feat(sense): T-G7 审计+前端（live 分流接线 + tier 指标卡 + Settings 灰度区块）

- proxy/routes：sense.mode=live 时转发前调 Grader 分流，档位 -> 池条目（tier_pool_hint 映射），响应头 X-Campus-Tier 透出（可解释性）； T1 决策按 review.sample_rate 抽样入 ReviewQueue（sense_t1 标签，人工核"判 T1 实为更高档"）；分级故障不影响代理可用性（D-G4） - /api/metrics：sense 段（enabled/mode/观察数/已标签/档位分布/shadow 一致率） - 前端：MetricsView「语义分级」卡片（观察/标签进度/一致率/晋升门提示）； SettingsView「Sense 灰度控制」区块（总开关 + collect/shadow/live 切换 + 标签进度），npm run build 产物更新 - 全量 410 passed
2026-09-05 15:06:27 +08:00
parent 204789aed6
commit a636b989e5
7 changed files with 171 additions and 6 deletions
@@ -1177,6 +1177,30 @@ try:
                "pending": get_review().count(status="pending"),
                "total": get_review().count(),
            }
+        # 语义分析器（T-G7）：灰度状态 + 观察统计 + 档位分布 + shadow 一致率
+        try:
+            from gateway.sense.config import build_sense_config
+            from gateway.sense.store import SenseStore
+            scfg = build_sense_config(settings_store().to_dict())
+            srows = SenseStore.init_db(scfg.db_path).all_observations(limit=100000)
+            by_decided: Dict[str, int] = {}
+            consistent = 0
+            labeled = 0
+            for r in srows:
+                by_decided[r["decided_tier"]] = by_decided.get(r["decided_tier"], 0) + 1
+                if r["executed_tier"] and r["decided_tier"] == r["executed_tier"]:
+                    consistent += 1
+                if r["true_tier"]:
+                    labeled += 1
+            out["sense"] = {
+                "enabled": scfg.enabled, "mode": scfg.mode,
+                "observations": len(srows), "labeled": labeled,
+                "by_decided_tier": by_decided,
+                "agreement": round(consistent / len(srows), 4) if srows else 0.0,
+                "min_labels": scfg.min_labels,
+            }
+        except Exception:
+            pass
        return out

 except ImportError:
@@ -78,11 +78,59 @@ def build_proxy_router(cfg: ProxyConfig, pool) -> APIRouter:
                           _limits(), time.time())
        if not _limits().acquire_slot(ctx["key_id"]):
            raise QuotaError("并发请求已达该 key 上限")
+
+        # 语义分析器 live 分流（D-G7：mode=live 才启用；任何异常不影响代理可用性）
+        tier_used = None
        try:
-            return await _run_chat(body, dict(request.headers), ctx, cfg, ledger, pool)
+            from gateway.sense.config import build_sense_config
+            from gateway.sense.grader import Grader
+            from gateway.sense.observer import get_observer
+            from gateway.sense.store import SenseStore
+            scfg = build_sense_config(settings_store().to_dict())
+            if scfg.enabled and scfg.mode == "live":
+                sstore = SenseStore.init_db(scfg.db_path)
+                grader = Grader(scfg, sstore, get_observer(sstore))
+                qtext = "\n".join(str(m.get("content") or "")
+                                  for m in (body.get("messages") or []))
+                d = await grader.decide(qtext or str(body.get("model") or ""),
+                                        "proxy")
+                tier_used = d.tier
+                # 档位 -> 池条目（§2 消费方表）：t1/t2/t3 映射模型替换请求模型
+                hint = scfg.tier_pool_hint("proxy", tier_used)
+                e = _entry_by_tier_hint(pool, hint)
+                if e is not None:
+                    body = {**body, "model": e["model"]}
+        except Exception:
+            pass  # D-G4：分级故障不影响代理可用性
+        try:
+            resp = await _run_chat(body, dict(request.headers), ctx, cfg, ledger, pool)
+            if tier_used:
+                resp.headers["x-campus-tier"] = tier_used
+                # T1 审计抽样（§9.4）：按 review.sample_rate 入队人工核
+                try:
+                    import random as _random
+                    rate = float(load_config().get("review", {}).get("sample_rate", 0.1))
+                    if _random.random() < rate:
+                        get_review().enqueue(
+                            request_id + "-sense", str(body.get("model") or "proxy"),
+                            "(sense T1 审计抽样)", tags=["sense_t1"],
+                            reason="sense_audit")
+                except Exception:
+                    pass
+            return resp
        finally:
            _limits().release_slot(ctx["key_id"])

+    def _entry_by_tier_hint(pool, hint: str) -> Optional[Dict[str, Any]]:
+        """按档位名选池条目（tier_hint 优先，缺省按 tier 字段映射）。"""
+        entries = pool.list().get("entries", [])
+        for e in entries:
+            if not e.get("enabled") or e.get("backend") == "mock":
+                continue
+            if (e.get("tier_hint") or e.get("tier")) == hint:
+                return e
+        return None
+
    # ---------------- 管理面 ----------------
    def _guard_admin(request: Request) -> None:
        host = request.client.host if request.client else ""
@@ -1,14 +1,14 @@
 {
  "schemaVersion": "mimosa-hook-status/v1",
-  "recordedAt": "2026-09-01T14:21:11.564Z",
+  "recordedAt": "2026-09-05T07:03:44.339Z",
  "sessionId": "sess_e50d4f25-3ac6-43f2-b2f3-8833b3150465",
  "event": "PostToolUse",
  "toolName": "Edit",
-  "file": "src/views/AgentView.vue",
+  "file": "src/views/MetricsView.vue",
  "outcome": "clear",
  "coverage": "complete",
  "findingCount": 0,
-  "durationMs": 7,
+  "durationMs": 2,
  "hostState": "hook_complete",
  "reportHint": ".mimosa/reports/"
 }
@@ -114,4 +114,13 @@ export interface Metrics {
  cache: Record<string, unknown>
  v2?: Record<string, unknown>
  review?: { pending: number; total: number }
+  sense?: {
+    enabled: boolean
+    mode: string
+    observations: number
+    labeled: number
+    by_decided_tier?: Record<string, number>
+    agreement?: number
+    min_labels?: number
+  }
 }
@@ -58,6 +58,21 @@
          <p class="hint">单价来自模型池条目（$/1M tokens）；经典设置下的模型成本不计入。</p>
        </div>

+        <div v-if="sense && sense.observations" class="metric-card">
+          <h3>语义分级（sense · {{ sense.mode }}）</h3>
+          <div class="kv-list">
+            <span>观察数</span><b>{{ sense.observations }}</b>
+            <span>已标签</span><b>{{ sense.labeled }} / {{ sense.min_labels }}</b>
+            <span>一致率</span><b>{{ ((sense.agreement ?? 0) * 100).toFixed(1) }}%</b>
+            <template v-for="(v, k) in sense.by_decided_tier" :key="k">
+              <span>判 {{ k }}</span><b>{{ v }}</b>
+            </template>
+          </div>
+          <p class="hint">
+            晋升门：一致率 ≥85% + 标签 ≥{{ sense.min_labels }} + 审计无误判（collect 攒满前不开 live）
+          </p>
+        </div>
+
        <div v-if="data.review" class="metric-card review-card">
          <h3>人工检验</h3>
          <div class="review-stats">
@@ -115,6 +130,8 @@ const byModel = computed(() => {
  return (v2?.by_model as Record<string, ModelBucket>) || null
 })

+const sense = computed(() => data.value?.sense || null)
+
 async function load() {
  loading.value = true
  error.value = ''
@@ -389,6 +389,45 @@
      </div>
    </section>

+    <!-- 🧠 语义分级（sense）灰度控制 -->
+    <section class="settings-section">
+      <div class="section-title">
+        <span>🧠 语义分级（Sense · 三级任务分级）</span>
+        <div class="title-right">
+          <span :class="['server-badge', senseCfg.enabled ? 'badge-running' : 'badge-stopped']">
+            {{ senseCfg.enabled ? `● ${senseCfg.mode}` : '○ 未启用' }}
+          </span>
+        </div>
+      </div>
+      <div class="section-body">
+        <p class="role-hint">
+          灰度纪律：collect（只记不决策）攒满 {{ senseCfg.min_labels }} 条标签并过晋升门
+          （一致率 ≥85% + conformal 覆盖 ≤α+2% + 审计无误判）后才可开 live。
+        </p>
+        <div class="form-grid">
+          <div class="field">
+            <label class="checkbox-label">
+              <input type="checkbox" v-model="senseCfg.enabled" @change="saveSense" />
+              启用语义分析器（总开关，改后保存生效）
+            </label>
+          </div>
+          <div class="field">
+            <label>灰度模式</label>
+            <select v-model="senseCfg.mode" @change="saveSense"
+              :disabled="!senseCfg.enabled">
+              <option value="collect">collect — 只采集观察</option>
+              <option value="shadow">shadow — 决策只比对不执行</option>
+              <option value="live">live — 真分流（需晋升门通过）</option>
+            </select>
+          </div>
+          <div class="field">
+            <label>观察/标签数（min_labels={{ senseCfg.min_labels }}）</label>
+            <input :value="`${senseCfg.obs} / ${senseCfg.labeled}`" disabled />
+          </div>
+        </div>
+      </div>
+    </section>
+
    <!-- 操作按钮 -->
    <div class="actions">
      <button class="btn-primary" :disabled="saving || !dirty" @click="save">
@@ -412,6 +451,7 @@ import {
  getLlamaStatus, listLocalModels, startLlama, stopLlama,
  downloadModel, watchDownload,
  getPool, upsertPoolEntry, deletePoolEntry, setPoolRoles, testPoolEntry, listPoolModels,
+  getMetrics,
 } from '@/api'
 import type {
  ModelSettings, ModelInfo, LlamaStatus, LocalModel,
@@ -961,7 +1001,34 @@ async function reset() {
  }
 }

-onMounted(load)
+// 语义分级（sense）灰度
+const senseCfg = reactive({ enabled: false, mode: 'collect', obs: 0, labeled: 0, min_labels: 500 })
+
+async function loadSense() {
+  try {
+    const [cfg, m] = await Promise.all([getConfig(), getMetrics()])
+    const s = (cfg as any).sense || {}
+    senseCfg.enabled = !!s.enabled
+    senseCfg.mode = s.mode || 'collect'
+    senseCfg.min_labels = s.policy?.min_labels ?? 500
+    const sm = (m as any).sense || {}
+    senseCfg.obs = sm.observations ?? 0
+    senseCfg.labeled = sm.labeled ?? 0
+  } catch { /* 静默 */ }
+}
+
+async function saveSense() {
+  try {
+    await updateConfig({ sense: { enabled: senseCfg.enabled, mode: senseCfg.mode } } as any)
+    saveMsg.value = { ok: true, msg: `✅ 语义分级已${senseCfg.enabled ? '启用（' + senseCfg.mode + '）' : '关闭'}` }
+  } catch (e: any) {
+    saveMsg.value = { ok: false, msg: '❌ 保存失败：' + (e?.response?.data?.detail || e?.message || e) }
+    await loadSense()
+  }
+  setTimeout(() => (saveMsg.value = null), 3000)
+}
+
+onMounted(() => { load(); loadSense() })
 onUnmounted(() => { _dlSSE?.close() })
 </script>

@@ -162,5 +162,5 @@ P0 完成后的能力：干净的后端抽象 + 可量化的评测 + 可追溯
 | T-G4 | 线性头：离线训练脚本 + LinearHead 纯 Python 推理 + 登记 | ✅ 完成 | T-G4 |
 | T-G5 | Grader：决策组合（特征门×概率×conformal）+ /v1/route + 三态 mode | ✅ 完成 | T-G5 |
 | T-G6 | live 分流：pipeline tier_fn 三档钩子 + proxy 档位映射 + T2 档升级阶梯 | ✅ 完成 | T-G6 |
-| T-G7 | 审计+前端：ReviewQueue 抽样 + tier 指标卡 + 客户端来源显示/一键升级 | ⬜ 待办 | |
+| T-G7 | 审计+前端：ReviewQueue 抽样 + tier 指标卡 + 客户端来源显示/一键升级 | ✅ 完成 | T-G7 |
 | T-G8 | 实验：E-G1/E-G3 报告；（可选）LoraRemote + E-G2 线性 vs LoRA | ⬜ 待办 | |