@hydranium/core 1.0.0-next.28 → 1.0.0-next.280

package/src/testing/playwright/flaky-network-proxy.ts

@@ -0,0 +1,240 @@
+/********************************************************************************
+ * Copyright (c) 2026 CrossBreeze, EclipseSource and others.
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the MIT License which is available in the project root.
+ *
+ * SPDX-License-Identifier: MIT
+ ********************************************************************************/
+
+import { createServer as createHttpServer, type Server as HttpServer } from 'node:http';
+import { connect, createServer as createTcpServer, type Server as TcpServer, type Socket } from 'node:net';
+
+/**
+ * A TCP proxy that can half-close a connection, so an e2e suite can reproduce a
+ * flaky network against a real browser session.
+ *
+ * Toggling the browser offline is not enough. On localhost the server sees the
+ * close immediately and reaps the socket before the browser reconnects, so the
+ * window where both sides disagree about the session — the one where reconnect
+ * defects live — never opens. Reaching it needs the browser's socket to die
+ * while the server still believes it is alive, and nothing inside the app or in
+ * Playwright's offline emulation can produce that asymmetry: both close cleanly
+ * at both ends.
+ *
+ * The proxy holds both halves of every connection and lets a test kill them
+ * independently:
+ *
+ *   `break`    destroy the BROWSER-side sockets, keep the server-side ones open.
+ *              The browser reconnects through a fresh pair while the server
+ *              still holds the old one.
+ *   `release`  destroy the orphaned server-side sockets `break` left. The
+ *              server's disconnect handler fires at once, rather than whenever
+ *              its next heartbeat write would have failed.
+ */
+
+/** Where the proxy listens, and the application it fronts. */
+export interface FlakyNetworkProxyOptions {
+   /** Port the browser connects to. */
+   readonly listenPort: number;
+   /** Port of the application being proxied. */
+   readonly targetPort: number;
+   /** Host of the application being proxied. Defaults to loopback. */
+   readonly targetHost?: string;
+   /** Port the HTTP control channel listens on. Defaults to `listenPort + 1000`. */
+   readonly controlPort?: number;
+}
+
+/** The commands the control channel accepts, and the proxy's own vocabulary. */
+export type FlakyNetworkCommand = 'break' | 'release' | 'status';
+
+/** Drives a running proxy. Implemented both in-process and over HTTP. */
+export interface FlakyNetworkControl {
+   /** Kill the browser side of every live connection; leave the server side up. */
+   break(): Promise<string>;
+   /** Kill the server side of every connection `break` orphaned. */
+   release(): Promise<string>;
+   /** How many connections are live and how many are orphaned. */
+   status(): Promise<string>;
+}
+
+export interface FlakyNetworkProxy extends FlakyNetworkControl {
+   /** Base URL a browser should be pointed at, i.e. Playwright's `baseURL`. */
+   readonly url: string;
+   /** Control channel, for workers that cannot reach this object directly. */
+   readonly controlUrl: string;
+   close(): Promise<void>;
+}
+
+interface ConnectionPair {
+   readonly id: number;
+   readonly browserSocket: Socket;
+   readonly targetSocket: Socket;
+   orphaned: boolean;
+}
+
+/**
+ * Starts the proxy in the calling process.
+ *
+ * Call it from a Playwright config, which runs in the main process before any
+ * worker forks. Workers are separate processes and cannot reach the returned
+ * object, which is what {@link controlUrl} and {@link flakyNetworkControlClient}
+ * are for.
+ */
+export async function startFlakyNetworkProxy(options: FlakyNetworkProxyOptions): Promise<FlakyNetworkProxy> {
+   const targetHost = options.targetHost ?? '127.0.0.1';
+   const controlPort = options.controlPort ?? options.listenPort + 1000;
+   const pairs = new Set<ConnectionPair>();
+   let sequence = 0;
+   let transcript: string[] = [];
+
+   const log = (message: string): void => {
+      transcript.push(message);
+      console.log(`[flaky-network] ${message}`);
+   };
+
+   const proxyServer: TcpServer = createTcpServer(browserSocket => {
+      const id = ++sequence;
+      const targetSocket = connect(options.targetPort, targetHost);
+      const pair: ConnectionPair = { id, browserSocket, targetSocket, orphaned: false };
+      pairs.add(pair);
+
+      // Forwarded by hand rather than with `pipe()`, so killing one side does not
+      // tear the other down — which is the entire point of the proxy.
+      browserSocket.on('data', chunk => {
+         if (!targetSocket.destroyed) {
+            targetSocket.write(chunk);
+         }
+      });
+      targetSocket.on('data', chunk => {
+         if (!pair.orphaned && !browserSocket.destroyed) {
+            browserSocket.write(chunk);
+         }
+      });
+
+      const teardown = (origin: 'browser' | 'target'): void => {
+         if (pair.orphaned && origin === 'browser') {
+            // Expected: `break` killed this side deliberately. The server side
+            // stays up until `release`.
+            return;
+         }
+         pairs.delete(pair);
+         browserSocket.destroy();
+         targetSocket.destroy();
+      };
+
+      browserSocket.on('close', () => teardown('browser'));
+      targetSocket.on('close', () => teardown('target'));
+      // A half-killed pair reliably raises ECONNRESET on the surviving side; an
+      // unhandled 'error' would take the whole runner down with it.
+      browserSocket.on('error', () => undefined);
+      targetSocket.on('error', () => undefined);
+   });
+
+   const commands: Record<FlakyNetworkCommand, () => void> = {
+      break: () => {
+         const live = [...pairs].filter(pair => !pair.orphaned);
+         if (live.length === 0) {
+            log('nothing to break');
+            return;
+         }
+         for (const pair of live) {
+            pair.orphaned = true;
+            pair.browserSocket.destroy();
+         }
+         log(`${live.length} connection(s) broken; the browser should reconnect shortly`);
+      },
+      release: () => {
+         const orphans = [...pairs].filter(pair => pair.orphaned);
+         if (orphans.length === 0) {
+            log('no orphaned connections; run `break` first');
+            return;
+         }
+         for (const pair of orphans) {
+            pairs.delete(pair);
+            pair.targetSocket.destroy();
+         }
+         log(`${orphans.length} orphaned connection(s) destroyed; their late disconnect fires now`);
+      },
+      status: () => {
+         const live = [...pairs].filter(pair => !pair.orphaned).length;
+         const orphaned = pairs.size - live;
+         log(`${live} live, ${orphaned} orphaned`);
+      }
+   };
+
+   const run = (name: string): string => {
+      transcript = [];
+      const command = commands[name as FlakyNetworkCommand];
+      if (command) {
+         command();
+      } else {
+         log(`unknown command '${name}'; use: ${Object.keys(commands).join(' | ')}`);
+      }
+      return transcript.join('\n');
+   };
+
+   const controlServer: HttpServer = createHttpServer((request, response) => {
+      const output = run((request.url ?? '').replace(/^\//, '').split('?')[0]);
+      response.writeHead(200, { 'content-type': 'text/plain' });
+      response.end(`${output}\n`);
+   });
+
+   await Promise.all([listen(proxyServer, options.listenPort), listen(controlServer, controlPort)]);
+   // Neither server keeps the process alive, so a runner that starts one from a
+   // config does not need a teardown hook to exit — and a caller that wants a
+   // deterministic shutdown still has `close()`.
+   proxyServer.unref();
+   controlServer.unref();
+
+   return {
+      url: `http://localhost:${options.listenPort}`,
+      controlUrl: `http://localhost:${controlPort}`,
+      break: async () => run('break'),
+      release: async () => run('release'),
+      status: async () => run('status'),
+      close: async () => {
+         for (const pair of pairs) {
+            pair.browserSocket.destroy();
+            pair.targetSocket.destroy();
+         }
+         pairs.clear();
+         await Promise.all([close(proxyServer), close(controlServer)]);
+      }
+   };
+}
+
+/**
+ * Drives a proxy running in another process, over its control channel.
+ *
+ * This is what a spec uses: Playwright workers are separate processes from the
+ * config that started the proxy.
+ */
+export function flakyNetworkControlClient(controlUrl: string): FlakyNetworkControl {
+   const send = async (command: FlakyNetworkCommand): Promise<string> => {
+      const response = await fetch(`${controlUrl}/${command}`);
+      if (!response.ok) {
+         throw new Error(`flaky-network proxy rejected '${command}': ${response.status}`);
+      }
+      return (await response.text()).trim();
+   };
+   return {
+      break: () => send('break'),
+      release: () => send('release'),
+      status: () => send('status')
+   };
+}
+
+function listen(server: TcpServer | HttpServer, port: number): Promise<void> {
+   return new Promise((resolve, reject) => {
+      server.once('error', reject);
+      server.listen(port, () => {
+         server.off('error', reject);
+         resolve();
+      });
+   });
+}
+
+function close(server: TcpServer | HttpServer): Promise<void> {
+   return new Promise(resolve => server.close(() => resolve()));
+}

package/src/testing/playwright/index.ts

@@ -11,6 +11,7 @@
 // `@hydranium/core/testing` barrel) so importing the framework's unit-test
 // helpers never pulls in `@playwright/test` — which is an OPTIONAL peer here.
 export * from './e2e-profiling.js';
+export * from './flaky-network-proxy.js';
 export * from './browser-capture-bridge.js';
 export * from './server-log-capture.js';
 export * from './server-log-fixture.js';

package/src/testing/playwright/server-log-capture.ts

@@ -10,7 +10,13 @@
 import { appendFileSync, existsSync, readdirSync, readFileSync, renameSync, unlinkSync, writeFileSync } from 'node:fs';
 import { join } from 'node:path';
 import { fileURLToPath } from 'node:url';
-import { DEFAULT_LOG_FILE_ENV, DEFAULT_LOG_LEVEL_ENV } from '@hydranium/protocol';
+import {
+   DEFAULT_LOG_FILE_ENV,
+   DEFAULT_LOG_FILE_LEVEL_ENV,
+   DEFAULT_LOG_LEVEL_ENV,
+   type LogThreshold,
+   parseLogLevel
+} from '@hydranium/protocol';
 import { toLogFileWorkspaceToken } from '../../langium/diagnostics/log-file-token.js';
 
 /**
@@ -69,11 +75,20 @@ export interface CaptureServerLogOptions {
     */
    dir?: string;
    /**
-    * Server log level set on `HYDRANIUM_LOG_LEVEL`. When omitted, respects a
-    * pre-set `HYDRANIUM_LOG_LEVEL` env (so a run can pick the level without editing
-    * the config), then falls back to `'debug'` (a capture at `'info'` is too thin).
+    * Threshold the log file is written at, set on `HYDRANIUM_LOG_FILE_LEVEL`;
+    * the client's log-level setting does not change it. When omitted, respects a
+    * pre-set `HYDRANIUM_LOG_FILE_LEVEL`, then `HYDRANIUM_LOG_LEVEL` env (so a run
+    * can pick the level without editing the config), then falls back to `'debug'`
+    * (a capture at `'info'` is too thin).
+    *
+    * The server's own threshold is left alone: it governs what the client sees,
+    * not the file, and the setting replaces it within moments of startup.
+    *
+    * The server runs every path guarded on a level this admits, its debug
+    * profiling included, so a timing run captured at a level that admits debug
+    * measures that work too.
     */
-   level?: string;
+   level?: LogThreshold;
 }
 
 /** A Playwright `reporter` config entry: `[modulePath, options]`. */
@@ -119,7 +134,11 @@ export function captureServerLog(options: CaptureServerLogOptions = {}): Capture
    const env: Record<string, string> = {
       [DEFAULT_LOG_FILE_ENV]: join(dir, `${WORKSPACE_PLACEHOLDER}.log`)
    };
-   env[DEFAULT_LOG_LEVEL_ENV] = options.level ?? process.env[DEFAULT_LOG_LEVEL_ENV] ?? 'debug';
+   env[DEFAULT_LOG_FILE_LEVEL_ENV] =
+      options.level ??
+      parseLogLevel(process.env[DEFAULT_LOG_FILE_LEVEL_ENV]) ??
+      parseLogLevel(process.env[DEFAULT_LOG_LEVEL_ENV]) ??
+      'debug';
    const reporterModule = fileURLToPath(new URL('./server-log-rename-reporter.js', import.meta.url));
    return { env, reporter: [reporterModule, { dir }] };
 }
@@ -160,16 +179,28 @@ export function serverLogSpecName(file: string): string {
  * Rename each `<token>.log` (backend) and `<token>.browser.log` (forwarded
  * browser console) to the spec name recorded in its `<token>.spec` sidecar
  * (written by {@link markServerLog}), then delete the sidecar. Colliding names —
- * a spec that opened more than one workspace — get a `-2`, `-3`, … suffix. Token
- * logs without a sidecar are left untouched. Run once, after all server processes
- * have exited (e.g. a reporter's `onEnd`). Best-effort.
+ * a spec that opened more than one workspace — get a `-2`, `-3`, … suffix. Run
+ * once, after all server processes have exited (e.g. a reporter's `onEnd`).
+ * Best-effort.
+ *
+ * Returns the workspace tokens left UNATTRIBUTED — a log the capture wrote that
+ * no test claimed, because that spec never reached
+ * {@link markServerLog} (no `serverLog` fixture, no `beforeEach` call).
+ *
+ * **Returned rather than ignored, because the unattributed state is the
+ * misleading one.** Such a log still exists, still has content, and is named by
+ * an opaque token with no test boundaries in it — so a capture run that covered
+ * nothing looks exactly like one that covered everything, and the logs are
+ * there to be read. A caller that can warn should warn.
  */
-export function renameServerLogs(dir: string): void {
+export function renameServerLogs(dir: string): string[] {
    let entries: string[];
    try {
       entries = readdirSync(dir);
    } catch {
-      return;
+      // An unreadable capture dir has nothing to report — including nothing
+      // unattributed, since no log was observed at all.
+      return [];
    }
    // Per suffix (`.log`, `.browser.log`) a `taken` set guards collisions, so the
    // backend and browser logs for one spec share the same friendly base name.
@@ -190,6 +221,11 @@ export function renameServerLogs(dir: string): void {
          // ignore
       }
    };
+   // Every token the capture produced a backend log for, so the ones no sidecar
+   // claims can be reported rather than silently left behind.
+   const unattributed = new Set(
+      entries.filter(entry => entry.endsWith('.log') && !entry.endsWith('.browser.log')).map(entry => entry.slice(0, -'.log'.length))
+   );
    for (const entry of entries) {
       if (!entry.endsWith('.spec')) {
          continue;
@@ -210,9 +246,11 @@ export function renameServerLogs(dir: string): void {
       if (!name) {
          continue;
       }
+      unattributed.delete(token);
       rename(token, name, '.log');
       rename(token, name, '.browser.log');
    }
+   return [...unattributed];
 }
 
 /**
@@ -377,6 +415,20 @@ export function forwardBrowserConsole(
    });
 }
 
+/** The label a `Warn` or `Error` entry starts with; continuation lines of a stack never do. */
+const WARN_OR_ERROR_LINE = /^\[(Warn|Error)\s*-/;
+
+/**
+ * The `Warn` and `Error` lines of `text` that no pattern of `allowlist`
+ * matches; its keys are the reasons the lines are expected. A clean run logs
+ * neither: `warn` is an anomaly that was handled and `error` is work that
+ * failed, so either one in a passing spec is a defect or a missing allowance.
+ */
+export function findUnexpectedServerLogLines(text: string, allowlist: Readonly<Record<string, RegExp>> = {}): string[] {
+   const patterns = Object.values(allowlist);
+   return text.split('\n').filter(line => WARN_OR_ERROR_LINE.test(line) && !patterns.some(pattern => pattern.test(line)));
+}
+
 /**
  * Per-test server-log capture, called from a spec's `afterEach` with the test's
  * workspace path:

package/src/testing/playwright/server-log-fixture.ts

@@ -11,7 +11,14 @@
 // The plain functions in `server-log-capture.ts` stay playwright-free; only
 // adopters that use this fixture need `@playwright/test` installed.
 import { test as base, type Fixtures, type PlaywrightTestArgs, type PlaywrightTestOptions } from '@playwright/test';
-import { attachServerLog, markServerLog } from './server-log-capture.js';
+import {
+   attachServerLog,
+   DEFAULT_SERVER_LOG_ATTACH_ENV,
+   findUnexpectedServerLogLines,
+   markServerLog,
+   resolveServerLogPath
+} from './server-log-capture.js';
+import { readServerLogRegion } from './server-log-region.js';
 
 export interface ServerLogFixtures {
    /**
@@ -22,6 +29,21 @@ export interface ServerLogFixtures {
     * `attachServerLog` in `beforeEach` / `afterEach` instead.
     */
    serverLogWorkspace: string;
+   /**
+    * Fail a test whose part of the server log holds a `Warn` or `Error` line
+    * that {@link serverLogAllowlist} does not match. Off by default; reads
+    * nothing when capture is off.
+    */
+   serverLogClean: boolean;
+   /**
+    * The `Warn` and `Error` lines {@link serverLogClean} lets through: each key
+    * is why the line is expected, and its pattern is matched against the whole
+    * line, label and timestamp included. A pattern takes neither the `g` nor
+    * the `y` flag, which make `test` resume from its last match. An object
+    * rather than a list, because Playwright reads an array whose second entry
+    * is an object as a fixture's `[value, options]`.
+    */
+   serverLogAllowlist: Readonly<Record<string, RegExp>>;
 }
 
 /** Full fixture set including the internal side-effect-only auto fixture. */
@@ -36,17 +58,35 @@ interface ServerLogTestFixtures extends ServerLogFixtures {
  * (teardown) — i.e. the `beforeEach` + `afterEach` pair fused into one auto
  * fixture. Compose onto any test base, or use the pre-extended {@link test}.
  * Override `serverLogWorkspace` to wire the workspace; otherwise it's inert.
+ * Set `serverLogClean` to also fail a test on an unexpected `Warn` or `Error`.
  */
 export const serverLogFixtures: Fixtures<ServerLogTestFixtures, object, PlaywrightTestArgs & PlaywrightTestOptions> = {
    serverLogWorkspace: ['', { option: true }],
+   serverLogClean: [false, { option: true }],
+   serverLogAllowlist: [{}, { option: true }],
    serverLogCapture: [
-      async ({ serverLogWorkspace }, use): Promise<void> => {
-         if (serverLogWorkspace) {
-            markServerLog(base.info(), serverLogWorkspace);
+      async ({ serverLogWorkspace, serverLogClean, serverLogAllowlist }, use): Promise<void> => {
+         if (!serverLogWorkspace) {
+            await use(undefined);
+            return;
          }
+         markServerLog(base.info(), serverLogWorkspace);
          await use(undefined);
-         if (serverLogWorkspace) {
-            await attachServerLog(base.info(), serverLogWorkspace);
+         const file = resolveServerLogPath(serverLogWorkspace);
+         const region = file ? readServerLogRegion(file) : '';
+         let problem: string | undefined;
+         if (serverLogClean) {
+            const unexpected = findUnexpectedServerLogLines(region, serverLogAllowlist);
+            if (unexpected.length > 0) {
+               problem = `The server logged ${unexpected.length} unexpected Warn or Error line(s):\n${unexpected.join('\n')}`;
+            }
+         }
+         // The failure-only policy reads the test's status, which failing here
+         // does not change; a run set to attach nothing still attaches nothing.
+         const attachOn = process.env[DEFAULT_SERVER_LOG_ATTACH_ENV] === 'never' ? 'never' : 'always';
+         await attachServerLog(base.info(), serverLogWorkspace, problem ? { attachOn } : {});
+         if (problem) {
+            throw new Error(problem);
          }
       },
       { auto: true }

package/src/testing/playwright/server-log-region.ts

@@ -0,0 +1,32 @@
+/********************************************************************************
+ * Copyright (c) 2026 EclipseSource and others.
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the MIT License which is available in the project root.
+ *
+ * SPDX-License-Identifier: MIT
+ ********************************************************************************/
+
+import { existsSync, readFileSync } from 'node:fs';
+
+/**
+ * How far each log has been read. A suite starts its app before its first
+ * test's marker, so reading from the marker would leave startup unread.
+ */
+const readUpTo = new Map<string, number>();
+
+/**
+ * The part of `file` written since the last read of it, up to its last complete
+ * line, since the server may be mid-write. Every test reads, checked or not, so
+ * a test that is not checked leaves no lines to the next one.
+ */
+export function readServerLogRegion(file: string): string {
+   if (!existsSync(file)) {
+      return '';
+   }
+   const log = readFileSync(file);
+   const end = log.lastIndexOf('\n') + 1;
+   const text = log.subarray(readUpTo.get(file) ?? 0, end).toString('utf-8');
+   readUpTo.set(file, end);
+   return text;
+}

package/src/testing/playwright/server-log-rename-reporter.ts

@@ -24,6 +24,17 @@ export default class ServerLogRenameReporter {
    }
 
    onEnd(): void {
-      renameServerLogs(this.dir);
+      const unattributed = renameServerLogs(this.dir);
+      if (unattributed.length > 0) {
+         // Loud, because the alternative is a capture run that reads as
+         // complete: these logs exist and have content, but are named by an
+         // opaque workspace token and carry no test boundaries, so nothing
+         // about them says which spec they belong to.
+         console.warn(
+            `[server-log] ${unattributed.length} captured log(s) belong to no test: ${unattributed.join(', ')}.\n` +
+               '[server-log] Those specs never marked a boundary — apply the `serverLog` fixture, ' +
+               'or call `markServerLog`/`attachServerLog` from the suite, or their logs cannot be read per test.'
+         );
+      }
    }
 }

package/src/testing/run-update-pipeline.ts

@@ -9,7 +9,7 @@
 
 /**
  * Driver for the structured-write transform chain — the rewrite registry plus
- * the serializer — run the way `ModelService.update` runs it.
+ * the serializer — run the way a session's `update` runs it.
  *
  * # Why a driver and not a documented recipe
  *
@@ -27,7 +27,8 @@
  * # Resolved per URI, not per language handle
  *
  * Both slots are resolved through `ServiceRegistry.getServices(uri)`, which is
- * what `ModelService.serialize` and `ModelService.rewriteModel` do. Taking a
+ * what `DefaultModelService.serialize` and `DefaultModelService.rewriteModel`
+ * do. Taking a
  * language-services handle instead would be shorter, but it re-hardcodes the
  * language at the call site and so cannot witness a multi-grammar workspace
  * routing a write to the wrong language's rewrite set or serializer — the one
@@ -38,7 +39,7 @@
  * This runs rewrite → serialize and stops. It does NOT open the document, push
  * text into the multi-client store, drive a build or return an
  * {@link AstDocument}: those need a live workspace, and a test that wants them
- * should call `ModelService.update` itself rather than a driver standing in for
+ * should call a session's `update` itself rather than a driver standing in for
  * it. What is left is the part with no observable of its own — the text and the
  * rewritten model — which is why it needed lifting.
  *
@@ -61,7 +62,7 @@ export interface UpdatePipelineArgs<TTransfer extends TransferElement = Transfer
     * rather than merely labelling it.
     */
    readonly uri: string;
-   /** Transfer-model payload, as an adopter caller would hand it to `ModelService.update`. */
+   /** Transfer-model payload, as an adopter caller would hand it to a session's `update`. */
    readonly model: TTransfer;
    /**
     * Previous AST root threaded into every rewrite, so a diff-based rewrite can
@@ -90,7 +91,7 @@ export interface UpdatePipelineResult<TTransfer extends TransferElement = Transf
     * surface but it makes the precise assertion unavailable.
     */
    readonly rewritten: TTransfer;
-   /** The serialized text, which is what `ModelService.update` applies to the store. */
+   /** The serialized text, which is what a session's `update` applies to the store. */
    readonly text: string;
 }