feat(v3): T26 两级智能体（大模型规划/审查 + 本地小模型执行，D7）

- run_dual 编排：规划者两阶段 JSON（plan/review，失败回喂重试一次再降级）， redo 时裁决意见回喂执行者，交接上限 agent.max_handoffs（默认 2） - 交接文档 agent_runs/{id}/handoff.json（智能体版交流文本：instructions/acceptance/exchanges） - /agent 新增 executor_pool_id；规划者==执行者条目拒绝；整体 token_cap 覆盖两级调用 - ToolLoop 增 emit_final 开关（内层循环不发终态，防前端 SSE 提前收口） - 前端：执行者选择器 + phase/message 事件渲染（阶段徽标 + 双色消息卡） - 测试 +3（done/redo/执行者故障），全量 277 passed - fix(tests): test_config_get_put_reset 增加设置备份/恢复隔离，防止清掉用户真实配置
2026-09-01 11:45:33 +08:00
parent 8e2123343c
commit 43e2bceae7
20 changed files with 568 additions and 59 deletions
@@ -244,3 +244,162 @@ def test_agent_workspace_and_file_accept_root(agent_env, client, tmp_path):
    # 非法 root -> 400
    r = client.get("/agent/workspace", params={"root": str(tmp_path / "nope")})
    assert r.status_code == 400
+
+
+# ---------------- 两级智能体（T26）：规划者 + 执行者 ----------------
+
+def _planner_resp(obj=None, raw=""):
+    content = raw or json.dumps(obj, ensure_ascii=False)
+    return {"content": content, "tool_calls": [],
+            "usage": {"prompt_tokens": 50, "completion_tokens": 20}}
+
+
+def _install_dual(agent_env, monkeypatch, planner_script, executor_script):
+    """注入假规划者（build_agent_chat）与假执行者（OpenAICompatChat）。"""
+
+    class FakePlanner:
+        api_key = "sk-fake"
+
+        def __init__(self, *a, **k):
+            self.script = list(planner_script)
+
+        async def __call__(self, messages, tools_spec):
+            if self.script:
+                return self.script.pop(0)
+            return _planner_resp({"verdict": "done", "final_answer": "（兜底）完成。"})
+
+    class FakeExecutorChat:
+        def __init__(self, *a, **k):
+            self.script = list(executor_script)
+
+        async def __call__(self, messages, tools_spec):
+            if self.script:
+                return self.script.pop(0)
+            return {"content": "（执行者兜底）没有更多动作。", "tool_calls": [], "usage": {}}
+
+    def fake_chat_factory(acfg):
+        return FakePlanner()
+
+    monkeypatch.setattr(ga, "build_agent_chat", fake_chat_factory)
+    monkeypatch.setattr(ag, "OpenAICompatChat", FakeExecutorChat)
+
+
+def test_dual_agent_done_flow(agent_env, client, monkeypatch, tmp_path):
+    """规划 -> 执行(写文件) -> 审查 done：事件/交接文档/状态全部落位。"""
+    _install_dual(
+        agent_env, monkeypatch,
+        planner_script=[
+            _planner_resp({"instructions": "在 data 目录创建 report.json",
+                           "acceptance": "文件存在且内容为合法 JSON"}),
+            _planner_resp({"verdict": "done", "reply_to_executor": "",
+                           "final_answer": "执行者已按指令创建数据文件，验收通过。"}),
+        ],
+        executor_script=[
+            {"content": None,
+             "tool_calls": [{"id": "e1", "name": "write_file",
+                             "arguments": {"path": "data/report.json",
+                                           "content": '{"ok": true}'}}],
+             "usage": {"prompt_tokens": 100, "completion_tokens": 10}},
+            {"content": "汇报：已创建 data/report.json，内容 {\"ok\": true}。",
+             "tool_calls": [], "usage": {"prompt_tokens": 120, "completion_tokens": 15}},
+        ])
+    r = client.post("/agent", json={"task": "建数据文件", "executor_pool_id": "no-such"})
+    # 执行者条目不存在 -> 400
+    assert r.status_code == 400
+
+    # 先放一个合法 llama_server 条目作为执行者
+    client.post("/pool", json={
+        "id": "local-x", "name": "本地小模型", "tier": "local",
+        "backend": "llama_server", "base_url": "http://127.0.0.1:8901/v1",
+        "model": "qwen-0.8b", "enabled": True})
+    r2 = client.post("/agent", json={"task": "建数据文件", "executor_pool_id": "local-x"})
+    assert r2.status_code == 200
+    assert r2.json()["mode"] == "dual"
+    assert "本地小模型" in r2.json()["executor_model"]
+
+    rid = r2.json()["request_id"]
+    info = _wait_done(agent_env["service"], rid)
+    assert info.state == "done", info.error
+    assert info.mode == "dual"
+    assert info.response == "执行者已按指令创建数据文件，验收通过。"
+
+    # 事件序列：规划 -> 执行(含工具) -> 审查 -> final
+    evs = agent_env["service"].read_events(rid)
+    phases = [e["phase"] for e in evs if e["type"] == "phase"]
+    assert phases == ["plan", "execute", "review"]
+    kinds = [e["type"] for e in evs]
+    assert "message" in kinds and "tool_call" in kinds
+    # 交接文档（智能体版交流文本）
+    ho = json.loads((agent_env["service"]._dir(rid) / "handoff.json").read_text(encoding="utf-8"))
+    assert ho["instructions"]
+    assert ho["exchanges"][0]["verdict"] == "done"
+    assert ho["executor_model"] == "本地小模型（qwen-0.8b）"
+
+    st = client.get(f"/agent/{rid}/status").json()
+    assert st["mode"] == "dual" and st["executor_model"]
+
+
+def test_dual_agent_redo_then_done(agent_env, client, monkeypatch):
+    """第一轮裁决 redo -> 执行者带补充指令再跑 -> 第二轮 done。"""
+    _install_dual(
+        agent_env, monkeypatch,
+        planner_script=[
+            _planner_resp({"instructions": "写 hello.txt"}),
+            _planner_resp({"verdict": "redo", "reply_to_executor": "文件内容不对，请写入 DONE",
+                           "final_answer": ""}),
+            _planner_resp({"verdict": "done", "reply_to_executor": "",
+                           "final_answer": "第二轮通过。"}),
+        ],
+        executor_script=[
+            {"content": "汇报：已写 hello.txt（内容空白）", "tool_calls": [],
+             "usage": {"prompt_tokens": 10, "completion_tokens": 5}},
+            {"content": None,
+             "tool_calls": [{"id": "e1", "name": "write_file",
+                             "arguments": {"path": "hello.txt", "content": "DONE"}}],
+             "usage": {"prompt_tokens": 10, "completion_tokens": 5}},
+            {"content": "汇报：已按补充指令重写 hello.txt 内容为 DONE",
+             "tool_calls": [], "usage": {"prompt_tokens": 10, "completion_tokens": 5}},
+        ])
+    client.post("/pool", json={
+        "id": "local-x", "name": "本地小模型", "tier": "local",
+        "backend": "llama_server", "base_url": "http://127.0.0.1:8901/v1",
+        "model": "qwen-0.8b", "enabled": True})
+    r = client.post("/agent", json={"task": "写 hello.txt", "executor_pool_id": "local-x"})
+    rid = r.json()["request_id"]
+    info = _wait_done(agent_env["service"], rid)
+    assert info.state == "done"
+    assert info.response == "第二轮通过。"
+    ho = json.loads((agent_env["service"]._dir(rid) / "handoff.json").read_text(encoding="utf-8"))
+    assert [x["verdict"] for x in ho["exchanges"]] == ["redo", "done"]
+    # 第二轮执行者应收到 redo 补充指令（消息历史含 reply_to_executor 内容）
+    evs = agent_env["service"].read_events(rid)
+    exec_phases = [e for e in evs if e["type"] == "phase" and e["phase"] == "execute"]
+    assert len(exec_phases) == 2
+
+
+def test_dual_agent_executor_error(agent_env, client, monkeypatch):
+    """执行者客户端异常 -> 任务 failed，错误透出。"""
+    class BoomChat:
+        def __init__(self, *a, **k):
+            pass
+
+        async def __call__(self, messages, tools_spec):
+            raise RuntimeError("本地模型连不上")
+
+    class PlanOK:
+        api_key = "sk-fake"
+
+        async def __call__(self, messages, tools_spec):
+            return _planner_resp({"instructions": "随便执行"})
+
+    monkeypatch.setattr(ga, "build_agent_chat", lambda acfg: PlanOK())
+    monkeypatch.setattr(ag, "OpenAICompatChat", BoomChat)
+    client.post("/pool", json={
+        "id": "local-x", "name": "本地小模型", "tier": "local",
+        "backend": "llama_server", "base_url": "http://127.0.0.1:8901/v1",
+        "model": "qwen-0.8b", "enabled": True})
+    r = client.post("/agent", json={"task": "t", "executor_pool_id": "local-x"})
+    rid = r.json()["request_id"]
+    info = _wait_done(agent_env["service"], rid)
+    assert info.state == "failed"
+    assert "RuntimeError" in (info.error or "")
@@ -86,18 +86,28 @@ def test_metrics(client):


 def test_config_get_put_reset(client):
-    # GET 默认
-    r = client.get("/config")
-    assert r.status_code == 200
-    assert r.json()["worker"]["backend"] in ("llama_server", "openai", "mock")
-    # PUT 更新 worker
-    r2 = client.put("/config", json={"worker": {"backend": "mock", "temperature": 0.5}})
-    assert r2.status_code == 200
-    assert r2.json()["worker"]["temperature"] == 0.5
-    # 重置
-    r3 = client.post("/config/reset")
-    assert r3.status_code == 200
-    assert r3.json()["worker"]["backend"] == "llama_server"
+    """/config 端到端。settings.json 是活文件（用户真实配置），
+    测试前后必须备份/恢复，禁止把用户配置清掉。"""
+    import json
+    store = ga.settings_store()
+    snapshot = json.loads(json.dumps(store._data, ensure_ascii=False))
+    try:
+        # GET 默认
+        r = client.get("/config")
+        assert r.status_code == 200
+        assert r.json()["worker"]["backend"] in ("llama_server", "openai", "mock")
+        # PUT 更新 worker
+        r2 = client.put("/config", json={"worker": {"backend": "mock", "temperature": 0.5}})
+        assert r2.status_code == 200
+        assert r2.json()["worker"]["temperature"] == 0.5
+        # 重置
+        r3 = client.post("/config/reset")
+        assert r3.status_code == 200
+        assert r3.json()["worker"]["backend"] == "llama_server"
+    finally:
+        store._data = snapshot
+        store.save()
+        ga.rebuild_pipeline()


 def test_workspace_not_found(client):