Add per-skill "Skill Value" view to the evaluation dashboard

.github/workflows/evaluation.yml

@@ -1829,21 +1829,40 @@ jobs:
               continue
             }
 
-            # Parse once, then group by the executor model recorded inside each
-            # results.json. For a one-model profile (e.g. opus48) all files share
-            # one model, so this yields exactly one group.
+            # Parse once, then group by the (executor model, judge model) pair
+            # recorded inside each results.json. Grouping by model ALONE collapsed
+            # every judge of one executor into a single entry stamped with the first
+            # file's judgeModel — so a dual-judge run mislabelled the second judge and
+            # concatenated both judges' verdicts under one label. Keying on model+judge
+            # keeps each judge's verdicts and its correct judgeModel, which is exactly
+            # the dimension the Skill Value view keys on. For a single-judge profile
+            # all files share one judge, so this still yields one group per model.
+            #
+            # Scope note: only the PRIMARY judge's results.json (vally-results-*) feed
+            # this benchmark data. The scheduled dual-judge cadence re-scores the same
+            # executor trajectories and uploads under vally-crossjudge-* (see the
+            # "Download second-judge" step), which only feeds judge-comparison.json —
+            # never this loop. So Skill Value and Quality/Efficiency both reflect the
+            # primary judge, and each executor model here carries a single judge.
+            # (The model+judge key and -SkillValueOnly dedup below stay correct even
+            # if a future change ever routes two primary judges into this set.)
             $parsed = @($resultsFiles | Sort-Object | ForEach-Object {
               [pscustomobject]@{ File = $_; Json = (Get-Content $_ -Raw | ConvertFrom-Json) }
             })
-            $modelGroups = @($parsed | Group-Object -Property { "$($_.Json.model)" })
+            $modelGroups = @($parsed | Group-Object -Property { "$($_.Json.model)|$($_.Json.judgeModel)" })
 
             $existingFile = "/tmp/eval-data/data/$plugin.json"
+            # Quality/Efficiency views key on executor model alone, so emit them
+            # once per model. Track which executor models already emitted them; the
+            # second+ judge of one model runs SkillValue-only to avoid duplicate
+            # same-model Quality/Efficiency points that would inflate those windows.
+            $qeEmittedModels = @{}
             foreach ($mg in $modelGroups) {
               $group = @($mg.Group)
-              # Merge this model's per-skill/per-shard verdicts into a single
-              # synthetic results.json. Schema: { model, verdicts[], ... }. Concat
-              # the verdicts arrays and keep top-level scalars from the first file
-              # of the group (all share the same model/judgeModel).
+              # Merge this (model, judge) pair's per-skill/per-shard verdicts into a
+              # single synthetic results.json. Schema: { model, judgeModel, verdicts[], ... }.
+              # Concat the verdicts arrays and keep top-level scalars from the first
+              # file of the group (all now share the same model AND judgeModel).
               $merged = $null
               $allVerdicts = [System.Collections.Generic.List[object]]::new()
               foreach ($item in $group) {
@@ -1854,12 +1873,16 @@ jobs:
               $merged.verdicts = $allVerdicts.ToArray()
               $safeModel = ("$($merged.model)" -replace '[^A-Za-z0-9._-]', '_')
               if (-not $safeModel) { $safeModel = 'default' }
+              # Include the judge in the filename so two judges of one executor write
+              # to distinct synthetic files instead of overwriting each other.
+              $safeJudge = ("$($merged.judgeModel)" -replace '[^A-Za-z0-9._-]', '_')
+              if (-not $safeJudge) { $safeJudge = 'nojudge' }
               $mergedDir = "all-results/_merged"
               New-Item -ItemType Directory -Force -Path $mergedDir | Out-Null
-              $resultsFile = Join-Path $mergedDir "$plugin--$safeModel.results.json"
+              $resultsFile = Join-Path $mergedDir "$plugin--$safeModel--$safeJudge.results.json"
               $merged | ConvertTo-Json -Depth 100 | Out-File -FilePath $resultsFile -Encoding utf8
 
-              Write-Host "`n=== Generating benchmark data for: $plugin (model '$($merged.model)', $($allVerdicts.Count) verdicts from $($group.Count) results.json) ==="
+              Write-Host "`n=== Generating benchmark data for: $plugin (model '$($merged.model)', judge '$($merged.judgeModel)', $($allVerdicts.Count) verdicts from $($group.Count) results.json) ==="
               $params = @{
                 ResultsFile    = $resultsFile
                 PluginName     = $plugin
@@ -1869,6 +1892,15 @@ jobs:
                 Source         = 'scheduled'
                 SkipTokenUsage = $true
               }
+              # Emit Quality/Efficiency once per executor model; SkillValue every
+              # (model, judge). The first judge of a model writes all three; later
+              # judges of the same model write SkillValue only.
+              $execModel = "$($merged.model)"
+              if ($qeEmittedModels.ContainsKey($execModel)) {
+                $params.SkillValueOnly = $true
+              } else {
+                $qeEmittedModels[$execModel] = $true
+              }
               # Accumulate: each model's entry appends to $plugin.json. The first
               # iteration reads the fetched history; later iterations read the file
               # the previous iteration just wrote (same path as OutputDir/$plugin.json).
@@ -2156,6 +2188,7 @@ jobs:
           cp ${{ github.workspace }}/eng/dashboard/dashboard.html index.html
           cp ${{ github.workspace }}/eng/dashboard/dashboard.js dashboard.js
           cp ${{ github.workspace }}/eng/dashboard/token-usage.js token-usage.js
+          cp ${{ github.workspace }}/eng/dashboard/skill-value.js skill-value.js
 
           # Deploy AGENTVIZ SPA (skip if already present and unchanged)
           if [ "${{ steps.check-replay.outputs.skip }}" != "true" ]; then

eng/dashboard/dashboard.html

@@ -222,6 +222,28 @@
     @media (max-width: 600px) {
       .charts-grid { grid-template-columns: 1fr; }
     }
+    /* Skill Value view */
+    .sv-controls { display: flex; flex-wrap: wrap; gap: 16px; align-items: center; margin-bottom: 16px; }
+    .sv-controls label { font-size: 13px; color: var(--text-muted); }
+    .sv-controls select {
+      background: var(--surface); color: var(--text); border: 1px solid var(--border);
+      border-radius: 6px; padding: 6px 8px; font-size: 13px; margin-left: 4px;
+    }
+    .sv-sub { display: block; color: var(--text-muted); font-size: 11px; font-weight: 400; margin-top: 2px; }
+    .sv-controls .sv-sub { display: inline; margin: 0; }
+    .sv-table td { vertical-align: top; }
+    .sv-value { font-size: 13px; max-width: 360px; }
+    .sv-insufficient { color: var(--text-muted); font-style: italic; }
+    .sv-drill { padding: 4px 0; }
+    .sv-arm { font-size: 13px; padding: 3px 0; font-family: 'SF Mono', SFMono-Regular, Consolas, monospace; }
+    .sv-arm-label { display: inline-block; min-width: 110px; color: var(--text-muted); font-family: inherit; }
+    .sv-diluted { opacity: 0.6; }
+    .sv-dilute-mark { color: var(--text-muted); margin-right: 3px; cursor: help; font-weight: 600; }
+    .sv-btn { background: var(--surface); color: var(--text); border: 1px solid var(--border); border-radius: 6px; padding: 6px 10px; font-size: 12px; cursor: pointer; }
+    .sv-btn:hover { border-color: var(--skilled); color: var(--skilled); }
+    /* Model leaf rows carry the real numbers, so undo the token-table level-2 muting. */
+    .sv-table tr.level-2 td { color: var(--text); }
+    .sv-table tr.sv-detail td { padding-left: 88px; }
   </style>
 </head>
 <body>
@@ -235,7 +257,8 @@ <h1>📊 Skills Evaluation Dashboard</h1>
     Generated by the skills evaluation pipeline · Verdict evidence shows the latest retained run per model; 0–10 score averages are triage only, not the pass gate.
   </div>
 
-  <script src="dashboard.js"></script>
   <script src="token-usage.js"></script>
+  <script src="skill-value.js"></script>
+  <script src="dashboard.js"></script>
 </body>
 </html>

eng/dashboard/dashboard.js

@@ -34,44 +34,63 @@
   const loadedPlugins = new Map(); // track loaded plugin data
 
   // Build tabs and placeholder panels
-  plugins.forEach((plugin, idx) => {
+  plugins.forEach((plugin) => {
     const tab = document.createElement('div');
-    tab.className = 'tab' + (idx === 0 ? ' active' : '');
+    tab.className = 'tab';
     tab.textContent = plugin;
     tab.dataset.plugin = plugin;
     tab.addEventListener('click', () => switchTab(plugin));
     tabBar.appendChild(tab);
 
     const panel = document.createElement('div');
-    panel.className = 'tab-content' + (idx === 0 ? ' active' : '');
+    panel.className = 'tab-content';
     panel.id = `panel-${plugin}`;
     panel.innerHTML = '<p style="color:#8b949e;text-align:center;padding:2rem;">Loading...</p>';
     tabContentContainer.appendChild(panel);
   });
 
   // Add Token Usage tab at the end
   const tokenTabId = '__token-usage__';
-  const noPlugins = plugins.length === 0;
   const tokenTab = document.createElement('div');
-  tokenTab.className = 'tab' + (noPlugins ? ' active' : '');
+  tokenTab.className = 'tab';
   tokenTab.textContent = '🔢 Token Usage';
   tokenTab.dataset.plugin = tokenTabId;
   tokenTab.addEventListener('click', () => switchTab(tokenTabId));
   tabBar.appendChild(tokenTab);
 
   const tokenPanel = document.createElement('div');
-  tokenPanel.className = 'tab-content' + (noPlugins ? ' active' : '');
+  tokenPanel.className = 'tab-content';
   tokenPanel.id = `panel-${tokenTabId}`;
   tokenPanel.innerHTML = '<div id="token-usage-content"><p style="color:#8b949e;text-align:center;padding:2rem;">Loading…</p></div>';
   tabContentContainer.appendChild(tokenPanel);
 
+  // Skill Value is the default landing tab, placed FIRST in the tab bar so the
+  // per-skill value story is the first thing a viewer sees.
+  const skillValueTabId = '__skill-value__';
+  const skillValueTab = document.createElement('div');
+  skillValueTab.className = 'tab active';
+  skillValueTab.textContent = '💡 Skill Value';
+  skillValueTab.dataset.plugin = skillValueTabId;
+  skillValueTab.addEventListener('click', () => switchTab(skillValueTabId));
+  tabBar.insertBefore(skillValueTab, tabBar.firstChild);
+
+  const skillValuePanel = document.createElement('div');
+  skillValuePanel.className = 'tab-content active';
+  skillValuePanel.id = `panel-${skillValueTabId}`;
+  skillValuePanel.innerHTML = '<div id="skill-value-content"><p style="color:#8b949e;text-align:center;padding:2rem;">Loading…</p></div>';
+  tabContentContainer.appendChild(skillValuePanel);
+
   async function switchTab(plugin) {
     tabBar.querySelectorAll('.tab').forEach(t => t.classList.toggle('active', t.dataset.plugin === plugin));
     tabContentContainer.querySelectorAll('.tab-content').forEach(p => p.classList.toggle('active', p.id === `panel-${plugin}`));
     if (plugin === tokenTabId) {
       if (window.initTokenUsage) window.initTokenUsage();
       return;
     }
+    if (plugin === skillValueTabId) {
+      if (window.initSkillValue) window.initSkillValue();
+      return;
+    }
     if (!loadedPlugins.has(plugin)) {
       await loadPlugin(plugin);
     }
@@ -1126,8 +1145,12 @@
     ]);
   }
 
-  // Load first plugin immediately (skip if no evaluation plugins)
-  if (plugins.length > 0) {
-    await loadPlugin(plugins[0]);
+  // Skill Value is the default active tab, so render it immediately. Plugin tabs
+  // load lazily on first click; Token Usage self-inits when its tab is shown.
+  if (window.initSkillValue) {
+    window.initSkillValue();
+  } else if (plugins.length > 0) {
+    // Defensive fallback: if skill-value.js failed to load, activate the first plugin.
+    await switchTab(plugins[0]);
   }
 })();