@opengeni/worker-bundle 1.1.0 → 1.2.0

package/dist/{activities-turn-4EEVNZCC.js → activities-turn-BMY3CDKL.js}

@@ -99,7 +99,7 @@ import {
   withFirstPartyTools,
   wrapLazyTurnBoxWithRouting,
   wrapTurnBoxWithRouting
-} from "./chunk-OT64XARC.js";
+} from "./chunk-JKMG5ZHH.js";
 import {
   StreamTimingMetrics,
   makeMachineOpObserver,
@@ -132,7 +132,7 @@ import {
   runtimeMetricsHooksForObservability,
   sandboxAutomaticRecoveryOutcome,
   turnLifecycleMetricsFor
-} from "./chunk-3ZTRWDMN.js";
+} from "./chunk-XONQEVAD.js";
 import {
   __export
 } from "./chunk-PZ5AY32C.js";
@@ -163,7 +163,7 @@ import {
 } from "@opengeni/db";
 import {
   sandboxOperationMetricObserver as sandboxOperationMetricObserver2,
-  turnExecutionTelemetryKey as turnExecutionTelemetryKey2,
+  turnExecutionTelemetryKey as turnExecutionTelemetryKey3,
   withTraceContext,
   withMcpTelemetry
 } from "@opengeni/observability";
@@ -3670,6 +3670,7 @@ import { fetchXaiSubscriptionQuota } from "@opengeni/xai-subscription";
 import { environmentsEncryptionKeyBytes } from "@opengeni/config";
 
 // src/activities/agent-turn/finalization-monitor.ts
+import { turnExecutionTelemetryKey } from "@opengeni/observability";
 var TURN_FINALIZATION_STAGES = [
   "tool_writers",
   "credential_renewals",
@@ -3701,6 +3702,16 @@ function diagnose(operation) {
 }
 function startTurnFinalizationMonitor(input) {
   const { observability, details } = input;
+  const correlation = {};
+  diagnose(() => {
+    if (input.execution) {
+      correlation.correlationId = turnExecutionTelemetryKey(
+        input.execution.workspaceId,
+        input.execution.sessionId,
+        input.execution.attemptId
+      );
+    }
+  });
   if (!initialized.has(observability)) {
     for (const stage2 of TURN_FINALIZATION_STAGES) {
       diagnose(() => observability.incrementGauge({ ...INFLIGHT, labels: { stage: stage2 }, amount: 0 }));
@@ -3743,7 +3754,8 @@ function startTurnFinalizationMonitor(input) {
             observability.error("turn finalization stalled; containing worker", {
               surface: "turn_finalization",
               outcome: "containment",
-              reason: next
+              reason: next,
+              ...correlation
             });
           } catch {
           }
@@ -3756,7 +3768,8 @@ function startTurnFinalizationMonitor(input) {
           observability.warn("turn finalization is taking longer than expected", {
             surface: "turn_finalization",
             outcome: "slow",
-            reason: next
+            reason: next,
+            ...correlation
           });
         }),
         input.slowAfterMs ?? 3e4
@@ -4618,7 +4631,12 @@ async function finalizeTurnAttempt(deps) {
       } catch {
       }
     },
-    requestWorkerDrain: deps.requestWorkerDrain
+    requestWorkerDrain: deps.requestWorkerDrain,
+    execution: {
+      workspaceId: deps.input.workspaceId,
+      sessionId: deps.input.sessionId,
+      attemptId: deps.input.attemptId
+    }
   });
   try {
     monitor.enter("tool_writers");
@@ -5259,7 +5277,7 @@ import {
 } from "@opengeni/codex";
 
 // src/activities/agent-turn/admission.ts
-import { checkWorkspaceAllowance, getBillingBalance, sumUsageQuantity } from "@opengeni/db";
+import { checkWorkspaceAllowance, getSpendableCreditBalance, sumUsageQuantity } from "@opengeni/db";
 import {
   configuredStaticUsageLimits
 } from "@opengeni/config";
@@ -5268,6 +5286,7 @@ import { directPersonalConnectionSubjectId } from "@opengeni/core";
 import { createHash as createHash7 } from "crypto";
 
 // src/activities/agent-turn/model-usage.ts
+import { canonicalizeConfiguredModelId } from "@opengeni/config";
 import {
   applyCreditDebitUpToBalance,
   recordUsageEvent,
@@ -5297,6 +5316,12 @@ function modelUsageSourceKey(input) {
   }
   return input.dispatchId ? `${input.dispatchId}:${input.positionalKey}` : input.positionalKey;
 }
+function aggregateCreditPolicyRevision(input) {
+  if (input.chargesOpenGeniCredits && input.responseRevisions.size > 1 && (input.totalTokens ?? 0) > 0) {
+    throw new Error("Aggregate model usage spans different credit policy revisions");
+  }
+  return input.responseRevisions.size === 1 ? input.responseRevisions.values().next().value : input.lastAdmittedRevision;
+}
 async function recordCompletedModelCallBeforeOwnershipFences(input) {
   await input.renewLease();
   await input.recordUsage();
@@ -5346,6 +5371,9 @@ async function processModelResponseTerminalEvent(input) {
   if (!terminal) {
     return { status: "not_response" };
   }
+  if (!terminal.usage && input.event.type === "raw_model_stream_event" && input.event.data.type !== "response_done") {
+    return { status: "not_response" };
+  }
   const responseOrdinal = input.state.responseCount + 1;
   const sourceKey = modelUsageSourceKey({
     responseId: terminal.responseId,
@@ -5378,6 +5406,7 @@ async function processModelResponseTerminalEvent(input) {
         turnId: input.turnId,
         turnAttemptId: input.turnAttemptId,
         model: input.model,
+        creditPolicyRevision: input.creditPolicyRevision,
         externallyBilled: input.externallyBilled,
         ...input.chargesOpenGeniCredits !== void 0 ? { chargesOpenGeniCredits: input.chargesOpenGeniCredits } : {},
         ...input.countsTowardTokenCap !== void 0 ? { countsTowardTokenCap: input.countsTowardTokenCap } : {},
@@ -5478,6 +5507,7 @@ async function processCompactionModelUsageEvent(input) {
         turnId: input.turnId,
         turnAttemptId: input.turnAttemptId,
         model: input.model,
+        creditPolicyRevision: input.creditPolicyRevision,
         externallyBilled: input.externallyBilled,
         ...input.chargesOpenGeniCredits !== void 0 ? { chargesOpenGeniCredits: input.chargesOpenGeniCredits } : {},
         ...input.countsTowardTokenCap !== void 0 ? { countsTowardTokenCap: input.countsTowardTokenCap } : {},
@@ -5745,6 +5775,8 @@ async function recordModelUsageAndDebitCredits(settings, db, input) {
       workspaceId: input.workspaceId,
       type: "model_usage_debit",
       requestedAmountMicros: costMicros,
+      modelId: canonicalizeConfiguredModelId(settings, input.model),
+      creditPolicyRevision: input.creditPolicyRevision,
       sourceType: "model_response",
       sourceId: `${input.turnId}:${input.sourceKey}`,
       idempotencyKey: `credit:model_usage_debit:${input.turnId}:${input.sourceKey}`,
@@ -5946,7 +5978,7 @@ var AllowanceExhaustedError = class extends Error {
 };
 async function ensureRunAllowedBetweenModelCalls(input) {
   try {
-    await ensureRunAllowed(
+    return await ensureRunAllowed(
       input.settings,
       input.db,
       input.accountId,
@@ -5955,7 +5987,8 @@ async function ensureRunAllowedBetweenModelCalls(input) {
       input.entitlements,
       input.chargesOpenGeniCredits,
       input.countsTowardTokenCap,
-      input.initiatingHumanSubjectId
+      input.initiatingHumanSubjectId,
+      input.modelId
     );
   } catch (limitError) {
     let serializedRunState = null;
@@ -5970,7 +6003,8 @@ async function ensureRunAllowedBetweenModelCalls(input) {
     );
   }
 }
-async function ensureRunAllowed(settings, db, accountId, workspaceId, isExternallyBilledTurn, entitlements, chargesOpenGeniCredits = !isExternallyBilledTurn, countsTowardTokenCap = !isExternallyBilledTurn, initiatingHumanSubjectId = null) {
+async function ensureRunAllowed(settings, db, accountId, workspaceId, isExternallyBilledTurn, entitlements, chargesOpenGeniCredits = !isExternallyBilledTurn, countsTowardTokenCap = !isExternallyBilledTurn, initiatingHumanSubjectId = null, modelId) {
+  let creditPolicyRevision;
   if (chargesOpenGeniCredits && entitlements && (settings.billingMode === "stripe" || settings.usageLimitsMode === "managed")) {
     const decision = await entitlements.admitRun({
       accountId,
@@ -5982,7 +6016,8 @@ async function ensureRunAllowed(settings, db, accountId, workspaceId, isExternal
       throw new Error(decision.reason || "insufficient Opengeni credits");
     }
   } else if (chargesOpenGeniCredits && (settings.billingMode === "stripe" || settings.usageLimitsMode === "managed")) {
-    const balance = await getBillingBalance(db, accountId);
+    const balance = await getSpendableCreditBalance(db, accountId, modelId);
+    creditPolicyRevision = balance.creditPolicyRevision;
     if (balance.balanceMicros <= 0) {
       throw new Error("insufficient Opengeni credits");
     }
@@ -6020,6 +6055,7 @@ async function ensureRunAllowed(settings, db, accountId, workspaceId, isExternal
       }
     }
   }
+  return creditPolicyRevision;
 }
 
 // src/activities/agent-turn/provider-recovery-metrics.ts
@@ -6132,9 +6168,13 @@ async function settleTurnFailure(deps) {
   try {
     return await settleTurnFailureInAttempt(deps);
   } catch (error51) {
+    if (deps.control.activityStatus === "recovering" && error51 instanceof ApplicationFailure && error51.type === "OpenGeniPostClaimDatabaseRecovery")
+      throw error51;
     if (deps.attempt.turnId && deps.attempt.triggerEventId) {
       const recovery = postClaimDatabaseRecoveryFailure({
-        error: error51,
+        // A failed rollback/terminal write must not erase no-replay evidence
+        // from the failure we were settling (notably unknown tool effects).
+        error: new AggregateError([error51, deps.error], "Turn failure settlement failed"),
         turnId: deps.attempt.turnId,
         triggerEventId: deps.attempt.triggerEventId,
         executionGeneration: deps.attempt.executionGeneration,
@@ -8895,6 +8935,8 @@ async function runTurnStreamAttempt(deps) {
   );
   let parallelSessionTitle = null;
   let parallelSessionTitleFinished = false;
+  let creditPolicyRevision;
+  let titleCreditPolicyRevision;
   const finishParallelSessionTitle = async () => {
     if (parallelSessionTitleFinished) return;
     parallelSessionTitleFinished = true;
@@ -8903,6 +8945,7 @@ async function runTurnStreamAttempt(deps) {
     if (generated.usage) {
       await processSessionTitleModelUsageEvent({
         usage: generated.usage,
+        creditPolicyRevision: titleCreditPolicyRevision,
         state: sessionTitleUsageState,
         dispatchId: modelUsageDispatchId,
         settings,
@@ -9069,7 +9112,8 @@ async function runTurnStreamAttempt(deps) {
   let finalReplyNudged = false;
   const revalidateModelCallAdmission = async () => {
     await historySink.reconcileConversationTruth({ requireDurable: true });
-    await ensureRunAllowedBetweenModelCalls({
+    creditPolicyRevision = await ensureRunAllowedBetweenModelCalls({
+      modelId: resolvedModel?.configured.id ?? turn.model,
       settings,
       db,
       accountId: input.accountId,
@@ -9091,6 +9135,7 @@ async function runTurnStreamAttempt(deps) {
       signal: runtimeCancellationSignal,
       admit: revalidateModelCallAdmission
     });
+    const responseCreditPolicyRevisions = /* @__PURE__ */ new Set();
     const responseCountBeforeStream = modelResponseState.responseCount;
     eventing.batcher = null;
     let currentToolBatchCallIds = /* @__PURE__ */ new Set();
@@ -9260,7 +9305,10 @@ async function runTurnStreamAttempt(deps) {
         attempt.modelRequestStarted = true;
         return await runtime.runStream(agent, runInput, eventing.modelRunSettings, {
           beforeModelRequest: modelCallAdmission.beforeModelRequest,
-          onModelResponse: modelCallAdmission.onModelResponse,
+          onModelResponse: (event) => {
+            responseCreditPolicyRevisions.add(creditPolicyRevision);
+            return modelCallAdmission.onModelResponse(event);
+          },
           signal: runtimeCancellationSignal,
           sandboxEnvironment,
           onModelVisibleContext: async (snapshot) => {
@@ -9450,6 +9498,7 @@ async function runTurnStreamAttempt(deps) {
         }
         const generatedImageReceipt = generatedImage ? await media.retainNativeGeneratedImage(generatedImage) : null;
         const responseResult = await processModelResponseTerminalEvent({
+          creditPolicyRevision,
           event: next.value,
           state: modelResponseState,
           dispatchId: modelUsageDispatchId,
@@ -9533,18 +9582,6 @@ async function runTurnStreamAttempt(deps) {
           await historySink.reconcileConversationTruth();
           turnLifecycleMetricsFor(observability).progress({ attemptId: input.attemptId });
           modelCheckpointMemoryCollector.schedule(observability);
-          await ensureRunAllowedBetweenModelCalls({
-            settings,
-            db,
-            accountId: input.accountId,
-            workspaceId: input.workspaceId,
-            isExternallyBilledTurn: billingState.isExternallyBilledTurn,
-            entitlements,
-            chargesOpenGeniCredits: billingState.chargesOpenGeniCredits,
-            countsTowardTokenCap: billingState.countsTowardTokenCap,
-            initiatingHumanSubjectId: turn.initiatingHumanSubjectId,
-            serializedRunState: () => media.compactMediaRunState(String(eventing.stream.state.toString()))
-          });
         }
         modelCallAdmission.settle(next.value);
         const durableSdkEvent = generatedImageReceipt ? compactGeneratedImageSdkEvent(next.value, generatedImageReceipt) : next.value;
@@ -9838,6 +9875,12 @@ async function runTurnStreamAttempt(deps) {
     if (!streamSawPerResponseUsage) {
       const aggregateUsage = eventing.stream.state.usage;
       const normalizedAggregateUsage = normalizeModelCallUsage2(aggregateUsage);
+      const aggregatePolicyRevision = aggregateCreditPolicyRevision({
+        responseRevisions: responseCreditPolicyRevisions,
+        lastAdmittedRevision: creditPolicyRevision,
+        chargesOpenGeniCredits: billingState.chargesOpenGeniCredits,
+        totalTokens: normalizedAggregateUsage.totalTokens
+      });
       const aggregateInput = normalizedAggregateUsage.telemetry.inputTokens;
       const aggregateSourceKey = modelUsageSourceKey({
         responseId: null,
@@ -9858,6 +9901,7 @@ async function runTurnStreamAttempt(deps) {
           leaseLostMessage: "Provider credential lease expired during the active turn",
           recordUsage: async () => {
             const billing = await recordModelUsageAndDebitCredits(settings, db, {
+              creditPolicyRevision: aggregatePolicyRevision,
               accountId: input.accountId,
               workspaceId: input.workspaceId,
               sessionId: input.sessionId,
@@ -10194,6 +10238,7 @@ async function runTurnStreamAttempt(deps) {
       turnExecutionPolicy.providerId,
       turnExecutionPolicy.latencyMode
     );
+    titleCreditPolicyRevision = creditPolicyRevision;
     parallelSessionTitle = startParallelSessionTitleGeneration({
       signal: runtimeCancellationSignal,
       generate: async (signal) => await withSessionTitleProviderRequestContext(
@@ -10399,7 +10444,7 @@ import {
   requireSession
 } from "@opengeni/db";
 import { appendAndPublishTurnEventsFenced as appendAndPublishTurnEventsFenced2, publishDurableSessionEvents as publishDurableSessionEvents4 } from "@opengeni/events";
-import { linkCurrentSpanToAdmission, turnExecutionTelemetryKey } from "@opengeni/observability";
+import { linkCurrentSpanToAdmission, turnExecutionTelemetryKey as turnExecutionTelemetryKey2 } from "@opengeni/observability";
 import {
   assertTurnExecutionPolicyMatchesConfigV1,
   settingsForAcceptedSubscriptionTurn,
@@ -10932,7 +10977,8 @@ async function claimTurnAttempt(deps) {
         entitlements,
         billingState.chargesOpenGeniCredits,
         billingState.countsTowardTokenCap,
-        turn.initiatingHumanSubjectId
+        turn.initiatingHumanSubjectId,
+        turnExecutionPolicy.productModelId
       ),
       cancellationSignal,
       void 0
@@ -11082,7 +11128,7 @@ async function claimTurnAttempt(deps) {
   throwIfTurnCancelled();
   recordTurnStartupPhase(observability, {
     phase: "claim_and_policy",
-    executionCorrelationId: turnExecutionTelemetryKey(
+    executionCorrelationId: turnExecutionTelemetryKey2(
       input.workspaceId,
       input.sessionId,
       input.attemptId
@@ -12820,9 +12866,11 @@ async function prepareCompaction(deps) {
   };
   const promptCacheKey = acceptsPromptCacheKeyForTurn(resolvedModel) ? input.sessionId : void 0;
   const compactionUsageState = createCompactionModelUsageEventState(claimedModelUsageSourceKeys);
+  let compactionCreditPolicyRevision;
   const recordCompactionUsage = async (usage) => {
     await processCompactionModelUsageEvent({
       usage,
+      creditPolicyRevision: compactionCreditPolicyRevision,
       state: compactionUsageState,
       dispatchId: modelUsageDispatchId,
       settings,
@@ -12850,7 +12898,7 @@ async function prepareCompaction(deps) {
     });
   };
   const compactionSummarizerFor = (systemInstructions) => {
-    const summarize = resolvedModel ? (s, m) => withProviderRequestContext(
+    const summarizeModel = resolvedModel ? (s, m) => withProviderRequestContext(
       () => summarizeContextForCompaction(s, m, {
         client: resolvedModel.client,
         provider: resolvedModel.provider,
@@ -12871,6 +12919,21 @@ async function prepareCompaction(deps) {
       ...systemInstructions ? { systemInstructions } : {},
       ...promptCacheKey ? { promptCacheKey } : {}
     });
+    const summarize = async (s, m) => {
+      compactionCreditPolicyRevision = await ensureRunAllowedBetweenModelCalls({
+        settings: s,
+        db,
+        accountId: input.accountId,
+        workspaceId: input.workspaceId,
+        modelId: resolvedModel?.configured.id ?? turn.model,
+        isExternallyBilledTurn: billingState.isExternallyBilledTurn,
+        chargesOpenGeniCredits: billingState.chargesOpenGeniCredits,
+        countsTowardTokenCap: billingState.countsTowardTokenCap,
+        initiatingHumanSubjectId: turn.initiatingHumanSubjectId,
+        entitlements: deps.entitlements
+      });
+      return await summarizeModel(s, m);
+    };
     summarize.estimatePrefixTokens = () => {
       if (resolvedModel?.provider.api === "chat") {
         return estimateSerializedValueTokens(systemInstructions ?? "");
@@ -38486,6 +38549,7 @@ function createRunAgentTurnActivity(services) {
             )
           );
           const compactionPrep = await prepareCompaction({
+            entitlements,
             input,
             settings: capabilitySettings,
             db,
@@ -39297,7 +39361,7 @@ These are publication receipts for the human's earlier answer. Do not call the r
     const servicesStartedAt = performance.now();
     const resolvedServices = await services();
     const servicesDurationSeconds = (performance.now() - servicesStartedAt) / 1e3;
-    const correlationId = turnExecutionTelemetryKey2(
+    const correlationId = turnExecutionTelemetryKey3(
       input.workspaceId,
       input.sessionId,
       input.attemptId
@@ -40443,4 +40507,4 @@ export {
   createTurnActivities,
   createTurnActivitiesFromServices
 };
-//# sourceMappingURL=activities-turn-4EEVNZCC.js.map
+//# sourceMappingURL=activities-turn-BMY3CDKL.js.map