test(matrix): force block-streaming marker separation

steipete · steipete · commit b01c6d4eaa0b · 2026-05-27T13:58:42.000-04:00
diff --git a/extensions/qa-matrix/src/runners/contract/scenario-runtime-shared.ts b/extensions/qa-matrix/src/runners/contract/scenario-runtime-shared.ts
@@ -137,8 +137,10 @@ export function buildMatrixBlockStreamingPrompt(
   return [
     `${sutUserId} Block streaming QA check: complete this whole sequence in one turn.`,
     `Step 1: send an assistant text block containing only this exact marker: \`${firstText}\`.`,
-    "Step 2: do not stop after that marker. Use the read tool exactly once on `QA_KICKOFF_TASK.md`.",
+    "That first marker block must be emitted before any tool call.",
+    "Step 2: after the first marker block, use the read tool exactly once on `QA_KICKOFF_TASK.md`.",
     `Step 3: after that read completes, send a final assistant text block containing only this exact marker: \`${secondText}\`.`,
+    "Never put both markers in the same assistant text block.",
   ].join("\n");
 }
 
diff --git a/extensions/qa-matrix/src/runners/contract/scenarios.test.ts b/extensions/qa-matrix/src/runners/contract/scenarios.test.ts
@@ -4164,12 +4164,14 @@ describe("matrix live qa scenarios", () => {
     expect(body).toMatch(
       /Step 1: send an assistant text block containing only this exact marker: `MATRIX_QA_BLOCK_ONE_[A-F0-9]{8}`/,
     );
+    expect(body).toContain("That first marker block must be emitted before any tool call.");
     expect(body).toContain(
-      "Step 2: do not stop after that marker. Use the read tool exactly once on `QA_KICKOFF_TASK.md`.",
+      "Step 2: after the first marker block, use the read tool exactly once on `QA_KICKOFF_TASK.md`.",
     );
     expect(body).toMatch(
       /Step 3: after that read completes, send a final assistant text block containing only this exact marker: `MATRIX_QA_BLOCK_TWO_[A-F0-9]{8}`/,
     );
+    expect(body).toContain("Never put both markers in the same assistant text block.");
     expect(scenario.configOverrides?.toolProfile).toBe("coding");
     expect(mockObjectArg(waitForRoomEvent, "waitForRoomEvent", 1).since).toBe(
       "driver-sync-block-one",
