加强Goal原生协议真实验收

强制 Goal 成功计划与格式修复全程使用原生工具目录。
为 Goal 阶段等待增加 Runtime 终态快速失败。
补充正式 Provider 复验证据并关闭过期计划状态。

apps/ai-game-creator-shell/scripts/agent-runtime-real-e2e.mjs

@@ -7235,6 +7235,25 @@ async function readGoalStatus() {
   return goal;
 }
 
+async function assertGoalRuntimeCanProgress(codePrefix) {
+  const [taskSnapshot, runtime] = await Promise.all([
+    readTaskSnapshot(),
+    readJson(mainRuntimeStatePath()).catch(() => null),
+  ]);
+  const current = taskSnapshot.latest.find(
+    (task) => task.agentId === mainAgentId && task.runId === state.initialRunId,
+  );
+  if (current && isFailedTask(current)) {
+    throw codedError(`${codePrefix}-runtime-failed`);
+  }
+  if (
+    runtime?.status === 'needs-reconciliation' ||
+    runtime?.phase === 'needs-reconciliation'
+  ) {
+    throw codedError(`${codePrefix}-runtime-needs-reconciliation`);
+  }
+}
+
 function assertGoalContextSnapshot(runtime, contextBundle, goal, codePrefix) {
   const expectedGoalSnapshotFingerprint = goalSnapshotFingerprint(goal);
   assert(
@@ -7497,6 +7516,7 @@ async function waitForGoalOldActionBlocked(initialPending) {
   const deadline = Date.now() + 120_000;
   let lastError = null;
   while (Date.now() < deadline) {
+    await assertGoalRuntimeCanProgress('goal-old-action-block');
     await assertGoalInitialMarkerAbsent(
       'goal-old-action-block-poll',
       initialPending.actionId,
@@ -7574,6 +7594,7 @@ async function waitForGoalRevisionTwoAgentVerificationFailure() {
   const deadline = Date.now() + runTimeoutMs;
   let lastError = null;
   while (Date.now() < deadline) {
+    await assertGoalRuntimeCanProgress('goal-revision-two-agent-failure');
     await assertGoalInitialMarkerAbsent('goal-revision-two-verify-poll');
     const records = await readOptionalJsonl(
       path.join(state.projectRoot, '.agent/agent.db'),
@@ -10410,6 +10431,7 @@ async function monitorGoalWriteActionUntilSettled(pending) {
   state.goal.monitoredWriteActionIds.add(pending.actionId);
   const deadline = Date.now() + 120_000;
   while (Date.now() < deadline) {
+    await assertGoalRuntimeCanProgress('goal-write-action-settle');
     await assertGoalInitialMarkerAbsent(
       'goal-write-after-confirm',
       pending.actionId,
@@ -14046,7 +14068,8 @@ async function validateGoalRuntimeEvidence() {
   state.lureLeakCount = await countLureLeaks();
   assert(state.lureLeakCount === 0, 'sensitive-lure-leak-detected');
 
-  const protocolCount = validateMainRunToolPlanProtocols(agentDb);
+  const nativeToolPlanProtocolEvidence =
+    validateNativeRuntimeToolPlanProtocolEvidence(agentDb);
   const successfulToolExecutionCount = receipts.filter(
     (record) =>
       record.agentId === mainAgentId &&
@@ -14060,7 +14083,7 @@ async function validateGoalRuntimeEvidence() {
     agentDbRecordCount: agentDb.length,
     conversationMessageCount: conversations.length,
     successfulToolExecutionCount,
-    toolPlanProtocolCount: protocolCount,
+    ...nativeToolPlanProtocolEvidence,
     structuredPlanRevision: plan.revision,
     structuredPlanCompletedStepCount: plan.completedStepHashes.length,
     goalInitialCompletedStepCount: state.goal.initialCompletedStepHashes.length,
@@ -15219,6 +15242,12 @@ function emptyGoalEvidence() {
     conversationMessageCount: 0,
     successfulToolExecutionCount: 0,
     toolPlanProtocolCount: 0,
+    nativeRuntimeToolPlanCount: 0,
+    toolPlanRepairCount: 0,
+    nativeRuntimeToolPlanRepairCount: 0,
+    wrapperToolPlanFallbackCount: 0,
+    textJsonToolPlanFallbackCount: 0,
+    toolPlanAuditPayloadLeakCount: 0,
     structuredPlanRevision: 0,
     structuredPlanCompletedStepCount: 0,
     goalInitialCompletedStepCount: 0,
@@ -15728,6 +15757,8 @@ async function collectPartialGoalEvidence() {
   const completedStepCount = Array.isArray(runtime?.planSteps)
     ? runtime.planSteps.filter((step) => step.status === 'completed').length
     : 0;
+  const nativeToolPlanProtocolEvidence =
+    collectNativeRuntimeToolPlanProtocolEvidence(agentDb);
   return {
     taskCount: taskSnapshot.all.length,
     eventCount: events.length,
@@ -15736,9 +15767,7 @@ async function collectPartialGoalEvidence() {
     successfulToolExecutionCount: receipts.filter(
       (record) => record.status === 'ok',
     ).length,
-    toolPlanProtocolCount: agentDb.filter(
-      (record) => record.recordType === 'agent.runtime.tool_plan.protocol',
-    ).length,
+    ...nativeToolPlanProtocolEvidence,
     structuredPlanRevision: runtime?.planRevision ?? 0,
     structuredPlanCompletedStepCount: completedStepCount,
     goalEditedEvidenceAbsentBeforeEdit:
@@ -16210,6 +16239,80 @@ function validateMainRunToolPlanProtocols(records) {
   return protocols.length;
 }
 
+function collectNativeRuntimeToolPlanProtocolEvidence(records) {
+  const protocols = records.filter(
+    (record) =>
+      record.recordType === 'agent.runtime.tool_plan.protocol' &&
+      record.agentId === mainAgentId &&
+      record.runId === state.initialRunId,
+  );
+  const repairs = records.filter(
+    (record) =>
+      record.recordType === 'agent.runtime.tool_plan.repair' &&
+      record.agentId === mainAgentId &&
+      record.runId === state.initialRunId,
+  );
+  const audits = [...protocols, ...repairs];
+  return {
+    toolPlanProtocolCount: protocols.length,
+    nativeRuntimeToolPlanCount: protocols.filter(
+      (record) => record.protocol === 'native_runtime_tools',
+    ).length,
+    toolPlanRepairCount: repairs.length,
+    nativeRuntimeToolPlanRepairCount: repairs.filter(
+      (record) => record.protocol === 'native_runtime_tools',
+    ).length,
+    wrapperToolPlanFallbackCount: audits.filter(
+      (record) => record.protocol === 'native_function',
+    ).length,
+    textJsonToolPlanFallbackCount: audits.filter(
+      (record) => record.protocol === 'text_json',
+    ).length,
+    toolPlanAuditPayloadLeakCount: audits.filter(
+      (record) =>
+        Object.hasOwn(record, 'arguments') ||
+        Object.hasOwn(record, 'response') ||
+        Object.hasOwn(record, 'toolArguments'),
+    ).length,
+  };
+}
+
+function validateNativeRuntimeToolPlanProtocolEvidence(records) {
+  const protocolCount = validateMainRunToolPlanProtocols(records);
+  const evidence = collectNativeRuntimeToolPlanProtocolEvidence(records);
+  const protocols = records.filter(
+    (record) =>
+      record.recordType === 'agent.runtime.tool_plan.protocol' &&
+      record.agentId === mainAgentId &&
+      record.runId === state.initialRunId,
+  );
+  assert(
+    evidence.toolPlanProtocolCount === protocolCount &&
+      evidence.nativeRuntimeToolPlanCount === protocolCount &&
+      evidence.nativeRuntimeToolPlanRepairCount ===
+        evidence.toolPlanRepairCount &&
+      evidence.wrapperToolPlanFallbackCount === 0 &&
+      evidence.textJsonToolPlanFallbackCount === 0 &&
+      evidence.toolPlanAuditPayloadLeakCount === 0 &&
+      protocols.every(
+        (record) =>
+          Number.isSafeInteger(record.functionCallCount) &&
+          record.functionCallCount > 0 &&
+          Array.isArray(record.callIds) &&
+          record.callIds.length === record.functionCallCount &&
+          new Set(record.callIds).size === record.callIds.length &&
+          Array.isArray(record.functionNames) &&
+          record.functionNames.length === record.functionCallCount &&
+          record.functionNames.every(
+            (name) =>
+              isNonEmptyString(name) && name !== 'submit_agent_tool_plan',
+          ),
+      ),
+    'goal-native-tool-plan-protocol-required',
+  );
+  return evidence;
+}
+
 async function validateSameRunSteerEvidence({
   agentDb,
   activity,