codebee 0.1.71 → 0.1.73

package/app/main.py

@@ -287,7 +287,27 @@ class Handler(BaseHTTPRequestHandler):
                 return self._json(200, {"versions": flows.flow_versions(fid)})
             if path == "/api/skills":
                 from core import skills
-                return self._json(200, skills.view())
+                v = skills.view()
+                try:
+                    # 技能行为卡（ClawHub Skill Card 模式）：市场安装的包按文件
+                    # 名关联拼卡——声明/来源/指纹/安检一次可见。skills(L1) 不能
+                    # 反向 import market（成环），拼接放路由层。
+                    from core import market as _market
+                    reg = _market._load_registry().get("installed") or {}
+                    by_file = {str(r.get("file") or ""): pid
+                               for pid, r in reg.items() if isinstance(r, dict)}
+                    for p in v.get("packs") or []:
+                        pid = by_file.get("market-%s.md" % p.get("id"))
+                        if pid:
+                            p["card"] = _market.skill_card(pid)
+                except Exception:
+                    pass
+                return self._json(200, v)
+            if path == "/api/plugins":
+                # 本地插件（.codex-plugin/plugin.json 契约）：发现+安装台账的
+                # 聚合视图。坏 manifest 以 blocked/broken 行呈现，不 500。
+                from core import plugins
+                return self._json(200, plugins.view())
             if path == "/api/knowledge":
                 from core import knowledge
                 return self._json(200, knowledge.view())
@@ -309,6 +329,14 @@ class Handler(BaseHTTPRequestHandler):
             if path == "/api/evalbench":
                 from core import evalbench
                 return self._json(200, evalbench.state())
+            if path == "/api/evalbench/report":
+                # 评测报告导出（Markdown，浏览器下载）
+                from core import evalbench
+                return self._send(200, evalbench.report_markdown().encode("utf-8"),
+                                  ctype="text/markdown; charset=utf-8",
+                                  headers={"Content-Disposition":
+                                           'attachment; filename="codebee-bench-%s.md"'
+                                           % time.strftime("%Y%m%d-%H%M%S")})
             if path == "/api/catalog":
                 return self._json(200, {"catalog": manager.catalog_view(),
                                         "checking": manager.updates_checking()})
@@ -533,11 +561,15 @@ class Handler(BaseHTTPRequestHandler):
                 return self._json(200, {"path": fp, "text": text})
             m = re.match(r"^/api/tasks/([^/]+)/book-meta$", path)
             if m:
-                # 作品信息（番茄/七猫建书表单资料）：读取任务上的生成状态与结果
+                # 作品信息（番茄/七猫建书表单资料）：读取生成状态与结果。
+                # 连载链沿链继承（都是用第一个）：续写批次读到的就是首批的结果。
                 task = store.get_task(m.group(1))
                 if not task:
                     return self._json(404, {"error": "not found"})
-                return self._json(200, {"book_meta": task.get("book_meta") or {}})
+                return self._json(200, {"book_meta": {
+                    p: store.inherited_book_meta(task, p)
+                    for p in ("fanqie", "qimao")
+                    if store.inherited_book_meta(task, p)}})
             if path == "/api/publish":
                 # 一键发布：平台连接状态 + 最近台账（详情页发布面板）
                 from core.publish import manager as pub
@@ -549,7 +581,8 @@ class Handler(BaseHTTPRequestHandler):
                 from core.publish import manager as pub
                 from core.publish import ledger as pub_ledger
                 tid = m.group(1)
-                books = pub_ledger.load_books().get(tid) or {}
+                # 绑定沿连载链继承（都是用第一个）：续写批次也要看到这本书的登记
+                books = pub_ledger.books_for(tid)
                 return self._json(200, {"history": pub.history(task_id=tid, limit=50),
                                         "books": books,
                                         "published": {p: pub.published_local(tid, p)
@@ -937,6 +970,15 @@ class Handler(BaseHTTPRequestHandler):
         m = re.match(r"^/api/publish/task/([^/]+)/(create-book|chapter|auto-publish)$", path)
         if m:
             return self._api_publish_task_op(m.group(1), m.group(2))
+        m = re.match(r"^/api/publish/task/([^/]+)/confirm-chapter$", path)
+        if m:
+            from core.publish import manager as pub
+            body = self._body() or {}
+            platform = (body.get("platform") or "").strip()
+            ok, err = pub.confirm_manual_chapter(
+                m.group(1), platform, body.get("chapter_no"))
+            return self._json(400, {"error": err or "确认失败"}) \
+                if not ok else self._json(200, {"ok": True})
         m = re.match(r"^/api/publish/task/([^/]+)/register-book$", path)
         if m:
             return self._api_publish_register_book(m.group(1))
@@ -960,7 +1002,9 @@ class Handler(BaseHTTPRequestHandler):
                 return self._json(404, {"error": "任务不存在"})
             body = self._body() or {}
             plat = str(body.get("platform") or "").strip()
-            books = pub_ledger.load_books().get(tid) or {}
+            # 绑定沿连载链继承（与 history 端点同口径）：续写批次发起校准也能
+            # 落到这本书的账上，否则前端看到继承绑定、这里却报「未登记」互相打架
+            books = pub_ledger.books_for(tid)
             if plat and plat not in books:
                 return self._json(400, {"error": "该平台未登记作品"})
             plats = [plat] if plat else list(books)
@@ -1267,6 +1311,23 @@ class Handler(BaseHTTPRequestHandler):
             from core import knowledge
             n = knowledge.learn_from_run((self._body().get("run_id") or ""))
             return self._json(200, {"ok": True, "learned": n})
+        m = re.match(r"^/api/plugins/([^/]+)/(install|enable|disable|remove)$", path)
+        if m:
+            # 本地插件生命周期：只搬文件+记账（skills 走 market.install_files、
+            # MCP 只暴露声明），永不执行插件代码。前端 loadPlugins 拉的就是
+            # 上面 GET /api/plugins；缺这组路由时设置页会显示「插件加载失败」。
+            from core import plugins
+            pid, op = m.group(1), m.group(2)
+            if op == "install":
+                result, err = plugins.install(pid)
+                if err:
+                    return self._json(400, {"error": err})
+                return self._json(200, result or {"ok": True})
+            if op == "remove":
+                err = plugins.remove(pid)
+                return self._json(400, {"error": err}) if err else self._json(200, {"ok": True})
+            err = plugins.toggle(pid, op == "enable")
+            return self._json(400, {"error": err}) if err else self._json(200, {"ok": True})
         m = re.match(r"^/api/flows/([^/]+)/reset$", path)
         if m:
             err = flows.reset_flow(m.group(1))
@@ -1568,6 +1629,14 @@ class Handler(BaseHTTPRequestHandler):
             except zentao.ZenError as e:
                 return self._json(404, {"error": str(e)})
             return self._json(200, {"ok": True, "claim": c})
+        if path == "/api/zentao/claims/retry":
+            from core import zentao
+            body = self._body() or {}
+            try:
+                c, msg = zentao.retry_claim(body.get("bug_id"))
+            except zentao.ZenError as e:
+                return self._json(409, {"error": str(e)})
+            return self._json(200, {"ok": True, "claim": c, "message": msg})
         if path == "/api/zentao/modules":
             from core import zentao
             pid = (self._body() or {}).get("product")
@@ -1714,7 +1783,13 @@ class Handler(BaseHTTPRequestHandler):
         if platform not in bookmeta.PLATFORMS:
             return self._json(400, {"error": "platform 必须是 fanqie 或 qimao"})
         if not bookmeta.needs_book_meta(task):
-            return self._json(400, {"error": "只有连载首批任务需要作品信息；续写批次沿用第一批的开书资料"})
+            # 生成权留在首批（一本书一份资料）：指路根任务，别让用户在续写批次上撞墙
+            root = next((store.get_task(t) for t
+                         in store.serial_chain_ids(task_id) if t != task_id), None)
+            return self._json(400, {"error": (
+                "开书资料在连载首批任务《%s》上生成与维护，请到首批任务的「作品信息」页操作"
+                % ((root or {}).get("title") or "首批")
+                if root else "只有连载首批任务需要作品信息；续写批次沿用第一批的开书资料")})
         if task.get("status") in ("queued", "running"):
             return self._json(400, {"error": "任务正在运行，请等本轮结束后再生成作品信息"})
         cur = ((task.get("book_meta") or {}).get(platform) or {})
@@ -1759,8 +1834,14 @@ class Handler(BaseHTTPRequestHandler):
         if platform not in ("fanqie", "qimao"):
             return self._json(400, {"error": "platform 必须是 fanqie 或 qimao"})
         auto_submit = bool(body.get("auto_submit"))
+        force = bool(body.get("force"))
+        force_confirmed = bool(body.get("force_confirmed"))
+        force_reason = str(body.get("force_reason") or "").strip()
         if op == "create-book":
-            ok, err = pub.create_book_async(task_id, platform, auto_submit)
+            kwargs = ({"force": force, "force_confirmed": force_confirmed,
+                       "force_reason": force_reason}
+                      if force or force_confirmed or force_reason else {})
+            ok, err = pub.create_book_async(task_id, platform, auto_submit, **kwargs)
         elif op == "auto-publish":
             # 定时发布配置（P2.5）：enabled=false 也落（保留 time 供再开）；
             # 校验/归一在 auto.norm_auto_publish，语义见 auto.py 头注
@@ -1789,9 +1870,16 @@ class Handler(BaseHTTPRequestHandler):
                 fp.relative_to(_P(wd).resolve())
             except (OSError, ValueError):
                 return self._json(400, {"error": "章节文件必须在任务工作目录内"})
-            ok, err = pub.upload_chapter_async(task_id, platform, str(fp), auto_submit)
+            kwargs = ({"force": force, "force_confirmed": force_confirmed,
+                       "force_reason": force_reason}
+                      if force or force_confirmed or force_reason else {})
+            ok, err = pub.upload_chapter_async(task_id, platform, str(fp), auto_submit,
+                                               **kwargs)
         if not ok:
-            return self._json(400, {"error": err or "操作失败"})
+            payload = {"error": err or "操作失败"}
+            if str(err or "").startswith("质量门禁拦截："):
+                payload["quality_gate"] = {"status": "blocked", "message": err}
+            return self._json(400, payload)
         return self._json(200, {"ok": True, "started": True})
 
     def _api_publish_register_book(self, task_id):
@@ -1821,10 +1909,19 @@ class Handler(BaseHTTPRequestHandler):
         platform = (body.get("platform") or "").strip()
         if platform not in ("fanqie", "qimao"):
             return self._json(400, {"error": "platform 必须是 fanqie 或 qimao"})
+        force = bool(body.get("force"))
+        force_confirmed = bool(body.get("force_confirmed"))
+        force_reason = str(body.get("force_reason") or "").strip()
         ok, err = pub_auto.publish_pending_async(
-            task_id, platform, auto_submit=bool(body.get("auto_submit")))
+            task_id, platform, auto_submit=bool(body.get("auto_submit")),
+            **({"force": force, "force_confirmed": force_confirmed,
+                "force_reason": force_reason}
+               if force or force_confirmed or force_reason else {}))
         if not ok:
-            return self._json(400, {"error": err or "操作失败"})
+            payload = {"error": err or "操作失败"}
+            if str(err or "").startswith("质量门禁拦截："):
+                payload["quality_gate"] = {"status": "blocked", "message": err}
+            return self._json(400, payload)
         return self._json(200, {"ok": True, "started": True})
 
     def _api_task_side(self, task_id):
@@ -3019,36 +3116,10 @@ def main():
     from core import notify as _notify
     usage_ledger.set_alert_push(_notify.push_text)  # 花费预警推送注入（防 usage→notify 静态环）
     from core import mcp_client as _mcp
+    from core.plugins import merged_servers_text as _plugins_mcp_text
     def _mcp_servers_text():
         """settings 清单 + 已启用插件的 MCP 服务器合成（plugin 名隔离防撞名）。"""
-        import json as _json
-        try:
-            raw = str(settings.load().get("mcp_servers") or "")
-            configured = _json.loads(raw) if raw.strip() else []
-        except Exception:
-            return raw or ""
-        if not isinstance(configured, list):
-            return raw or ""
-        try:
-            from core import plugins as _plugins
-            plugin_servers = _plugins.active_mcp_servers()
-        except Exception:
-            plugin_servers = []          # 坏插件不能连累用户显式配置的服务器
-        import hashlib as _hashlib
-        import re as _re
-        for item in plugin_servers or []:
-            if not isinstance(item, dict):
-                continue
-            plugin_id = str(item.pop("plugin_id", "plugin"))
-            server_name = str(item.get("name") or "server")
-            digest = _hashlib.sha256(
-                (plugin_id + ":" + server_name).encode("utf-8")).hexdigest()[:6]
-            safe = _re.sub(r"[^a-z0-9_-]", "-", plugin_id.lower())
-            item["name"] = ("plg_" + safe[:11] + "_" + digest)[:24]
-            if not any(x.get("name") == item["name"] for x in configured
-                       if isinstance(x, dict)):
-                configured.append(item)
-        return _json.dumps(configured, ensure_ascii=False)
+        return _plugins_mcp_text(settings.load().get("mcp_servers"))
     _mcp.set_settings_text(_mcp_servers_text)  # MCP 清单注入（防 mcp_client→settings 边）
     _step("正在回填用量台账…")
     n_bf = usage.backfill_from_runs()  # 历史运行 token 回填台账（幂等，仅补缺失步骤）

package/app/mcp_server.py

@@ -1,175 +1,175 @@
-# -*- coding: utf-8 -*-
-"""CodeBee 的 MCP 服务器形态：其他 AI 客户端（Claude Code / ZCode / Cursor）
-把这里当工具服务器挂载，即可直接创建任务、查状态、看用量与评测榜单——
-CodeBee 从编排工具变成可以被别的智能体调度的生态节点。
-
-stdio JSON-RPC 2.0（按行分帧，与 core/mcp_client 同一传输约定）。启动：
-  python -X utf8 <CodeBee目录>/app/mcp_server.py
-客户端配置示例（Claude Code / 兼容客户端）：
-  {"mcpServers": {"codebee": {"command": "python",
-                              "args": ["-X", "utf8", "<CodeBee目录>/app/mcp_server.py"]}}}
-仅本机 stdio，无网络监听，凭据不经过此通道；create_task 会真实调度智能体
-执行（与界面上建任务同一条链）。
-"""
-from __future__ import annotations
-
-import json
-import sys
-
-PROTOCOL_VERSION = "2024-11-05"
-SERVER_INFO = {"name": "codebee", "version": "1.0"}
-
-TOOLS = [
-    {"name": "create_task", "description": "创建并启动一个 CodeBee 编排任务（真实调度智能体执行，与界面建任务同链）。返回 task_id 与 run_id。",
-     "inputSchema": {"type": "object", "properties": {
-         "goal": {"type": "string", "description": "任务目标（一句话说清要做什么）"},
-         "type": {"type": "string", "description": "任务类型 id（如 code/novel/doc/direct），默认 doc"},
-         "workdir": {"type": "string", "description": "工作目录绝对路径；留空用 CodeBee 默认保存路径"},
-         "title": {"type": "string", "description": "任务标题；留空取目标前 30 字"}},
-         "required": ["goal"]}},
-    {"name": "get_status", "description": "查任务与最近一次运行的状态：进度、当前步骤、评分、错误。",
-     "inputSchema": {"type": "object", "properties": {
-         "task_id": {"type": "string", "description": "任务 ID"}}, "required": ["task_id"]}},
-    {"name": "list_recent", "description": "列出最近的运行（id/标题/状态/时间）。",
-     "inputSchema": {"type": "object", "properties": {
-         "limit": {"type": "integer", "description": "条数，默认 10，最大 30"}}}},
-    {"name": "usage_summary", "description": "用量台账汇总：今日/本月花费（¥）与 token 总量。",
-     "inputSchema": {"type": "object", "properties": {}}},
-    {"name": "bench_leaderboard", "description": "模型评测基准台能力榜（综合分 Top N）。",
-     "inputSchema": {"type": "object", "properties": {
-         "top": {"type": "integer", "description": "取前 N，默认 5"}}}},
-]
-
-
-def _enqueue(job):
-    """thin wrapper：便于单测 mock（真链会拉起 worker 执行）。"""
-    from core import jobs
-    jobs.enqueue(job)
-
-
-def _tool_create_task(args):
-    from core import settings as settings_mod, store
-    goal = str(args.get("goal") or "").strip()
-    if not goal:
-        return {"isError": True, "content": "goal 不能为空"}
-    payload = {"type": str(args.get("type") or "doc").strip()[:40] or "doc",
-               "title": (str(args.get("title") or "").strip() or goal[:30])[:40],
-               "goal": goal[:4000],
-               "workdir": str(args.get("workdir") or "").strip()
-               or settings_mod.default_workdir()}
-    task = store.create_task(payload)
-    run = store.create_run("orchestration", task["title"], task_id=task["id"])
-    store.update_task_status(task["id"], "queued")
-    _enqueue({"kind": "orchestration", "run_id": run["id"], "task_id": task["id"]})
-    return {"content": "已创建并排队：task_id=%s run_id=%s（可用 get_status 查进度）"
-            % (task["id"], run["id"])}
-
-
-def _tool_get_status(args):
-    from core import store
-    task = store.get_task(str(args.get("task_id") or "").strip())
-    if not task:
-        return {"isError": True, "content": "任务不存在"}
-    lines = ["任务：%s（%s）" % (task.get("title"), task.get("status"))]
-    run = store.latest_run_by_task().get(task["id"])
-    if run:
-        lines.append("最近运行：%s（%s）" % (run.get("id"), run.get("status")))
-        steps = [s for s in (run.get("steps") or []) if isinstance(s, dict)]
-        if steps:
-            cur = steps[-1]
-            lines.append("步骤 %d/%d：%s（%s）" % (len(steps), run.get("total_steps") or len(steps),
-                                                  cur.get("title"), cur.get("status")))
-        v = run.get("verdict") or {}
-        if v.get("overall") is not None:
-            lines.append("评分 %.1f（%s）" % (float(v["overall"]),
-                                             "达标" if v.get("publishable") else "未达标"))
-        if run.get("error"):
-            lines.append("错误：%s" % str(run["error"])[:200])
-    if task.get("git_state") == "isolated":
-        lines.append("⚠ 任务分支待裁决（合并/丢弃）")
-    return {"content": "\n".join(lines)}
-
-
-def _tool_list_recent(args):
-    from core import store
-    try:
-        limit = max(1, min(30, int(args.get("limit") or 10)))
-    except (TypeError, ValueError):
-        limit = 10
-    rows = [{"id": r.get("id"), "title": r.get("title"), "status": r.get("status"),
-             "ts": r.get("started_at") or ""} for r in store.list_runs(limit=limit)]
-    return {"content": json.dumps(rows, ensure_ascii=False, indent=1)}
-
-
-def _tool_usage_summary(args):
-    from core import usage
-    today, month = usage.cost_snapshot()
-    tokens = sum(int(r.get("total") or 0) for r in usage._iter_records(31))
-    return {"content": "今日花费 ¥%.2f\n本月花费 ¥%.2f\n近 31 天 tokens：%d（未标定单价的调用不计钱）"
-            % (today, month, tokens)}
-
-
-def _tool_bench(args):
-    from core import evalbench
-    try:
-        top = max(1, min(10, int(args.get("top") or 5)))
-    except (TypeError, ValueError):
-        top = 5
-    board = evalbench.state().get("leaderboard") or []
-    rows = [{"rank": r["rank"], "model": r["model"], "overall": r["overall"],
-             "provider": r.get("provider_name")} for r in board[:top]]
-    return {"content": json.dumps(rows, ensure_ascii=False, indent=1)}
-
-
-_TOOL_IMPL = {"create_task": _tool_create_task, "get_status": _tool_get_status,
-              "list_recent": _tool_list_recent, "usage_summary": _tool_usage_summary,
-              "bench_leaderboard": _tool_bench}
-
-
-def handle_request(msg):
-    """处理一条客户端请求，返回响应 dict（通知返回 None）。"""
-    if not isinstance(msg, dict) or msg.get("method") is None:
-        return None
-    rid = msg.get("id")
-    method = str(msg["method"])
-    if method == "initialize":
-        return {"jsonrpc": "2.0", "id": rid, "result": {
-            "protocolVersion": PROTOCOL_VERSION, "capabilities": {"tools": {}},
-            "serverInfo": SERVER_INFO}}
-    if method.startswith("notifications/"):
-        return None
-    if method == "tools/list":
-        return {"jsonrpc": "2.0", "id": rid, "result": {"tools": TOOLS}}
-    if method == "tools/call":
-        params = msg.get("params") or {}
-        name = str(params.get("name") or "")
-        fn = _TOOL_IMPL.get(name)
-        if fn is None:
-            return {"jsonrpc": "2.0", "id": rid, "result": {
-                "isError": True, "content": "未知工具: %s" % name}}
-        try:
-            out = fn(params.get("arguments") or {})
-        except Exception as e:
-            out = {"isError": True, "content": "工具执行失败: %s" % e}
-        return {"jsonrpc": "2.0", "id": rid, "result": out}
-    return {"jsonrpc": "2.0", "id": rid,
-            "error": {"code": -32601, "message": "no such method: %s" % method}}
-
-
-def main():
-    for raw in sys.stdin:
-        line = raw.strip()
-        if not line.startswith("{"):
-            continue
-        try:
-            msg = json.loads(line)
-        except Exception:
-            continue
-        resp = handle_request(msg)
-        if resp is not None:
-            sys.stdout.write(json.dumps(resp, ensure_ascii=False) + "\n")
-            sys.stdout.flush()
-
-
-if __name__ == "__main__":
-    main()
+# -*- coding: utf-8 -*-
+"""CodeBee 的 MCP 服务器形态：其他 AI 客户端（Claude Code / ZCode / Cursor）
+把这里当工具服务器挂载，即可直接创建任务、查状态、看用量与评测榜单——
+CodeBee 从编排工具变成可以被别的智能体调度的生态节点。
+
+stdio JSON-RPC 2.0（按行分帧，与 core/mcp_client 同一传输约定）。启动：
+  python -X utf8 <CodeBee目录>/app/mcp_server.py
+客户端配置示例（Claude Code / 兼容客户端）：
+  {"mcpServers": {"codebee": {"command": "python",
+                              "args": ["-X", "utf8", "<CodeBee目录>/app/mcp_server.py"]}}}
+仅本机 stdio，无网络监听，凭据不经过此通道；create_task 会真实调度智能体
+执行（与界面上建任务同一条链）。
+"""
+from __future__ import annotations
+
+import json
+import sys
+
+PROTOCOL_VERSION = "2024-11-05"
+SERVER_INFO = {"name": "codebee", "version": "1.0"}
+
+TOOLS = [
+    {"name": "create_task", "description": "创建并启动一个 CodeBee 编排任务（真实调度智能体执行，与界面建任务同链）。返回 task_id 与 run_id。",
+     "inputSchema": {"type": "object", "properties": {
+         "goal": {"type": "string", "description": "任务目标（一句话说清要做什么）"},
+         "type": {"type": "string", "description": "任务类型 id（如 code/novel/doc/direct），默认 doc"},
+         "workdir": {"type": "string", "description": "工作目录绝对路径；留空用 CodeBee 默认保存路径"},
+         "title": {"type": "string", "description": "任务标题；留空取目标前 30 字"}},
+         "required": ["goal"]}},
+    {"name": "get_status", "description": "查任务与最近一次运行的状态：进度、当前步骤、评分、错误。",
+     "inputSchema": {"type": "object", "properties": {
+         "task_id": {"type": "string", "description": "任务 ID"}}, "required": ["task_id"]}},
+    {"name": "list_recent", "description": "列出最近的运行（id/标题/状态/时间）。",
+     "inputSchema": {"type": "object", "properties": {
+         "limit": {"type": "integer", "description": "条数，默认 10，最大 30"}}}},
+    {"name": "usage_summary", "description": "用量台账汇总：今日/本月花费（¥）与 token 总量。",
+     "inputSchema": {"type": "object", "properties": {}}},
+    {"name": "bench_leaderboard", "description": "模型评测基准台能力榜（综合分 Top N）。",
+     "inputSchema": {"type": "object", "properties": {
+         "top": {"type": "integer", "description": "取前 N，默认 5"}}}},
+]
+
+
+def _enqueue(job):
+    """thin wrapper：便于单测 mock（真链会拉起 worker 执行）。"""
+    from core import jobs
+    jobs.enqueue(job)
+
+
+def _tool_create_task(args):
+    from core import settings as settings_mod, store
+    goal = str(args.get("goal") or "").strip()
+    if not goal:
+        return {"isError": True, "content": "goal 不能为空"}
+    payload = {"type": str(args.get("type") or "doc").strip()[:40] or "doc",
+               "title": (str(args.get("title") or "").strip() or goal[:30])[:40],
+               "goal": goal[:4000],
+               "workdir": str(args.get("workdir") or "").strip()
+               or settings_mod.default_workdir()}
+    task = store.create_task(payload)
+    run = store.create_run("orchestration", task["title"], task_id=task["id"])
+    store.update_task_status(task["id"], "queued")
+    _enqueue({"kind": "orchestration", "run_id": run["id"], "task_id": task["id"]})
+    return {"content": "已创建并排队：task_id=%s run_id=%s（可用 get_status 查进度）"
+            % (task["id"], run["id"])}
+
+
+def _tool_get_status(args):
+    from core import store
+    task = store.get_task(str(args.get("task_id") or "").strip())
+    if not task:
+        return {"isError": True, "content": "任务不存在"}
+    lines = ["任务：%s（%s）" % (task.get("title"), task.get("status"))]
+    run = store.latest_run_by_task().get(task["id"])
+    if run:
+        lines.append("最近运行：%s（%s）" % (run.get("id"), run.get("status")))
+        steps = [s for s in (run.get("steps") or []) if isinstance(s, dict)]
+        if steps:
+            cur = steps[-1]
+            lines.append("步骤 %d/%d：%s（%s）" % (len(steps), run.get("total_steps") or len(steps),
+                                                  cur.get("title"), cur.get("status")))
+        v = run.get("verdict") or {}
+        if v.get("overall") is not None:
+            lines.append("评分 %.1f（%s）" % (float(v["overall"]),
+                                             "达标" if v.get("publishable") else "未达标"))
+        if run.get("error"):
+            lines.append("错误：%s" % str(run["error"])[:200])
+    if task.get("git_state") == "isolated":
+        lines.append("⚠ 任务分支待裁决（合并/丢弃）")
+    return {"content": "\n".join(lines)}
+
+
+def _tool_list_recent(args):
+    from core import store
+    try:
+        limit = max(1, min(30, int(args.get("limit") or 10)))
+    except (TypeError, ValueError):
+        limit = 10
+    rows = [{"id": r.get("id"), "title": r.get("title"), "status": r.get("status"),
+             "ts": r.get("started_at") or ""} for r in store.list_runs(limit=limit)]
+    return {"content": json.dumps(rows, ensure_ascii=False, indent=1)}
+
+
+def _tool_usage_summary(args):
+    from core import usage
+    today, month = usage.cost_snapshot()
+    tokens = sum(int(r.get("total") or 0) for r in usage._iter_records(31))
+    return {"content": "今日花费 ¥%.2f\n本月花费 ¥%.2f\n近 31 天 tokens：%d（未标定单价的调用不计钱）"
+            % (today, month, tokens)}
+
+
+def _tool_bench(args):
+    from core import evalbench
+    try:
+        top = max(1, min(10, int(args.get("top") or 5)))
+    except (TypeError, ValueError):
+        top = 5
+    board = evalbench.state().get("leaderboard") or []
+    rows = [{"rank": r["rank"], "model": r["model"], "overall": r["overall"],
+             "provider": r.get("provider_name")} for r in board[:top]]
+    return {"content": json.dumps(rows, ensure_ascii=False, indent=1)}
+
+
+_TOOL_IMPL = {"create_task": _tool_create_task, "get_status": _tool_get_status,
+              "list_recent": _tool_list_recent, "usage_summary": _tool_usage_summary,
+              "bench_leaderboard": _tool_bench}
+
+
+def handle_request(msg):
+    """处理一条客户端请求，返回响应 dict（通知返回 None）。"""
+    if not isinstance(msg, dict) or msg.get("method") is None:
+        return None
+    rid = msg.get("id")
+    method = str(msg["method"])
+    if method == "initialize":
+        return {"jsonrpc": "2.0", "id": rid, "result": {
+            "protocolVersion": PROTOCOL_VERSION, "capabilities": {"tools": {}},
+            "serverInfo": SERVER_INFO}}
+    if method.startswith("notifications/"):
+        return None
+    if method == "tools/list":
+        return {"jsonrpc": "2.0", "id": rid, "result": {"tools": TOOLS}}
+    if method == "tools/call":
+        params = msg.get("params") or {}
+        name = str(params.get("name") or "")
+        fn = _TOOL_IMPL.get(name)
+        if fn is None:
+            return {"jsonrpc": "2.0", "id": rid, "result": {
+                "isError": True, "content": "未知工具: %s" % name}}
+        try:
+            out = fn(params.get("arguments") or {})
+        except Exception as e:
+            out = {"isError": True, "content": "工具执行失败: %s" % e}
+        return {"jsonrpc": "2.0", "id": rid, "result": out}
+    return {"jsonrpc": "2.0", "id": rid,
+            "error": {"code": -32601, "message": "no such method: %s" % method}}
+
+
+def main():
+    for raw in sys.stdin:
+        line = raw.strip()
+        if not line.startswith("{"):
+            continue
+        try:
+            msg = json.loads(line)
+        except Exception:
+            continue
+        resp = handle_request(msg)
+        if resp is not None:
+            sys.stdout.write(json.dumps(resp, ensure_ascii=False) + "\n")
+            sys.stdout.flush()
+
+
+if __name__ == "__main__":
+    main()