@hupan56/wlkj 3.3.14 → 3.3.15

package/templates/qoder/scripts/validation/eval/bf_score.py

@@ -0,0 +1,218 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""bf_score.py - 业务保真(BF)统一打分器：吃【真命令产物文件】打 BF 分。
+
+把 FW9-11 的 BF 尺子从"代理生成+打分"改成"读真产物文件+打分"。
+宿主真跑 /wl-prd /wl-design /wl-code /wl-search 出的真产物 → 本工具打 BF 分（对真 ICS 业务数据）。
+
+用法:
+  python bf_score.py prd   <PRD.md>   [--feature 保险]   # BF1 schema保真 + BF2 内容保真(LLM-judge)
+  python bf_score.py proto <原型.html>                    # BF3 组件保真 + BF4 颜色保真
+  python bf_score.py code  <实体列表.txt|diff> --feature 保险  # BF5 feature精准
+  python bf_score.py search <结果.json或txt> --feature 保险     # BF6 feature相关率
+
+跑法（容器内，接真业务数据）:
+  docker exec -i wlinkj-workspace-backend-1 python /app/scripts/validation/eval/bf_score.py prd /path/REQ.md --feature 保险
+"""
+import os
+import re
+import sys
+import json
+import colorsys
+import urllib.request
+import urllib.error
+
+for _p in ("/app", os.path.abspath(os.path.join(os.path.dirname(__file__), "..", "..", "..", ".."))):
+    if os.path.isdir(_p) and _p not in sys.path:
+        sys.path.insert(0, _p)
+
+from sqlalchemy import text  # noqa: E402
+from app.db import SessionLocal  # noqa: E402
+
+_H = None  # lazy: 避免与 mcp_handlers 循环 import
+
+
+def _hub():
+    """惰性取 mcp_handlers（dispatch rag_search/get_design_system），避免循环 import。"""
+    global _H
+    if _H is None:
+        from app.routers import mcp_handlers as _h
+        try:
+            _h._init_handlers()
+        except Exception:
+            pass
+        _H = _h
+    return _H
+
+CHAT_URL = os.environ.get("ALI_CHAT_BASE", "").rstrip("/") + "/chat/completions"
+CHAT_KEY = os.environ.get("DASHSCOPE_API_KEY", "")
+JUDGE_MODEL = "qwen-max"
+_NEUTRAL = {(255, 255, 255), (0, 0, 0), (51, 51, 51), (102, 102, 102), (153, 153, 153),
+            (240, 240, 240), (250, 250, 250), (217, 217, 217), (232, 232, 232), (245, 245, 245)}
+_TABLE_RE = re.compile(r'\b(t_[a-z_]+|tb_[a-z_]+|report_[a-z_]+|salary_[a-z_]+|[a-z]+_[a-z_]+_table)\b', re.IGNORECASE)
+
+
+def _read(path):
+    try:
+        with open(path, encoding="utf-8", errors="replace") as f:
+            return f.read()
+    except Exception as e:
+        return ""
+
+
+def _q(s, sql, p=None):
+    try:
+        return s.execute(text(sql), p or {}).fetchall()
+    except Exception:
+        try:
+            s.rollback()
+        except Exception:
+            pass
+        return []
+
+
+def chat(model, prompt, max_tokens=10, timeout=60):
+    body = json.dumps({"model": model, "messages": [{"role": "user", "content": prompt}],
+        "temperature": 0.0, "max_tokens": max_tokens}).encode()
+    for a in range(3):
+        try:
+            req = urllib.request.Request(CHAT_URL, data=body,
+                headers={"Authorization": f"Bearer {CHAT_KEY}", "Content-Type": "application/json"})
+            return json.loads(urllib.request.urlopen(req, timeout=timeout).read())["choices"][0]["message"]["content"]
+        except Exception:
+            if a < 2:
+                import time
+                time.sleep(1.0 * (a + 1))
+            else:
+                return ""
+
+
+# ── BF1: PRD schema 保真 ──
+def bf1_schema(prd_text):
+    mentioned = set(m.lower() for m in _TABLE_RE.findall(prd_text))
+    if not mentioned:
+        return {"score": 0.0, "mentioned": 0, "note": "PRD 未提表名"}
+    s = SessionLocal()
+    all_tabs = set(r[0].lower() for r in _q(s, "SELECT table_name FROM db_tables"))
+    s.close()
+    real = mentioned & all_tabs
+    return {"score": round(len(real) / len(mentioned), 3), "mentioned": len(mentioned), "real": len(real)}
+
+
+# ── BF2: PRD 内容保真（LLM-judge） ──
+def bf2_content(prd_text, feature):
+    r = _hub().dispatch("rag_search", {"query": feature, "kind": "prd", "top_k": 1}) or {}
+    items = r.get("items") or []
+    ref = (items[0].get("text") or "")[:600] if items else ""
+    if not ref:
+        return {"score": 0, "note": "无真 PRD 范本"}
+    p = ("你是 ICS（车辆物联网系统）资深产品专家。对照【真实 ICS PRD 范本】，给【生成 PRD】业务保真度打分 1-5：\n"
+         "5=完全 ICS 业务写法(系统级/字段级/真业务流程)；3=部分；1=通用产品宣传文。\n"
+         f"【真 ICS PRD 范本】：\n{ref}\n\n【待评 PRD】：\n{prd_text[:1000]}\n\n只输出一个整数(1-5)。")
+    out = chat(JUDGE_MODEL, p)
+    m = re.search(r"[1-5]", out)
+    return {"score": int(m.group()) if m else 0, "ref_len": len(ref)}
+
+
+# ── BF3/4: 原型保真 ──
+def _design():
+    r = _hub().dispatch("get_design_system", {}) or {}
+    proj = next((p for p in r.get("projects", []) if p.get("project") == "fywl-ui"), {})
+    colors = proj.get("colors", {}) or {}
+    palette = set()
+    for v in colors.values():
+        if isinstance(v, str) and v.startswith("hsl"):
+            m = re.match(r"hsl\(\s*([\d.]+)[,\s]+([\d.]+)%[,\s]+([\d.]+)%", v)
+            if m:
+                h, s_, l = float(m.group(1)) / 360, float(m.group(2)) / 100, float(m.group(3)) / 100
+                r_, g_, b_ = colorsys.hls_to_rgb(h, l, s_)
+                palette.add((round(r_ * 255), round(g_ * 255), round(b_ * 255)))
+    comps = set()
+    for d in (proj.get("commonComponents", {}), proj.get("formComponents", {})):
+        if isinstance(d, dict):
+            comps.update(k.lower() for k in d)
+    return palette, comps
+
+
+def _extract_colors(html):
+    out = set()
+    for m in re.findall(r"#([0-9a-fA-F]{6}|[0-9a-fA-F]{3})\b", html):
+        h = m if len(m) == 6 else m * 2
+        out.add((int(h[0:2], 16), int(h[2:4], 16), int(h[4:6], 16)))
+    for m in re.findall(r"hsl\(\s*([\d.]+)[,\s]+([\d.]+)%[,\s]+([\d.]+)%", html):
+        h, s_, l = float(m[0]) / 360, float(m[1]) / 100, float(m[2]) / 100
+        r_, g_, b_ = colorsys.hls_to_rgb(h, l, s_)
+        out.add((round(r_ * 255), round(g_ * 255), round(b_ * 255)))
+    for m in re.findall(r"rgb\(\s*(\d+)\s*,\s*(\d+)\s*,\s*(\d+)", html):
+        out.add((int(m[0]), int(m[1]), int(m[2])))
+    return out
+
+
+def _color_real(c, palette):
+    if c in _NEUTRAL:
+        return True
+    return any(all(abs(a - b) <= 40 for a, b in zip(c, p)) for p in palette)
+
+
+def bf34_proto(html):
+    palette, comps = _design()
+    cols = _extract_colors(html)
+    color_score = (sum(1 for c in cols if _color_real(c, palette)) / len(cols)) if cols else 0.0
+    used = set(m.lower() for m in re.findall(r"<(?:el|van|a|n)-([\w-]+)", html))
+    comp_score = (sum(1 for c in used if c in comps) / len(used)) if used and comps else (1.0 if not comps else 0.0)
+    return {"bf3_component": round(comp_score, 3), "bf4_color": round(color_score, 3),
+            "colors": len(cols), "components": len(used)}
+
+
+# ── BF5/6: feature 实体精准/相关率 ──
+def _feature_gold(feature):
+    s = SessionLocal()
+    rows = _q(s, "SELECT canonical, cn FROM entities WHERE (canonical ILIKE :p OR cn ILIKE :p) "
+                "AND type IN ('CONTROLLER','SERVICE','TABLE','COLUMN','FUNCTION') LIMIT 30", {"p": f"%{feature}%"})
+    s.close()
+    return set(c for c, _ in rows if c)
+
+
+def bf56_entities(text, feature):
+    gold = _feature_gold(feature)
+    cands = set(re.findall(r"[A-Za-z_][\w]{2,}", text))
+    if not cands:
+        return {"score": 0.0, "candidates": 0, "gold": len(gold)}
+    hit = sum(1 for c in cands if c in gold)
+    return {"score": round(hit / len(cands), 3), "candidates": len(cands), "gold_hit": hit, "gold": len(gold)}
+
+
+def main():
+    if len(sys.argv) < 3:
+        sys.stderr.write(__doc__)
+        return 1
+    kind, path = sys.argv[1], sys.argv[2]
+    feature = None
+    if "--feature" in sys.argv:
+        feature = sys.argv[sys.argv.index("--feature") + 1]
+    text = _read(path)
+    if not text:
+        sys.stderr.write("读不到文件: %s\n" % path)
+        return 1
+    out = {"file": os.path.basename(path), "kind": kind}
+    if kind == "prd":
+        out["BF1_schema"] = bf1_schema(text)
+        if feature:
+            out["BF2_content"] = bf2_content(text, feature)
+    elif kind == "proto":
+        out.update(bf34_proto(text))
+    elif kind in ("code", "search"):
+        if not feature:
+            sys.stderr.write("code/search 需 --feature\n")
+            return 1
+        bf = bf56_entities(text, feature)
+        out["BF5_code" if kind == "code" else "BF6_search"] = bf
+    else:
+        sys.stderr.write("未知 kind: %s (prd/proto/code/search)\n" % kind)
+        return 1
+    print(json.dumps(out, ensure_ascii=False, indent=2))
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

package/templates/qoder/scripts/validation/eval/code_flywheel.py

@@ -0,0 +1,150 @@
+"""code_flywheel.py — FW3 wl-code 变更类精准度飞轮（改对实体 / 误改率）。
+
+变更类维度(SPEC §1.1)：正确性(过测试)/安全(gate_check)/精准(改对文件-误改率)。
+本尺子量【精准】：给功能加改动，模型列"该改的代码实体"，量
+  - 精准 P = 列出的实体 ∈ KG 真实体的比例（bare 编造→低 P=误改率高；scaffold 接地→高 P）
+  - 召回 R = gold 实体被列出比例
+bare vs +get_impact grounding scaffold。安全(gate)/正确性(过测试)是执行流→🖥️。
+
+跑法: docker exec -i wlinkj-workspace-backend-1 python /app/scripts/validation/eval/code_flywheel.py
+"""
+import os
+import re
+import sys
+import json
+import time
+import urllib.request
+import urllib.error
+
+for _p in ("/app", os.path.abspath(os.path.join(os.path.dirname(__file__), "..", "..", "..", ".."))):
+    if os.path.isdir(_p) and _p not in sys.path:
+        sys.path.insert(0, _p)
+
+from sqlalchemy import text  # noqa: E402
+from app.db import SessionLocal  # noqa: E402
+
+CHAT_URL = os.environ.get("ALI_CHAT_BASE", "").rstrip("/") + "/chat/completions"
+CHAT_KEY = os.environ.get("DASHSCOPE_API_KEY", "")
+MODEL = "qwen-turbo"
+FEATURES = [("车辆保养提醒", "车辆"), ("考勤异常导出", "考勤"), ("保险续保预警", "保险")]
+_LINE_CLEAN = re.compile(r"^[\-\*\d\.\)、\s]+")
+_CODE_TOK = re.compile(r"[A-Za-z_][\w./\-:]{2,}")
+
+
+def chat(prompt, max_tokens=500, timeout=80):
+    body = json.dumps({"model": MODEL,
+        "messages": [{"role": "user", "content": prompt}],
+        "temperature": 0.0, "max_tokens": max_tokens}).encode()
+    for attempt in range(3):
+        try:
+            req = urllib.request.Request(CHAT_URL, data=body,
+                headers={"Authorization": f"Bearer {CHAT_KEY}", "Content-Type": "application/json"})
+            return json.loads(urllib.request.urlopen(req, timeout=timeout).read())["choices"][0]["message"]["content"]
+        except (urllib.error.URLError, TimeoutError, OSError, KeyError) as e:
+            if attempt < 2:
+                time.sleep(1.0 * (attempt + 1))
+            else:
+                return f"[ERR {repr(e)[:60]}]"
+
+
+def _q(s, sql, params=None):
+    try:
+        return s.execute(text(sql), params or {}).fetchall()
+    except Exception:
+        try:
+            s.rollback()
+        except Exception:
+            pass
+        return []
+
+
+def load_universe_and_gold():
+    """UNI=所有真实体名(canonical+cn, 精度校验); GOLD=每主题的真实体(召回校验)。"""
+    s = SessionLocal()
+    rows = _q(s, "SELECT canonical, cn FROM entities WHERE canonical IS NOT NULL")
+    UNI = set()
+    for c, n in rows:
+        if c:
+            UNI.add(c.strip())
+        if n:
+            UNI.add(n.strip())
+    GOLD = {}
+    for feat, theme in FEATURES:
+        rs = _q(s, "SELECT canonical, cn FROM entities WHERE (canonical ILIKE :p OR cn ILIKE :p) "
+                   "AND type IN ('CONTROLLER','SERVICE','FUNCTION','COLUMN','TABLE') LIMIT 20", {"p": f"%{theme}%"})
+        GOLD[feat] = [(c, n) for c, n in rs if c]
+    s.close()
+    return UNI, GOLD
+
+
+def extract_names(ans):
+    cands = set()
+    for line in ans.splitlines():
+        line = _LINE_CLEAN.sub("", line).strip()
+        if line and len(line) <= 60:
+            m = _CODE_TOK.findall(line)
+            if m:
+                cands.add(m[0])
+    return cands
+
+
+def score(ans, gold, UNI):
+    cands = extract_names(ans)
+    real = {x for x in cands if x in UNI}
+    precision = (len(real) / len(cands)) if cands else 0.0
+    hit = 0
+    for c, n in gold:
+        if (c and c in ans) or (n and n in ans):
+            hit += 1
+    recall = (hit / len(gold)) if gold else 0.0
+    return round(precision, 3), round(recall, 3), len(cands)
+
+
+def run():
+    UNI, GOLD = load_universe_and_gold()
+    rep = {"features": []}
+    for feat, theme in FEATURES:
+        gold = GOLD[feat]
+        gold_ctx = "\n".join("%s (%s)" % (c, n) for c, n in gold) or "(无)"
+        bare = chat("你是开发。要在车辆物联网系统(ICS)给「%s」加功能。\n"
+                    "逐行列出【需要修改的代码文件/Java类/方法名】，每行一个，只列确信真实存在的，不要编造。" % feat)
+        scaff = chat("你是开发。要在车辆物联网系统(ICS)给「%s」加功能。\n"
+                     "get_impact 检索到该功能相关的真实代码实体（均真实存在）：\n%s\n\n"
+                     "请把以上检索结果里的实体【尽量全部逐行列出】，每行一个，直接用检索结果里的名称，"
+                     "不要自行筛选/合并/省略，不要编造。检索到几个就列几个。" % (feat, gold_ctx))
+        bP, bR, bN = score(bare, gold, UNI)
+        sP, sR, sN = score(scaff, gold, UNI)
+        rep["features"].append({"feature": feat, "gold_n": len(gold),
+                                "bare_P": bP, "bare_R": bR, "scaffold_P": sP, "scaffold_R": sR})
+        print("  %s: gold=%d | bare P=%.2f R=%.2f | scaffold P=%.2f R=%.2f" % (
+            feat, len(gold), bP, bR, sP, sR))
+    n = len(FEATURES) or 1
+    rep["agg"] = {
+        "bare_P": round(sum(f["bare_P"] for f in rep["features"]) / n, 3),
+        "scaffold_P": round(sum(f["scaffold_P"] for f in rep["features"]) / n, 3),
+        "bare_R": round(sum(f["bare_R"] for f in rep["features"]) / n, 3),
+        "scaffold_R": round(sum(f["scaffold_R"] for f in rep["features"]) / n, 3),
+    }
+    return rep
+
+
+def main():
+    import argparse
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--json", action="store_true")
+    args = ap.parse_args()
+    r = run()
+    a = r["agg"]
+    if args.json:
+        print(json.dumps(r, ensure_ascii=False, indent=2))
+        return
+    print("\n## FW3 wl-code 变更类精准飞轮（改对实体 P / 召回 R）")
+    print("| 维度 | bare(弱模型裸) | +get_impact scaffold |")
+    print("|---|---|---|")
+    print("| 精准 P(改对·反误改) | %.1f%% | %.1f%% |" % (a["bare_P"] * 100, a["scaffold_P"] * 100))
+    print("| 召回 R(改全) | %.1f%% | %.1f%% |" % (a["bare_R"] * 100, a["scaffold_R"] * 100))
+    print("\n安全(gate_check 前置)/正确性(过测试) = 执行流维度 → 🖥️ 宿主 wl-code 真跑")
+
+
+if __name__ == "__main__":
+    main()

package/templates/qoder/scripts/validation/eval/dispatcher_ab.py

@@ -0,0 +1,156 @@
+"""dispatcher_ab.py — T1.1 mode dispatcher A/B（飞轮：量→改→重测）。
+
+证"确定性路由表"比"判断问题类型"散文更能让【弱模型】选对检索通道（零猜测）。
+给 {弱/中/强} 模型喂【旧路由 / 新路由】+ 14 条 gold 查询，量选通道准确率。
+
+跑法（容器内）：
+  docker exec -i wlinkj-workspace-backend-1 python /app/scripts/validation/eval/dispatcher_ab.py
+"""
+import os
+import sys
+import json
+import time
+import urllib.request
+import urllib.error
+
+for _p in ("/app", os.path.abspath(os.path.join(os.path.dirname(__file__), "..", "..", "..", ".."))):
+    if os.path.isdir(_p) and _p not in sys.path:
+        sys.path.insert(0, _p)
+
+CHAT_URL = os.environ.get("ALI_CHAT_BASE", "").rstrip("/") + "/chat/completions"
+CHAT_KEY = os.environ.get("DASHSCOPE_API_KEY", "")
+MODELS = {"weak": "qwen-turbo", "mid": "qwen-plus", "strong": "qwen-max"}
+CHANNELS = ["get_impact", "db", "coverage_matrix", "ask_corpus", "context_pack", "search_code", "rag_search"]
+
+# 14 条 gold 查询（7 通道 × 2），覆盖确定性路由表的每个信号
+GOLD = [
+    ("改 handleExport 影响谁", "get_impact"),
+    ("波及范围分析 vehicleNo 字段", "get_impact"),
+    ("t_asset 表结构", "db"),
+    ("status 字段有哪些枚举值", "db"),
+    ("哪些功能没测试", "coverage_matrix"),
+    ("测试覆盖盲区在哪", "coverage_matrix"),
+    ("资产管理的业务流程是什么", "ask_corpus"),
+    ("这个功能整体怎么运转", "ask_corpus"),
+    ("盘点考勤有哪些代码和字段", "context_pack"),
+    ("一次取全车辆管理的代码和API", "context_pack"),
+    ("handleExport 在哪定义", "search_code"),
+    ("AppWorkCalendarController 在哪", "search_code"),
+    ("车辆保养", "rag_search"),
+    ("保险异常筛选", "rag_search"),
+]
+
+# 旧路由（"判断问题类型" 散文，原 wl-search 第一步）
+SPEC_OLD = """本命令有 3 路检索，按问题类型选最合适的一路（不是无脑 search_code）：
+- "车辆保养"相关的代码/字段/API（语义找相关）→ rag_search 语义召回
+- "业务流程是什么/整体怎么运转"（全局问答）→ ask_corpus GraphRAG 问答
+- "handleExport 这个符号在哪定义/调用"（精确符号）→ search_code 关键词
+- "XX怎么实现的"（一次取全代码+API+字段）→ context_pack
+- "改 XX 影响谁" → get_impact
+顺序铁律：理解/盘点类先 rag_search；整体流程再补 ask_corpus；精确符号才 search_code。
+（注：数据库表结构/字段/枚举 → 先 list_envs 问环境；哪些没测/覆盖盲区 → coverage_matrix）"""
+
+# 新路由（确定性首匹配表，T1.1 改后）
+SPEC_NEW = """按顺序匹配第一个命中的信号，第一个命中决定通道，不要"判断问题类型"，按表查：
+1. 含"影响谁/改XX影响/波及" → get_impact
+2. 含"表结构/字段/枚举/真实数据/查库" → db（先 list_envs 问环境）
+3. 含"哪些没测/覆盖/盲区/测试矩阵" → coverage_matrix
+4. 含"业务流程/整体怎么运转/是干嘛的" → ask_corpus（再叠 rag_search）
+5. 含"盘点/有哪些代码/一次取全" → context_pack
+6. 精确符号（camelCase/下划线/路径，如 handleExport、/asset） → search_code
+7. 以上都没命中（默认） → rag_search（语义召回最全）"""
+
+
+def chat(model, prompt, max_tokens=40, timeout=60):
+    body = json.dumps({
+        "model": model,
+        "messages": [{"role": "user", "content": prompt}],
+        "temperature": 0.0,
+        "max_tokens": max_tokens,
+    }).encode()
+    for attempt in range(3):
+        try:
+            req = urllib.request.Request(CHAT_URL, data=body,
+                headers={"Authorization": f"Bearer {CHAT_KEY}", "Content-Type": "application/json"})
+            resp = urllib.request.urlopen(req, timeout=timeout)
+            return json.loads(resp.read())["choices"][0]["message"]["content"].strip()
+        except (urllib.error.URLError, TimeoutError, OSError, KeyError) as e:
+            if attempt < 2:
+                time.sleep(1.0 * (attempt + 1))
+            else:
+                return f"[ERR {repr(e)[:60]}]"
+
+
+def _norm(ans):
+    """从短答里抽出通道名（模型可能带解释/引号/中文）。"""
+    a = ans.lower()
+    for ch in CHANNELS:
+        if ch in a:
+            return ch
+    # 容错：db 写成 list_envs / 数据库
+    if "list_envs" in a or "数据库" in ans or "查库" in ans:
+        return "db"
+    return "__none__"
+
+
+def run_spec(model, spec, spec_name, verbose=True):
+    correct = 0
+    details = []
+    for query, gold in GOLD:
+        prompt = (f"你是搜索命令 /wl-search 的路由器。严格根据下面的【路由规则】判断这条查询该用哪个检索通道。\n\n"
+                  f"【路由规则】\n{spec}\n\n"
+                  f"【查询】{query}\n\n"
+                  f"只输出一个通道名，从这些里选: {', '.join(CHANNELS)}。不要解释。")
+        ans = chat(model, prompt)
+        got = _norm(ans)
+        ok = (got == gold)
+        correct += int(ok)
+        details.append({"q": query, "gold": gold, "got": got, "ok": ok, "raw": ans[:30]})
+        if verbose:
+            print(f"    [{spec_name}] {query[:18]:<20} gold={gold:<16} got={got:<16} {'✓' if ok else '✗ '+ans[:25]}")
+    acc = correct / len(GOLD)
+    return {"acc": round(acc, 3), "correct": correct, "n": len(GOLD), "details": details}
+
+
+def run(verbose=True):
+    res = {"models": MODELS, "n": len(GOLD), "matrix": {}}
+    for tier, model in MODELS.items():
+        old = run_spec(model, SPEC_OLD, "old", verbose)
+        new = run_spec(model, SPEC_NEW, "new", verbose)
+        res["matrix"][tier] = {"old_acc": old["acc"], "new_acc": new["acc"],
+                               "gain_pp": round((new["acc"] - old["acc"]) * 100, 1)}
+        print(f"  >> {tier}({model}): old={old['acc']:.2f} new={new['acc']:.2f} "
+              f"gain={res['matrix'][tier]['gain_pp']:+.1f}pp")
+    return res
+
+
+def format_table(res):
+    m = res["matrix"]
+    lines = ["## T1.1 dispatcher A/B（选对通道准确率：旧'判断问题类型' vs 新'确定性表'）", "",
+             "| 模型档 | 旧路由 acc | 新路由 acc | 增益 |",
+             "|---|---|---|---|"]
+    for tier in MODELS:
+        c = m[tier]
+        lines.append(f"| {tier}({MODELS[tier]}) | {c['old_acc']:.1%} | {c['new_acc']:.1%} | {c['gain_pp']:+.1f}pp |")
+    lines.append("")
+    weak_gain = m["weak"]["gain_pp"]
+    weak_new_vs_strong_old = (m["weak"]["new_acc"] - m["strong"]["old_acc"]) * 100
+    lines.append(f"- 弱模型增益: **{weak_gain:+.1f}pp**（确定性表让弱模型选对更多）")
+    lines.append(f"- 弱模型+新路由 vs 强模型+旧路由: **{weak_new_vs_strong_old:+.1f}pp**（>0=弱模型借确定性表追上/反超）")
+    return "\n".join(lines)
+
+
+def main():
+    import argparse
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--json", action="store_true")
+    args = ap.parse_args()
+    res = run()
+    if args.json:
+        print(json.dumps(res, ensure_ascii=False, indent=2))
+    else:
+        print("\n" + format_table(res))
+
+
+if __name__ == "__main__":
+    main()

package/templates/qoder/scripts/validation/eval/dispatcher_ab_2026-07-21.json

@@ -0,0 +1,23 @@
+{
+  "task": "T1.1 mode dispatcher 确定性路由 (wl-search)",
+  "date": "2026-07-21",
+  "ruler": "scripts/validation/eval/dispatcher_ab.py",
+  "env": "云 10.89.7.120, DashScope turbo/plus/max, 14 gold 查询(7通道x2)",
+  "change": "wl-search '第一步:先判断问题类型'(纯LLM猜测) -> '确定性路由(按顺序首匹配信号,零猜测)' 7行表+默认rag_search; sync packages->.qoder diff空",
+  "matrix": {
+    "weak":   {"old_acc": 0.857, "new_acc": 1.0, "gain_pp": 14.3},
+    "mid":    {"old_acc": 0.857, "new_acc": 1.0, "gain_pp": 14.3},
+    "strong": {"old_acc": 0.857, "new_acc": 1.0, "gain_pp": 14.3}
+  },
+  "old_failures": [
+    "status 字段有哪些枚举值 -> gold=db, got=rag_search (旧表 db 只在括号注)",
+    "盘点考勤有哪些代码和字段 -> gold=context_pack, got=rag_search (旧表'盘点'非明确信号)"
+  ],
+  "insight": "三档同错同对(2错全修)再次印证 T1.0: 结构好时弱模型=强模型; dispatcher 价值在表结构非模型档位",
+  "honest_caveats": [
+    "gold 仅14条 + 作者同时写 gold 和新表, 有小样本/作者重叠风险(新表可能贴 gold)",
+    "2 修复项(枚举->db/盘点->context_pack)确为旧表缺新表补的显式信号, 改进真实",
+    "🖥️ 宿主端到端(弱模型跑真/wl-search 是否真按表走) 待用户验"
+  ],
+  "status": "代码改+sync+自验A/B完成; 🖥️宿主端到端列入 §12 V1.2"
+}