@bitkyc08/opencodex 2.100.4 → 2.100.5-preview.20261118.1

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 (102) hide show
  1. package/gui/dist/assets/{App-DcpyI78f.js → App-BiOGtxDS.js} +21 -21
  2. package/gui/dist/assets/{Tray-CzjPJprT.js → Tray-BwsMejUO.js} +1 -1
  3. package/gui/dist/assets/{brand-mark-BTIi7e69.js → brand-mark-BY1dgACL.js} +1 -1
  4. package/gui/dist/assets/index-CwHSKMdG.js +138 -0
  5. package/gui/dist/assets/{segment-thumb-features-Bn6Y-g3R.js → segment-thumb-features-Bds9Hb_7.js} +1 -1
  6. package/gui/dist/classic/assets/{App-eYJAyNHM.js → App-CGLOiqQL.js} +20 -20
  7. package/gui/dist/classic/assets/{Tray-xHMYxLkV.js → Tray-Da4r2O70.js} +1 -1
  8. package/gui/dist/classic/assets/index-CvoKXKL9.js +138 -0
  9. package/gui/dist/classic/assets/{quota-summary-DyzHlkFl.js → quota-summary-DvqzfCmr.js} +1 -1
  10. package/gui/dist/classic/index.html +1 -1
  11. package/gui/dist/index.html +1 -1
  12. package/native/ocx-child/src/link_ws.rs +24 -19
  13. package/native/ocx-child/src/link_ws_frame.rs +136 -50
  14. package/package.json +3 -3
  15. package/src/adapters/anthropic-input-schema.ts +270 -0
  16. package/src/adapters/anthropic.ts +1 -63
  17. package/src/adapters/command-code-tool-text.ts +9 -1
  18. package/src/adapters/command-code.ts +18 -2
  19. package/src/adapters/devin/cloud-direct/chat.ts +24 -0
  20. package/src/adapters/devin/live-models.ts +11 -1
  21. package/src/adapters/kiro/reasoning.ts +3 -0
  22. package/src/adapters/kiro-tools.ts +14 -9
  23. package/src/adapters/openai-responses/passthrough.ts +3 -1
  24. package/src/adapters/openai-responses/request-strips.ts +1 -0
  25. package/src/adapters/tool-catalog-nudge.ts +1 -1
  26. package/src/claude/chat-tool-dispatch.ts +4 -1
  27. package/src/cli/logout-command.ts +23 -4
  28. package/src/cli/provider.ts +3 -2
  29. package/src/client/connect.ts +10 -2
  30. package/src/clients/config-export.ts +12 -7
  31. package/src/codex/account-lifecycle.ts +3 -1
  32. package/src/codex/catalog/metadata.ts +8 -0
  33. package/src/codex/catalog/native-models.ts +16 -6
  34. package/src/codex/catalog/pinned-models.ts +1 -0
  35. package/src/codex/catalog/provider-models.ts +5 -5
  36. package/src/codex/catalog-auto-refresh.ts +13 -2
  37. package/src/codex/config-drift-heal.ts +40 -12
  38. package/src/codex/config-write-lock.ts +39 -1
  39. package/src/codex/data/roster-pinned-models.json +346 -1
  40. package/src/codex/inject/config-write-section.ts +39 -1
  41. package/src/codex/inject/plan.ts +10 -8
  42. package/src/codex/inject/remove.ts +21 -19
  43. package/src/codex/inject/restore.ts +5 -1
  44. package/src/codex/inject-coordination.ts +58 -7
  45. package/src/codex/inject.ts +10 -12
  46. package/src/codex/journal-config-restore.ts +42 -0
  47. package/src/codex/journal.ts +62 -8
  48. package/src/codex/main-account.ts +7 -2
  49. package/src/codex/native-main-credentials.ts +11 -5
  50. package/src/combos/failover.ts +183 -0
  51. package/src/config/subagent-models.ts +20 -12
  52. package/src/generated/compatibility-version.json +99 -83
  53. package/src/generated/model-metadata.ts +3 -3
  54. package/src/lib/client-error-body.ts +6 -0
  55. package/src/lib/developer-role-rejection.ts +15 -0
  56. package/src/lib/upstream-retry.ts +1 -1
  57. package/src/messaging/remote-ports.ts +22 -10
  58. package/src/oauth/credential-siblings.ts +32 -0
  59. package/src/oauth/generic-account-failover.ts +25 -13
  60. package/src/oauth/index.ts +22 -14
  61. package/src/oauth/store.ts +38 -7
  62. package/src/providers/anthropic-reset-grants.ts +4 -0
  63. package/src/providers/claude-cli-usage.ts +2 -0
  64. package/src/providers/codebuddy-models.ts +11 -0
  65. package/src/providers/google-vertex-model-discovery.ts +43 -14
  66. package/src/providers/kiro-models.ts +5 -0
  67. package/src/providers/registry/entries-core.ts +9 -3
  68. package/src/providers/registry/entries-extended.ts +20 -9
  69. package/src/providers/registry/model-seeds.ts +4 -1
  70. package/src/providers/stale-context-window-migration.ts +2 -0
  71. package/src/responses/citation-markers.ts +26 -23
  72. package/src/server/management/link-enrollment.ts +3 -3
  73. package/src/server/management/link-route-request.ts +11 -5
  74. package/src/server/management/oauth-account-routes.ts +21 -8
  75. package/src/server/management/provider-credential-release.ts +44 -0
  76. package/src/server/management/provider-routes.ts +8 -8
  77. package/src/server/management/usage-aggregate-cache.ts +6 -3
  78. package/src/server/messages-native.ts +16 -2
  79. package/src/server/request-log-conversation.ts +6 -2
  80. package/src/server/request-log-failure-detail.ts +26 -9
  81. package/src/server/responses/adapter-continuation.ts +59 -25
  82. package/src/server/responses/adapter-dispatch.ts +8 -8
  83. package/src/server/responses/codex-tool-catalog.ts +47 -10
  84. package/src/server/responses/combo-session-recall.ts +26 -1
  85. package/src/server/responses/combo-stream-preflight.ts +35 -26
  86. package/src/server/responses/core-combo-failure.ts +6 -0
  87. package/src/server/responses/core-combo-native.ts +2 -1
  88. package/src/server/responses/core-combo.ts +43 -19
  89. package/src/server/responses/core-options.ts +2 -0
  90. package/src/server/responses/input-admission.ts +41 -3
  91. package/src/server/responses/request-send-budget.ts +1 -1
  92. package/src/server/responses/request-transport.ts +10 -2
  93. package/src/server/responses/sidecar-execution.ts +6 -4
  94. package/src/service/windows-checked-stop-live.ts +28 -5
  95. package/src/types/request.ts +4 -3
  96. package/src/usage/expected-prices.ts +35 -0
  97. package/src/usage/ledger-scanner.ts +44 -10
  98. package/src/web-search/loop.ts +20 -2
  99. package/src/web-search/parse.ts +2 -12
  100. package/src/web-search/sources.ts +13 -0
  101. package/gui/dist/assets/index-DVgU6S3y.js +0 -138
  102. package/gui/dist/classic/assets/index-DSaVRW4F.js +0 -138
@@ -120,7 +120,7 @@ import {
120
120
  import type { CodexAuthContext } from "../../codex/auth-context";
121
121
  import type { ResponsesTerminalStatus } from "../../bridge";
122
122
  import { beginRequestAttempt, sealRequestAttemptIdentity, finishRequestAttempt } from "../request-log";
123
- import { rememberComboForLane } from "./combo-session-recall";
123
+ import { recallComboTurnEffortForLane, rememberComboForLane, type ComboTurnEffort } from "./combo-session-recall";
124
124
  import { COMBO_TOOL_LOOP_PIN_REASON, toolLoopPinnedPick } from "./combo-tool-loop-pin";
125
125
  import { JEV_SESSION_PIN_ROUTE_REASON, jevSessionPinScope } from "./jev-session-pin";
126
126
  import { runTurnAdapterSseResponses } from "./core-lifetime";
@@ -716,6 +716,17 @@ export async function executeComboResponses(
716
716
  const jevPin = options.compactionRoutingOverride || options.mainHardLockFallbackOverride
717
717
  ? undefined
718
718
  : jevSessionPinScope({ combo, comboId, admission: options.admission, lane: sessionLaneIdFromRequest(req.headers) });
719
+ // A tool-loop continuation never asks JEV again, so under `decide` it keeps the effort the
720
+ // decision set for its turn while the target still offers it. Without this the rest of the
721
+ // turn ran at the caller's effort and only the first request at the decided one.
722
+ let toolLoopTurnEffort: ComboTurnEffort | undefined;
723
+ if (toolLoopPin && combo.strategy === "jev" && combo.jevEffortMode !== "client") {
724
+ const carried = recallComboTurnEffortForLane(sessionLaneIdFromRequest(req.headers), comboId, toolLoopPin.target);
725
+ const offered = carried ? jevChoiceEfforts(config, toolLoopPin.target, false)?.reasoningEfforts : undefined;
726
+ if (carried && offered && (carried.effort === null ? offered.length === 0 : offered.includes(carried.effort))) {
727
+ toolLoopTurnEffort = carried;
728
+ }
729
+ }
719
730
 
720
731
  if (unreadableEncryptedAgentTask && !pick) {
721
732
  pick = await pickWithWait({
@@ -1113,10 +1124,14 @@ export async function executeComboResponses(
1113
1124
  modelId: targetRoute.modelId,
1114
1125
  });
1115
1126
  const initialJevDecision = firstComboTarget ? jevDecision : undefined;
1127
+ // The effort a decision owns on this dispatch: its own first request, or the turn a
1128
+ // tool-loop continuation resumes. Failover targets keep ordinary Combo policy either way.
1129
+ const turnEffort: ComboTurnEffort | undefined = initialJevDecision
1130
+ ?? (firstComboTarget ? toolLoopTurnEffort : undefined);
1116
1131
  // `client` keeps the decision's target but not its tier: the caller and the target ladder
1117
1132
  // own the wire effort, so a provider whose effort swaps the upstream model id is never
1118
1133
  // re-tiered mid-session and its prompt cache survives the turn.
1119
- const jevEffortApplies = initialJevDecision !== undefined && combo.jevEffortMode !== "client";
1134
+ const jevEffortApplies = turnEffort !== undefined && combo.jevEffortMode !== "client";
1120
1135
  const childInput = options.droidDefaultEffort && isPlainObject(body) ? { ...body } : body;
1121
1136
  applyDroidResponsesReasoningDefault(childInput, options.droidDefaultEffort, {
1122
1137
  provider: targetRoute.provider,
@@ -1125,21 +1140,21 @@ export async function executeComboResponses(
1125
1140
  const childBody = concreteComboRequestBody(
1126
1141
  childInput,
1127
1142
  pick.target,
1128
- jevEffortApplies ? initialJevDecision.effort : comboDefaultEffort(config, comboId),
1129
- jevEffortApplies && initialJevDecision.effort === null ? [] : targetReasoningEfforts,
1143
+ jevEffortApplies ? turnEffort.effort : comboDefaultEffort(config, comboId),
1144
+ jevEffortApplies && turnEffort.effort === null ? [] : targetReasoningEfforts,
1130
1145
  combo.reasoningEffortMode,
1131
1146
  jevEffortApplies
1132
- ? (initialJevDecision.effort === null ? "fallback" : "force")
1147
+ ? (turnEffort.effort === null ? "fallback" : "force")
1133
1148
  : combo.defaultEffortMode,
1134
1149
  );
1135
- if (initialJevDecision) {
1150
+ if (turnEffort) {
1136
1151
  if (jevEffortApplies) delete childBody.service_tier;
1137
- if (jevEffortApplies && initialJevDecision.effort !== null) {
1152
+ if (jevEffortApplies && turnEffort.effort !== null) {
1138
1153
  const childReasoning = childBody.reasoning;
1139
1154
  const preservedReasoning = childReasoning && typeof childReasoning === "object" && !Array.isArray(childReasoning)
1140
1155
  ? childReasoning as Record<string, unknown>
1141
1156
  : {};
1142
- childBody.reasoning = { ...preservedReasoning, effort: initialJevDecision.effort };
1157
+ childBody.reasoning = { ...preservedReasoning, effort: turnEffort.effort };
1143
1158
  delete childBody.reasoning_effort;
1144
1159
  delete childBody.thinking_budget;
1145
1160
  delete childBody.thinking;
@@ -1173,6 +1188,10 @@ export async function executeComboResponses(
1173
1188
  childLog.activeAttemptStartedAt = started;
1174
1189
  childLog.attempts = logCtx.attempts ??= [];
1175
1190
  childLog.attempts.push(attempt);
1191
+ // A key or account rotation inside this target seals the previous segment as its own attempt
1192
+ // and moves `childLog.activeAttemptStartedAt` forward (noteProviderAttemptSend). Timing the
1193
+ // live segment from the target's start would count every sealed segment a second time.
1194
+ const segmentStartedAt = (): number => childLog.activeAttemptStartedAt ?? started;
1176
1195
  let attemptRetained = false;
1177
1196
  const retainCancelledAttempt = (): void => {
1178
1197
  if (attemptRetained) return;
@@ -1182,7 +1201,7 @@ export async function executeComboResponses(
1182
1201
  childLog.providerAdapter ?? attempt.adapter,
1183
1202
  childLog.accountLogLabel,
1184
1203
  );
1185
- finishRequestAttempt(attempt, 499, Date.now() - started, childLog.usage);
1204
+ finishRequestAttempt(attempt, 499, Date.now() - segmentStartedAt(), childLog.usage);
1186
1205
  attemptRetained = true;
1187
1206
  };
1188
1207
  // The row is written from `logCtx`, which copies `childLog` only when the child commits. A
@@ -1212,7 +1231,10 @@ export async function executeComboResponses(
1212
1231
  && Object.hasOwn(config.providers, completedTarget.provider)
1213
1232
  && provider && provider.disabled !== true
1214
1233
  && currentCombo?.targets.some(target => targetKey(target) === targetKey(completedTarget))) {
1215
- rememberComboForLane(sessionLaneIdFromRequest(req.headers), comboId, completedTarget, model, writerGeneration);
1234
+ rememberComboForLane(
1235
+ sessionLaneIdFromRequest(req.headers), comboId, completedTarget, model, writerGeneration,
1236
+ jevEffortApplies ? turnEffort : undefined,
1237
+ );
1216
1238
  jevPin?.complete(targetKey(completedTarget), initialJevDecision, writerGeneration);
1217
1239
  }
1218
1240
  options.onResponseComplete?.(model);
@@ -1243,7 +1265,7 @@ export async function executeComboResponses(
1243
1265
  && targetEligible(target)
1244
1266
  && !isComboTargetInCooldown(comboId, target),
1245
1267
  );
1246
- const nativeChild = protocolLanes?.nativeChild(pick.target, targetRoute, targetSendBudget, initialJevDecision, jevEffortApplies);
1268
+ const nativeChild = protocolLanes?.nativeChild(pick.target, targetRoute, targetSendBudget, turnEffort, jevEffortApplies);
1247
1269
  response = nativeChild ? await dispatchNativeComboChild({
1248
1270
  source: options.protocolSource!,
1249
1271
  plan: nativeChild,
@@ -1256,7 +1278,7 @@ export async function executeComboResponses(
1256
1278
  // Attempt-relative TTFT, recorded here for the same reason as the bridge child below.
1257
1279
  onFirstOutput: () => {
1258
1280
  if (attempt.firstOutputMs === undefined) {
1259
- attempt.firstOutputMs = Math.max(0, Date.now() - started);
1281
+ attempt.firstOutputMs = Math.max(0, Date.now() - segmentStartedAt());
1260
1282
  }
1261
1283
  options.onFirstOutput?.();
1262
1284
  },
@@ -1287,7 +1309,7 @@ export async function executeComboResponses(
1287
1309
  // Object.assign(logCtx, childLog) would overwrite the request-relative value).
1288
1310
  onFirstOutput: () => {
1289
1311
  if (attempt.firstOutputMs === undefined) {
1290
- attempt.firstOutputMs = Math.max(0, Date.now() - started);
1312
+ attempt.firstOutputMs = Math.max(0, Date.now() - segmentStartedAt());
1291
1313
  }
1292
1314
  options.onFirstOutput?.();
1293
1315
  },
@@ -1300,7 +1322,7 @@ export async function executeComboResponses(
1300
1322
  onResponseComplete: callbackGate.onResponseComplete,
1301
1323
  });
1302
1324
  // Both lanes applied the initial choice before provider pins/caps; retain those transitions.
1303
- restoreOriginalRequestedEffort(childLog, jevEffortApplies ? initialJevDecision?.effort : undefined);
1325
+ restoreOriginalRequestedEffort(childLog, jevEffortApplies ? turnEffort?.effort : undefined);
1304
1326
  } catch (error) {
1305
1327
  callbackGate.discard();
1306
1328
  if (options.abortSignal?.aborted) {
@@ -1308,7 +1330,7 @@ export async function executeComboResponses(
1308
1330
  return clientCancelledResponse();
1309
1331
  }
1310
1332
  jevPin?.forget();
1311
- finishRequestAttempt(attempt, 502, Date.now() - started, childLog.usage);
1333
+ finishRequestAttempt(attempt, 502, Date.now() - segmentStartedAt(), childLog.usage);
1312
1334
  throw error;
1313
1335
  }
1314
1336
 
@@ -1336,7 +1358,7 @@ export async function executeComboResponses(
1336
1358
  return clientCancelledResponse();
1337
1359
  }
1338
1360
  jevPin?.forget();
1339
- finishRequestAttempt(attempt, 502, Date.now() - started, childLog.usage);
1361
+ finishRequestAttempt(attempt, 502, Date.now() - segmentStartedAt(), childLog.usage);
1340
1362
  throw error;
1341
1363
  }
1342
1364
  if (preflight.kind === "failed") {
@@ -1378,7 +1400,7 @@ export async function executeComboResponses(
1378
1400
  routeDecision: logCtx.routeDecision,
1379
1401
  attempts: logCtx.attempts,
1380
1402
  activeAttempt: attempt,
1381
- activeAttemptStartedAt: started,
1403
+ activeAttemptStartedAt: segmentStartedAt(),
1382
1404
  resolvedModel: childLog.resolvedModel ?? childLog.model,
1383
1405
  });
1384
1406
  options.onCodexAuthContextResolved?.(resolvedAuth);
@@ -1406,7 +1428,7 @@ export async function executeComboResponses(
1406
1428
  return clientCancelledResponse();
1407
1429
  }
1408
1430
  jevPin?.forget();
1409
- finishRequestAttempt(attempt, 502, Date.now() - started, childLog.usage);
1431
+ finishRequestAttempt(attempt, 502, Date.now() - segmentStartedAt(), childLog.usage);
1410
1432
  throw error;
1411
1433
  }
1412
1434
  if (options.abortSignal?.aborted) {
@@ -1426,7 +1448,7 @@ export async function executeComboResponses(
1426
1448
  finishRequestAttempt(
1427
1449
  attempt,
1428
1450
  failure.response.status,
1429
- Date.now() - started,
1451
+ Date.now() - segmentStartedAt(),
1430
1452
  failure.usage ?? childLog.usage,
1431
1453
  );
1432
1454
  if (isReplayRefusalResponse(failure.response)) {
@@ -1478,6 +1500,7 @@ export async function executeComboResponses(
1478
1500
  verifiedPlanRefusal: failure.verifiedPlanRefusal,
1479
1501
  malformedResponseFormat: failure.malformedResponseFormat,
1480
1502
  openRouterOpaque400: failure.openRouterOpaque400,
1503
+ unofferedToolCallRefusal: failure.unofferedToolCallRefusal,
1481
1504
  providerModerationRefusal: failure.providerModerationRefusal,
1482
1505
  // The adapter the failed leg actually resolved: a configured `kiro` provider can route a
1483
1506
  // model through another wire via `modelAdapters`, and an unresolved leg grants no hop.
@@ -1557,6 +1580,7 @@ export async function executeComboResponses(
1557
1580
  code: failure.upstreamCode,
1558
1581
  malformedResponseFormat: failure.malformedResponseFormat,
1559
1582
  openRouterOpaque400: failure.openRouterOpaque400,
1583
+ unofferedToolCallRefusal: failure.unofferedToolCallRefusal,
1560
1584
  providerModerationRefusal: failure.providerModerationRefusal,
1561
1585
  });
1562
1586
  const failedTargetKey = targetKey(pick.target);
@@ -37,6 +37,8 @@ export interface ConsumedComboFailure {
37
37
  malformedResponseFormat?: boolean;
38
38
  /** Complete bounded-envelope OpenRouter opaque-400 verdict captured before display truncation. */
39
39
  openRouterOpaque400?: boolean;
40
+ /** Complete bounded-envelope unoffered-tool-call verdict captured before display truncation. */
41
+ unofferedToolCallRefusal?: boolean;
40
42
  /** Complete bounded-envelope provider moderation verdict captured before display truncation. */
41
43
  providerModerationRefusal?: boolean;
42
44
  /** Structured upstream `error.code` when present in the failure body. */
@@ -11,6 +11,7 @@ import { clientFacingModelText, type ClientFacingModelLabel } from "../../lib/er
11
11
  * catches the pathological case and stays out of the way otherwise. Every uncertainty
12
12
  * resolves toward admitting.
13
13
  */
14
+ import { collapseDevinModelUid, DEVIN_MODEL_CONTEXT_WINDOWS } from "../../adapters/devin/live-models";
14
15
  import {
15
16
  nativeOpenAiContextWindow,
16
17
  nativeOpenAiMaxInputTokens,
@@ -171,6 +172,35 @@ interface ResolvedContextLimits {
171
172
  ceiling: number | null;
172
173
  }
173
174
 
175
+ /**
176
+ * Devin spells one model dotted in catalog families (`swe-1.7`) and hyphenated in uids
177
+ * (`swe-1-7`), with optional effort suffixes (`swe-2-flash-medium`). A per-model record is
178
+ * matched on the exact id first, then the effort-collapsed id, then both hyphenated forms, and
179
+ * finally on record keys folded the same way (`swe-1.7` serves a `swe-1-7-medium` request), so
180
+ * an operator override on any of those spellings binds before a provider-wide default.
181
+ */
182
+ function devinRecordValue(record: Record<string, number> | undefined, modelId: string): number | null {
183
+ const normalized = modelId.replace(/\./g, "-");
184
+ // Lower-case before collapsing so an upper-case effort suffix (SWE-1-7-MEDIUM) is removed too.
185
+ const collapsed = collapseDevinModelUid(normalized.toLowerCase());
186
+ for (const key of new Set([modelId, collapseDevinModelUid(modelId), normalized, collapsed])) {
187
+ const value = positive(modelRecordValue(record, key));
188
+ if (value !== null) return value;
189
+ }
190
+ if (!record) return null;
191
+ // Exact spellings missed; fold dotted record keys only now so an exact key always wins.
192
+ for (const target of new Set([normalized, collapsed])) {
193
+ // Fold case too, like modelRecordValue's exact-spelling lookup.
194
+ const folded = target.toLowerCase();
195
+ for (const [key, raw] of Object.entries(record)) {
196
+ if (key.replace(/\./g, "-").toLowerCase() !== folded) continue;
197
+ const value = positive(raw);
198
+ if (value !== null) return value;
199
+ }
200
+ }
201
+ return null;
202
+ }
203
+
174
204
  function resolveContextLimits(
175
205
  provider: OcxProviderConfig,
176
206
  providerName: string,
@@ -180,7 +210,10 @@ function resolveContextLimits(
180
210
  // `modelRecordValue`, not a bare lookup: the catalog resolves these same two maps that
181
211
  // way, so a `gpt-oss` entry covers `gpt-oss:120b`. Reading raw here made the gate fall
182
212
  // back to the provider-wide window and refuse turns the model can plainly hold.
183
- const configured = positive(modelRecordValue(provider.modelContextWindows, modelId))
213
+ const devin = provider.adapter === "devin";
214
+ const configured = (devin
215
+ ? devinRecordValue(provider.modelContextWindows, modelId)
216
+ : positive(modelRecordValue(provider.modelContextWindows, modelId)))
184
217
  ?? positive(provider.contextWindow);
185
218
 
186
219
  // The canonical `openai` registry entry declares no context fields, so without this the
@@ -202,9 +235,14 @@ function resolveContextLimits(
202
235
  : null;
203
236
  const nativeMaxInput = canonicalNativeBare ? positive(nativeOpenAiMaxInputTokens(modelId, nativeLimits)) : null;
204
237
 
205
- const window = canonicalNativeBare ? (native ?? generatedNativeWindow(modelId, configured, nativeContextCap)) : configured;
238
+ // Devin's live catalog windows reach the Codex model list but never `provider.modelContextWindows`;
239
+ // without the adapter table every devin combo target resolved no window and the strict gate stayed open.
240
+ const window = canonicalNativeBare ? (native ?? generatedNativeWindow(modelId, configured, nativeContextCap))
241
+ : (configured ?? (devin ? devinRecordValue(DEVIN_MODEL_CONTEXT_WINDOWS, modelId) : null));
206
242
  // modelMaxInputTokens is an input-only cap, so it can only tighten the window.
207
- const configuredMaxInput = positive(modelRecordValue(provider.modelMaxInputTokens, modelId));
243
+ const configuredMaxInput = devin
244
+ ? devinRecordValue(provider.modelMaxInputTokens, modelId)
245
+ : positive(modelRecordValue(provider.modelMaxInputTokens, modelId));
208
246
  const limits = [window, configuredMaxInput, nativeMaxInput].filter((v): v is number => v !== null);
209
247
  return { window, ceiling: limits.length === 0 ? null : Math.min(...limits) };
210
248
  }
@@ -321,7 +321,7 @@ export function createResponsesSendBudget(
321
321
  /**
322
322
  * One credential hop of this logical request, admitted by the INTERSECTION of two bounds.
323
323
  *
324
- * The snapshotted generic OAuth roster cap and `ANTHROPIC_POOL_MAX_FAILOVERS_PER_REQUEST`
324
+ * The snapshotted generic OAuth and Anthropic roster caps
325
325
  * bound rotation within one credential roster. What neither
326
326
  * can see is everything else this request already sent, so three hops layered on a spent
327
327
  * budget still reached upstream three more times. A hop now happens only when its own layer
@@ -36,6 +36,7 @@ import type { AnthropicAccountSelectionReason } from "../../oauth/anthropic-rout
36
36
  import { resolveAnthropicModelRouteForInstance, routeCandidates, type AnthropicRouteDecision } from "../../oauth/anthropic-model-routes";
37
37
  import {
38
38
  anthropicRoutingFor,
39
+ ANTHROPIC_POOL_MAX_FAILOVERS_PER_REQUEST,
39
40
  AnthropicAccountCooldownError,
40
41
  formatAnthropicProviderForLog,
41
42
  } from "../../oauth/anthropic-routing";
@@ -1425,10 +1426,17 @@ export async function prepareResponsesTransport(
1425
1426
  ]).size : 0;
1426
1427
  const fundedAccounts = Math.min(genericRosterSize, GENERIC_OAUTH_MAX_ACCOUNTS_PER_REQUEST);
1427
1428
  const genericFailoverLimit = Math.max(GENERIC_OAUTH_MAX_FAILOVERS_PER_REQUEST, fundedAccounts - 1);
1428
- if (fundingAccountId) expandInferenceOAuthSendBudget(options.sendBudget, fundedAccounts);
1429
+ const anthropicRosterSize = anthropicRouting && anthropicPoolAccountId ? new Set([
1430
+ anthropicPoolAccountId,
1431
+ ...anthropicRouting.getEligibleAnthropicAccounts(Date.now(), route.modelId),
1432
+ ]).size : 0;
1433
+ const anthropicFundedAccounts = Math.min(anthropicRosterSize, GENERIC_OAUTH_MAX_ACCOUNTS_PER_REQUEST);
1434
+ const anthropicFailoverLimit = Math.max(ANTHROPIC_POOL_MAX_FAILOVERS_PER_REQUEST, anthropicFundedAccounts - 1);
1435
+ if (fundingAccountId || anthropicPoolAccountId)
1436
+ expandInferenceOAuthSendBudget(options.sendBudget, Math.max(fundedAccounts, anthropicFundedAccounts));
1429
1437
 
1430
1438
  return {
1431
- genericFailoverLimit,
1439
+ genericFailoverLimit, anthropicFailoverLimit,
1432
1440
  isOAuth401ReplayProvider,
1433
1441
  get sentOAuthSnapshot(): OAuthAccessSnapshot | undefined {
1434
1442
  return sentOAuthSnapshot;
@@ -44,8 +44,8 @@ import { noteKiroMonthlyRefusal, noteKiroServedSuccess } from "../../providers/k
44
44
  import { persistKiroAccountState } from "../../providers/kiro-account-state-disk";
45
45
  import { anthropicPoolCooldownResponse, readDisplaySafeErrorText } from "./core-errors";
46
46
  import { sidecarContextOverflowResponse } from "./context-overflow";
47
+ import { emptyCompletionRetryEnabled } from "./empty-completion-guard";
47
48
  import {
48
- ANTHROPIC_POOL_MAX_FAILOVERS_PER_REQUEST,
49
49
  anthropicRoutingFor,
50
50
  formatAnthropicProviderForLog,
51
51
  } from "../../oauth/anthropic-routing";
@@ -86,7 +86,7 @@ export async function executeResponsesSidecars(
86
86
  | "adapter"
87
87
  | "genericFailoverAccountId"
88
88
  | "genericFailovers"
89
- | "genericFailoverLimit"
89
+ | "genericFailoverLimit" | "anthropicFailoverLimit"
90
90
  | "replayOAuthCredentialSnapshot"
91
91
  | "applyFailoverSnapshot"
92
92
  | "anthropicInstance"
@@ -110,7 +110,7 @@ export async function executeResponsesSidecars(
110
110
  | "notifyResponseComplete"
111
111
  | "cancelResponseCompletion"
112
112
  >,
113
- sendBudgetState: Pick<ResponsesSendBudget, "sendBudgetExhausted" | "reserveCredentialHop" | "refundableAdapterDispatchBudget" | "adapterDispatchBudget" | "noteInitialDispatch" | "noteAdapterPhysicalSend" | "noteAdapterRecoveryWithheld" | "pendingHopPermit" | "dispatchSidecarSend">,
113
+ sendBudgetState: Pick<ResponsesSendBudget, "initialSendAllowance" | "sendBudgetExhausted" | "reserveCredentialHop" | "refundableAdapterDispatchBudget" | "adapterDispatchBudget" | "noteInitialDispatch" | "noteAdapterPhysicalSend" | "noteAdapterRecoveryWithheld" | "pendingHopPermit" | "dispatchSidecarSend">,
114
114
  ) {
115
115
  const { config, options, logCtx } = requestContext;
116
116
  const antigravityPoolActivated = requestState.route.providerName === "google-antigravity"
@@ -422,7 +422,7 @@ export async function executeResponsesSidecars(
422
422
  config, accountId: transportState.anthropicPoolAccountId, sessionKey: anthropicSessionKey,
423
423
  model: route.modelId, requestKey: transportState, decision: transportState.anthropicRouteDecision, currentDecision: currentAnthropicRouteDecision, signal: options.abortSignal,
424
424
  allow429Recovery: !sidecarHasCommittedOutput(), allowAccountRefusal: !sidecarHasCommittedOutput(),
425
- canRetry: hop.allowed && transportState.anthropicPoolFailovers < ANTHROPIC_POOL_MAX_FAILOVERS_PER_REQUEST,
425
+ canRetry: hop.allowed && transportState.anthropicPoolFailovers < transportState.anthropicFailoverLimit,
426
426
  });
427
427
  if (!nextAccountId) {
428
428
  hop.permit?.release();
@@ -503,6 +503,7 @@ export async function executeResponsesSidecars(
503
503
  if (existing) existing();
504
504
  else {
505
505
  const prepaid = sendBudgetState.pendingHopPermit;
506
+ if (anthropicInstance && !prepaid && sendBudgetState.initialSendAllowance(1) === 0) throw new SendBudgetExhaustedError();
506
507
  sendBudgetState.pendingHopPermit = undefined;
507
508
  sendBudgetState.noteInitialDispatch(prepaid);
508
509
  }
@@ -701,6 +702,7 @@ export async function executeResponsesSidecars(
701
702
  ...sidecarBudget.fetchOptions(transportState.adapter),
702
703
  })(input, init)) as typeof globalThis.fetch;
703
704
  const loopResponse = await runWithWebSearch({
705
+ emptyCompletionRetry: emptyCompletionRetryEnabled(config) && !options.comboAttempt && !routedCompaction,
704
706
  dispatchSidecarSend: sendBudgetState.dispatchSidecarSend,
705
707
  parsed, adapter: transportState.adapter,
706
708
  ...(webSearchAdvisor ? { advisor: createAdvisorSidecarStep(webSearchAdvisor as Parameters<typeof createAdvisorSidecarStep>[0]) } : {}),
@@ -121,14 +121,25 @@ export function classifyTaskEnd(result: WindowsStopCommandResult): WindowsTaskEn
121
121
  return toWindowsSchtasksError(error, ["/end", "/tn", TASK]).reason === "access-denied" ? "denied" : "failed";
122
122
  }
123
123
 
124
- /** schtasks /query /fo CSV /nh status column. Localized values fall back to PowerShell. */
124
+ /**
125
+ * Status words of `schtasks /query /fo CSV /nh`. schtasks localizes the column from its .mui
126
+ * resources, and the PowerShell fallback cannot start inside Part A's 1 s manager-state cap on a
127
+ * cold machine (#8863), so the words observed from shipped resources are read here: zh-CN from
128
+ * System32\zh-CN\schtasks.exe.mui, ko-KR from two machines. Any other word stays unknown and
129
+ * falls back to Get-ScheduledTask's locale-independent State enum.
130
+ */
131
+ const TASK_CSV_RUNNING = new Set(["running", "正在运行", "실행 중"]);
132
+ const TASK_CSV_NOT_RUNNING = new Set(["ready", "disabled", "queued", "就绪", "已禁用", "已排队", "준비"]);
133
+ /** ERROR_ACCESS_DENIED in the locales whose status words are read above (zh-CN, ko-KR). */
134
+ const LOCALIZED_ACCESS_DENIED = /拒绝访问|액세스가 거부/;
135
+
125
136
  export function parseTaskCsvState(csv: string): "running" | "not-running" | "unknown" {
126
137
  for (const line of csv.split(/\r?\n/)) {
127
138
  if (!windowsSchedulerCsvIncludesTask(line, TASK)) continue;
128
139
  const fields = [...line.matchAll(/"([^"]*)"/g)].map(match => match[1]!.trim());
129
- const status = fields[fields.length - 1] ?? "";
130
- if (/^running$/i.test(status)) return "running";
131
- if (/^(ready|disabled|queued)$/i.test(status)) return "not-running";
140
+ const status = (fields[fields.length - 1] ?? "").normalize("NFC").toLowerCase();
141
+ if (TASK_CSV_RUNNING.has(status)) return "running";
142
+ if (TASK_CSV_NOT_RUNNING.has(status)) return "not-running";
132
143
  return "unknown";
133
144
  }
134
145
  return "unknown";
@@ -291,7 +302,19 @@ export function windowsCheckedStopLiveDeps(options: WindowsCheckedStopLiveOption
291
302
  },
292
303
  async endTask(timeoutMs) {
293
304
  if (!gate()) return "failed";
294
- return classifyTaskEnd(await run("schtasks", ["/end", "/tn", TASK], timeoutMs));
305
+ const started = Date.now();
306
+ const ended = await run("schtasks", ["/end", "/tn", TASK], timeoutMs);
307
+ const verdict = classifyTaskEnd(ended);
308
+ // schtasks localizes "no running instance" as well (#8863). After a real nonzero exit, an
309
+ // idle task is not-running only when its own CSV state says so; a running or unreadable
310
+ // state, a timeout, or an /end that never ran keeps the failure. A localized denial stays a
311
+ // denial: Part A stops at it whatever the task's state.
312
+ if (verdict !== "failed" || ended.timedOut || ended.status === null) return verdict;
313
+ if (LOCALIZED_ACCESS_DENIED.test(ended.stderr + "\n" + ended.stdout)) return "denied";
314
+ const left = Math.floor(timeoutMs - (Date.now() - started));
315
+ if (left < 1) return "failed";
316
+ const csv = await run("schtasks", ["/query", "/tn", TASK, "/fo", "CSV", "/nh"], left);
317
+ return csv.status === 0 && !csv.timedOut && parseTaskCsvState(csv.stdout) === "not-running" ? "not-running" : "failed";
295
318
  },
296
319
  async taskState(timeoutMs) {
297
320
  const started = Date.now();
@@ -38,9 +38,10 @@ export interface OcxReasoningReplayScopeRef {
38
38
  */
39
39
  readonly clientPrincipalId?: string;
40
40
  /**
41
- * Conversation namespace for replay state. Historically this was always the Codex parent-thread
42
- * id; headerless Responses callers use a raw sanitized thread/Cursor/session fallback, never the
43
- * hashed request-log conversation id.
41
+ * Conversation namespace for replay state. The request's own thread-id wins so a child agent
42
+ * cannot write its route into the parent's record (#8632); otherwise this is the Codex
43
+ * parent-thread id, and headerless Responses callers use a raw sanitized thread/Cursor/session
44
+ * fallback, never the hashed request-log conversation id.
44
45
  */
45
46
  readonly clientThreadId: string;
46
47
  current?: Readonly<OcxReasoningReplayIdentity>;
@@ -317,7 +317,26 @@ const DEVIN_PRICING = "https://docs.devin.ai/desktop/models (official modelCostD
317
317
  const DEVIN_FABLE_52_SOURCE = "https://docs.devin.ai/desktop/models (Claude Fable 5.2 listed at 8 / 40, cache read 0.20, cache write 10, 2026-11-10; uid claude-fable-5-2-<effort>, seed collapses to the base)";
318
318
  const DEVIN_SWE2_NOTE = "list rate; $0 self-serve through 2026-10-08 and 75%-off enterprise through 2026-12-31 are time-boxed promos, not baked in";
319
319
 
320
+ const GPT62_PRICING = [
321
+ { modelId: "gpt-6.2-sol", cost4: { input: 2, output: 10, cacheRead: 0.1, cacheWrite: 2.5 } },
322
+ { modelId: "gpt-6.2-astra", cost4: { input: 8, output: 40, cacheRead: 0.4, cacheWrite: 10 } },
323
+ ];
324
+ const GPT62_VERIFIED_AT = "2026-11-18";
325
+
320
326
  export const EXPECTED_PRICE_OVERLAYS: readonly ExpectedPriceOverlay[] = [
327
+ ...GPT62_PRICING.flatMap(({ modelId, cost4 }): ExpectedPriceOverlay[] => {
328
+ const source = `https://developers.openai.com/api/docs/models/${modelId}`;
329
+ return [
330
+ { provider: "openai-apikey", modelId, cost4, source, verifiedAt: GPT62_VERIFIED_AT, status: "verified" },
331
+ { provider: "openai", modelId, cost4, source: `API-reference comparison estimate: ${source}`, verifiedAt: GPT62_VERIFIED_AT, status: "verified-derived" },
332
+ { provider: "openrouter", modelId: `openai/${modelId}`, cost4, source: "https://openrouter.ai/api/v1/models", verifiedAt: GPT62_VERIFIED_AT, status: "verified" },
333
+ // 261118 preemptive: Devin uid follows gpt-6-1-sol; API-reference estimate, not billing.
334
+ ...(["devin-cli", "devin"] as const).map((provider): ExpectedPriceOverlay => ({
335
+ provider, modelId: modelId.replaceAll(".", "-"), cost4,
336
+ source: `derived: OpenAI API list price ${source}; Devin uid mirrors gpt-6-1-sol`, verifiedAt: GPT62_VERIFIED_AT, status: "verified-derived",
337
+ })),
338
+ ];
339
+ }),
321
340
  { provider: "openai-apikey", modelId: "gpt-6-astra", cost4: GPT6_ASTRA, source: ASTRA_API_PRICING, verifiedAt: "2026-09-05", status: "verified" },
322
341
  // Display estimates use API prices for both login and API-key routes, including cache writes.
323
342
  { provider: "openai", modelId: "gpt-6-astra", cost4: GPT6_ASTRA, source: `API-reference comparison estimate: ${ASTRA_API_PRICING}`, verifiedAt: "2026-09-05", status: "verified-derived" },
@@ -719,6 +738,10 @@ const ANTHROPIC_FAST_PRICING = "https://platform.claude.com/docs/en/about-claude
719
738
  * because they reuse its model slug. Multipliers apply uniformly after cache discounts.
720
739
  */
721
740
  export const PRIORITY_PRICING_RULES: readonly PriorityPricingRule[] = [
741
+ ...GPT62_PRICING.flatMap(({ modelId }) => ["openai", "openai-apikey"].map((provider): PriorityPricingRule => ({
742
+ provider, modelId, multiplier: 2,
743
+ source: "https://developers.openai.com/api/docs/pricing", verifiedAt: GPT62_VERIFIED_AT,
744
+ }))),
722
745
  ...["openai", "openai-apikey"].flatMap(provider =>
723
746
  Object.entries(PRIORITY_MULTIPLIERS)
724
747
  // Daybreak's `gpt-` selector belongs to ChatGPT/Codex accounts; its bare selector belongs
@@ -841,6 +864,18 @@ const OPENAI_CONTEXT_MODELS = [
841
864
  ];
842
865
 
843
866
  export const CONTEXT_TIERS: readonly ContextTier[] = [
867
+ ...GPT62_PRICING.flatMap(({ modelId }): ContextTier[] => [
868
+ ...["openai", "openai-apikey"].map((provider): ContextTier => ({
869
+ provider, modelId, thresholdInputTokens: 272_000, inclusive: false,
870
+ multiplier: OPENAI_LONG_CONTEXT, confirmedPriorityRelation: "stack",
871
+ source: `https://developers.openai.com/api/docs/models/${modelId}`, verifiedAt: GPT62_VERIFIED_AT,
872
+ })),
873
+ {
874
+ provider: "openrouter", modelId: `openai/${modelId}`, thresholdInputTokens: 272_000, inclusive: true,
875
+ multiplier: OPENAI_LONG_CONTEXT,
876
+ source: "https://openrouter.ai/api/v1/models", verifiedAt: GPT62_VERIFIED_AT,
877
+ },
878
+ ]),
844
879
  // Reference bands on preemptive/subscription surfaces follow their reference price, not billing.
845
880
  ...[
846
881
  ...["anthropic", "anthropic-apikey", "opencode-go", "opencode-zen", "venice", "github-copilot", "devin", "devin-cli", "claude-cli"].map(provider => [provider, "claude-haiku-5-5"]),
@@ -486,6 +486,12 @@ export async function scanUsageLedgerCooperatively(
486
486
  export interface ReadUsageLedgerRowsOptions {
487
487
  /** Ascending, distinct row-start offsets reported by an earlier scan of the same file. */
488
488
  offsets: readonly number[];
489
+ /**
490
+ * Folds rows as they are decoded instead of retaining them in `entries`. Callback work shares
491
+ * the cooperative time slice. Results are provisional until this read resolves: discard them
492
+ * if identity or content validation fails, as with the full-ledger scanner's callback.
493
+ */
494
+ onEntry?: (entry: PersistedUsageEntry, rowStartBytes: number) => void;
489
495
  /** Identity those offsets were captured under; a replaced file invalidates all of them. */
490
496
  expectedIdentityKey: string;
491
497
  /** Revision the offsets were captured under. Any movement re-validates the boundary. */
@@ -503,6 +509,7 @@ export interface ReadUsageLedgerRowsOptions {
503
509
  }
504
510
 
505
511
  export interface UsageLedgerRowRead {
512
+ /** Empty when an onEntry callback consumes the rows. */
506
513
  entries: PersistedUsageEntry[];
507
514
  /** Offsets that no longer frame a complete, normalizable row. */
508
515
  unreadableOffsets: number[];
@@ -626,27 +633,32 @@ export async function readUsageLedgerRowsAtOffsets(
626
633
  return null;
627
634
  };
628
635
 
629
- const walk = (visit: (bytes: Buffer, offset: number) => boolean): number[] => {
636
+ const walk = async (visit: (bytes: Buffer, offset: number) => boolean): Promise<number[]> => {
630
637
  const unreadableOffsets: number[] = [];
638
+ let lastYieldAt = performance.now();
631
639
  for (const offset of options.offsets) {
632
640
  throwIfAborted(options.signal);
633
641
  if (!Number.isSafeInteger(offset) || offset < 0 || offset >= size) {
634
642
  unreadableOffsets.push(offset);
635
- continue;
643
+ } else {
644
+ const bytes = lineAt(offset);
645
+ if (!bytes || !visit(bytes, offset)) unreadableOffsets.push(offset);
636
646
  }
637
- const bytes = lineAt(offset);
638
- if (!bytes) {
639
- unreadableOffsets.push(offset);
640
- continue;
647
+ // A bounded candidate count does not bound decoding or aggregation time. Include the
648
+ // visitor in the same slice as the full scanner, and let timers/cancellation run even
649
+ // within one page. A single row remains atomic.
650
+ if (performance.now() - lastYieldAt >= MAX_SCAN_SLICE_MS) {
651
+ await new Promise<void>(resolve => setTimeout(resolve, 0));
652
+ throwIfAborted(options.signal);
653
+ lastYieldAt = performance.now();
641
654
  }
642
- if (!visit(bytes, offset)) unreadableOffsets.push(offset);
643
655
  }
644
656
  return unreadableOffsets;
645
657
  };
646
658
 
647
659
  const entries: PersistedUsageEntry[] = [];
648
660
  const captured = createHash("sha256");
649
- const unreadableOffsets = walk((bytes, offset) => {
661
+ const unreadableOffsets = await walk((bytes, offset) => {
650
662
  captured.update(bytes);
651
663
  captured.update("\0");
652
664
  captured.update(String(offset));
@@ -660,7 +672,8 @@ export async function readUsageLedgerRowsAtOffsets(
660
672
  // normalizes means the ledger moved under the projection, so it is unreadable rather than
661
673
  // silently dropped: the caller rebuilds instead of reporting a smaller complete result.
662
674
  if (!entry) return false;
663
- entries.push(entry);
675
+ if (options.onEntry) options.onEntry(entry, offset);
676
+ else entries.push(entry);
664
677
  return true;
665
678
  });
666
679
  const capturedDigest = captured.digest("hex");
@@ -675,6 +688,7 @@ export async function readUsageLedgerRowsAtOffsets(
675
688
  if (!pathRevision || usageLogIdentityKey(pathRevision) !== options.expectedIdentityKey) {
676
689
  throw rebuildRequired("identity_mismatch");
677
690
  }
691
+ if (pathRevision.size < size) throw rebuildRequired("shrink");
678
692
  // Any write during the read can be an in-place rewrite, not just an append. Re-read the
679
693
  // same offsets and compare, so a rewrite that lands between two batches cannot publish a
680
694
  // mixture of old and new rows. A pure append leaves every selected row unchanged.
@@ -686,8 +700,11 @@ export async function readUsageLedgerRowsAtOffsets(
686
700
  if (physical !== undefined && endingRevision.size <= physical) {
687
701
  throw rebuildRequired("content_changed");
688
702
  }
703
+ // Verification must read the file again, not re-hash a small page still in the read cache.
704
+ windowStart = -1;
705
+ windowLength = 0;
689
706
  const verified = createHash("sha256");
690
- const recheck = walk((bytes, offset) => {
707
+ const recheck = await walk((bytes, offset) => {
691
708
  verified.update(bytes);
692
709
  verified.update("\0");
693
710
  verified.update(String(offset));
@@ -697,7 +714,24 @@ export async function readUsageLedgerRowsAtOffsets(
697
714
  || verified.digest("hex") !== capturedDigest) {
698
715
  throw rebuildRequired("content_changed");
699
716
  }
717
+ const verifiedFdRevision = revisionFromStat(path, fd, fstatSync(fd));
718
+ const verifiedPathRevision = currentUsageLogRevision();
719
+ if (!verifiedPathRevision
720
+ || usageLogIdentityKey(verifiedFdRevision) !== options.expectedIdentityKey
721
+ || usageLogIdentityKey(verifiedPathRevision) !== options.expectedIdentityKey) {
722
+ throw rebuildRequired("identity_mismatch");
723
+ }
724
+ if (verifiedFdRevision.size < size || verifiedPathRevision.size < size) {
725
+ throw rebuildRequired("shrink");
726
+ }
727
+ // Verification yields too. A later write could change bytes already checked, so retry
728
+ // rather than publish a result whose verification raced another mutation.
729
+ if (usageLogRevisionKey(verifiedFdRevision) !== usageLogRevisionKey(endingRevision)
730
+ || usageLogRevisionKey(verifiedPathRevision) !== usageLogRevisionKey(pathRevision)) {
731
+ throw rebuildRequired("content_changed");
732
+ }
700
733
  }
734
+ throwIfAborted(options.signal);
701
735
  return { entries, unreadableOffsets };
702
736
  } finally {
703
737
  closeSync(fd);