@piagent/platform 1.5.5 → 1.6.1

package/packages/piagent-core/benchmark/benchmark-execution-guard.js

@@ -25,7 +25,7 @@ function assetError(stage, asset, cause) {
   return error;
 }
 
-export function createBenchmarkExecutionGuard({ candidateGuard, suiteRoot, suiteIdentity, piAgentHome, codexCredential, runtimeDependencies, commands }) {
+export function createBenchmarkExecutionGuard({ candidateGuard, suiteRoot, suiteIdentity, piAgentHome, codexCredential, runtimeDependencies, commands, verifyCommandAssets = true }) {
   const observeAssets = (stage, runtimeHomes = []) => {
     try {
       const suite = benchmarkTreeIdentity(suiteRoot, { rejectSymlinks: true });
@@ -61,10 +61,12 @@ export function createBenchmarkExecutionGuard({ candidateGuard, suiteRoot, suite
         throw assetError(stage, "codex-credential", error);
       }
     }
-    for (const [label, identity] of Object.entries(commands ?? {})) {
-      if (!identity) continue;
-      try { verifyBenchmarkCommandIdentity(identity, label, { fullPackageClosure: stage === "finalization" || stage === "prepublish" }); }
-      catch (error) { throw assetError(stage, `command:${label}`, error); }
+    if (verifyCommandAssets) {
+      for (const [label, identity] of Object.entries(commands ?? {})) {
+        if (!identity) continue;
+        try { verifyBenchmarkCommandIdentity(identity, label, { fullPackageClosure: stage === "finalization" || stage === "prepublish" }); }
+        catch (error) { throw assetError(stage, `command:${label}`, error); }
+      }
     }
   };
   const receipt = (stage, runtimeHomes = []) => {

package/packages/piagent-core/benchmark/benchmark-forensics.js

@@ -10,6 +10,7 @@ import {
 } from "./benchmark-candidate.js";
 import { assertBenchmarkLedgerBinding, inspectBenchmarkLedger } from "./benchmark-ledger.js";
 import { benchmarkPhaseAttribution } from "./benchmark-phase-attribution.js";
+import { causalRuntimeEvidence } from "./benchmark-runtime-causal.js";
 
 export { materializeBenchmarkCandidate } from "./benchmark-candidate.js";
 
@@ -329,6 +330,438 @@ export function benchmarkOperationalEvidence(events) {
   };
 }
 
+const causalContextReadTools = new Set(["read", "grep", "find", "ls"]);
+const causalContextShellTools = new Set(["bash", "shell", "exec"]);
+const causalContextCoverageLanesV1 = Object.freeze([
+  "telemetry-window",
+  "session-lifecycle",
+  "criterion-initial-pack",
+  "pack-lifecycle",
+  "direct-fallback-rereads",
+  "managed-prefix"
+]);
+const causalContextCoverageLanes = Object.freeze([
+  ...causalContextCoverageLanesV1,
+  "edit-recovery-context"
+]);
+const causalContextCriterionReasons = new Set([
+  "selected",
+  "auto-context-disabled",
+  "criterion-graph-unavailable",
+  "no-candidates",
+  "no-readable-selection"
+]);
+const causalContextMaximumAggregate = 1_000_000_000;
+
+function boundedCausalInteger(value) {
+  return Number.isSafeInteger(value) && value >= 0 && value <= causalContextMaximumAggregate;
+}
+
+function boundedCausalSum(values) {
+  let total = 0;
+  for (const value of values) {
+    if (!boundedCausalInteger(value) || total + value > causalContextMaximumAggregate) return null;
+    total += value;
+  }
+  return total;
+}
+
+function normalizeCausalPath(value) {
+  if (typeof value !== "string" || !value || value.length > 4_096) return null;
+  const normalized = path.posix.normalize(value.replaceAll("\\", "/").replace(/^\.\/+/, "").replace(/\/+/g, "/"));
+  if (!normalized || normalized === "." || path.posix.isAbsolute(normalized)
+    || normalized === ".." || normalized.startsWith("../")) return null;
+  return normalized;
+}
+
+function causalContextPackEvent(event, idField) {
+  const id = event?.[idField];
+  if (typeof id !== "string" || !id || id.length > 200 || !Array.isArray(event.selectedItems)
+    || event.selectedItems.length > 10_000 || !boundedCausalInteger(event.estimatedTokens)
+    || typeof event.source !== "string" || !event.source || event.source.length > 100) return null;
+  const paths = event.selectedItems.map((item) => normalizeCausalPath(item?.path));
+  if (paths.some((item) => item === null)) return null;
+  const selectedItemEstimatedTokens = boundedCausalSum(event.selectedItems.map((item) => item?.estimatedTokens));
+  if (selectedItemEstimatedTokens === null) return null;
+  let binding;
+  try {
+    binding = JSON.stringify([event.source, event.queryHash ?? null, event.confidence ?? null,
+      event.estimatedTokens, event.selectedItems]);
+  } catch {
+    return null;
+  }
+  return {
+    id,
+    estimatedTokens: event.estimatedTokens,
+    selectedItems: event.selectedItems.length,
+    selectedItemEstimatedTokens,
+    criterion: event.source === "criterion-pack",
+    paths,
+    binding
+  };
+}
+
+function causalContextTask(event, taskByTurn) {
+  if (typeof event?.taskRunId === "string" && event.taskRunId) return event.taskRunId;
+  if (typeof event?.turnId !== "string" || !event.turnId) return "";
+  return taskByTurn.get(event.turnId) ?? "";
+}
+
+function causalEditRecoveryEvent(event) {
+  const toolCallId = event?.toolCallId;
+  const targetPath = normalizeCausalPath(event.targetPath);
+  if (typeof toolCallId !== "string" || !toolCallId || toolCallId.length > 200
+    || targetPath === null
+    || !/^[a-f0-9]{64}$/.test(String(event.contentHash ?? ""))
+    || !boundedCausalInteger(event.originalChars)
+    || !boundedCausalInteger(event.injectedChars) || event.injectedChars === 0
+    || event.injectedChars <= event.originalChars
+    || !boundedCausalInteger(event.injectedEstimatedTokens) || event.injectedEstimatedTokens === 0
+    || event.sensitiveContentRedacted !== false) return null;
+  return {
+    toolCallId,
+    targetPath,
+    taskRunId: typeof event.taskRunId === "string" ? event.taskRunId : "",
+    injectedChars: event.injectedChars,
+    injectedEstimatedTokens: event.injectedEstimatedTokens
+  };
+}
+
+/**
+ * Persistable causal context evidence for one benchmark session. The projection
+ * deliberately drops every path, prompt, hash, identifier and timestamp. Any
+ * incomplete lane nulls the aggregate payload so missing telemetry cannot be
+ * interpreted as zero context work.
+ */
+export function benchmarkCausalContextReceipt(events, {
+  surface,
+  sessionId,
+  criterionExpected = true,
+  telemetryTruncated = false,
+  telemetryExists = true,
+  telemetryIntegrityFailures = 0,
+  recoverableTailBytes = 0
+} = {}) {
+  if (surface !== "piagent") return {
+    schemaVersion: 1,
+    evidenceSource: "not-applicable",
+    applicability: "not-applicable",
+    available: false,
+    coverage: {
+      status: "not-applicable",
+      telemetryTruncated: false,
+      telemetryIntegrityFailures: 0,
+      recoverableTailBytes: 0,
+      criterionExpected: false,
+      sessionEventsObserved: 0,
+      observedLanes: 0,
+      requiredLanes: 0,
+      missingLanes: []
+    },
+    aggregates: null
+  };
+
+  const visible = Array.isArray(events) && typeof sessionId === "string" && sessionId
+    ? events.filter((event) => event && typeof event === "object" && event.sessionId === sessionId)
+    : [];
+  const envelopeFailures = visible.filter((event) => event.schemaVersion !== 2 || event.telemetrySource !== "piagent").length;
+  const suppliedIntegrityFailures = boundedCausalInteger(telemetryIntegrityFailures)
+    ? telemetryIntegrityFailures
+    : causalContextMaximumAggregate;
+  const totalIntegrityFailures = boundedCausalSum([suppliedIntegrityFailures, envelopeFailures])
+    ?? causalContextMaximumAggregate;
+  const promptIndexByTurn = new Map();
+  for (const [index, event] of visible.entries()) {
+    if (event.event !== "agent_prompt" || typeof event.turnId !== "string" || !event.turnId) continue;
+    if (promptIndexByTurn.has(event.turnId)) promptIndexByTurn.set(event.turnId, null);
+    else promptIndexByTurn.set(event.turnId, index);
+  }
+  const offered = new Map(), delivered = new Map(), injected = new Map();
+  const criterion = {
+    attempts: 0,
+    selectedAttempts: 0,
+    candidates: 0,
+    selectedItems: 0,
+    estimatedTokens: 0,
+    zeroSelectionReasonCounts: {
+      autoContextDisabled: 0,
+      criterionGraphUnavailable: 0,
+      noCandidates: 0,
+      noReadableSelection: 0
+    }
+  };
+  let criterionEvidenceComplete = true, packLifecycleComplete = true;
+  for (const [index, event] of visible.entries()) {
+    if (event.event === "criterion_context_pack") {
+      const promptIndex = promptIndexByTurn.get(event.turnId);
+      const valid = boundedCausalInteger(event.selected) && boundedCausalInteger(event.candidates)
+        && boundedCausalInteger(event.estimatedTokens) && causalContextCriterionReasons.has(event.reasonCode)
+        && event.candidates >= event.selected
+        && Number.isInteger(promptIndex) && promptIndex < index
+        && ((event.selected > 0 && event.reasonCode === "selected") || (event.selected === 0 && event.reasonCode !== "selected"));
+      if (!valid) {
+        criterionEvidenceComplete = false;
+        continue;
+      }
+      criterion.attempts += 1;
+      criterion.selectedAttempts += event.selected > 0 ? 1 : 0;
+      criterion.candidates += event.candidates;
+      criterion.selectedItems += event.selected;
+      criterion.estimatedTokens += event.estimatedTokens;
+      if (event.selected === 0) {
+        const key = {
+          "auto-context-disabled": "autoContextDisabled",
+          "criterion-graph-unavailable": "criterionGraphUnavailable",
+          "no-candidates": "noCandidates",
+          "no-readable-selection": "noReadableSelection"
+        }[event.reasonCode];
+        criterion.zeroSelectionReasonCounts[key] += 1;
+      }
+      if (![criterion.attempts, criterion.selectedAttempts, criterion.candidates, criterion.selectedItems, criterion.estimatedTokens]
+        .every(boundedCausalInteger)) criterionEvidenceComplete = false;
+    } else if (event.event === "context_pack_offered") {
+      const value = causalContextPackEvent(event, "deliveryId");
+      if (!value || offered.has(value.id)) packLifecycleComplete = false;
+      else offered.set(value.id, { ...value, index });
+    } else if (event.event === "context_delivery_confirmed") {
+      const id = event.deliveryId;
+      if (typeof id !== "string" || !id || id.length > 200 || !boundedCausalInteger(event.selected) || delivered.has(id)) {
+        packLifecycleComplete = false;
+      } else delivered.set(id, { id, selectedItems: event.selected, index });
+    } else if (event.event === "context_pack_injected") {
+      const value = causalContextPackEvent(event, "injectionId");
+      if (!value || injected.has(value.id)) packLifecycleComplete = false;
+      else injected.set(value.id, { ...value, index });
+    }
+  }
+  for (const [id, delivery] of delivered) {
+    const offer = offered.get(id), injection = injected.get(id);
+    if (!offer || !injection || offer.index >= delivery.index || delivery.index >= injection.index
+      || offer.selectedItems !== delivery.selectedItems || delivery.selectedItems !== injection.selectedItems
+      || offer.estimatedTokens !== injection.estimatedTokens
+      || offer.selectedItemEstimatedTokens !== injection.selectedItemEstimatedTokens
+      || offer.binding !== injection.binding) packLifecycleComplete = false;
+  }
+  for (const id of offered.keys()) if (!delivered.has(id) || !injected.has(id)) packLifecycleComplete = false;
+  for (const id of delivered.keys()) if (!offered.has(id) || !injected.has(id)) packLifecycleComplete = false;
+  for (const id of injected.keys()) if (!delivered.has(id)) packLifecycleComplete = false;
+  const criterionOffered = [...offered.values()].filter((item) => item.criterion);
+  const criterionDelivered = [...delivered.keys()].filter((id) => offered.get(id)?.criterion);
+  const criterionInjected = [...injected.values()].filter((item) => item.criterion);
+  if (criterionExpected === true && criterion.attempts === 0) criterionEvidenceComplete = false;
+  if (criterion.selectedAttempts !== criterionOffered.length
+    || criterion.selectedItems !== boundedCausalSum(criterionOffered.map((item) => item.selectedItems))
+    || criterion.estimatedTokens !== boundedCausalSum(criterionOffered.map((item) => item.estimatedTokens))
+    || criterionOffered.length !== criterionDelivered.length
+    || criterionDelivered.length !== criterionInjected.length) criterionEvidenceComplete = false;
+
+  const taskByTurn = new Map(visible
+    .filter((event) => event.event === "turn_task_bound" && typeof event.turnId === "string"
+      && event.turnId && typeof event.taskRunId === "string" && event.taskRunId)
+    .map((event) => [event.turnId, event.taskRunId]));
+  const activePackByTask = new Map();
+  const pendingDirectReads = new Map(), seenDirectReadIds = new Set();
+  let directFallbackRereadCalls = 0, shellToolCallsObserved = 0;
+  let fallbackEvidenceComplete = packLifecycleComplete;
+  for (const event of visible) {
+    if (event.event === "context_pack_injected") {
+      const taskRunId = causalContextTask(event, taskByTurn);
+      const paths = injected.get(event.injectionId)?.paths;
+      if (!taskRunId || !paths) {
+        fallbackEvidenceComplete = false;
+        continue;
+      }
+      const selected = activePackByTask.get(taskRunId) ?? new Set();
+      for (const selectedPath of paths) selected.add(selectedPath);
+      activePackByTask.set(taskRunId, selected);
+      continue;
+    }
+    if (event.event === "tool_call") {
+      const toolName = String(event.toolName ?? "").toLowerCase();
+      const taskRunId = causalContextTask(event, taskByTurn);
+      const selected = activePackByTask.get(taskRunId);
+      if (selected && causalContextShellTools.has(toolName)) shellToolCallsObserved += 1;
+      if (!selected || !causalContextReadTools.has(toolName)) continue;
+      const target = normalizeCausalPath(event.targetPath);
+      if (!target || typeof event.toolCallId !== "string" || !event.toolCallId || event.toolCallId.length > 200) {
+        fallbackEvidenceComplete = false;
+        continue;
+      }
+      if (!selected.has(target)) continue;
+      if (seenDirectReadIds.has(event.toolCallId)) fallbackEvidenceComplete = false;
+      else {
+        seenDirectReadIds.add(event.toolCallId);
+        pendingDirectReads.set(event.toolCallId, target);
+      }
+      continue;
+    }
+    const toolCallId = typeof event.toolCallId === "string" ? event.toolCallId : "";
+    if (!toolCallId || !pendingDirectReads.has(toolCallId)) continue;
+    if (event.event === "tool_decision" && event.decision === "blocked") {
+      pendingDirectReads.delete(toolCallId);
+      continue;
+    }
+    if (event.event === "tool_result") {
+      if (event.isError !== true) directFallbackRereadCalls += 1;
+      pendingDirectReads.delete(toolCallId);
+    }
+  }
+  if (pendingDirectReads.size > 0) fallbackEvidenceComplete = false;
+
+  const recoverySessionStarts = visible.filter((event) => event.event === "session_start");
+  const firstRecoverySessionStart = visible.findIndex((event) => event.event === "session_start");
+  const recoveryCapabilityComplete = recoverySessionStarts.length > 0
+    && recoverySessionStarts.every((event) => event.editRecoveryContextTelemetryVersion === 1);
+  const editRecoveryReceipts = new Map(), editRecoveryResults = new Map();
+  let editRecoveryEvidenceComplete = recoveryCapabilityComplete;
+  for (const [index, event] of visible.entries()) {
+    if (event.event === "edit_recovery_context") {
+      const value = causalEditRecoveryEvent(event);
+      if (!value || index <= firstRecoverySessionStart || editRecoveryReceipts.has(value.toolCallId)) editRecoveryEvidenceComplete = false;
+      else editRecoveryReceipts.set(value.toolCallId, { ...value, index });
+      continue;
+    }
+    if (event.event !== "tool_result") continue;
+    const recoveryFailure = ["edit-anchor-not-unique", "edit-anchor-stale"].includes(event.reasonCode);
+    if (!recoveryFailure) {
+      if (event.editRecoveryContext !== undefined) editRecoveryEvidenceComplete = false;
+      continue;
+    }
+    const toolCallId = event.toolCallId;
+    if (typeof toolCallId !== "string" || !toolCallId || toolCallId.length > 200
+      || editRecoveryResults.has(toolCallId)
+      || index <= firstRecoverySessionStart
+      || String(event.toolName ?? "").toLowerCase() !== "edit"
+      || event.isError !== true
+      || typeof event.editRecoveryContext !== "boolean"
+      || normalizeCausalPath(event.targetPath) === null
+      || (event.editRecoveryContext === true
+        && (!boundedCausalInteger(event.editRecoveryInjectedChars) || event.editRecoveryInjectedChars === 0
+          || !boundedCausalInteger(event.editRecoveryEstimatedTokens) || event.editRecoveryEstimatedTokens === 0))
+      || (event.editRecoveryContext === false
+        && (event.editRecoveryInjectedChars !== undefined || event.editRecoveryEstimatedTokens !== undefined))) {
+      editRecoveryEvidenceComplete = false;
+      continue;
+    }
+    editRecoveryResults.set(toolCallId, {
+      index,
+      injected: event.editRecoveryContext,
+      targetPath: normalizeCausalPath(event.targetPath),
+      taskRunId: typeof event.taskRunId === "string" ? event.taskRunId : "",
+      injectedChars: event.editRecoveryInjectedChars,
+      injectedEstimatedTokens: event.editRecoveryEstimatedTokens
+    });
+  }
+  for (const [toolCallId, receipt] of editRecoveryReceipts) {
+    const result = editRecoveryResults.get(toolCallId);
+    if (!result || result.injected !== true || receipt.index >= result.index
+      || receipt.targetPath !== result.targetPath
+      || receipt.taskRunId !== result.taskRunId
+      || receipt.injectedChars !== result.injectedChars
+      || receipt.injectedEstimatedTokens !== result.injectedEstimatedTokens) editRecoveryEvidenceComplete = false;
+  }
+  for (const [toolCallId, result] of editRecoveryResults) {
+    if (result.injected !== editRecoveryReceipts.has(toolCallId)) editRecoveryEvidenceComplete = false;
+  }
+  const editRecoveryInjectedChars = boundedCausalSum([...editRecoveryReceipts.values()].map((item) => item.injectedChars));
+  const editRecoveryEstimatedTokens = boundedCausalSum([...editRecoveryReceipts.values()].map((item) => item.injectedEstimatedTokens));
+  editRecoveryEvidenceComplete &&= editRecoveryInjectedChars !== null && editRecoveryEstimatedTokens !== null;
+
+  const prompts = visible.map((event, index) => ({ event, index }))
+    .filter((item) => item.event.event === "agent_prompt");
+  const managedPrefixComplete = prompts.length > 0
+    && prompts.every((item) => typeof item.event.managedInstructionsCompacted === "boolean");
+  const compactedPrompts = prompts.filter((item) => item.event.managedInstructionsCompacted === true).length;
+  const lastPromptIndex = prompts.at(-1)?.index ?? -1;
+  const lifecycleComplete = prompts.length > 0
+    && visible.some((event, index) => index > lastPromptIndex
+      && (event.event === "agent_settled" || event.event === "session_shutdown"));
+  const offeredTokens = boundedCausalSum([...offered.values()].map((item) => item.estimatedTokens));
+  const deliveredTokens = boundedCausalSum([...delivered.keys()].map((id) => offered.get(id)?.estimatedTokens));
+  const injectedTokens = boundedCausalSum([...injected.values()].map((item) => item.estimatedTokens));
+  const offeredItemTokens = boundedCausalSum([...offered.values()].map((item) => item.selectedItemEstimatedTokens));
+  const deliveredItemTokens = boundedCausalSum([...delivered.keys()].map((id) => offered.get(id)?.selectedItemEstimatedTokens));
+  const injectedItemTokens = boundedCausalSum([...injected.values()].map((item) => item.selectedItemEstimatedTokens));
+  const offeredItems = boundedCausalSum([...offered.values()].map((item) => item.selectedItems));
+  const deliveredItems = boundedCausalSum([...delivered.values()].map((item) => item.selectedItems));
+  const injectedItems = boundedCausalSum([...injected.values()].map((item) => item.selectedItems));
+  const aggregateBoundsComplete = [offeredTokens, deliveredTokens, injectedTokens, offeredItemTokens, deliveredItemTokens,
+    injectedItemTokens, offeredItems, deliveredItems, injectedItems]
+    .every((value) => value !== null) && boundedCausalInteger(directFallbackRereadCalls)
+    && boundedCausalInteger(shellToolCallsObserved);
+  packLifecycleComplete &&= aggregateBoundsComplete;
+  fallbackEvidenceComplete &&= aggregateBoundsComplete;
+
+  const lanes = {
+    "telemetry-window": telemetryExists === true && telemetryTruncated !== true
+      && totalIntegrityFailures === 0 && recoverableTailBytes === 0,
+    "session-lifecycle": lifecycleComplete,
+    "criterion-initial-pack": criterionEvidenceComplete,
+    "pack-lifecycle": packLifecycleComplete,
+    "direct-fallback-rereads": fallbackEvidenceComplete,
+    "managed-prefix": managedPrefixComplete,
+    "edit-recovery-context": editRecoveryEvidenceComplete
+  };
+  const missingLanes = causalContextCoverageLanes.filter((lane) => lanes[lane] !== true);
+  const observedLanes = causalContextCoverageLanes.length - missingLanes.length;
+  const available = missingLanes.length === 0;
+  const compactions = visible.filter((event) => event.event === "session_compact").length;
+  const managedPrefixState = compactedPrompts === 0
+    ? "uncompacted"
+    : compactedPrompts === prompts.length ? "compacted" : "mixed";
+  return {
+    schemaVersion: 3,
+    evidenceSource: "context-telemetry-closed-aggregate-v3",
+    applicability: "piagent",
+    available,
+    coverage: {
+      status: available ? "complete" : visible.length === 0 || !lifecycleComplete ? "unavailable" : "partial",
+      telemetryTruncated: telemetryTruncated === true,
+      telemetryIntegrityFailures: totalIntegrityFailures,
+      recoverableTailBytes: boundedCausalInteger(recoverableTailBytes) ? recoverableTailBytes : causalContextMaximumAggregate,
+      criterionExpected: criterionExpected === true,
+      sessionEventsObserved: visible.length,
+      observedLanes,
+      requiredLanes: causalContextCoverageLanes.length,
+      missingLanes
+    },
+    aggregates: available ? {
+      packCounts: { offered: offered.size, delivered: delivered.size, injected: injected.size },
+      estimatedTokens: { offered: offeredTokens, delivered: deliveredTokens, injected: injectedTokens },
+      selectedItemEstimatedTokens: { offered: offeredItemTokens, delivered: deliveredItemTokens, injected: injectedItemTokens },
+      selectedItemCounts: { offered: offeredItems, delivered: deliveredItems, injected: injectedItems },
+      criterionInitialPack: {
+        ...criterion,
+        offered: criterionOffered.length,
+        delivered: criterionDelivered.length,
+        injected: criterionInjected.length
+      },
+      directFallbackRereads: {
+        successfulCalls: directFallbackRereadCalls,
+        shellToolCallsObserved,
+        definition: "successful-direct-path-tool-call-v1"
+      },
+      editRecoveryContext: {
+        count: editRecoveryReceipts.size,
+        failuresObserved: editRecoveryResults.size,
+        suppressedFailures: [...editRecoveryResults.values()].filter((item) => item.injected === false).length,
+        injectedChars: editRecoveryInjectedChars,
+        injectedEstimatedTokens: editRecoveryEstimatedTokens,
+        evidenceCoverage: {
+          status: "complete",
+          observed: editRecoveryReceipts.size,
+          comparable: editRecoveryReceipts.size,
+          rate: 1
+        },
+        definition: "matched-edit-recovery-context-receipt-v1"
+      },
+      runtimeCausal: causalRuntimeEvidence(visible),
+      compaction: { eventsObserved: compactions, state: compactions > 0 ? "observed" : "not-observed" },
+      managedPrefix: { promptsObserved: prompts.length, compactedPrompts, state: managedPrefixState }
+    } : null
+  };
+}
+
 export function loadReplayFailurePlan(reportPath) {
   let report;
   try {
@@ -440,12 +873,13 @@ export function classifyPreUsageFailure(agent, usage, diagnosticInput, { termina
   const measuredUsage = Number.isInteger(usage?.sessions) && usage.sessions > 0
     && ["input", "output", "cacheRead", "cacheWrite", "reasoning", "total", "fresh"]
       .every((field) => Number.isFinite(usage?.[field]) && Number(usage[field]) >= 0)
-    && Number.isFinite(usage?.cost) && Number(usage.cost) >= 0;
+    && usage.fresh === usage.input + usage.output
+    && usage.total === usage.input + usage.output + usage.cacheRead + usage.cacheWrite;
   const measuredZeroUsage = measuredUsage
-    && ["input", "output", "cacheRead", "cacheWrite", "reasoning", "total", "fresh", "cost"]
+    && ["input", "output", "cacheRead", "cacheWrite", "reasoning", "total", "fresh"]
       .every((field) => Number(usage[field]) === 0);
   const providerUnavailable = /\b(?:server(?:s)? (?:are )?(?:currently )?overloaded|temporarily unavailable|service unavailable|try again later)\b/.test(diagnostic);
-  if (agent.code === 0 && terminalProviderError && measuredUsage && providerUnavailable) {
+  if ((terminalProviderError || agent.code !== 0) && measuredUsage && providerUnavailable) {
     const afterUsage = Number(usage.fresh) > 0;
     return {
       failure: afterUsage
@@ -467,7 +901,10 @@ export function classifyPreUsageFailure(agent, usage, diagnosticInput, { termina
   }
   if (agent.code === 0) return undefined;
   if (/\b(?:provider|safety|policy|refus(?:al|ed|e)|disallowed|not allowed|cannot assist|can't assist|cyber safety)\b/.test(diagnostic)) {
-    return { failure: "provider-policy-refusal-with-usage-unavailable", class: "provider-policy", usageStatus: "unknown-after-provider-start", retryable: false };
+    return measuredUsage
+      ? { failure: "provider-policy-refusal-after-measured-usage", class: "provider-policy", usageStatus: "measured-but-unaccepted", retryable: false }
+      : { failure: "provider-policy-refusal-with-usage-unavailable", class: "provider-policy", usageStatus: "unknown-after-provider-start", retryable: false };
   }
+  if (measuredUsage) return { failure: `agent-exit-${agent.code}-after-measured-usage`, class: "agent-process", usageStatus: "measured-but-unaccepted", retryable: false };
   return { failure: `agent-exit-${agent.code}-with-usage-unavailable`, class: "unknown-cost", usageStatus: "unknown-after-provider-start", retryable: true };
 }