codebee 0.1.95 → 0.1.96

package/app/core/aiflavor.py

@@ -1,197 +1,197 @@
-# -*- coding: utf-8 -*-
-"""AI 味确定性检测（借鉴 oh-story-claudecode 的去AI味思路）。
-
-模型评审是主观分，这里补一道客观参考线：统计译制腔/套话高频词的
-密度，产出报告行注入评审提示词，让评审官把「套话密度」纳入评分
-参考。只做参考不做门禁——「仿佛」「缓缓」在正常小说里也常见，
-误伤风险由评审官（能读懂上下文）兜底，脚本只负责指认位置与密度。
-"""
-from __future__ import annotations
-
-import re
-
-# 译制腔/套话词表（按误导性排序，持续运维）
-AI_PHRASES = (
-    "空气仿佛凝固", "眸中闪过一丝", "嘴角勾起一抹", "心中一动",
-    "值得注意的是", "总而言之", "综上所述", "不难发现",
-    "在这一刻", "仿佛在诉说着", "无声地诉说着", "见证着",
-    "深深地看了一眼", "似乎在", "空气仿佛", "一抹微笑",
-    "不禁", "仿佛",
-)
-
-# 密度告警线（每千字命中次数）：超过即提示评审官重点关注
-ALERT_PER_KILO = 8.0
-
-# 叙事架构层信号（借鉴 sepia 2.7k★/StoryScope 研究 2026：AI 小说 93.2% 靠
-# 叙事架构特征检出，人工改写措辞后检出率仅从 95.5% 降到 93.9%——措辞层
-# 改不掉的架构级指纹才是真破绽）。三类可确定性检测的架构信号：
-NARRATIVE_TELLS = {
-    "顿悟说教": ("终于明白", "这才明白", "明白了，", "意识到，自己", "懂得了",
-               "原来，成长", "原来，生活", "原来，所谓"),
-    "情绪身体化": ("心脏猛地", "指尖冰凉", "指尖发凉", "喉咙发紧", "喉头发紧",
-                "胃里一阵", "胃部一阵", "后背一凉", "血液仿佛", "呼吸一滞"),
-    "成长式收束": ("释然", "和解", "放下了", "接纳了", "与自己和解", "轻轻松了口气",
-                "内心归于平静"),
-}
-# 架构信号告警线比措辞层低：这些表达在好小说里本就该稀缺
-NARRATIVE_ALERT_PER_KILO = 2.0
-# 「成长式收束」只在结尾才构成架构指纹（中段出现多半是剧情词），只扫尾部
-ENDING_SCAN_CHARS = 600
-
-
-def analyze(text):
-    """统计套话命中。返回 {hits: {短语: 次数}, per_kilo: 每千字密度, alert: bool}。"""
-    text = text or ""
-    total = len(text)
-    hits = {}
-    for p in AI_PHRASES:
-        p = p.strip()
-        if not p:
-            continue
-        n = text.count(p)
-        if n:
-            hits[p] = n
-    n_hits = sum(hits.values())
-    per_kilo = round(n_hits * 1000.0 / total, 2) if total else 0.0
-    return {"hits": hits, "per_kilo": per_kilo, "alert": per_kilo >= ALERT_PER_KILO}
-
-
-def narrative_analyze(text):
-    """统计叙事架构层信号。返回 {cats: {类: 次数}, per_kilo, alert, ending_hits}。
-
-    - 前两类全篇统计；「成长式收束」只统计末尾 ENDING_SCAN_CHARS 字
-      （中段的「和解/放下」是剧情词，结尾的才是成长式收束指纹）。
-    - per_kilo 为三类合计密度；alert 判据：合计 ≥ 告警线 或 收束类命中 ≥ 2。
-    """
-    text = text or ""
-    total = len(text)
-    cats, ending_hits = {}, 0
-    if total:
-        tail = text[-ENDING_SCAN_CHARS:]
-        for cat, phrases in NARRATIVE_TELLS.items():
-            n = 0
-            for p in phrases:
-                if cat == "成长式收束":
-                    n += tail.count(p)
-                else:
-                    n += text.count(p)
-            if n:
-                cats[cat] = n
-        ending_hits = cats.get("成长式收束", 0)
-    per_kilo = round(sum(cats.values()) * 1000.0 / total, 2) if (total and cats) else 0.0
-    alert = bool(cats) and (per_kilo >= NARRATIVE_ALERT_PER_KILO or ending_hits >= 2)
-    return {"cats": cats, "per_kilo": per_kilo, "alert": alert, "ending_hits": ending_hits}
-
-
-# ---------- 节奏/钩子密度（借鉴 web-novel-pacing-analyzer：连载作品的章级
-# 节奏审计——读者留存看「每章有没有钩子、段落有没有淤积、是不是通篇独白」，
-# 这些是能确定性统计的客观量。与套话检测同路：只给评审官一条参考线，
-# 不直接扣分——短段落多的快节奏文也可能开篇没有冲突词。）
-NARRATIVE_FLOWS = ("novel", "serial_novel", "video_script")
-
-# 冲突/悬念信号词表：开篇用来看「抓不抓人」，章末用来看「留不留钩子」
-HOOK_WORDS = (
-    "突然", "忽然", "竟然", "没想到", "万万没想到", "难道", "究竟", "为什么",
-    "是谁", "怎么回事", "就在这时", "就在此时", "然而", "可是", "却见", "却听",
-    "不对劲", "不对！", "脸色一变", "猛地", "陡然", "下一秒", "醒来", "睁开眼",
-    "电话响", "门被推开", "传来", "砸", "吼", "喊",
-)
-DIALOG_RE = re.compile(r"[「『“\"]")
-OPENING_SCAN_CHARS = 300        # 开篇抓力只看前 300 字（黄金三章 / 短视频黄金 3 秒）
-PACING_TAIL_CHARS = 200         # 章末钩子只看最后 200 字
-PARA_LONG_CHARS = 400           # 单段超过此字数算「信息淤积」
-PARA_AVG_ALERT = 300            # 段均长告警线
-DIALOG_ALERT_RATIO = 0.12       # 含对话段占比低于此值算「通篇叙述独白」
-PACING_MIN_CHARS = 400          # 太短（摘要/片段）不做节奏统计，避免噪声
-
-
-def _paras(text):
-    """切段：空行或单换行都算分隔；丢掉标题行（# 开头）与空白。"""
-    out = []
-    for p in re.split(r"\n\s*\n|\n", text or ""):
-        p = p.strip()
-        if p and not p.startswith("#"):
-            out.append(p)
-    return out
-
-
-def pacing_analyze(text):
-    """统计叙事节奏客观量。返回 {ok, paras, avg_len, longest, dialog_ratio,
-    open_hits, tail_hits, alerts, alert}；文本过短时 ok=False。"""
-    text = text or ""
-    paras = _paras(text)
-    if len(text) < PACING_MIN_CHARS or not paras:
-        return {"ok": False}
-    lens = [len(p) for p in paras]
-    avg = round(sum(lens) * 1.0 / len(lens), 1)
-    longest = max(lens)
-    dialog = sum(1 for p in paras if DIALOG_RE.search(p))
-    ratio = round(dialog * 1.0 / len(paras), 3)
-    opening, tail = text[:OPENING_SCAN_CHARS], text[-PACING_TAIL_CHARS:]
-    open_hits = sum(1 for w in HOOK_WORDS if w in opening)
-    tail_hits = sum(1 for w in HOOK_WORDS if w in tail)
-    alerts = []
-    if avg >= PARA_AVG_ALERT:
-        alerts.append("段落偏长（均 %.0f 字），阅读有淤积感" % avg)
-    if longest >= PARA_LONG_CHARS:
-        alerts.append("存在超长段（最长 %d 字）" % longest)
-    if ratio < DIALOG_ALERT_RATIO:
-        alerts.append("对话段占比偏低（%.0f%%），易成通篇独白" % (ratio * 100))
-    if not tail_hits:
-        alerts.append("章末 200 字无悬念信号，缺钩子")
-    if not open_hits:
-        alerts.append("开篇 300 字无冲突信号，开头偏平")
-    return {"ok": True, "paras": len(paras), "avg_len": avg, "longest": longest,
-            "dialog_ratio": ratio, "open_hits": open_hits, "tail_hits": tail_hits,
-            "alerts": alerts, "alert": bool(alerts)}
-
-
-def inject_into_prompt(tpl, text, kind=None):
-    """把确定性检测块插进评审提示词，返回新提示词（无命中原样返回）。
-
-    位置固定为「## 待评审稿件」之前——与技能包/圣经/知识库注入同一位置，
-    保持前缀稳定（§07 T1.2' 前缀缓存）。单稿件评审与连载逐章评审共用本
-    函数：此前连载逐章完全没有这条参考线（旗舰场景漏挂）。
-    """
-    det = report_line(text, kind)
-    if not det:
-        return tpl
-    return tpl.replace("## 待评审稿件",
-                       "## 确定性检测结果（供评审参考）\n%s\n\n## 待评审稿件" % det, 1)
-
-
-def report_line(text, kind=None):
-    """生成注入评审提示词的报告行（措辞层 + 叙事架构层 + 节奏层）；
-    全部无命中返回空串。kind 为流程类型：只有叙事类流程才下发节奏检测
-    （邮件/汇报/翻译等非叙事文体统计段落节奏没有意义）。"""
-    lines = []
-    r = analyze(text)
-    if r["hits"]:
-        top = "、".join("「%s」×%d" % (k, v)
-                        for k, v in sorted(r["hits"].items(), key=lambda x: -x[1])[:8])
-        line = "- [AI味检测] 确定性统计：套话密度 %.1f/千字（%s）。" % (r["per_kilo"], top)
-        if r["alert"]:
-            line += "密度超过告警线 %.0f/千字，请重点评审译制腔与套话问题。" % ALERT_PER_KILO
-        lines.append(line)
-    nr = narrative_analyze(text)
-    if nr["cats"]:
-        top = "、".join("「%s」×%d" % (k, v) for k, v in nr["cats"].items())
-        line = ("- [叙事架构信号] 确定性统计（措辞改写不掉的架构级指纹）："
-                "%s，合计 %.1f/千字。" % (top, nr["per_kilo"]))
-        if nr["alert"]:
-            line += ("出现架构级 AI 指纹（顿悟说教/情绪只写身体反应/成长式收束），"
-                     "请评审情节结构：主题是否被叙述者直接说破、情绪是否只有身体描写、"
-                     "结尾是否靠主角想通收束。")
-        lines.append(line)
-    if kind in NARRATIVE_FLOWS:
-        pr = pacing_analyze(text)
-        if pr["ok"]:
-            line = ("- [节奏检测] 确定性统计：%d 段（均 %.0f 字 / 最长 %d 字），"
-                    "含对话段 %.0f%%，开篇 300 字冲突信号 %d 处，章末 200 字悬念信号 %d 处。"
-                    % (pr["paras"], pr["avg_len"], pr["longest"],
-                       pr["dialog_ratio"] * 100, pr["open_hits"], pr["tail_hits"]))
-            if pr["alert"]:
-                line += ("提示：%s。连载留存取决于每章的钩子与段落节奏，"
-                         "请评审官结合剧情判断是否真问题。" % "、".join(pr["alerts"]))
-            lines.append(line)
-    return "\n".join(lines)
+# -*- coding: utf-8 -*-
+"""AI 味确定性检测（借鉴 oh-story-claudecode 的去AI味思路）。
+
+模型评审是主观分，这里补一道客观参考线：统计译制腔/套话高频词的
+密度，产出报告行注入评审提示词，让评审官把「套话密度」纳入评分
+参考。只做参考不做门禁——「仿佛」「缓缓」在正常小说里也常见，
+误伤风险由评审官（能读懂上下文）兜底，脚本只负责指认位置与密度。
+"""
+from __future__ import annotations
+
+import re
+
+# 译制腔/套话词表（按误导性排序，持续运维）
+AI_PHRASES = (
+    "空气仿佛凝固", "眸中闪过一丝", "嘴角勾起一抹", "心中一动",
+    "值得注意的是", "总而言之", "综上所述", "不难发现",
+    "在这一刻", "仿佛在诉说着", "无声地诉说着", "见证着",
+    "深深地看了一眼", "似乎在", "空气仿佛", "一抹微笑",
+    "不禁", "仿佛",
+)
+
+# 密度告警线（每千字命中次数）：超过即提示评审官重点关注
+ALERT_PER_KILO = 8.0
+
+# 叙事架构层信号（借鉴 sepia 2.7k★/StoryScope 研究 2026：AI 小说 93.2% 靠
+# 叙事架构特征检出，人工改写措辞后检出率仅从 95.5% 降到 93.9%——措辞层
+# 改不掉的架构级指纹才是真破绽）。三类可确定性检测的架构信号：
+NARRATIVE_TELLS = {
+    "顿悟说教": ("终于明白", "这才明白", "明白了，", "意识到，自己", "懂得了",
+               "原来，成长", "原来，生活", "原来，所谓"),
+    "情绪身体化": ("心脏猛地", "指尖冰凉", "指尖发凉", "喉咙发紧", "喉头发紧",
+                "胃里一阵", "胃部一阵", "后背一凉", "血液仿佛", "呼吸一滞"),
+    "成长式收束": ("释然", "和解", "放下了", "接纳了", "与自己和解", "轻轻松了口气",
+                "内心归于平静"),
+}
+# 架构信号告警线比措辞层低：这些表达在好小说里本就该稀缺
+NARRATIVE_ALERT_PER_KILO = 2.0
+# 「成长式收束」只在结尾才构成架构指纹（中段出现多半是剧情词），只扫尾部
+ENDING_SCAN_CHARS = 600
+
+
+def analyze(text):
+    """统计套话命中。返回 {hits: {短语: 次数}, per_kilo: 每千字密度, alert: bool}。"""
+    text = text or ""
+    total = len(text)
+    hits = {}
+    for p in AI_PHRASES:
+        p = p.strip()
+        if not p:
+            continue
+        n = text.count(p)
+        if n:
+            hits[p] = n
+    n_hits = sum(hits.values())
+    per_kilo = round(n_hits * 1000.0 / total, 2) if total else 0.0
+    return {"hits": hits, "per_kilo": per_kilo, "alert": per_kilo >= ALERT_PER_KILO}
+
+
+def narrative_analyze(text):
+    """统计叙事架构层信号。返回 {cats: {类: 次数}, per_kilo, alert, ending_hits}。
+
+    - 前两类全篇统计；「成长式收束」只统计末尾 ENDING_SCAN_CHARS 字
+      （中段的「和解/放下」是剧情词，结尾的才是成长式收束指纹）。
+    - per_kilo 为三类合计密度；alert 判据：合计 ≥ 告警线 或 收束类命中 ≥ 2。
+    """
+    text = text or ""
+    total = len(text)
+    cats, ending_hits = {}, 0
+    if total:
+        tail = text[-ENDING_SCAN_CHARS:]
+        for cat, phrases in NARRATIVE_TELLS.items():
+            n = 0
+            for p in phrases:
+                if cat == "成长式收束":
+                    n += tail.count(p)
+                else:
+                    n += text.count(p)
+            if n:
+                cats[cat] = n
+        ending_hits = cats.get("成长式收束", 0)
+    per_kilo = round(sum(cats.values()) * 1000.0 / total, 2) if (total and cats) else 0.0
+    alert = bool(cats) and (per_kilo >= NARRATIVE_ALERT_PER_KILO or ending_hits >= 2)
+    return {"cats": cats, "per_kilo": per_kilo, "alert": alert, "ending_hits": ending_hits}
+
+
+# ---------- 节奏/钩子密度（借鉴 web-novel-pacing-analyzer：连载作品的章级
+# 节奏审计——读者留存看「每章有没有钩子、段落有没有淤积、是不是通篇独白」，
+# 这些是能确定性统计的客观量。与套话检测同路：只给评审官一条参考线，
+# 不直接扣分——短段落多的快节奏文也可能开篇没有冲突词。）
+NARRATIVE_FLOWS = ("novel", "serial_novel", "video_script")
+
+# 冲突/悬念信号词表：开篇用来看「抓不抓人」，章末用来看「留不留钩子」
+HOOK_WORDS = (
+    "突然", "忽然", "竟然", "没想到", "万万没想到", "难道", "究竟", "为什么",
+    "是谁", "怎么回事", "就在这时", "就在此时", "然而", "可是", "却见", "却听",
+    "不对劲", "不对！", "脸色一变", "猛地", "陡然", "下一秒", "醒来", "睁开眼",
+    "电话响", "门被推开", "传来", "砸", "吼", "喊",
+)
+DIALOG_RE = re.compile(r"[「『“\"]")
+OPENING_SCAN_CHARS = 300        # 开篇抓力只看前 300 字（黄金三章 / 短视频黄金 3 秒）
+PACING_TAIL_CHARS = 200         # 章末钩子只看最后 200 字
+PARA_LONG_CHARS = 400           # 单段超过此字数算「信息淤积」
+PARA_AVG_ALERT = 300            # 段均长告警线
+DIALOG_ALERT_RATIO = 0.12       # 含对话段占比低于此值算「通篇叙述独白」
+PACING_MIN_CHARS = 400          # 太短（摘要/片段）不做节奏统计，避免噪声
+
+
+def _paras(text):
+    """切段：空行或单换行都算分隔；丢掉标题行（# 开头）与空白。"""
+    out = []
+    for p in re.split(r"\n\s*\n|\n", text or ""):
+        p = p.strip()
+        if p and not p.startswith("#"):
+            out.append(p)
+    return out
+
+
+def pacing_analyze(text):
+    """统计叙事节奏客观量。返回 {ok, paras, avg_len, longest, dialog_ratio,
+    open_hits, tail_hits, alerts, alert}；文本过短时 ok=False。"""
+    text = text or ""
+    paras = _paras(text)
+    if len(text) < PACING_MIN_CHARS or not paras:
+        return {"ok": False}
+    lens = [len(p) for p in paras]
+    avg = round(sum(lens) * 1.0 / len(lens), 1)
+    longest = max(lens)
+    dialog = sum(1 for p in paras if DIALOG_RE.search(p))
+    ratio = round(dialog * 1.0 / len(paras), 3)
+    opening, tail = text[:OPENING_SCAN_CHARS], text[-PACING_TAIL_CHARS:]
+    open_hits = sum(1 for w in HOOK_WORDS if w in opening)
+    tail_hits = sum(1 for w in HOOK_WORDS if w in tail)
+    alerts = []
+    if avg >= PARA_AVG_ALERT:
+        alerts.append("段落偏长（均 %.0f 字），阅读有淤积感" % avg)
+    if longest >= PARA_LONG_CHARS:
+        alerts.append("存在超长段（最长 %d 字）" % longest)
+    if ratio < DIALOG_ALERT_RATIO:
+        alerts.append("对话段占比偏低（%.0f%%），易成通篇独白" % (ratio * 100))
+    if not tail_hits:
+        alerts.append("章末 200 字无悬念信号，缺钩子")
+    if not open_hits:
+        alerts.append("开篇 300 字无冲突信号，开头偏平")
+    return {"ok": True, "paras": len(paras), "avg_len": avg, "longest": longest,
+            "dialog_ratio": ratio, "open_hits": open_hits, "tail_hits": tail_hits,
+            "alerts": alerts, "alert": bool(alerts)}
+
+
+def inject_into_prompt(tpl, text, kind=None):
+    """把确定性检测块插进评审提示词，返回新提示词（无命中原样返回）。
+
+    位置固定为「## 待评审稿件」之前——与技能包/圣经/知识库注入同一位置，
+    保持前缀稳定（§07 T1.2' 前缀缓存）。单稿件评审与连载逐章评审共用本
+    函数：此前连载逐章完全没有这条参考线（旗舰场景漏挂）。
+    """
+    det = report_line(text, kind)
+    if not det:
+        return tpl
+    return tpl.replace("## 待评审稿件",
+                       "## 确定性检测结果（供评审参考）\n%s\n\n## 待评审稿件" % det, 1)
+
+
+def report_line(text, kind=None):
+    """生成注入评审提示词的报告行（措辞层 + 叙事架构层 + 节奏层）；
+    全部无命中返回空串。kind 为流程类型：只有叙事类流程才下发节奏检测
+    （邮件/汇报/翻译等非叙事文体统计段落节奏没有意义）。"""
+    lines = []
+    r = analyze(text)
+    if r["hits"]:
+        top = "、".join("「%s」×%d" % (k, v)
+                        for k, v in sorted(r["hits"].items(), key=lambda x: -x[1])[:8])
+        line = "- [AI味检测] 确定性统计：套话密度 %.1f/千字（%s）。" % (r["per_kilo"], top)
+        if r["alert"]:
+            line += "密度超过告警线 %.0f/千字，请重点评审译制腔与套话问题。" % ALERT_PER_KILO
+        lines.append(line)
+    nr = narrative_analyze(text)
+    if nr["cats"]:
+        top = "、".join("「%s」×%d" % (k, v) for k, v in nr["cats"].items())
+        line = ("- [叙事架构信号] 确定性统计（措辞改写不掉的架构级指纹）："
+                "%s，合计 %.1f/千字。" % (top, nr["per_kilo"]))
+        if nr["alert"]:
+            line += ("出现架构级 AI 指纹（顿悟说教/情绪只写身体反应/成长式收束），"
+                     "请评审情节结构：主题是否被叙述者直接说破、情绪是否只有身体描写、"
+                     "结尾是否靠主角想通收束。")
+        lines.append(line)
+    if kind in NARRATIVE_FLOWS:
+        pr = pacing_analyze(text)
+        if pr["ok"]:
+            line = ("- [节奏检测] 确定性统计：%d 段（均 %.0f 字 / 最长 %d 字），"
+                    "含对话段 %.0f%%，开篇 300 字冲突信号 %d 处，章末 200 字悬念信号 %d 处。"
+                    % (pr["paras"], pr["avg_len"], pr["longest"],
+                       pr["dialog_ratio"] * 100, pr["open_hits"], pr["tail_hits"]))
+            if pr["alert"]:
+                line += ("提示：%s。连载留存取决于每章的钩子与段落节奏，"
+                         "请评审官结合剧情判断是否真问题。" % "、".join(pr["alerts"]))
+            lines.append(line)
+    return "\n".join(lines)

package/app/core/analytics.py

@@ -1,38 +1,38 @@
-"""Task cost/quality rollups for operator views."""
-from __future__ import annotations
-
-from collections import defaultdict
-
-
-def summary(store):
-    by_task = defaultdict(lambda: {"runs": 0, "done": 0, "failed": 0,
-                                   "tokens": 0, "cost_usd": 0.0, "quality_sum": 0.0,
-                                   "quality_count": 0})
-    for run in store.list_runs(limit=None):
-        tid = run.get("task_id") or "_unowned"
-        row = by_task[tid]
-        row["runs"] += 1
-        row["done"] += int(run.get("status") == "done")
-        row["failed"] += int(run.get("status") == "failed")
-        row["tokens"] += int(run.get("tokens") or 0)
-        row["cost_usd"] += float(run.get("cost_usd") or 0)
-        verdict = run.get("verdict") or {}
-        quality = verdict.get("overall")
-        if isinstance(quality, (int, float)):
-            row["quality_sum"] += float(quality)
-            row["quality_count"] += 1
-    rows = []
-    for task_id, row in by_task.items():
-        item = dict(row)
-        item["task_id"] = task_id
-        item["cost_usd"] = round(item["cost_usd"], 6)
-        item["quality_avg"] = round(item.pop("quality_sum") / item.pop("quality_count"), 2) \
-            if item["quality_count"] else None
-        item.pop("quality_count", None)
-        rows.append(item)
-    rows.sort(key=lambda x: (-x["cost_usd"], x["task_id"]))
-    return {"tasks": rows, "totals": {
-        "runs": sum(x["runs"] for x in rows),
-        "tokens": sum(x["tokens"] for x in rows),
-        "cost_usd": round(sum(x["cost_usd"] for x in rows), 6),
-    }}
+"""Task cost/quality rollups for operator views."""
+from __future__ import annotations
+
+from collections import defaultdict
+
+
+def summary(store):
+    by_task = defaultdict(lambda: {"runs": 0, "done": 0, "failed": 0,
+                                   "tokens": 0, "cost_usd": 0.0, "quality_sum": 0.0,
+                                   "quality_count": 0})
+    for run in store.list_runs(limit=None):
+        tid = run.get("task_id") or "_unowned"
+        row = by_task[tid]
+        row["runs"] += 1
+        row["done"] += int(run.get("status") == "done")
+        row["failed"] += int(run.get("status") == "failed")
+        row["tokens"] += int(run.get("tokens") or 0)
+        row["cost_usd"] += float(run.get("cost_usd") or 0)
+        verdict = run.get("verdict") or {}
+        quality = verdict.get("overall")
+        if isinstance(quality, (int, float)):
+            row["quality_sum"] += float(quality)
+            row["quality_count"] += 1
+    rows = []
+    for task_id, row in by_task.items():
+        item = dict(row)
+        item["task_id"] = task_id
+        item["cost_usd"] = round(item["cost_usd"], 6)
+        item["quality_avg"] = round(item.pop("quality_sum") / item.pop("quality_count"), 2) \
+            if item["quality_count"] else None
+        item.pop("quality_count", None)
+        rows.append(item)
+    rows.sort(key=lambda x: (-x["cost_usd"], x["task_id"]))
+    return {"tasks": rows, "totals": {
+        "runs": sum(x["runs"] for x in rows),
+        "tokens": sum(x["tokens"] for x in rows),
+        "cost_usd": round(sum(x["cost_usd"] for x in rows), 6),
+    }}