Thin-channel/thick-runner refactor: strangler-fig migration + auth/reliability hardening + MCP interface

.gitattributes

@@ -0,0 +1,9 @@
+# Force LF everywhere. Some frontend files in this repo's history carried
+# mixed CRLF/LF line endings (pre-dating this file); an editor touching any
+# line of such a file makes git's line-based diff show the whole file as
+# changed, even when the real content delta is tiny (see
+# frontend/src/components/ChannelConfigForm.tsx and
+# frontend/src/pages/SourcesPage.tsx in refactor/thin-channel-thick-runner —
+# `git diff -w` shows the real, small diff). This doesn't rewrite already-
+# pushed history; it just stops new churn from the same cause.
+* text=auto eol=lf

TESTING.md

@@ -483,12 +483,20 @@ playwright install chromium
 
 ## 技能执行环路 e2e（live marker，Windows）
 
-`tests/skills/test_execute_loop_live.py` 是唯一一个跑**真实本地 Chrome**的端到端测试：它通过
+`tests/skills/test_execute_loop_live.py` 跑**真实本地 Chromium 内核浏览器**的端到端测试：它通过
 CDP 把整条 `perceive → act → extract → done` 环路、以及 headless 写前确认闸门，对着一个**本机静态
-页面**真跑一遍（issue 07）。整个文件标了 `@pytest.mark.live`，所以默认的 `pytest -m "not live"`
-（带 `--cov-fail-under=80`）**永远不需要浏览器**。浏览器是真的；只有便宜模型的**动作选择**被脚本
+页面**真跑一遍（issue 07）。`tests/skills/test_cookie_persistence_live.py` 补一条：证 cookie/登录态
+在两次独立 `connect_over_cdp`（attach → detach → 重新 attach，等价生产里两次任务复用同一 endpoint）
+之间真的没丢——不是靠"CDP 挂接不落地新 context"这条架构推理去信,是真设 cookie 断开重连读回来验证过。
+
+两个文件都标了 `@pytest.mark.live`，所以默认的 `pytest -m "not live"`（带 `--cov-fail-under=80`）
+**永远不需要浏览器**。浏览器是真的；只有 `test_execute_loop_live.py` 里便宜模型的**动作选择**被脚本
 固定（patch `backend.channels.skill_channel._build_model_call`），避免模型抖动让 live 测试变 flaky。
 
+**Chrome 和 Edge 都真跑过**（2026-07-02）：`connect_over_cdp` 走纯 CDP 协议，不认浏览器厂商——只要
+给个 `--remote-debugging-port` endpoint 就行，不是 Chrome 专属。以下步骤两个浏览器通用，示例用 Chrome，
+Edge 见下方"多浏览器"小节。
+
 为新机器（win32）从零复现：
 
 1. 安装 Playwright + 其 Chromium 驱动（每台机器一次性）：
@@ -509,17 +517,31 @@ CDP 把整条 `perceive → act → extract → done` 环路、以及 headless 
      --no-first-run --no-default-browser-check
    ```
 
-3. 指向它，并**只**跑这个 live 技能测试：
+3. 指向它，跑 live 技能测试（连带 cookie 持久化那条）：
 
    ```powershell
    $env:SKILL_LIVE_CDP_ENDPOINT = "http://127.0.0.1:9222"
-   uv run pytest -m live tests/skills/test_execute_loop_live.py
+   uv run pytest -m live tests/skills/test_execute_loop_live.py tests/skills/test_cookie_persistence_live.py
    ```
 
    测试从 `SKILL_LIVE_CDP_ENDPOINT` 读 endpoint（回退到 `OPENCLI_CDP_ENDPOINT`）。**未设置时**它会
    `pytest.skip(...)` 并给出可操作的提示，而不是 fail。被测页面由测试内嵌的 `ThreadingHTTPServer`
    起在 `127.0.0.1:<随机端口>`，不依赖任何外部站点。
 
+   **多浏览器（Edge / 其他 Chromium 内核）：**
+
+   ```powershell
+   & "C:\Program Files (x86)\Microsoft\Edge\Application\msedge.exe" `
+     --remote-debugging-port=9223 --remote-debugging-address=127.0.0.1 `
+     --no-first-run --no-default-browser-check `
+     --user-data-dir="<某个持久目录>"
+   ```
+
+   ⚠️ **踩坑**：Edge 即使给了 `--remote-debugging-address=127.0.0.1`，实测仍只绑
+   `::1`（IPv6 loopback），`http://127.0.0.1:9223/json/version` 会超时。用
+   `http://localhost:9223`（Windows 上 `localhost` 优先解析到 `::1`）或直接
+   `http://[::1]:9223`。Chrome 没这毛病，`127.0.0.1` 直接通。
+
 4. 默认套件**不含**它（CI / 本地日常都走这条，不需要 Chrome）：
 
    ```powershell
@@ -529,7 +551,7 @@ CDP 把整条 `perceive → act → extract → done` 环路、以及 headless 
    确认它确实被默认排除：
 
    ```powershell
-   uv run pytest -m "not live" --collect-only -q | Select-String "test_execute_loop_live"   # 应无匹配
+   uv run pytest -m "not live" --collect-only -q | Select-String "test_execute_loop_live|test_cookie_persistence_live"   # 应无匹配
    ```
 
 5. DB 说明：live 测试把一个**临时 SQLite**（默认内存库，`StaticPool` 单连接共享）绑进

backend/api/v1/__init__.py

@@ -6,13 +6,15 @@
     agents,
     browsers,
     chat,
+    cookies,
     dashboard,
     nodes,
     notifications,
     providers,
     records,
     schedules,
     skill_bridge,
+    skill_record,
     skills,
     sources,
     system,
@@ -26,6 +28,7 @@
 v1_router.include_router(agents.router)
 v1_router.include_router(browsers.router)
 v1_router.include_router(chat.router)
+v1_router.include_router(cookies.router)
 v1_router.include_router(nodes.router)
 v1_router.include_router(providers.router)
 v1_router.include_router(sources.router)
@@ -34,6 +37,7 @@
 v1_router.include_router(schedules.router)
 v1_router.include_router(skills.router)
 v1_router.include_router(skill_bridge.router)
+v1_router.include_router(skill_record.router)
 v1_router.include_router(webhooks.router)
 v1_router.include_router(notifications.router)
 v1_router.include_router(workers.router)

backend/api/v1/chat.py

@@ -27,6 +27,7 @@
 from backend.schemas.schedule import CronScheduleUpdate
 from backend.schemas.source import DataSourceUpdate
 from backend.services import schedule_service, source_service, task_service
+from backend.skills.toolcall import _is_xml_tool_model, _parse_tool_use, _safe_json
 
 logger = logging.getLogger(__name__)
 
@@ -40,7 +41,9 @@
 规则:
 - 需要知道有哪些数据源时, 调 list_sources。
 - 用户要启用/停用某个数据源时, 调 toggle_source。这是写操作, 系统不会立即执行, 会先让用户确认。
-- 不要编造数据源 id; 先用 list_sources 拿到真实 id 再 toggle。
+- 用户要配置 AI 处理(富化)阶段时(换模型 / 开关 AI), 先 list_providers 看现有提供商, 再 update_provider。
+  启用一个 provider = 采集成功后自动用它跑 AI 富化; 全部停用 = 不跑 AI。换模型改 default_model。
+- 不要编造 id; 先用 list_* 拿到真实 id 再做写操作。
 - 用中文简洁回答。"""
 
 
@@ -113,9 +116,33 @@
             },
         },
     },
+    {
+        "type": "function",
+        "function": {
+            "name": "list_providers",
+            "description": "列出所有模型提供商 (返回 id / name / provider_type / default_model / base_url / enabled)。AI 富化阶段用哪个模型由 provider 决定。只读, 立即执行。",
+            "parameters": {"type": "object", "properties": {}},
+        },
+    },
+    {
+        "type": "function",
+        "function": {
+            "name": "update_provider",
+            "description": "配置 AI 处理阶段: 改某个模型提供商的默认模型, 或启用/停用它。启用一个 provider 后, 采集成功会自动用它跑 AI 富化; 全部停用则不跑 AI。写操作, 需用户确认。",
+            "parameters": {
+                "type": "object",
+                "properties": {
+                    "provider_id": {"type": "string", "description": "模型提供商 id"},
+                    "default_model": {"type": "string", "description": "默认模型名 (可选, 如 qwen3:4b)"},
+                    "enabled": {"type": "boolean", "description": "启用/停用 (可选)"},
+                },
+                "required": ["provider_id"],
+            },
+        },
+    },
 ]
 
-WRITE_TOOLS = {"toggle_source", "trigger_task", "update_schedule"}
+WRITE_TOOLS = {"toggle_source", "trigger_task", "update_schedule", "update_provider"}
 
 
 # ── request / response 模型 ─────────────────────────────────────────────────
@@ -195,6 +222,15 @@ async def _run_read_tool(db: AsyncSession, name: str, args: dict[str, Any]) -> A
             {"id": t.id, "source_id": t.source_id, "status": t.status, "trigger_type": t.trigger_type}
             for t in tasks
         ]
+    if name == "list_providers":
+        result = await db.execute(select(ModelProvider).order_by(ModelProvider.created_at.asc()))
+        return [
+            {
+                "id": p.id, "name": p.name, "provider_type": p.provider_type,
+                "default_model": p.default_model, "base_url": p.base_url, "enabled": p.enabled,
+            }
+            for p in result.scalars().all()
+        ]
     return {"error": f"unknown read tool: {name}"}
 
 
@@ -247,6 +283,29 @@ async def _build_proposal(db: AsyncSession, name: str, args: dict[str, Any]) ->
             summary=f"修改调度「{schedule.name}」",
             diff="; ".join(changes),
         )
+    if name == "update_provider":
+        provider_id = args.get("provider_id", "")
+        provider = await db.get(ModelProvider, provider_id)
+        if not provider:
+            raise HTTPException(status_code=404, detail=f"模型提供商 {provider_id} 不存在")
+        out_args: dict[str, Any] = {"provider_id": provider_id}
+        changes: list[str] = []
+        if args.get("default_model") is not None:
+            new_model = str(args["default_model"])
+            out_args["default_model"] = new_model
+            changes.append(f"default_model {provider.default_model} → {new_model}")
+        if args.get("enabled") is not None:
+            out_args["enabled"] = bool(args["enabled"])
+            state = "启用" if out_args["enabled"] else "停用"
+            changes.append(f"{state} (enabled {provider.enabled} → {out_args['enabled']})")
+        if not changes:
+            raise HTTPException(status_code=400, detail="update_provider 未指定要改的字段 (default_model 或 enabled)")
+        return Proposal(
+            tool=name,
+            args=out_args,
+            summary=f"配置 AI 模型提供商「{provider.name}」",
+            diff="; ".join(changes),
+        )
     raise HTTPException(status_code=400, detail=f"unknown write tool: {name}")
 
 
@@ -353,20 +412,31 @@ async def confirm(body: ConfirmRequest, db: AsyncSession = Depends(get_db)) -> A
         logger.info("chat confirm | update_schedule %s %s", schedule.id, fields)
         return ApiResponse.ok({"applied": True, "tool": proposal.tool, "summary": proposal.summary})
 
+    if proposal.tool == "update_provider":
+        provider = await db.get(ModelProvider, args.get("provider_id", ""))
+        if not provider:
+            raise HTTPException(status_code=404, detail="模型提供商不存在")
+        if "default_model" in args:
+            provider.default_model = str(args["default_model"])
+        if "enabled" in args:
+            provider.enabled = bool(args["enabled"])
+        await db.commit()
+        logger.info(
+            "chat confirm | update_provider %s %s",
+            provider.id, {k: args[k] for k in ("default_model", "enabled") if k in args},
+        )
+        return ApiResponse.ok({"applied": True, "tool": proposal.tool, "summary": proposal.summary})
+
     raise HTTPException(status_code=400, detail=f"unknown proposal tool: {proposal.tool}")
 
 
 # ── XML-style tool models (e.g. Qwable-v1: emits <tool_use> XML, not OpenAI tool_calls) ──
-# Qwable-v1 (Qwen3.6-35B distill + Claude Fable-5 tool-use) emits custom
-#   <tool_use name="X" id="...">{json}</tool_use>
-# XML in the message content instead of OpenAI `tool_calls`. We describe the
-# tools in the system prompt as text and parse the XML ourselves.
-XML_TOOL_MODELS = ("qwable",)
-
-# matches both <tool_use name="X" .../> (self-closing) and <tool_use name="X" ...>{json}</tool_use>
-_TOOL_USE_RE = re.compile(
-    r'<tool_use\s+name="([^"]+)"[^>]*?(?:/\s*>|>\s*(\{.*?\}|)\s*</tool_use>)', re.DOTALL
-)
+# XML tool-call parsing (XML_TOOL_MODELS / _TOOL_USE_RE / _is_xml_tool_model /
+# _parse_tool_use / _safe_json) is shared with the skill execute core — the single
+# source of truth lives in backend.skills.toolcall (imported above). Qwable-v1
+# emits custom <tool_use name="X">{json}</tool_use> in the message content instead
+# of OpenAI tool_calls; we describe the tools in the system prompt as text
+# (XML_TOOL_TEXT) and parse the XML ourselves via the imported helpers.
 _THINK_RE = re.compile(r"<think>.*?</think>", re.DOTALL)
 
 XML_TOOL_TEXT = (
@@ -377,23 +447,13 @@ async def confirm(body: ConfirmRequest, db: AsyncSession = Depends(get_db)) -> A
     "- toggle_source(source_id, enabled): 启用/停用数据源 (写)。\n"
     "- trigger_task(source_id): 立即触发一次采集 (写)。\n"
     "- update_schedule(schedule_id, cron_expression?, enabled?): 改调度 cron 或启停 (写)。\n"
+    "- list_providers(): 列出模型提供商 (id/name/default_model/enabled)。\n"
+    "- update_provider(provider_id, default_model?, enabled?): 配置 AI 富化阶段的模型提供商, 改模型或启停 (写)。\n"
     '需要调用工具时, 严格输出 XML: <tool_use name="工具名" id="toolu_1">{json 参数}</tool_use>\n'
     "先用 list_* 拿到真实 id 再做写操作。不要用 markdown 代码块。"
 )
 
 
-def _is_xml_tool_model(model: str) -> bool:
-    m = model.lower()
-    return any(k in m for k in XML_TOOL_MODELS)
-
-
-def _parse_tool_use(content: str) -> list[tuple[str, dict[str, Any]]]:
-    calls: list[tuple[str, dict[str, Any]]] = []
-    for match in _TOOL_USE_RE.finditer(content or ""):
-        calls.append((match.group(1), _safe_json(match.group(2) or "{}")))
-    return calls
-
-
 async def _chat_xml(client: Any, model: str, system: str, body: ChatRequest, db: AsyncSession) -> ApiResponse:
     """Tool loop for XML-style models (parse <tool_use> from content, feed results back as text)."""
     messages: list[dict[str, Any]] = [{"role": "system", "content": system + XML_TOOL_TEXT}]
@@ -428,11 +488,3 @@ async def _chat_xml(client: Any, model: str, system: str, body: ChatRequest, db:
             )
 
     return ApiResponse.ok(ChatReply(type="message", content="(达到工具调用步数上限, 请换个说法再试)"))
-
-
-def _safe_json(raw: str) -> dict[str, Any]:
-    try:
-        value = json.loads(raw or "{}")
-        return value if isinstance(value, dict) else {}
-    except json.JSONDecodeError:
-        return {}

backend/api/v1/cookies.py

@@ -0,0 +1,28 @@
+"""Admin endpoint for CookieCloud sync — manual trigger only (v1; no scheduled
+sync, see backend/auth/cookiecloud_sync.py docstring). Credentials are passed
+per call, not persisted, matching the "no new settings storage" tight scope."""
+
+from __future__ import annotations
+
+from fastapi import APIRouter, HTTPException
+from pydantic import BaseModel
+
+from backend.auth.cookiecloud_sync import CookieCloudSyncError, sync_from_cookiecloud
+from backend.schemas.common import ApiResponse
+
+router = APIRouter(prefix="/cookies", tags=["cookies"])
+
+
+class CookieCloudSyncRequest(BaseModel):
+    url: str
+    uuid: str
+    password: str
+
+
+@router.post("/sync", response_model=ApiResponse[dict])
+async def sync_cookies(body: CookieCloudSyncRequest) -> ApiResponse:
+    try:
+        synced = await sync_from_cookiecloud(body.url, body.uuid, body.password)
+    except CookieCloudSyncError as exc:
+        raise HTTPException(status_code=502, detail=str(exc)) from exc
+    return ApiResponse.ok({"synced": synced})