@superblocksteam/sdk 2.0.155 → 2.0.156-next.0

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.
Files changed (62) hide show
  1. package/.turbo/turbo-build.log +1 -1
  2. package/dist/cli-replacement/automatic-upgrades.d.ts.map +1 -1
  3. package/dist/cli-replacement/automatic-upgrades.js +4 -8
  4. package/dist/cli-replacement/automatic-upgrades.js.map +1 -1
  5. package/dist/cli-replacement/automatic-upgrades.test.js +2 -1
  6. package/dist/cli-replacement/automatic-upgrades.test.js.map +1 -1
  7. package/dist/cli-replacement/dev-s3-restore.test.mjs +6 -2
  8. package/dist/cli-replacement/dev-s3-restore.test.mjs.map +1 -1
  9. package/dist/cli-replacement/dev-startup-git-before-dbfs-order.test.mjs +6 -2
  10. package/dist/cli-replacement/dev-startup-git-before-dbfs-order.test.mjs.map +1 -1
  11. package/dist/cli-replacement/dev-startup-lock-acquisition.test.d.mts +2 -0
  12. package/dist/cli-replacement/dev-startup-lock-acquisition.test.d.mts.map +1 -0
  13. package/dist/cli-replacement/dev-startup-lock-acquisition.test.mjs +487 -0
  14. package/dist/cli-replacement/dev-startup-lock-acquisition.test.mjs.map +1 -0
  15. package/dist/cli-replacement/dev.d.mts +6 -6
  16. package/dist/cli-replacement/dev.d.mts.map +1 -1
  17. package/dist/cli-replacement/dev.interception.test.mjs +12 -12
  18. package/dist/cli-replacement/dev.interception.test.mjs.map +1 -1
  19. package/dist/cli-replacement/dev.mjs +100 -42
  20. package/dist/cli-replacement/dev.mjs.map +1 -1
  21. package/dist/dev-utils/dev-server.d.mts +34 -1
  22. package/dist/dev-utils/dev-server.d.mts.map +1 -1
  23. package/dist/dev-utils/dev-server.mjs +69 -70
  24. package/dist/dev-utils/dev-server.mjs.map +1 -1
  25. package/dist/dev-utils/dev-server.shutdown-lock.test.d.mts +2 -0
  26. package/dist/dev-utils/dev-server.shutdown-lock.test.d.mts.map +1 -0
  27. package/dist/dev-utils/dev-server.shutdown-lock.test.mjs +48 -0
  28. package/dist/dev-utils/dev-server.shutdown-lock.test.mjs.map +1 -0
  29. package/dist/dev-utils/fatal-process-barrier.d.mts +71 -0
  30. package/dist/dev-utils/fatal-process-barrier.d.mts.map +1 -0
  31. package/dist/dev-utils/fatal-process-barrier.mjs +161 -0
  32. package/dist/dev-utils/fatal-process-barrier.mjs.map +1 -0
  33. package/dist/dev-utils/fatal-process-barrier.spawn-fixture.d.mts +2 -0
  34. package/dist/dev-utils/fatal-process-barrier.spawn-fixture.d.mts.map +1 -0
  35. package/dist/dev-utils/fatal-process-barrier.spawn-fixture.mjs +9 -0
  36. package/dist/dev-utils/fatal-process-barrier.spawn-fixture.mjs.map +1 -0
  37. package/dist/dev-utils/fatal-process-barrier.test.d.mts +2 -0
  38. package/dist/dev-utils/fatal-process-barrier.test.d.mts.map +1 -0
  39. package/dist/dev-utils/fatal-process-barrier.test.mjs +281 -0
  40. package/dist/dev-utils/fatal-process-barrier.test.mjs.map +1 -0
  41. package/dist/telemetry/logging.js +1 -1
  42. package/dist/telemetry/logging.js.map +1 -1
  43. package/dist/telemetry/safe-stringify.d.ts.map +1 -1
  44. package/dist/telemetry/safe-stringify.js +25 -9
  45. package/dist/telemetry/safe-stringify.js.map +1 -1
  46. package/package.json +10 -6
  47. package/src/cli-replacement/automatic-upgrades.test.ts +2 -1
  48. package/src/cli-replacement/automatic-upgrades.ts +4 -11
  49. package/src/cli-replacement/dev-s3-restore.test.mts +6 -2
  50. package/src/cli-replacement/dev-startup-git-before-dbfs-order.test.mts +6 -2
  51. package/src/cli-replacement/dev-startup-lock-acquisition.test.mts +573 -0
  52. package/src/cli-replacement/dev.interception.test.mts +12 -12
  53. package/src/cli-replacement/dev.mts +112 -56
  54. package/src/dev-utils/dev-server.mts +93 -88
  55. package/src/dev-utils/dev-server.shutdown-lock.test.mts +58 -0
  56. package/src/dev-utils/fatal-process-barrier.mts +262 -0
  57. package/src/dev-utils/fatal-process-barrier.spawn-fixture.mts +12 -0
  58. package/src/dev-utils/fatal-process-barrier.test.mts +346 -0
  59. package/src/telemetry/logging.ts +1 -1
  60. package/src/telemetry/safe-stringify.ts +25 -10
  61. package/test/safe-stringify.test.mts +25 -0
  62. package/tsconfig.tsbuildinfo +1 -1
@@ -16,10 +16,10 @@ import type {
16
16
  DependencyUpgradeError,
17
17
  ServerError,
18
18
  } from "@superblocksteam/library-shared/types";
19
+ import type { ConflictError } from "@superblocksteam/shared";
19
20
  import {
20
21
  buildGithubSuperblocksSyncWorkflow,
21
22
  buildGithubSuperblocksSyncWorkflowFromBaseUrl,
22
- ConflictError,
23
23
  DEFAULT_NON_GIT_BRANCH,
24
24
  isGitAuthErrorMessage,
25
25
  isGitHubRemoteUrl,
@@ -176,6 +176,19 @@ async function handleInactiveLockTakeover(
176
176
  throw new Error("Lock takeover prompt shown");
177
177
  }
178
178
 
179
+ // Only ask if somebody can answer. `promptUser` reads stdin, which on a
180
+ // sandbox nothing is attached to — it would wait forever for a keystroke
181
+ // that cannot come. Reachable now that conflicts surface when a client
182
+ // attaches rather than only during a CLI startup.
183
+ if (!process.stdin.isTTY) {
184
+ logger.warn(`${warningMessage} Run with --force-takeover to take it.`);
185
+ span.setStatus({
186
+ code: SpanStatusCode.ERROR,
187
+ message: "Lock held by an inactive user, nobody available to prompt",
188
+ });
189
+ throw new Error(warningMessage);
190
+ }
191
+
179
192
  // Prompt user to force takeover (CLI only)
180
193
  logger.warn(warningMessage);
181
194
  const shouldTakeover = await getBooleanUserResponse(prompt);
@@ -1019,12 +1032,12 @@ export async function flushTelemetryWithTimeout(): Promise<boolean> {
1019
1032
  * MUST BE CALLED LAST in any shutdown sequence. This does not merely flush: it
1020
1033
  * awaits `shutdownTelemetry()`, which tears the provider down, so every signal
1021
1034
  * emitted after it goes into a dead pipeline. It can also block for up to
1022
- * `TELEMETRY_FLUSH_TIMEOUT_MS`, and delaying `lockService.shutdown()` by that
1023
- * much delays the lock release that lets a replacement sandbox take over —
1035
+ * `TELEMETRY_FLUSH_TIMEOUT_MS`, and delaying `lockService.relinquishLock()` by
1036
+ * that much delays the lock release that lets a replacement sandbox take over —
1024
1037
  * which is the entire recovery story this exit path exists to enable.
1025
1038
  *
1026
1039
  * "Last" is only reachable because the callers drive the lock service with
1027
- * `shutdown()` rather than `shutdownAndExit()`. The latter ends in an
1040
+ * `relinquishLock()` rather than `shutdownAndExit()`. The latter ends in an
1028
1041
  * unconditional `finally { process.exit(1) }`, and `process.exit` is
1029
1042
  * synchronous, so anything sequenced after it never runs at all. See the two
1030
1043
  * fatal-exit blocks in `dev()`.
@@ -1075,8 +1088,8 @@ const FATAL_EXIT_STEP_TIMEOUT_MS = 5000;
1075
1088
 
1076
1089
  /**
1077
1090
  * One teardown step. Returning `undefined` is allowed so callers can pass the
1078
- * optional-chained call directly (`() => lockService?.shutdown(...)`) without
1079
- * having to re-narrow a mutable binding at the call site.
1091
+ * optional-chained call directly (`() => lockService?.relinquishLock()`)
1092
+ * without having to re-narrow a mutable binding at the call site.
1080
1093
  */
1081
1094
  type FatalExitStep = () => Promise<unknown> | undefined;
1082
1095
 
@@ -1130,7 +1143,7 @@ async function runFatalExitStep(
1130
1143
  *
1131
1144
  * Order is load-bearing:
1132
1145
  * 1. `removeIntegrationCache` — ai-service state that must not outlive us.
1133
- * 2. `shutdownLockService` — releases the AppBranchLock. Recovery-critical:
1146
+ * 2. `relinquishLock` — hands the AppBranchLock back. Recovery-critical:
1134
1147
  * the heartbeat dies with the process, so skipping this makes the
1135
1148
  * replacement sandbox's `acquireLock()` fail with a ConflictError until
1136
1149
  * the server-side TTL expires, and the replacement is the entire point of
@@ -1145,15 +1158,15 @@ async function runFatalExitStep(
1145
1158
  *
1146
1159
  * This does NOT exit. The caller owns `process.exit(1)` from a `finally`, so
1147
1160
  * the exit still happens even if this function is somehow bypassed. That is
1148
- * also why callers must drive the lock service through `shutdown()` and not
1149
- * `shutdownAndExit()`: the latter ends in its own unconditional
1161
+ * also why callers must drive the lock service through `relinquishLock()` and
1162
+ * not `shutdownAndExit()`: the latter ends in its own unconditional
1150
1163
  * `finally { process.exit(1) }`, and `process.exit` is synchronous, so
1151
1164
  * everything sequenced after it — including the flush — silently never ran.
1152
1165
  */
1153
1166
  export async function runFatalExitShutdown(steps: {
1154
1167
  logger: Logger;
1155
1168
  removeIntegrationCache?: FatalExitStep;
1156
- shutdownLockService?: FatalExitStep;
1169
+ relinquishLock?: FatalExitStep;
1157
1170
  }): Promise<void> {
1158
1171
  const { logger } = steps;
1159
1172
  await runFatalExitStep(
@@ -1161,11 +1174,7 @@ export async function runFatalExitShutdown(steps: {
1161
1174
  "ai-service integration cache removal",
1162
1175
  steps.removeIntegrationCache,
1163
1176
  );
1164
- await runFatalExitStep(
1165
- logger,
1166
- "lock service shutdown",
1167
- steps.shutdownLockService,
1168
- );
1177
+ await runFatalExitStep(logger, "lock hand-back", steps.relinquishLock);
1169
1178
  await flushTelemetryBeforeExit(logger);
1170
1179
  }
1171
1180
 
@@ -1635,35 +1644,84 @@ export async function dev(options: {
1635
1644
 
1636
1645
  logger.info("Checking if local files are synced with the server");
1637
1646
 
1647
+ const lockSvc = lockService;
1648
+
1649
+ // Nobody takes the lock to boot: nothing startup does is gated on
1650
+ // holding one, and a lock taken before any client exists is held on
1651
+ // behalf of nobody. `LockService` takes it when the first client
1652
+ // attaches, and tells us here if somebody else already has it —
1653
+ // this is the `ConflictError` catch that used to wrap the boot-time
1654
+ // acquire, moved to where acquisition actually happens (ENG-6090).
1655
+ lockSvc.setOnLockConflict(async (error) => {
1656
+ let stopForExtension = false;
1657
+ await tracer.startActiveSpan("handlingLockConflict", async (span) => {
1658
+ try {
1659
+ await handleLockConflict(error, lockSvc, span, logger);
1660
+ } catch (conflictError) {
1661
+ // Reporting is all this owes the caller. The holder is someone
1662
+ // else's live session and the pod is still serving preview, so
1663
+ // there is nothing to tear down — the lock does not decide how
1664
+ // long this process lives (ENG-6032).
1665
+ logger.warn(
1666
+ `Could not take the live-edit lock: ${conflictError instanceof Error ? conflictError.message : String(conflictError)}`,
1667
+ );
1668
+ // Except under the VS Code extension, which reads the `cliprompt`
1669
+ // above, asks the user, and answers by restarting this command
1670
+ // with `--force-takeover`. That hand-off has always ended with
1671
+ // this process gone: it used to abort startup before anything
1672
+ // bound a port. Staying up now would leave the replacement
1673
+ // fighting us for :5173.
1674
+ if (process.env.SUPERBLOCKS_VSCODE === "true") {
1675
+ logger.info(
1676
+ "Stopping so the extension can restart with --force-takeover",
1677
+ );
1678
+ stopForExtension = true;
1679
+ }
1680
+ } finally {
1681
+ span.end();
1682
+ }
1683
+ });
1684
+ if (!stopForExtension) {
1685
+ return;
1686
+ }
1687
+ // Outside the span, so `span.end()` above has already run and the
1688
+ // flush below can carry it. This is the same teardown the two
1689
+ // fatal-exit blocks in `dev()` use, and for the same reason: the
1690
+ // `shutdownAndExit()` that used to be inside the `catch` exits
1691
+ // synchronously from its own `finally`, which made both the
1692
+ // `span.end()` and the telemetry flush unreachable — the trace for
1693
+ // the one event this handler exists to record never left the process.
1694
+ try {
1695
+ await runFatalExitShutdown({
1696
+ logger,
1697
+ removeIntegrationCache: () => aiService?.removeIntegrationCache(),
1698
+ relinquishLock: () => lockSvc.relinquishLock(),
1699
+ });
1700
+ } finally {
1701
+ process.exit(1);
1702
+ }
1703
+ });
1704
+
1638
1705
  try {
1639
1706
  await maskUnixSignals(async () => {
1640
- if (lockService) {
1641
- const lockSvc = lockService;
1707
+ // `--force-takeover` is the one thing that still acquires at boot:
1708
+ // it is an explicit "take it from them now" rather than a pod
1709
+ // helping itself to a lock nobody asked it to hold.
1710
+ if (options.forceTakeover) {
1642
1711
  await tracer.startActiveSpan(
1643
- "acquiringInitialLock",
1712
+ "forcingInitialLockTakeover",
1644
1713
  async (span) => {
1645
1714
  try {
1646
- // If --force-takeover flag is set, immediately force takeover instead of acquiring
1647
- if (options.forceTakeover) {
1648
- logger.info(
1649
- "Force takeover requested, taking over lock...",
1650
- );
1651
- await lockSvc.forceTakeover();
1652
- logger.info(
1653
- "Lock acquired successfully via force takeover.",
1654
- );
1655
- } else {
1656
- await lockSvc.acquireLock();
1657
- }
1715
+ logger.info(
1716
+ "Force takeover requested, taking over lock...",
1717
+ );
1718
+ await lockSvc.forceTakeover();
1719
+ logger.info(
1720
+ "Lock acquired successfully via force takeover.",
1721
+ );
1658
1722
  } catch (error) {
1659
- // If we got a conflict error (409), inspect the lock for more information
1660
- if (error instanceof ConflictError) {
1661
- await handleLockConflict(error, lockSvc, span, logger);
1662
- return;
1663
- }
1664
-
1665
1723
  logger.error(
1666
- "Failed to acquire lock on application",
1724
+ "Failed to force takeover of lock on application",
1667
1725
  getErrorMeta(error),
1668
1726
  );
1669
1727
  span.setStatus({
@@ -2337,7 +2395,7 @@ export async function dev(options: {
2337
2395
  try {
2338
2396
  logger.info("Releasing lock before restarting the dev server");
2339
2397
  await aiService?.removeIntegrationCache();
2340
- await lockService?.shutdown({ serverInitiated: false });
2398
+ await lockService?.relinquishLock();
2341
2399
  } catch (e) {
2342
2400
  logger.error(
2343
2401
  "Error releasing lock before restarting the dev server",
@@ -2414,13 +2472,13 @@ export async function dev(options: {
2414
2472
  }),
2415
2473
  );
2416
2474
  try {
2417
- // `shutdown()`, NOT `shutdownAndExit()`. The latter ends in an
2418
- // unconditional `finally { process.exit(1) }`; `process.exit` is
2419
- // synchronous, so every statement below it would be dead code
2475
+ // `relinquishLock()`, NOT `shutdownAndExit()`. The latter ends in
2476
+ // an unconditional `finally { process.exit(1) }`; `process.exit`
2477
+ // is synchronous, so every statement below it would be dead code
2420
2478
  // whenever `lockService` is set — which is the normal state in a
2421
2479
  // SABS pod, exactly the population this exit path targets. The
2422
- // telemetry flush was silently never running. `shutdown()` does
2423
- // the same teardown without owning the exit, matching how
2480
+ // telemetry flush was silently never running. `relinquishLock()`
2481
+ // does the same teardown without owning the exit, matching how
2424
2482
  // `runGracefulShutdown` in dev-server.mts drives it; the
2425
2483
  // `finally` below performs the exit instead.
2426
2484
  //
@@ -2431,8 +2489,7 @@ export async function dev(options: {
2431
2489
  logger,
2432
2490
  removeIntegrationCache: () =>
2433
2491
  aiService?.removeIntegrationCache(),
2434
- shutdownLockService: () =>
2435
- lockService?.shutdown({ serverInitiated: false }),
2492
+ relinquishLock: () => lockService?.relinquishLock(),
2436
2493
  });
2437
2494
  } finally {
2438
2495
  // The only exit on this path now that the lock service no longer
@@ -2489,15 +2546,14 @@ export async function dev(options: {
2489
2546
  }),
2490
2547
  );
2491
2548
  try {
2492
- // `shutdown()`, NOT `shutdownAndExit()` — same reason as the
2549
+ // `relinquishLock()`, NOT `shutdownAndExit()` — same reason as the
2493
2550
  // sync/lock catch above: `shutdownAndExit()` exits the process
2494
2551
  // synchronously in its own `finally`, which made everything
2495
2552
  // sequenced after it unreachable.
2496
2553
  await runFatalExitShutdown({
2497
2554
  logger,
2498
2555
  removeIntegrationCache: () => aiService?.removeIntegrationCache(),
2499
- shutdownLockService: () =>
2500
- lockService?.shutdown({ serverInitiated: false }),
2556
+ relinquishLock: () => lockService?.relinquishLock(),
2501
2557
  });
2502
2558
  } finally {
2503
2559
  // The only exit on this path; here so a thrown shutdown path can't
@@ -2695,7 +2751,8 @@ export async function dev(options: {
2695
2751
  logger.warn(`Error removing integration cache: ${error}`);
2696
2752
  });
2697
2753
  lockService?.shutdownAndExit().catch(() => {
2698
- // this is redundant, but it's here to make sure the lock service is shutdown and the process exits
2754
+ // Redundant `shutdownAndExit` exits from its own `finally` but
2755
+ // kept so the process still goes away if that ever stops being true.
2699
2756
  process.exit(1);
2700
2757
  });
2701
2758
  });
@@ -3094,7 +3151,7 @@ async function ensureRuntimeDbfsBranchConsistency({
3094
3151
  currentBranchName: string;
3095
3152
  }): Promise<string> {
3096
3153
  interface BranchSwitchingLockService {
3097
- isLocked: boolean;
3154
+ isLockedWhenSettled(): Promise<boolean>;
3098
3155
  switchBranch(
3099
3156
  nextBranchName: string,
3100
3157
  options?: { reacquireLock?: boolean },
@@ -3124,7 +3181,8 @@ async function ensureRuntimeDbfsBranchConsistency({
3124
3181
 
3125
3182
  const lockSvc = lockService as unknown as BranchSwitchingLockService;
3126
3183
  const syncSvc = syncService as unknown as BranchSwitchingSyncService;
3127
- const wasLocked = lockSvc.isLocked;
3184
+ // Retain the pre-switch state for explicit fallback and rollback paths.
3185
+ const wasLocked = await lockSvc.isLockedWhenSettled();
3128
3186
  logger.info(
3129
3187
  `Switching runtime DBFS branch context from '${currentBranchName}' to '${targetBranchName}'`,
3130
3188
  );
@@ -3132,14 +3190,12 @@ async function ensureRuntimeDbfsBranchConsistency({
3132
3190
  try {
3133
3191
  syncSvc.setBranchName(targetBranchName);
3134
3192
  try {
3135
- await lockSvc.switchBranch(targetBranchName, {
3136
- reacquireLock: wasLocked,
3137
- });
3193
+ await lockSvc.switchBranch(targetBranchName);
3138
3194
  } catch (switchError) {
3139
3195
  // Disconnecting native git can invalidate the previous live branch
3140
3196
  // context. Fall back to explicit release/reacquire on the target branch.
3141
3197
  logger.warn(
3142
- `Lock branch transfer failed for '${currentBranchName}' -> '${targetBranchName}', performing shutdown-style lock reset: ${switchError instanceof Error ? switchError.message : String(switchError)}`,
3198
+ `Lock branch transfer failed for '${currentBranchName}' -> '${targetBranchName}', performing an explicit release and reacquire: ${switchError instanceof Error ? switchError.message : String(switchError)}`,
3143
3199
  );
3144
3200
  await lockSvc.switchBranch(targetBranchName, { reacquireLock: false });
3145
3201
  if (wasLocked) {
@@ -3156,7 +3212,7 @@ async function ensureRuntimeDbfsBranchConsistency({
3156
3212
  syncSvc.setBranchName(currentBranchName);
3157
3213
  try {
3158
3214
  await lockSvc.switchBranch(currentBranchName, {
3159
- reacquireLock: lockSvc.isLocked || wasLocked,
3215
+ reacquireLock: (await lockSvc.isLockedWhenSettled()) || wasLocked,
3160
3216
  });
3161
3217
  } catch (rollbackError) {
3162
3218
  logger.warn(
@@ -69,11 +69,12 @@ import {
69
69
  type DevServerFailureType,
70
70
  devServerMetrics,
71
71
  } from "./dev-server-metrics.mjs";
72
+ import { isRecoverableShutdownError, parseIsWarm } from "./fatal-exit.mjs";
72
73
  import {
73
- buildFatalExitLog,
74
- isRecoverableShutdownError,
75
- parseIsWarm,
76
- } from "./fatal-exit.mjs";
74
+ fatalAwareExitCode,
75
+ installFatalProcessBarrier,
76
+ type FatalProcessHandler,
77
+ } from "./fatal-process-barrier.mjs";
77
78
  import { OPTIMIZE_DEPS_CONFIG } from "./optimize-deps-config.mjs";
78
79
  import { prepareInPlaceRestart } from "./prepare-in-place-restart.mjs";
79
80
  import { handleRecoverableBuildError } from "./recoverable-build-error.mjs";
@@ -546,7 +547,7 @@ export function buildStatusPayload<T extends object>(
546
547
  return { ...base, serverErrors: devServerStatus?.serverErrors ?? [] };
547
548
  }
548
549
 
549
- type ProcessErrorHandler = "uncaughtException" | "unhandledRejection";
550
+ type ProcessErrorHandler = FatalProcessHandler;
550
551
 
551
552
  // Both listeners log a recoverable shutdown, but with different wording, and
552
553
  // these lines are what on-call greps when a pod stays up instead of restarting.
@@ -557,6 +558,59 @@ const RECOVERABLE_SHUTDOWN_LOG_MESSAGE: Record<ProcessErrorHandler, string> = {
557
558
  "Ignoring recoverable shutdown rejection (unhandledRejection)",
558
559
  };
559
560
 
561
+ /**
562
+ * What the dev server needs of the lock service on its way out. Structural so a
563
+ * test does not have to stand up a real one.
564
+ */
565
+ type ExitingLockService = Pick<
566
+ LockService,
567
+ "notifyClosedByServer" | "relinquishLock"
568
+ >;
569
+
570
+ /**
571
+ * Give the AppBranchLock to whoever should have it next, then let the caller
572
+ * exit.
573
+ *
574
+ * Which of the two things to do is the dev server's knowledge, not the lock
575
+ * service's — it used to be a `serverInitiated` flag passed down into a
576
+ * `shutdown()` that branched on it, which is how a lock service came to have an
577
+ * opinion about why a process was stopping.
578
+ *
579
+ * Normally this process is the one stopping, so the lock goes back: the
580
+ * heartbeat dies with the process, and anything not released here makes the
581
+ * replacement sandbox's `acquireLock()` fail until the server-side TTL expires
582
+ * — and the replacement is the entire point of exiting.
583
+ *
584
+ * When the server closed us, the lock is already gone. `/_sb_disconnect` is
585
+ * usually a session moving across local↔cloud, so releasing would take the lock
586
+ * from whoever is taking over. The editor is told instead, because it has to
587
+ * distinguish "the server closed you" from a socket dropping.
588
+ *
589
+ * Extracted so both branches are testable: `runGracefulShutdown` is a closure
590
+ * over a booted dev server, so a unit test cannot otherwise reach either one.
591
+ */
592
+ export async function handOverLockOnExit(
593
+ lockService: ExitingLockService | undefined,
594
+ {
595
+ serverInitiated,
596
+ switchingTo,
597
+ initiatedByEmail,
598
+ }: {
599
+ serverInitiated: boolean;
600
+ switchingTo?: "local" | "cloud" | "none";
601
+ initiatedByEmail?: string;
602
+ },
603
+ ): Promise<void> {
604
+ if (!lockService) {
605
+ return;
606
+ }
607
+ if (serverInitiated) {
608
+ await lockService.notifyClosedByServer({ switchingTo, initiatedByEmail });
609
+ return;
610
+ }
611
+ await lockService.relinquishLock();
612
+ }
613
+
560
614
  /**
561
615
  * Decides whether a process-level error keeps the pod alive, and performs the
562
616
  * recording for the cases that do. Returns `true` when the caller must return
@@ -871,7 +925,7 @@ export async function createDevServer({
871
925
  }): Promise<void> {
872
926
  try {
873
927
  stopMemoryMetrics();
874
- await lockService?.shutdown({
928
+ await handOverLockOnExit(lockService, {
875
929
  serverInitiated,
876
930
  switchingTo,
877
931
  initiatedByEmail,
@@ -882,16 +936,23 @@ export async function createDevServer({
882
936
  } catch (e) {
883
937
  logger.error(`Error during graceful shutdown: ${e}`);
884
938
  } finally {
939
+ // The statement about this process, made by the thing that owns its exit.
940
+ // It used to be the last line of `LockService.shutdown()`, prefixed
941
+ // `[lock-service]` — a lock service announcing a process exit, which is
942
+ // the tangle this change removes.
943
+ //
944
+ // Load bearing: cli-system-tests waits on this exact substring.
945
+ logger.info("[dev-server] shutting down and exiting dev server");
885
946
  const closeTimeout = setTimeout(() => {
886
947
  logger.info("Dev server closed");
887
- process.exit(0);
948
+ process.exit(fatalAwareExitCode());
888
949
  }, 1000); // Fallback to exit if the HTTP server doesn't close within 1 second
889
950
 
890
951
  // Then close the HTTP server
891
952
  httpServer.close(() => {
892
953
  logger.info("Dev server closed");
893
954
  clearTimeout(closeTimeout);
894
- process.exit(0);
955
+ process.exit(fatalAwareExitCode());
895
956
  });
896
957
  }
897
958
  }
@@ -1266,18 +1327,11 @@ export async function createDevServer({
1266
1327
  // even if the pod is activated later.
1267
1328
  const isWarm = parseIsWarm(process.env.SUPERBLOCKS_WARM_STANDBY);
1268
1329
 
1269
- // Only the first fatal event writes a fatal-exit line and starts shutdown.
1270
- // Without this, one crash could write several lines: an unhandled rejection
1271
- // fires once per rejected promise, so a burst of failures (say a dead DB pool
1272
- // rejecting every in-flight query) would each write a line and each start
1273
- // shutdown, making one crash look like many in the logs and the crash count.
1274
- let fatalExitHandled = false;
1275
-
1276
1330
  // Signal handlers attach `.catch` so a synchronous throw in
1277
1331
  // `runGracefulShutdown` (e.g. logger init or lockService shutdown throwing
1278
1332
  // before the first `await`) is logged rather than surfacing as an
1279
1333
  // unhandled rejection across N handlers (one per fired signal). Symmetric
1280
- // with the `uncaughtException` handler below.
1334
+ // with the process barrier below.
1281
1335
  process.on("SIGINT", () => {
1282
1336
  logger.info("SIGINT received");
1283
1337
  gracefulShutdown({
@@ -1311,86 +1365,37 @@ export async function createDevServer({
1311
1365
  );
1312
1366
  });
1313
1367
 
1314
- // The `.finally` always ends the process once shutdown settles. If an earlier
1315
- // signal already ran shutdown, this handler still needs to exit; without the
1316
- // final `process.exit` the process could hang instead of stopping.
1317
- process.on("uncaughtException", (error) => {
1318
- if (fatalExitHandled) {
1319
- return;
1320
- }
1321
- if (
1368
+ // Preload may already have armed this from a separate module instance
1369
+ // (node_modules export vs this esbuild bundle). A second call updates
1370
+ // shutdown hooks on the shared host state and does not register another
1371
+ // pair of listeners.
1372
+ installFatalProcessBarrier({
1373
+ isNonFatal: (handler, reason) =>
1322
1374
  handleNonFatalProcessError({
1323
- devServerStatus,
1324
1375
  devServerMetrics,
1325
- handler: "uncaughtException",
1326
- logger,
1327
- reason: error,
1328
- })
1329
- ) {
1330
- return;
1331
- }
1332
- fatalExitHandled = true;
1333
- // Write the fatal-exit line first, and synchronously, so the reason reaches
1334
- // the logs before anything else runs. Shutdown below may finish first and
1335
- // call `process.exit(0)`, so the container's real exit code can be 0 even
1336
- // though this was a crash. That is why `exit_code=1` here is the code this
1337
- // handler means to use, not a promise of what the container reports: for the
1338
- // JS-handler lines, trust that the line exists (and the `signal` field) over
1339
- // its `exit_code` when comparing against the pod's exit status.
1340
- logger.error(
1341
- buildFatalExitLog({ handler: "uncaughtException", exitCode: 1, isWarm }),
1342
- getErrorMeta(error),
1343
- );
1344
- gracefulShutdown({
1345
- logger,
1346
- serverInitiated: false,
1347
- source: "uncaughtException",
1348
- })
1349
- .catch((shutdownError) => {
1350
- logger.error(
1351
- "Error during shutdown after uncaught exception:",
1352
- getErrorMeta(shutdownError),
1353
- );
1354
- })
1355
- .finally(() => process.exit(1));
1356
- });
1357
-
1358
- // Without its own listener, Node turns an unhandled promise rejection into an
1359
- // uncaught exception, which would log it under the wrong cause. Handle it here
1360
- // so a rejection is labelled as a rejection, then exit through the same
1361
- // shutdown path as an uncaught exception.
1362
- process.on("unhandledRejection", (reason) => {
1363
- if (fatalExitHandled) {
1364
- return;
1365
- }
1366
- if (
1367
- handleNonFatalProcessError({
1368
1376
  devServerStatus,
1369
- devServerMetrics,
1370
- handler: "unhandledRejection",
1377
+ handler,
1371
1378
  logger,
1372
1379
  reason,
1373
- })
1374
- ) {
1375
- return;
1376
- }
1377
- fatalExitHandled = true;
1378
- logger.error(
1379
- buildFatalExitLog({ handler: "unhandledRejection", exitCode: 1, isWarm }),
1380
- getErrorMeta(reason),
1381
- );
1382
- gracefulShutdown({
1383
- logger,
1384
- serverInitiated: false,
1385
- source: "unhandledRejection",
1386
- })
1387
- .catch((shutdownError) => {
1380
+ }),
1381
+ isWarm,
1382
+ log: (message, meta) => logger.error(message, meta),
1383
+ onFatal: async (handler) => {
1384
+ try {
1385
+ await gracefulShutdown({
1386
+ logger,
1387
+ serverInitiated: false,
1388
+ source: handler,
1389
+ });
1390
+ } catch (shutdownError) {
1388
1391
  logger.error(
1389
- "Error during shutdown after unhandled rejection:",
1392
+ handler === "uncaughtException"
1393
+ ? "Error during shutdown after uncaught exception:"
1394
+ : "Error during shutdown after unhandled rejection:",
1390
1395
  getErrorMeta(shutdownError),
1391
1396
  );
1392
- })
1393
- .finally(() => process.exit(1));
1397
+ }
1398
+ },
1394
1399
  });
1395
1400
 
1396
1401
  if (existingServer) {
@@ -0,0 +1,58 @@
1
+ import { describe, expect, it, vi } from "vitest";
2
+
3
+ import { handOverLockOnExit } from "./dev-server.mjs";
4
+
5
+ /**
6
+ * ENG-6128. The dev server's exit path used to drive the lock service through a
7
+ * `shutdown({ serverInitiated })` that decided, inside the lock service, whether
8
+ * to release. That flag is the dev server's knowledge, not the lock service's,
9
+ * and getting it wrong in either direction is a live-edit outage: releasing a
10
+ * lock the server has already reassigned takes the app away from the session
11
+ * taking over, and failing to release on a normal exit makes the replacement
12
+ * sandbox wait out the TTL.
13
+ */
14
+ describe("handOverLockOnExit", () => {
15
+ const makeLockService = () => ({
16
+ notifyClosedByServer: vi.fn(async () => undefined),
17
+ relinquishLock: vi.fn(async () => undefined),
18
+ });
19
+
20
+ it("hands the lock back when the dev server is the one stopping", async () => {
21
+ const lockService = makeLockService();
22
+
23
+ await handOverLockOnExit(lockService, { serverInitiated: false });
24
+
25
+ // Prompt release is the whole recovery story for this path: the heartbeat
26
+ // dies with the process, so anything not released here blocks the
27
+ // replacement sandbox for the two-minute TTL.
28
+ expect(lockService.relinquishLock).toHaveBeenCalledOnce();
29
+ expect(lockService.notifyClosedByServer).not.toHaveBeenCalled();
30
+ });
31
+
32
+ it("only notifies when the server closed the session", async () => {
33
+ const lockService = makeLockService();
34
+
35
+ await handOverLockOnExit(lockService, {
36
+ serverInitiated: true,
37
+ switchingTo: "local",
38
+ initiatedByEmail: "someone@example.com",
39
+ });
40
+
41
+ // The server has already taken the lock — on a local↔cloud switch it now
42
+ // belongs to the session taking over, so releasing would take it away from
43
+ // them.
44
+ expect(lockService.relinquishLock).not.toHaveBeenCalled();
45
+ expect(lockService.notifyClosedByServer).toHaveBeenCalledWith({
46
+ switchingTo: "local",
47
+ initiatedByEmail: "someone@example.com",
48
+ });
49
+ });
50
+
51
+ it("is a no-op when the dev server never had a lock service", async () => {
52
+ // Startup can fail before the lock service exists, and the exit path still
53
+ // runs.
54
+ await expect(
55
+ handOverLockOnExit(undefined, { serverInitiated: false }),
56
+ ).resolves.toBeUndefined();
57
+ });
58
+ });