@bitkyc08/opencodex 2.87.0 → 2.89.0

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 (246) hide show
  1. package/README.md +11 -1
  2. package/bin/ocx.mjs +15 -4
  3. package/gui/dist/assets/App-BnRSujOq.js +58 -0
  4. package/gui/dist/assets/App-dnkJ85te.css +1 -0
  5. package/gui/dist/assets/{Tray-DCd-A1az.js → Tray-7z1COnmz.js} +1 -1
  6. package/gui/dist/assets/index-uchD-ngf.js +94 -0
  7. package/gui/dist/assets/{quota-summary-ChvmKGOk.js → quota-summary-BY_10IEz.js} +1 -1
  8. package/gui/dist/index.html +1 -1
  9. package/package.json +1 -1
  10. package/src/adapters/anthropic/cache-retention.ts +117 -0
  11. package/src/adapters/anthropic/passthrough.ts +7 -4
  12. package/src/adapters/anthropic-rejection-memo.ts +71 -0
  13. package/src/adapters/anthropic-replay.ts +135 -0
  14. package/src/adapters/anthropic-tool-catalog.ts +11 -0
  15. package/src/adapters/anthropic.ts +86 -48
  16. package/src/adapters/coding-agent/protocol.ts +6 -0
  17. package/src/adapters/google-http.ts +3 -2
  18. package/src/adapters/google-progress-history.ts +53 -0
  19. package/src/adapters/google-vertex-endpoint.ts +67 -0
  20. package/src/adapters/google.ts +52 -30
  21. package/src/adapters/openai-responses/compaction-search-history.ts +27 -0
  22. package/src/adapters/openai-responses/passthrough.ts +17 -12
  23. package/src/adapters/openai-responses/reasoning.ts +68 -1
  24. package/src/adapters/registry.ts +1 -1
  25. package/src/adapters/upstream-http-error.ts +3 -2
  26. package/src/adapters/zed.ts +7 -4
  27. package/src/bridge/internal.ts +17 -3
  28. package/src/bridge/response-json.ts +4 -4
  29. package/src/bridge/sse.ts +17 -10
  30. package/src/claude/inbound-content-options.ts +66 -0
  31. package/src/claude/inbound.ts +12 -26
  32. package/src/claude/outbound.ts +18 -5
  33. package/src/cli/account-api.ts +13 -1
  34. package/src/cli/account-extended.ts +17 -5
  35. package/src/cli/account-key-quota.ts +4 -1
  36. package/src/cli/account.ts +1 -1
  37. package/src/cli/capabilities-agents-routing.ts +45 -8
  38. package/src/cli/capabilities-base.ts +5 -5
  39. package/src/cli/capabilities-provider-models.ts +5 -4
  40. package/src/cli/combo-result.ts +8 -2
  41. package/src/cli/combo.ts +17 -0
  42. package/src/cli/cross-home-owner.ts +16 -2
  43. package/src/cli/help.ts +1 -1
  44. package/src/cli/loopback-family.ts +31 -0
  45. package/src/cli/message-command.ts +4 -1
  46. package/src/cli/message-remote-args.ts +69 -0
  47. package/src/cli/message-remote-command.ts +77 -0
  48. package/src/cli/provider-runtime.ts +69 -1
  49. package/src/cli/registry.ts +3 -3
  50. package/src/codex/account-pause.ts +6 -1
  51. package/src/codex/account-runtime-state.ts +5 -0
  52. package/src/codex/account-store.ts +30 -3
  53. package/src/codex/auth-api/account-list.ts +3 -0
  54. package/src/codex/auth-api/main-account-probe.ts +3 -2
  55. package/src/codex/auth-context.ts +9 -4
  56. package/src/codex/catalog/aggregation.ts +15 -0
  57. package/src/codex/catalog/auto-review.ts +2 -1
  58. package/src/codex/catalog/build-entries.ts +17 -7
  59. package/src/codex/catalog/derive-entry.ts +3 -1
  60. package/src/codex/catalog/metadata.ts +2 -0
  61. package/src/codex/catalog/model-visibility.ts +0 -2
  62. package/src/codex/catalog/ownership.ts +71 -0
  63. package/src/codex/catalog/parsing.ts +12 -3
  64. package/src/codex/catalog/retained-sync.ts +2 -1
  65. package/src/codex/catalog/routed-gather.ts +3 -1
  66. package/src/codex/catalog/routed-removal.ts +4 -4
  67. package/src/codex/convergence.ts +2 -1
  68. package/src/codex/low-quota-desktop-notifier.ts +125 -0
  69. package/src/codex/low-quota-protection.ts +57 -31
  70. package/src/codex/main-account-cache.ts +77 -3
  71. package/src/codex/main-account.ts +6 -2
  72. package/src/codex/native-residue.ts +2 -3
  73. package/src/codex/quota.ts +26 -2
  74. package/src/codex/websocket-activity.ts +161 -0
  75. package/src/codex/websocket-registry.ts +25 -1
  76. package/src/combos/failover.ts +2 -1
  77. package/src/combos/jev-model-backend.ts +30 -7
  78. package/src/combos/jev-model-request.ts +20 -0
  79. package/src/combos/jev-quota-activation.ts +153 -0
  80. package/src/combos/jev-quota-config.ts +62 -0
  81. package/src/combos/jev-quota-route.ts +26 -0
  82. package/src/combos/jev-quota.ts +88 -0
  83. package/src/combos/jev.ts +25 -12
  84. package/src/combos/types.ts +27 -0
  85. package/src/compatibility/index.ts +3 -0
  86. package/src/compatibility/openai-responses.ts +8 -1
  87. package/src/compatibility/opengateway-responses.ts +61 -0
  88. package/src/config/account-auto-pause-revisions.ts +136 -0
  89. package/src/config/diagnostics.ts +13 -0
  90. package/src/config/live-reconcile.ts +9 -0
  91. package/src/config/opencode-zen-upgrade.ts +9 -8
  92. package/src/config/persist-unlocked.ts +10 -1
  93. package/src/config/removed-opencode-free.ts +135 -0
  94. package/src/config/schema/account-auto-pause.ts +14 -0
  95. package/src/config/schema/config-schema.ts +15 -0
  96. package/src/config/schema/leaf-validators.ts +7 -0
  97. package/src/config.ts +10 -3
  98. package/src/generated/compatibility-version.json +411 -175
  99. package/src/images/loop.ts +101 -27
  100. package/src/lib/account-auto-pause-response.ts +16 -0
  101. package/src/lib/bun-runtime.ts +18 -1
  102. package/src/lib/bun-transpiler-cache-scope.d.mts +8 -0
  103. package/src/lib/bun-transpiler-cache-scope.mjs +68 -0
  104. package/src/lib/decision-quota-sink.ts +76 -0
  105. package/src/lib/request-failure-model.ts +2 -0
  106. package/src/lib/state-store-registrations.ts +25 -0
  107. package/src/link/ssh-argv.ts +3 -0
  108. package/src/messaging/envelope.ts +6 -3
  109. package/src/messaging/remote-auth.ts +83 -0
  110. package/src/messaging/remote-bridge.ts +270 -0
  111. package/src/messaging/remote-contract.ts +108 -0
  112. package/src/messaging/remote-enrollment-recovery.ts +59 -0
  113. package/src/messaging/remote-enrollment.ts +244 -0
  114. package/src/messaging/remote-envelope.ts +40 -0
  115. package/src/messaging/remote-files.ts +117 -0
  116. package/src/messaging/remote-owner.ts +169 -0
  117. package/src/messaging/remote-ports.ts +56 -0
  118. package/src/messaging/remote-process.ts +131 -0
  119. package/src/messaging/remote-replies.ts +98 -0
  120. package/src/messaging/remote-rpc-admission.ts +32 -0
  121. package/src/messaging/remote-send.ts +72 -0
  122. package/src/messaging/remote-store.ts +111 -0
  123. package/src/messaging/remote-tunnels.ts +65 -0
  124. package/src/messaging/rpc.ts +13 -1
  125. package/src/oauth/account-auto-pause-policy.ts +66 -0
  126. package/src/oauth/account-auto-pause.ts +171 -0
  127. package/src/oauth/account-quota-rank.ts +2 -1
  128. package/src/oauth/anthropic-account-refusal.ts +6 -0
  129. package/src/oauth/anthropic-model-family.ts +8 -0
  130. package/src/oauth/anthropic-model-quota.ts +2 -6
  131. package/src/oauth/api-key-pool-preserve.ts +37 -0
  132. package/src/oauth/generic-account-failover.ts +7 -4
  133. package/src/oauth/index.ts +5 -35
  134. package/src/oauth/store.ts +35 -8
  135. package/src/oauth/types.ts +2 -0
  136. package/src/plugins/upstream-hooks.ts +14 -1
  137. package/src/protocols/encoders/adapter-events.ts +5 -5
  138. package/src/protocols/encoders/messages.ts +2 -2
  139. package/src/providers/api-key-pause.ts +32 -0
  140. package/src/providers/api-key-resolve.ts +10 -4
  141. package/src/providers/api-key-selection.ts +6 -1
  142. package/src/providers/api-keys.ts +8 -4
  143. package/src/providers/claude-cli-usage.ts +5 -0
  144. package/src/providers/key-failover.ts +11 -11
  145. package/src/providers/model-rename-fields.ts +1 -0
  146. package/src/providers/opencode-zen-rate-limit.ts +12 -73
  147. package/src/providers/quota/account-cache.ts +15 -1
  148. package/src/providers/quota/anthropic-account-quota.ts +2 -1
  149. package/src/providers/quota/anthropic-cooldown-recovery.ts +22 -4
  150. package/src/providers/quota/anthropic-decision-windows.ts +47 -0
  151. package/src/providers/quota/antigravity.ts +39 -6
  152. package/src/providers/quota/report-cache.ts +14 -0
  153. package/src/providers/quota/vendor-probes-oauth.ts +3 -3
  154. package/src/providers/quota-decision-publication.ts +77 -0
  155. package/src/providers/quota-decision-snapshot.ts +145 -0
  156. package/src/providers/quota-routing-cache.ts +3 -0
  157. package/src/providers/quota-types.ts +6 -1
  158. package/src/providers/quota.ts +3 -2
  159. package/src/providers/registry/entries-extended.ts +8 -52
  160. package/src/providers/registry/model-seeds.ts +1 -1
  161. package/src/providers/registry.ts +1 -2
  162. package/src/providers/stale-vision-classification-migration.ts +1 -1
  163. package/src/responses/parser.ts +18 -10
  164. package/src/responses/reasoning-replay-cache.ts +50 -4
  165. package/src/responses/web-search-replay.ts +138 -0
  166. package/src/routing/analytics.ts +7 -15
  167. package/src/routing/compatibility/behavior.ts +4 -1
  168. package/src/server/auth-cors.ts +7 -0
  169. package/src/server/background-lifecycle.ts +2 -1
  170. package/src/server/chat-native.ts +28 -13
  171. package/src/server/claude-messages.ts +4 -2
  172. package/src/server/index/live-sideband.ts +3 -2
  173. package/src/server/index/optional-listeners.ts +6 -0
  174. package/src/server/index/serve-options.ts +11 -5
  175. package/src/server/index/websocket-handler.ts +18 -4
  176. package/src/server/index.ts +7 -7
  177. package/src/server/lifecycle.ts +13 -1
  178. package/src/server/management/combo-routes.ts +17 -2
  179. package/src/server/management/oauth-account-routes.ts +10 -1
  180. package/src/server/management/provider-completion-test.ts +289 -0
  181. package/src/server/management/provider-routes.ts +9 -1
  182. package/src/server/management/route-registry.ts +2 -0
  183. package/src/server/messages-native.ts +34 -7
  184. package/src/server/messages-shared.ts +4 -1
  185. package/src/server/request-log.ts +7 -2
  186. package/src/server/responses/account-auto-pause-recovery.ts +26 -0
  187. package/src/server/responses/adapter-continuation.ts +34 -18
  188. package/src/server/responses/adapter-delivery.ts +53 -6
  189. package/src/server/responses/adapter-dispatch.ts +104 -29
  190. package/src/server/responses/anthropic-thinking-recovery.ts +153 -0
  191. package/src/server/responses/combo-session-recall.ts +25 -0
  192. package/src/server/responses/combo-stream-preflight.ts +6 -3
  193. package/src/server/responses/combo-tool-loop-pin.ts +70 -0
  194. package/src/server/responses/compact.ts +8 -5
  195. package/src/server/responses/compaction-recovery.ts +2 -0
  196. package/src/server/responses/context-overflow.ts +141 -2
  197. package/src/server/responses/core-codex-account.ts +9 -3
  198. package/src/server/responses/core-combo.ts +34 -7
  199. package/src/server/responses/core-lifetime.ts +3 -2
  200. package/src/server/responses/core-options.ts +2 -0
  201. package/src/server/responses/core-replay.ts +5 -1
  202. package/src/server/responses/core.ts +12 -9
  203. package/src/server/responses/empty-completion-guard.ts +2 -0
  204. package/src/server/responses/fetch-helpers.ts +35 -4
  205. package/src/server/responses/jev-model-invoke.ts +13 -14
  206. package/src/server/responses/native-steering-log.ts +2 -2
  207. package/src/server/responses/passthrough-delivery.ts +3 -2
  208. package/src/server/responses/passthrough-dispatch.ts +30 -33
  209. package/src/server/responses/passthrough-execution.ts +3 -2
  210. package/src/server/responses/passthrough-observations.ts +38 -0
  211. package/src/server/responses/request-compression.ts +98 -0
  212. package/src/server/responses/request-prepare.ts +1 -0
  213. package/src/server/responses/request-transport.ts +32 -3
  214. package/src/server/responses/response-effects.ts +1 -0
  215. package/src/server/responses/response-tool-ownership.ts +21 -0
  216. package/src/server/responses/run-turn-execution.ts +2 -0
  217. package/src/server/responses/sidecar-execution.ts +59 -10
  218. package/src/server/responses/sidecar-send-budget.ts +2 -2
  219. package/src/server/responses/terminal-guard.ts +2 -0
  220. package/src/server/responses-undeclared-tool-guard.ts +1 -1
  221. package/src/server/startup-action-control.ts +2 -2
  222. package/src/server/startup-health-cache.ts +2 -2
  223. package/src/server/windows-tray-control.ts +2 -2
  224. package/src/server/ws-bridge.ts +2 -0
  225. package/src/service/definition.ts +9 -1
  226. package/src/types/account-auto-pause.ts +15 -0
  227. package/src/types/config.ts +23 -1
  228. package/src/types/opencode-zen.ts +9 -3
  229. package/src/types/provider.ts +9 -5
  230. package/src/types/request.ts +30 -4
  231. package/src/types/wire.ts +4 -2
  232. package/src/types.ts +2 -0
  233. package/src/usage/cost.ts +17 -1
  234. package/src/usage/expected-prices.ts +3 -1
  235. package/src/usage/jev-stats.ts +4 -0
  236. package/src/usage/log.ts +2 -0
  237. package/src/usage/summary.ts +16 -3
  238. package/src/usage/telemetry-contract.ts +1 -0
  239. package/src/web-search/anthropic-history.ts +59 -0
  240. package/src/web-search/loop.ts +212 -73
  241. package/src/web-search/progress-stream.ts +14 -1
  242. package/src/web-search/run-turn-loop.ts +9 -3
  243. package/src/web-search/sidecar-search-replay.ts +96 -0
  244. package/gui/dist/assets/App-ZfYbLHkB.js +0 -58
  245. package/gui/dist/assets/App-wT3mXqFz.css +0 -1
  246. package/gui/dist/assets/index-CfYDMj34.js +0 -94
@@ -13,6 +13,7 @@
13
13
  * other.
14
14
  */
15
15
  import { sniffImageDimensions } from "../adapters/anthropic-image-guard";
16
+ import { anthropicCacheWrite1hTokens } from "../adapters/anthropic/cache-retention";
16
17
  import { createToolCallIdAllocator } from "../adapters/tool-call-id";
17
18
  import { idleDeadline } from "../lib/abort";
18
19
  import {
@@ -29,12 +30,13 @@ import type { RequestLogContext } from "./request-log";
29
30
  import { recordGenerationEvent } from "./request-log-generation-window";
30
31
  import type { OcxConfig } from "../types";
31
32
 
32
- export function anthropicUsageToOcx(usage: Rec | undefined): { inputTokens: number; outputTokens: number; cachedInputTokens?: number; cacheReadInputTokens?: number; cacheCreationInputTokens?: number } | undefined {
33
+ export function anthropicUsageToOcx(usage: Rec | undefined): { inputTokens: number; outputTokens: number; cachedInputTokens?: number; cacheReadInputTokens?: number; cacheCreationInputTokens?: number; cacheCreation1hInputTokens?: number } | undefined {
33
34
  if (!usage) return undefined;
34
35
  const num = (v: unknown) => typeof v === "number" ? v : 0;
35
36
  const hasCache = usage.cache_read_input_tokens !== undefined || usage.cache_creation_input_tokens !== undefined;
36
37
  const read = num(usage.cache_read_input_tokens);
37
38
  const write = num(usage.cache_creation_input_tokens);
39
+ const write1h = anthropicCacheWrite1hTokens(usage, write);
38
40
  // Anthropic input_tokens excludes cache read/write; normalize to the canonical
39
41
  // inclusive convention (types.ts OcxUsage / devlog 070). cached = READS only.
40
42
  return {
@@ -44,6 +46,7 @@ export function anthropicUsageToOcx(usage: Rec | undefined): { inputTokens: numb
44
46
  cachedInputTokens: read,
45
47
  cacheReadInputTokens: read,
46
48
  cacheCreationInputTokens: write,
49
+ ...(write1h !== undefined ? { cacheCreation1hInputTokens: write1h } : {}),
47
50
  } : {}),
48
51
  };
49
52
  }
@@ -982,7 +982,7 @@ export function usageFromResponsesPayload(usage: unknown): OcxUsage | undefined
982
982
  const raw = usage as {
983
983
  input_tokens?: unknown;
984
984
  output_tokens?: unknown;
985
- input_tokens_details?: { cached_tokens?: unknown; cache_write_tokens?: unknown };
985
+ input_tokens_details?: { cached_tokens?: unknown; cache_write_tokens?: unknown; cache_write_1h_tokens?: unknown };
986
986
  output_tokens_details?: { reasoning_tokens?: unknown };
987
987
  total_tokens?: unknown;
988
988
  prompt_tokens?: unknown;
@@ -1004,6 +1004,9 @@ export function usageFromResponsesPayload(usage: unknown): OcxUsage | undefined
1004
1004
  ...(typeof raw.input_tokens_details?.cache_write_tokens === "number"
1005
1005
  ? { cacheCreationInputTokens: raw.input_tokens_details.cache_write_tokens }
1006
1006
  : {}),
1007
+ ...(typeof raw.input_tokens_details?.cache_write_1h_tokens === "number"
1008
+ ? { cacheCreation1hInputTokens: raw.input_tokens_details.cache_write_1h_tokens }
1009
+ : {}),
1007
1010
  ...(typeof raw.output_tokens_details?.reasoning_tokens === "number"
1008
1011
  ? { reasoningOutputTokens: raw.output_tokens_details.reasoning_tokens }
1009
1012
  : {}),
@@ -1864,7 +1867,7 @@ export function aggregateAttemptUsage(
1864
1867
  if (usages.length === 0) return { status };
1865
1868
 
1866
1869
  const sumOptional = (
1867
- key: "cachedInputTokens" | "cacheReadInputTokens" | "cacheCreationInputTokens"
1870
+ key: "cachedInputTokens" | "cacheReadInputTokens" | "cacheCreationInputTokens" | "cacheCreation1hInputTokens"
1868
1871
  | "reasoningOutputTokens" | "providerCredits",
1869
1872
  ): number | undefined => {
1870
1873
  const present = usages.flatMap(usage => (
@@ -1875,6 +1878,7 @@ export function aggregateAttemptUsage(
1875
1878
  const cachedInputTokens = sumOptional("cachedInputTokens");
1876
1879
  const cacheReadInputTokens = sumOptional("cacheReadInputTokens");
1877
1880
  const cacheCreationInputTokens = sumOptional("cacheCreationInputTokens");
1881
+ const cacheCreation1hInputTokens = sumOptional("cacheCreation1hInputTokens");
1878
1882
  const reasoningOutputTokens = sumOptional("reasoningOutputTokens");
1879
1883
  const providerCredits = sumOptional("providerCredits");
1880
1884
  const totalTokens = usages.reduce(
@@ -1888,6 +1892,7 @@ export function aggregateAttemptUsage(
1888
1892
  ...(cachedInputTokens !== undefined ? { cachedInputTokens } : {}),
1889
1893
  ...(cacheReadInputTokens !== undefined ? { cacheReadInputTokens } : {}),
1890
1894
  ...(cacheCreationInputTokens !== undefined ? { cacheCreationInputTokens } : {}),
1895
+ ...(cacheCreation1hInputTokens !== undefined ? { cacheCreation1hInputTokens } : {}),
1891
1896
  ...(reasoningOutputTokens !== undefined ? { reasoningOutputTokens } : {}),
1892
1897
  ...(providerCredits !== undefined ? { providerCredits } : {}),
1893
1898
  ...(status === "estimated" ? { estimated: true } : {}),
@@ -0,0 +1,26 @@
1
+ import type { ResponsesSendBudget } from "./request-send-budget";
2
+
3
+ const moves = new WeakMap<object, number>();
4
+ /** One same-provider credential move, funded by the existing recovery owner. */
5
+ export async function recoverAutoPausedAccount<T>(options: {
6
+ requestKey: object;
7
+ remaining?: number;
8
+ reserve: ResponsesSendBudget["reserveCredentialHop"];
9
+ refresh: () => Promise<boolean>;
10
+ refetch: () => Promise<T>;
11
+ setPermit: (permit: ResponsesSendBudget["pendingHopPermit"]) => void;
12
+ }): Promise<T | undefined> {
13
+ const count = moves.get(options.requestKey) ?? 0;
14
+ if (count >= 3 || options.remaining !== undefined && options.remaining <= 0) return;
15
+ const hop = options.reserve("auth-recovery", "account-auto-pause", true);
16
+ if (!hop.allowed) return;
17
+ try {
18
+ if (!await options.refresh()) return;
19
+ moves.set(options.requestKey, count + 1);
20
+ options.setPermit(hop.permit);
21
+ return await options.refetch();
22
+ } finally {
23
+ options.setPermit(undefined);
24
+ hop.permit?.release();
25
+ }
26
+ }
@@ -1,3 +1,4 @@
1
+ import { bindAnthropicRejectionResponse } from "../../adapters/anthropic-rejection-memo";
1
2
  import { rotateAnthropicAccountOnResponseForInstance } from "../../oauth/anthropic-account-refusal";
2
3
  import type { ResponsesRequestContext } from "./core-options";
3
4
  import type { PreparedResponsesRequest } from "./request-prepare";
@@ -29,6 +30,7 @@ import {
29
30
  applyUpstreamRecoveryInit,
30
31
  isNonReplayableResponse,
31
32
  prepareSameTarget429Wait,
33
+ SendBudgetExhaustedError,
32
34
  } from "../../lib/upstream-retry";
33
35
  import { unwrapUpstreamRetryEvidenceError } from "../../codex/auth-context";
34
36
  import { AdmissionModelDeniedError, MODEL_NOT_ALLOWED_FOR_KEY } from "../admission-model-scope";
@@ -121,6 +123,7 @@ export function createAdapterContinuations(
121
123
  | "rateLimitRetries"
122
124
  | "stallTimeoutMs"
123
125
  | "localUpstream"
126
+ | "thinkingSignatureRecovery"
124
127
  >,
125
128
  ) {
126
129
  const { options, logCtx, config } = requestContext;
@@ -159,6 +162,10 @@ export function createAdapterContinuations(
159
162
  errorType: MODEL_NOT_ALLOWED_FOR_KEY, code: MODEL_NOT_ALLOWED_FOR_KEY, retryable: false,
160
163
  };
161
164
  }
165
+ if (refusal instanceof SendBudgetExhaustedError) return {
166
+ type: "error", status: 429, errorType: "rate_limit_error", code: refusal.code,
167
+ message: refusal.message, retryable: false,
168
+ };
162
169
  return {
163
170
  type: "error",
164
171
  message: `Provider continuation failed: ${redactSecretString(error instanceof Error ? error.message : String(error))}`,
@@ -309,6 +316,14 @@ export function createAdapterContinuations(
309
316
  const recoveryKind = nextContinuationRecoveryKind;
310
317
  nextContinuationRecoveryKind = undefined;
311
318
  response = await fetchContinuation(recoveryKind);
319
+ bindAnthropicRejectionResponse(transportState.sameTargetRequest, response);
320
+ if (initialRecoveryKind === "empty-completion" && await adapterExchange.thinkingSignatureRecovery.prepareHttp(
321
+ response, transportState.activeAdapter.name, nextParsed,
322
+ )) {
323
+ try { void response.body?.cancel().catch(() => {}); } catch { /* closed */ }
324
+ invalidateSameTargetRequest();
325
+ response = await fetchContinuation("opaque-blob-rejection");
326
+ }
312
327
  if (kiroRefusalPendingReplay) {
313
328
  try { void kiroRefusalPendingReplay.body?.cancel().catch(() => {}); } catch { /* already closed */ }
314
329
  kiroRefusalPendingReplay = undefined;
@@ -704,24 +719,25 @@ export function createAdapterContinuations(
704
719
  // listener that can cancel its reader even while the parser holds the body lock.
705
720
  // The shared signal is not aborted on timeout: this generator must still emit
706
721
  // the error that the already-live Responses bridge turns into a terminal.
707
- if (nextParsed.stream) {
708
- yield* readResponseStreamWithInactivity(
709
- response,
710
- upstream.signal,
711
- bodyInactivityMs,
712
- guarded => transportState.activeAdapter.parseStream(guarded, translatorBudget, logCtx.activeTierMetadata),
713
- );
714
- } else if (transportState.activeAdapter.parseResponse) {
715
- yield* await readResponseBodyWithInactivity(
716
- response,
717
- upstream.signal,
718
- bodyInactivityMs,
719
- guarded => transportState.activeAdapter.parseResponse!(guarded, translatorBudget, logCtx.activeTierMetadata),
720
- );
721
- } else {
722
- try { void response.body?.cancel().catch(() => {}); } catch { /* already closed */ }
723
- yield { type: "error", message: "Provider continuation does not support response parsing" };
724
- }
722
+ const parse = async function* (current: Response): AsyncGenerator<AdapterEvent> {
723
+ if (!current.ok) {
724
+ yield await adapterExchange.thinkingSignatureRecovery.projectHttpFailure(current, upstream.signal);
725
+ } else if (nextParsed.stream) {
726
+ yield* readResponseStreamWithInactivity(current, upstream.signal, bodyInactivityMs,
727
+ guarded => transportState.activeAdapter.parseStream(guarded, translatorBudget, logCtx.activeTierMetadata));
728
+ } else if (transportState.activeAdapter.parseResponse) {
729
+ yield* await readResponseBodyWithInactivity(current, upstream.signal, bodyInactivityMs,
730
+ guarded => transportState.activeAdapter.parseResponse!(guarded, translatorBudget, logCtx.activeTierMetadata));
731
+ } else {
732
+ try { void current.body?.cancel().catch(() => {}); } catch { /* closed */ }
733
+ yield { type: "error", message: "Provider continuation does not support response parsing" };
734
+ }
735
+ };
736
+ yield* adapterExchange.thinkingSignatureRecovery.wrap(parse(response), async function* () {
737
+ invalidateSameTargetRequest();
738
+ yield* parse(await fetchContinuation("opaque-blob-rejection"));
739
+ }, () => transportState.activeAdapter.name,
740
+ initialRecoveryKind === "empty-completion" && !isNonReplayableResponse(response), response, nextParsed);
725
741
  } catch (error) {
726
742
  // Classify the thrown error before reading signal state: cancelling the stalled
727
743
  // source can synchronously abort the shared signal, which would otherwise report
@@ -23,6 +23,9 @@ import { resolveStallTimeoutMs } from "../../stall-timeout";
23
23
  import { clientEncoderForDelivery, deliverClientEncodedResponse } from "../inference/client-encoder-delivery";
24
24
  import { noteKiroServedSuccess } from "../../providers/kiro-usage";
25
25
  import { persistKiroAccountState } from "../../providers/kiro-account-state-disk";
26
+ import { readDisplaySafeErrorText, normalizeUpstreamErrorText } from "./core-errors";
27
+ import { sanitizeNonReplayableUpstreamError } from "./non-replayable-error";
28
+ import { isNonReplayableResponse } from "../../lib/upstream-retry";
26
29
  import { isCanonicalOpenAiForwardProvider } from "../../providers/openai-tiers";
27
30
  import { guardEmptyToolCatalogEvents, guardEmptyToolCatalogStream } from "../../lib/empty-tool-catalog";
28
31
 
@@ -49,7 +52,7 @@ export async function deliverAdapterResponse(
49
52
  | "notifyResponseComplete"
50
53
  >,
51
54
  completionPolicy: Pick<ResponsesCompletionPolicy, "emptyCompletionGuardEnabled">,
52
- adapterExchange: Pick<AdapterExchange, "upstreamResponse" | "upstream" | "cleanupUpstreamAbort" | "localUpstream">,
55
+ adapterExchange: Pick<AdapterExchange, "upstreamResponse" | "upstream" | "cleanupUpstreamAbort" | "localUpstream" | "thinkingSignatureRecovery" | "refetchThinkingSignature">,
53
56
  continuationState: Pick<AdapterContinuations, "terminalGuardEnabled" | "fetchTerminalGuardContinuation" | "fetchGuardedEmptyCompletionRetry">,
54
57
  ): Promise<Response> {
55
58
  const { logCtx, options, config } = requestContext;
@@ -61,7 +64,7 @@ export async function deliverAdapterResponse(
61
64
  rememberKiroDeliveredFinalAnswer,
62
65
  responseStateOptions,
63
66
  } = requestState;
64
- const { upstreamResponse, upstream, cleanupUpstreamAbort, localUpstream } = adapterExchange;
67
+ const { upstreamResponse, upstream, cleanupUpstreamAbort, localUpstream, thinkingSignatureRecovery, refetchThinkingSignature } = adapterExchange;
65
68
  const {
66
69
  terminalGuardEnabled,
67
70
  fetchTerminalGuardContinuation,
@@ -77,6 +80,43 @@ export async function deliverAdapterResponse(
77
80
  const { routedCompaction } = sidecarState;
78
81
  const bodyInactivityMs = resolveStallTimeoutMs(config.stallTimeoutSec, { localUpstream });
79
82
  const upstreamRequestsStream = parsed.stream || isCanonicalOpenAiForwardProvider(route.provider);
83
+ const recoverThinkingSignature = (events: AsyncIterable<AdapterEvent>): AsyncIterable<AdapterEvent> =>
84
+ thinkingSignatureRecovery.wrap(events, async function* () {
85
+ const result = await refetchThinkingSignature();
86
+ let response = "failed" in result ? result.failed : result;
87
+ // A replacement can be an HTTP/auth/budget refusal; never feed its JSON body to an SSE parser.
88
+ if (!response.ok) {
89
+ // A local refusal closes the upstream signal before returning its generated JSON.
90
+ // Reading that bounded local body still honors caller cancellation.
91
+ const errorSignal = "failed" in result ? options.abortSignal ?? new AbortController().signal : upstream.signal;
92
+ const nonReplayable = isNonReplayableResponse(response);
93
+ if (nonReplayable) response = await sanitizeNonReplayableUpstreamError(response, errorSignal);
94
+ const body = await readDisplaySafeErrorText(response, errorSignal, "Upstream recovery failed");
95
+ const error = normalizeUpstreamErrorText(body, "Upstream recovery failed");
96
+ yield { type: "error", status: response.status, message: error.message ?? error.safeText,
97
+ errorType: error.type ?? "upstream_error", ...(error.code ? { code: error.code } : {}),
98
+ ...(nonReplayable ? { retryable: false } : {}) };
99
+ return;
100
+ }
101
+ try {
102
+ const retryEvents = upstreamRequestsStream
103
+ ? readResponseStreamWithInactivity(response, upstream.signal, bodyInactivityMs,
104
+ guarded => transportState.activeAdapter.parseStream(guarded, translatorBudget, logCtx.activeTierMetadata))
105
+ : (async function* () {
106
+ yield* await readResponseBodyWithInactivity(response, upstream.signal, bodyInactivityMs,
107
+ guarded => transportState.activeAdapter.parseResponse!(guarded, translatorBudget, logCtx.activeTierMetadata));
108
+ })();
109
+ for await (const event of retryEvents) {
110
+ options.onCompactionRecoveryAdapterEvent?.(event);
111
+ yield event;
112
+ }
113
+ } catch (error) {
114
+ if (!(error instanceof ResponseBodyInactivityError)) throw error;
115
+ yield { type: "error", message: "Upstream recovery response body stalled before completing",
116
+ status: 504, errorType: "upstream_error" };
117
+ }
118
+ }, () => transportState.activeAdapter.name, !isNonReplayableResponse(upstreamResponse), upstreamResponse);
119
+
80
120
  const clientToolCatalog = requestState.clientToolCatalogRestriction && route.provider.authMode !== "forward"
81
121
  ? requestState.clientToolCatalogRestriction : undefined;
82
122
  const deliveryToolBridgeMaps = clientToolCatalog
@@ -110,15 +150,16 @@ export async function deliverAdapterResponse(
110
150
  throw error;
111
151
  }
112
152
  })();
153
+ const recoveredEvents = recoverThinkingSignature(initialEventStream);
113
154
  const eventStream = terminalGuardEnabled
114
155
  ? guardTerminalEventStream({
115
156
  parsed,
116
- firstEvents: initialEventStream,
157
+ firstEvents: recoveredEvents,
117
158
  adapterName: transportState.activeAdapter.name,
118
159
  maxAutoContinuations: 1,
119
160
  continuation: next => fetchTerminalGuardContinuation(next, undefined, !parsed.stream),
120
161
  })
121
- : initialEventStream;
162
+ : recoveredEvents;
122
163
  // Optional advisor guard (registered per request by the sidecar planner; absent unless the
123
164
  // user enabled the advisor): holds synthetic `advisor` tool calls, consults the configured
124
165
  // expert model, and re-dispatches the worker. Sits inside the empty-completion guard so an
@@ -147,6 +188,7 @@ export async function deliverAdapterResponse(
147
188
  // One completion owner for both deliveries: the bridge calls it from its terminal, the
148
189
  // direct client encoder from the fold of the same events.
149
190
  const onCompletedResponse = (response: Record<string, unknown>, providerState?: OcxProviderContinuationState) => {
191
+ if (!options.abortSignal?.aborted) thinkingSignatureRecovery.rememberSuccess(response);
150
192
  const served = transportState.sentOAuthSnapshot;
151
193
  if (transportState.activeAdapter.name === "kiro" && response.status === "completed" && served
152
194
  && noteKiroServedSuccess(served.accountId, served.generation)) persistKiroAccountState();
@@ -250,18 +292,22 @@ export async function deliverAdapterResponse(
250
292
  ));
251
293
  }
252
294
  for (const event of initialEvents) options.onCompactionRecoveryAdapterEvent?.(event);
295
+ const recoveredEvents: AdapterEvent[] = [];
296
+ for await (const event of recoverThinkingSignature((async function* () { yield* initialEvents; })())) {
297
+ recoveredEvents.push(event);
298
+ }
253
299
  let guardedEvents: AdapterEvent[];
254
300
  if (terminalGuardEnabled) {
255
301
  guardedEvents = [];
256
302
  for await (const event of guardTerminalEventStream({
257
303
  parsed,
258
- firstEvents: (async function* () { yield* initialEvents; })(),
304
+ firstEvents: (async function* () { yield* recoveredEvents; })(),
259
305
  adapterName: transportState.activeAdapter.name,
260
306
  maxAutoContinuations: 1,
261
307
  continuation: next => fetchTerminalGuardContinuation(next, undefined, !parsed.stream),
262
308
  })) guardedEvents.push(event);
263
309
  } else {
264
- guardedEvents = initialEvents;
310
+ guardedEvents = recoveredEvents;
265
311
  }
266
312
  // Optional advisor guard — same semantics as the streaming branch above.
267
313
  if (parsed._advisorGuard) {
@@ -325,6 +371,7 @@ export async function deliverAdapterResponse(
325
371
  // #1926 gap 2: same buffered-path durability bound as the primary branch.
326
372
  await awaitThoughtSignatureDurability();
327
373
  if (adapterResponseReachedServingTerminal(events, json)) {
374
+ if (!options.abortSignal?.aborted) thinkingSignatureRecovery.rememberSuccess(json);
328
375
  const served = transportState.sentOAuthSnapshot;
329
376
  if (transportState.activeAdapter.name === "kiro" && json.status === "completed" && served
330
377
  && noteKiroServedSuccess(served.accountId, served.generation)) persistKiroAccountState();
@@ -1,4 +1,8 @@
1
+ import { recoverAutoPausedAccount } from "./account-auto-pause-recovery";
2
+ import { wasAccountAutoPaused, wasPermanentAccountRefusal } from "../../oauth/account-auto-pause";
3
+ import { bindAnthropicRejectionResponse } from "../../adapters/anthropic-rejection-memo";
1
4
  import { deferPhysicalSendReporter } from "./deferred-send-reporter";
5
+ import { usesApiKeyAccount } from "../../providers/label";
2
6
  import { spendLedgerStorageErrorResponse } from "./spend-storage-error";
3
7
  import { classifyAnthropic429 } from "../../oauth/anthropic-rate-limit-policy";
4
8
  import { rotateAnthropicAccountOnResponseForInstance } from "../../oauth/anthropic-account-refusal";
@@ -87,6 +91,7 @@ import {
87
91
  reasoningEffortRejectionText,
88
92
  anthropicFastRefused,
89
93
  } from "./core-opaque-recovery";
94
+ import { createAnthropicThinkingRecovery } from "./anthropic-thinking-recovery";
90
95
  import { shouldAttemptImageTierRetry } from "../image-retry";
91
96
  import {
92
97
  isTransientConsoleGoUploadRejection,
@@ -139,6 +144,7 @@ export async function prepareAdapterExchange(
139
144
  | "sameTargetToken"
140
145
  | "transportToken"
141
146
  | "oauthDispatch"
147
+ | "recoverAutoPausedSelection"
142
148
  | "imageTierBias"
143
149
  | "oauth401ReplayAttempted"
144
150
  | "isOAuth401ReplayProvider"
@@ -392,11 +398,12 @@ export async function prepareAdapterExchange(
392
398
  upstreamResponse = new Response(text, { status: original.status, statusText: original.statusText, headers });
393
399
  return upstreamResponse;
394
400
  };
401
+ const autoPauseEnabled = config.providers[route.providerName]?.accountAutoPause?.enabled === true;
395
402
  const billingPoolRequest = isAnthropicBillingKeyPool(route.provider);
396
403
  let billingRecoveryActive = false;
397
- const genericRefetchCountsSends = () => sendBudgetState.adapterSendBudget?.spendEnforced === true
404
+ const genericRefetchCountsSends = () => options.comboAttempt === true || sendBudgetState.adapterSendBudget?.spendEnforced === true
398
405
  || transientRetryPolicyFor(route.provider) !== null || resetReplayPolicyFor(route.provider) !== null
399
- || billingPoolRequest || billingRecoveryActive;
406
+ || billingPoolRequest || billingRecoveryActive || autoPauseEnabled;
400
407
  let handoffPermit: ResponsesSendBudget["pendingHopPermit"];
401
408
  try {
402
409
  if (transportState.activeAdapter.fetchResponse) {
@@ -518,10 +525,10 @@ export async function prepareAdapterExchange(
518
525
  label: safeHostLabel(builtInitialRequest.url),
519
526
  claimAmbiguousResend: claimPreHeaderResend,
520
527
  ...(sendBudgetState.adapterSendBudget?.spendEnforced || options.comboDispatchPermit || transientPolicy || resetPolicy || compactPrepaid || options.comboAttempt
521
- || billingPoolRequest
528
+ || billingPoolRequest || autoPauseEnabled
522
529
  ? {
523
530
  // A pending compaction permit already booked this leg's first physical send.
524
- ...(transientPolicy || resetPolicy || compactPrepaid || options.comboAttempt
531
+ ...(transientPolicy || resetPolicy || compactPrepaid || options.comboAttempt || autoPauseEnabled
525
532
  ? {
526
533
  attempts: Math.min(initialSendCap,
527
534
  sendBudgetState.initialSendAllowance(initialSendCap) + (compactPrepaid ? 1 : 0)),
@@ -591,7 +598,7 @@ export async function prepareAdapterExchange(
591
598
  let rateLimitRetries = 0;
592
599
  // Shared with the terminal-guard continuation below: an image-tier reduction that let the
593
600
  // main request clear a 413 must not be forgotten on the very next continuation build.
594
- if (!upstreamResponse.ok) {
601
+ // Recovery helpers also serve initial HTTP-200 adapter errors during delivery.
595
602
  // Recovery loop: multi-key 429 failover + at most ONE opaque-state rebuild and ONE
596
603
  // anthropic 413 tightened retry
597
604
  // (devlog/260714_image_normalization_pipeline/030). One mutable activeAdapter serves
@@ -600,6 +607,9 @@ export async function prepareAdapterExchange(
600
607
  // 413→429 rotation cannot silently undo the tightening.
601
608
  let imageRetryAttempted = false;
602
609
  const opaqueBlobRecoveryGuard: OpaqueBlobRecoveryGuard = { attempted: false };
610
+ const thinkingSignatureRecovery = createAnthropicThinkingRecovery(
611
+ parsed, upstream.signal, opaqueBlobRecoveryGuard, () => !sendBudgetExhausted(),
612
+ );
603
613
  // Console Go answers a transient 400 "Invalid upload request." for bodies it accepts
604
614
  // moments later; at most one byte-identical replay is allowed per request.
605
615
  const consoleGoUploadRetryGuard: { attempted: boolean } = { attempted: false };
@@ -753,7 +763,9 @@ export async function prepareAdapterExchange(
753
763
  : undefined;
754
764
  let refetchPermitUsed = false;
755
765
  const refetchReporter = refetchAllowance ? transientSendReporter(refetchAllowance.permit) : undefined;
756
- const deferredRefetch = billingPoolRequest && refetchReporter ? deferPhysicalSendReporter(refetchReporter) : undefined;
766
+ const deferredRefetch = (billingPoolRequest || options.comboAttempt) && refetchReporter
767
+ ? deferPhysicalSendReporter(refetchReporter) : undefined;
768
+ const pendingKeyMetadata = deferredRefetch && usesApiKeyAccount(route.provider);
757
769
  try {
758
770
  return await refetchWithPolicy(
759
771
  recoveryKind => {
@@ -762,14 +774,14 @@ export async function prepareAdapterExchange(
762
774
  });
763
775
  // Prepare key metadata before the guarded transport commits its attempt;
764
776
  // counters and permits still wait for the physical executor after pacing.
765
- if (deferredRefetch) transportState.noteRoutedAttemptSend(retryEstimate, recoveryKind ?? recovery);
777
+ if (pendingKeyMetadata) transportState.noteRoutedAttemptSend(retryEstimate, recoveryKind ?? recovery);
766
778
  const confirmDispatch = () => {
767
779
  if (refetchAllowance?.permit && !refetchPermitUsed) {
768
780
  if (!refetchAllowance.permit.use()) throw new SendBudgetExhaustedError(safeHostLabel(retryRequest.url));
769
781
  refetchPermitUsed = true;
770
782
  }
771
- if (!deferredRefetch) transportState.noteRoutedAttemptSend(retryEstimate, recoveryKind ?? recovery);
772
- // Eligible pools confirm after pacing at the physical executor; legacy
783
+ if (!pendingKeyMetadata) transportState.noteRoutedAttemptSend(retryEstimate, recoveryKind ?? recovery);
784
+ // Counted Combo/pool sends confirm at the physical executor; legacy
773
785
  // helper paths retain their existing thunk admission boundary.
774
786
  onDispatch?.();
775
787
  if (preserveFailureResponse) replacementAdmitted = true;
@@ -863,6 +875,11 @@ export async function prepareAdapterExchange(
863
875
  return { failed: formatErrorResponse(502, "upstream_error", msg) };
864
876
  }
865
877
  };
878
+ const refetchThinkingSignature = () => {
879
+ invalidateSameTargetRequest();
880
+ return rebuildAndRefetch("opaque-blob-rejection");
881
+ };
882
+ if (!upstreamResponse.ok) {
866
883
  const rotateAntigravityAuth = async (
867
884
  failedResponse: Response,
868
885
  recovery: AttemptRecoveryKind,
@@ -877,7 +894,7 @@ export async function prepareAdapterExchange(
877
894
  const adapterOwnsDispatch = transportState.activeAdapter.fetchResponse !== undefined;
878
895
  const hop = reserveCredentialHop("auth-recovery",
879
896
  `${route.providerName}|${route.modelId}|antigravity-auth`,
880
- !adapterOwnsDispatch && (transientRetryPolicyFor(route.provider) !== null || resetReplayPolicyFor(route.provider) !== null));
897
+ !adapterOwnsDispatch && genericRefetchCountsSends());
881
898
  if (!hop.allowed) return null;
882
899
  try {
883
900
  const snapshot = await failoverAccountSnapshot(route.providerName, nextId);
@@ -913,6 +930,15 @@ export async function prepareAdapterExchange(
913
930
  };
914
931
  // Keep recovery kinds in sync with the native Responses `passthroughRecovery:` loop above.
915
932
  recovery: for (;;) {
933
+ if (wasAccountAutoPaused(upstreamResponse) && route.provider.authMode !== "forward") {
934
+ const recovered = await recoverAutoPausedAccount({ requestKey: transportState, reserve: reserveCredentialHop,
935
+ remaining: transientSendCapFor(transientRetryPolicyFor(route.provider)?.attempts, sendBudgetState.targetSendsUsed),
936
+ refresh: () => transportState.recoverAutoPausedSelection(parsed), refetch: () => rebuildAndRefetch("oauth-account-429"), setPermit: permit => { sendBudgetState.pendingHopPermit = permit; } });
937
+ if (!recovered) break;
938
+ if ("failed" in recovered) return recovered.failed;
939
+ upstreamResponse = recovered; continue recovery;
940
+ }
941
+
916
942
  // Preserve the terminal verdict through adapter and combo error formatting.
917
943
  // This also covers a reset reached by a 401/429/413 recovery refetch.
918
944
  if (isNonReplayableResponse(upstreamResponse)) {
@@ -920,8 +946,19 @@ export async function prepareAdapterExchange(
920
946
  return await sanitizeNonReplayableUpstreamError(upstreamResponse, upstream.signal);
921
947
  } finally { cleanupUpstreamAbort(); }
922
948
  }
949
+ bindAnthropicRejectionResponse(transportState.sameTargetRequest, upstreamResponse);
950
+ if (await thinkingSignatureRecovery.prepareHttp(
951
+ upstreamResponse, transportState.activeAdapter.name,
952
+ )) {
953
+ try { void upstreamResponse.body?.cancel().catch(() => {}); } catch { /* closed */ }
954
+ invalidateSameTargetRequest();
955
+ const result = await rebuildAndRefetch("opaque-blob-rejection");
956
+ if ("failed" in result) return result.failed;
957
+ upstreamResponse = result;
958
+ continue recovery;
959
+ }
923
960
  if (
924
- upstreamResponse.status === 401
961
+ upstreamResponse.status === 401 && !wasPermanentAccountRefusal(upstreamResponse)
925
962
  && isOAuth401ReplayProvider
926
963
  && transportState.sentOAuthSnapshot
927
964
  && !oauth401ReplayAttempted
@@ -957,7 +994,7 @@ export async function prepareAdapterExchange(
957
994
  const adapterOwnsDispatch = transportState.activeAdapter.fetchResponse !== undefined;
958
995
  const hop = reserveCredentialHop("auth-recovery",
959
996
  `${route.providerName}|${route.modelId}|terminal-refresh-account`,
960
- !adapterOwnsDispatch && (transientRetryPolicyFor(route.provider) !== null || resetReplayPolicyFor(route.provider) !== null));
997
+ !adapterOwnsDispatch && genericRefetchCountsSends());
961
998
  if (hop.allowed) {
962
999
  try {
963
1000
  const admitted = await applyFailoverSnapshot(alternate);
@@ -1041,7 +1078,7 @@ export async function prepareAdapterExchange(
1041
1078
  continue recovery;
1042
1079
  }
1043
1080
 
1044
- if (route.providerName === "google-antigravity" && upstreamResponse.status === 401
1081
+ if (route.providerName === "google-antigravity" && upstreamResponse.status === 401 && !wasPermanentAccountRefusal(upstreamResponse)
1045
1082
  && oauth401ReplayAttempted && !antigravityAuthRotationAttempted
1046
1083
  && !isNonReplayableResponse(upstreamResponse)) {
1047
1084
  antigravityAuthRotationAttempted = true;
@@ -1327,12 +1364,31 @@ export async function prepareAdapterExchange(
1327
1364
  && anthropicInstance
1328
1365
  && transportState.anthropicPoolAccountId
1329
1366
  ) {
1330
- const nextAccountId = await rotateAnthropicAccountOnResponseForInstance(anthropicInstance, upstreamResponse, {
1331
- config, accountId: transportState.anthropicPoolAccountId, sessionKey: anthropicSessionKey,
1332
- model: route.modelId, requestKey: transportState, decision: transportState.anthropicRouteDecision, currentDecision: currentAnthropicRouteDecision, signal: upstream.signal,
1333
- canRetry: !sendBudgetExhausted() && transportState.anthropicPoolFailovers < ANTHROPIC_POOL_MAX_FAILOVERS_PER_REQUEST,
1334
- });
1335
- if (!nextAccountId) break;
1367
+ // Same admission as the other rotation ladders: the roster cap cannot see the shared
1368
+ // send budget. A canRetry term, not a break -- the call also records the refusal.
1369
+ const adapterOwnsDispatch = transportState.activeAdapter.fetchResponse !== undefined;
1370
+ const hop = reserveCredentialHop(
1371
+ "auth-recovery",
1372
+ `${route.providerName}|${route.modelId}|anthropic-pool-rotation`,
1373
+ !adapterOwnsDispatch && genericRefetchCountsSends(),
1374
+ );
1375
+ let nextAccountId: string | null;
1376
+ try {
1377
+ nextAccountId = await rotateAnthropicAccountOnResponseForInstance(anthropicInstance, upstreamResponse, {
1378
+ config, accountId: transportState.anthropicPoolAccountId, sessionKey: anthropicSessionKey,
1379
+ model: route.modelId, requestKey: transportState, decision: transportState.anthropicRouteDecision, currentDecision: currentAnthropicRouteDecision, signal: upstream.signal,
1380
+ // A denied reservation must not rotate on base headroom its bounds cannot see.
1381
+ canRetry: hop.allowed && (hop.permit !== undefined || !sendBudgetExhausted())
1382
+ && transportState.anthropicPoolFailovers < ANTHROPIC_POOL_MAX_FAILOVERS_PER_REQUEST,
1383
+ });
1384
+ } catch (rotationError) {
1385
+ hop.permit?.release();
1386
+ throw rotationError;
1387
+ }
1388
+ if (!nextAccountId) {
1389
+ hop.permit?.release();
1390
+ break;
1391
+ }
1336
1392
  try {
1337
1393
  const admitted = await commitResolvedOAuthSelection(await anthropicRoutingFor(anthropicInstance).getAnthropicPoolAccessSnapshot(nextAccountId));
1338
1394
  if (!admitted) throw new Error("OAuth selection changed during recovery");
@@ -1355,12 +1411,26 @@ export async function prepareAdapterExchange(
1355
1411
  });
1356
1412
  sealRequestAttemptIdentity(logCtx.activeAttempt, logCtx.provider, transportState.activeAdapter.name, logCtx.accountLogLabel);
1357
1413
  recordAttemptCredentialSource(logCtx.activeAttempt, route.providerName, route.provider, transportState.activeAdapter.name);
1358
- const result = await rebuildAndRefetch(upstreamResponse.status === 401 ? "oauth-401"
1359
- : upstreamResponse.status === 403 ? "oauth-account-403" : "anthropic-oauth-429");
1360
- if ("failed" in result) return result.failed;
1361
- upstreamResponse = result;
1414
+ // The replay IS this hop's send; the dispatching layer settles the reservation.
1415
+ sendBudgetState.pendingHopPermit = hop.permit;
1416
+ let result: Response | { failed: Response };
1417
+ try {
1418
+ result = await rebuildAndRefetch(upstreamResponse.status === 401 ? "oauth-401"
1419
+ : upstreamResponse.status === 403 ? "oauth-account-403" : "anthropic-oauth-429",
1420
+ () => { if (!adapterOwnsDispatch) hop.permit?.use(); });
1421
+ } finally {
1422
+ sendBudgetState.pendingHopPermit = undefined;
1423
+ }
1424
+ if ("failed" in result) {
1425
+ // A no-op once the boundary was reached; a refund when the rebuild died first.
1426
+ hop.permit?.release();
1427
+ return result.failed;
1428
+ }
1429
+ upstreamResponse = result;
1362
1430
  if (isNonReplayableResponse(upstreamResponse)) continue recovery;
1363
- } catch {
1431
+ } catch {
1432
+ // release() refunds a pre-send hop and is a no-op once the send settled it.
1433
+ hop.permit?.release();
1364
1434
  break;
1365
1435
  }
1366
1436
  }
@@ -1400,9 +1470,9 @@ export async function prepareAdapterExchange(
1400
1470
  const hop = reserveCredentialHop(
1401
1471
  "auth-recovery",
1402
1472
  `${route.providerName}|${route.modelId}|adapter-recovery-oauth-429`,
1403
- // Transient and opted-in reset helpers report physical sends; an adapter-owned
1473
+ // Counted generic helpers report physical sends; an adapter-owned
1404
1474
  // ladder settles its own booking.
1405
- !adapterOwnsDispatch && (transientRetryPolicyFor(route.provider) !== null || resetReplayPolicyFor(route.provider) !== null),
1475
+ !adapterOwnsDispatch && genericRefetchCountsSends(),
1406
1476
  );
1407
1477
  if (!hop.allowed) break;
1408
1478
  const nextAccountId = rotateGenericOAuthAccountOnRefusal(
@@ -1504,9 +1574,9 @@ export async function prepareAdapterExchange(
1504
1574
  const hop = reserveCredentialHop(
1505
1575
  "auth-recovery",
1506
1576
  `${route.providerName}|${route.modelId}|adapter-recovery-oauth-429`,
1507
- // Transient and opted-in reset helpers report physical sends; an adapter-owned
1577
+ // Counted generic helpers report physical sends; an adapter-owned
1508
1578
  // ladder settles its own booking.
1509
- !adapterOwnsDispatch && (transientRetryPolicyFor(route.provider) !== null || resetReplayPolicyFor(route.provider) !== null),
1579
+ !adapterOwnsDispatch && genericRefetchCountsSends(),
1510
1580
  );
1511
1581
  if (!hop.allowed) break;
1512
1582
  const nextAccountId = rotateGenericOAuthAccountOn429(
@@ -1625,7 +1695,7 @@ export async function prepareAdapterExchange(
1625
1695
  const hop = reserveCredentialHop(
1626
1696
  "auth-recovery",
1627
1697
  `${route.providerName}|${route.modelId}|adapter-recovery-oauth-verify`,
1628
- !adapterOwnsDispatch && (transientRetryPolicyFor(route.provider) !== null || resetReplayPolicyFor(route.provider) !== null),
1698
+ !adapterOwnsDispatch && genericRefetchCountsSends(),
1629
1699
  );
1630
1700
  if (!hop.allowed) break;
1631
1701
  // Auth-refusal rotation, not the rate-limit one: a verification refusal
@@ -1877,6 +1947,8 @@ export async function prepareAdapterExchange(
1877
1947
  authMode: route.provider.authMode,
1878
1948
  hasApiKey: Boolean(route.provider.apiKey?.trim()),
1879
1949
  upstreamRetryAfter,
1950
+ upstreamErrorType: normalized.type,
1951
+ upstreamErrorCode: normalized.code,
1880
1952
  // This recovery path is the HTTP Responses wire; custom runTurn transports
1881
1953
  // never reach enrichOpenCodeZenUpstreamMessage here.
1882
1954
  supportsHttpSameKeyRetry: true,
@@ -1902,6 +1974,7 @@ export async function prepareAdapterExchange(
1902
1974
  }
1903
1975
  }
1904
1976
 
1977
+ bindAnthropicRejectionResponse(transportState.sameTargetRequest, upstreamResponse);
1905
1978
  cancelBodyOnAbort(upstreamResponse.body, upstream.signal);
1906
1979
 
1907
1980
  return {
@@ -1911,6 +1984,8 @@ export async function prepareAdapterExchange(
1911
1984
  stallTimeoutMs,
1912
1985
  localUpstream,
1913
1986
  upstreamResponse,
1987
+ thinkingSignatureRecovery,
1988
+ refetchThinkingSignature,
1914
1989
  rateLimitPolicy,
1915
1990
  get rateLimitRetries(): typeof rateLimitRetries {
1916
1991
  return rateLimitRetries;