@giamme/forge 0.10.0 → 0.11.0

package/ui/forge_ui/static/app.js

@@ -19,7 +19,7 @@ window.ForgeApp = (() => {
     replay: null, playing: null, speed: store.get('speed', 60), stage: null, stageFailed: null, stageLoading: false, timer: null,
     lastStatuses: new Map(), announced: 0, rendered: null, cloud: null, cloudUI: null, cloudPrefs: store.get('cloud', {}),
     settings: false, lastKey: null, settingsRun: null, settingsPickedAt: null, settingsView: null,
-    models: false, modelsData: null, modelsText: null, modelsError: null, modelsPrefs: store.get('models', {}),
+    models: false, modelsData: null, modelsText: null, modelsError: null, reviewerOutcomes: null, modelsPrefs: store.get('models', {}),
     files: new Map(), findingsCount: new Map(), filesPrefs: store.get('files', {}), runSort: store.get('run-sort', 'updated'), tablePrefs: store.get('tables', {}),
     findings: new Map() };
   state.stageMode = state.mode === 'exec' ? 'exec' : 'plan';   // the stage's own mode, kept while Files is shown
@@ -106,9 +106,9 @@ window.ForgeApp = (() => {
   function link(text, key, sel) { const a = el('a', text); a.href = hashFor(key, sel); return a; }
 
   const statuses = new Set(['merged', 'pass', 'running', 'pending', 'fail', 'error', 'infra',
-    'timeout', 'unknown', 'invalidated', 'blocked', 'interrupted', 'nochanges', 'conflict']);
-  function chip(value) {
-    const status = String(value || 'unknown').toLowerCase();
+    'timeout', 'unknown', 'invalidated', 'blocked', 'interrupted', 'nochanges', 'conflict', 'accepted', 'outdated']);
+  function chip(value, cls) {
+    const status = String(cls || value || 'unknown').toLowerCase();
     return el('span', value || '—', 'chip st-' + (statuses.has(status) ? status : 'unknown'));
   }
   function time(value) {
@@ -251,7 +251,8 @@ window.ForgeApp = (() => {
         if (!solo) {
           const verification = el('span', undefined, 'verification');
           verification.append(document.createTextNode('Verification '), chip(run.verification_status));
-          if (run.final_review) verification.append(document.createTextNode(' Review '), chip(run.final_review));
+          const reviewStatus = RM.runReviewStatus(run);
+          if (reviewStatus) verification.append(document.createTextNode(' Review '), chip(reviewStatus));
           item.append(verification);
         }
         $('run-list').append(item);
@@ -338,7 +339,7 @@ window.ForgeApp = (() => {
     add('Kind', sel('kind', [['', 'any'], ['plan', 'plan'], ['solo', 'solo']]));
     add('Model', sel('model', [['', 'any'], ...opts.models.map(m => [m, m])]));
     add('Age', sel('ageDays', [[0, 'any'], [1, 'last day'], [7, 'last week'], [30, 'last month']]));
-    add('Review', sel('verdict', [['', 'any'], ['PASS', 'passed'], ['FAIL', 'failed'], ['none', 'no verdict']]));
+    add('Review', sel('verdict', [['', 'any'], ['PASS', 'passed'], ['FAIL', 'failed'], ['OUTDATED', 'outdated'], ['none', 'no verdict']]));
     const all = el('input'); all.type = 'checkbox'; all.checked = state.showAll; all.dataset.f = 'showAll';
     const allL = el('label', undefined, 'filter check'); allL.append(all, document.createTextNode(' older runs too')); bar.append(allL);
     const reset = el('button', 'Reset'); reset.type = 'button'; reset.addEventListener('click', () => { Object.assign(state.mapFilter, { text: '', state: '', repo: '', kind: '', model: '', ageDays: 0, verdict: '' }); state.showAll = false; filterChanged(true); });
@@ -368,7 +369,12 @@ window.ForgeApp = (() => {
         card.className = `map-card st-${({ live: 'running', failed: 'fail', done: 'merged', pending: 'pending' })[m.state]}${m.match ? '' : ' dim'}`;
         card.append(el('strong', String(r.goal || 'Untitled run').split('\n')[0]), el('span', `${r.run_id || ''}${r.kind === 'solo' ? ' · solo' : ''}`, 'mono muted'));
         const leds = el('span', undefined, 'leds');
-        for (const t of m.tasks.slice(0, 60)) { const l = el('i', undefined, 'led st-' + String(t.status || 'pending').toLowerCase()); l.title = `${t.id}: ${t.status}`; leds.append(l); }
+        for (const t of m.tasks.slice(0, 60)) {
+          const status = RM.taskStatus(t);
+          const l = el('i', undefined, 'led st-' + String(status || 'pending').toLowerCase());
+          l.title = `${t.id}: ${status}`;
+          leds.append(l);
+        }
         card.append(leds, el('span', relative(r.updated), 'muted'));
         grid.append(card);
       }
@@ -474,15 +480,23 @@ window.ForgeApp = (() => {
       runs: state.runs, prefs: state.modelsPrefs, now: () => Date.now() / 1000,
       savePrefs: p => { state.modelsPrefs = p; store.set('models', p); },
       tablePrefs: id => state.tablePrefs[id], saveTablePrefs: (id, p) => { state.tablePrefs = { ...state.tablePrefs, [id]: p }; store.set('tables', state.tablePrefs); },
-      modelHref: alias => '#/cloud/models/' + encodeURIComponent(alias),
+      modelHref: alias => '#/cloud/models/' + encodeURIComponent(alias), reviewerOutcomes: state.reviewerOutcomes,
     });
   }
+  // How each reviewer's PASSed work held up (forge dataset reviewers). Advisory and optional: read once per
+  // change of the models data, and a failure (an export, no dataset) only leaves the column empty.
+  function loadReviewerOutcomes() {
+    if (SNAP) return;
+    fetchJSON('/api/models/reviewers').then(result => { state.reviewerOutcomes = result.data; },
+      () => { state.reviewerOutcomes = null; }).then(() => { if (state.models) renderModels(); });
+  }
   function loadModels() {
     return request('/api/models', result => {
       state.modelsError = null;
       if (result.text === state.modelsText) return;
       state.modelsText = result.text; state.modelsData = result.data;
       if (state.models) renderModels();
+      loadReviewerOutcomes();
     }, 'Could not refresh the model results', error => { state.modelsError = error.message; if (state.models) renderModels(); });
   }
 
@@ -741,11 +755,11 @@ window.ForgeApp = (() => {
       { id: 'title', label: 'Title', value: t => t.title },
       { id: 'difficulty', label: 'Difficulty', value: t => DIFF_RANK[t.difficulty] ?? null, cell: t => display(t.difficulty) },
       { id: 'wave', label: 'Wave', value: t => Number.isFinite(Number(t.wave)) && t.wave !== null && t.wave !== '' ? Number(t.wave) : null, cell: t => display(t.wave) },
-      { id: 'status', label: 'Status', value: t => t.status, cell: t => chip(t.status) },
+      { id: 'status', label: 'Status', value: t => RM.taskStatus(t), cell: t => chip(RM.taskStatus(t)) },
       { id: 'dwarf', label: 'Dwarf', value: t => t.dwarf, cell: t => window.ForgeIcons.modelPlate(t.dwarf) },
       { id: 'qa', label: 'QA', value: t => t.qa, cell: t => window.ForgeIcons.modelPlate(t.qa) },
       { id: 'attempt', label: 'Attempt', value: t => Number(t.attempt) || null, cell: t => display(t.attempt) },
-      { id: 'verdict', label: 'Verdict', value: t => t.verdict, cell: t => chip(t.verdict) },
+      { id: 'verdict', label: 'Verdict', value: t => RM.taskVerdict(t).text, cell: t => { const v = RM.taskVerdict(t); return chip(v.text, v.status); } },
       { id: 'findings', label: 'Findings', title: 'The worst confirmed severity and how many; ~n plausible only',
         value: t => t.findings && t.findings.worst ? SEV_RANK[t.findings.worst] * 1000 - (t.findings[t.findings.worst] || 0) : t.findings && t.findings.plausible ? 4000 : null,
         cell: t => { const f = t.findings, v = sev(f); if (!f || !f.worst) return v; const n = el('span', v, 'sev-chip sev-' + f.worst); return n; } },

package/ui/forge_ui/static/cloud3d.js

@@ -427,9 +427,12 @@ export function createCloud(E, host, cb = {}) {
   function info(el) {
     const f = el.f, r = f.run, counts = Object.entries(r.counts || {}).map(([s, n]) => `${s.toLowerCase()} ${n}`).join(' · ');
     const proj = st.L && st.L.projects.find(p => p.id === f.project);
+    const verdict = r.final_review && typeof r.final_review === 'object' ? r.final_review.verdict : r.final_review;
+    const isOutdated = verdict && RM.runReviewStatus(r) === 'OUTDATED';
+    const reviewText = verdict ? (isOutdated ? `${verdict} (outdated)` : verdict) : null;
     return [['meta', `${proj ? proj.name : RM.basename(r.repo)} · ${r.run_id || ''} · ${r.kind}`], ['title', String(r.goal || 'Untitled run').split('\n')[0]],
       ['body', `${f.n} task${f.n > 1 ? 's' : ''} · ${counts}`], ['body', `${[...f.dwarf].join(', ') || '—'} → ${[...f.qa].join(', ') || '—'}`],
-      ['foot', `started ${new Date(f.t * 1000).toLocaleString()}${r.final_review ? ' · review ' + r.final_review : ''}${r.verification_status ? ' · verify ' + r.verification_status : ''}`],
+      ['foot', `started ${new Date(f.t * 1000).toLocaleString()}${reviewText ? ' · review ' + reviewText : ''}${r.verification_status ? ' · verify ' + r.verification_status : ''}`],
       ['foot', (store.layouts[st.layout] || {})[el.key] ? 'moved by you · drag to move, Shift = depth, Alt = height' : 'drag to move · double-click to open']];
   }
   function showTip(el, cx, cy) {

package/ui/forge_ui/static/cloudmodel.js

@@ -63,7 +63,7 @@
     const aliases = specs => new Set(list(specs).map(RM.alias).filter(Boolean));
     return { key: run.key, run, project: hashId(run.repo || '?'), n, bucket: bucket(n), diff, dominant, state, t: startOf(run),
       dwarf: aliases(run.dwarf_models || run.models), qa: aliases(run.qa_models || []), all: aliases(run.models),
-      review: run.final_review || null, verification: run.verification_status || null, decompose: run.decompose_level || null, kind: run.kind };
+      review: RM.runReviewStatus(run) || null, verification: run.verification_status || null, decompose: run.decompose_level || null, kind: run.kind };
   }
 
   /* ---------- filters: the cloud shows exactly the runs that match ---------- */

package/ui/forge_ui/static/cloudui.js

@@ -254,7 +254,7 @@ window.ForgeCloudUI = (() => {
       }
       const more = el('fieldset', undefined, 'cloud-field cloud-more'); more.append(el('legend', 'More'));
       more.append(select('Active', 'active', [[0, 'any time'], [1, 'in the last day'], [7, 'in the last week'], [30, 'in the last month'], [90, 'in the last 3 months']], () => f.activeDays, v => { f.activeDays = Number(v) || 0; }));
-      more.append(select('Final review', 'review', [['', 'any'], ['PASS', 'passed'], ['FAIL', 'failed'], ['none', 'no verdict']], () => f.review, v => { f.review = v; }));
+      more.append(select('Final review', 'review', [['', 'any'], ['PASS', 'passed'], ['FAIL', 'failed'], ['OUTDATED', 'outdated'], ['none', 'no verdict']], () => f.review, v => { f.review = v; }));
       more.append(select('Verification', 'verification', [['', 'any'], ['PASS', 'passed'], ['FAIL', 'failed'], ['none', 'not run']], () => f.verification, v => { f.verification = v; }));
       const dec = Object.keys(O.decompose);
       if (dec.length > 1) more.append(chips('Decompose', dec.map(k => [k, k === 'none' ? 'off' : k, O.decompose[k]]), () => f.decompose, v => { f.decompose = v; }));
@@ -290,13 +290,22 @@ window.ForgeCloudUI = (() => {
       head.append(shut);
       card.append(head, el('h2', String(r.goal || 'Untitled run').split('\n')[0]), el('p', r.run_id || '', 'mono muted'));
       const leds = el('span', undefined, 'leds');
-      for (const t of list(r.task_statuses).slice(0, 120)) { const l = el('i', undefined, 'led st-' + String(t.status || 'pending').toLowerCase()); l.title = `${t.id}: ${t.status}${t.title ? ' · ' + t.title : ''}`; leds.append(l); }
+      for (const t of list(r.task_statuses).slice(0, 120)) {
+        const st = RM.taskStatus(t);
+        const l = el('i', undefined, 'led st-' + String(st || 'pending').toLowerCase());
+        l.title = `${t.id}: ${st}${t.title ? ' · ' + t.title : ''}`;
+        leds.append(l);
+      }
       card.append(leds);
       const dl = el('dl', undefined, 'fields');
       const mix = ['high', 'medium', 'low', 'none'].filter(k => f.diff[k]).map(k => `${f.diff[k]} ${k === 'none' ? 'unset' : k}`).join(' · ');
       const plates = specs => { if (!specs.length || !window.ForgeIcons) return specs.join(', ') || '—'; const s = el('span', undefined, 'plates'); for (const x of specs) s.append(window.ForgeIcons.modelPlate(x)); return s; };
+      const verdict = r.final_review && typeof r.final_review === 'object' ? r.final_review.verdict : r.final_review;
+      const isOutdated = verdict && RM.runReviewStatus(r) === 'OUTDATED';
+      const reviewText = verdict ? (isOutdated ? `${verdict} (outdated)` : verdict) : '—';
+      const reviewVal = isOutdated ? el('span', reviewText, 'chip st-outdated') : reviewText;
       for (const [k, v, mono] of [['Tasks', `${f.n}${mix ? ' · ' + mix : ''}`], ['Implementers', plates([...f.dwarf])], ['Reviewers', plates([...f.qa])],
-        ['Started', when(r.created)], ['Updated', when(r.updated)], ['Final review', r.final_review || '—'], ['Verification', r.verification_status || '—'], ['Decompose', r.decompose_level || 'off']]) {
+        ['Started', when(r.created)], ['Updated', when(r.updated)], ['Final review', reviewVal], ['Verification', r.verification_status || '—'], ['Decompose', r.decompose_level || 'off']]) {
         const dd = el('dd', undefined, mono ? 'mono' : ''); dd.append(v instanceof Node ? v : document.createTextNode(String(v))); dl.append(el('dt', k), dd);
       }
       card.append(dl);
@@ -345,7 +354,7 @@ window.ForgeCloudUI = (() => {
           a.href = '#/runs/' + encodeURIComponent(r.key);
           a.append(el('strong', String(r.goal || 'Untitled run').split('\n')[0]), el('span', `${r.run_id || ''} · ${f.n} task${f.n === 1 ? '' : 's'}${r.kind === 'solo' ? ' · solo' : ''}`, 'mono muted'));
           const leds = el('span', undefined, 'leds');
-          for (const t of list(r.task_statuses).slice(0, 60)) leds.append(el('i', undefined, 'led st-' + String(t.status || 'pending').toLowerCase()));
+          for (const t of list(r.task_statuses).slice(0, 60)) leds.append(el('i', undefined, 'led st-' + String(RM.taskStatus(t) || 'pending').toLowerCase()));
           a.append(leds); grid.append(a);
         }
         g.append(grid); listHost.append(g);

package/ui/forge_ui/static/findingsview.js

@@ -16,7 +16,7 @@
   const MAX_FPS = 25, MAX_LABELS = 5, MAX_LABEL = 50, MAX_REASON = 500;
   const list = v => Array.isArray(v) ? v : [];
   const SEVERITIES = ['P0', 'P1', 'P2', 'P3'];
-  const ORIGINS = ['tolerated', 'review', 'plausible'];
+  const ORIGINS = ['tolerated', 'accepted', 'review', 'plausible'];
   const STATUSES = ['open', 'filed', 'fixed', 'dismissed'];
 
   /* ---------- pure ---------- */
@@ -106,6 +106,20 @@
   }
   // An untrusted URL becomes a link only when it is a plain https address.
   const safeUrl = u => typeof u === 'string' && /^https:\/\/[^\s"'<>]+$/.test(u) ? u : null;
+  // The marker for a finding whose whole-run review is outdated, or null. The lamp is the hollow unlit
+  // `.st-outdated`. `title` is the same sentence the issue body adds, ids cut after 5, so a row can show it
+  // without building HTML. A current review, and any finding that is not from a whole-run review, has none.
+  function reviewMarker(row) {
+    const review = row && row.review;
+    if (!review || review.outdated !== true) return null;
+    const ids = list(review.merged_since).map(id => String(id));
+    const shown = ids.slice(0, 5), extra = ids.length - shown.length;
+    const names = shown.join(', ') + (extra > 0 ? ', +' + extra + ' more' : '');
+    const n = ids.length;
+    const sentence = `The review is outdated: ${n} ${n === 1 ? 'task' : 'tasks'} merged into the integration branch after it`
+      + (names ? ` (${names})` : '') + ', so check whether they fixed it first.';
+    return { lamp: 'st-outdated', label: 'outdated review', title: sentence, sentence };
+  }
 
   /* ---------- DOM ---------- */
   function render(host, ctx) {
@@ -198,6 +212,16 @@
     }
 
     // cells
+    const originCell = r => {
+      const marker = reviewMarker(r);
+      if (!marker) return r.origin || '—';
+      const box = el('span', undefined, 'fnd-status');
+      const lamp = el('i', undefined, 'lamp ' + marker.lamp);
+      lamp.title = marker.title;
+      box.title = marker.title;
+      box.append(lamp, el('span', r.origin || '—'), el('span', marker.label));
+      return box;
+    };
     const statusCell = r => {
       const st = statusOf(r), box = el('span', undefined, 'fnd-status fnd-' + st);
       box.append(el('i', undefined, 'lamp ' + (st === 'filed' ? 'fnd-amber' : st === 'fixed' ? 'st-pass' : 'lamp-off')));
@@ -223,6 +247,8 @@
       more.append(el('span', '▸', 'fnd-caret'));
       const detail = el('div', undefined, 'fnd-detail'); detail.hidden = true;
       detail.append(el('pre', r.text || r.summary || '', 'fnd-text'));
+      const marker = reviewMarker(r);
+      if (marker) detail.append(el('p', marker.sentence, 'muted'));
       const meta = el('dl', undefined, 'fnd-meta');
       for (const [k, v] of [['Reviewer', r.reviewer], ['Commit', r.commit], ['Fingerprint', r.fp], ['Plan', r.plan_dir || r.plan],
         ['Dismissed', r.dismissed ? `${r.dismissed.at || ''} ${r.dismissed.reason || ''}`.trim() : null]]) {
@@ -245,7 +271,7 @@
       cols.push(
         { id: 'severity', label: 'Severity', value: r => SEVERITIES.indexOf(r.severity), cell: sevCell },
         { id: 'confidence', label: 'Confidence', value: r => r.confidence, cell: r => r.confidence || '—' },
-        { id: 'origin', label: 'Origin', value: r => r.origin, cell: r => r.origin || '—' },
+        { id: 'origin', label: 'Origin', value: r => r.origin, cell: originCell },
         { id: 'task', label: 'Task', value: r => r.task, mono: true, cell: r => r.task || 'whole run' });
       if (s.scope === 'repo') cols.push({ id: 'plan', label: 'Plan', value: r => r.plan, mono: true, cell: r => r.plan || '—' });
       cols.push(
@@ -408,5 +434,5 @@
   }
 
   return { render, statusOf, filterRows, mergeFiled, selectable, selection, issuesUrl, issuesPath, dismissPath, fileBody, dismissBody,
-    parseLabels, resultLines, safeUrl, MAX_FPS };
+    parseLabels, resultLines, safeUrl, reviewMarker, ORIGINS, MAX_FPS };
 });

package/ui/forge_ui/static/graph.js

@@ -201,8 +201,9 @@
       if (drawn[t.id]) return;
       drawn[t.id] = true;
       var p = pos[t.id];
-      var status = str(t.status, 'pending');
-      var stClass = 'st-' + status.toLowerCase().replace(/[^a-z0-9_-]/g, '');
+      var status = window.ForgeRunModel.taskStatus(t);
+      var verdict = window.ForgeRunModel.taskVerdict(t);
+      var stClass = 'st-' + String(status || 'pending').toLowerCase().replace(/[^a-z0-9_-]/g, '');
       var selected = options.selected === t.id;
       var id = String(t.id);
 
@@ -217,7 +218,7 @@
 
       var tip = [str(t.title, id), 'status: ' + status,
         'dwarf: ' + str(t.dwarf, '—') + ' / qa: ' + str(t.qa, '—')];
-      if (t.verdict) tip.push('verdict: ' + t.verdict);
+      if (verdict.text) tip.push('verdict: ' + verdict.text);
       if (t.explored) tip.push('explored');
       var title = el('title', null, g);
       title.textContent = tip.join('\n');
@@ -238,8 +239,8 @@
         x: 12, y: 36, 'font-size': 12, fill: 'currentColor', 'fill-opacity': 0.75
       });
 
-      if (t.verdict === 'PASS' || t.verdict === 'FAIL') {
-        var pass = t.verdict === 'PASS';
+      if (verdict.mark === 'PASS' || verdict.mark === 'FAIL') {
+        var pass = verdict.mark === 'PASS';
         // a drawn tick or cross, in the verdict's colour
         el('path', { 'class': 'fg-verdict' + (pass ? ' fg-pass' : ''), d: pass ? 'M' + (NODE_W - 20) + ' 12l4 4 7-8' : 'M' + (NODE_W - 19) + ' 8l8 8m0-8l-8 8',
           fill: 'none', 'stroke-width': 1.8, 'stroke-linecap': 'round', 'stroke-linejoin': 'round' }, g);

package/ui/forge_ui/static/leaderboard.js

@@ -149,6 +149,7 @@ window.ForgeLeaderboard = (() => {
     ['Verdicts', '', 'PASS and FAIL verdicts over every review'],
     ...MS.SEVERITIES.map(s => [s, 'lb-num', SEV_TITLE(s)]),
     ['Shadow agreement', 'lb-num', 'How often this model\'s shadow verdict matched the primary reviewer\'s (both PASS or FAIL)'],
+    ['Did not hold', 'lb-num', 'Of the work this model PASSed that was integrated, the share later reverted or with under 80% of its lines kept (forge dataset reviewers, every run in the dataset: the filters do not apply). Advisory: routing never reads it'],
     ['Median time', 'lb-num', 'Median duration of a review'],
     ['Tokens in / out', 'lb-num', 'Summed over the reviews that recorded tokens'],
   ];
@@ -182,7 +183,23 @@ window.ForgeLeaderboard = (() => {
     td.title = `${a.k} of ${a.n} shadow verdicts matched the primary's; 95% Wilson interval ${pct(a.lo)}–${pct(a.hi)}${a.unknown ? ` · ${a.unknown} without two comparable verdicts` : ''}`;
     return td;
   }
+  function heldCell(h) {
+    const days = h.horizon ? ` ${h.horizon} days after integrate` : '';
+    if (h.state === 'unavailable') return numCell('—', 'The run dataset could not be read (forge dataset reviewers)');
+    if (h.state === 'none') return numCell('—', `No work this model PASSed has a survival outcome${days} yet (forge dataset outcomes records them)`);
+    const detail = `${h.reverted} reverted, ${h.rewritten} rewritten, of ${h.n} PASSed and integrated${days}`;
+    if (h.state === 'few') {
+      const td = numCell('not enough data', `${detail}; a rate needs ${h.need || 'more'} outcomes`);
+      td.classList.add('muted');
+      return td;
+    }
+    const td = el('td', undefined, 'lb-num');
+    td.append(el('strong', pct(h.p)), document.createTextNode(' '), el('span', `${h.k}/${h.n}`, 'muted'));
+    td.title = `${detail}; 95% Wilson interval ${pct(h.lo)}–${pct(h.hi)}`;
+    return td;
+  }
   function reviewerTable(rows, ctx) {
+    const held = r => MS.heldUp(ctx.reviewerOutcomes, r.alias);
     const columns = colsOf(REV_COLS, [
       { id: 'model', value: r => r.alias, cell: r => modelCell(r.alias, ctx, '') },
       { id: 'primary', value: r => r.reviews.primary, cell: r => numCell(r.reviews.primary) },
@@ -191,6 +208,7 @@ window.ForgeLeaderboard = (() => {
       { id: 'verdicts', value: r => { const v = r.verdicts, t = v.PASS + v.FAIL + v.other; return t ? v.PASS / t : null; }, cell: r => verdictCell(r.verdicts) },
       ...MS.SEVERITIES.map(sv => ({ id: 'sev-' + sv, value: r => r.findings[sv], cell: r => findingCell(r, sv) })),
       { id: 'agreement', value: r => r.agreement.n ? r.agreement.p : null, cell: r => agreementCell(r.agreement) },
+      { id: 'held', value: r => { const h = held(r); return h.state === 'rate' ? h.p : null; }, cell: r => heldCell(held(r)) },
       { id: 'time', value: r => num(r.medianDurationS), cell: r => numCell(dur(r.medianDurationS)) },
       { id: 'tokens', value: r => (r.tokens.attempts ?? r.tokens.reviews) ? (r.tokens.input || 0) + (r.tokens.output || 0) : null, cell: r => tokenCell(r.tokens, r.reviews.total, 'review') },
     ]);

package/ui/forge_ui/static/modelstats.js

@@ -245,5 +245,23 @@
     return { implementers: imp.rows, reviewers: rev.rows, byDifficulty, totals, caveats, filter: f };
   }
 
-  return { DIFFS, KINDS, SEVERITIES, DEFAULT_MIN_N, wilson, emptyFilter, normFilter, taskMatch, options, leaderboard };
+  // How the work a reviewer PASSed held up, from GET /api/models/reviewers (`forge dataset reviewers --json`):
+  // { state: 'unavailable' } without a report, 'none' when the alias has no landed outcome, 'few' below the
+  // report's minimum (no rate is shown then), else 'rate' with the share that did not hold and its interval.
+  function heldUp(report, alias) {
+    const r = obj(report);
+    if (!r || !Array.isArray(r.reviewers)) return { state: 'unavailable' };
+    const horizon = num(r.horizon_days), row = r.reviewers.find(x => obj(x) && x.reviewer === alias);
+    if (!row) return { state: 'none', horizon };
+    const n = num(row.landed), k = num(row.did_not_hold);
+    if (n === null || k === null) return { state: 'unavailable' };
+    const base = { n, k, reverted: num(row.reverted) || 0, rewritten: num(row.rewritten) || 0, horizon };
+    const ci = list(row.ci);
+    if (row.enough !== true || num(row.rate) === null || num(ci[0]) === null || num(ci[1]) === null) {
+      return { state: 'few', need: num(r.min_landed), ...base };
+    }
+    return { state: 'rate', p: row.rate, lo: ci[0], hi: ci[1], ...base };
+  }
+
+  return { DIFFS, KINDS, SEVERITIES, DEFAULT_MIN_N, wilson, emptyFilter, normFilter, taskMatch, options, leaderboard, heldUp };
 });

package/ui/forge_ui/static/panel.js

@@ -9,7 +9,7 @@ window.ForgePanel = (() => {
   const num = v => typeof v === 'number' && Number.isFinite(v) ? v : null;
   const display = v => v === null || v === undefined || v === '' ? '—' : typeof v === 'object' ? JSON.stringify(v, null, 2) : String(v);
   const STATUSES = new Set(['merged', 'pass', 'running', 'pending', 'fail', 'error', 'infra', 'timeout', 'unknown', 'invalidated', 'blocked',
-    'interrupted', 'nochanges', 'conflict', 'accepted', 'none']);
+    'interrupted', 'nochanges', 'conflict', 'accepted', 'none', 'outdated']);
   const cache = new Map();                                  // path → {data, error, at}
   function el(tag, text, className) {
     const node = document.createElement(tag);
@@ -51,6 +51,7 @@ window.ForgePanel = (() => {
   function section(parent, title) { const s = el('section', undefined, 'insp-section'); s.append(el('h3', title)); parent.append(s); return s; }
   function details(parent, title, value, key, { open = false, cls = '' } = {}) {
     const d = el('details', undefined, cls); d.dataset.keep = key; if (open) d.open = true;
+    if (key && String(key).endsWith('known_issues')) d.id = key;
     const pre = el('pre', value); pre.dataset.scroll = key + '-text';
     d.append(el('summary', title), pre); parent.append(d); return d;
   }
@@ -60,6 +61,21 @@ window.ForgePanel = (() => {
     return a;
   }
   function muted(parent, text) { parent.append(el('p', text, 'muted')); }
+  // The acceptance chip is a key to the Known issues text in this panel, not another route.
+  function acceptedChip(task) {
+    const over = task.accepted_over === 'FAIL' || task.accepted_over === 'UNKNOWN' ? task.accepted_over : null;
+    const id = 'task-' + task.id + '-known_issues';
+    const a = el('a', over ? 'accepted over ' + over : 'accepted', 'chip st-accepted');
+    a.href = '#' + id;
+    a.addEventListener('click', event => {
+      event.preventDefault();
+      const node = document.getElementById(id);
+      if (!node) return;
+      node.open = true;
+      node.scrollIntoView({ block: 'nearest' });
+    });
+    return a;
+  }
   // who decided an escalation: policy, a human's retry --dwarf (retry) or a pin; older rows read as policy
   const sourcePlate = source => el('span', source || 'policy', 'plate src-plate src-' + String(source || 'policy').replace(/[^a-z]/g, ''));
   // the write controls of a plan run (task undefined) or of one of its tasks; a page without controls.js says it is read-only
@@ -218,11 +234,23 @@ window.ForgePanel = (() => {
     }
     const f = section(host, 'Finish');
     const F = RM.finishState(run);
-    if (run.kind === 'solo') fields(f, [['Verdict', chip(list(run.tasks)[0] && list(run.tasks)[0].verdict)]]);
+    if (run.kind === 'solo') {
+      const soloTask = list(run.tasks)[0];
+      const verdict = soloTask ? RM.taskVerdict(soloTask) : { text: null, status: 'unknown' };
+      fields(f, [['Verdict', chip(verdict.text, verdict.status)]]);
+    }
     else {
+      const finalSel = { kind: 'final' }, openFinal = () => ctx.onSelect(finalSel);
+      let finalValue;
+      if (F.final.status === 'OUTDATED') {
+        finalValue = el('a'); finalValue.href = ctx.hash(finalSel);
+        finalValue.addEventListener('click', e => { e.preventDefault(); openFinal(); });
+        finalValue.append(chip('OUTDATED'));
+      } else finalValue = link(F.final.status === 'NONE' ? 'not requested' : F.final.status, ctx.hash(finalSel), openFinal);
       fields(f, [['Integration', link(`${F.integration.merged}/${F.integration.total} merged`, ctx.hash({ kind: 'integration' }), () => ctx.onSelect({ kind: 'integration' }))],
         ['Verification', link(F.verification.status, ctx.hash({ kind: 'verification' }), () => ctx.onSelect({ kind: 'verification' }))],
-        ['Final review', link(F.final.status === 'NONE' ? 'not requested' : F.final.status, ctx.hash({ kind: 'final' }), () => ctx.onSelect({ kind: 'final' }))]]);
+        ['Final review', finalValue]]);
+      if (F.final.status === 'OUTDATED') muted(f, RM.reviewSinceLine(F.final));
       if (F.final.findings) findingCounts(f, F.final.findings);
     }
     if (list(run.deferrals).length) details(host, `Deferred tasks (${run.deferrals.length})`, run.deferrals.join('\n'), 'deferrals');
@@ -236,10 +264,13 @@ window.ForgePanel = (() => {
     const head = el('div', undefined, 'insp-head');
     head.append(el('h2', solo ? (task.title || 'Solo run') : `${task.id} · ${task.title || 'Untitled task'}`), el('div', solo ? 'solo task' : `task · wave ${display(task.wave)}`, 'insp-kind'));
     const chips = el('div', undefined, 'chips');
-    chips.append(chip(task.status));
-    if (task.verdict) chips.append(chip('review ' + task.verdict, task.verdict.toLowerCase()));
+    if (RM.taskStatus(task) === 'ACCEPTED') chips.append(acceptedChip(task));
+    else {
+      chips.append(chip(RM.taskStatus(task)));
+      if (task.accepted === true) chips.append(acceptedChip(task));
+      else if (task.verdict) chips.append(chip('review ' + task.verdict, task.verdict.toLowerCase()));
+    }
     if (task.explored) chips.append(el('span', '* explored', 'tag'));
-    if (task.accepted_by) chips.append(chip('accepted over FAIL', 'nochanges'));
     head.append(chips);
     const dl = diffLink(ctx, solo ? 'View the diff →' : 'View its diff, findings pinned to their lines →', { kind: 'task', task: task.id, source: 'current' }); if (dl) head.append(dl);
     host.append(head);
@@ -270,7 +301,7 @@ window.ForgePanel = (() => {
     for (const z of tp.zones) {
       const box = el('div', undefined, 'attempt');
       const row = list(task.attempts).filter(a => a.attempt === z.attempt).pop();
-      const h = el('div', undefined, 'attempt-head'); h.append(el('strong', z.name.toLowerCase())); if (row) h.append(chip(row.status), el('span', time(row.at), 'muted'));
+      const h = el('div', undefined, 'attempt-head'); h.append(el('strong', z.name.toLowerCase())); if (row) h.append(chip(RM.attemptStatus(task, row.attempt, row.status)), el('span', time(row.at), 'muted'));
       box.append(h);
       for (const pid of z.parts) {
         const p = tp.parts.find(x => x.id === pid); if (!p || p.planned) continue;
@@ -329,7 +360,8 @@ window.ForgePanel = (() => {
       const cs = section(host, 'Clef predictions');
       for (const [label, p] of preds) {
         const n = Number(String(label).match(/\d+/) || 1), outcome = list(task.attempts).filter(a => a.attempt === n).pop();
-        fields(cs, [[label, `p(pass) = ${display(p.p)} → ${outcome ? outcome.status : 'not finished'}`], ['Assignment', p.assignment ? plates([p.assignment.dwarf, p.assignment.qa]) : '—']]);
+        const happened = outcome ? RM.attemptStatus(task, outcome.attempt, outcome.status) : 'not finished';
+        fields(cs, [[label, `p(pass) = ${display(p.p)} → ${happened}`], ['Assignment', p.assignment ? plates([p.assignment.dwarf, p.assignment.qa]) : '—']]);
       }
     }
     // the prompts and the final messages
@@ -338,10 +370,18 @@ window.ForgePanel = (() => {
     else {
       const t = detail.data.texts || {};
       for (const [title, key] of [['Requirements (prompt.md)', 'prompt'], ['Intended approach', 'approach'], ['Retry findings', 'retry_findings'],
-        ['Dwarf prompt as assembled (dwarf.input)', 'dwarf_input'], ['Review prompt as assembled (qa.input)', 'qa_input'], ['Known issues', 'known_issues'], ['Accepted by', 'accepted_by']]) {
+        ['Dwarf prompt as assembled (dwarf.input)', 'dwarf_input'], ['Review prompt as assembled (qa.input)', 'qa_input']]) {
         if (t[key] != null) details(ps, title, t[key], `task-${task.id}-${key}`);
       }
     }
+    // Known issues is the text the acceptance chip opens. Accepted by keeps the reason, who, and when.
+    const texts = detail.data && detail.data.texts || null;
+    if (task.accepted === true || (texts && texts.known_issues != null)) {
+      const known = texts ? texts.known_issues : null;
+      details(ps, 'Known issues', known != null ? known : (detail.data ? '—' : 'Loading…'), `task-${task.id}-known_issues`);
+    }
+    const acceptedBy = texts && texts.accepted_by != null ? texts.accepted_by : task.accepted_by;
+    if (acceptedBy != null) details(ps, 'Accepted by', acceptedBy, `task-${task.id}-accepted_by`);
     if (task.dwarf_tail != null) details(ps, 'Dwarf final message', task.dwarf_tail, `task-${task.id}-dwarf_tail`);
     if (task.qa_tail != null) details(ps, 'Reviewer final message', task.qa_tail, `task-${task.id}-qa_tail`);
   }
@@ -475,10 +515,12 @@ window.ForgePanel = (() => {
       return;
     }
     const chips = el('div', undefined, 'chips');
-    chips.append(chip(pick.state === 'running' ? 'RUNNING' : pick.verdict || pick.state || 'unknown'));
+    const status = RM.finalStatus(pick);
+    chips.append(chip(status));
     if (pick.published) chips.append(el('span', 'published', 'tag'));
     chips.append(el('span', pick.source === 'integrate' ? 'during integrate' : 'forge review', 'tag'));
     head.append(chips);
+    if (status === 'OUTDATED') muted(host, RM.reviewSinceLine(pick));
     fields(host, [['Reviewer', plate(pick.spec)], ['Threshold', pick.threshold], ['Reviewed', pick.base || pick.tip ? `${String(pick.base || '?').slice(0, 10)}..${String(pick.tip || '?').slice(0, 10)}` : '—', true],
       ['Started', time(pick.started ?? pick.at)], ['Ended', time(pick.ended)], ['Duration', RM.fmtDur(pick.duration_s)]]);
     if (pick.execution) { const s = { kind: 'execution', task: null, exec: pick.execution }; host.append(link('Model, timings and exact prompt →', ctx.hash(s), () => ctx.onSelect(s))); }