feat(challenge): 反驳卡片加 Hermes 二次验证, 评估反驳是否成立

nieao · claude · nieao · commit 45bd0bdda548 · 2026-05-03T17:39:23.000+08:00
用户图 43 反馈: LLM 生成的反驳脱离实际 (PDF 不会超 1GB 但反驳照搬 OOM)。
ChallengeNode 待办区下方加"⚖ 二次验证"按钮:
  - 触发 challenge:verify-hermes 事件
  - KnowledgeGraph 监听后调 LLM 做 sanity check (system prompt 明确"避免
    LLM 编造不切实际论点")
  - 返回 { score, verdict, reason }, 写到节点 data.verification
  - UI 显示彩色可信度数字 + 一句话理由
  - 70+ 绿 / 40+ 暖 / &lt;40 红粉 — 视觉直接看出反驳质量

Co-Authored-By: Claude Opus 4.7 (1M context) &lt;noreply@anthropic.com&gt;
diff --git a/src/components/canvas/ChallengeNode.jsx b/src/components/canvas/ChallengeNode.jsx
@@ -29,6 +29,8 @@ function ChallengeNodeImpl({ id, data, selected }) {
   const evidence = Array.isArray(data.evidence) ? data.evidence : []
   const todos = Array.isArray(data.todos) ? data.todos : []
   const chainRunning = data.chainRunning === true
+  const verifyRunning = data.verifyRunning === true
+  const verification = data.verification || null  // { score, verdict, reason }
 
   const onChainTodos = (e) => {
     e.stopPropagation()
@@ -38,6 +40,14 @@ function ChallengeNodeImpl({ id, data, selected }) {
     }))
   }
 
+  const onVerify = (e) => {
+    e.stopPropagation()
+    if (verifyRunning) return
+    window.dispatchEvent(new CustomEvent('challenge:verify-hermes', {
+      detail: { challengeId: id, claim, angle, severity, sourceTitle },
+    }))
+  }
+
   return (
     <div
       className="relative shadow-sm transition-all duration-300"
@@ -152,6 +162,54 @@ function ChallengeNodeImpl({ id, data, selected }) {
           </div>
         )}
 
+        {/* === 二次验证: Hermes sanity check === */}
+        {/* 让 LLM 做"反思"判断这个反驳是否切合实际 (用户图 43 反馈反驳脱离实际, 比如说 PDF 超 1GB) */}
+        <div className="mt-2.5 pt-2" style={{ borderTop: `1px dashed ${meta.color}30` }}>
+          <div className="flex items-center justify-between gap-2">
+            <span
+              className="text-[9px]"
+              style={{ color: 'var(--text-muted)', letterSpacing: '0.15em' }}
+            >
+              {verification ? `验证 · ${verification.verdict || '已审'}` : '反驳是否成立?'}
+            </span>
+            <button
+              type="button"
+              onClick={onVerify}
+              disabled={verifyRunning}
+              className="text-[9px] px-2 py-0.5 rounded transition-all duration-300"
+              style={{
+                border: '1px solid var(--accent)',
+                color: verifyRunning ? 'var(--text-faint)' : 'var(--accent)',
+                background: verification ? 'var(--accent-bg, rgba(245,240,235,0.4))' : 'transparent',
+                cursor: verifyRunning ? 'wait' : 'pointer',
+                letterSpacing: '0.05em',
+              }}
+              title="让 LLM 二次审查这个反驳: 是否切合实际, 给可信度评分"
+            >
+              {verifyRunning ? '审中…' : verification ? '重审' : '⚖ 二次验证'}
+            </button>
+          </div>
+          {verification && (
+            <div className="mt-1.5 p-1.5 rounded" style={{
+              background: verification.score >= 70 ? 'rgba(123,196,127,0.08)' : verification.score >= 40 ? 'rgba(200,168,130,0.08)' : 'rgba(178,124,139,0.08)',
+              border: `1px solid ${verification.score >= 70 ? '#7bc47f55' : verification.score >= 40 ? '#c8a88255' : '#b27c8b55'}`,
+            }}>
+              <div className="flex items-center gap-1.5" style={{ marginBottom: 3 }}>
+                <span style={{
+                  fontSize: 12,
+                  fontFamily: '"Noto Serif SC", Georgia, serif',
+                  color: verification.score >= 70 ? '#3a8a3e' : verification.score >= 40 ? 'var(--accent)' : '#7a3a4a',
+                  fontWeight: 600,
+                }}>{Math.round(verification.score)}</span>
+                <span style={{ fontSize: 9, color: 'var(--text-faint)', letterSpacing: '0.15em' }}>可信度 / 100</span>
+              </div>
+              <div style={{ fontSize: 10, color: 'var(--text-secondary)', lineHeight: 1.4 }}>
+                {verification.reason}
+              </div>
+            </div>
+          )}
+        </div>
+
         {/* 攻击对象 (源节点) */}
         {sourceTitle && (
           <div className="text-[9px] mt-2.5 pt-1.5" style={{ color: 'var(--text-muted)', borderTop: '1px dashed var(--border-subtle)' }}>
diff --git a/src/pages/KnowledgeGraph.jsx b/src/pages/KnowledgeGraph.jsx
@@ -509,6 +509,42 @@ export default function KnowledgeGraph() {
       }
     }
 
+    // 反驳节点的"二次验证"按钮 — 让 LLM 反思这个反驳是否切合实际, 给可信度评分
+    // (用户图 43 反馈: 反驳脱离实际, 比如 PDF 不会超 1GB 但反驳照搬 OOM 风险)
+    const onVerifyHermes = async (e) => {
+      const { challengeId, claim, angle, severity, sourceTitle } = e.detail || {}
+      if (!challengeId || !claim) return
+      const store = useCanvasStore.getState()
+      store.updateNode(challengeId, { verifyRunning: true })
+      try {
+        const { callLLM } = await import('../services/aiProvider')
+        const { tryParseLLMJson } = await import('../services/aiService')
+        const system = `你是 Hermes 二次审查官. 给定一个对原方案的反驳, 你要判断这个反驳是否站得住脚 — 是否切合实际语境 (避免 LLM 编造不切实际的论点, 比如假设 PDF 超 1GB 之类).
+按 JSON 输出: { "score": 0-100 整数 (反驳的可信度), "verdict": "成立"|"勉强"|"不成立", "reason": "一句话理由 (50 字内)" }`
+        const prompt = `原方案 (针对): ${sourceTitle || '未知'}
+反驳角度: ${angle || ''}
+反驳论点: ${claim}
+反驳严重度: ${severity || 'medium'}
+
+请审视这个反驳是否切合实际, 还是 LLM 想当然脱离场景的臆测? 评分 + 一句话理由.`
+        const raw = await callLLM({ system, prompt, jsonMode: true })
+        const parsed = tryParseLLMJson(raw)
+        const score = typeof parsed?.score === 'number' ? Math.max(0, Math.min(100, parsed.score)) : 50
+        const verdict = parsed?.verdict || (score >= 70 ? '成立' : score >= 40 ? '勉强' : '不成立')
+        const reason = parsed?.reason || '审查官未给出理由'
+        useCanvasStore.getState().updateNode(challengeId, {
+          verifyRunning: false,
+          verification: { score, verdict, reason, ts: Date.now() },
+        })
+      } catch (err) {
+        console.error('[verify-hermes] 失败:', err)
+        useCanvasStore.getState().updateNode(challengeId, {
+          verifyRunning: false,
+          verification: { score: 0, verdict: '审查失败', reason: err?.message || String(err) },
+        })
+      }
+    }
+
     // 反驳节点的"下一步"按钮 — 把 todos 派给元认知, 串成下一轮可执行项目
     const onChainTodos = (e) => {
       const { challengeId, todos = [], claim = '', sourceTitle = '' } = e.detail || {}
@@ -543,6 +579,7 @@ export default function KnowledgeGraph() {
     window.addEventListener('node-change-type', onNodeChangeType)
     window.addEventListener('group-color-change', onGroupColorChange)
     window.addEventListener('challenge:chain-todos', onChainTodos)
+    window.addEventListener('challenge:verify-hermes', onVerifyHermes)
     window.addEventListener('canvas-auto-save', onAutoSave)
 
     return () => {
@@ -557,6 +594,8 @@ export default function KnowledgeGraph() {
       window.removeEventListener('node-change-type', onNodeChangeType)
       window.removeEventListener('group-color-change', onGroupColorChange)
       window.removeEventListener('challenge:chain-todos', onChainTodos)
+      window.removeEventListener('challenge:verify-hermes', onVerifyHermes)
+      window.removeEventListener('canvas-auto-save', onAutoSave)
     }
   }, [handleFileDrop, addBookmarkNode, addConceptNode, addNoteNode, addImageNode, addVideoNode, addFileNode, onNodesChange])
 
