not undefined: `JSON.stringify(undefined)` yields undefined rather
157
- // than a string, putting a non-string into the hash inputs.
158
- return fingerprintHash([
159
- JSON.stringify(transferDocument.root ?? null),
160
- FINGERPRINT_SEPARATOR,
161
- JSON.stringify(transferDocument.diagnostics)
162
- ]);
188
+ /**
189
+ * Hash an encoded model's root + diagnostics — the `'transfer-document'` strategy.
190
+ * Absent diagnostics must hash apart from `[]`: a client skipping a render on an
191
+ * equal hash would otherwise miss a validation that found nothing.
192
+ */
193
+ function transferDocumentFingerprint(model: EncodedTransferDocument<TransferElement, unknown>['model']): string {
194
+ return textHash([JSON.stringify(model.root), FINGERPRINT_SEPARATOR, JSON.stringify(model.diagnostics ?? null)]);
163
195
  }
164
196
  /**
165
197
  * `logAfterMs` threshold passed to {@link Tracer.time} around
@@ -180,12 +212,13 @@ export interface DataServerOptions extends LogNameOptions {
180
212
  * (`DataClientProtocol.onDocumentUpdated`).
181
213
  *
182
214
  * **Subscription dispatch only** — distinct from the synchronous RPC
183
- * response phase. The lifecycle reads/writes (`getModelDocument` /
184
- * `updateModelDocument` / `saveModelDocument`) settle at the
185
- * integrity-settled landmark (`IntegrityService.SettledState`); a
186
- * read additionally upgrades to `Validated` per call when
187
- * {@link GetModelDocumentArgs.includeDiagnostics} is set. This option
188
- * controls only the async notification phase.
215
+ * response phase. A read (`getModelDocument`) settles at the
216
+ * integrity-settled landmark (`IntegrityService.SettledState`), or at
217
+ * `Validated` when {@link GetModelDocumentArgs.includeDiagnostics} is set.
218
+ * A write (`updateModelDocument` / `updateModelDocuments` /
219
+ * `saveModelDocument`) answers once its document is validated, as a
220
+ * session's write does. This option controls only the async notification
221
+ * phase.
189
222
  *
190
223
  * Default: `DocumentState.Validated` — validation is the last builder
191
224
  * phase and produces the full diagnostic set, so subscription events fired
@@ -226,6 +259,22 @@ export interface DataServerOptions extends LogNameOptions {
226
259
  */
227
260
  readonly fingerprintStrategy?: FingerprintStrategy;
228
261
 
262
+ /**
263
+ * How long a reference question waits for the build to reach `Linked` before
264
+ * failing with {@link REFERENCE_SETTLE_TIMEOUT_CODE}. Bounded because the
265
+ * wait resolves off a build-phase notification, so a build cancelled below
266
+ * `Linked` with nothing after it leaves the request unanswered for the life
267
+ * of the server.
268
+ *
269
+ * Expiry rejects rather than reading the index anyway: the unsettled index
270
+ * is what the wait keeps the caller away from, so falling through to it on a
271
+ * slow build reinstates the fault — and on a name proposal persists it.
272
+ *
273
+ * Default {@link DataServer.DEFAULT_OPTIONS}, sized for a cold workspace
274
+ * still reaching `Linked` for the first time.
275
+ */
276
+ readonly referenceSettleTimeoutMs?: number;
277
+
229
278
  /**
230
279
  * Wire-method namespace for both inbound request handlers and the
231
280
  * outbound notification client proxy. Defaults to
@@ -283,15 +332,20 @@ export interface DataServerOptions extends LogNameOptions {
283
332
  * When supplied, every inbound data-server RPC is timed into this collector
284
333
  * (via the `createRpcProxy` binding) and exposed through
285
334
  * {@link DataServerDiagnosticsProtocol.getLatency}. A head that also runs an
286
- * LSP connection can pass the SAME collector to
287
- * `instrumentLspConnection(connection, latency)` so one report covers both
288
- * heads. Absent by default (no timing overhead).
335
+ * LSP connection can pass the SAME collector through `lspLatencyOptions` to
336
+ * that connection, so one report covers both heads. Absent by default (no
337
+ * timing overhead).
289
338
  */
290
339
  readonly latency?: LatencyCollector;
291
340
  }
292
341
 
293
- /** Fully-resolved variant — every field set, used internally after merging defaults. */
294
- interface ResolvedDataServerOptions {
342
+ /**
343
+ * Fully-resolved variant of {@link DataServerOptions} — every field set, after
344
+ * merging defaults. Returned by the `protected`
345
+ * {@link DataServer.resolveOptions} and held on {@link DataServer.options}, so
346
+ * an adopter changing how defaults resolve has to name it.
347
+ */
348
+ export interface ResolvedDataServerOptions {
295
349
  readonly subscriptionPhase: DocumentState;
296
350
  readonly fingerprintStrategy: FingerprintStrategy;
297
351
  readonly methodNamespace: string;
@@ -299,6 +353,41 @@ interface ResolvedDataServerOptions {
299
353
  readonly excludedMethods: readonly string[];
300
354
  /** Always resolved — to the caller's, or to the platform default. */
301
355
  readonly diagnostics: DataServerDiagnosticsProvider;
356
+ readonly referenceSettleTimeoutMs: number;
357
+ }
358
+
359
+ /**
360
+ * What a {@link DataServer} keeps per canonical URI, in the `protected`
361
+ * {@link DataServer.uriWatchRecords}. A record lives while it has a watcher
362
+ * or a release mark; {@link DataServer.pruneUriWatchRecord} enforces that on
363
+ * every path that removes either, so a record with no watcher has a release
364
+ * broadcast pending. Test the size of {@link DataServerUriWatchRecord.watchers}, not the
365
+ * record's presence, for "watched".
366
+ */
367
+ export interface DataServerUriWatchRecord {
368
+ /** The client ids watching the URI; save, dirty and phase events go out only while it is non-empty, a pending release mark aside. */
369
+ readonly watchers: Set<string>;
370
+ /** Digest of the last emitted state, or of the state a first watch found; see {@link DataServer.dispatchPhaseEvent} for why it de-duplicates, {@link DataServer.computeDocumentFingerprint} for its inputs. */
371
+ fingerprint?: string;
372
+ /**
373
+ * The version {@link DataServerUriWatchRecord.fingerprint} was taken at; an
374
+ * event goes out when either moves. Not {@link DataServerUriWatchRecord.sentVersion}:
375
+ * a first watch sets this and sends nothing, so the next event at its
376
+ * version still carries the manager's attribution.
377
+ */
378
+ fingerprintVersion?: ModelVersion;
379
+ /** Set by {@link DataServer.subscribeToModelReleases} once a release happened: its broadcast, the next phase event, goes out even without a watcher. */
380
+ releaseBroadcastPending?: boolean;
381
+ /**
382
+ * The version of an open document the last event sent was built at. The
383
+ * version sent again goes out as `'rebuilt'` from no client: the manager
384
+ * counts a version delivered only once its `Validated` listeners ran, so a
385
+ * build cancelled after the subscription phase would send it as a change
386
+ * twice. Only while the store has the document open: without it, a
387
+ * document can keep one version through changes of its file, so its events
388
+ * keep the manager's attribution.
389
+ */
390
+ sentVersion?: number;
302
391
  }
303
392
 
304
393
  /**
@@ -322,13 +411,19 @@ interface ResolvedDataServerOptions {
322
411
  * adopters subclass ONLY when they need to decorate wire returns or
323
412
  * notifications. The usual adoption path is DI rebinds alone.
324
413
  *
414
+ * **`TTransfer` is not checked.** It declares the union of the transfer roots
415
+ * the workspace's languages produce. Which root a document has is known only
416
+ * at runtime, from its language, so a `TTransfer` missing a language compiles.
417
+ *
325
418
  * Lifecycle: the constructor registers framework request handlers (plus
326
419
  * any names supplied via {@link DataServerOptions.additionalMethods}) and
327
420
  * builds the {@link DataClientProtocol} notification proxy on the same
328
421
  * connection in one {@link createRpcProxy} call (binding `this` as its
329
- * `localTarget`) under the configured namespace, and subscribes one
330
- * `DocumentBuilder.onDocumentPhase` listener per configured phase that
331
- * dispatches subscription events via `clientProxy.onDocumentUpdated`.
422
+ * `localTarget`) under the configured namespace, and follows documents
423
+ * through `ModelService`'s events: at the configured phase, `onModelUpdated`
424
+ * dispatches per-document `clientProxy.onDocumentUpdated` for watched URIs and
425
+ * `onModelsBuilt` a single `clientProxy.onDocumentsBuilt` naming the batch's
426
+ * unwatched documents.
332
427
  *
333
428
  * The data-server requires the full {@link ServerSharedServices}
334
429
  * shape — `HydraniumTextDocuments` for client-attributed updates,
@@ -356,87 +451,30 @@ export class DataServer<
356
451
  * Framework defaults, exposed so an adopter can spread them when
357
452
  * extending rather than restate a value the framework may change.
358
453
  */
359
- static readonly DEFAULT_OPTIONS: Required<Pick<DataServerOptions, 'subscriptionPhase'>> = {
360
- subscriptionPhase: DocumentState.Validated
454
+ static readonly DEFAULT_OPTIONS: Required<Pick<DataServerOptions, 'subscriptionPhase' | 'referenceSettleTimeoutMs'>> = {
455
+ subscriptionPhase: DocumentState.Validated,
456
+ referenceSettleTimeoutMs: 60_000
361
457
  };
362
458
 
363
459
  protected readonly options: ResolvedDataServerOptions;
364
- /** Subscription bookkeeping: URI → set of clientIds that subscribed for that URI. */
365
- protected readonly subscriptions = new Map<string, Set<string>>();
366
- /**
367
- * Open-document bookkeeping: URI → set of clientIds that opened it over THIS
368
- * connection.
369
- *
370
- * Kept because the document store releases a client's hold only from an
371
- * explicit close, and a client that dies without closing would otherwise keep
372
- * the document open forever — resident, and with its last-close revert
373
- * suppressed. The store has no notion of which connection a clientId reached
374
- * it over, so the connection-scoped set has to live here; teaching it a client
375
- * identity is the more robust and much larger alternative.
376
- */
377
- protected readonly openedDocuments = new Map<string, Set<string>>();
378
- /** Typed client proxy — sends `data-server/on*` notifications back over the same wire. */
379
- protected readonly clientProxy: DataClientProtocol<TTransfer, TDiagnostic, TProject>;
380
- protected readonly disposables = new DisposableCollection();
460
+ /** This connection's {@link DataServerUriWatchRecord} per canonical URI; see there for when a record lives. */
461
+ protected readonly uriWatchRecords = new Map<string, DataServerUriWatchRecord>();
381
462
  /**
382
- * Snapshot of the most recent `DocumentBuilder.onUpdate` event. Drives
383
- * `reason` discrimination on outbound `onDocumentUpdated` notifications:
384
- * a URI in the `changed` list emits `'changed'`, in `deleted` emits
385
- * `'deleted'`, otherwise `'rebuilt'` (cascade rebuild from a dependent
386
- * URI's change). The same mechanism `AstDocumentManager.onUpdate` uses on
387
- * the LSP side, so reason fidelity stays consistent between heads.
463
+ * The live client sessions this connection registered, by client id. A
464
+ * document request carrying one of these ids acts as that session; one
465
+ * carrying any other id fails. A session leaves once it ends, so a
466
+ * connection that registers many over its life does not accumulate them.
388
467
  */
389
- protected lastBuildUpdate?: { changed: readonly URI[]; deleted: readonly URI[] };
468
+ protected readonly clientSessions = new Map<string, ClientSession<AstNode, AstDiagnostic, TTransfer>>();
390
469
  /**
391
- * Per-URI hash of the last emitted (text + diagnostics) state,
392
- * used by {@link dispatchPhaseEvent} to suppress duplicate
393
- * `onDocumentUpdated` notifications for rebuilds that produce no
394
- * observable change since the last emit.
395
- *
396
- * Two scenarios routinely produce such rebuilds:
397
- * 1. **Refresh-triggered rebuilds.** When a second client attaches to a
398
- * document already open in another client,
399
- * `HydraniumTextDocuments.refreshContent` fires `onDidChangeContent`
400
- * purely to re-trigger the build pipeline — the text and diagnostics
401
- * are identical to the prior emit.
402
- * 2. **Cascade rebuilds with no diagnostic delta.** Langium rebuilds a
403
- * document when a dependency changes; if the rebuild produces the
404
- * same diagnostics, there is nothing new to communicate to subscribers.
405
- *
406
- * Without this filter, both scenarios reach subscribers as wire-side
407
- * `'changed'` events. A subscriber that interprets `'changed'` as
408
- * "another client wrote new content" and resets its in-memory root to
409
- * the server view then misreads the spurious event as a concurrent
410
- * third-party write and loses the user's edits.
411
- *
412
- * Initialised on {@link watchModelDocument} to the current
413
- * document fingerprint so the FIRST phase event after a fresh
414
- * subscription is also de-duplicated against the state subscribers
415
- * obtained via `getModelDocument` — they do not need a redundant phase
416
- * notification for the state they just fetched.
417
- *
418
- * Cleared in {@link unwatchModelDocument} when the last
419
- * subscriber for a URI leaves so memory does not accumulate.
420
- *
421
- * Stored as a cyrb53 hex digest (16 chars per URI) rather than the raw
422
- * text + diagnostics serialisation so the memory cost is constant in
423
- * document size. See {@link computeDocumentFingerprint} for the inputs.
424
- */
425
- protected readonly lastEmittedFingerprint = new Map<string, string>();
426
- /**
427
- * URIs whose LAST client just closed. The update handler rebuilds such a
428
- * document from its disk content (discarding unsaved in-session edits),
429
- * but {@link dispatchPhaseEvent} gates on the subscription map — and the
430
- * last close typically also removed the last watcher, so consumers that
431
- * only ever fetch via `getModelDocument` would keep showing the
432
- * discarded state forever. Marked on the last-close transition (see
433
- * {@link subscribeToTextDocumentCloses}) and consumed by
434
- * {@link dispatchPhaseEvent}, which broadcasts the following rebuild's
435
- * phase event even without a subscription — de-duplicated against
436
- * {@link lastEmittedFingerprint} where an entry survives, so a close
437
- * whose disk state equals the last emitted state stays silent.
470
+ * Typed client proxy — sends `data-server/on*` notifications back over the same wire.
471
+ * The proxy sends any `on[A-Z]*` name other than `onDidOpenConnection` and
472
+ * `onDidCloseConnection`, for which it returns its own events, so a subclass
473
+ * with notifications of its own narrows only the type:
474
+ * `declare protected readonly clientProxy: TheirClient`.
438
475
  */
439
- protected readonly pendingRevertBroadcasts = new Set<string>();
476
+ protected readonly clientProxy: DataClientProtocol<TTransfer, TDiagnostic, TProject>;
477
+ protected readonly disposables = new DisposableCollection();
440
478
  protected readonly tracer: Tracer;
441
479
  /** The in-flight interactive profile capture, held between {@link startProfiling} and {@link stopProfiling}. */
442
480
  protected activeProfile?: DataServerProfileCapture;
@@ -445,27 +483,23 @@ export class DataServer<
445
483
  /** Per-method RPC latency collector, when the head opted in via {@link DataServerOptions.latency}. */
446
484
  protected readonly latency?: LatencyCollector;
447
485
  /** Encoder pulled from DI. Adopters rebind `services.model.TransferEncoder` to a typed-overlay subclass. */
448
- protected readonly encoder: TransferEncoder<unknown, TDiagnostic>;
486
+ protected readonly encoder: TransferEncoder<TDiagnostic>;
449
487
  /**
450
488
  * In-process workspace facade — the lifecycle delegate `get` / `update` / `save` go through.
451
- * The facade's AstDocument diagnostic shape is intentionally typed `unknown` here: adopters
452
- * carry LSP-shape diagnostics in their AstDocument (e.g. an adopter's LSP-shape diagnostic type)
453
- * while the wire shape stays `TDiagnostic extends TransferDiagnostic`. The encoder's
454
- * `astDocumentToTransferDocument` accepts both shapes (wire-shape or LSP-shape) and projects
455
- * to wire shape on the return — see `TransferEncoder.astDocumentToTransferDocument`.
489
+ * Its documents carry the AST-layer {@link AstDiagnostic}, never the wire `TDiagnostic`:
490
+ * {@link encoder} converts one into the other on the way out.
456
491
  */
457
- protected readonly modelService: ModelService<AstNode, unknown, TTransfer>;
492
+ protected readonly modelService: ModelService<AstNode, AstDiagnostic, TTransfer>;
458
493
 
459
494
  constructor(
460
495
  protected readonly connection: MessageConnection,
461
- protected readonly services: ServerSharedServices<TProject>,
496
+ protected readonly services: ServerSharedServices<TProject, TDiagnostic>,
462
497
  options: DataServerOptions = {}
463
498
  ) {
464
499
  this.options = this.resolveOptions(options);
465
500
  this.tracer = this.services.Tracer.for(options.logName ?? 'DataServer').trace('instantiated');
466
501
  // DI-bound: adopters rebind `services.model.TransferEncoder` /
467
- // `services.model.ModelService` with their own subclasses. Slot types use the
468
- // framework upper bounds; the casts below narrow to this instance's generics.
502
+ // `services.model.ModelService` with their own subclasses.
469
503
  const { TransferEncoder: encoder, ModelService: modelService } = this.services.model;
470
504
  if (!encoder || !modelService) {
471
505
  throw new Error(
@@ -473,8 +507,8 @@ export class DataServer<
473
507
  'Did you compose `createServerSharedModule(ctx)` into your shared module?'
474
508
  );
475
509
  }
476
- this.encoder = encoder as TransferEncoder<unknown, TDiagnostic>;
477
- this.modelService = modelService as ModelService<AstNode, unknown, TTransfer>;
510
+ this.encoder = encoder;
511
+ this.modelService = modelService;
478
512
  const excluded = new Set<string>(this.options.excludedMethods);
479
513
  const registeredMethods = [
480
514
  ...DATA_SERVER_PROTOCOL_METHODS,
@@ -490,38 +524,43 @@ export class DataServer<
490
524
  methodNamespace: this.options.methodNamespace,
491
525
  localTarget: this,
492
526
  localMethods: registeredMethods as readonly (keyof this & string)[],
493
- latency: this.latency
527
+ latency: this.latency,
528
+ // Bound at the chokepoint rather than per handler, so an adopter's
529
+ // `additionalMethods` are rendered on the same terms as the
530
+ // framework's own rejections.
531
+ renderErrorMessage: error => this.services.MessageRenderer.renderError(error)
494
532
  });
495
- this.disposables.push(
496
- this.services.workspace.DocumentBuilder.onUpdate((changed, deleted) => {
497
- this.lastBuildUpdate = { changed, deleted };
498
- })
499
- );
500
- this.subscribeToDocumentBuilder();
533
+ this.disposables.push(this.modelService.onModelDeleted(event => this.dispatchDeleteEvent(event)));
534
+ this.subscribeToModelUpdates();
501
535
  this.subscribeToTextDocumentSaves();
502
- this.subscribeToTextDocumentCloses();
536
+ this.subscribeToDirtyChanges();
537
+ this.subscribeToModelReleases();
503
538
  this.subscribeToProjectManager();
504
539
  // Self-register teardown so an adopter that keeps no reference to the
505
540
  // server still releases per-connection state, with no lifecycle hook of
506
- // its own. See `dispose`.
507
- this.disposables.push(connection.onClose(() => this.dispose()));
541
+ // its own. The client did not close anything itself, so its sessions end
542
+ // as lost and each document it was the last to have open waits out the
543
+ // store's release grace. See `dispose`.
544
+ this.disposables.push(connection.onClose(() => this.dispose('lost')));
508
545
  }
509
546
 
510
547
  /**
511
548
  * Release everything the constructor wired up — the bound protocol
512
- * handlers and every listener — and clear the subscription map and the
513
- * per-URI emission-fingerprint cache, so a long-lived shared services
514
- * bundle does not retain per-connection memory after the connection
515
- * closes. Also closes every document still open over this connection (see
516
- * {@link closeOpenDocuments}), which is the SHARED store's state rather than
517
- * this server's and so outlives the connection unless released here.
549
+ * handlers and every listener — and clear {@link uriWatchRecords}, so a
550
+ * long-lived shared services bundle does not retain per-connection memory
551
+ * after the connection closes. Also ends every session this connection registered, which closes
552
+ * every document it has open: the SHARED store's state rather than this
553
+ * server's, which outlives the connection unless released here.
518
554
  *
519
- * Idempotent: subsequent calls are no-ops. Self-fires on
520
- * `connection.onClose` so adopters who don't hold a reference still
521
- * get per-connection cleanup; adopters that DO hold a reference may
522
- * call `dispose()` directly for early teardown.
555
+ * Idempotent: subsequent calls are no-ops. Runs by itself when the
556
+ * connection closes, so adopters who don't hold a reference still get
557
+ * per-connection cleanup; adopters that DO hold a reference may call
558
+ * `dispose()` directly for early teardown. The sessions end with `cause`:
559
+ * `'closed'` releases every document whose last open they close at once,
560
+ * and `'lost'`, the connection's own close, lets each wait out the release
561
+ * grace.
523
562
  */
524
- dispose(): void {
563
+ dispose(cause: SessionEndCause = 'closed'): void {
525
564
  this.disposed = true;
526
565
  // Release an in-flight interactive capture so the process-wide inspector
527
566
  // singleton is not left active after the connection closes.
@@ -530,104 +569,148 @@ export class DataServer<
530
569
  this.activeProfile = undefined;
531
570
  void capture.stop({}).catch(() => undefined);
532
571
  }
533
- // AFTER `disposables.dispose()`, deliberately: closing a document fires
534
- // `onDidClose`, and this server's own close listener would otherwise mark a
535
- // revert broadcast for a connection that is already gone.
572
+ // AFTER `disposables.dispose()`, deliberately: ending a session releases
573
+ // each document it was the last to hold, and this server's own
574
+ // `onModelReleased` subscription would otherwise mark a release broadcast
575
+ // for a connection that is already gone.
536
576
  this.disposables.dispose();
537
- this.closeOpenDocuments();
538
- this.subscriptions.clear();
539
- this.lastEmittedFingerprint.clear();
540
- this.pendingRevertBroadcasts.clear();
577
+ for (const [clientId, session] of this.clientSessions) {
578
+ this.endSession(clientId, session, cause);
579
+ }
580
+ this.uriWatchRecords.clear();
541
581
  }
542
582
 
543
583
  // ============================================================
544
- // DataServerProtocol implementation — delegates to ModelService.
584
+ // DataServerProtocol implementation
545
585
  // ============================================================
546
586
  //
547
- // The lifecycle methods (`get` / `update` / `save` / `ready`) forward
548
- // straight to `ModelService` and encode the returned `AstDocument` on
549
- // the wire boundary via `encoder.astDocumentToTransferDocument`.
550
- // Adopters customising lifecycle behaviour (normalisation, supersession,
551
- // settled-phase choice, etc.) override on the ModelService subclass —
552
- // no override on DataServer is needed.
553
-
554
- async openModelDocument(args: OpenModelArgs): Promise<TransferDocument<TTransfer, TDiagnostic>> {
555
- // Register the editor session (idempotent — `ModelService.open` refreshes
556
- // an already-open document rather than re-opening), then return the built
557
- // state at the configured target phase. The one-shot snapshot; subsequent
558
- // build-phase events arrive via `watchModelDocument`.
559
- await this.modelService.open(args);
560
- // Record the hold so `dispose` can release it for a client that never
561
- // closes. Keyed by the URI as given, because that is what `close` takes.
562
- let holders = this.openedDocuments.get(args.uri);
563
- if (!holders) {
564
- holders = new Set<string>();
565
- this.openedDocuments.set(args.uri, holders);
587
+ // Reads go to `ModelService`; a request that opens or writes a document
588
+ // goes to the session this connection registered under its `clientId`.
589
+ // Each answer is encoded on the wire boundary via
590
+ // `encoder.astDocumentToTransferDocument`. To change how documents are
591
+ // read, override on the ModelService subclass; to change how they are
592
+ // written, on the session class the bound `ClientSessionFactory` builds.
593
+
594
+ async createSession(args: CreateSessionArgs): Promise<void> {
595
+ // After teardown nothing would ever end a session registered now, and its
596
+ // id would stay taken for the life of the process.
597
+ if (this.disposed) {
598
+ throw new SessionClosedError(args.clientId);
566
599
  }
567
- holders.add(args.clientId);
568
- const document = await this.getModelDocument({ uri: args.uri });
569
- // Project the authoritative client-facing version onto the open snapshot.
570
- // `getModelDocument` encodes the freshly-built AST snapshot, whose version
571
- // is the `LangiumDocument`'s own `textDocument.version`. That lags the
572
- // multi-client synced version whenever the open seeded the synced document
573
- // with a caller-supplied version id — and the synced version is what
574
- // `ModelService.update`'s optimistic-concurrency gate compares
575
- // `baseVersion` against, so reporting the snapshot's would make the
576
- // caller's first tagged write self-conflict. A genuine concurrent edit
577
- // still trips the gate and is reconciled by the caller's replay rather
578
- // than predicted here.
579
- return { ...document, version: this.services.workspace.TextDocuments.version(args.uri) };
600
+ const session = this.modelService.createSession(args.label, args.clientId, { resumeToken: args.resumeToken });
601
+ this.clientSessions.set(args.clientId, session);
602
+ // A session also ends without closeSession, as when a client registers
603
+ // again on another connection, and its watches and entry here would
604
+ // otherwise outlive it.
605
+ session.onDidDispose(cause => {
606
+ // A session class may report its end late, once a replacement holds
607
+ // the id here; the replacement's entry and watches stay.
608
+ if (this.clientSessions.get(args.clientId) !== session) {
609
+ return;
610
+ }
611
+ this.dropWatches(args.clientId);
612
+ this.clientSessions.delete(args.clientId);
613
+ this.tracer.debug(`Forget session ${args.clientId}: it ended as ${cause}`);
614
+ });
580
615
  }
581
616
 
582
- async closeModelDocument(args: CloseModelArgs): Promise<void> {
583
- // Closing a session also releases its watch for (uri, clientId) — a
584
- // forgotten unwatch would otherwise leak phase-event dispatch until the
585
- // connection closes. Idempotent: a close without a prior watch is a no-op.
586
- await this.unwatchModelDocument({ uri: args.uri, clientId: args.clientId });
587
- this.forgetOpenDocument(args.uri, args.clientId);
588
- await this.modelService.close(args);
617
+ /** End a session this connection registered. */
618
+ async closeSession(args: CloseSessionArgs): Promise<void> {
619
+ const session = this.clientSessions.get(args.clientId);
620
+ if (session) {
621
+ this.endSession(args.clientId, session);
622
+ }
623
+ }
624
+
625
+ /** End `session`, registered here under `clientId`, with `cause`, and drop its watches on this connection. */
626
+ protected endSession(
627
+ clientId: string,
628
+ session: ClientSession<AstNode, AstDiagnostic, TTransfer>,
629
+ cause: SessionEndCause = 'closed'
630
+ ): void {
631
+ this.dropWatches(clientId);
632
+ session.dispose(cause);
633
+ }
634
+
635
+ protected dropWatches(clientId: string): void {
636
+ for (const [uri, record] of this.uriWatchRecords) {
637
+ if (record.watchers.delete(clientId)) {
638
+ this.pruneUriWatchRecord(uri);
639
+ }
640
+ }
641
+ }
642
+
643
+ async createModelDocument(args: CreateModelDocumentArgs): Promise<TransferDocument<TTransfer, TDiagnostic>> {
644
+ const session = this.requireSession(args.clientId);
645
+ await session.create(args.uri, args.text);
646
+ return this.openedSnapshot(args.uri, () => session.close(args.uri));
647
+ }
648
+
649
+ async updateModelDocuments(args: TransferUpdateDocumentsArgs<TTransfer>): Promise<TransferDocument<TTransfer, TDiagnostic>[]> {
650
+ const astDocuments = await this.requireSession(args.clientId).updateAll({
651
+ updates: args.updates.map(update => this.toSessionWrite(update))
652
+ });
653
+ return astDocuments.map(astDocument => this.encodeDocument(astDocument));
589
654
  }
590
655
 
591
656
  /**
592
- * Drop the recorded hold for `(uri, clientId)` so {@link dispose} does not
593
- * close it a second time. Idempotent.
657
+ * The session write a wire write request carries. Only the fields a session
658
+ * takes: a field an adopter adds to a request reaches the session only
659
+ * through an override of this.
594
660
  */
595
- protected forgetOpenDocument(uri: string, clientId: string): void {
596
- const holders = this.openedDocuments.get(uri);
597
- if (!holders) {
598
- return;
599
- }
600
- holders.delete(clientId);
601
- if (holders.size === 0) {
602
- this.openedDocuments.delete(uri);
661
+ protected toSessionWrite(request: Omit<TransferUpdateDocumentArgs<TTransfer>, 'clientId'>): ClientSessionWriteArgs<TTransfer> {
662
+ return { uri: request.uri, model: request.model, baseVersion: request.baseVersion };
663
+ }
664
+
665
+ /**
666
+ * The session this connection registered under `clientId`, which every
667
+ * document request acts as. An id that is not one fails with the
668
+ * closed-session code: opening or writing under it would leave an open that
669
+ * no session end ever closes.
670
+ */
671
+ protected requireSession(clientId: string): ClientSession<AstNode, AstDiagnostic, TTransfer> {
672
+ const session = this.clientSessions.get(clientId);
673
+ if (!session) {
674
+ throw new SessionClosedError(clientId, 'The client session is not registered on this connection.');
603
675
  }
676
+ return session;
677
+ }
678
+
679
+ async openModelDocument(args: Pick<OpenModelArgs, 'uri' | 'clientId' | 'options'>): Promise<TransferDocument<TTransfer, TDiagnostic>> {
680
+ const session = this.requireSession(args.clientId);
681
+ const wasOpen = this.services.workspace.TextDocuments.isOpenInClient(args.uri, args.clientId);
682
+ await session.open(args.uri, args.options);
683
+ // A repeat open keeps the earlier one, which is still in use.
684
+ return this.openedSnapshot(args.uri, wasOpen ? undefined : () => session.close(args.uri));
604
685
  }
605
686
 
606
687
  /**
607
- * Close every document still open over this connection, for the clients that
608
- * opened it here.
609
- *
610
- * The document store releases a per-URI hold only from an explicit close, so
611
- * without this a client that dies mid-session keeps its documents open for the
612
- * lifetime of the process: `isOpenInAnyClient` stays true, the document stays
613
- * resident, and the last-close revert never runs. A long-lived multi-client
614
- * head is the configuration where a dead client is normal rather than
615
- * exceptional, so the leak accumulates there.
616
- *
617
- * Runs from {@link dispose}, which is synchronous, so each close is fired and
618
- * its failure swallowed — a teardown must not reject, and a URI whose close
619
- * fails is no worse off than it was before this drain existed.
688
+ * The snapshot a session's open or create answers with. When the read
689
+ * fails, `rollback` undoes the open this call made before the failure is
690
+ * rethrown: the caller sees no open, so nothing on its side would ever close
691
+ * it.
620
692
  */
621
- protected closeOpenDocuments(): void {
622
- const held = [...this.openedDocuments];
623
- this.openedDocuments.clear();
624
- for (const [uri, clientIds] of held) {
625
- for (const clientId of clientIds) {
626
- void Promise.resolve(this.modelService.close({ uri, clientId })).catch(() => undefined);
627
- }
693
+ protected async openedSnapshot(uri: string, rollback?: () => Promise<void>): Promise<TransferDocument<TTransfer, TDiagnostic>> {
694
+ try {
695
+ return await this.getModelDocument({ uri });
696
+ } catch (error: unknown) {
697
+ // The open's failure is the one the caller needs, not the cleanup's.
698
+ await rollback?.().catch(() => undefined);
699
+ throw error;
628
700
  }
629
701
  }
630
702
 
703
+ async closeModelDocument(args: CloseModelArgs): Promise<void> {
704
+ // Through the handle, so a close under an ended session fails as that
705
+ // session's calls do, and a session class's own close is honoured.
706
+ const session = this.requireSession(args.clientId);
707
+ // Closing also releases the watch for (uri, clientId) — a forgotten
708
+ // unwatch would otherwise leak phase-event dispatch until the connection
709
+ // closes. Idempotent: a close without a prior watch is a no-op.
710
+ await this.unwatchModelDocument({ uri: args.uri, clientId: args.clientId });
711
+ await session.close(args.uri);
712
+ }
713
+
631
714
  async getModelDocument(args: GetModelDocumentArgs): Promise<TransferDocument<TTransfer, TDiagnostic>> {
632
715
  // Smart dispatch: a warm document (already in LangiumDocuments) is
633
716
  // returned at its settle phase without a redundant build; a cold URI
@@ -640,18 +723,32 @@ export class DataServer<
640
723
  // unsubscribed caller that needs diagnostics inline passes
641
724
  // `includeDiagnostics: true` to settle at `Validated` instead.
642
725
  const state = args.includeDiagnostics ? DocumentState.Validated : undefined;
643
- const astDocument = await this.modelService.ensureDocumentState(args.uri, state);
644
- return this.encoder.astDocumentToTransferDocument(astDocument as never) as unknown as TransferDocument<TTransfer, TDiagnostic>;
726
+ try {
727
+ return this.encodeDocument(await this.modelService.ensureDocumentState(args.uri, state));
728
+ } catch (error: unknown) {
729
+ // The protocol answers a read of a URI with no document with an
730
+ // envelope that has no model.
731
+ if (this.isMissingDocument(error, args.uri)) {
732
+ return this.envelope(UriUtils.toUri(args.uri));
733
+ }
734
+ throw error;
735
+ }
645
736
  }
646
737
 
647
738
  async updateModelDocument(args: TransferUpdateDocumentArgs<TTransfer>): Promise<TransferDocument<TTransfer, TDiagnostic>> {
648
- const astDocument = await this.modelService.update(args);
649
- return this.encoder.astDocumentToTransferDocument(astDocument as never) as unknown as TransferDocument<TTransfer, TDiagnostic>;
739
+ const astDocument = await this.requireSession(args.clientId).update(this.toSessionWrite(args));
740
+ return this.encodeDocument(astDocument);
650
741
  }
651
742
 
652
- async saveModelDocument(args: TransferSaveDocumentArgs<TTransfer>): Promise<TransferDocument<TTransfer, TDiagnostic>> {
653
- const astDocument = await this.modelService.save(args);
654
- return this.encoder.astDocumentToTransferDocument(astDocument as never) as unknown as TransferDocument<TTransfer, TDiagnostic>;
743
+ async saveModelDocument(args: TransferSaveDocumentArgs<TTransfer>): Promise<TransferSavedDocument<TTransfer, TDiagnostic>> {
744
+ const astDocument = await this.requireSession(args.clientId).save(this.toSessionWrite(args));
745
+ return { ...this.encodeDocument(astDocument), persisted: astDocument.persisted };
746
+ }
747
+
748
+ async persistModelDocument(args: TransferPersistDocumentArgs): Promise<TransferSavedDocument<TTransfer, TDiagnostic>> {
749
+ const version = await this.requireSession(args.clientId).persist({ uri: args.uri, baseVersion: args.baseVersion });
750
+ const astDocument = await this.modelService.ensureDocumentState(args.uri, this.services.workspace.DocumentBuilder.finalBuildState());
751
+ return { ...this.encodeDocument(astDocument), persisted: { version } };
655
752
  }
656
753
 
657
754
  /**
@@ -659,28 +756,25 @@ export class DataServer<
659
756
  * fan out to the wired `clientProxy.onDocumentUpdated`. The
660
757
  * bidirectional pattern means the event channel is the client
661
758
  * notification surface, not a returned handle — this method only
662
- * registers the URI in the dispatch table.
759
+ * registers the watcher in {@link uriWatchRecords}.
663
760
  *
664
- * Also baselines the per-URI emission fingerprint (see
665
- * {@link lastEmittedFingerprint}) to the document's current state when
666
- * the first watcher for the URI registers. This guarantees that any
667
- * phase event firing immediately after the watch with no observable
668
- * change is suppressed — watchers obtain initial state via
669
- * `getModelDocument` (or {@link openModelDocument}) and do not need a
670
- * redundant phase notification for that same state.
761
+ * Also baselines the URI's emission fingerprint (see
762
+ * {@link dispatchPhaseEvent}) to the document's current state when the
763
+ * first watcher for the URI registers. This guarantees that any phase event firing immediately after
764
+ * the watch with no observable change is suppressed — watchers obtain
765
+ * initial state via `getModelDocument` (or {@link openModelDocument}) and
766
+ * do not need a redundant phase notification for that same state.
671
767
  */
672
768
  async watchModelDocument(args: WatchModelDocumentArgs): Promise<void> {
673
769
  const uri = this.canonicalKey(args.uri);
674
- let subscribers = this.subscriptions.get(uri);
675
- if (!subscribers) {
676
- subscribers = new Set();
677
- this.subscriptions.set(uri, subscribers);
678
- }
679
- subscribers.add(args.clientId);
680
- if (!this.lastEmittedFingerprint.has(uri)) {
681
- const document = this.services.workspace.LangiumDocuments.getDocument(UriUtils.toUri(uri));
682
- if (document) {
683
- this.lastEmittedFingerprint.set(uri, this.computeDocumentFingerprint(document));
770
+ const record = this.uriWatchRecords.get(uri) ?? { watchers: new Set<string>() };
771
+ this.uriWatchRecords.set(uri, record);
772
+ record.watchers.add(args.clientId);
773
+ if (record.fingerprint === undefined) {
774
+ const document = this.modelService.getDocument(args.uri);
775
+ if (document && !this.services.workspace.ModelLedger.isPlaceholder(document.parseResult.value)) {
776
+ record.fingerprint = this.computeDocumentFingerprint(document.parseResult.value, this.encoder.toTransferDocument(document));
777
+ record.fingerprintVersion = this.services.workspace.ModelLedger.versionOf(document.parseResult.value);
684
778
  }
685
779
  }
686
780
  }
@@ -688,29 +782,41 @@ export class DataServer<
688
782
  /**
689
783
  * Remove a watch for `(uri, clientId)` previously created by
690
784
  * {@link watchModelDocument}. Idempotent — unwatching twice is a no-op.
691
- * Dispatch for `uri` stops once no watchers remain, at which point the
692
- * per-URI emission fingerprint is also cleared so the next first-watch
693
- * re-baselines against the then-current document state rather than a
694
- * stale snapshot from the previous watch.
785
+ * Dispatch for `uri` stops once no watchers remain, and the emission
786
+ * fingerprint goes with the last watcher (see {@link pruneUriWatchRecord}).
695
787
  */
696
788
  async unwatchModelDocument(args: WatchModelDocumentArgs): Promise<void> {
697
789
  const uri = this.canonicalKey(args.uri);
698
- const subscribers = this.subscriptions.get(uri);
699
- if (!subscribers) {
790
+ if (this.uriWatchRecords.get(uri)?.watchers.delete(args.clientId)) {
791
+ this.pruneUriWatchRecord(uri);
792
+ }
793
+ }
794
+
795
+ /**
796
+ * Drop what `uri` no longer needs: the
797
+ * {@link DataServerUriWatchRecord.fingerprint} once no client watches it,
798
+ * and the whole record once it holds no
799
+ * {@link DataServerUriWatchRecord.releaseBroadcastPending} mark either.
800
+ * Every path that removes a watcher or consumes a mark ends here.
801
+ */
802
+ protected pruneUriWatchRecord(uri: string): void {
803
+ const record = this.uriWatchRecords.get(uri);
804
+ if (!record || record.watchers.size > 0) {
700
805
  return;
701
806
  }
702
- subscribers.delete(args.clientId);
703
- if (subscribers.size === 0) {
704
- this.subscriptions.delete(uri);
705
- this.lastEmittedFingerprint.delete(uri);
807
+ if (record.releaseBroadcastPending) {
808
+ record.fingerprint = undefined;
809
+ record.fingerprintVersion = undefined;
810
+ } else {
811
+ this.uriWatchRecords.delete(uri);
706
812
  }
707
813
  }
708
814
 
709
815
  /**
710
- * Canonicalise a URI string for use as a subscription / fingerprint
711
- * map key, via the shared `DocumentUriPolicy`. Callers may send
712
- * non-canonical URIs (drive-letter casing, percent-encoding differences,
713
- * or a symlink path) over the wire; the dispatch side keys by
816
+ * Canonicalise a URI string for use as a {@link uriWatchRecords} key, via the
817
+ * shared `DocumentUriPolicy`. Callers may send non-canonical URIs
818
+ * (drive-letter casing, percent-encoding differences, or a symlink path)
819
+ * over the wire; the dispatch side keys by
714
820
  * `document.uri.toString()` from `LangiumDocuments`, so writer keys must
715
821
  * canonicalise to the same form or events silently fail to deliver.
716
822
  * Routing through the seam (rather than a bare `UriUtils.normalize`)
@@ -720,6 +826,7 @@ export class DataServer<
720
826
  * LSP head resolves.
721
827
  *
722
828
  * Adopters can still override for head-specific URI policy by subclassing.
829
+ * The result keys watches only; a document is looked up by its own URI.
723
830
  */
724
831
  protected canonicalKey(uri: string): string {
725
832
  return this.services.workspace.DocumentUriPolicy.canonicalUri(uri);
@@ -759,27 +866,60 @@ export class DataServer<
759
866
  // (see resolveReferenceServices) and delegate.
760
867
  // ============================================================
761
868
 
762
- async findReferenceCandidates(ctx: ReferenceContext): Promise<ReferenceCandidate[]> {
763
- // Candidates are derived from resolved cross-references, so the relevant document(s) must
764
- // have finished the Linked phase — otherwise a query issued right after a model update races
765
- // the asynchronous rebuild and returns stale candidates. Wait the source document
766
- // specifically, but ONLY when a document is actually loaded at that URI: a synthetic source
767
- // can address a URI with no document (a directory URI, typically), and a per-URI `waitUntil`
768
- // there throws "No document found". Fall back to a global Linked settle, matching the
769
- // id-based ElementSource (no URI) path.
770
- const uri = isDocumentSource(ctx.source) || isSyntheticSource(ctx.source) ? UriUtils.toUri(ctx.source.uri) : undefined;
869
+ /**
870
+ * Settle the build before a reference question is answered from the index.
871
+ *
872
+ * A query issued right after a model update otherwise races the rebuild and
873
+ * reads an index missing the symbols it asks about. The cost is not confined
874
+ * to this call: a scope built inside that window is cached, and the build's
875
+ * remaining documents link against it.
876
+ *
877
+ * Waits the source document specifically, but ONLY when one is loaded at
878
+ * that URI: a synthetic source can address a URI with no document (a
879
+ * directory URI, typically), where a per-URI `waitUntil` throws "No document
880
+ * found". Falls back to a global settle, matching the id-based
881
+ * `ElementSource` path.
882
+ *
883
+ * Bounded by {@link DataServerOptions.referenceSettleTimeoutMs}.
884
+ */
885
+ protected async awaitReferencesLinked(source: ReferenceSource): Promise<void> {
886
+ const uri = isDocumentSource(source) || isSyntheticSource(source) ? UriUtils.toUri(source.uri) : undefined;
771
887
  const waitUri = uri && this.services.workspace.LangiumDocuments.hasDocument(uri) ? uri : undefined;
772
- await this.services.workspace.DocumentBuilder.waitUntil(DocumentState.Linked, waitUri);
888
+ const stopwatch = this.services.Clock.stopwatch();
889
+ const linked = this.services.workspace.DocumentBuilder.waitUntil(DocumentState.Linked, waitUri);
890
+ if ((await this.services.Clock.raceTimer(linked, this.options.referenceSettleTimeoutMs)) !== TIMED_OUT) {
891
+ return;
892
+ }
893
+ const elapsedMs = Math.round(stopwatch.elapsedMs);
894
+ // A per-URI wait resolves off a document-phase notification, so a
895
+ // missed one strands a wait on a document that HAS reached the
896
+ // phase. Re-read before failing; the workspace-wide wait has no
897
+ // equivalent reading and can only reject.
898
+ const document = waitUri ? this.services.workspace.LangiumDocuments.getDocument(waitUri) : undefined;
899
+ if (waitUri && document && document.state >= DocumentState.Linked) {
900
+ this.tracer
901
+ .withUri(waitUri.toString())
902
+ .warn(`Missed the 'Linked' notification after ${elapsedMs}ms; the document already reached it`);
903
+ return;
904
+ }
905
+ throw referenceSettleTimeoutError(elapsedMs);
906
+ }
907
+
908
+ async findReferenceCandidates(ctx: ReferenceContext): Promise<ReferenceCandidate[]> {
909
+ await this.awaitReferencesLinked(ctx.source);
773
910
  return this.resolveReferenceServices(ctx.source).CandidateProvider.find(ctx);
774
911
  }
775
912
 
776
- async resolveReference(ref: ReferenceRequest): Promise<ReferenceTarget<TTransfer> | undefined> {
913
+ async resolveReference<TElement extends TransferElement = TransferElement>(
914
+ ref: ReferenceRequest
915
+ ): Promise<ReferenceTarget<TElement> | undefined> {
916
+ await this.awaitReferencesLinked(ref.source);
777
917
  const resolved = this.resolveReferenceServices(ref.source).CandidateProvider.resolveCandidate(ref);
778
918
  if (!resolved) {
779
919
  return undefined;
780
920
  }
781
- const element = this.encoder.toTransfer(resolved.node) as unknown as TTransfer;
782
- return { ...resolved.candidate, element };
921
+ // `TElement` is the caller's claim about the target's type; nothing checks it.
922
+ return { ...resolved.candidate, element: this.encoder.toTransfer(resolved.node) as TElement };
783
923
  }
784
924
 
785
925
  async findNextName(args: FindNextNameArgs): Promise<string> {
@@ -789,15 +929,31 @@ export class DataServer<
789
929
  // synthetic source — and the create-element flow driving this method is
790
930
  // the one that produces directory URIs, on which a bare
791
931
  // `getServices(uri)` throws while `findReferenceCandidates` succeeds.
792
- const nameProvider = this.resolveReferenceServices(ReferenceSource.synthetic(args.uri, args.type)).NameProvider;
932
+ const source = ReferenceSource.synthetic(args.uri, args.type, args.language);
933
+ // Mid-rebuild the names of documents not yet re-indexed are absent from
934
+ // the taken set, and the caller PERSISTS what it is handed. Nothing
935
+ // reserves the answer either, so a name is free of the collisions the
936
+ // index knows about and of nothing else — two callers racing get the same
937
+ // proposal, which only uniqueness at the write can catch.
938
+ await this.awaitReferencesLinked(source);
939
+ const nameProvider = this.resolveReferenceServices(source).NameProvider;
793
940
  const tier = args.tier ?? 'project';
794
941
  if (tier === 'public') {
795
942
  return nameProvider.findNextProjectQualifiedName(args.type, args.proposal);
796
943
  }
797
944
  if (tier === 'local') {
798
- // Document-scoped uniqueness: the document root is the container.
799
- const document = await this.modelService.ensureDocumentState(args.uri);
800
- return document.root ? nameProvider.findNextName(args.type, args.proposal, document.root as AstNode) : args.proposal;
945
+ // Document-scoped uniqueness: the document root is the container. A
946
+ // URI with no document has nothing to collide with, like a document
947
+ // with no root.
948
+ try {
949
+ const document = await this.modelService.ensureDocumentState(args.uri);
950
+ return document.root ? nameProvider.findNextName(args.type, args.proposal, document.root) : args.proposal;
951
+ } catch (error: unknown) {
952
+ if (this.isMissingDocument(error, args.uri)) {
953
+ return args.proposal;
954
+ }
955
+ throw error;
956
+ }
801
957
  }
802
958
  const project = this.services.workspace.ProjectManager.getProject(uri);
803
959
  if (!project) {
@@ -834,67 +990,32 @@ export class DataServer<
834
990
  }
835
991
 
836
992
  /**
837
- * Pick the language that owns a reference source, in order:
993
+ * Pick the language that owns a reference source: the registry's
994
+ * `getLanguageFor` over the source's declared language, URI and AST type,
995
+ * with an {@link isElementSource} routed by the document that holds the
996
+ * element (see {@link findElementDocumentUri}) before its type, and
997
+ * {@link fallbackReferenceLanguage}, adopter policy, last.
838
998
  *
839
- * 1. A URI-bearing source ({@link isDocumentSource}/{@link isSyntheticSource})
840
- * whose URI resolves to a registered language routes by URI.
841
- * 2. A single-language workspace always routes to that one language — so
842
- * single-language adopters never reach the steps below, and never pay
843
- * for them.
844
- * 3. An {@link isElementSource} source carries no URI, so in a
845
- * multi-language workspace it is routed via the document that holds the
846
- * element (see {@link findElementDocumentUri}).
847
- * 4. A source carrying an AST type is routed by that type when exactly one
848
- * registered grammar can produce it (see
849
- * {@link resolveReferenceLanguageByType}, over the registry's own type
850
- * index). This is what resolves a create-element flow's synthetic source
851
- * on a bare directory URI without asking the adopter.
852
- * 5. Anything still unresolved falls to {@link fallbackReferenceLanguage},
853
- * which is adopter policy.
999
+ * An unregistered declared language falls through rather than failing: the
1000
+ * id crosses the wire, so a client built against a head with one more
1001
+ * grammar would otherwise lose queries the other signals answer.
854
1002
  */
855
1003
  protected resolveReferenceLanguage(source: ReferenceSource): HydraniumLanguageServices | undefined {
856
1004
  const registry = this.services.ServiceRegistry;
857
- // `getServicesFor` (non-throwing, one ladder walk) gates the URI lookup:
858
- // an extensionless / unregistered URI falls through to the steps below
859
- // rather than throwing "no services for the extension ''".
860
- const uri = isDocumentSource(source) || isSyntheticSource(source) ? UriUtils.toUri(source.uri) : undefined;
861
- const byUri = uri && registry.getServicesFor(uri);
862
- if (byUri) {
863
- return byUri;
864
- }
865
- const all = registry.all;
866
- if (all.length === 1) {
867
- return all[0] as HydraniumLanguageServices;
868
- }
869
1005
  if (isElementSource(source)) {
870
- const documentUri = this.findElementDocumentUri(source);
871
- const byDocument = documentUri && registry.getServicesFor(documentUri);
872
- if (byDocument) {
873
- return byDocument;
874
- }
1006
+ // The index lookup runs only once the cheaper signals abstained.
1007
+ return (
1008
+ registry.getLanguageFor({}) ??
1009
+ registry.getLanguageFor({ target: this.findElementDocumentUri(source), type: source.type }) ??
1010
+ this.fallbackReferenceLanguage(source)
1011
+ );
875
1012
  }
876
- const byType = this.resolveReferenceLanguageByType(source);
877
- return byType ?? this.fallbackReferenceLanguage(source);
878
- }
879
-
880
- /**
881
- * Route a reference source by the AST type it carries — a
882
- * {@link isSyntheticSource}'s `type` (the transient node being created) or an
883
- * {@link isElementSource}'s optional narrowing `type`.
884
- *
885
- * Answers only when EXACTLY ONE registered grammar can produce the type.
886
- * Several can when the type comes from a grammar both import, and then the
887
- * type genuinely does not identify a language — that is a fall-through to
888
- * adopter policy, not a coin toss. Note this asks which grammar can *produce*
889
- * the type, not which mentions it: a grammar that merely cross-references a
890
- * type can never hold a node of it.
891
- */
892
- protected resolveReferenceLanguageByType(source: ReferenceSource): HydraniumLanguageServices | undefined {
893
- const type = isSyntheticSource(source) ? source.type : isElementSource(source) ? source.type : undefined;
894
- // The registry owns the type index — it is the one place that knows when
895
- // the registered set changed, so unlike a private memo here it cannot go
896
- // stale on a language registered after the first lookup.
897
- return type ? this.services.ServiceRegistry.soleServicesByType(type) : undefined;
1013
+ const hint = isSyntheticSource(source)
1014
+ ? { languageId: source.language, target: source.uri, type: source.type }
1015
+ : isDocumentSource(source)
1016
+ ? { target: source.uri }
1017
+ : {};
1018
+ return registry.getLanguageFor(hint) ?? this.fallbackReferenceLanguage(source);
898
1019
  }
899
1020
 
900
1021
  /**
@@ -904,15 +1025,14 @@ export class DataServer<
904
1025
  * disambiguates names that repeat across types (honouring grammar
905
1026
  * subtyping through `AstReflection.isSubtype`).
906
1027
  *
907
- * Only reached on the multi-language, name-based path (step 3 of
908
- * {@link resolveReferenceLanguage}); single-language adopters return at
909
- * step 2.
1028
+ * Only reached in a multi-language workspace, once the declared language
1029
+ * and the URI have abstained.
910
1030
  *
911
1031
  * Abstains when the name matches elements in more than one DOCUMENT — the
912
1032
  * index spans every language and is filled in build order, so "the
913
- * first match" would be file-watch order rather than an answer. Step 3
914
- * then falls through to the type-based step 4, which abstains on ties
915
- * in the same way, and finally to adopter policy.
1033
+ * first match" would be file-watch order rather than an answer. Routing
1034
+ * then falls through to the element's type, which abstains on ties in the
1035
+ * same way, and finally to adopter policy.
916
1036
  */
917
1037
  protected findElementDocumentUri(source: ElementSource): URI | undefined {
918
1038
  const matches = this.services.workspace.IndexManager.getElementsByName(source.name, source.type);
@@ -947,61 +1067,144 @@ export class DataServer<
947
1067
  // Internal plumbing
948
1068
  // ============================================================
949
1069
 
1070
+ /**
1071
+ * Whether `error`, from waiting on `uri`'s document, means that `uri` has
1072
+ * no document: a URI with neither a file nor text builds nothing, and
1073
+ * Langium's wait after the build rejects for want of one, with a
1074
+ * `ServerCancelled` response error. Any other failure, such as a read
1075
+ * that fails before a document is registered, a cancelled wait or a
1076
+ * `ReentrantWriteLockError`, is the caller's to see.
1077
+ */
1078
+ protected isMissingDocument(error: unknown, uri: string): boolean {
1079
+ return isResponseError(error) && error.code === SERVER_CANCELLED && this.modelService.getDocument(uri) === undefined;
1080
+ }
1081
+
950
1082
  /**
951
1083
  * Build a {@link TransferDocument} envelope from the current document state,
952
1084
  * delegating root + diagnostic encoding to {@link encoder} (see
953
- * `TransferEncoder.toTransferDocument` for the walk).
954
- *
955
- * The encoder field's generic-map binding is widened to
956
- * `Record<string, TransferElement>` at the framework-default level, and an
957
- * adopter supplying a typed-overlay encoder narrows the runtime shape to
958
- * its wire types. The cast on the return is where that invariant — the
959
- * adopter's `TTransfer` matches its encoder's overlay — is asserted, at a
960
- * single boundary point rather than spread across the callers.
1085
+ * `TransferEncoder.toTransferDocument` for the walk). `fingerprint` is the
1086
+ * caller's {@link computeDocumentFingerprint} of that same state, which
1087
+ * spares hashing it twice.
961
1088
  */
962
- protected envelope(uri: URI): TransferDocument<TTransfer, TDiagnostic> {
1089
+ protected envelope(uri: URI, fingerprint?: string): TransferDocument<TTransfer, TDiagnostic> {
963
1090
  // Resolve through the model service's canonicalizing gateway rather than
964
1091
  // reaching into `LangiumDocuments` directly, so a divergent (symlink) URI
965
1092
  // still finds the document the build keys by its real path — and so the
966
1093
  // data-server never has to remember to canonicalize this lookup itself.
967
1094
  const document = this.modelService.getDocument(uri.toString());
968
- if (!document) {
1095
+ if (!document || this.services.workspace.ModelLedger.isPlaceholder(document.parseResult.value)) {
969
1096
  // No document — a shaped envelope rather than a throw, so the caller
970
- // decides policy at its own layer; `root` is optional on the envelope
1097
+ // decides policy at its own layer; `model` is optional on the envelope
971
1098
  // so the compiler forces that decision. Adopters preferring to throw
972
- // override `envelope`.
973
- return {
974
- uri: uri.toString(),
975
- version: 0,
976
- root: undefined,
977
- diagnostics: [] as TDiagnostic[]
978
- };
1099
+ // override `envelope`. The builder's unparsed placeholder is no
1100
+ // document either: its root parses no text.
1101
+ return TransferDocument.absent<TTransfer, TDiagnostic>(uri.toString());
979
1102
  }
980
- return this.encoder.toTransferDocument<AstNode>(document) as unknown as TransferDocument<TTransfer, TDiagnostic>;
1103
+ const encoded = this.encoder.toTransferDocument(document);
1104
+ return this.withServerState(encoded, fingerprint ?? this.computeDocumentFingerprint(document.parseResult.value, encoded));
981
1105
  }
982
1106
 
983
1107
  /**
984
- * Subscribe one listener at the configured {@link DataServerOptions.subscriptionPhase}.
985
- * The listener dispatches subscription events for every matching URI. A
986
- * single listener (rather than one per subscription) keeps the cost flat
987
- * regardless of subscriber count.
1108
+ * The transfer document a request answers with for `astDocument`, its model
1109
+ * hashed and the text stamped on it by {@link withServerState}. Every
1110
+ * document a request answers with goes through here, so none goes out
1111
+ * without them.
988
1112
  */
989
- protected subscribeToDocumentBuilder(): void {
990
- this.disposables.push(
991
- this.services.workspace.DocumentBuilder.onDocumentPhase(this.options.subscriptionPhase, (document, cancelToken) =>
992
- this.dispatchPhaseEvent(document, cancelToken)
993
- )
994
- );
1113
+ protected encodeDocument(astDocument: AstDocument<AstNode, AstDiagnostic>): TransferDocument<TTransfer, TDiagnostic> {
1114
+ const encoded = this.encoder.astDocumentToTransferDocument(astDocument);
1115
+ return this.withServerState(encoded, this.computeDocumentFingerprint(astDocument.root, encoded));
1116
+ }
1117
+
1118
+ /**
1119
+ * `encoded` with `hash` on its model and the text the server holds for it:
1120
+ * the store's while it holds the document, the build's otherwise. The text
1121
+ * is read when the document is sent rather than kept with the build: a save
1122
+ * changes `text.dirty` without a rebuild, and the store's text moves on before
1123
+ * the build that follows it.
1124
+ */
1125
+ protected withServerState(
1126
+ encoded: EncodedTransferDocument<TransferElement, TDiagnostic>,
1127
+ hash: string
1128
+ ): TransferDocument<TTransfer, TDiagnostic> {
1129
+ const text = this.services.workspace.TextDocuments.textState(encoded.uri) ?? this.builtTextState(encoded.uri);
1130
+ if (text && text.version > encoded.model.version) {
1131
+ this.tracer.withUri(encoded.uri).debug(`Send a model behind its text: model v${encoded.model.version} / text v${text.version}`);
1132
+ }
1133
+ return {
1134
+ uri: encoded.uri,
1135
+ // The encoder returns the structural base; `TTransfer` is the adopter's
1136
+ // declaration of the roots its languages produce, and nothing checks it.
1137
+ model: { ...encoded.model, hash } as TransferModelSnapshot<TTransfer, TDiagnostic>,
1138
+ ...(text ? { text } : {})
1139
+ };
1140
+ }
1141
+
1142
+ /**
1143
+ * The text a build read for a document the store never held, which is
1144
+ * clean: the store holds every text a client gave it.
1145
+ */
1146
+ protected builtTextState(uri: string): TextState | undefined {
1147
+ const textDocument = this.modelService.getDocument(uri)?.textDocument;
1148
+ return textDocument && { version: textDocument.version, hash: textHash(textDocument.getText()), dirty: false };
1149
+ }
1150
+
1151
+ /**
1152
+ * Follow every document at the configured
1153
+ * {@link DataServerOptions.subscriptionPhase}, once per connection rather
1154
+ * than once per watch, so the cost stays flat in the number of watches.
1155
+ */
1156
+ protected subscribeToModelUpdates(): void {
1157
+ const phase = this.options.subscriptionPhase;
1158
+ this.disposables.push(this.modelService.onModelUpdated(event => this.dispatchPhaseEvent(event), { phase }));
1159
+ // One message per build rather than one per document.
1160
+ this.disposables.push(this.modelService.onModelsBuilt(event => this.dispatchBuiltEvent(event.uris), { phase }));
1161
+ }
1162
+
1163
+ /**
1164
+ * Report the documents that reached {@link DataServerOptions.subscriptionPhase}
1165
+ * and that NOBODY on this
1166
+ * connection is watching — the ones a client was not told about through
1167
+ * {@link dispatchPhaseEvent}, which is gated per URI.
1168
+ *
1169
+ * **The gap this closes is the one no other source can observe.** A document
1170
+ * the client watches, it hears about already. A file changed on disk, the
1171
+ * host's own filesystem watcher reports — and on a browser host, where the
1172
+ * workspace lives behind this head, an external change cannot happen at all.
1173
+ * What is left, and what nothing outside the server can see, is a document
1174
+ * rebuilt because something it DEPENDS ON changed: its file never changed, so
1175
+ * a filesystem watcher is silent by construction, and it has no subscriber, so
1176
+ * the update channel is silent by design. A consumer displaying data derived
1177
+ * from that document — a tree label, a decorator — otherwise goes stale with
1178
+ * no signal from any source, and repairs itself the moment someone opens the
1179
+ * file to investigate, which is what makes it expensive to diagnose later.
1180
+ *
1181
+ * Payload-free on purpose: URIs only, so a client re-reads what it displays
1182
+ * rather than being handed transfer documents it did not ask for. That is the
1183
+ * property the watcher gate protects, and it is preserved here by
1184
+ * carrying no document rather than by gating the message.
1185
+ *
1186
+ * Quiet in the common case. Editing a document that an editor has open leaves
1187
+ * that URI watched and therefore out of this set, so a build with no
1188
+ * dependents produces nothing at all, and workspace initialisation produces
1189
+ * nothing because it does not build to this phase. The ceiling is a
1190
+ * whole-workspace rebuild at the subscription phase: one message, URIs only.
1191
+ */
1192
+ protected dispatchBuiltEvent(built: readonly CanonicalUri[]): void {
1193
+ const uris = built.map(uri => this.canonicalKey(uri)).filter(uri => !this.uriWatchRecords.get(uri)?.watchers.size);
1194
+ if (uris.length === 0) {
1195
+ return;
1196
+ }
1197
+ this.tracer.debug(`Emit onDocumentsBuilt: ${uris.length} unwatched document(s)`);
1198
+ this.clientProxy.onDocumentsBuilt({ uris });
995
1199
  }
996
1200
 
997
1201
  /**
998
- * Subscribe to the universal save event on `HydraniumTextDocuments` so a
999
- * single `onDocumentSaved` wire notification fires for ANY save of a
1000
- * subscribed URI — regardless of whether the save originated from the
1001
- * data-server's RPC `saveModelDocument`, the LSP head's text-editor save,
1002
- * or any other client writing through `notifyDidSaveTextDocument`.
1202
+ * Follow every save so a single `onDocumentSaved` wire notification fires
1203
+ * for ANY save of a subscribed URI — regardless of whether the save
1204
+ * originated from the data-server's RPC `saveModelDocument`, the LSP head's
1205
+ * text-editor save, or any other client.
1003
1206
  *
1004
- * Architectural symmetry with {@link subscribeToDocumentBuilder}: every
1207
+ * Architectural symmetry with {@link subscribeToModelUpdates}: every
1005
1208
  * subscribed client sees every state change to documents they care about,
1006
1209
  * regardless of which client triggered it. Firing `onDocumentSaved` only
1007
1210
  * from the data-server's own RPC path is a bug, not an optimisation: an
@@ -1009,41 +1212,114 @@ export class DataServer<
1009
1212
  * clients, which never clear their dirty state.
1010
1213
  */
1011
1214
  protected subscribeToTextDocumentSaves(): void {
1012
- this.disposables.push(this.services.workspace.TextDocuments.onDidSave(event => this.dispatchSaveEvent(event)));
1215
+ this.disposables.push(this.modelService.onModelSaved(event => this.dispatchSaveEvent(event)));
1216
+ }
1217
+
1218
+ /** Relay each change of a document's dirty state; see {@link dispatchDirtyEvent}. */
1219
+ protected subscribeToDirtyChanges(): void {
1220
+ this.disposables.push(this.modelService.onDirtyChanged(event => this.dispatchDirtyEvent(event)));
1221
+ }
1222
+
1223
+ /**
1224
+ * Send a dirty flip to the connection, gated on the URI's watchers as
1225
+ * {@link dispatchSaveEvent} is: the answer at any one moment travels on
1226
+ * every document sent, so a client that watches nothing reads it there.
1227
+ */
1228
+ protected dispatchDirtyEvent(event: ModelDirtyChangedEvent): void {
1229
+ const uri = this.canonicalKey(event.uri);
1230
+ if (this.uriWatchRecords.get(uri)?.watchers.size) {
1231
+ this.clientProxy.onDocumentDirtyChanged({ uri, text: event.text });
1232
+ }
1013
1233
  }
1014
1234
 
1015
1235
  /**
1016
- * Mark the last-close transition per URI (see
1017
- * {@link pendingRevertBroadcasts}). The listener consults
1018
- * `isOpenInAnyClient` AFTER the store decremented the closing client's
1019
- * hold, so a `false` answer means this close was the last one.
1236
+ * Mark each document the store releases, so {@link dispatchPhaseEvent}
1237
+ * broadcasts the build that follows even without a watcher. By default
1238
+ * that build reverts the document to its disk content, discarding unsaved
1239
+ * in-session edits, and the last close typically also
1240
+ * removed the last watcher: without the broadcast, a consumer that only
1241
+ * ever fetches via `getModelDocument` keeps showing the discarded state.
1242
+ * The broadcast is de-duplicated against the fingerprint while a watcher
1243
+ * holds one, so a close whose disk state equals the last emitted state
1244
+ * stays silent; with no watcher there is no fingerprint, and every release
1245
+ * build is sent. The first build after the release consumes the mark, and
1246
+ * goes out as the release unless it is stamped past the version a client
1247
+ * opened the document at again: a client's write or an integrity repair
1248
+ * that rides it keeps its author.
1249
+ *
1250
+ * A release, not the close itself: a document whose last client lost its
1251
+ * connection is released only once the release grace runs out, or when
1252
+ * another client opens it meanwhile, and not at all when a client lost
1253
+ * from it opens it again within its own grace.
1020
1254
  */
1021
- protected subscribeToTextDocumentCloses(): void {
1022
- const textDocuments = this.services.workspace.TextDocuments;
1255
+ protected subscribeToModelReleases(): void {
1023
1256
  this.disposables.push(
1024
- textDocuments.onDidClose(event => {
1025
- if (!textDocuments.isOpenInAnyClient(event.document.uri)) {
1026
- this.pendingRevertBroadcasts.add(this.canonicalKey(event.document.uri));
1027
- }
1257
+ this.modelService.onModelReleased(event => {
1258
+ const uri = this.canonicalKey(event.uri);
1259
+ const record = this.uriWatchRecords.get(uri) ?? { watchers: new Set<string>() };
1260
+ record.releaseBroadcastPending = true;
1261
+ this.uriWatchRecords.set(uri, record);
1262
+ this.tracer.withUri(uri).debug('Release marked: the first build after it is broadcast, watched or not');
1028
1263
  })
1029
1264
  );
1030
1265
  }
1031
1266
 
1032
- /** Fan out a save event for the document's URI, gated by the subscription map. */
1033
- protected dispatchSaveEvent(event: ClientTextDocumentChangeEvent<TextDocument>): void {
1034
- // The save event arrives under the CLIENT URI the text store keys by (e.g. a
1035
- // symlink path S); the subscription map and `dispatchPhaseEvent` key by the
1036
- // CANONICAL identity R. Canonicalize before both the gate and the envelope so
1037
- // a save of a symlinked file isn't silently dropped (and the envelope resolves
1038
- // the R-keyed document rather than missing into an empty one).
1267
+ /**
1268
+ * Fan out a deletion for each removed URI, to EVERY client on the connection
1269
+ * rather than only to the ones watching that URI.
1270
+ *
1271
+ * **Deliberately ungated, where {@link dispatchPhaseEvent} and
1272
+ * {@link dispatchSaveEvent} are gated.** Those two carry a built document and
1273
+ * fire on every build, so the watcher gate is what keeps bandwidth
1274
+ * proportional to what a client asked for. A deletion is neither: it is one
1275
+ * URI, it is rare, and it reports the workspace's STRUCTURE rather than a
1276
+ * document's content. Gating it forces any consumer that displays the
1277
+ * workspace — a model tree, a file decorator — to learn about disappearances
1278
+ * from a filesystem watcher instead, which a client hosted in a browser has
1279
+ * no way to run: its workspace lives behind the head. A client that does not
1280
+ * care filters on the URI, which costs it a comparison.
1281
+ *
1282
+ * The precedent is the release broadcast (see
1283
+ * {@link subscribeToModelReleases}), where the same judgement was already made
1284
+ * in the other direction: a transition that matters enough is delivered
1285
+ * without a subscription.
1286
+ *
1287
+ * A deletion is announced before the build that follows it, so this goes
1288
+ * out ahead of the `'rebuilt'` events for the dependents whose references
1289
+ * the deletion just broke.
1290
+ *
1291
+ * The URI's fingerprint and release mark are dropped: they describe a
1292
+ * document that no longer exists. A kept fingerprint is adopted as the
1293
+ * baseline and suppresses the first emit after the file returns with its
1294
+ * previous content.
1295
+ *
1296
+ * The watchers SURVIVE: a recreated file resumes delivering to them with no
1297
+ * re-subscription, and a client that answers the deletion by closing
1298
+ * releases the watch through `closeModelDocument` anyway.
1299
+ */
1300
+ protected dispatchDeleteEvent(event: ModelDeletedEvent): void {
1301
+ const uri = this.canonicalKey(event.uri);
1302
+ const record = this.uriWatchRecords.get(uri);
1303
+ if (record) {
1304
+ record.fingerprint = undefined;
1305
+ record.fingerprintVersion = undefined;
1306
+ record.releaseBroadcastPending = undefined;
1307
+ record.sentVersion = undefined;
1308
+ this.pruneUriWatchRecord(uri);
1309
+ }
1310
+ this.clientProxy.onDocumentDeleted({ uri });
1311
+ }
1312
+
1313
+ /** Fan out a save event for the document's URI, gated on the URI's watchers. */
1314
+ protected dispatchSaveEvent(event: AstDocumentSavedEvent<AstNode>): void {
1039
1315
  const uri = this.canonicalKey(event.document.uri);
1040
- if (!this.subscriptions.has(uri)) {
1316
+ if (!this.uriWatchRecords.get(uri)?.watchers.size) {
1041
1317
  return;
1042
1318
  }
1043
- const response = this.envelope(UriUtils.toUri(uri));
1319
+ const response = this.envelope(UriUtils.toUri(event.document.uri));
1044
1320
  const wireEvent: TransferDocumentSavedEvent<TTransfer, TDiagnostic> = {
1045
1321
  document: response,
1046
- sourceClientId: event.clientId
1322
+ sourceClientId: event.sourceClientId
1047
1323
  };
1048
1324
  this.clientProxy.onDocumentSaved(wireEvent);
1049
1325
  }
@@ -1057,116 +1333,114 @@ export class DataServer<
1057
1333
  * out to multiple local subscribers with an `Emitter<T>` — the
1058
1334
  * framework deliberately does NOT promise multi-listener semantics.
1059
1335
  *
1060
- * The dispatch is guarded by two filters:
1061
- * 1. **Subscription map**: events for URIs no subscriber registered for
1062
- * are NOT sent over the wire (bandwidth scales with subscribed URIs,
1063
- * not phase events).
1064
- * 2. **Emission fingerprint**: events for rebuilds that produce no
1065
- * observable change since the last emit are suppressed. See
1066
- * {@link lastEmittedFingerprint} for the rationale.
1336
+ * An event for a URI no client watches is NOT sent over the wire, so
1337
+ * bandwidth scales with watched URIs rather than with phase events; a
1338
+ * pending release mark is the exception (see
1339
+ * {@link subscribeToModelReleases}).
1340
+ *
1341
+ * An event for a rebuild with no observable change since the last emit is
1342
+ * suppressed against the URI's fingerprint and the version it was taken at;
1343
+ * a new version with the same fingerprint goes out with the same
1344
+ * `model.hash`, so a watcher still learns the version. Such rebuilds are routine: a
1345
+ * second client attaching to an open document makes
1346
+ * `HydraniumTextDocuments.refreshContent` fire `onDidChangeContent` purely
1347
+ * to re-trigger the build, and Langium rebuilds a document when a
1348
+ * dependency changes even if its diagnostics come out the same. Sent at a
1349
+ * version this head has not sent, they reach watchers as `'changed'`, and a
1350
+ * watcher that reads `'changed'` as another client's write and resets its
1351
+ * root to the server view loses the user's edits. The first {@link watchModelDocument} baselines the
1352
+ * fingerprint, so the first event after a watch is de-duplicated against
1353
+ * the state the watcher just fetched; it goes with the last watcher, so the
1354
+ * next first watch re-baselines rather than adopting a stale digest. A
1355
+ * digest keeps its memory constant in document size; see
1356
+ * {@link computeDocumentFingerprint} for the inputs.
1067
1357
  */
1068
- protected dispatchPhaseEvent(document: LangiumDocument, cancelToken: CancellationToken): void {
1069
- if (cancelToken.isCancellationRequested) {
1070
- // Build preempted by a concurrent write lock (or other cancel source) —
1071
- // the subscription event is stale by the time it would fire. Skip
1072
- // emission so RPC subscribers don't surface intermediate states.
1073
- // A pending revert mark is deliberately NOT consumed here: the
1074
- // follow-up build re-fires this phase and broadcasts then.
1358
+ protected dispatchPhaseEvent(event: ModelUpdatedEvent<AstNode>): void {
1359
+ const uri = this.canonicalKey(event.document.uri);
1360
+ const record = this.uriWatchRecords.get(uri);
1361
+ const afterRelease = record?.releaseBroadcastPending === true;
1362
+ if (!record || (record.watchers.size === 0 && !afterRelease)) {
1075
1363
  return;
1076
1364
  }
1077
- const uri = document.uri.toString();
1078
- // Consume the last-close mark even when subscriptions exist — the
1079
- // regular dispatch below serves those watchers, and the mark's author
1080
- // attribution is more precise than the post-close `UNKNOWN_CLIENT_ID`
1081
- // the author lookup would yield.
1082
- const revertedOnClose = this.pendingRevertBroadcasts.delete(uri);
1083
- if (!this.subscriptions.has(uri) && !revertedOnClose) {
1365
+ if (afterRelease) {
1366
+ // Consumed even when watchers exist — the regular dispatch below
1367
+ // serves them, and the mark's attribution is more precise than the
1368
+ // post-close unknown client the manager's attribution would yield.
1369
+ // With no watcher the record goes now, so the fingerprint written
1370
+ // below lands on a detached record and the next first watch
1371
+ // re-baselines instead of adopting it.
1372
+ record.releaseBroadcastPending = undefined;
1373
+ this.pruneUriWatchRecord(uri);
1374
+ }
1375
+ // The encoder reads the built document, which the event's holds for this
1376
+ // synchronous run.
1377
+ const document = this.modelService.getDocument(event.document.uri);
1378
+ if (!document) {
1084
1379
  return;
1085
1380
  }
1086
- const fingerprint = this.computeDocumentFingerprint(document);
1087
- if (this.lastEmittedFingerprint.get(uri) === fingerprint) {
1381
+ const fingerprint = this.computeDocumentFingerprint(document.parseResult.value, this.encoder.toTransferDocument(document));
1382
+ const stamped = this.services.workspace.ModelLedger.versionOf(document.parseResult.value);
1383
+ if (record.fingerprint === fingerprint && record.fingerprintVersion === stamped) {
1088
1384
  // A rebuild that produced no observable change since the last emit —
1089
1385
  // suppressed so RPC subscribers don't see a no-op broadcast. Logged at
1090
1386
  // debug so a *needed* re-broadcast wrongly suppressed by this dedup
1091
1387
  // (the failure mode the fingerprint strategy must avoid) is visible.
1092
- this.tracer.withUri(uri).debug(`Suppress onDocumentUpdated v${document.textDocument.version}: fingerprint unchanged`);
1388
+ this.tracer.withUri(uri).debug(`Suppress onDocumentUpdated v${stamped}: fingerprint and version unchanged`);
1093
1389
  return;
1094
1390
  }
1095
- this.lastEmittedFingerprint.set(uri, fingerprint);
1096
- const event: TransferDocumentUpdatedEvent<TTransfer, TDiagnostic> = {
1097
- document: this.envelope(document.uri),
1098
- sourceClientId: revertedOnClose ? REVERT_ON_CLOSE_CLIENT_ID : this.resolveSourceClientId(document),
1099
- reason: this.resolveUpdateReason(document.uri)
1391
+ record.fingerprint = fingerprint;
1392
+ record.fingerprintVersion = stamped;
1393
+ // The event's attribution, so this head names the same client as the
1394
+ // in-process heads do for one build, except for a version this head
1395
+ // already sent; see `DataServerUriWatchRecord.sentVersion`.
1396
+ const version = this.modelService.isOpen(event.document.uri) ? document.textDocument.version : undefined;
1397
+ const { reason, sourceClientId } =
1398
+ version !== undefined && record.sentVersion === version
1399
+ ? { reason: 'rebuilt' as const, sourceClientId: UNKNOWN_CLIENT_ID }
1400
+ : event;
1401
+ record.sentVersion = version;
1402
+ // The stamp, not the live version: a write after the build's parse moves
1403
+ // only the latter. At the opened version too, which a reopen with the
1404
+ // released text keeps.
1405
+ const openedVersion = afterRelease ? this.modelService.openedVersion(event.document.uri) : undefined;
1406
+ const releaseBuild = afterRelease && (openedVersion === undefined || stamped <= openedVersion);
1407
+ const wireEvent: TransferDocumentUpdatedEvent<TTransfer, TDiagnostic> = {
1408
+ document: this.envelope(document.uri, fingerprint),
1409
+ sourceClientId: releaseBuild ? DOCUMENT_RELEASE_CLIENT_ID : sourceClientId,
1410
+ reason
1100
1411
  };
1412
+ const mark = afterRelease ? `, release mark: opened v${openedVersion ?? '-'}, attributed ${sourceClientId}` : '';
1101
1413
  this.tracer
1102
1414
  .withUri(uri)
1103
- .debug(`Emit onDocumentUpdated v${event.document.version} (reason=${event.reason}, sourceClientId=${event.sourceClientId})`);
1104
- this.clientProxy.onDocumentUpdated(event);
1415
+ .debug(`Emit onDocumentUpdated v${stamped} (reason=${wireEvent.reason}, sourceClientId=${wireEvent.sourceClientId}${mark})`);
1416
+ this.clientProxy.onDocumentUpdated(wireEvent);
1105
1417
  }
1106
1418
 
1107
1419
  /**
1108
- * Fingerprint of the document's observable state, used to de-dup
1109
- * `onDocumentUpdated` emissions. The {@link FingerprintStrategy} option
1110
- * selects what is hashed (default `'transfer-document'` — see the type), and
1111
- * {@link additionalFingerprintInputs} folds in any extra adopter signal.
1420
+ * Fingerprint of the snapshot `root` and its encoding `encoded`, used to
1421
+ * de-dup `onDocumentUpdated` emissions and sent as every document's
1422
+ * `model.hash`, so it must not depend on the version or on live state. The
1423
+ * {@link FingerprintStrategy} option selects what is hashed (default
1424
+ * `'transfer-document'` — see the type).
1112
1425
  *
1113
- * The default `'transfer-document'` strategy goes through the encoder's
1114
- * {@link TransferEncoder.toTransferDocument}, which is cached per build —
1115
- * so within one phase event the fingerprint and the subsequently-emitted
1116
- * `envelope` share a single encode walk.
1426
+ * {@link TransferEncoder.toTransferDocument} is cached per build, so within
1427
+ * one phase event the fingerprint and the subsequently-emitted `envelope`
1428
+ * share a single encode walk.
1117
1429
  *
1118
1430
  * Wrapped in {@link Tracer.time} against {@link FINGERPRINT_LOG_AFTER_MS}.
1119
1431
  */
1120
- protected computeDocumentFingerprint(document: LangiumDocument): string {
1121
- return this.tracer.withUri(document.uri.toString()).time(
1122
- 'Compute fingerprint',
1123
- () => {
1124
- const base =
1432
+ protected computeDocumentFingerprint(root: AstNode, encoded: EncodedTransferDocument<TransferElement, TDiagnostic>): string {
1433
+ return this.tracer
1434
+ .withUri(encoded.uri)
1435
+ .time(
1436
+ 'Compute fingerprint',
1437
+ () =>
1125
1438
  this.options.fingerprintStrategy === 'text-diagnostics'
1126
- ? textDiagnosticsFingerprint(document)
1127
- : transferDocumentFingerprint(this.encoder.toTransferDocument(document));
1128
- const extra = this.additionalFingerprintInputs(document);
1129
- return extra.length === 0 ? base : fingerprintHash([base, FINGERPRINT_SEPARATOR, ...extra.map(value => JSON.stringify(value))]);
1130
- },
1131
- 'debug',
1132
- { logAfterMs: FINGERPRINT_LOG_AFTER_MS }
1133
- );
1134
- }
1135
-
1136
- /**
1137
- * Extra inputs folded into {@link computeDocumentFingerprint} alongside the
1138
- * chosen {@link FingerprintStrategy}. Default: none. Override to contribute a
1139
- * signal that lives outside the document's text / root / diagnostics — each
1140
- * entry need only be stable across equivalent emissions and
1141
- * JSON-serialisable.
1142
- */
1143
- protected additionalFingerprintInputs(_document: LangiumDocument): readonly unknown[] {
1144
- return [];
1145
- }
1146
-
1147
- /**
1148
- * Discriminate the reason for a phase-event-driven update notification.
1149
- * Uses the most recent `DocumentBuilder.onUpdate` snapshot:
1150
- * - URI in the `deleted` list → `'deleted'`
1151
- * - URI in the `changed` list → `'changed'` (the URI was passed to
1152
- * `documentBuilder.update(changed, deleted)`, which spans `didChange`
1153
- * text-document events and programmatic `update([uri], [])` calls).
1154
- * - Otherwise → `'rebuilt'` (cascade re-derivation: this URI was rebuilt
1155
- * because something it depends on changed; its own text wasn't flagged).
1156
- *
1157
- * `'saved'` is NOT emitted from this code path — saves take the dedicated
1158
- * `DataClientProtocol.onDocumentSaved` channel; adopters that want a
1159
- * unified update stream synthesise `'saved'` in their bridge layer.
1160
- */
1161
- protected resolveUpdateReason(uri: URI): TransferDocumentUpdatedEvent<TTransfer, TDiagnostic>['reason'] {
1162
- const last = this.lastBuildUpdate;
1163
- if (last?.deleted.some(deleted => UriUtils.equals(deleted, uri))) {
1164
- return 'deleted';
1165
- }
1166
- if (last?.changed.some(changed => UriUtils.equals(changed, uri))) {
1167
- return 'changed';
1168
- }
1169
- return 'rebuilt';
1439
+ ? textDiagnosticsFingerprint(this.services.workspace.ModelLedger.textOf(root), encoded.model)
1440
+ : transferDocumentFingerprint(encoded.model),
1441
+ 'debug',
1442
+ { logAfterMs: FINGERPRINT_LOG_AFTER_MS }
1443
+ );
1170
1444
  }
1171
1445
 
1172
1446
  /**
@@ -1212,20 +1486,6 @@ export class DataServer<
1212
1486
  }
1213
1487
  }
1214
1488
 
1215
- /**
1216
- * Resolve the wire-level `sourceClientId` for `document`'s events — the client
1217
- * that authored its current version, from the version-author history on
1218
- * `HydraniumTextDocuments`. The protocol-level counterpart of the internal
1219
- * `AstDocumentManager.getAuthor`: a framework-internal rebuild has no author,
1220
- * so this surfaces the {@link UNKNOWN_CLIENT_ID} presentation default. Adopters
1221
- * that rebuild through non-text-document channels override to derive a
1222
- * source id of their own.
1223
- */
1224
- protected resolveSourceClientId(document: LangiumDocument): string {
1225
- const author = this.services.workspace.TextDocuments.getAuthor(document.textDocument.uri, document.textDocument.version);
1226
- return author ?? UNKNOWN_CLIENT_ID;
1227
- }
1228
-
1229
1489
  // --- Diagnostics (DataServerDiagnosticsProtocol) ---------------------------
1230
1490
  // Computed in THIS process: the data-server child holds the model store, so
1231
1491
  // these snapshots reflect the heap that actually carries the workspace
@@ -1287,11 +1547,11 @@ export class DataServer<
1287
1547
  /** Resolve a partial options object into a fully-defaulted form. */
1288
1548
  protected resolveOptions(partial: DataServerOptions): ResolvedDataServerOptions {
1289
1549
  const additionalMethods = partial.additionalMethods ?? [];
1290
- const builtIn = [...DATA_SERVER_PROTOCOL_METHODS, ...DATA_SERVER_DIAGNOSTICS_METHODS] as readonly string[];
1550
+ const builtIn: readonly string[] = [...DATA_SERVER_PROTOCOL_METHODS, ...DATA_SERVER_DIAGNOSTICS_METHODS];
1291
1551
  const overlap = additionalMethods.filter(name => builtIn.includes(name));
1292
1552
  if (overlap.length > 0) {
1293
1553
  throw new Error(
1294
- `DataServer.additionalMethods overlaps with built-in data-server methods: ${overlap.join(', ')}. ` +
1554
+ `DataServerOptions.additionalMethods overlaps with built-in data-server methods: ${overlap.join(', ')}. ` +
1295
1555
  'Adopter protocols must not redeclare framework method names; rename the adopter method or remove it from additionalMethods.'
1296
1556
  );
1297
1557
  }
@@ -1301,7 +1561,8 @@ export class DataServer<
1301
1561
  methodNamespace: partial.methodNamespace ?? DATA_SERVER_WIRE_PREFIX,
1302
1562
  additionalMethods,
1303
1563
  excludedMethods: partial.excludedMethods ?? [],
1304
- diagnostics: partial.diagnostics ?? defaultDataServerDiagnostics()
1564
+ diagnostics: partial.diagnostics ?? defaultDataServerDiagnostics(),
1565
+ referenceSettleTimeoutMs: partial.referenceSettleTimeoutMs ?? DataServer.DEFAULT_OPTIONS.referenceSettleTimeoutMs
1305
1566
  };
1306
1567
  }
1307
1568
  }