@north-light/crouter 0.3.273 → 0.3.275

package/dist/builtin-pi-packages/pi-crtr-extensions/extensions/provider-rotation.d.ts

@@ -6,6 +6,17 @@ export interface ProviderRotationOptions {
     nodeCwd?: string;
     profileId?: string;
 }
+type ProviderErrorClassification = {
+    kind: "rate_limit" | "transient" | "fatal";
+    status?: number;
+    reason: string;
+    hadRetryAfterHeader: boolean;
+};
+/** Test seam for the policy mapping over the shared recognition signal. */
+export declare function __classifyProviderErrorForTest(error: unknown, observed?: {
+    status?: number;
+    retryAfterMs?: number;
+}): ProviderErrorClassification;
 declare function defaultSleep(ms: number): Promise<void>;
 /** Test-only seam: override the sleep implementation used for in-place/wait-out retries. Pass undefined to restore the default. */
 export declare function __setSleepForTest(fn: typeof defaultSleep | undefined): void;

package/dist/builtin-pi-packages/pi-crtr-extensions/extensions/provider-rotation.js

@@ -6,6 +6,7 @@ import { readMergedLaunchConfig } from '../../../core/config.js';
 import { expandModelCandidates, modelRequestFromConfig, probeRouteAvailability } from '../../../core/model-routes.js';
 import { envModelExact } from '../../../shared/env.js';
 import { ManagedProviderCoolingError } from '../../../core/runtime/managed-provider-cooling.js';
+import { recognizeProviderError } from '../../../core/provider-error-recognition.js';
 // A response without Retry-After proves only that THIS attempt was throttled; it
 // does not authorize crouter to invent a five-minute provider-wide outage. Keep
 // a short turn-local probe backoff to avoid an immediate retry loop, then let a
@@ -35,39 +36,42 @@ function configuredValue(state, key, envName) {
 function isExactModel(state) {
     return state.options === undefined && envModelExact();
 }
-const TRANSIENT_MESSAGE_PATTERN = /connection|network|fetch failed|socket|econnreset|econnrefused|enotfound|etimedout|timed? out|timeout|overloaded|service.?unavailable|temporarily unavailable/i;
-const RATE_LIMIT_MESSAGE_PATTERN = /\b429\b|rate.?limit|too many requests/i;
-// Precise terminal usage-limit signals (as opposed to the bare `usage`/`quota` substrings
-// dropped above): pi-ai's Codex adapter surfaces `usage_limit_reached` / `usage_not_included`
-// as the friendly "You have hit your ChatGPT usage limit..." message, often on a non-429
-// status, so the status-code check above misses it. These are specific enough not to false-
-// positive on ordinary assistant text mentioning "usage".
-const USAGE_LIMIT_MESSAGE_PATTERN = /usage[ _]limit|usage_not_included|usagelimiterror|insufficient_quota|quota exceeded/i;
+function classifyProviderSignal(signal) {
+    const hadRetryAfterHeader = signal.retryAfterMs !== undefined;
+    if (signal.kind === 'rate-limit' && signal.matchedOn === 'status') {
+        return { kind: 'rate_limit', status: signal.status, reason: 'http 429', hadRetryAfterHeader };
+    }
+    if (signal.kind === 'transient-http') {
+        if (signal.status === 503 && hadRetryAfterHeader) {
+            return { kind: 'rate_limit', status: signal.status, reason: 'http 503 with retry-after', hadRetryAfterHeader };
+        }
+        return { kind: 'transient', status: signal.status, reason: `http ${signal.status}`, hadRetryAfterHeader };
+    }
+    const textMatches = signal.textMatches ?? (signal.matchedOn === 'text' && signal.match ? [{ kind: signal.kind, match: signal.match }] : []);
+    const transientMatch = textMatches.find(({ kind }) => kind === 'connection' || kind === 'socket' || kind === 'service-unavailable' || kind === 'overloaded');
+    if (transientMatch) {
+        return { kind: 'transient', status: signal.status, reason: transientMatch.match.toLowerCase(), hadRetryAfterHeader };
+    }
+    const rateLimitMatch = textMatches.find(({ kind }) => kind === 'rate-limit');
+    if (rateLimitMatch) {
+        return { kind: 'rate_limit', status: signal.status, reason: rateLimitMatch.match.toLowerCase(), hadRetryAfterHeader };
+    }
+    return { kind: 'fatal', status: signal.status, reason: 'unclassified', hadRetryAfterHeader };
+}
 function classifyProviderError(error, observed) {
     const value = error;
-    const errorStatus = typeof value?.status === "number" ? value.status : undefined;
-    const status = observed.status ?? errorStatus;
-    const message = [value?.errorMessage, value?.message, error instanceof Error ? error.message : ""]
-        .filter((part) => typeof part === "string" && part.length > 0)
-        .join(" ");
-    const hadRetryAfterHeader = observed.retryAfterMs !== undefined;
-    if (status === 429)
-        return { kind: "rate_limit", status, reason: "http 429", hadRetryAfterHeader };
-    if (status === 503 && hadRetryAfterHeader)
-        return { kind: "rate_limit", status, reason: "http 503 with retry-after", hadRetryAfterHeader };
-    if (status === 500 || status === 502 || status === 503 || status === 504) {
-        return { kind: "transient", status, reason: `http ${status}`, hadRetryAfterHeader };
-    }
-    const transientMatch = message.match(TRANSIENT_MESSAGE_PATTERN);
-    if (transientMatch)
-        return { kind: "transient", status, reason: transientMatch[0].toLowerCase(), hadRetryAfterHeader };
-    const rateLimitMatch = message.match(RATE_LIMIT_MESSAGE_PATTERN);
-    if (rateLimitMatch)
-        return { kind: "rate_limit", status, reason: rateLimitMatch[0].toLowerCase(), hadRetryAfterHeader };
-    const usageLimitMatch = message.match(USAGE_LIMIT_MESSAGE_PATTERN);
-    if (usageLimitMatch)
-        return { kind: "rate_limit", status, reason: usageLimitMatch[0].toLowerCase(), hadRetryAfterHeader };
-    return { kind: "fatal", status, reason: "unclassified", hadRetryAfterHeader };
+    const message = [value?.errorMessage, value?.message, error instanceof Error ? error.message : '']
+        .filter((part) => typeof part === 'string' && part.length > 0)
+        .join(' ');
+    return classifyProviderSignal(recognizeProviderError({
+        status: observed.status ?? value?.status,
+        retryAfterMs: observed.retryAfterMs,
+        message,
+    }));
+}
+/** Test seam for the policy mapping over the shared recognition signal. */
+export function __classifyProviderErrorForTest(error, observed = {}) {
+    return classifyProviderError(error, observed);
 }
 function providerErrorMessage(error) {
     const value = error;

package/dist/builtin-pi-packages/pi-crtr-extensions/extensions/provider-rotation.ts

@@ -41,6 +41,7 @@ import { readMergedLaunchConfig } from '../../../core/config.js';
 import { expandModelCandidates, modelRequestFromConfig, probeRouteAvailability } from '../../../core/model-routes.js';
 import { envModelExact } from '../../../shared/env.js';
 import { ManagedProviderCoolingError } from '../../../core/runtime/managed-provider-cooling.js';
+import { recognizeProviderError, type ProviderErrorSignal } from '../../../core/provider-error-recognition.js';
 
 // A response without Retry-After proves only that THIS attempt was throttled; it
 // does not authorize crouter to invent a five-minute provider-wide outage. Keep
@@ -123,37 +124,44 @@ type ProviderErrorClassification = {
   hadRetryAfterHeader: boolean;
 };
 
-const TRANSIENT_MESSAGE_PATTERN =
-  /connection|network|fetch failed|socket|econnreset|econnrefused|enotfound|etimedout|timed? out|timeout|overloaded|service.?unavailable|temporarily unavailable/i;
-const RATE_LIMIT_MESSAGE_PATTERN = /\b429\b|rate.?limit|too many requests/i;
-// Precise terminal usage-limit signals (as opposed to the bare `usage`/`quota` substrings
-// dropped above): pi-ai's Codex adapter surfaces `usage_limit_reached` / `usage_not_included`
-// as the friendly "You have hit your ChatGPT usage limit..." message, often on a non-429
-// status, so the status-code check above misses it. These are specific enough not to false-
-// positive on ordinary assistant text mentioning "usage".
-const USAGE_LIMIT_MESSAGE_PATTERN = /usage[ _]limit|usage_not_included|usagelimiterror|insufficient_quota|quota exceeded/i;
+function classifyProviderSignal(signal: ProviderErrorSignal): ProviderErrorClassification {
+  const hadRetryAfterHeader = signal.retryAfterMs !== undefined;
+  if (signal.kind === 'rate-limit' && signal.matchedOn === 'status') {
+    return { kind: 'rate_limit', status: signal.status, reason: 'http 429', hadRetryAfterHeader };
+  }
+  if (signal.kind === 'transient-http') {
+    if (signal.status === 503 && hadRetryAfterHeader) {
+      return { kind: 'rate_limit', status: signal.status, reason: 'http 503 with retry-after', hadRetryAfterHeader };
+    }
+    return { kind: 'transient', status: signal.status, reason: `http ${signal.status}`, hadRetryAfterHeader };
+  }
+  const textMatches = signal.textMatches ?? (signal.matchedOn === 'text' && signal.match ? [{ kind: signal.kind, match: signal.match }] : []);
+  const transientMatch = textMatches.find(({ kind }) => kind === 'connection' || kind === 'socket' || kind === 'service-unavailable' || kind === 'overloaded');
+  if (transientMatch) {
+    return { kind: 'transient', status: signal.status, reason: transientMatch.match.toLowerCase(), hadRetryAfterHeader };
+  }
+  const rateLimitMatch = textMatches.find(({ kind }) => kind === 'rate-limit');
+  if (rateLimitMatch) {
+    return { kind: 'rate_limit', status: signal.status, reason: rateLimitMatch.match.toLowerCase(), hadRetryAfterHeader };
+  }
+  return { kind: 'fatal', status: signal.status, reason: 'unclassified', hadRetryAfterHeader };
+}
 
 function classifyProviderError(error: unknown, observed: { status?: number; retryAfterMs?: number }): ProviderErrorClassification {
   const value = error as { status?: unknown; errorMessage?: unknown; message?: unknown } | undefined;
-  const errorStatus = typeof value?.status === "number" ? value.status : undefined;
-  const status = observed.status ?? errorStatus;
-  const message = [value?.errorMessage, value?.message, error instanceof Error ? error.message : ""]
-    .filter((part): part is string => typeof part === "string" && part.length > 0)
-    .join(" ");
-  const hadRetryAfterHeader = observed.retryAfterMs !== undefined;
+  const message = [value?.errorMessage, value?.message, error instanceof Error ? error.message : '']
+    .filter((part): part is string => typeof part === 'string' && part.length > 0)
+    .join(' ');
+  return classifyProviderSignal(recognizeProviderError({
+    status: observed.status ?? value?.status,
+    retryAfterMs: observed.retryAfterMs,
+    message,
+  }));
+}
 
-  if (status === 429) return { kind: "rate_limit", status, reason: "http 429", hadRetryAfterHeader };
-  if (status === 503 && hadRetryAfterHeader) return { kind: "rate_limit", status, reason: "http 503 with retry-after", hadRetryAfterHeader };
-  if (status === 500 || status === 502 || status === 503 || status === 504) {
-    return { kind: "transient", status, reason: `http ${status}`, hadRetryAfterHeader };
-  }
-  const transientMatch = message.match(TRANSIENT_MESSAGE_PATTERN);
-  if (transientMatch) return { kind: "transient", status, reason: transientMatch[0].toLowerCase(), hadRetryAfterHeader };
-  const rateLimitMatch = message.match(RATE_LIMIT_MESSAGE_PATTERN);
-  if (rateLimitMatch) return { kind: "rate_limit", status, reason: rateLimitMatch[0].toLowerCase(), hadRetryAfterHeader };
-  const usageLimitMatch = message.match(USAGE_LIMIT_MESSAGE_PATTERN);
-  if (usageLimitMatch) return { kind: "rate_limit", status, reason: usageLimitMatch[0].toLowerCase(), hadRetryAfterHeader };
-  return { kind: "fatal", status, reason: "unclassified", hadRetryAfterHeader };
+/** Test seam for the policy mapping over the shared recognition signal. */
+export function __classifyProviderErrorForTest(error: unknown, observed: { status?: number; retryAfterMs?: number } = {}): ProviderErrorClassification {
+  return classifyProviderError(error, observed);
 }
 
 function providerErrorMessage(error: unknown): string | undefined {