replicas-engine 0.1.1310 → 0.1.1312

package/dist/src/{asp-host-3CZEGHOZ.js → asp-host-5YBXALSS.js}

@@ -8,9 +8,9 @@ import {
   listRunningCodexModels,
   restartCodexAspHost,
   restartCodexAspHostIfRunning
-} from "./chunk-77XPI2HV.js";
+} from "./chunk-4FC4NBAH.js";
 import "./chunk-FJGC3SKY.js";
-import "./chunk-GEKCOYIC.js";
+import "./chunk-D6BGWZZE.js";
 import "./chunk-PKLNM4OL.js";
 import "./chunk-UZSNFLDQ.js";
 import "./chunk-BG2GDBAV.js";

package/dist/src/{chunk-UGRKIEEG.js → chunk-3KI4Z7U5.js}

@@ -3,7 +3,7 @@ import { createRequire as __createRequire } from 'node:module';
 const require = __createRequire(import.meta.url);
 import {
   AspClient
-} from "./chunk-GEKCOYIC.js";
+} from "./chunk-D6BGWZZE.js";
 import {
   ASTER_BASE_URL,
   ASTER_PROVIDER,

package/dist/src/{chunk-77XPI2HV.js → chunk-4FC4NBAH.js}

@@ -14,7 +14,7 @@ import {
 import {
   AppServerProcess,
   buildCodexAgentEnv
-} from "./chunk-GEKCOYIC.js";
+} from "./chunk-D6BGWZZE.js";
 import {
   SUBPROCESS_MAX_BUFFER,
   execFileAsync
@@ -306,7 +306,7 @@ var CodexTokenManager = class extends BaseRefreshManager {
     await this.applyCredentialUpdate(++this.credentialGeneration, async () => {
       await this.applyCredentialsResponse(data);
       setAgentCredentialSnapshot("codex", data);
-      const { restartCodexAspHostIfRunning: restartCodexAspHostIfRunning2 } = await import("./asp-host-3CZEGHOZ.js");
+      const { restartCodexAspHostIfRunning: restartCodexAspHostIfRunning2 } = await import("./asp-host-5YBXALSS.js");
       await restartCodexAspHostIfRunning2();
     });
     await this.start(true);
@@ -328,7 +328,7 @@ var CodexTokenManager = class extends BaseRefreshManager {
       const accountChanged = ENGINE_ENV.REPLICAS_AGENT_CREDENTIALS.codex?.credentialId !== data.credentialId;
       await this.applyCredentialsResponse(data);
       if (restartOnChange && (accountChanged || CODEX_AUTH_ENV_KEYS.some((key, index) => process.env[key] !== previousEnv[index]))) {
-        const { restartCodexAspHostIfRunning: restartCodexAspHostIfRunning2 } = await import("./asp-host-3CZEGHOZ.js");
+        const { restartCodexAspHostIfRunning: restartCodexAspHostIfRunning2 } = await import("./asp-host-5YBXALSS.js");
         await restartCodexAspHostIfRunning2();
       }
       setAgentCredentialSnapshot("codex", data);

package/dist/src/{chunk-GEKCOYIC.js → chunk-D6BGWZZE.js}

@@ -240,7 +240,7 @@ function getCodexModelProvider(method) {
   if (method === "foundry") return "azure";
   return method === "ollama" || method === "openai-compatible" ? "replicas_openai_compatible" : "openai";
 }
-var ENGINE_PACKAGE_VERSION = "0.1.1310";
+var ENGINE_PACKAGE_VERSION = "0.1.1312";
 var INITIALIZE_METHOD = "initialize";
 var INITIALIZED_NOTIFICATION = "initialized";
 var ACCOUNT_LOGIN_START_METHOD = "account/login/start";

package/dist/src/headless-agent.js

@@ -14,11 +14,11 @@ import {
   recoverCompletedTurn,
   spawnMuseMsp,
   startMuseSession
-} from "./chunk-UGRKIEEG.js";
+} from "./chunk-3KI4Z7U5.js";
 import {
   AppServerProcess,
   buildCodexAgentEnv
-} from "./chunk-GEKCOYIC.js";
+} from "./chunk-D6BGWZZE.js";
 import "./chunk-PKLNM4OL.js";
 import {
   AGENT,

package/dist/src/index.js

@@ -65,7 +65,7 @@ import {
   spawnMuseMsp,
   startMuseSession,
   uuidV7
-} from "./chunk-UGRKIEEG.js";
+} from "./chunk-3KI4Z7U5.js";
 import {
   publishedPullRequestUrls
 } from "./chunk-CPBB3JKK.js";
@@ -143,7 +143,7 @@ import {
   timestampFromSeconds,
   upsertCredentialFileLines,
   writeSecureCredentialFile
-} from "./chunk-77XPI2HV.js";
+} from "./chunk-4FC4NBAH.js";
 import {
   ENGINE_ENV,
   monolithRequest,
@@ -156,7 +156,7 @@ import {
 } from "./chunk-FJGC3SKY.js";
 import {
   getCodexModelProvider
-} from "./chunk-GEKCOYIC.js";
+} from "./chunk-D6BGWZZE.js";
 import {
   SUBPROCESS_MAX_BUFFER,
   execAsync,
@@ -11330,12 +11330,12 @@ var AuthFallbackCoordinator = class {
   record;
   context;
   /** `attempted` is the credential the failed turn ran on, which another chat may already have replaced. */
-  fallBack(reason, detail, attempted, resetsAt) {
+  fallBack(reason, detail, attempted, resetsAt, afterLimit = false) {
     let lock = fallbackLocks.get(this.provider);
     if (!lock) fallbackLocks.set(this.provider, lock = new AsyncLock());
-    return lock.run(() => this.fallBackNow(reason, detail, attempted, resetsAt));
+    return lock.run(() => this.fallBackNow(reason, detail, attempted, resetsAt, afterLimit));
   }
-  async fallBackNow(reason, detail, attempted, resetsAt) {
+  async fallBackNow(reason, detail, attempted, resetsAt, afterLimit) {
     const tokenManager = SWAPPABLE_AGENTS[this.provider];
     const live = ENGINE_ENV.REPLICAS_AGENT_CREDENTIALS[this.provider];
     const exhausted = attempted ?? live;
@@ -11357,7 +11357,7 @@ var AuthFallbackCoordinator = class {
     for (const step of steps) {
       if (step.kind === "harness") {
         if (!forkable || !isAgentAvailable(step.provider)) continue;
-        if (reason === "auth_failed" || reason === "model_unavailable") continue;
+        if ((reason === "auth_failed" || reason === "model_unavailable") && !afterLimit) continue;
         return { outcome: "fork", target: step };
       }
       if (alreadyReplaced && live) {
@@ -11737,9 +11737,9 @@ var CodingAgentManager = class _CodingAgentManager {
       this.getHistorySink()?.append(event);
     }, options.fallbackContext ?? (() => ({ excludeProviders: [], forkable: false }))) : null;
   }
-  async fallBackCredential(reason, detail, attempted, resetsAt) {
+  async fallBackCredential(reason, detail, attempted, resetsAt, afterLimit = false) {
     if (!this.authFallback) return { outcome: "unavailable" };
-    const result = await this.authFallback.fallBack(reason, detail, attempted, resetsAt);
+    const result = await this.authFallback.fallBack(reason, detail, attempted, resetsAt, afterLimit);
     if (result.outcome === "switched" && result.model) await this.onModelSwitched(result.model);
     return result;
   }
@@ -13479,6 +13479,7 @@ var ClaudeManager = class _ClaudeManager extends CodingAgentManager {
     let loginRevoked = false;
     let namedLogin;
     let failedCredentials = [];
+    let walkLimit;
     let attempt = 0;
     let authRetries = 0;
     let transientRetries = 0;
@@ -13538,8 +13539,10 @@ var ClaudeManager = class _ClaudeManager extends CodingAgentManager {
             const detail = error instanceof ClaudeQuotaError || error instanceof ClaudeModelUnavailableError ? error.message : extractErrorText(error);
             const repeated = !!attemptedCredential && failedCredentials.some((failed) => isSameCredential(failed, attemptedCredential));
             if (attemptedCredential) failedCredentials.push(attemptedCredential);
-            const fallback = repeated ? { outcome: "unavailable" } : await this.fallBackCredential(reason, detail, attemptedCredential, error instanceof ClaudeQuotaError ? error.resetsAt : void 0);
+            const limit = quotaKind ? { kind: quotaKind, credential: attemptedCredential } : walkLimit;
+            const fallback = repeated ? { outcome: "unavailable" } : await this.fallBackCredential(reason, detail, attemptedCredential, error instanceof ClaudeQuotaError ? error.resetsAt : void 0, !!walkLimit);
             if (fallback.outcome === "switched") {
+              walkLimit = limit;
               if (isAuthError) {
                 this.turnLoginFailure ??= describeLoginFailure("claude", attemptedCredential ?? namedLogin, resolveClaudeAuthMethodForMessage(), loginRevoked);
               }
@@ -13551,8 +13554,11 @@ var ClaudeManager = class _ClaudeManager extends CodingAgentManager {
               attempt++;
               continue;
             }
-            if (quotaKind) {
-              await this.emitQuotaExhaustedEvent(quotaKind, fallback, detail, attemptedCredential);
+            if (limit && (quotaKind || fallback.outcome === "fork")) {
+              if (isAuthError) {
+                this.turnLoginFailure ??= describeLoginFailure("claude", attemptedCredential ?? namedLogin, resolveClaudeAuthMethodForMessage(), loginRevoked);
+              }
+              await this.emitQuotaExhaustedEvent(limit.kind, fallback, detail, limit.credential);
               return;
             }
             if (modelUnavailable) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "replicas-engine",
-  "version": "0.1.1310",
+  "version": "0.1.1312",
   "description": "Lightweight API server for Replicas workspaces",
   "type": "module",
   "main": "dist/src/index.js",