diff --git a/.changeset/silver-coins-deny.md b/.changeset/silver-coins-deny.md
new file mode 100644
index 0000000000..a845151cc8
--- /dev/null
+++ b/.changeset/silver-coins-deny.md
@@ -0,0 +1,2 @@
+---
+---
diff --git a/.github/scripts/aggregate-e2e-results.js b/.github/scripts/aggregate-e2e-results.js
index b322d83f10..764ca1ba46 100644
--- a/.github/scripts/aggregate-e2e-results.js
+++ b/.github/scripts/aggregate-e2e-results.js
@@ -54,6 +54,7 @@ function findResultFiles(dir) {
     'e2e-metadata-',
     'e2e-failures-',
     'e2e-diagnostics-',
+    'e2e-runtime-logs-',
   ]);
 }
 
diff --git a/.github/scripts/fetch-e2e-runtime-logs.mjs b/.github/scripts/fetch-e2e-runtime-logs.mjs
new file mode 100644
index 0000000000..ce1ca97e81
--- /dev/null
+++ b/.github/scripts/fetch-e2e-runtime-logs.mjs
@@ -0,0 +1,226 @@
+#!/usr/bin/env node
+/**
+ * Fetches Vercel runtime request logs relevant to a failed e2e lane and
+ * writes them to a JSON sidecar that CI uploads as an artifact. Intended to
+ * run only when the e2e test step failed.
+ *
+ * Runtime logs are the missing piece when triaging e2e flakes: the test
+ * harness captures workflow run state and event timelines, but not what the
+ * deployed functions actually logged (cold starts, crashes, stderr). This
+ * uses the same request-logs API as `vercel logs` (one entry per request,
+ * with log messages, status code, and — for workflow routes — the
+ * workflowRunId attached), so failed tests can be correlated to server-side
+ * behavior directly from the artifact.
+ *
+ * The deployment is shared with concurrent CI runs and a test window can
+ * span tens of thousands of request rows, so instead of a bulk dump this
+ * makes targeted queries:
+ *   1. all rows with error/fatal-level log lines in the test window, and
+ *   2. all rows belonging to each failed test's workflow run ID (from the
+ *      github-reporter failures sidecar).
+ *
+ * Required env: WORKFLOW_VERCEL_AUTH_TOKEN, WORKFLOW_VERCEL_TEAM,
+ *               WORKFLOW_VERCEL_PROJECT, VERCEL_DEPLOYMENT_ID, APP_NAME
+ * Optional env: E2E_START_MS (window start; defaults to 90 minutes ago),
+ *               E2E_END_MS (window end; defaults to now — useful for
+ *               re-fetching a historical window when triaging locally),
+ *               WORKFLOW_VERCEL_ENV (production | preview)
+ */
+
+import fs from 'node:fs';
+import path from 'node:path';
+
+const API_URL = 'https://vercel.com/api/logs/request-logs';
+// Each page returns up to 50 rows; rows come newest-first, so hitting a cap
+// drops the oldest rows. Caps bound artifact size and API load.
+const MAX_PAGES_PER_QUERY = 40;
+// Cap per-row console output; the full data is in the JSON artifact.
+const MAX_PRINTED_ROWS = 60;
+
+const {
+  WORKFLOW_VERCEL_AUTH_TOKEN: token,
+  WORKFLOW_VERCEL_TEAM: teamId,
+  WORKFLOW_VERCEL_PROJECT: projectId,
+  VERCEL_DEPLOYMENT_ID: deploymentId,
+  WORKFLOW_VERCEL_ENV: environment,
+  APP_NAME: appName,
+  E2E_START_MS: e2eStartMs,
+  E2E_END_MS: e2eEndMs,
+} = process.env;
+
+if (!token || !teamId || !projectId || !deploymentId || !appName) {
+  console.log(
+    '[runtime-logs] Missing required env (token/team/project/deployment/app); skipping capture.'
+  );
+  process.exit(0);
+}
+
+// Small buffer before the recorded start so requests already in flight when
+// the test step began are included.
+const startDate = e2eStartMs
+  ? Number(e2eStartMs) - 2 * 60 * 1000
+  : Date.now() - 90 * 60 * 1000;
+const endDate = e2eEndMs ? Number(e2eEndMs) : Date.now() + 60 * 1000;
+
+/** Read failed-test run IDs from the github-reporter failures sidecar, if present. */
+function readFailedRunIds() {
+  const failedRunIds = new Map(); // runId -> testName
+  const failuresPath = path.resolve(
+    process.cwd(),
+    `e2e-failures-${appName}-vercel.json`
+  );
+  try {
+    const failures = JSON.parse(fs.readFileSync(failuresPath, 'utf-8'));
+    for (const failure of failures) {
+      if (failure.runId) failedRunIds.set(failure.runId, failure.testName);
+    }
+  } catch {
+    // Sidecar absent (no failures, or reporter didn't run) — fine.
+  }
+  return failedRunIds;
+}
+
+async function fetchPage(extraParams, page) {
+  const url = new URL(API_URL);
+  url.searchParams.set('projectId', projectId);
+  url.searchParams.set('ownerId', teamId);
+  url.searchParams.set('deploymentId', deploymentId);
+  url.searchParams.set('startDate', String(startDate));
+  url.searchParams.set('endDate', String(endDate));
+  url.searchParams.set('page', String(page));
+  if (environment) url.searchParams.set('environment', environment);
+  for (const [key, value] of Object.entries(extraParams)) {
+    url.searchParams.set(key, value);
+  }
+
+  const response = await fetch(url, {
+    headers: { Authorization: `Bearer ${token}` },
+  });
+  if (!response.ok) {
+    throw new Error(
+      `request-logs API returned ${response.status}: ${(await response.text()).slice(0, 300)}`
+    );
+  }
+  return response.json();
+}
+
+async function fetchAllPages(label, extraParams) {
+  const rows = [];
+  let truncated = false;
+  for (let page = 0; page < MAX_PAGES_PER_QUERY; page++) {
+    let data;
+    try {
+      data = await fetchPage(extraParams, page);
+    } catch (error) {
+      console.log(
+        `[runtime-logs] ${label}: fetch failed on page ${page}: ${error}`
+      );
+      break;
+    }
+    rows.push(...(data.rows || []));
+    if (!data.hasMoreRows) break;
+    if (page === MAX_PAGES_PER_QUERY - 1) truncated = true;
+  }
+  if (truncated) {
+    console.log(
+      `[runtime-logs] WARNING: ${label} hit the ${MAX_PAGES_PER_QUERY}-page cap; older rows were dropped.`
+    );
+  }
+  return rows;
+}
+
+async function main() {
+  console.log(
+    `[runtime-logs] Capturing logs for ${deploymentId} (${new Date(startDate).toISOString()} → ${new Date(endDate).toISOString()})`
+  );
+
+  const failedRunIds = readFailedRunIds();
+  const rowsByRequestId = new Map();
+  const addRows = (rows) => {
+    for (const row of rows) {
+      rowsByRequestId.set(
+        row.requestId ?? `${row.timestamp}-${rowsByRequestId.size}`,
+        row
+      );
+    }
+  };
+
+  // 1. Everything that logged at error/fatal level in the window — catches
+  //    crashes and failures not attributable to a tracked workflow run.
+  addRows(await fetchAllPages('error sweep', { level: 'error,fatal' }));
+
+  // 2. Every request belonging to a failed test's workflow run.
+  for (const [runId, testName] of failedRunIds) {
+    const rows = await fetchAllPages(`run ${runId}`, { search: runId });
+    console.log(
+      `[runtime-logs] ${runId} (${testName}): ${rows.length} request(s)`
+    );
+    addRows(rows);
+  }
+
+  const rows = [...rowsByRequestId.values()].sort((a, b) =>
+    String(a.timestamp).localeCompare(String(b.timestamp))
+  );
+
+  const outPath = path.resolve(
+    process.cwd(),
+    `e2e-runtime-logs-${appName}-vercel.json`
+  );
+  fs.writeFileSync(
+    outPath,
+    JSON.stringify(
+      {
+        fetchedAt: new Date().toISOString(),
+        deploymentId,
+        projectId,
+        environment,
+        window: {
+          start: new Date(startDate).toISOString(),
+          end: new Date(endDate).toISOString(),
+        },
+        failedRunIds: Object.fromEntries(failedRunIds),
+        rows,
+      },
+      null,
+      2
+    )
+  );
+  console.log(`[runtime-logs] Wrote ${rows.length} rows to ${outPath}`);
+
+  // Console summary so the most important signals are visible in the job log
+  // without downloading the artifact.
+  let printed = 0;
+  let suppressed = 0;
+  for (const row of rows) {
+    if (printed >= MAX_PRINTED_ROWS) {
+      suppressed++;
+      continue;
+    }
+    printed++;
+
+    const isFailedRun =
+      row.workflowRunId && failedRunIds.has(row.workflowRunId);
+    const marker = isFailedRun
+      ? `failed-test run ${row.workflowRunId} (${failedRunIds.get(row.workflowRunId)})`
+      : row.workflowRunId || 'no run id';
+    console.log(
+      `[runtime-logs] ${row.timestamp} ${row.requestMethod} ${row.requestPath} → ${row.statusCode} [${marker}]`
+    );
+    for (const log of row.logs || []) {
+      console.log(
+        `    [${log.level}] ${(log.message || '').split('\n')[0].slice(0, 300)}`
+      );
+    }
+  }
+  if (suppressed > 0) {
+    console.log(
+      `[runtime-logs] ...and ${suppressed} more rows — see the e2e-runtime-logs artifact.`
+    );
+  }
+}
+
+main().catch((error) => {
+  // Diagnostics capture must never turn a red job into a different red job.
+  console.log(`[runtime-logs] Capture failed: ${error}`);
+  process.exit(0);
+});
diff --git a/.github/workflows/tests.yml b/.github/workflows/tests.yml
index 57cfaef595..1665b31aaa 100644
--- a/.github/workflows/tests.yml
+++ b/.github/workflows/tests.yml
@@ -377,6 +377,10 @@ jobs:
           check-interval: 15
           environment: ${{ github.ref == 'refs/heads/main' && 'production' || 'preview' }}
 
+      - name: Record E2E start time
+        id: e2eStart
+        run: echo "ms=$(($(date +%s) * 1000))" >> "$GITHUB_OUTPUT"
+
       - name: Run E2E Tests
         run: pnpm run test:e2e --reporter=default --reporter=json --reporter=./packages/core/e2e/github-reporter.ts "--outputFile=e2e-vercel-prod-$APP_NAME.json"
         env:
@@ -401,6 +405,18 @@ jobs:
           # so production runs hit the public vercel-workflow.com URL.
           VERCEL_WORKFLOW_SERVER_URL: ${{ github.ref != 'refs/heads/main' && secrets.VERCEL_WORKFLOW_SERVER_URL || '' }}
 
+      - name: Capture runtime logs on failure
+        if: failure()
+        env:
+          APP_NAME: ${{ matrix.app.name }}
+          WORKFLOW_VERCEL_AUTH_TOKEN: ${{ secrets.VERCEL_LABS_TOKEN }}
+          WORKFLOW_VERCEL_TEAM: "team_nO2mCG4W8IxPIeKoSsqwAxxB"
+          WORKFLOW_VERCEL_PROJECT: ${{ matrix.app.project-id }}
+          WORKFLOW_VERCEL_ENV: ${{ github.ref == 'refs/heads/main' && 'production' || 'preview' }}
+          VERCEL_DEPLOYMENT_ID: ${{ steps.waitForDeployment.outputs.deployment-id }}
+          E2E_START_MS: ${{ steps.e2eStart.outputs.ms }}
+        run: node .github/scripts/fetch-e2e-runtime-logs.mjs
+
       - name: Generate E2E summary
         if: always()
         env:
@@ -417,6 +433,7 @@ jobs:
             e2e-metadata-${{ matrix.app.name }}-vercel.json
             e2e-failures-${{ matrix.app.name }}-vercel.json
             e2e-diagnostics-${{ matrix.app.name }}-vercel.json
+            e2e-runtime-logs-${{ matrix.app.name }}-vercel.json
           retention-days: 7
           if-no-files-found: ignore