loop-benchmarking

commit 7f450138909c4de86b94981aca1ebbb1a9defa82
parent 08a16207bbb7cb2cc8da116886da99149ea093da
Author: Brian Graham <brian@buildingbetterteams.de>
Date:   Sat,  4 Apr 2026 08:17:46 +0200

Add HTML validation, duplication detection, accessibility, page load time

code-analysis.sh:
- HTML validation via html-validate (error count, pass/fail)
- Code duplication percentage via jscpd
- Both factor into the code analysis score

gameplay bot:
- Page load time measurement (ms)
- Lightweight accessibility checks: page title, headings, canvas
  aria-label, image alt text, color contrast
- Results included in bot JSON report
- Timeout bumped to 3 minutes

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Diffstat:
M
tasks/tetris/eval/code-analysis.sh
 | 
47
+++++++++++++++++++++++++++++++++++++++++++++++
M
tasks/tetris/eval/gameplay-bot/index.ts
 | 
67
++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++-
M
tasks/tetris/eval/gameplay-bot/types.ts
 | 
8
++++++++

3 files changed, 121 insertions(+), 1 deletion(-)
diff --git a/tasks/tetris/eval/code-analysis.sh b/tasks/tetris/eval/code-analysis.sh
@@ -74,6 +74,40 @@ results=$(echo "$results" | jq --arg c "$complexity" '. + {complexity: $c}')
 console_logs=$(grep -r "console\.log" --include="*.ts" --include="*.js" . 2>/dev/null | grep -v node_modules | wc -l)
 results=$(echo "$results" | jq --argjson cl "$console_logs" '. + {console_logs: $cl}')
 
+# --- HTML validation ---
+html_valid="unknown"
+html_errors=0
+if [ -f "index.html" ] || find . -name "*.html" -not -path "*/node_modules/*" | head -1 | grep -q .; then
+  npm install --save-dev html-validate > /dev/null 2>&1
+  html_file=$(find . -name "index.html" -not -path "*/node_modules/*" | head -1)
+  if [ -n "$html_file" ]; then
+    html_output=$(npx html-validate --formatter json "$html_file" 2>/dev/null) || true
+    if echo "$html_output" | jq . > /dev/null 2>&1; then
+      html_errors=$(echo "$html_output" | jq '[.[].errorCount] | add // 0')
+      html_errors=${html_errors:-0}
+      if [ "$html_errors" -eq 0 ]; then
+        html_valid="true"
+      else
+        html_valid="false"
+      fi
+    fi
+  fi
+fi
+results=$(echo "$results" | jq --arg v "$html_valid" --argjson e "$html_errors" \
+  '. + {html_validation: {valid: ($v == "true"), errors: $e}}')
+
+# --- Code duplication ---
+duplication_pct=0
+npm install --save-dev jscpd > /dev/null 2>&1
+dupl_output=$(npx jscpd --min-lines 5 --min-tokens 50 --reporters json \
+  --ignore "node_modules,package-lock.json" . 2>/dev/null) || true
+if echo "$dupl_output" | jq . > /dev/null 2>&1; then
+  duplication_pct=$(echo "$dupl_output" | jq '.statistics.total.percentage // 0')
+  duplication_pct=${duplication_pct:-0}
+fi
+results=$(echo "$results" | jq --argjson d "$duplication_pct" \
+  '. + {duplication_percentage: $d}')
+
 # --- Compute score ---
 # Scoring: fewer unnecessary files, fewer deps, moderate LOC, no debug noise
 score=100
@@ -100,6 +134,19 @@ if [ "$complexity" = "over-engineered" ]; then
   score=$((score - 10))
 fi
 
+# Penalty for invalid HTML (5 points)
+if [ "$html_valid" = "false" ]; then
+  score=$((score - 5))
+fi
+
+# Penalty for high duplication (over 10%)
+dup_int=$(awk "BEGIN {printf \"%d\", $duplication_pct}")
+if [ "$dup_int" -gt 10 ]; then
+  score=$((score - 10))
+elif [ "$dup_int" -gt 5 ]; then
+  score=$((score - 5))
+fi
+
 # Normalize to 0-1
 score_normalized=$(awk "BEGIN {s = $score / 100; if (s < 0) s = 0; printf \"%.2f\", s}")
 
diff --git a/tasks/tetris/eval/gameplay-bot/index.ts b/tasks/tetris/eval/gameplay-bot/index.ts
@@ -92,10 +92,67 @@ test.describe("Tetris Gameplay Bot", () => {
   });
 
   test("run gameplay bot", async ({ page }) => {
-    test.setTimeout(120_000); // 2-minute total timeout
+    test.setTimeout(180_000); // 3-minute total timeout
+
+    // Measure page load time
+    let loadTimeMs = -1;
+    try {
+      const loadStart = Date.now();
+      await page.goto(serverUrl, { waitUntil: "load", timeout: 10000 });
+      loadTimeMs = Date.now() - loadStart;
+      // Navigate away so runAllTests starts fresh
+      await page.goto("about:blank");
+    } catch {
+      // Load time measurement failed, not critical
+    }
 
     const { testResults, calibration, gameplay } = await runAllTests(page, serverUrl);
 
+    // Accessibility check via page evaluation (lightweight, no axe-core dependency)
+    let a11yIssues: string[] = [];
+    try {
+      await page.goto(serverUrl, { timeout: 10000 });
+      await page.waitForTimeout(1000);
+      a11yIssues = await page.evaluate(() => {
+        const issues: string[] = [];
+        // Check page title
+        if (!document.title || document.title.trim() === "") {
+          issues.push("missing page title");
+        }
+        // Check for headings
+        if (document.querySelectorAll("h1, h2, h3, [role='heading']").length === 0) {
+          issues.push("no headings found");
+        }
+        // Check images have alt text
+        document.querySelectorAll("img").forEach((img) => {
+          if (!img.alt && !img.getAttribute("aria-label")) {
+            issues.push("image without alt text");
+          }
+        });
+        // Check canvas has accessible label
+        document.querySelectorAll("canvas").forEach((canvas) => {
+          if (!canvas.getAttribute("aria-label") && !canvas.getAttribute("role")) {
+            issues.push("canvas without aria-label or role");
+          }
+        });
+        // Check for focus indicators
+        const focusable = document.querySelectorAll("button, a, input, [tabindex]");
+        if (focusable.length === 0 && document.querySelectorAll("canvas").length === 0) {
+          issues.push("no focusable elements");
+        }
+        // Check color contrast on text elements (basic check)
+        const body = window.getComputedStyle(document.body);
+        const bgColor = body.backgroundColor;
+        const textColor = body.color;
+        if (bgColor === textColor) {
+          issues.push("text color matches background color");
+        }
+        return issues;
+      });
+    } catch {
+      // a11y check failed, not critical
+    }
+
     const passed = testResults.filter((t) => t.pass).length;
     const failed = testResults.filter((t) => !t.pass).length;
     const total = testResults.length;
@@ -117,6 +174,14 @@ test.describe("Tetris Gameplay Bot", () => {
         score: total > 0 ? Math.round((passed / total) * 100) / 100 : 0,
       },
       gameplay,
+      performance: {
+        load_time_ms: loadTimeMs,
+      },
+      accessibility: {
+        issues: a11yIssues,
+        issue_count: a11yIssues.length,
+        pass: a11yIssues.length === 0,
+      },
     };
 
     // Write report to file
diff --git a/tasks/tetris/eval/gameplay-bot/types.ts b/tasks/tetris/eval/gameplay-bot/types.ts
@@ -81,6 +81,14 @@ export interface BotReport {
     score: number;
   };
   gameplay: GameplayStats;
+  performance?: {
+    load_time_ms: number;
+  };
+  accessibility?: {
+    issues: string[];
+    issue_count: number;
+    pass: boolean;
+  };
 }
 
 /** Context passed through calibration, play, and reporting phases. */