fix(qa): align runtime parity evidence with Codex

extensions/qa-lab/src/runtime-parity.test.ts

@@ -168,23 +168,42 @@ describe("runtime parity", () => {
     const scoped = __testing.filterMockRequestsForParentPrompt(
       [
         {
+          prompt: "Fanout worker alpha: inspect the QA workspace and finish with exactly ALPHA-OK.",
+          allInputText:
+            "Delegate one bounded QA task to a subagent. Fanout worker alpha: inspect the QA workspace and finish with exactly ALPHA-OK.",
+          plannedToolName: "read",
+        },
+        {
+          prompt: "Delegate one bounded QA task to a subagent.",
           allInputText: "Delegate one bounded QA task to a subagent.",
           plannedToolName: "sessions_spawn",
         },
+        {
+          prompt: "Continue the bounded QA task with the retained child result.",
+          allInputText:
+            "Delegate one bounded QA task to a subagent. Continue the bounded QA task with the retained child result.",
+          plannedToolName: "sessions_spawn",
+        },
         {
           allInputText: "Inspect the QA workspace and return one concise protocol note.",
           plannedToolName: "read",
         },
         {
+          prompt: "Delegate one bounded QA task to a subagent.",
           allInputText: "Delegate one bounded QA task to a subagent. Tool result: child accepted.",
           toolOutput: "child accepted",
         },
       ],
       "Delegate one bounded QA task to a subagent.",
+      [
+        "Delegate one bounded QA task to a subagent.",
+        "Continue the bounded QA task with the retained child result.",
+      ],
     );
 
-    expect(scoped).toHaveLength(2);
+    expect(scoped).toHaveLength(3);
     expect(scoped.map((request) => request.plannedToolName ?? "result")).toEqual([
+      "sessions_spawn",
       "sessions_spawn",
       "result",
     ]);

extensions/qa-lab/src/runtime-parity.ts

@@ -120,6 +120,7 @@ type RuntimeParityTranscriptRecord = {
 };
 
 type RuntimeParityMockRequestSnapshot = {
+  prompt?: string;
   allInputText?: string;
   plannedToolName?: string;
   plannedToolArgs?: unknown;
@@ -759,14 +760,22 @@ function resolveRuntimeParityToolCalls(params: {
 function filterMockRequestsForParentPrompt(
   requests: RuntimeParityMockRequestSnapshot[],
   parentPrompt: string,
+  parentPrompts: readonly string[] = [parentPrompt],
 ) {
-  const normalizedParentPrompt = normalizeTextForParity(parentPrompt);
-  if (!normalizedParentPrompt) {
+  const normalizedParentPrompts = parentPrompts
+    .map(normalizeTextForParity)
+    .filter((prompt) => prompt.length > 0);
+  if (normalizedParentPrompts.length === 0) {
     return requests;
   }
-  const matching = requests.filter((request) =>
-    normalizeTextForParity(request.allInputText ?? "").includes(normalizedParentPrompt),
-  );
+  const matching = requests.filter((request) => {
+    const normalizedPrompt = normalizeTextForParity(request.prompt ?? "");
+    if (normalizedPrompt) {
+      return normalizedParentPrompts.some((prompt) => normalizedPrompt.includes(prompt));
+    }
+    const normalizedHistory = normalizeTextForParity(request.allInputText ?? "");
+    return normalizedParentPrompts.some((prompt) => normalizedHistory.includes(prompt));
+  });
   return matching.length > 0 ? matching : requests;
 }
 
@@ -966,6 +975,7 @@ async function loadRuntimeParityTranscripts(params: {
 async function loadRuntimeParityMockToolCalls(
   mockBaseUrl: string | undefined,
   parentPrompt: string,
+  parentPrompts: readonly string[] = [parentPrompt],
 ): Promise<RuntimeParityToolCall[] | null> {
   const normalizedBaseUrl = mockBaseUrl?.trim().replace(/\/+$/u, "");
   if (!normalizedBaseUrl) {
@@ -991,6 +1001,7 @@ async function loadRuntimeParityMockToolCalls(
     }
     const requests = payload.filter(isMessageRecord).map(
       (entry): RuntimeParityMockRequestSnapshot => ({
+        prompt: readNonEmptyString(entry.prompt),
         allInputText: readNonEmptyString(entry.allInputText),
         plannedToolName: readNonEmptyString(entry.plannedToolName),
         plannedToolArgs: entry.plannedToolArgs ?? null,
@@ -998,7 +1009,7 @@ async function loadRuntimeParityMockToolCalls(
       }),
     );
     return resolveToolCallOrderFromMockRequests(
-      filterMockRequestsForParentPrompt(requests, parentPrompt),
+      filterMockRequestsForParentPrompt(requests, parentPrompt, parentPrompts),
     );
   } catch {
     return null;
@@ -1015,12 +1026,16 @@ export async function captureRuntimeParityCell(
   });
   const transcriptRecords = buildTranscriptRecords(transcriptBytes);
   const transcriptToolCalls = resolveToolCallOrder(transcriptRecords);
-  const parentPrompt =
-    transcriptRecords
-      .filter((record) => record.role === "user" && !isToolResultLikeMessage(record.message))
-      .map((record) => extractAssistantText(record.message))
-      .find(Boolean) ?? "";
-  const mockToolCalls = await loadRuntimeParityMockToolCalls(params.mockBaseUrl, parentPrompt);
+  const parentPrompts = transcriptRecords
+    .filter((record) => record.role === "user")
+    .map((record) => extractAssistantText(record.message))
+    .filter((prompt) => prompt.length > 0);
+  const parentPrompt = parentPrompts[0] ?? "";
+  const mockToolCalls = await loadRuntimeParityMockToolCalls(
+    params.mockBaseUrl,
+    parentPrompt,
+    parentPrompts,
+  );
   const gatewayLogs = params.gateway.logs?.();
   const sentinelFindings = [
     ...scanGatewayLogSentinels(gatewayLogs),

qa/scenarios/runtime/long-context-cache-stability.yaml

@@ -26,7 +26,9 @@ scenario:
     config:
       sessionKey: agent:qa:long-context-cache-stability
       fixtureFile: large-cache-fixture.txt
-      cacheEvidenceNeedle: CACHE-FIXTURE-0550
+      cacheEvidenceNeedle: CACHE-FIXTURE-0050
+      cacheEvidenceLine: "CACHE-FIXTURE-0050: stable tool-result evidence for prompt-cache reuse across long sessions."
+      followupPromptNeedle: Using the already-read
       warmupMarker: QA-LARGE-CACHE-WARMUP-OK
       hitMarker: QA-LARGE-CACHE-HIT-OK
 
@@ -84,8 +86,17 @@ flow:
         - set: debugRequests
           value:
             expr: "env.mock ? [...(await fetchJson(`${env.mock.baseUrl}/debug/requests`))] : []"
+        - set: cappedReadOutputIndex
+          value:
+            expr: "debugRequests.reduce((found, planned, index) => { if (found >= 0 || !planned.plannedToolCallId || planned.plannedToolName !== 'read' || planned.plannedToolArgs?.path !== config.fixtureFile) return found; const outputOffset = debugRequests.slice(index + 1).findIndex((candidate) => Boolean(candidate.toolOutputCallId) && candidate.toolOutputCallId === planned.plannedToolCallId); if (outputOffset < 0) return found; const output = debugRequests[index + 1 + outputOffset]; const evidence = [planned.allInputText, output.allInputText, output.toolOutput].filter((value) => typeof value === 'string').join('\\n'); const hasCodexFormattedTruncation = evidence.includes('Warning: truncated output') && (evidence.includes('chars truncated') || evidence.includes('tokens truncated')); return evidence.includes(config.cacheEvidenceLine) && (evidence.includes('[Read output capped at 50KB') || evidence.includes('...(OpenClaw truncated dynamic tool result') || evidence.includes('...(truncated)...') || hasCodexFormattedTruncation) ? index + 1 + outputOffset : found; }, -1)"
+        - set: hasCappedReadEvidence
+          value:
+            expr: "cappedReadOutputIndex >= 0"
+        - set: hasFollowupCacheEvidence
+          value:
+            expr: "cappedReadOutputIndex >= 0 && debugRequests.some((request, index) => index > cappedReadOutputIndex && String(request.prompt ?? '').includes(config.followupPromptNeedle) && String(request.allInputText ?? '').includes(config.cacheEvidenceLine))"
         - assert:
-            expr: "!env.mock || debugRequests.some((request, index) => request.plannedToolName === 'read' && request.plannedToolArgs?.path === config.fixtureFile && typeof request.plannedToolCallId === 'string' && debugRequests.slice(index + 1).some((result, resultOffset) => result.toolOutputCallId === request.plannedToolCallId && String(result.toolOutput ?? '').includes(config.cacheEvidenceNeedle) && (String(result.toolOutput ?? '').includes('[Read output capped at 50KB') || (String(result.toolOutput ?? '').includes('...(truncated)...') && String(result.toolOutput ?? '').length <= 13000)) && debugRequests.slice(index + resultOffset + 2).some((followup) => followup.plannedToolName === 'read' && followup.plannedToolArgs?.path === config.fixtureFile && String(followup.allInputText ?? '').includes(config.cacheEvidenceNeedle) && (String(followup.allInputText ?? '').includes('[Read output capped at 50KB') || String(followup.allInputText ?? '').includes('...(truncated)...')))))"
+            expr: "!env.mock || (hasCappedReadEvidence && hasFollowupCacheEvidence)"
             message:
-              expr: "`large capped read tool result was not observed: ${JSON.stringify(debugRequests.slice(-8).map((request) => ({ plannedToolName: request.plannedToolName ?? null, plannedToolArgs: request.plannedToolArgs ?? null, plannedToolCallId: request.plannedToolCallId ?? null, toolOutputCallId: request.toolOutputCallId ?? null, toolOutputLength: String(request.toolOutput ?? '').length, toolOutputHasNeedle: String(request.toolOutput ?? '').includes(config.cacheEvidenceNeedle), toolOutputHasReadCap: String(request.toolOutput ?? '').includes('[Read output capped at 50KB'), toolOutputHasCodexTruncation: String(request.toolOutput ?? '').includes('...(truncated)...'), inputHasNeedle: String(request.allInputText ?? '').includes(config.cacheEvidenceNeedle), inputHasReadCap: String(request.allInputText ?? '').includes('[Read output capped at 50KB'), inputHasCodexTruncation: String(request.allInputText ?? '').includes('...(truncated)...') })))}`"
+              expr: "`large capped read cache evidence was not observed: ${JSON.stringify({ hasCappedReadEvidence, hasFollowupCacheEvidence, requests: debugRequests.slice(-8).map((request) => ({ prompt: request.prompt ?? null, plannedToolName: request.plannedToolName ?? null, plannedToolArgs: request.plannedToolArgs ?? null, plannedToolCallId: request.plannedToolCallId ?? null, toolOutputCallId: request.toolOutputCallId ?? null, toolOutputLength: String(request.toolOutput ?? '').length, outputHasReadCap: String(request.toolOutput ?? '').includes('[Read output capped at 50KB'), outputHasCodexTruncation: String(request.toolOutput ?? '').includes('...(truncated)...'), inputHasEvidenceLine: String(request.allInputText ?? '').includes(config.cacheEvidenceLine) })) })}`"
       detailsExpr: "outbound?.text ?? config.hitMarker"