codebee 0.1.10 → 0.1.11

package/app/core/pipeline.py

@@ -19,7 +19,7 @@ import re
 import threading
 import time
 
-from . import catalog, history, jobs, manager, modelhub, mocks, planner, registry, router, runner, skills, store, usage
+from . import aiflavor, catalog, history, jobs, manager, modelhub, mocks, planner, registry, router, runner, skills, store, usage
 from . import builtin_agent
 from . import diagnostics
 from . import paths as paths_mod
@@ -730,6 +730,118 @@ def _format_issues(review_json, verify_pass, verify_failed_note):
 
 # ---------------------------------------------------------------- 代码流水线
 
+def _shortstat_files(diffstat_line):
+    """` 3 files changed, 10 insertions(+)` → 文件数（解析失败给大数排最后）。"""
+    import re as _re
+    m = _re.search(r"(\d+) files? changed", diffstat_line or "")
+    return int(m.group(1)) if m else 9999
+
+
+def _code_bestof(run, task, impl, difficulty, ev):
+    """代码任务 Best-of-N（借鉴 orca 并行 worktree 择优）。启用条件：best_of≥2、
+    无续会话、非手动指定实现者、工作目录是 git 仓库。
+
+    按基线（git_rev 或 HEAD）建 N 个临时 worktree → 各路并行跑实现子任务链 →
+    各跑验证命令 → 择优（验证通过 > 改动文件更少）→ 胜者 diff 应用回主工作区。
+    任一环节失败一律返回 False，由调用方回落原单路实现（绝不因赛马挡任务）。"""
+    from . import gitmod
+    run_id = run["id"]
+    n = max(2, min(3, int(task.get("best_of") or 1)))
+    wd = task["workdir"]
+    if not gitmod._git(wd, "rev-parse", "--is-inside-work-tree")["ok"]:
+        return False
+    base = task.get("git_rev") or ""
+    if base and not gitmod.valid_rev(base):
+        return False
+    base_rev = (gitmod._git(wd, "rev-parse", base)["stdout"].strip() if base
+                else gitmod._git(wd, "rev-parse", "HEAD")["stdout"].strip())
+    if not base_rev:
+        return False
+    wts = []
+    results = {}
+
+    def _cleanup():
+        for k, br, wt_path in wts:
+            gitmod._git(wd, "worktree", "remove", "--force", wt_path, timeout=60)
+            gitmod._git(wd, "branch", "-D", br, timeout=60)
+
+    try:
+        import tempfile
+        race_root = os.path.join(tempfile.gettempdir(), "codebee-race", run_id)
+        os.makedirs(race_root, exist_ok=True)
+        for k in range(n):
+            br = "codebee-race-%s-%d" % (run_id, k)
+            wt_path = os.path.join(race_root, "v%d" % k)
+            r = gitmod._git(wd, "worktree", "add", "--detach", wt_path, base_rev, timeout=120)
+            if not r["ok"]:
+                raise RuntimeError(r["stderr"] or "worktree add 失败")
+            wts.append((k, br, wt_path))
+
+        subtasks = [{"title": task.get("title") or task.get("goal") or "实现",
+                     "detail": task.get("goal") or ""}]
+
+        def _race_one(k, wt_path):
+            agt_b = modelhub.bind_agent(impl, difficulty)
+            res = None
+            for i, sub in enumerate(subtasks):
+                prompt = (CODE_IMPL_PROMPT
+                          .replace("__GOAL__", task["goal"])
+                          .replace("__SUBTASK__", sub["detail"] if sub["detail"] else sub["title"])
+                          .replace("__CONTEXT__", task.get("context") or "（无）")
+                          .replace("__VERIFY_HINT__", _verify_hint(task)))
+                res = _run_step(run_id, "race%d-impl" % k, agt_b, prompt, wt_path,
+                                readonly=False, ev=ev,
+                                note="赛马路 %d/%d（worktree 隔离）" % (k + 1, n),
+                                images=_task_images(task, wd), require_tools=True)
+                if not res["ok"]:
+                    break
+            v_pass = bool(res and res["ok"])
+            if v_pass and task.get("verify_command"):
+                r = runner.run_process(shell_cmd=task["verify_command"], cwd=wt_path,
+                                       timeout=600, cancel_event=ev)
+                v_pass = bool(r["ok"])
+            diffstat = (gitmod._git(wt_path, "diff", "--shortstat", base_rev)["stdout"] or "").strip()
+            results[k] = {"ok": bool(res and res["ok"]), "verify": v_pass,
+                          "wt": wt_path, "diffstat": diffstat}
+
+        threads = []
+        for k, br, wt_path in wts:
+            th = threading.Thread(target=_race_one, args=(k, wt_path),
+                                  name="codebestof-%s-%d" % (run_id, k), daemon=True)
+            threads.append(th)
+            th.start()
+        for th in threads:
+            th.join(3600)
+        _check_cancel(ev)
+
+        usable = [r for k, r in sorted(results.items()) if r["ok"] and r["verify"]]
+        if not usable:
+            return False   # 全败：回落单路实现（保持原有换将/报错行为）
+        usable.sort(key=lambda r: _shortstat_files(r["diffstat"]))
+        winner = usable[0]
+        diff = gitmod._git(winner["wt"], "diff", base_rev, timeout=120)
+        if diff["ok"] and diff["stdout"].strip():
+            apply_r = runner.run_process(
+                argv=["git", "apply", "--whitespace=nowarn"],
+                cwd=wd, timeout=60, stdin_text=diff["stdout"])
+            if not apply_r["ok"]:
+                return False
+        store.update_run(run_id, bestof={
+            "kind": "code-worktree", "candidates": len(results),
+            "winner_files": _shortstat_files(winner["diffstat"]),
+            "diffstat": winner["diffstat"],
+        })
+        return True
+    except Exception:
+        log.warning("code bestof aborted run=%s", run_id, exc_info=True)
+        return False
+    finally:
+        try:
+            _cleanup()
+        except Exception:
+            pass
+
+
 def _run_code(run, task, agents, ev, stats, mode):
     run_id = run["id"]
     workdir = task["workdir"]
@@ -842,6 +954,13 @@ def _run_code(run, task, agents, ev, stats, mode):
         ok, res = _run_one(impl_agent)
         if ok:
             return True
+        # 单路实现失败：Best-of-N 赛马兜底（借鉴 orca worktree 择优）——多路并行
+        # 各自 worktree 隔离实现+验证，胜者 diff 回主工作区；失败回落走换将
+        if (mode == "auto" and impl_agent.get("mode") == "real"
+                and max(1, min(3, int(task.get("best_of") or 1))) >= 2
+                and resume_ctx is None
+                and _code_bestof(run, task, impl_agent, difficulty, ev)):
+            return True
         # 实现步失败不立刻判死：2026-09-16 实测配额烧干时 5 连跑全在同一条 CLI 上
         # 失败收场，而健康的 opencode 一直在旁观望——跨 CLI 换将重试一次
         # （mode=manual 尊重用户指定，不换）。
@@ -957,6 +1076,7 @@ def _run_code(run, task, agents, ev, stats, mode):
         lines.append("（无）")
     lines += ["", "## 评审总评", "", review_json.get("summary", ""), ""]
     store.write_report(run_id, "\n".join(lines))
+    _write_task_evidence(run_id, task, workdir, _evidence_lines_from_run(run_id, task))
     store.update_run(run_id, status="done", verdict=verdict,
                      summary="代码任务%s（验证%s / 评审%s%s）" % (
                          "通过" if overall_pass else "未通过",
@@ -1224,6 +1344,7 @@ def _run_direct(run, task, agents, ev, stats, mode):
     if last_text:
         report += ["## 最近一轮输出", "", last_text[-5000:], ""]
     store.write_report(run_id, "\n".join(report))
+    _write_task_evidence(run_id, task, workdir, _evidence_lines_from_run(run_id, task))
     store.update_run(run_id, status="done", verdict=verdict,
                      summary="直连完成（%d 轮）：%s" % (turns, last_text[:160]),
                      ended_at=_now())
@@ -1292,6 +1413,8 @@ def _tpl(task, key, default):
 
 BIBLE_FILE = "story-bible.md"
 _BIBLE_MAX_CHARS = 20000
+MODULES_FILE = "plot-modules.md"
+_MODULES_MAX_CHARS = 12000
 
 # 评审视角播种（dev-3.0 式 bug hunters）：N 个评审各领一个深挖镜头，
 # 避免全员盯着同一处。按评审序号取模分配——同一评审每轮同一镜头，
@@ -1321,6 +1444,23 @@ def _story_bible(workdir):
             "与其冲突处以圣经为准）\n\n" + txt)
 
 
+def _plot_modules(workdir):
+    """剧情模块库（oh-story 拆文沉淀式）：工作目录里的 plot-modules.md
+    （可复用的桥段/冲突/爽点/名场面素材模块），作者手工维护，每章起草与
+    评审前自动注入。不存在/为空返回 ""——约定式功能，零配置零噪音。"""
+    p = os.path.abspath(os.path.join(str(workdir or ""), MODULES_FILE))
+    if not _inside(workdir, p) or not os.path.isfile(p):
+        return ""
+    try:
+        txt = _read_text_any_enc(p)[:_MODULES_MAX_CHARS].strip()
+    except OSError:
+        return ""
+    if not txt:
+        return ""
+    return ("## 剧情模块库（plot-modules.md：可复用的桥段/冲突/爽点素材模块，"
+            "鼓励化用，不要照抄原句）\n\n" + txt)
+
+
 def _critic_lens(critics, agent):
     """该评审的专属视角；单评审/手动指定时不播种（无从轮换，也别稀释注意力）。"""
     try:
@@ -1521,6 +1661,11 @@ def _run_serial_review(run, task, agents, ev, stats, mode, critics, impl, route,
     start = int(serial.get("start_chapter") or 1)
     # 故事圣经：工作目录里的 story-bible.md，整个 run 内字节稳定（前缀缓存友好）
     bible = _story_bible(workdir)
+    # 剧情模块库：plot-modules.md（拆文沉淀的可复用素材模块）并入同一注入块，
+    # 同样要求 run 内字节稳定；无模块库时零噪音
+    mods = _plot_modules(workdir)
+    if mods:
+        bible = (bible + "\n\n" + mods) if bible else mods
 
 
     def crit_prompt_for(text, note=""):
@@ -1689,6 +1834,7 @@ def _run_serial_review(run, task, agents, ev, stats, mode, critics, impl, route,
         critic_sids = {}   # §07 T1.1：每评审的会话 id（第 2 轮复用，前缀走缓存读）
         race_cj = None     # 变体赛马已评审胜者：直接作为第 1 轮结果，不重评
         race_scored = 0
+        race_losers = []   # 赛马败稿文本（收卷前留存）：首轮修订时提炼败者精华
 
         reuse = i in done_set and os.path.exists(os.path.join(workdir, ch_file))
         if reuse:
@@ -1914,6 +2060,13 @@ def _run_serial_review(run, task, agents, ev, stats, mode, critics, impl, route,
                                          error="第 %d 章赛马收卷失败: %r" % (i, e), ended_at=_now())
                         return
                 for v in scored_variants[1:]:
+                    # 败者精华：删除前留存文本，供首轮修订提炼（懒调用）
+                    try:
+                        race_losers.append({
+                            "variant": v["variant"],
+                            "text": (_read_text_any_enc(os.path.join(workdir, v["file"])) or "")[:8000]})
+                    except OSError:
+                        pass
                     try:
                         os.remove(os.path.join(workdir, v["file"]))
                     except OSError:
@@ -1966,6 +2119,27 @@ def _run_serial_review(run, task, agents, ev, stats, mode, critics, impl, route,
             crit_lines = ["- %s：%.1f（章阈值 %.1f）" % (d, means[d], threshold_ch) for d in dims]
             crit_lines += ["- [%s] %s" % (x.get("dim", "?"), str(x.get("note", ""))[:140])
                            for x in majors]
+            if rnd == 1 and race_losers and impl.get("mode") == "real":
+                # 赛马败者精华回收（连载版）：一次选择器调用提炼落选稿优点，
+                # 与评审意见一并喂给首轮修订；失败静默不影响修订
+                try:
+                    cands = "### 胜者稿（已选定）\n\n%s" % _read_chapter(workdir, i)[:8000]
+                    for li, lo in enumerate(race_losers, 1):
+                        cands += "\n\n### 落选稿 %d\n\n%s" % (li, lo["text"])
+                    sel_prompt = (BESTOF_SELECTOR_PROMPT
+                                  .replace("__GOAL__", task["goal"])
+                                  .replace("__CANDIDATES__", cands))
+                    sel = _run_step(run_id, "race-select-c%d" % i,
+                                    modelhub.bind_agent(critics[0] if critics else impl, difficulty),
+                                    sel_prompt, step_wd, readonly=True, ev=ev,
+                                    note="赛马败者精华提炼")
+                    _selcj = runner.extract_json(sel.get("text") or "")
+                    _imp = str((_selcj or {}).get("improvements") or "").strip()
+                    if _imp:
+                        crit_lines.append("- [赛马精华] 终审从落选候选稿提炼出值得吸收的优点：%s"
+                                          % _imp[:400])
+                except Exception:
+                    pass
             if impl.get("mode") == "mock":
                 step, log_abs = store.add_step(run_id, "revise-c%d" % i, impl["id"],
                                                impl.get("label"))
@@ -2198,6 +2372,7 @@ def _run_serial_review(run, task, agents, ev, stats, mode, critics, impl, route,
     else:
         lines.append("（无 major 问题）")
     store.write_report(run_id, "\n".join(lines))
+    _write_task_evidence(run_id, task, workdir, _evidence_lines_from_run(run_id, task))
     store.update_run(run_id, status="done", verdict=verdict,
                      summary="连载任务%s（%s，约 %d 字，综合 %.1f）" % (
                          "达标" if publishable else "未达标", scope_txt,
@@ -2429,6 +2604,11 @@ def _run_content_review(run, task, agents, ev, stats, mode):
             _tpl(task, "critique_prompt", NOVEL_CRITIQUE_PROMPT))
             .replace("__DIMKEYS__", dimkey)
             .replace("__MANUSCRIPT__", manuscript or "（稿件为空！）"))
+        # AI 味确定性检测（借鉴 oh-story 去AI味）：客观参考线随评审下发，
+        # 命中才追加——评审官结合上下文判断是否真问题，脚本不直接扣分
+        _aiflavor_line = aiflavor.report_line(manuscript)
+        if _aiflavor_line:
+            crit_prompt += "\n\n## 确定性检测结果（供评审参考）\n" + _aiflavor_line
         for agent in critics:
             role = "critique-r%d" % r
             if agent.get("mode") == "mock":
@@ -2538,6 +2718,7 @@ def _run_content_review(run, task, agents, ev, stats, mode):
         lines.append("（无）")
     lines += ["", "## 稿件位置", "", "`%s`" % ms_path, ""]
     store.write_report(run_id, "\n".join(lines))
+    _write_task_evidence(run_id, task, workdir, _evidence_lines_from_run(run_id, task))
     store.update_run(run_id, status="done", verdict=verdict,
                      summary="评审任务%s（综合 %.1f）" % ("达标" if publishable else "未达标", overall),
                      ended_at=_now())
@@ -2593,6 +2774,103 @@ def _run_serial_qa(run, task, agents, ev):
                      error="答疑失败（执行/评审链不可用）——" + "；".join(errors[-3:]))
 
 
+def _write_task_spec(task, workdir):
+    """任务规格落盘 .codebee/spec.md（借鉴 agent-orchestrator 的 .spec/PROMPT.md 与
+    planning-with-files 的文件化计划）：任务定义随工作目录留存、随任务分支版本化，
+    追话/复盘/续跑时可见原始意图。失败静默返回空串——规格文件永远不能挡住任务执行。"""
+    try:
+        from pathlib import Path as _P
+        root = _P(workdir).resolve()
+        target = (root / ".codebee" / "spec.md").resolve()
+        if root not in target.parents:   # 守卫：spec 必须落在工作目录内（../、symlink 出逃弃写）
+            return ""
+        os.makedirs(str(target.parent), exist_ok=True)
+        lines = [
+            "# 任务规格", "",
+            "- 标题：%s" % (task.get("title") or ""),
+            "- 类型：%s" % (task.get("type") or ""),
+            "- 创建：%s" % (task.get("created_at") or ""),
+            "- 目标：%s" % str(task.get("goal") or "").replace("\n", " "),
+        ]
+        if task.get("context"):
+            lines.append("- 背景：%s" % str(task["context"]).replace("\n", " "))
+        if task.get("difficulty"):
+            lines.append("- 难度：%s" % task["difficulty"])
+        if task.get("mode"):
+            lines.append("- 路由模式：%s" % task["mode"])
+        label = {"rounds": "评审轮数", "threshold": "发布阈值", "best_of": "赛马候选数"}
+        for k in ("rounds", "threshold", "best_of"):
+            if task.get(k) is not None:
+                lines.append("- %s：%s" % (label[k], task[k]))
+        if task.get("rubric"):
+            lines.append("- 评审维度：%s" % "、".join(task["rubric"]))
+        if task.get("serial"):
+            s = task["serial"]
+            lines.append("- 连载：%s 章 × %s 字（赛马变体 %s）"
+                         % (s.get("chapters"), s.get("words_per_chapter"), s.get("variants", 1)))
+        if task.get("verify_command"):
+            lines.append("- 验证命令：`%s`" % task["verify_command"])
+        if not str(target).startswith(str(root) + os.sep):   # sink 侧复检：路径必须仍在工作目录内
+            return ""
+        with open(str(target), "w", encoding="utf-8") as f:
+            f.write("\n".join(lines) + "\n")
+        return str(target)
+    except Exception:
+        return ""
+
+
+def _write_task_evidence(run_id, task, workdir, summary_lines):
+    """验证证据持久化（借鉴 gsd-pi 的 validation evidence）：run 收尾把确定性
+    证据追加进 .codebee/evidence.md——验证命令结果、评审分数、AI 味检测等。
+    与 spec.md（任务意图）呼应成「任务档案」；失败静默，绝不挡收尾。"""
+    if not summary_lines:
+        return ""
+    try:
+        spec_dir = os.path.join(workdir, ".codebee")
+        os.makedirs(spec_dir, exist_ok=True)
+        path = os.path.join(spec_dir, "evidence.md")
+        header_needed = not os.path.exists(path)
+        with open(path, "a", encoding="utf-8") as f:
+            if header_needed:
+                f.write("# 验证证据（每次运行追加一节）\n\n")
+            f.write("## %s · run %s\n\n" % (_now(), run_id))
+            for ln in summary_lines:
+                f.write("- %s\n" % str(ln)[:300])
+            f.write("\n")
+        return path
+    except Exception:
+        return ""
+
+
+def _evidence_lines_from_run(run_id, task):
+    """从 run 步骤与 verdict 提取证据行（确定性事实，不抄模型输出）。"""
+    run = store.get_run(run_id) or {}
+    lines = []
+    for s in run.get("steps") or []:
+        role = str(s.get("role") or "")
+        if role == "verify":
+            lines.append("验证命令 `%s` → %s%s" % (
+                task.get("verify_command") or "", s.get("status"),
+                "（%s）" % s.get("summary") if s.get("summary") else ""))
+    verdict = run.get("verdict") or {}
+    if verdict:
+        means = verdict.get("scores") or {}
+        if means:
+            lines.append("评审均分：%s（阈值 %s，%s）" % (
+                "、".join("%s %.1f" % (d, v) for d, v in means.items()),
+                verdict.get("threshold"),
+                "达标" if verdict.get("publishable") else "未达标"))
+        if verdict.get("overall") is not None:
+            lines.append("综合分 %.1f / %d 轮" % (verdict.get("overall") or 0.0,
+                                                  verdict.get("rounds_used") or 0))
+        bestof = run.get("bestof")
+        if bestof:
+            lines.append("赛马：%s（胜者 %s）" % (
+                bestof.get("kind") or "内容候选",
+                bestof.get("pick", bestof.get("winner_files", "?"))))
+    return lines
+
+
 def execute_run(run_id):
     run = store.get_run(run_id)
     if not run:
@@ -2627,6 +2905,9 @@ def execute_run(run_id):
         store.update_run(run_id, git=gitinfo)
         # 任务分支裁决状态：新一轮 run 产生新分支内容，重置回「待裁决」
         store.set_task_git_state(task["id"], "isolated")
+    # 任务规格文件化（借鉴 planning-with-files/agent-orchestrator）：任何任务都在
+    # 工作目录留一份 .codebee/spec.md——原始意图可见、随任务分支版本化
+    _write_task_spec(task, task["workdir"])
     agents = _agents()
     global _CURRENT_AGENTS
     _CURRENT_AGENTS = agents

package/app/core/publish/__init__.py

@@ -0,0 +1,7 @@
+# -*- coding: utf-8 -*-
+"""一键发布：CDP 驱动平台后台（番茄/七猫）建书与章节发布。
+
+分层：ws.py（最小 WebSocket）→ browser.py（CDP 驱动）→ ledger.py（台账）
+→ 平台流程（数据驱动的步骤表，selector 可外置覆盖）→ manager.py（状态机）。
+"""
+from . import browser, ledger  # noqa: F401