open-multi-agent-kit 0.95.2 → 0.96.0

package/dist/core/agent-session.js

@@ -50,6 +50,8 @@ import { buildRuntimeProvenance } from "./runtime-provenance.js";
 import { TurnMetricsSink } from "./turn-metrics.js";
 /** Values that disable a session-level feature through its environment variable. */
 const DISABLED_ENV_VALUES = new Set(["0", "false", "off", "disable", "disabled"]);
+const MAX_OVERFLOW_RECOVERY_ATTEMPTS = 2;
+const OVERFLOW_RECOVERY_EMERGENCY_TOKENS = 4_096;
 function isDisabledEnvValue(value) {
     return value !== undefined && DISABLED_ENV_VALUES.has(value.trim().toLowerCase());
 }
@@ -70,7 +72,7 @@ function firstTextContent(result) {
 import { redactCredentialShapedContent } from "./compaction/transaction.js";
 import { findExactModelReferenceMatch } from "./model-resolver.js";
 import { expandPromptTemplate } from "./prompt-templates.js";
-import { isContentSafetyStopMessage, isStickySafetyModel, isTransientProviderErrorMessage, pickFailoverCandidate, resolveProviderResilience, stickySafetyBlockMessage, } from "./provider-resilience.js";
+import { isContentSafetyStopMessage, isQuotaExhaustionMessage, isStickySafetyModel, isTransientProviderErrorMessage, pickFailoverCandidate, resolveProviderResilience, stickySafetyBlockMessage, } from "./provider-resilience.js";
 import { getBiasStepsForCell, parseRouterBiasSnapshot } from "./reasoning-router-bias.js";
 import { classifyTaskV4, deriveRouterFeedbackFeaturesV4, resolveThinkingLevelV4WithUncertainty, TASK_CLASS_THINKING_LEVELS_V4, } from "./reasoning-router-v4.js";
 import { redactSensitiveText, redactSensitiveTextForced } from "./redaction.js";
@@ -302,7 +304,7 @@ export class AgentSession {
     // Compaction state
     _compactionAbortController = undefined;
     _autoCompactionAbortController = undefined;
-    _overflowRecoveryAttempted = false;
+    _overflowRecoveryAttempts = 0;
     _thresholdCompactionEmergency = false;
     _compactionHysteresisState = createCompactionHysteresisState();
     // Branch summarization state
@@ -585,12 +587,15 @@ export class AgentSession {
         if (/stop_reason\s*=\s*(refusal|sensitive)|content\/safety stop|safety stop|provider\.refusal|kind=provider_refusal/i.test(text)) {
             return { area: "provider", code: "refusal" };
         }
+        // Quota/billing exhaustion is checked BEFORE the generic 401/403 auth
+        // patterns: "403 ... usage limit for this billing cycle" is transient per
+        // cycle and must fail over, not terminate the turn as an auth error.
+        if (isQuotaExhaustionMessage(text) || /rate.?limit|too many requests|429/i.test(text)) {
+            return { area: "provider", code: "rate_limit" };
+        }
         if (/auth|unauthori[sz]ed|forbidden|invalid.?api.?key|no api key|401|403|\/login/i.test(text)) {
             return { area: "provider", code: "auth" };
         }
-        if (/rate.?limit|too many requests|429|quota|available balance|billing/i.test(text)) {
-            return { area: "provider", code: "rate_limit" };
-        }
         // Kimi/K3 + OpenAI-compat: orphan tool results after dropped error assistants.
         // Sanitize-and-retry (transform-messages drops orphans), not a hard tool fatal.
         if (/tool_call_id\s+is\s+not\s+found|tool_call_id\s+not\s+found|unknown\s+tool_call_id/i.test(text)) {
@@ -609,6 +614,9 @@ export class AgentSession {
         if (!this.model || /no model|model selected|model is required/i.test(message)) {
             return { area: "configuration", code: "invalid" };
         }
+        if (isQuotaExhaustionMessage(message)) {
+            return { area: "provider", code: "rate_limit" };
+        }
         if (/auth|api key|unauthori[sz]ed|forbidden|401|403|\/login/i.test(message)) {
             return { area: "provider", code: "auth" };
         }
@@ -966,7 +974,7 @@ export class AgentSession {
         // When a user message starts, check if it's from either queue and remove it BEFORE emitting
         // This ensures the UI sees the updated queue state
         if (event.type === "message_start" && event.message.role === "user") {
-            this._overflowRecoveryAttempted = false;
+            this._overflowRecoveryAttempts = 0;
             const messageText = this._getUserMessageText(event.message);
             if (messageText) {
                 // Check steering queue first
@@ -1017,7 +1025,7 @@ export class AgentSession {
                 this._lastAssistantMessage = finalizedEvent.message;
                 const assistantMsg = finalizedEvent.message;
                 if (assistantMsg.stopReason !== "error") {
-                    this._overflowRecoveryAttempted = false;
+                    this._overflowRecoveryAttempts = 0;
                 }
                 // Reset retry counter immediately on successful assistant response
                 // This prevents accumulation across multiple LLM calls within a turn
@@ -2977,18 +2985,18 @@ export class AgentSession {
         }
         // Case 1: Overflow - LLM returned context overflow error
         if ((sameModel || visionRouteOverflow) && isContextOverflow(assistantMessage, contextWindow)) {
-            if (this._overflowRecoveryAttempted) {
+            if (this._overflowRecoveryAttempts >= MAX_OVERFLOW_RECOVERY_ATTEMPTS) {
                 this._emit({
                     type: "compaction_end",
                     reason: "overflow",
                     result: undefined,
                     aborted: false,
                     willRetry: false,
-                    errorMessage: "Context overflow recovery failed after one compact-and-retry attempt. Try reducing context or switching to a larger-context model.",
+                    errorMessage: "Context overflow recovery failed after two staged compact-and-retry attempts. Reduce the latest input or switch to a model with a larger effective context window.",
                 });
                 return false;
             }
-            this._overflowRecoveryAttempted = true;
+            this._overflowRecoveryAttempts++;
             // Remove the error message from agent state (it IS saved to session for history,
             // but we don't want it in context for the retry)
             const messages = this.agent.state.messages;
@@ -3026,11 +3034,24 @@ export class AgentSession {
         }
         return false;
     }
+    _overflowCompactionSettings(settings, attempt) {
+        if (attempt < MAX_OVERFLOW_RECOVERY_ATTEMPTS)
+            return settings;
+        return {
+            ...settings,
+            reserveTokens: Math.min(settings.reserveTokens, OVERFLOW_RECOVERY_EMERGENCY_TOKENS),
+            reservedOutputTokens: Math.min(settings.reservedOutputTokens ?? settings.reserveTokens, OVERFLOW_RECOVERY_EMERGENCY_TOKENS),
+            keepRecentTokens: Math.min(settings.keepRecentTokens, OVERFLOW_RECOVERY_EMERGENCY_TOKENS),
+        };
+    }
     /**
      * Internal: Run auto-compaction with events.
      */
     async _runAutoCompaction(reason, willRetry, emergency = reason === "overflow" || this._thresholdCompactionEmergency) {
-        const settings = this.settingsManager.getCompactionSettings();
+        const configuredSettings = this.settingsManager.getCompactionSettings();
+        const settings = reason === "overflow"
+            ? this._overflowCompactionSettings(configuredSettings, this._overflowRecoveryAttempts)
+            : configuredSettings;
         this._emit({ type: "compaction_start", reason });
         this._autoCompactionAbortController = new AbortController();
         try {
@@ -3666,9 +3687,6 @@ export class AgentSession {
     // =========================================================================
     // Auto-Retry
     // =========================================================================
-    _isNonRetryableProviderLimitError(errorMessage) {
-        return /GoUsageLimitError|FreeUsageLimitError|Monthly usage limit reached|available balance|insufficient_quota|out of budget|quota exceeded|billing/i.test(errorMessage);
-    }
     /**
      * Check if an error is retryable (overloaded, rate limit, server errors).
      * Context overflow errors are NOT retryable (handled by compaction instead).
@@ -3681,8 +3699,11 @@ export class AgentSession {
         if (isContextOverflow(message, contextWindow))
             return false;
         const err = message.errorMessage;
-        if (this._isNonRetryableProviderLimitError(err))
-            return false;
+        // Billing/quota exhaustion: useless to retry the same model until the
+        // cycle resets, but the turn may still be saved by failing over to the
+        // next candidate, so let _prepareRetry run its failover path.
+        if (isQuotaExhaustionMessage(err))
+            return true;
         // Shared contract with provider-resilience.ts (Fable safety stop, K3 orphan tool_call_id, terminated).
         return isTransientProviderErrorMessage(err);
     }
@@ -3752,8 +3773,11 @@ export class AgentSession {
         const resilience = resolveProviderResilience(this.settingsManager.getProviderResilienceSettings());
         if (!resilience.autoFailoverOnSafetyStop)
             return undefined;
-        if (!isContentSafetyStopMessage(message.errorMessage))
+        // Fire on safety-stop FPs AND on billing/quota exhaustion: both mean the
+        // current model cannot finish this turn and same-model retry is useless.
+        if (!isContentSafetyStopMessage(message.errorMessage) && !isQuotaExhaustionMessage(message.errorMessage)) {
             return undefined;
+        }
         const current = this.model;
         // v10.0-Ω: do NOT gate on isStickySafetyModel(source).
         // claude-opus-5 emits the same stop_reason=refusal FP; same-model retry is useless.