feat(v2): 模型设置（用户可调小模型/大模型/管线，/config + UI）

.gitignore

@@ -12,6 +12,7 @@ htmlcov/
 .env
 *.env
 api_keys*.json
+config/settings.json
 
 # Models / data / runtime
 models/

gateway/api.py

@@ -39,6 +39,7 @@ _router: Optional[Router] = None
 _pipeline: Optional["CollaborativePipeline"] = None
 _v2stats = V2Stats() if _V2_OK else None
 _review = None
+_settings = None
 
 
 def get_router() -> Router:
@@ -56,8 +57,23 @@ def get_review() -> "ReviewQueue":
     return _review
 
 
+def settings_store():
+    """用户可调整设置（懒加载单例）。"""
+    global _settings
+    if _settings is None:
+        from gateway.settings import load_settings
+        _settings = load_settings()
+    return _settings
+
+
+def rebuild_pipeline() -> None:
+    """清除管线单例，下次调用重建（配置改动后生效）。"""
+    global _pipeline
+    _pipeline = None
+
+
 def build_v2_pipeline(worker_cfg_override: Optional[dict] = None):
-    """从配置构建 v2 协作管线（architect + worker + pipeline）。
+    """从配置 + 用户设置构建 v2 协作管线（architect + worker + pipeline）。
 
     worker_cfg_override 可注入（测试/演示用 mock）。无 API key 时 /chat 会走
     本地降级路径（不崩溃）。
@@ -65,13 +81,27 @@ def build_v2_pipeline(worker_cfg_override: Optional[dict] = None):
     global _pipeline
     if _pipeline is None:
         cfg = load_config()
+        s = settings_store().to_dict() if _V2_OK else {}
         kb = KnowledgeBase()
-        architect = build_architect(cfg.get("architect", {}))
+
+        # architect（合并用户设置）
+        acfg = dict(cfg.get("architect", {}))
+        acfg.update(s.get("architect", {}))
+        architect = build_architect(acfg)
+
+        # worker（合并用户设置；backend 可 mock/openai/llama_server）
         wcfg = dict(cfg.get("worker", {}))
+        wcfg.update(s.get("worker", {}))
         if worker_cfg_override:
             wcfg.update(worker_cfg_override)
         worker = build_worker(wcfg, kb=kb)
-        _pipeline = build_pipeline(cfg, architect, worker)
+
+        # pipeline（合并用户设置）
+        cfg2 = dict(cfg)
+        pcfg = dict(cfg.get("pipeline", {}))
+        pcfg.update(s.get("pipeline", {}))
+        cfg2["pipeline"] = pcfg
+        _pipeline = build_pipeline(cfg2, architect, worker)
     return _pipeline
 
 
@@ -221,6 +251,31 @@ try:
             raise HTTPException(status_code=404, detail=f"审核记录不存在或已审核: {review_id}")
         return {"ok": True, "review_id": review_id, "verdict": verdict}
 
+    # ---------------- 模型设置（用户可调整） ----------------
+    @app.get("/config", tags=["settings"])
+    async def get_config():
+        """读取当前可调整设置（小模型 / 大模型 / 管线）。"""
+        return settings_store().to_dict()
+
+    @app.put("/config", tags=["settings"])
+    async def put_config(patch: dict):
+        """部分更新设置并重建管线。示例：
+        {"worker": {"backend": "openai", "base_url": "http://127.0.0.1:11434/v1", "temperature": 0.4}}
+        """
+        try:
+            merged = settings_store().update(patch)
+        except Exception as e:
+            raise HTTPException(status_code=400, detail=f"设置非法: {e}")
+        rebuild_pipeline()
+        return merged
+
+    @app.post("/config/reset", tags=["settings"])
+    async def reset_config():
+        """恢复默认设置并重建管线。"""
+        merged = settings_store().reset()
+        rebuild_pipeline()
+        return merged
+
     # ---------------- metrics ----------------
     @app.get("/metrics", tags=["system"])
     async def metrics():

gateway/settings.py

@@ -0,0 +1,114 @@
+"""可调整的运行设置（SettingsStore）—— 让用户自定义内置小模型 / 大模型 / 管线。
+
+用户可在 Web 界面"模型设置"里调整并持久化到 config/settings.json（gitignore），
+重启后保留。调整会触发 v2 管线重建（gateway.api.build_v2_pipeline 重新读取）。
+
+默认值与 config/config.yaml 的 v2 段一致；settings.json 只存用户改动覆盖项。
+"""
+from __future__ import annotations
+
+import json
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+_SETTINGS_PATH = Path(__file__).resolve().parent.parent / "config" / "settings.json"
+
+# 可调整项（含默认值）；用户改动存这里
+DEFAULTS: Dict[str, Any] = {
+    "worker": {
+        "backend": "llama_server",      # mock | openai/api | llama_server
+        "model": "models/qwen3.5-4b-q4_k_m.gguf",
+        "base_url": "",                 # openai 后端填 http://127.0.0.1:11434/v1 等
+        "port": 8901,                   # llama_server 端口
+        "temperature": 0.3,
+        "max_fix_attempts": 2,
+        "code_timeout_s": 10,
+    },
+    "architect": {
+        "model": "deepseek-chat",
+        "base_url": "https://api.deepseek.com/v1",
+        "api_key_env": "DEEPSEEK_API_KEY",
+    },
+    "pipeline": {
+        "fast_path": True,
+        "rounds_cap": 6,
+        "api_token_cap": 8000,
+        "breach_policy": "architect_do",
+    },
+}
+
+
+class SettingsStore:
+    """用户可调整设置（内存 + settings.json 持久化）。"""
+
+    def __init__(self, path: Optional[Path] = None):
+        self._path = Path(path) if path else _SETTINGS_PATH
+        self._data: Dict[str, Any] = {}
+        self.load()
+
+    # ---------- 持久化 ----------
+    def load(self) -> None:
+        if self._path.exists():
+            try:
+                self._data = json.loads(self._path.read_text(encoding="utf-8"))
+            except Exception:
+                self._data = {}
+        else:
+            self._data = {}
+
+    def save(self) -> None:
+        self._path.parent.mkdir(parents=True, exist_ok=True)
+        self._path.write_text(
+            json.dumps(self._data, ensure_ascii=False, indent=2), encoding="utf-8")
+
+    # ---------- 访问 ----------
+    def to_dict(self) -> Dict[str, Any]:
+        """返回 默认值 + 用户覆盖 合并后的完整设置。"""
+        merged: Dict[str, Any] = {}
+        for section, defaults in DEFAULTS.items():
+            ov = self._data.get(section, {})
+            merged[section] = {**defaults, **(ov if isinstance(ov, dict) else {})}
+        return merged
+
+    def get(self, section: str, key: str, default: Any = None) -> Any:
+        merged = self.to_dict()
+        return merged.get(section, {}).get(key, default)
+
+    def update(self, patch: Dict[str, Any]) -> Dict[str, Any]:
+        """应用部分更新（可只传改动的 section/key）。返回合并后的完整设置。"""
+        for section, values in patch.items():
+            if section not in DEFAULTS or not isinstance(values, dict):
+                continue
+            cur = self._data.setdefault(section, {})
+            for k, v in values.items():
+                if k in DEFAULTS[section]:
+                    cur[k] = _coerce(v, DEFAULTS[section][k])
+        self.save()
+        return self.to_dict()
+
+    def reset(self) -> Dict[str, Any]:
+        """恢复默认。"""
+        self._data = {}
+        self.save()
+        return self.to_dict()
+
+
+def _coerce(value: Any, template: Any) -> Any:
+    """按默认值的类型把输入转成一致类型（数值容错）。"""
+    if isinstance(template, bool):
+        return bool(value)
+    if isinstance(template, int):
+        try:
+            return int(float(value))
+        except (TypeError, ValueError):
+            return template
+    if isinstance(template, float):
+        try:
+            return float(value)
+        except (TypeError, ValueError):
+            return template
+    return value
+
+
+def load_settings(path: Optional[Path] = None) -> SettingsStore:
+    return SettingsStore(path)

gateway/static/index.html

@@ -58,6 +58,9 @@ input,select{background:var(--panel2);border:1px solid var(--line);color:var(--t
 .muted{color:var(--muted)}
 .history{max-height:220px;overflow:auto;margin-top:8px}
 table{width:100%;border-collapse:collapse;font-size:13px}
+.form-grid{display:grid;grid-template-columns:repeat(auto-fill,minmax(220px,1fr));gap:12px}
+.form-grid label{display:flex;flex-direction:column;gap:5px;font-size:12px;color:var(--muted)}
+.form-grid input,.form-grid select{width:100%}
 th,td{text-align:left;padding:8px;border-bottom:1px solid var(--line)}
 th{color:var(--muted);font-weight:600}
 footer{padding:8px 20px;color:var(--muted);font-size:11px;border-top:1px solid var(--line);background:var(--panel)}
@@ -74,6 +77,7 @@ footer{padding:8px 20px;color:var(--muted);font-size:11px;border-top:1px solid v
   <button data-tab="work">🔗 协作过程</button>
   <button data-tab="review">🧑‍💻 人工检验</button>
   <button data-tab="metrics">📊 指标</button>
+  <button data-tab="settings">⚙️ 模型设置</button>
 </nav>
 <main>
   <!-- 对话 -->
@@ -115,6 +119,53 @@ footer{padding:8px 20px;color:var(--muted);font-size:11px;border-top:1px solid v
     <div class="row-btns" style="margin-bottom:10px"><button class="small" id="refreshMetrics">刷新</button></div>
     <div id="metricsResult"></div>
   </section>
+  <!-- 模型设置 -->
+  <section class="panel" id="panel-settings">
+    <div class="card">
+      <h3>⚙️ 内置小模型（Worker）</h3>
+      <div class="form-grid">
+        <label>后端 backend
+          <select id="set-worker-backend">
+            <option value="llama_server">llama_server（内置本地模型）</option>
+            <option value="openai">openai/api（Ollama / vLLM 等兼容端点）</option>
+            <option value="mock">mock（演示，零运行时）</option>
+          </select>
+        </label>
+        <label>模型路径/名称 <input id="set-worker-model" type="text" placeholder="models/xxx.gguf 或模型名"></label>
+        <label>端点 base_url（openai 用） <input id="set-worker-baseurl" type="text" placeholder="http://127.0.0.1:11434/v1"></label>
+        <label>端口（llama_server） <input id="set-worker-port" type="number"></label>
+        <label>temperature <input id="set-worker-temperature" type="number" step="0.1"></label>
+        <label>max_fix_attempts <input id="set-worker-fix" type="number"></label>
+      </div>
+    </div>
+    <div class="card">
+      <h3>🧠 大模型（Architect / API）</h3>
+      <div class="form-grid">
+        <label>model <input id="set-arch-model" type="text"></label>
+        <label>base_url <input id="set-arch-baseurl" type="text"></label>
+        <label>api_key_env <input id="set-arch-keyenv" type="text"></label>
+      </div>
+    </div>
+    <div class="card">
+      <h3>🔀 管线 pipeline</h3>
+      <div class="form-grid">
+        <label>fast_path（快路径直答） <input id="set-pipe-fast" type="checkbox"></label>
+        <label>rounds_cap <input id="set-pipe-rounds" type="number"></label>
+        <label>api_token_cap <input id="set-pipe-tokens" type="number"></label>
+        <label>breach_policy
+          <select id="set-pipe-breach">
+            <option value="architect_do">architect_do（兜底代做）</option>
+            <option value="local_only">local_only（本地降级）</option>
+          </select>
+        </label>
+      </div>
+    </div>
+    <div class="row-btns">
+      <button class="primary" id="saveSettingsBtn">保存并生效</button>
+      <button class="small" id="resetSettingsBtn">恢复默认</button>
+      <span class="muted" id="settingsMsg"></span>
+    </div>
+  </section>
 </main>
 <footer>端云协同 LLM 协作系统 · v2 · 交流文本协议 · 北极星：token 下降 ≥80%</footer>
 <script>
@@ -123,7 +174,7 @@ let lastRequestId=null;
 const esc=s=>{const d=document.createElement("div");d.textContent=s==null?"":String(s);return d.innerHTML;};
 function el(tag,cls,txt){const n=document.createElement(tag);if(cls)n.className=cls;if(txt!=null)n.textContent=txt;return n;}
 function setTab(t){document.querySelectorAll("nav button").forEach(b=>b.classList.toggle("active",b.dataset.tab===t));document.querySelectorAll(".panel").forEach(p=>p.classList.toggle("active",p.id==="panel-"+t));}
-document.querySelectorAll("nav button").forEach(b=>b.onclick=()=>{setTab(b.dataset.tab);if(b.dataset.tab==="review")loadReview();if(b.dataset.tab==="metrics")loadMetrics();});
+document.querySelectorAll("nav button").forEach(b=>b.onclick=()=>{setTab(b.dataset.tab);if(b.dataset.tab==="review")loadReview();if(b.dataset.tab==="metrics")loadMetrics();if(b.dataset.tab==="settings")loadSettings();});
 
 // ---- 健康状态 ----
 async function health(){try{const r=await fetch("/health");const d=await r.json();$("healthPill").textContent="健康 · "+d.domains.length+" 领域";}catch(e){$("healthPill").textContent="离线";}}
@@ -154,6 +205,12 @@ async function sendLegacy(){const q=$("query").value.trim();if(!q)return;const h
 const SAMPLES=["用 Python 实现快速排序，并分析时间与空间复杂度","求解方程 x^2 - 5x + 6 = 0","高血压患者日常饮食需要注意什么","解释一下深度学习中的注意力机制"];
 function sample(){$("query").value=SAMPLES[Math.floor(Math.random()*SAMPLES.length)];}
 
+// ---- 模型设置 ----
+async function loadSettings(){try{const r=await fetch("/config");const s=await r.json();const w=s.worker||{};$("set-worker-backend").value=w.backend||"llama_server";$("set-worker-model").value=w.model||"";$("set-worker-baseurl").value=w.base_url||"";$("set-worker-port").value=w.port||8901;$("set-worker-temperature").value=w.temperature||0.3;$("set-worker-fix").value=w.max_fix_attempts||2;const a=s.architect||{};$("set-arch-model").value=a.model||"";$("set-arch-baseurl").value=a.base_url||"";$("set-arch-keyenv").value=a.api_key_env||"";const p=s.pipeline||{};$("set-pipe-fast").checked=!!p.fast_path;$("set-pipe-rounds").value=p.rounds_cap||6;$("set-pipe-tokens").value=p.api_token_cap||8000;$("set-pipe-breach").value=p.breach_policy||"architect_do";$("settingsMsg").textContent="";}catch(e){$("settingsMsg").textContent="读取失败："+e.message;}}
+async function saveSettings(){const patch={worker:{backend:$("set-worker-backend").value,model:$("set-worker-model").value,base_url:$("set-worker-baseurl").value,port:parseInt($("set-worker-port").value||8901),temperature:parseFloat($("set-worker-temperature").value||0.3),max_fix_attempts:parseInt($("set-worker-fix").value||2)},architect:{model:$("set-arch-model").value,base_url:$("set-arch-baseurl").value,api_key_env:$("set-arch-keyenv").value},pipeline:{fast_path:$("set-pipe-fast").checked,rounds_cap:parseInt($("set-pipe-rounds").value||6),api_token_cap:parseInt($("set-pipe-tokens").value||8000),breach_policy:$("set-pipe-breach").value}};try{const r=await fetch("/config",{method:"PUT",headers:{"Content-Type":"application/json"},body:JSON.stringify(patch)});const d=await r.json();$("settingsMsg").textContent=r.ok?"✅ 已保存并重建管线":"❌ "+(d.detail||"保存失败");}catch(e){$("settingsMsg").textContent="保存失败："+e.message;}}
+async function resetSettings(){try{await fetch("/config/reset",{method:"POST"});$("settingsMsg").textContent="已恢复默认设置";loadSettings();}catch(e){$("settingsMsg").textContent="重置失败："+e.message;}}
+$("saveSettingsBtn").onclick=saveSettings;
+$("resetSettingsBtn").onclick=resetSettings;
 $("sendBtn").onclick=sendChat;
 $("legacyBtn").onclick=sendLegacy;
 $("sampleBtn").onclick=sample;

router_system/worker.py

@@ -155,11 +155,15 @@ class WorkerLoop:
 def build_worker(cfg: Dict[str, Any], kb: Any = None,
                  generate: Optional[Callable[[str], Awaitable[str]]] = None) -> WorkerLoop:
     """cfg 为 config.worker 段。generate 缺省时按 backend 选择：
-    mock（零运行时演示）| llama_server（真实本地模型，惰性连接）。"""
+    mock（零运行时演示）| openai/api（任意 OpenAI 兼容端点，如 Ollama/vLLM）|
+    llama_server（内置本地 llama-server）。"""
     backend = cfg.get("backend", "llama_server")
     if generate is None:
         if backend == "mock":
             generate = _mock_generate()
+        elif backend in ("openai", "api"):
+            generate = _make_llama_generate(
+                cfg, default_base_url=cfg.get("base_url") or "http://127.0.0.1:11434/v1")
         else:
             generate = _make_llama_generate(cfg)
     verifier = Verifier(code_timeout_s=float(cfg.get("code_timeout_s", 10)))
@@ -182,10 +186,13 @@ def _mock_generate() -> Callable[[str], Awaitable[str]]:
     return _gen
 
 
-def _make_llama_generate(cfg: Dict[str, Any]) -> Callable[[str], Awaitable[str]]:
+def _make_llama_generate(cfg: Dict[str, Any],
-    """返回调用本地 llama-server（OpenAI 兼容 /v1/chat/completions）的生成器。"""
+                           default_base_url: Optional[str] = None) -> Callable[[str], Awaitable[str]]:
-    base_url = cfg.get("base_url", f"http://127.0.0.1:{cfg.get('port', 8901)}/v1")
+    """返回调用本地 OpenAI 兼容端点（llama-server / Ollama / vLLM）的生成器。"""
-    model = cfg.get("model", "local")
+    if default_base_url is None:
+        default_base_url = f"http://127.0.0.1:{cfg.get('port', 8901)}/v1"
+    base_url = cfg.get("base_url") or default_base_url
+    model = cfg.get("model") or "local"
     temperature = float(cfg.get("temperature", 0.3))
     timeout_s = float(cfg.get("per_step_timeout_s", 300))
 
@@ -201,8 +208,8 @@ def _make_llama_generate(cfg: Dict[str, Any]) -> Callable[[str], Awaitable[str]]
                 resp.raise_for_status()
                 return resp.json()["choices"][0]["message"]["content"]
         except Exception as e:  # noqa: BLE001
-            # 连不上本地模型 -> 优雅降级（不抛 500），提示用户准备运行时
+            # 连不上本地模型 -> 优雅降级（不抛 500），提示用户检查模型端点
-            return ("（本地降级）无法连接本地 llama-server，未能生成该步骤内容。"
+            return ("（本地降级）无法连接本地模型端点，未能生成该步骤内容。"
-                    f"请先运行 scripts/setup_runtime.py 并启动模型。错误：{type(e).__name__}")
+                    f"请检查模型后端配置或启动服务。错误：{type(e).__name__}")
 
     return _gen

tests/test_gateway.py

@@ -66,6 +66,24 @@ def test_metrics(client):
     assert "review" in data
 
 
+
+
+
+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"
+
+
 def test_workspace_not_found(client):
     resp = client.get("/runs/nonexistent/workspace")
     assert resp.status_code == 404

tests/test_settings.py

@@ -0,0 +1,50 @@
+"""SettingsStore 单测（封闭：临时路径）。"""
+import json
+
+from gateway.settings import DEFAULTS, SettingsStore, _coerce
+
+
+def test_defaults(tmp_path):
+    s = SettingsStore(path=tmp_path / "s.json")
+    d = s.to_dict()
+    assert d["worker"]["backend"] == "llama_server"
+    assert "pipeline" in d and "architect" in d
+
+
+def test_update_and_coerce(tmp_path):
+    s = SettingsStore(path=tmp_path / "s.json")
+    s.update({"worker": {"backend": "openai", "temperature": "0.4", "max_fix_attempts": "3"}})
+    d = s.to_dict()
+    assert d["worker"]["backend"] == "openai"
+    assert d["worker"]["temperature"] == 0.4      # 字符串转 float
+    assert d["worker"]["max_fix_attempts"] == 3   # 字符串转 int
+
+
+def test_ignores_unknown_keys(tmp_path):
+    s = SettingsStore(path=tmp_path / "s.json")
+    s.update({"worker": {"not_a_key": 1}, "unknown_section": {"x": 1}})
+    d = s.to_dict()
+    assert "not_a_key" not in d["worker"]
+    assert "unknown_section" not in d
+
+
+def test_persistence_roundtrip(tmp_path):
+    p = tmp_path / "s.json"
+    s = SettingsStore(path=p)
+    s.update({"worker": {"temperature": 0.5}})
+    s2 = SettingsStore(path=p)  # 重新加载
+    assert s2.to_dict()["worker"]["temperature"] == 0.5
+
+
+def test_reset(tmp_path):
+    s = SettingsStore(path=tmp_path / "s.json")
+    s.update({"worker": {"backend": "mock"}})
+    d = s.reset()
+    assert d["worker"]["backend"] == "llama_server"
+
+
+def test_coerce():
+    assert _coerce("true", True) is True
+    assert _coerce("42", 0) == 42
+    assert _coerce("bad", 0) == 0
+    assert _coerce(1.5, 0.0) == 1.5