codebee 0.1.64 → 0.1.65

package/app/core/pipeline.py

@@ -1060,7 +1060,9 @@ def _retry_content_draft_with_cli(run_id, task, agents, impl, difficulty, mode,
     failures = ["%s：%s" % (impl.get("id") or "CLI",
                             (draft_res.get("error") or "起草失败")[:180])]
     dead_upstreams = []
-    first_upstreams = router.agent_upstreams(impl.get("id"))
+    first_upstreams = router.agent_upstreams(
+        impl.get("id"), difficulty=difficulty,
+        task_type=task.get("type") or "novel", role="implement")
     if first_upstreams:
         dead_upstreams.append(first_upstreams)
     current = impl
@@ -1072,7 +1074,7 @@ def _retry_content_draft_with_cli(run_id, task, agents, impl, difficulty, mode,
         _check_cancel(ev)
         candidate, reason = router.pick_switch_candidate(
             agents, "implement", task.get("type") or "novel", stats,
-            exclude=tried, dead_upstreams=dead_upstreams)
+            exclude=tried, dead_upstreams=dead_upstreams, difficulty=difficulty)
         if candidate is None or candidate.get("id") in tried:
             break
         tried.add(candidate.get("id"))
@@ -1097,7 +1099,9 @@ def _retry_content_draft_with_cli(run_id, task, agents, impl, difficulty, mode,
                                     (last_result.get("error") or "起草失败")[:180]))
         if (last_result.get("raw") or {}).get("cancelled"):
             break
-        ups = router.agent_upstreams(candidate.get("id"))
+        ups = router.agent_upstreams(
+            candidate.get("id"), difficulty=difficulty,
+            task_type=task.get("type") or "novel", role="implement")
         if ups:
             dead_upstreams.append(ups)
         current = candidate
@@ -1309,8 +1313,11 @@ def _run_code(run, task, agents, ev, stats, mode):
             # 原实现者若死于配额，其上游（网关/账号）也一并判入死池——否则第一棒
             # 换将不知道同网关候选已死，评审者重选更会把「与死者同上游」的候选
             # 继续当活口，评审撞同一条失效网关白烧一轮后误报「评审器故障」。
-            if runner._quota_error(res.get("error") or ""):
-                ups0 = router.agent_upstreams(impl_agent["id"])
+            if (runner._quota_error(res.get("error") or "")
+                    or runner._permission_error(res.get("error") or "")):
+                ups0 = router.agent_upstreams(
+                    impl_agent["id"], difficulty=difficulty,
+                    task_type="code", role="implement")
                 if ups0:
                     dead_ups.append(ups0)
             prev_id = impl_agent["id"]
@@ -1318,7 +1325,8 @@ def _run_code(run, task, agents, ev, stats, mode):
             while True:
                 other, other_reason = router.pick_switch_candidate(
                     agents, "implement", "code", stats,
-                    exclude=tried, dead_upstreams=dead_ups)
+                    exclude=tried, dead_upstreams=dead_ups,
+                    difficulty=difficulty)
                 if other is None or other.get("mode") != "real":
                     other = None
                     break
@@ -1348,8 +1356,11 @@ def _run_code(run, task, agents, ev, stats, mode):
                 tried.add(other["id"])
                 dead_ids.add(other["id"])
                 prev_id = other["id"]
-                if runner._quota_error(res.get("error") or ""):
-                    ups = router.agent_upstreams(other["id"])
+                if (runner._quota_error(res.get("error") or "")
+                        or runner._permission_error(res.get("error") or "")):
+                    ups = router.agent_upstreams(
+                        other["id"], difficulty=difficulty,
+                        task_type="code", role="implement")
                     if ups:
                         dead_ups.append(ups)
                 else:

package/app/core/router.py

@@ -2,8 +2,9 @@
 """智能路由：能力基线 × 历史胜率 × 角色约束 → 选智能体，并给出可解释的理由。"""
 from __future__ import annotations
 
-import os
-import re
+import os
+import re
+import json
 
 from . import dispatch, history
 
@@ -144,14 +145,16 @@ def pick(agents, role, ttype, stats=None, exclude=()):
 # 空链 CLI 的真实上游在各自本机配置里（注册表里的 provider_id 是 UI 残留，
 # 2026-09-22 实测 qwencode 注册表指 prov-36 本机却指公司网关）。正则探针
 # 只为换将时识别「同上游」，读不到就算未知——未知不参与剔除，宁白试不误杀。
-_LOCAL_ENDPOINT_PROBES = (
+_LOCAL_ENDPOINT_PROBES = (
     ("kimi-code", "~/.kimi-code/config.toml", r'baseUrl\s*=\s*"([^"]+)"'),
     ("claude-code", "~/.claude/settings.json", r'"ANTHROPIC_BASE_URL"\s*:\s*"([^"]+)"'),
     ("qwencode", "~/.qwen/settings.json", r'"OPENAI_BASE_URL"\s*:\s*"([^"]+)"'),
     ("opencode", "~/.config/opencode/opencode.json", r'"baseURL"\s*:\s*"([^"]+)"'),
     ("opencode", "~/.config/opencode/opencode.jsonc", r'"baseURL"\s*:\s*"([^"]+)"'),
     ("codex-cli", "~/.codex/config.toml", r'base_url\s*=\s*"([^"]+)"'),
-)
+)
+
+_PI_SETTINGS_PATH = "~/.pi/agent/settings.json"
 
 
 def _host_of(url):
@@ -160,27 +163,59 @@ def _host_of(url):
     return m.group(1).lower() if m else ""
 
 
-def agent_upstreams(agent_id):
-    """候选实际会用到的上游集合（host:port 归一）。显式链 > CLI 本机配置
-    端点；两者都拿不到返回空集 = 未知。"""
-    ups = set()
-    try:
-        from . import modelhub
-        pref = modelhub._binding_for(agent_id)
-        chain = modelhub._binding_chain(pref)
-        provs = {p.get("id"): p for p in modelhub.providers()}
-        if chain:
-            for item in chain:
-                pid = (item.get("provider_id") or "").strip()
-                h = _host_of((provs.get(pid) or {}).get("base_url"))
-                if h:
-                    ups.add(h)
-    except Exception:
-        pass
-    if not ups:
-        for aid, path, pat in _LOCAL_ENDPOINT_PROBES:
-            if aid != agent_id:
-                continue
+def agent_upstreams(agent_id, difficulty="default", task_type="", role=""):
+    """候选实际会用到的上游集合（host:port 归一）。显式链 > CLI 本机配置
+    端点；无显式绑定时按运行期自动推荐链判断；都拿不到返回空集 = 未知。"""
+    ups = set()
+    try:
+        from . import modelhub
+        pref = modelhub._binding_for(agent_id)
+        provs = {p.get("id"): p for p in modelhub.providers()}
+
+        def _add_chain(chain):
+            for item in chain or ():
+                pid = (item.get("provider_id") or "").strip()
+                provider = item.get("provider") or provs.get(pid) or {}
+                h = _host_of(provider.get("base_url"))
+                if h:
+                    ups.add(h)
+
+        configured = bool(modelhub._binding_chain(pref) or pref.get("provider_id"))
+        # 保留原始显式链中的 endpoint，即使某条已禁用/失效；同一个死上游仍
+        # 不值得在另一个 CLI 上立刻再撞一次。
+        if configured:
+            chain = modelhub._binding_chain(pref)
+            _add_chain(chain)
+            resolved = modelhub.resolve_binding(agent_id, difficulty) or {}
+            _add_chain(resolved.get("call_chain"))
+        else:
+            # bind_agent 在空绑定时会自动推荐供应商链。换将选路必须观察同一
+            # 生效链，否则“空绑定”的多个 CLI 可能被排成不同候选却共用一个 403 网关。
+            recommended = modelhub.recommend_binding(
+                agent_id, difficulty, task_type=task_type, role=role) or {}
+            _add_chain(recommended.get("call_chain"))
+    except Exception:
+        pass
+    if not ups:
+        if agent_id == "pi":
+            # Pi 将 endpoint 放在 providers.<defaultProvider>.baseUrl，不能只扫
+            # 所有 baseUrl（那会把未选中的 provider 也误算成当前上游）。
+            try:
+                with open(os.path.expanduser(_PI_SETTINGS_PATH), "r",
+                          encoding="utf-8") as fh:
+                    cfg = json.load(fh)
+                selected = (cfg.get("providers") or {}).get(
+                    cfg.get("defaultProvider") or "") or {}
+                h = _host_of(selected.get("baseUrl") or selected.get("baseURL") or
+                             selected.get("base_url"))
+                if h:
+                    ups.add(h)
+            except (OSError, ValueError, TypeError):
+                pass
+    if not ups:
+        for aid, path, pat in _LOCAL_ENDPOINT_PROBES:
+            if aid != agent_id:
+                continue
             try:
                 with open(os.path.expanduser(path), "r", encoding="utf-8",
                           errors="ignore") as fh:
@@ -194,8 +229,8 @@ def agent_upstreams(agent_id):
     return ups
 
 
-def pick_switch_candidate(agents, role, ttype, stats, exclude=(),
-                          dead_upstreams=()):
+def pick_switch_candidate(agents, role, ttype, stats, exclude=(),
+                          dead_upstreams=(), difficulty="default"):
     """换将选将。配额类死亡背景下：与死者同上游的候选依次让位，直到找到
     异上游/上游未知的候选（2026-09-22 实案：kimi 与 claude 同骑智谱，0.1 分
     之差把异上游 qwencode 压在下面，换将=换壳不换命）。异上游耗尽后同上游
@@ -204,9 +239,10 @@ def pick_switch_candidate(agents, role, ttype, stats, exclude=(),
     best, reason = pick(agents, role, ttype, stats, exclude=exclude)
     if best is None or not dead_upstreams:
         return best, reason
-    deferred = []
-    while best is not None:
-        ups = agent_upstreams(best.get("id"))
+    deferred = []
+    while best is not None:
+        ups = agent_upstreams(best.get("id"), difficulty=difficulty,
+                              task_type=ttype, role=role)
         if not (ups and any(ups & dead for dead in dead_upstreams)):
             break  # 异上游或上游未知：就用它
         deferred.append((best, reason))

package/app/core/runner.py

@@ -1043,6 +1043,17 @@ def _auth_error(err):
     return any(k in err for k in _AUTH)
 
 
+_FORBIDDEN_HINTS = ("forbidden", "request not allowed", "permission denied",
+                    "access denied")
+
+
+def _permission_error(err):
+    """识别明确的 HTTP 403 权限拒绝；它是上游故障，不等同于无效 API Key。"""
+    err = (err or "").lower()
+    return bool(re.search(r"(?<!\d)403(?!\d)", err) and
+                any(marker in err for marker in _FORBIDDEN_HINTS))
+
+
 def _attempt_credential(att):
     """绑定链中凭据的非敏感身份；不把密钥值写入日志或尝试记录。"""
     return (str((att or {}).get("provider_id") or ""),
@@ -1621,6 +1632,7 @@ def run_agent(agent, prompt, workdir=None, readonly=True,
     grace_left = RATE_LIMIT_GRACE_N  # 链尾限流宽限预算：整链撞 429 时原地等一个窗口再试一次
     timed_out_upstreams = {}  # 同一上游两次超时后跳过剩余模型
     skipped_upstreams = set()
+    forbidden_upstreams = set()
     auth_failed_credentials = set()
     for ai, att in enumerate(attempts):
         credential = _attempt_credential(att)
@@ -1638,7 +1650,7 @@ def run_agent(agent, prompt, workdir=None, readonly=True,
                 out["attempts"] = list(attempt_history)
             continue
         upstream = _attempt_upstream(att)
-        if upstream in skipped_upstreams:
+        if upstream in skipped_upstreams or upstream in forbidden_upstreams:
             continue
         # 有任务 deadline 时，所有候选共享同一个截止时刻；否则每个候选
         # 使用 catalog 配置或 caller 提供的单模型时限。
@@ -1763,7 +1775,8 @@ def run_agent(agent, prompt, workdir=None, readonly=True,
                     except Exception:
                         pass
                 if not (_quota_error(out["error"]) or _transient_error(out["error"])
-                        or _rate_limited(out["error"])):
+                        or _rate_limited(out["error"])
+                        or _permission_error(out["error"])):
                     # 尾段清洗后没有错误信号而流式审计日志里有：补日志错误行。
                     # 2026-09-22 实测 kimi 撞 429 时 stderr/stdout 只剩启动横幅
                     # （真错误只走了流式管道），路由层看不到配额关键词，换将
@@ -1867,6 +1880,15 @@ def run_agent(agent, prompt, workdir=None, readonly=True,
                        for candidate in attempts[ai + 1:]):
                 return out
             continue
+        if _permission_error(out.get("error")):
+            # 403 通常是该网关/上游拒绝当前请求；同一 host 换模型只会重撞，
+            # 但不同 host 的候选仍可能有权限，应继续尝试异上游链项。
+            if upstream != "unknown":
+                forbidden_upstreams.add(upstream)
+            if not any(_attempt_upstream(candidate) not in forbidden_upstreams
+                       for candidate in attempts[ai + 1:]):
+                return out
+            continue
         # 换将闸门：什么失败值得烧链上下一个候选。
         # · 瞬态/配额（文本特征，含 CLI 内部 api_retry 的 429/5xx 字样）照旧；
         # · 超时：显式可换（看 res 标志位）——慢上游换快候选可能救回。
@@ -1879,6 +1901,7 @@ def run_agent(agent, prompt, workdir=None, readonly=True,
         if (raw.get("cancelled")   # 取消是用户意志，绝不降级（尾段可能带瞬态字样）
                 or not (_transient_error(out.get("error")) or _quota_error(out.get("error"))
                         or _refusal_error(out.get("error"))
+                        or _permission_error(out.get("error"))
                         or raw.get("timed_out"))):
             return out  # 取消/解析失败等真终态不降级
         # 同一上游的模型连续两次超时就收手——

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "codebee",
-  "version": "0.1.64",
+  "version": "0.1.65",
   "description": "CodeBee · 多智能体编排台 — 统一编排本机 AI 编码 CLI（Codex / Claude Code / QwenCode / OpenCode / Aider…）：目标→拆解→路由→执行→验证→跨厂商评审→修复→报告。纯 Python 标准库实现，本地 Web UI。",
   "license": "MIT",
   "bin": {