@juspay/neurolink 10.12.5 → 10.12.7
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.
- package/CHANGELOG.md +12 -0
- package/dist/browser/neurolink.min.js +329 -329
- package/dist/cli/commands/proxy.js +13 -17
- package/dist/cli/commands/proxyAnalyze.js +8 -3
- package/dist/lib/providers/googleVertex/client.js +9 -1
- package/dist/lib/proxy/logCleanupScheduler.d.ts +12 -0
- package/dist/lib/proxy/logCleanupScheduler.js +74 -0
- package/dist/lib/proxy/logCleanupWorkerEntry.d.ts +1 -0
- package/dist/lib/proxy/logCleanupWorkerEntry.js +15 -0
- package/dist/lib/proxy/proxyAnalysis.js +6 -0
- package/dist/lib/proxy/requestLogger.d.ts +6 -0
- package/dist/lib/proxy/requestLogger.js +57 -47
- package/dist/lib/proxy/rollingWorkerSupervisor.d.ts +2 -0
- package/dist/lib/proxy/rollingWorkerSupervisor.js +40 -2
- package/dist/lib/server/routes/claudeProxyRoutes.d.ts +2 -0
- package/dist/lib/server/routes/claudeProxyRoutes.js +33 -2
- package/dist/lib/types/proxy.d.ts +32 -1
- package/dist/lib/utils/messageBuilder.js +12 -0
- package/dist/providers/googleVertex/client.js +9 -1
- package/dist/proxy/logCleanupScheduler.d.ts +12 -0
- package/dist/proxy/logCleanupScheduler.js +73 -0
- package/dist/proxy/logCleanupWorkerEntry.d.ts +1 -0
- package/dist/proxy/logCleanupWorkerEntry.js +14 -0
- package/dist/proxy/proxyAnalysis.js +6 -0
- package/dist/proxy/requestLogger.d.ts +6 -0
- package/dist/proxy/requestLogger.js +57 -47
- package/dist/proxy/rollingWorkerSupervisor.d.ts +2 -0
- package/dist/proxy/rollingWorkerSupervisor.js +40 -2
- package/dist/server/routes/claudeProxyRoutes.d.ts +2 -0
- package/dist/server/routes/claudeProxyRoutes.js +33 -2
- package/dist/types/proxy.d.ts +32 -1
- package/dist/utils/messageBuilder.js +12 -0
- package/package.json +4 -2
|
@@ -646,6 +646,7 @@ export type AnthropicSuccessResult = {
|
|
|
646
646
|
failure?: {
|
|
647
647
|
message: string;
|
|
648
648
|
rateLimit: boolean;
|
|
649
|
+
retryDelayMs?: number;
|
|
649
650
|
};
|
|
650
651
|
} | {
|
|
651
652
|
response: Response | unknown;
|
|
@@ -702,6 +703,8 @@ export type AnthropicAuthRetryResult = {
|
|
|
702
703
|
response?: Response | unknown;
|
|
703
704
|
holdsAccountAdmission?: boolean;
|
|
704
705
|
continueLoop: boolean;
|
|
706
|
+
/** Failure-path pacing before rotating after provider-wide overload. */
|
|
707
|
+
retryDelayMs?: number;
|
|
705
708
|
lastError: unknown;
|
|
706
709
|
authFailureMessage: string | null;
|
|
707
710
|
sawRateLimit: boolean;
|
|
@@ -713,6 +716,8 @@ export type AnthropicNonOkResult = {
|
|
|
713
716
|
response?: Response | unknown;
|
|
714
717
|
continueLoop: boolean;
|
|
715
718
|
retrySameAccount?: boolean;
|
|
719
|
+
/** Failure-path pacing before rotating after provider-wide overload. */
|
|
720
|
+
retryDelayMs?: number;
|
|
716
721
|
lastError: unknown;
|
|
717
722
|
authFailureMessage: string | null;
|
|
718
723
|
sawTransientFailure: boolean;
|
|
@@ -1457,6 +1462,9 @@ export type ProxyAnalysisReport = {
|
|
|
1457
1462
|
attemptLatency: boolean;
|
|
1458
1463
|
cacheUsage: boolean;
|
|
1459
1464
|
routingDecisions: boolean;
|
|
1465
|
+
/** True only when every stream needed for cross-stream request/attempt
|
|
1466
|
+
* reconciliation begins at or before the requested analysis window. */
|
|
1467
|
+
comparableRequestAttempts: boolean;
|
|
1460
1468
|
};
|
|
1461
1469
|
dataQuality: {
|
|
1462
1470
|
linesRead: number;
|
|
@@ -1468,6 +1476,8 @@ export type ProxyAnalysisReport = {
|
|
|
1468
1476
|
observedFrom: string | null;
|
|
1469
1477
|
observedTo: string | null;
|
|
1470
1478
|
startsAtOrBeforeRequestedWindow: boolean;
|
|
1479
|
+
/** Whether this stream can support claims covering the full window. */
|
|
1480
|
+
completeWindow: boolean;
|
|
1471
1481
|
}>;
|
|
1472
1482
|
bodyArtifacts: {
|
|
1473
1483
|
capturesIndexed: number;
|
|
@@ -2073,6 +2083,14 @@ export type RollingWorkerFailureDetails = {
|
|
|
2073
2083
|
workerExitSignal?: string | null;
|
|
2074
2084
|
supervisorAction?: "none" | "sigkill_after_transfer_failure";
|
|
2075
2085
|
};
|
|
2086
|
+
export type RollingWorkerSupervisorEvent = {
|
|
2087
|
+
at: string;
|
|
2088
|
+
type: "activated" | "failure" | "failed_transfer" | "rejected_socket";
|
|
2089
|
+
generation: number | null;
|
|
2090
|
+
version: string | null;
|
|
2091
|
+
phase?: "startup" | "activation" | "runtime" | "transfer";
|
|
2092
|
+
reason?: string;
|
|
2093
|
+
};
|
|
2076
2094
|
export type RollingWorkerSupervisorSnapshot = {
|
|
2077
2095
|
generation: number;
|
|
2078
2096
|
active: {
|
|
@@ -2093,6 +2111,8 @@ export type RollingWorkerSupervisorSnapshot = {
|
|
|
2093
2111
|
queuedSockets: number;
|
|
2094
2112
|
rejectedSockets: number;
|
|
2095
2113
|
failedTransfers: number;
|
|
2114
|
+
/** Bounded generation-scoped evidence for attributing lifetime counters. */
|
|
2115
|
+
recentEvents: RollingWorkerSupervisorEvent[];
|
|
2096
2116
|
lastFailure: ({
|
|
2097
2117
|
at: string;
|
|
2098
2118
|
generation: number;
|
|
@@ -2305,7 +2325,18 @@ export type ProxyNeurolinkRuntime = {
|
|
|
2305
2325
|
neurolink: {
|
|
2306
2326
|
getToolRegistry(): MCPToolRegistry;
|
|
2307
2327
|
};
|
|
2308
|
-
|
|
2328
|
+
logsDir: string;
|
|
2329
|
+
};
|
|
2330
|
+
/** Data passed to the isolated proxy log-retention worker. */
|
|
2331
|
+
export type ProxyLogCleanupWorkerData = {
|
|
2332
|
+
logsDir: string;
|
|
2333
|
+
maxAgeDays: number;
|
|
2334
|
+
maxSizeMb: number;
|
|
2335
|
+
};
|
|
2336
|
+
/** Lifecycle handle for non-blocking proxy log retention. */
|
|
2337
|
+
export type ProxyLogCleanupScheduler = {
|
|
2338
|
+
trigger: () => boolean;
|
|
2339
|
+
stop: () => Promise<void>;
|
|
2309
2340
|
};
|
|
2310
2341
|
/** Hono app + readiness state created by the proxy start command. */
|
|
2311
2342
|
export type ProxyStartApp = {
|
|
@@ -1073,6 +1073,18 @@ export async function processUnifiedFilesArray(options, maxSize, provider) {
|
|
|
1073
1073
|
options.systemPrompt = existingSystem
|
|
1074
1074
|
? `${existingSystem}${filePromptAugmentation}`
|
|
1075
1075
|
: filePromptAugmentation.trim();
|
|
1076
|
+
// Keep options.prompt in sync with the enriched input.text.
|
|
1077
|
+
// neurolink.ts snapshots `prompt: options.input?.text` at baseOptions
|
|
1078
|
+
// creation — BEFORE this function appends file content — and some
|
|
1079
|
+
// native paths (executeNativeAnthropicGenerate historically) read
|
|
1080
|
+
// `options.prompt` first. Leaving the stale snapshot in place means
|
|
1081
|
+
// every appended file is silently dropped on those paths and the model
|
|
1082
|
+
// answers "no file attached". Same dual-write processCSVFilesForNativeSDK
|
|
1083
|
+
// already documents for exactly this trap.
|
|
1084
|
+
const promptCarrier = options;
|
|
1085
|
+
if (typeof promptCarrier.prompt === "string") {
|
|
1086
|
+
promptCarrier.prompt = inp2.text;
|
|
1087
|
+
}
|
|
1076
1088
|
}
|
|
1077
1089
|
});
|
|
1078
1090
|
}
|
|
@@ -4815,7 +4815,15 @@ export class GoogleVertexProvider extends BaseProvider {
|
|
|
4815
4815
|
});
|
|
4816
4816
|
// Build messages from input
|
|
4817
4817
|
const messages = [];
|
|
4818
|
-
|
|
4818
|
+
// input.text FIRST — the file preprocessors (processUnifiedFilesArray,
|
|
4819
|
+
// processCSVFilesForNativeSDK) append attached-file content to input.text,
|
|
4820
|
+
// while options.prompt is snapshotted from the ORIGINAL input.text at
|
|
4821
|
+
// baseOptions creation (neurolink.ts). With prompt-first precedence every
|
|
4822
|
+
// attached file was silently dropped on the Claude generate path — the
|
|
4823
|
+
// model answered "no file attached" to a request that carried one. The
|
|
4824
|
+
// native Gemini path already reads input.text first for exactly this
|
|
4825
|
+
// reason.
|
|
4826
|
+
const inputText = options.input?.text || options.prompt || "Please respond.";
|
|
4819
4827
|
// Replay conversationMessages (with tool turns), else the legacy text-only conversationHistory.
|
|
4820
4828
|
if (options.conversationMessages &&
|
|
4821
4829
|
options.conversationMessages.length > 0) {
|
|
@@ -0,0 +1,12 @@
|
|
|
1
|
+
import type { ProxyLogCleanupScheduler } from "../types/index.js";
|
|
2
|
+
/**
|
|
3
|
+
* Runs retention in a worker so large log trees cannot delay readiness or
|
|
4
|
+
* block active streams. Concurrent runs are coalesced into the active scan.
|
|
5
|
+
*/
|
|
6
|
+
export declare function startProxyLogCleanupScheduler(params: {
|
|
7
|
+
logsDir: string;
|
|
8
|
+
maxAgeDays?: number;
|
|
9
|
+
maxSizeMb?: number;
|
|
10
|
+
initialDelayMs?: number;
|
|
11
|
+
intervalMs?: number;
|
|
12
|
+
}): ProxyLogCleanupScheduler;
|
|
@@ -0,0 +1,73 @@
|
|
|
1
|
+
import { Worker } from "node:worker_threads";
|
|
2
|
+
import { withTimeout } from "../utils/async/withTimeout.js";
|
|
3
|
+
import { logger } from "../utils/logger.js";
|
|
4
|
+
const DEFAULT_INITIAL_DELAY_MS = 30_000;
|
|
5
|
+
const DEFAULT_INTERVAL_MS = 60 * 60 * 1000;
|
|
6
|
+
const WORKER_TERMINATION_TIMEOUT_MS = 5_000;
|
|
7
|
+
/**
|
|
8
|
+
* Runs retention in a worker so large log trees cannot delay readiness or
|
|
9
|
+
* block active streams. Concurrent runs are coalesced into the active scan.
|
|
10
|
+
*/
|
|
11
|
+
export function startProxyLogCleanupScheduler(params) {
|
|
12
|
+
const maxAgeDays = params.maxAgeDays ?? 7;
|
|
13
|
+
const maxSizeMb = params.maxSizeMb ?? 500;
|
|
14
|
+
let activeWorker;
|
|
15
|
+
let stopped = false;
|
|
16
|
+
const trigger = () => {
|
|
17
|
+
if (stopped || activeWorker) {
|
|
18
|
+
return false;
|
|
19
|
+
}
|
|
20
|
+
let worker;
|
|
21
|
+
try {
|
|
22
|
+
worker = new Worker(params.workerUrl ??
|
|
23
|
+
new URL("./logCleanupWorkerEntry.js", import.meta.url), {
|
|
24
|
+
execArgv: process.execArgv.filter((argument) => !argument.startsWith("--input-type")),
|
|
25
|
+
workerData: {
|
|
26
|
+
logsDir: params.logsDir,
|
|
27
|
+
maxAgeDays,
|
|
28
|
+
maxSizeMb,
|
|
29
|
+
},
|
|
30
|
+
});
|
|
31
|
+
}
|
|
32
|
+
catch (error) {
|
|
33
|
+
logger.debug(`[proxy] could not start background log cleanup: ${error instanceof Error ? error.message : String(error)}`);
|
|
34
|
+
return false;
|
|
35
|
+
}
|
|
36
|
+
activeWorker = worker;
|
|
37
|
+
worker.unref();
|
|
38
|
+
worker.once("error", (error) => {
|
|
39
|
+
logger.debug(`[proxy] background log cleanup failed: ${error instanceof Error ? error.message : String(error)}`);
|
|
40
|
+
});
|
|
41
|
+
worker.once("exit", (code) => {
|
|
42
|
+
if (activeWorker === worker) {
|
|
43
|
+
activeWorker = undefined;
|
|
44
|
+
}
|
|
45
|
+
if (code !== 0 && !stopped) {
|
|
46
|
+
logger.debug(`[proxy] background log cleanup exited with code ${code}`);
|
|
47
|
+
}
|
|
48
|
+
});
|
|
49
|
+
return true;
|
|
50
|
+
};
|
|
51
|
+
const initialTimer = setTimeout(trigger, params.initialDelayMs ?? DEFAULT_INITIAL_DELAY_MS);
|
|
52
|
+
initialTimer.unref();
|
|
53
|
+
const interval = setInterval(trigger, params.intervalMs ?? DEFAULT_INTERVAL_MS);
|
|
54
|
+
interval.unref();
|
|
55
|
+
return {
|
|
56
|
+
trigger,
|
|
57
|
+
stop: async () => {
|
|
58
|
+
stopped = true;
|
|
59
|
+
clearTimeout(initialTimer);
|
|
60
|
+
clearInterval(interval);
|
|
61
|
+
const worker = activeWorker;
|
|
62
|
+
activeWorker = undefined;
|
|
63
|
+
if (worker) {
|
|
64
|
+
try {
|
|
65
|
+
await withTimeout(worker.terminate(), WORKER_TERMINATION_TIMEOUT_MS, "Timed out terminating the proxy log cleanup worker");
|
|
66
|
+
}
|
|
67
|
+
catch (error) {
|
|
68
|
+
logger.debug(`[proxy] background log cleanup termination failed: ${error instanceof Error ? error.message : String(error)}`);
|
|
69
|
+
}
|
|
70
|
+
}
|
|
71
|
+
},
|
|
72
|
+
};
|
|
73
|
+
}
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
export {};
|
|
@@ -0,0 +1,14 @@
|
|
|
1
|
+
import { parentPort, workerData } from "node:worker_threads";
|
|
2
|
+
import { cleanupLogsAt } from "./requestLogger.js";
|
|
3
|
+
const data = workerData;
|
|
4
|
+
try {
|
|
5
|
+
cleanupLogsAt(data.logsDir, data.maxAgeDays, data.maxSizeMb);
|
|
6
|
+
parentPort?.postMessage({ ok: true });
|
|
7
|
+
}
|
|
8
|
+
catch (error) {
|
|
9
|
+
parentPort?.postMessage({
|
|
10
|
+
ok: false,
|
|
11
|
+
error: error instanceof Error ? error.message : String(error),
|
|
12
|
+
});
|
|
13
|
+
process.exitCode = 1;
|
|
14
|
+
}
|
|
@@ -737,6 +737,10 @@ export async function analyzeProxyLogs(options) {
|
|
|
737
737
|
const artifactsReferenced = artifactsPresent + artifactsMissing;
|
|
738
738
|
const finalSummary = summarizeFinalRequests(finalRequests, terminalStreamErrors, attemptsByRequest, accounts);
|
|
739
739
|
const routingSummary = summarizeRouting(finalRequests);
|
|
740
|
+
const streamComplete = (stream) => {
|
|
741
|
+
const range = observedRanges[stream];
|
|
742
|
+
return range.from !== null && range.from <= sinceMs;
|
|
743
|
+
};
|
|
740
744
|
return {
|
|
741
745
|
generatedAt: new Date(nowMs).toISOString(),
|
|
742
746
|
since: new Date(sinceMs).toISOString(),
|
|
@@ -755,6 +759,7 @@ export async function analyzeProxyLogs(options) {
|
|
|
755
759
|
attemptLatency: attemptLatency.length > 0,
|
|
756
760
|
cacheUsage: finalSummary.cache.requestsWithUsage > 0,
|
|
757
761
|
routingDecisions: routingSummary.totalRecords > 0,
|
|
762
|
+
comparableRequestAttempts: streamComplete("requests") && streamComplete("attempts"),
|
|
758
763
|
},
|
|
759
764
|
dataQuality: {
|
|
760
765
|
linesRead,
|
|
@@ -768,6 +773,7 @@ export async function analyzeProxyLogs(options) {
|
|
|
768
773
|
observedFrom: range.from === null ? null : new Date(range.from).toISOString(),
|
|
769
774
|
observedTo: range.to === null ? null : new Date(range.to).toISOString(),
|
|
770
775
|
startsAtOrBeforeRequestedWindow: range.from !== null && range.from <= sinceMs,
|
|
776
|
+
completeWindow: range.from !== null && range.from <= sinceMs,
|
|
771
777
|
},
|
|
772
778
|
])),
|
|
773
779
|
bodyArtifacts: {
|
|
@@ -70,3 +70,9 @@ export declare function logStreamError(entry: {
|
|
|
70
70
|
* Non-fatal — proxy keeps working even if cleanup fails.
|
|
71
71
|
*/
|
|
72
72
|
export declare function cleanupLogs(maxAgeDays?: number, maxSizeMb?: number): void;
|
|
73
|
+
/**
|
|
74
|
+
* Path-scoped retention implementation used by the proxy cleanup worker.
|
|
75
|
+
* This function is intentionally synchronous: callers must run it outside the
|
|
76
|
+
* request-serving process when the directory can contain many artifacts.
|
|
77
|
+
*/
|
|
78
|
+
export declare function cleanupLogsAt(activeLogDir: string, maxAgeDays?: number, maxSizeMb?: number): void;
|
|
@@ -700,60 +700,70 @@ export async function logStreamError(entry) {
|
|
|
700
700
|
* Non-fatal — proxy keeps working even if cleanup fails.
|
|
701
701
|
*/
|
|
702
702
|
export function cleanupLogs(maxAgeDays = 7, maxSizeMb = 500) {
|
|
703
|
-
if (!logDir
|
|
703
|
+
if (!logDir) {
|
|
704
704
|
return;
|
|
705
705
|
}
|
|
706
706
|
try {
|
|
707
|
-
|
|
708
|
-
|
|
709
|
-
|
|
710
|
-
|
|
711
|
-
|
|
712
|
-
|
|
713
|
-
|
|
714
|
-
|
|
715
|
-
|
|
716
|
-
|
|
717
|
-
|
|
718
|
-
|
|
719
|
-
|
|
720
|
-
|
|
721
|
-
|
|
722
|
-
|
|
723
|
-
|
|
724
|
-
|
|
725
|
-
|
|
726
|
-
|
|
727
|
-
|
|
728
|
-
|
|
729
|
-
|
|
730
|
-
|
|
731
|
-
|
|
732
|
-
|
|
733
|
-
|
|
734
|
-
const deletionCandidates = remaining.filter(canDelete);
|
|
735
|
-
// Current-day metadata is the only reliable source for final-request,
|
|
736
|
-
// attempt, lifecycle, and body-index reconciliation. Keep those indexes
|
|
737
|
-
// intact during size cleanup; body artifacts and older indexes remain
|
|
738
|
-
// eligible for eviction.
|
|
739
|
-
while (totalSize > maxBytes && deletionCandidates.length > 0) {
|
|
740
|
-
const oldest = deletionCandidates.shift();
|
|
741
|
-
if (!oldest) {
|
|
742
|
-
break;
|
|
743
|
-
}
|
|
744
|
-
unlinkSync(oldest.path);
|
|
745
|
-
totalSize -= oldest.size;
|
|
707
|
+
cleanupLogsAt(logDir, maxAgeDays, maxSizeMb);
|
|
708
|
+
}
|
|
709
|
+
catch {
|
|
710
|
+
// Non-fatal for legacy in-process callers.
|
|
711
|
+
}
|
|
712
|
+
}
|
|
713
|
+
/**
|
|
714
|
+
* Path-scoped retention implementation used by the proxy cleanup worker.
|
|
715
|
+
* This function is intentionally synchronous: callers must run it outside the
|
|
716
|
+
* request-serving process when the directory can contain many artifacts.
|
|
717
|
+
*/
|
|
718
|
+
export function cleanupLogsAt(activeLogDir, maxAgeDays = 7, maxSizeMb = 500) {
|
|
719
|
+
if (!existsSync(activeLogDir)) {
|
|
720
|
+
return;
|
|
721
|
+
}
|
|
722
|
+
const files = collectManagedLogFiles(activeLogDir).sort((a, b) => a.mtime - b.mtime); // oldest first
|
|
723
|
+
const currentDate = new Date().toISOString().split("T")[0];
|
|
724
|
+
const currentMetadataLogs = new Set(["proxy", "proxy-attempts", "proxy-debug", "proxy-lifecycle"].map((prefix) => join(activeLogDir, `${prefix}-${currentDate}.jsonl`)));
|
|
725
|
+
const canDelete = (file) => !currentMetadataLogs.has(file.path);
|
|
726
|
+
const cutoff = Date.now() - maxAgeDays * 24 * 60 * 60 * 1000;
|
|
727
|
+
let deletedCount = 0;
|
|
728
|
+
let freedBytes = 0;
|
|
729
|
+
// Pass 1: delete files older than maxAgeDays
|
|
730
|
+
const remaining = [];
|
|
731
|
+
for (const file of files) {
|
|
732
|
+
if (file.mtime < cutoff && canDelete(file)) {
|
|
733
|
+
unlinkSync(file.path);
|
|
746
734
|
deletedCount++;
|
|
747
|
-
freedBytes +=
|
|
735
|
+
freedBytes += file.size;
|
|
748
736
|
}
|
|
749
|
-
|
|
750
|
-
|
|
737
|
+
else {
|
|
738
|
+
remaining.push(file);
|
|
751
739
|
}
|
|
752
|
-
|
|
753
|
-
|
|
740
|
+
}
|
|
741
|
+
const bodiesDir = join(activeLogDir, "bodies");
|
|
742
|
+
if (existsSync(bodiesDir)) {
|
|
743
|
+
pruneEmptyDirectories(bodiesDir, bodiesDir);
|
|
744
|
+
}
|
|
745
|
+
// Pass 2: if total size exceeds maxSizeMb, delete oldest until under limit
|
|
746
|
+
const maxBytes = maxSizeMb * 1024 * 1024;
|
|
747
|
+
let totalSize = remaining.reduce((sum, f) => sum + f.size, 0);
|
|
748
|
+
const deletionCandidates = remaining.filter(canDelete);
|
|
749
|
+
// Current-day metadata is the only reliable source for final-request,
|
|
750
|
+
// attempt, lifecycle, and body-index reconciliation. Keep those indexes
|
|
751
|
+
// intact during size cleanup; body artifacts and older indexes remain
|
|
752
|
+
// eligible for eviction.
|
|
753
|
+
while (totalSize > maxBytes && deletionCandidates.length > 0) {
|
|
754
|
+
const oldest = deletionCandidates.shift();
|
|
755
|
+
if (!oldest) {
|
|
756
|
+
break;
|
|
754
757
|
}
|
|
758
|
+
unlinkSync(oldest.path);
|
|
759
|
+
totalSize -= oldest.size;
|
|
760
|
+
deletedCount++;
|
|
761
|
+
freedBytes += oldest.size;
|
|
755
762
|
}
|
|
756
|
-
|
|
757
|
-
|
|
763
|
+
if (existsSync(bodiesDir)) {
|
|
764
|
+
pruneEmptyDirectories(bodiesDir, bodiesDir);
|
|
765
|
+
}
|
|
766
|
+
if (deletedCount > 0) {
|
|
767
|
+
logger.info(`[proxy] log cleanup: deleted ${deletedCount} file(s), freed ${(freedBytes / 1024 / 1024).toFixed(1)} MB`);
|
|
758
768
|
}
|
|
759
769
|
}
|
|
@@ -14,6 +14,7 @@ export declare class RollingWorkerSupervisor {
|
|
|
14
14
|
private replacement;
|
|
15
15
|
private rejectedSockets;
|
|
16
16
|
private failedTransfers;
|
|
17
|
+
private readonly recentEvents;
|
|
17
18
|
private lastFailure;
|
|
18
19
|
private closed;
|
|
19
20
|
private shutdownPromise;
|
|
@@ -38,5 +39,6 @@ export declare class RollingWorkerSupervisor {
|
|
|
38
39
|
private describeTransferError;
|
|
39
40
|
private extractLifecycleFailureDetails;
|
|
40
41
|
private recordFailure;
|
|
42
|
+
private recordEvent;
|
|
41
43
|
private publishState;
|
|
42
44
|
}
|
|
@@ -3,6 +3,7 @@ const DEFAULT_READY_TIMEOUT_MS = 30_000;
|
|
|
3
3
|
const DEFAULT_SOCKET_QUEUE_LIMIT = 1_024;
|
|
4
4
|
const DEFAULT_SOCKET_QUEUE_TIMEOUT_MS = 30_000;
|
|
5
5
|
const DEFAULT_SHUTDOWN_TIMEOUT_MS = 30_000;
|
|
6
|
+
const MAX_RECENT_SUPERVISOR_EVENTS = 100;
|
|
6
7
|
/**
|
|
7
8
|
* Owns worker generations while the caller owns the public listening socket.
|
|
8
9
|
* Sockets are transferred once to the active worker, so response bytes never
|
|
@@ -18,6 +19,7 @@ export class RollingWorkerSupervisor {
|
|
|
18
19
|
replacement = null;
|
|
19
20
|
rejectedSockets = 0;
|
|
20
21
|
failedTransfers = 0;
|
|
22
|
+
recentEvents = [];
|
|
21
23
|
lastFailure = null;
|
|
22
24
|
closed = false;
|
|
23
25
|
shutdownPromise = null;
|
|
@@ -56,6 +58,7 @@ export class RollingWorkerSupervisor {
|
|
|
56
58
|
queuedSockets: this.queuedSockets.length,
|
|
57
59
|
rejectedSockets: this.rejectedSockets,
|
|
58
60
|
failedTransfers: this.failedTransfers,
|
|
61
|
+
recentEvents: [...this.recentEvents],
|
|
59
62
|
lastFailure: this.lastFailure,
|
|
60
63
|
};
|
|
61
64
|
}
|
|
@@ -302,6 +305,11 @@ export class RollingWorkerSupervisor {
|
|
|
302
305
|
this.maybeDrainWorker(previous);
|
|
303
306
|
}
|
|
304
307
|
this.options.log?.(`[proxy-supervisor] activated generation=${generation} pid=${handle.pid} version=${expectedVersion}`);
|
|
308
|
+
this.recordEvent({
|
|
309
|
+
type: "activated",
|
|
310
|
+
generation,
|
|
311
|
+
version: expectedVersion,
|
|
312
|
+
});
|
|
305
313
|
this.publishState();
|
|
306
314
|
finish();
|
|
307
315
|
});
|
|
@@ -410,6 +418,13 @@ export class RollingWorkerSupervisor {
|
|
|
410
418
|
handleTransferFailure(worker, socket, error) {
|
|
411
419
|
this.failedTransfers += 1;
|
|
412
420
|
const detail = this.describeTransferError(error);
|
|
421
|
+
this.recordEvent({
|
|
422
|
+
type: "failed_transfer",
|
|
423
|
+
generation: worker.generation,
|
|
424
|
+
version: worker.version,
|
|
425
|
+
phase: "transfer",
|
|
426
|
+
reason: detail,
|
|
427
|
+
});
|
|
413
428
|
const lifecycle = this.extractLifecycleFailureDetails(error, worker.handle.pid);
|
|
414
429
|
this.recordFailure(worker.generation, worker.version, "transfer", `worker ${worker.handle.pid} failed to accept a transferred socket: ${detail}`, {
|
|
415
430
|
...lifecycle.details,
|
|
@@ -432,10 +447,16 @@ export class RollingWorkerSupervisor {
|
|
|
432
447
|
}
|
|
433
448
|
this.publishState();
|
|
434
449
|
}
|
|
435
|
-
this.rejectSocket(socket);
|
|
450
|
+
this.rejectSocket(socket, worker.generation, worker.version, "transfer_failure");
|
|
436
451
|
}
|
|
437
|
-
rejectSocket(socket) {
|
|
452
|
+
rejectSocket(socket, generation = this.active?.generation ?? null, version = this.active?.version ?? null, reason = "unavailable") {
|
|
438
453
|
this.rejectedSockets += 1;
|
|
454
|
+
this.recordEvent({
|
|
455
|
+
type: "rejected_socket",
|
|
456
|
+
generation,
|
|
457
|
+
version,
|
|
458
|
+
reason,
|
|
459
|
+
});
|
|
439
460
|
socket.destroy();
|
|
440
461
|
this.publishState();
|
|
441
462
|
}
|
|
@@ -479,6 +500,23 @@ export class RollingWorkerSupervisor {
|
|
|
479
500
|
message: message.slice(0, 1_000),
|
|
480
501
|
...details,
|
|
481
502
|
};
|
|
503
|
+
this.recordEvent({
|
|
504
|
+
type: "failure",
|
|
505
|
+
generation,
|
|
506
|
+
version,
|
|
507
|
+
phase,
|
|
508
|
+
reason: message,
|
|
509
|
+
});
|
|
510
|
+
}
|
|
511
|
+
recordEvent(event) {
|
|
512
|
+
this.recentEvents.push({
|
|
513
|
+
at: new Date().toISOString(),
|
|
514
|
+
...event,
|
|
515
|
+
...(event.reason ? { reason: event.reason.slice(0, 1_000) } : {}),
|
|
516
|
+
});
|
|
517
|
+
if (this.recentEvents.length > MAX_RECENT_SUPERVISOR_EVENTS) {
|
|
518
|
+
this.recentEvents.splice(0, this.recentEvents.length - MAX_RECENT_SUPERVISOR_EVENTS);
|
|
519
|
+
}
|
|
482
520
|
}
|
|
483
521
|
publishState() {
|
|
484
522
|
try {
|
|
@@ -310,6 +310,7 @@ declare function shouldAttemptClaudeFallback(loopState: AnthropicLoopState): boo
|
|
|
310
310
|
export declare function createClaudeProxyRoutes(modelRouter?: ModelRouterInterface, basePath?: string, accountStrategy?: "round-robin" | "fill-first", passthroughMode?: boolean, primaryAccountKey?: string, accountAllowlistOrRuntimeOptions?: AccountAllowlist | ClaudeProxyRouteRuntimeOptions): RouteGroup;
|
|
311
311
|
declare function reconcileEligibleAccountRuntimeState(account: ProxyPassthroughAccount): void;
|
|
312
312
|
export declare function getTransientSameAccountRetryDelayMs(retryNumber: number): number;
|
|
313
|
+
export declare function getOverloadRotationDelayMs(attemptNumber: number): number;
|
|
313
314
|
declare function describeTransportError(error: unknown): string;
|
|
314
315
|
/**
|
|
315
316
|
* Determine whether a POST can be retried without risking duplicate provider
|
|
@@ -400,6 +401,7 @@ export declare const __testHooks: {
|
|
|
400
401
|
describeTransportError: typeof describeTransportError;
|
|
401
402
|
redactProviderErrorMessage: typeof redactProviderErrorMessage;
|
|
402
403
|
isUpstreamOverload: typeof isUpstreamOverload;
|
|
404
|
+
getOverloadRotationDelayMs: typeof getOverloadRotationDelayMs;
|
|
403
405
|
shouldAttemptClaudeFallback: typeof shouldAttemptClaudeFallback;
|
|
404
406
|
executeClaudeFallbackWithRetry: typeof executeClaudeFallbackWithRetry;
|
|
405
407
|
buildClaudeAnthropicFailureResponse: typeof buildClaudeAnthropicFailureResponse;
|
|
@@ -63,6 +63,7 @@ let lastKnownAccountCount = 0;
|
|
|
63
63
|
const MAX_AUTH_RETRIES = 5;
|
|
64
64
|
const MAX_TRANSIENT_SAME_ACCOUNT_RETRIES = 2;
|
|
65
65
|
const TRANSIENT_SAME_ACCOUNT_RETRY_DELAYS_MS = [250, 1_000];
|
|
66
|
+
const OVERLOAD_ACCOUNT_ROTATION_DELAYS_MS = [250, 500, 1_000, 2_000];
|
|
66
67
|
const MAX_FALLBACK_NETWORK_RETRIES = 1;
|
|
67
68
|
const FALLBACK_STREAM_IDLE_TIMEOUT_MS = 2 * 60 * 1000;
|
|
68
69
|
/** Maximum upstream 429 attempts per account before rotating — for a TRANSIENT
|
|
@@ -2883,7 +2884,13 @@ async function handleAnthropicStreamingSuccessResponse(args) {
|
|
|
2883
2884
|
});
|
|
2884
2885
|
return {
|
|
2885
2886
|
retryNextAccount: true,
|
|
2886
|
-
failure: {
|
|
2887
|
+
failure: {
|
|
2888
|
+
message: preflight.message,
|
|
2889
|
+
rateLimit: isRateLimit,
|
|
2890
|
+
...(preflight.errorType === "overloaded_error"
|
|
2891
|
+
? { retryDelayMs: getOverloadRotationDelayMs(attemptNumber) }
|
|
2892
|
+
: {}),
|
|
2893
|
+
},
|
|
2887
2894
|
};
|
|
2888
2895
|
}
|
|
2889
2896
|
logAttempt(response.status, undefined, undefined, {
|
|
@@ -3484,6 +3491,9 @@ async function handleAnthropicAuthRetry(args) {
|
|
|
3484
3491
|
const failure = successResult.failure;
|
|
3485
3492
|
return {
|
|
3486
3493
|
continueLoop: true,
|
|
3494
|
+
...(failure?.retryDelayMs
|
|
3495
|
+
? { retryDelayMs: failure.retryDelayMs }
|
|
3496
|
+
: {}),
|
|
3487
3497
|
lastError: failure?.message ?? currentLastError,
|
|
3488
3498
|
authFailureMessage: currentAuthFailureMessage,
|
|
3489
3499
|
sawRateLimit: currentSawRateLimit || Boolean(failure?.rateLimit),
|
|
@@ -3941,6 +3951,9 @@ async function handleAnthropicNonOkResponse(args) {
|
|
|
3941
3951
|
return {
|
|
3942
3952
|
continueLoop: true,
|
|
3943
3953
|
retrySameAccount: !upstreamOverload,
|
|
3954
|
+
...(upstreamOverload
|
|
3955
|
+
? { retryDelayMs: getOverloadRotationDelayMs(attemptNumber) }
|
|
3956
|
+
: {}),
|
|
3944
3957
|
lastError: currentLastError,
|
|
3945
3958
|
authFailureMessage: currentAuthFailureMessage,
|
|
3946
3959
|
sawTransientFailure: currentSawTransientFailure,
|
|
@@ -4553,7 +4566,8 @@ async function handleAnthropicRoutedClaudeRequest(args) {
|
|
|
4553
4566
|
effectiveAccounts.every((account) => !isAccountAdmissionAvailable(account.key, accountAdmissionCapacity))) {
|
|
4554
4567
|
queuedAccountAdmission = await acquireFirstAvailableAccountAdmission(effectiveAccounts.map((account) => account.key), accountAdmissionCapacity, ctx.abortSignal);
|
|
4555
4568
|
}
|
|
4556
|
-
accountLoop: for (const account of effectiveAccounts) {
|
|
4569
|
+
accountLoop: for (const [accountIndex, account,] of effectiveAccounts.entries()) {
|
|
4570
|
+
const hasNextAccount = accountIndex < effectiveAccounts.length - 1;
|
|
4557
4571
|
const accountState = getOrCreateRuntimeState(account.key);
|
|
4558
4572
|
let transientSameAccountRetries = 0;
|
|
4559
4573
|
let rateLimitSameAccountRetries = 0;
|
|
@@ -4763,6 +4777,10 @@ async function handleAnthropicRoutedClaudeRequest(args) {
|
|
|
4763
4777
|
return authRetryResult.response;
|
|
4764
4778
|
}
|
|
4765
4779
|
if (authRetryResult.continueLoop) {
|
|
4780
|
+
if (hasNextAccount && authRetryResult.retryDelayMs) {
|
|
4781
|
+
logger.always(`[proxy] pacing cross-account SSE overload rotation for ${authRetryResult.retryDelayMs}ms after auth refresh`);
|
|
4782
|
+
await sleep(authRetryResult.retryDelayMs);
|
|
4783
|
+
}
|
|
4766
4784
|
continue accountLoop;
|
|
4767
4785
|
}
|
|
4768
4786
|
}
|
|
@@ -4804,6 +4822,10 @@ async function handleAnthropicRoutedClaudeRequest(args) {
|
|
|
4804
4822
|
if (nonOkResult.retrySameAccount) {
|
|
4805
4823
|
logger.always(`[proxy] exhausted transient same-account retries for account=${account.label}; rotating`);
|
|
4806
4824
|
}
|
|
4825
|
+
if (hasNextAccount && nonOkResult.retryDelayMs) {
|
|
4826
|
+
logger.always(`[proxy] pacing cross-account overload rotation for ${nonOkResult.retryDelayMs}ms`);
|
|
4827
|
+
await sleep(nonOkResult.retryDelayMs);
|
|
4828
|
+
}
|
|
4807
4829
|
continue accountLoop;
|
|
4808
4830
|
}
|
|
4809
4831
|
break accountLoop;
|
|
@@ -4845,6 +4867,10 @@ async function handleAnthropicRoutedClaudeRequest(args) {
|
|
|
4845
4867
|
loopState.lastError = successResult.failure.message;
|
|
4846
4868
|
loopState.sawRateLimit ||= successResult.failure.rateLimit;
|
|
4847
4869
|
loopState.sawTransientFailure ||= !successResult.failure.rateLimit;
|
|
4870
|
+
if (hasNextAccount && successResult.failure.retryDelayMs) {
|
|
4871
|
+
logger.always(`[proxy] pacing cross-account SSE overload rotation for ${successResult.failure.retryDelayMs}ms`);
|
|
4872
|
+
await sleep(successResult.failure.retryDelayMs);
|
|
4873
|
+
}
|
|
4848
4874
|
}
|
|
4849
4875
|
continue accountLoop;
|
|
4850
4876
|
}
|
|
@@ -5315,6 +5341,10 @@ export function getTransientSameAccountRetryDelayMs(retryNumber) {
|
|
|
5315
5341
|
const index = Math.min(Math.max(retryNumber - 1, 0), TRANSIENT_SAME_ACCOUNT_RETRY_DELAYS_MS.length - 1);
|
|
5316
5342
|
return TRANSIENT_SAME_ACCOUNT_RETRY_DELAYS_MS[index] ?? 0;
|
|
5317
5343
|
}
|
|
5344
|
+
export function getOverloadRotationDelayMs(attemptNumber) {
|
|
5345
|
+
const index = Math.min(Math.max(attemptNumber - 1, 0), OVERLOAD_ACCOUNT_ROTATION_DELAYS_MS.length - 1);
|
|
5346
|
+
return jitteredDelay(OVERLOAD_ACCOUNT_ROTATION_DELAYS_MS[index] ?? 250);
|
|
5347
|
+
}
|
|
5318
5348
|
async function sleep(ms) {
|
|
5319
5349
|
await new Promise((resolve) => setTimeout(resolve, ms));
|
|
5320
5350
|
}
|
|
@@ -5586,6 +5616,7 @@ export const __testHooks = {
|
|
|
5586
5616
|
describeTransportError,
|
|
5587
5617
|
redactProviderErrorMessage,
|
|
5588
5618
|
isUpstreamOverload,
|
|
5619
|
+
getOverloadRotationDelayMs,
|
|
5589
5620
|
shouldAttemptClaudeFallback,
|
|
5590
5621
|
executeClaudeFallbackWithRetry,
|
|
5591
5622
|
buildClaudeAnthropicFailureResponse,
|