codebee 0.1.5 → 0.1.7

package/app/core/pipeline.py

@@ -273,18 +273,45 @@ def _wait_gate(run_id, ev):
         time.sleep(1.0)
 
 
+def _binding_dead_msg(agent):
+    """死链失败文案：委托 modelhub 单一真源（告警 sync 共用同一文案）。"""
+    try:
+        from . import modelhub
+        return modelhub.binding_dead_msg(agent.get("id") or "")
+    except Exception:
+        return ("绑定链全部失效，本步判失败、不回落 CLI 本机默认——"
+                "请在「CLI 绑定」页为该 CLI 绑定已启用的供应商")
+
+
 def _run_step(run_id, role, agent, prompt, workdir, readonly, ev, timeout=runner.DEFAULT_TIMEOUT, note="", resume=None, images=None, require_tools=False):
     """执行一个智能体步骤并记录。返回 runner 统一结果。"""
     _wait_gate(run_id, ev)
-    # 绑定解析为空 → CLI 将回落本机默认配置（用户配置的模型/供应商全部不生效）。
-    # 2026-09-16 实测：这种状态下烧干配额的本机默认供应商被静默使用，用户以为
-    # 在用自己配的模型。首次出现时在步骤备注里醒目标出。
-    if agent.get("mode") == "real" and not (agent.get("call_chain") or agent.get("env")):
-        note = ((note + "；") if note else "") + \
-               "⚠ 未解析到绑定链，本步回落 CLI 本机默认配置（请在模型接入页检查该 CLI 的供应商绑定）"
+    # 绑定解析为空分两种（2026-09-18 起 区分对待）：
+    # · 从没配过链（binding_configured=False）：回落 CLI 本机默认照跑——
+    #   用户根本没在 CodeBee 里配供应商，谈不到「烧自己配的配额」；判失败
+    #   反而把用本地登录的普通用户全挡在门外（0.1.6 真实装机误伤案例）。
+    # · 配过链但全死（binding_configured=True）：本步判失败不静默降级——
+    #   宁可失败不偷跑本机默认；auto 流程实现步的既有换将会接手健康 CLI。
+    #   只记在步骤备注/错误里，不再产生全局健康告警胶囊（同上案例：红胶囊
+    #   吓不到也帮不到普通用户，2026-09-18 用户拍板移除）。
+    dead_binding = (agent.get("mode") == "real"
+                    and agent.get("binding_configured")
+                    and not (agent.get("call_chain") or agent.get("env")))
+    dead_msg = _binding_dead_msg(agent) if dead_binding else ""
+    if dead_binding:
+        note = ((note + "；") if note else "") + "⚠ " + dead_msg
     step, log_abs = store.add_step(run_id, role, agent["id"],
                                    agent.get("label", agent["id"]), note=note)
     start = time.time()
+    if dead_binding:
+        from .error_codes import ErrorCode
+        res = {"ok": False, "text": "", "json": None, "cost_usd": 0.0,
+               "tokens": 0, "usage": None, "error": dead_msg,
+               "error_code": ErrorCode.ENV_BLOCK,
+               "raw": {"exit_code": None}, "kind": agent.get("kind", "generic"),
+               "model": agent.get("model")}
+        _finish_step_result(run_id, step, res, role, agent, start)
+        return res
     if agent.get("mode") == "mock":
         time.sleep(0.3)
         res = {"ok": True, "text": "[mock] %s" % prompt[:80], "json": None,
@@ -1601,12 +1628,19 @@ def _run_serial_review(run, task, agents, ev, stats, mode, critics, impl, route,
                 res = None
                 good = False
                 txt = ""
+                use_prompt = prompt
                 for draft_attempt in range(3):
                     if draft_attempt:
                         if ev is not None and ev.is_set():
                             break
                         time.sleep(30 * draft_attempt)   # 30s / 60s 退避
-                    res = _run_step(run_id, "draft-c%d" % i, modelhub.bind_agent(impl, difficulty), prompt,
+                    if draft_attempt and len(prompt) > 12000 and sk_block and sk_block in prompt:
+                        # 长提示词在容量受限通道（讯飞托管 35B 等）上会挂起/秒拒
+                        # ——降级重试：经验库块截到 4K 字，保留大纲/前情/本章要点
+                        # （2026-09-17 七猫实测：全量 30KB 对讯飞必挂）
+                        use_prompt = prompt.replace(
+                            sk_block, sk_block[:4000] + "\n\n（经验库已因通道容量限制精简）")
+                    res = _run_step(run_id, "draft-c%d" % i, modelhub.bind_agent(impl, difficulty), use_prompt,
                                     step_wd, readonly=False, ev=ev, timeout=2400,
                                     resume=resume_ctx["session"] if resume_ctx else None,
                                     images=_task_images(task, workdir),
@@ -1626,15 +1660,23 @@ def _run_serial_review(run, task, agents, ev, stats, mode, critics, impl, route,
                             break
                     if ev is not None and ev.is_set():
                         break
-                # 同作者重试穷尽 → 起草换将（连载不断档优先，风格差异交评审门与
-                # 后续 revise 拉回）——与代码流程换将同款语义（2026-09-17 c34 实测）
+                # 同作者重试穷尽 → 起草换将：按路由分序逐个试备选（最多 2 个，
+                # 只试一个会让第二名没机会——2026-09-17 c35 实测 opencode 顶在
+                # 前面，能干活的 kimi 永远轮不上）。连载不断档优先，风格差异交
+                # 评审门与后续 revise 拉回。
                 if not good:
-                    other, other_reason = router.pick(
-                        agents, "implement", task.get("type") or "serial", None,
-                        exclude={impl["id"], "mock-a", "mock-b"})
-                    if other and other.get("mode") == "real":
+                    tried = {impl["id"], "mock-a", "mock-b"}
+                    for _alt in range(2):
+                        if good or (ev is not None and ev.is_set()):
+                            break
+                        other, other_reason = router.pick(
+                            agents, "implement", task.get("type") or "serial", None,
+                            exclude=tried)
+                        if not (other and other.get("mode") == "real"):
+                            break
+                        tried.add(other["id"])
                         res = _run_step(run_id, "draft-c%d" % i,
-                                        modelhub.bind_agent(other, difficulty), prompt,
+                                        modelhub.bind_agent(other, difficulty), use_prompt,
                                         step_wd, readonly=False, ev=ev, timeout=2400,
                                         images=_task_images(task, workdir),
                                         note="起草换将 %s → %s：%s" % (
@@ -1649,6 +1691,9 @@ def _run_serial_review(run, task, agents, ev, stats, mode, critics, impl, route,
                                 pass
                         if good:
                             draft_sid = ""   # 换将作者无本任会话，revise 另起
+                if not good:
+                    time.sleep(3)   # 落盘竞态宽限：CLI 崩溃退出前写的文件可能晚于
+                    good, txt = _chapter_state()   # 退出检查零点几秒才可见（c34 实测）
                 if not good:
                     store.update_run(run_id, status="failed",
                                      error="第 %d 章起草失败: %s" % (i, (res or {}).get("error")), ended_at=_now())
@@ -1668,53 +1713,68 @@ def _run_serial_review(run, task, agents, ev, stats, mode, critics, impl, route,
                 # 均分最高者为正稿。变体写隔离文件 chapter-XX-vK.md，赢家改名、
                 # 败稿删除；变体 0 = 本任作者（revise 会话沿用），其余取跨族优先的
                 # 其他真实智能体，不足时同作者开新会话凑数。
-                pool = [impl]
-                others = [a for a in agents if a.get("mode") == "real" and a["id"] != impl["id"]]
-                others.sort(key=lambda a: 0 if a.get("kind") != impl.get("kind") else 1)
-                pool += others[:n_variants - 1]
-                while len(pool) < n_variants:
-                    pool.append(impl)   # 不够就同作者再开一路（新会话天然出不同稿）
-                results = {}
-
-                def _draft_one(kk, agent):
-                    vfile = "chapter-%02d-v%d.md" % (i, kk)
-                    r = _run_step(run_id, "draft-c%d-v%d" % (i, kk),
-                                  modelhub.bind_agent(agent, difficulty),
-                                  _draft_prompt(vfile), step_wd, readonly=False, ev=ev,
-                                  timeout=2400,
-                                  # 赛马只在全新起草时启用（无续会话），每路都是新会话
-                                  images=_task_images(task, workdir),
-                                  note="赛马变体 %d/%d（%s）" % (kk + 1, len(pool), agent.get("id")))
-                    results[kk] = (vfile, agent, r)
-
-                threads = []
-                for kk, agent in enumerate(pool):
-                    th = threading.Thread(target=_draft_one, args=(kk, agent),
-                                          name="race-%s-c%d-v%d" % (run_id, i, kk), daemon=True)
-                    threads.append(th)
-                    th.start()
-                for th in threads:
-                    th.join(3000)
-                _check_cancel(ev)
-
                 scored_variants = []
-                for kk in range(len(pool)):
-                    vfile, agent, r = results.get(kk, (None, None, None))
-                    if vfile is None:
-                        continue
-                    txt = _read_variant(workdir, i, kk)
-                    ok_text = txt and _wc(txt) >= int(wpc * 0.6)
-                    if r is not None and not r["ok"] and not ok_text:
-                        continue   # 这一路彻底失败（无成品也不够长）
-                    if not ok_text:
-                        continue
-                    cj_map, sc, sids2 = run_critique(
-                        txt, 1, note_extra="（本稿为同章赛马变体 %d/%d，只评这一份）" % (kk + 1, len(pool)))
-                    m = means_of(cj_map)
-                    avg = round(sum(m.values()) / max(1, len(m)), 2) if m else 0.0
-                    scored_variants.append({"variant": kk, "agent": agent.get("id"),
-                                            "file": vfile, "means": m, "avg": avg,
-                                            "cj": cj_map, "scored": sc, "sids": sids2})
+                for race_round in range(2):
+                    # 全变体失败（网关突发限流）→ 60s 退避重赛一轮，别一章判死
+                    if race_round:
+                        if ev is not None and ev.is_set():
+                            break
+                        time.sleep(60)
+                        for kk in range(n_variants):
+                            # 清上一轮残稿：防陈旧半成品被本轮评分误认成新成品
+                            try:
+                                os.remove(os.path.join(workdir, "chapter-%02d-v%d.md" % (i, kk)))
+                            except OSError:
+                                pass
+                    pool = [impl]
+                    others = [a for a in agents if a.get("mode") == "real" and a["id"] != impl["id"]]
+                    others.sort(key=lambda a: 0 if a.get("kind") != impl.get("kind") else 1)
+                    pool += others[:n_variants - 1]
+                    while len(pool) < n_variants:
+                        pool.append(impl)   # 不够就同作者再开一路（新会话天然出不同稿）
+                    results = {}
+
+                    def _draft_one(kk, agent):
+                        vfile = "chapter-%02d-v%d.md" % (i, kk)
+                        r = _run_step(run_id, "draft-c%d-v%d" % (i, kk),
+                                      modelhub.bind_agent(agent, difficulty),
+                                      _draft_prompt(vfile), step_wd, readonly=False, ev=ev,
+                                      timeout=2400,
+                                      # 赛马只在全新起草时启用（无续会话），每路都是新会话
+                                      images=_task_images(task, workdir),
+                                      note="赛马变体 %d/%d（%s）" % (kk + 1, len(pool), agent.get("id")))
+                        results[kk] = (vfile, agent, r)
+
+                    threads = []
+                    for kk, agent in enumerate(pool):
+                        th = threading.Thread(target=_draft_one, args=(kk, agent),
+                                              name="race-%s-c%d-v%d" % (run_id, i, kk), daemon=True)
+                        threads.append(th)
+                        th.start()
+                    for th in threads:
+                        th.join(3000)
+                    _check_cancel(ev)
+
+                    scored_variants = []
+                    for kk in range(len(pool)):
+                        vfile, agent, r = results.get(kk, (None, None, None))
+                        if vfile is None:
+                            continue
+                        txt = _read_variant(workdir, i, kk)
+                        ok_text = txt and _wc(txt) >= int(wpc * 0.6)
+                        if r is not None and not r["ok"] and not ok_text:
+                            continue   # 这一路彻底失败（无成品也不够长）
+                        if not ok_text:
+                            continue
+                        cj_map, sc, sids2 = run_critique(
+                            txt, 1, note_extra="（本稿为同章赛马变体 %d/%d，只评这一份）" % (kk + 1, len(pool)))
+                        m = means_of(cj_map)
+                        avg = round(sum(m.values()) / max(1, len(m)), 2) if m else 0.0
+                        scored_variants.append({"variant": kk, "agent": agent.get("id"),
+                                                "file": vfile, "means": m, "avg": avg,
+                                                "cj": cj_map, "scored": sc, "sids": sids2})
+                    if scored_variants:
+                        break
                 if not scored_variants:
                     store.update_run(run_id, status="failed",
                                      error="第 %d 章赛马全部变体起草失败" % i, ended_at=_now())

package/app/core/registry.py

@@ -66,6 +66,10 @@ def _build_agent(entry):
         # 小时级 token 配额（可选，0/缺省=不限）：路由时对本小时用量超标的
         # 智能体降权（munder-difflin 式配额感知），订阅型 CLI 不至于被单任务打爆
         "quota_tokens_per_hour": int(orch.get("quota_tokens_per_hour") or 0),
+        # 整包透传 orch：per-agent 运行时开关（timeout_ms 5E、stall_timeout_s
+        # 看门狗等）都在 runner 侧读取——此前只挑字段，catalog 上配的
+        # timeout_ms/stall 根本流不到 runner（2026-09-17 连载 c35 实测）
+        "orch": orch,
     }