fix(grok): use Responses search tools

app/dashboard/server.py

@@ -347,6 +347,7 @@ class ReusableThreadingHTTPServer(ThreadingHTTPServer):
     {"name": "CROSSDESK_BASE_URL", "label": "Crossdesk Base URL", "group": "上游模型覆盖", "kind": "text", "default": "", "effect": "next_run"},
     {"name": "CROSSDESK_API_KEY", "label": "Crossdesk API Key", "group": "上游模型覆盖", "kind": "secret", "default": "", "effect": "next_run"},
     {"name": "DASHBOARD_GROK_MODEL", "label": "Grok 模型", "group": "牛牛美股", "kind": "text", "default": "grok-4.20-multi-agent-xhigh", "effect": "next_run"},
+    {"name": "DASHBOARD_GROK_API_MODE", "label": "Grok 搜索工具接口模式", "group": "牛牛美股", "kind": "api_mode", "default": "auto", "effect": "next_run"},
     {"name": "DASHBOARD_GROK_CONTEXT_LENGTH", "label": "Grok 模型上下文长度", "group": "牛牛美股", "kind": "context_length", "default": DEFAULT_MODEL_CONTEXT_LENGTH, "effect": "next_run"},
     {"name": "DASHBOARD_GROK_MAX_TOKENS", "label": "Grok 最大输出长度", "group": "牛牛美股", "kind": "max_tokens", "default": DEFAULT_MODEL_MAX_TOKENS, "effect": "next_run"},
     {"name": "DASHBOARD_GROK_BASE_URL", "label": "Grok API 地址", "group": "牛牛美股", "kind": "text", "default": "", "effect": "next_run"},
@@ -386,6 +387,7 @@ class ReusableThreadingHTTPServer(ThreadingHTTPServer):
 
     {"name": "X_WATCHLIST_STRICT_CONTEXT_HOLD", "label": "X 上下文缺失时暂缓发送", "group": "X 监控", "kind": "bool", "default": "0", "effect": "next_run"},
     {"name": "X_WATCHLIST_DEADLINE_SECONDS", "label": "X 总截止秒数", "group": "X 监控", "kind": "int", "default": "135", "effect": "next_run"},
+    {"name": "X_WATCHLIST_REQUEST_TIMEOUT_SECONDS", "label": "X 单账号请求超时秒数", "group": "牛牛美股", "kind": "int", "default": "45", "effect": "next_run"},
     {"name": "X_WATCHLIST_SCRIPT_ALARM_SECONDS", "label": "X 脚本 alarm 秒数", "group": "X 监控", "kind": "int", "default": "90", "effect": "next_run"},
     {"name": "X_WATCHLIST_MAX_WORKERS", "label": "X 抓取并发", "group": "X 监控", "kind": "int", "default": "5", "effect": "next_run"},
     {"name": "X_WATCHLIST_MAX_ATTEMPTS", "label": "X 抓取重试次数", "group": "X 监控", "kind": "int", "default": "1", "effect": "next_run"},
@@ -407,12 +409,14 @@ class ReusableThreadingHTTPServer(ThreadingHTTPServer):
     "DASHBOARD_ADMIN_PASSWORD",
     "DASHBOARD_US_FEATURES_ENABLED",
     "DASHBOARD_GROK_MODEL",
+    "DASHBOARD_GROK_API_MODE",
     "DASHBOARD_GROK_CONTEXT_LENGTH",
     "DASHBOARD_GROK_MAX_TOKENS",
     "DASHBOARD_GROK_BASE_URL",
     "DASHBOARD_GROK_API_KEY",
     "X_WATCHLIST_ACCOUNTS",
     "X_WATCHLIST_DAEMON_INTERVAL_SECONDS",
+    "X_WATCHLIST_REQUEST_TIMEOUT_SECONDS",
     "DASHBOARD_US_RATING_CRON",
     "US_RATING_CONTEXT_LENGTH",
     "US_RATING_MAX_TOKENS",
@@ -2369,6 +2373,19 @@ def normalize_env_update(name: str, value: str, kind: str) -> str:
         int(value)
     if kind in {"max_tokens", "context_length"}:
         return normalize_context_length_update(value)
+    if kind == "api_mode":
+        normalized = value.lower().replace("-", "_") or "auto"
+        aliases = {
+            "auto": "auto",
+            "responses": "responses",
+            "response": "responses",
+            "chat": "chat",
+            "chat_completions": "chat",
+            "chat_completion": "chat",
+        }
+        if normalized not in aliases:
+            raise ValueError("API 接口模式必须是 auto、responses 或 chat")
+        return aliases[normalized]
     if kind == "time":
         normalized = normalize_hhmm(value)
         if value and not normalized:
@@ -2751,13 +2768,15 @@ def removed_notification_config_names(channel_ids: set[str] | list[str] | tuple[
     "US_RATING_CONTEXT_LENGTH",
     "US_RATING_MAX_TOKENS",
     "DASHBOARD_GROK_MODEL",
+    "DASHBOARD_GROK_API_MODE",
     "DASHBOARD_GROK_CONTEXT_LENGTH",
     "DASHBOARD_GROK_MAX_TOKENS",
     "DASHBOARD_GROK_BASE_URL",
     "DASHBOARD_GROK_API_KEY",
     "X_WATCHLIST_ACCOUNTS",
     "X_WATCHLIST_MAX_TOKENS",
     "X_WATCHLIST_DAEMON_INTERVAL_SECONDS",
+    "X_WATCHLIST_REQUEST_TIMEOUT_SECONDS",
     "DASHBOARD_US_RATING_CRON",
     "US_RATING_DEADLINE_SECONDS",
     "US_RATING_REQUEST_TIMEOUT_SECONDS",
@@ -2952,6 +2971,8 @@ def normalize_business_updates(updates: dict[str, str]) -> dict[str, str]:
             normalized[name] = normalize_preset_strategy_text_update(normalized[name])
         elif ENV_CONFIG_BY_NAME.get(name, {}).get("kind") == "trade_discipline_text":
             normalized[name] = normalize_trade_discipline_text_update(normalized[name])
+        elif ENV_CONFIG_BY_NAME.get(name, {}).get("kind") == "api_mode":
+            normalized[name] = normalize_env_update(name, normalized[name], "api_mode")
         elif ENV_CONFIG_BY_NAME.get(name, {}).get("kind") in {"max_tokens", "context_length"}:
             normalized[name] = normalize_context_length_update(normalized[name])
     return normalized
@@ -3018,6 +3039,10 @@ def validate_business_updates(updates: dict[str, str]) -> None:
             timeout = int(value)
             if timeout < 1 or timeout > 30:
                 raise ValueError(f"{name} 必须在 1 到 30 之间")
+        elif name == "X_WATCHLIST_REQUEST_TIMEOUT_SECONDS" and str(value or "").strip():
+            timeout = int(value)
+            if timeout < 8 or timeout > 120:
+                raise ValueError(f"{name} 必须在 8 到 120 之间")
         elif name in {
             "DASHBOARD_MAX_SINGLE_POSITION_PCT",
             "DASHBOARD_MAX_TOTAL_POSITION_PCT",

app/monitoring/x/monitor_service.py

@@ -97,6 +97,7 @@
 def load_dashboard_env() -> None:
     allowed = {
         "DASHBOARD_GROK_MODEL",
+        "DASHBOARD_GROK_API_MODE",
         "DASHBOARD_GROK_CONTEXT_LENGTH",
         "DASHBOARD_GROK_BASE_URL",
         "DASHBOARD_GROK_API_KEY",
@@ -106,6 +107,7 @@ def load_dashboard_env() -> None:
         "X_WATCHLIST_BASE_URL",
         "X_WATCHLIST_API_KEY",
         "X_WATCHLIST_ACCOUNTS",
+        "X_WATCHLIST_REQUEST_TIMEOUT_SECONDS",
         "CROSSDESK_BASE_URL",
         "CROSSDESK_API_KEY",
     }
@@ -150,6 +152,7 @@ def configured_max_tokens(default: int) -> int:
 
 
 MODEL = os.environ.get("X_WATCHLIST_MODEL") or os.environ.get("DASHBOARD_GROK_MODEL") or "grok-4.20-multi-agent-xhigh"
+GROK_API_MODE = os.environ.get("DASHBOARD_GROK_API_MODE") or "auto"
 X_WATCHLIST_CONTEXT_LENGTH = env_token_count("X_WATCHLIST_CONTEXT_LENGTH", "DASHBOARD_GROK_CONTEXT_LENGTH", default=128000)
 X_WATCHLIST_MAX_TOKENS = env_token_count("X_WATCHLIST_MAX_TOKENS", default=4096)
 CROSSDESK_PROVIDER_NAME = "Crossdesk.ccwu.cc"
@@ -159,7 +162,13 @@ def configured_max_tokens(default: int) -> int:
 # Grok-backed X fetching can intermittently return empty/non-JSON content or run slow;
 # those should be treated as transient poll misses, not user-visible job failures.
 TOTAL_DEADLINE_SECONDS = 135
-REQUEST_TIMEOUT_SECONDS = 25
+try:
+    REQUEST_TIMEOUT_SECONDS = max(
+        8,
+        min(120, int(os.environ.get("X_WATCHLIST_REQUEST_TIMEOUT_SECONDS") or "45")),
+    )
+except (TypeError, ValueError):
+    REQUEST_TIMEOUT_SECONDS = 45
 DETAIL_REQUEST_TIMEOUT_SECONDS = 8
 REPAIR_REQUEST_TIMEOUT_SECONDS = 10
 HELD_CONTEXT_REPAIR_TIMEOUT_SECONDS = 8
@@ -257,14 +266,63 @@ def is_temporary_error(exc):
     return False
 
 
-def openai_chat_json(base_url, api_key, prompt, max_tokens, timeout=REQUEST_TIMEOUT_SECONDS):
-    payload = {
-        "model": MODEL,
-        "messages": [{"role": "user", "content": prompt}],
-        "max_tokens": max_tokens,
-    }
+def responses_output_text(data):
+    direct = str(data.get("output_text") or "").strip()
+    if direct:
+        return direct
+    parts = []
+    for item in data.get("output") or []:
+        if not isinstance(item, dict):
+            continue
+        for content in item.get("content") or []:
+            if not isinstance(content, dict):
+                continue
+            if content.get("type") in {"output_text", "text"} and content.get("text"):
+                parts.append(str(content["text"]))
+    return "\n".join(parts).strip()
+
+
+def uses_responses_api(mode, model):
+    normalized = str(mode or "auto").strip().lower().replace("-", "_")
+    if normalized in {"responses", "response"}:
+        return True
+    if normalized in {"chat", "chat_completions", "chat_completion"}:
+        return False
+    # Preserve zero-config compatibility for the first Grok generation that
+    # requires tool-backed Responses requests, while allowing gateways and
+    # future model aliases to opt in explicitly.
+    return str(model or "").strip().lower().startswith("grok-4.5")
+
+
+def openai_chat_json(base_url, api_key, prompt, max_tokens, timeout=REQUEST_TIMEOUT_SECONDS, x_handles=None):
+    use_responses_tools = uses_responses_api(GROK_API_MODE, MODEL)
+    if use_responses_tools:
+        handles = []
+        for raw_handle in x_handles or []:
+            handle = str(raw_handle or "").strip().lstrip("@").lower()
+            if handle and handle not in handles:
+                handles.append(handle)
+        tool = {"type": "x_search"}
+        if handles:
+            tool["allowed_x_handles"] = handles[:20]
+        payload = {
+            "model": MODEL,
+            "input": [{"role": "user", "content": prompt}],
+            "tools": [tool],
+            "reasoning": {"effort": "low"},
+            "max_output_tokens": max_tokens,
+            "stream": False,
+        }
+        endpoint = base_url + "/responses"
+    else:
+        payload = {
+            "model": MODEL,
+            "messages": [{"role": "user", "content": prompt}],
+            "max_tokens": max_tokens,
+        }
+        endpoint = base_url + "/chat/completions"
     req = urllib.request.Request(
-        base_url + "/chat/completions",
+        endpoint,
         data=json.dumps(payload).encode("utf-8"),
         headers={
             "Authorization": "Bearer " + api_key,
@@ -297,7 +355,11 @@ def openai_chat_json(base_url, api_key, prompt, max_tokens, timeout=REQUEST_TIME
         content = "".join(content_parts)
     else:
         data = json.loads(raw)
-        content = data.get("choices", [{}])[0].get("message", {}).get("content", "")
+        content = (
+            responses_output_text(data)
+            if use_responses_tools
+            else data.get("choices", [{}])[0].get("message", {}).get("content", "")
+        )
     return extract_json(content)
 
 
@@ -325,6 +387,7 @@ def call_grok_once(base_url, api_key, account_handles, latest_by_handle, timeout
         prompt,
         configured_max_tokens(min(3000, 1000 + 500 * len(account_handles))),
         timeout=timeout,
+        x_handles=account_handles,
     )
     return parsed.get("accounts", [])
 
@@ -370,6 +433,8 @@ def hydrate_posts(base_url, api_key, new_items, timeout=DETAIL_REQUEST_TIMEOUT_S
 - 如果推文或引用/回复里有图片/视频/GIF，尽量返回可打开的媒体 URL；不需要识图、OCR 或图片内容描述。
 """
     try:
+        # Context may belong to an unmonitored reply or quote author, so this
+        # lookup must not inherit the initial account-fetch allowlist.
         parsed = openai_chat_json(
             base_url,
             api_key,
@@ -609,7 +674,15 @@ def repair_one_context(base_url, api_key, display_name, post, post_id, handle, t
 - reply_to_chinese_text / quoted_chinese_text 只填中文：外文原帖只给中文翻译，中文原帖只给中文原文；不要中英双语，不要加“翻译：”。
 - 如果原推包含图片/视频/GIF，尽量返回 reply_to_media/quoted_media 的可打开 URL；不需要识图、OCR 或图片内容描述。
 """
-    parsed = openai_chat_json(base_url, api_key, prompt, configured_max_tokens(3000), timeout=timeout)
+    # The parent or quoted post can belong to a different account. Restricting
+    # X Search to the monitored handle would make that context unreachable.
+    parsed = openai_chat_json(
+        base_url,
+        api_key,
+        prompt,
+        configured_max_tokens(3000),
+        timeout=timeout,
+    )
     if not isinstance(parsed, dict):
         return post
     merged = dict(post)

app/reports/us/rating_report.py

@@ -32,6 +32,7 @@
 def load_dashboard_env() -> None:
     allowed = {
         "DASHBOARD_GROK_MODEL",
+        "DASHBOARD_GROK_API_MODE",
         "DASHBOARD_GROK_CONTEXT_LENGTH",
         "DASHBOARD_GROK_BASE_URL",
         "DASHBOARD_GROK_API_KEY",
@@ -77,6 +78,7 @@ def load_dashboard_env() -> None:
 JOB_NAME = "每日美股机构买入评级汇报"
 CONFIG_PATH = Path(os.environ.get("DASHBOARD_CONFIG") or str(DASHBOARD_HOME / "config.yaml")).expanduser()
 US_RATING_MODEL = os.environ.get("US_RATING_MODEL") or os.environ.get("DASHBOARD_GROK_MODEL") or "grok-4.20-multi-agent-xhigh"
+GROK_API_MODE = os.environ.get("DASHBOARD_GROK_API_MODE") or "auto"
 
 
 def _int_env(name: str, default: int, *, min_value: int) -> int:
@@ -143,16 +145,55 @@ def _is_transient_error(err):
     return any(s in text for s in ("timed out", "timeout", "temporarily", "connection reset", "empty stream", "ssl"))
 
 
+def _responses_output_text(data: dict) -> str:
+    direct = str(data.get("output_text") or "").strip()
+    if direct:
+        return direct
+    parts: list[str] = []
+    for item in data.get("output") or []:
+        if not isinstance(item, dict):
+            continue
+        for content in item.get("content") or []:
+            if not isinstance(content, dict):
+                continue
+            if content.get("type") in {"output_text", "text"} and content.get("text"):
+                parts.append(str(content["text"]))
+    return "\n".join(parts).strip()
+
+
+def _uses_responses_api(mode: str, model: str) -> bool:
+    normalized = str(mode or "auto").strip().lower().replace("-", "_")
+    if normalized in {"responses", "response"}:
+        return True
+    if normalized in {"chat", "chat_completions", "chat_completion"}:
+        return False
+    return str(model or "").strip().lower().startswith("grok-4.5")
+
+
 def _call_api(base_url, api_key, messages, max_tokens=US_RATING_MAX_TOKENS):
-    body = json.dumps({
-        "model": US_RATING_MODEL,
-        "messages": messages,
-        "max_tokens": max_tokens,
-        "stream": False,
-    }).encode("utf-8")
+    use_responses_tools = _uses_responses_api(GROK_API_MODE, US_RATING_MODEL)
+    if use_responses_tools:
+        endpoint = f"{base_url}/responses"
+        payload = {
+            "model": US_RATING_MODEL,
+            "input": messages,
+            "tools": [{"type": "web_search"}],
+            "reasoning": {"effort": "low"},
+            "max_output_tokens": max_tokens,
+            "stream": False,
+        }
+    else:
+        endpoint = f"{base_url}/chat/completions"
+        payload = {
+            "model": US_RATING_MODEL,
+            "messages": messages,
+            "max_tokens": max_tokens,
+            "stream": False,
+        }
+    body = json.dumps(payload).encode("utf-8")
 
     req = Request(
-        f"{base_url}/chat/completions",
+        endpoint,
         data=body,
         headers={
             "Content-Type": "application/json",
@@ -173,7 +214,11 @@ def _call_api(base_url, api_key, messages, max_tokens=US_RATING_MAX_TOKENS):
             timeout_seconds = min(max(10, US_RATING_REQUEST_TIMEOUT_SECONDS), max(10, remaining - 2))
             with urlopen(req, timeout=timeout_seconds, context=_SSL_CONTEXT) as resp:
                 data = json.loads(resp.read().decode("utf-8", "ignore"))
-                content = data.get("choices", [{}])[0].get("message", {}).get("content", "")
+                content = (
+                    _responses_output_text(data)
+                    if use_responses_tools
+                    else data.get("choices", [{}])[0].get("message", {}).get("content", "")
+                )
                 if str(content or "").strip():
                     return content
                 last_err = RuntimeError("API returned empty content")

dashboard.env.example

@@ -68,6 +68,8 @@ DASHBOARD_TELEGRAM_CHAT_ID=
 
 DASHBOARD_US_FEATURES_ENABLED=0
 DASHBOARD_GROK_MODEL=grok-4.20-multi-agent-xhigh
+# auto: Grok 4.5 uses Responses search tools; responses/chat force either API.
+DASHBOARD_GROK_API_MODE=auto
 DASHBOARD_GROK_CONTEXT_LENGTH=128000
 DASHBOARD_GROK_MAX_TOKENS=4096
 DASHBOARD_GROK_BASE_URL=
@@ -109,5 +111,6 @@ A_SHARE_MODEL_SUMMARY_REQUEST_TIMEOUT_SECONDS=45
 X_WATCHLIST_CONTEXT_LENGTH=128000
 X_WATCHLIST_MAX_TOKENS=4096
 X_WATCHLIST_DAEMON_INTERVAL_SECONDS=1200
+X_WATCHLIST_REQUEST_TIMEOUT_SECONDS=45
 DASHBOARD_US_RATING_CRON="0 11 * * *"
 DASHBOARD_INDICES_TTL_SECONDS=60

docs/OPERATIONS.md

@@ -78,7 +78,7 @@ NiuOne 需要大模型驱动完整工作流。X 关注列表监控和美股机
 | 场景 | 配置项 |
 |---|---|
 | 牛牛美股总开关 | `DASHBOARD_US_FEATURES_ENABLED` |
-| Grok API | `DASHBOARD_GROK_BASE_URL`、`DASHBOARD_GROK_API_KEY`、`DASHBOARD_GROK_MODEL`、`DASHBOARD_GROK_CONTEXT_LENGTH` |
+| Grok API | `DASHBOARD_GROK_BASE_URL`、`DASHBOARD_GROK_API_KEY`、`DASHBOARD_GROK_MODEL`、`DASHBOARD_GROK_API_MODE`、`DASHBOARD_GROK_CONTEXT_LENGTH` |
 | A 股盘面模型总结单独覆盖 | `A_SHARE_MODEL_SUMMARY_BASE_URL`、`A_SHARE_MODEL_SUMMARY_API_KEY`、`A_SHARE_MODEL_SUMMARY_MODEL`、`A_SHARE_MODEL_SUMMARY_MAX_TOKENS` |
 | 消息面预检 API | `DASHBOARD_NEWS_BASE_URL`、`DASHBOARD_NEWS_API_KEY`、`DASHBOARD_NEWS_MODEL`、`DASHBOARD_NEWS_MAX_TOKENS`、`DASHBOARD_NEWS_CONCURRENCY` |
 | 买卖决策 API | `DASHBOARD_DECISION_BASE_URL`、`DASHBOARD_DECISION_API_KEY`、`DASHBOARD_DECISION_MODEL` |
@@ -89,6 +89,7 @@ NiuOne 需要大模型驱动完整工作流。X 关注列表监控和美股机
 | X 关注列表单独覆盖 | `X_WATCHLIST_BASE_URL`、`X_WATCHLIST_API_KEY`、`X_WATCHLIST_MODEL`、`X_WATCHLIST_MAX_TOKENS` |
 
 完成管理员认证后，优先通过页面上的设置按钮进入设置页维护。推文监控和美股评级相关设置由“开启牛牛美股”开关控制；关闭时设置页会隐藏这些项，后台 X 监控和美股评级定时任务也会跳过。也可以直接编辑 `.local-data/dashboard.env`，保存后按配置影响范围重启或等待下一轮任务读取。
+`DASHBOARD_GROK_API_MODE` 可设为 `auto`、`responses` 或 `chat`。默认 `auto` 会为 Grok 4.5 使用带 `web_search`/`x_search` 工具的 Responses API，其他模型保持 Chat Completions；兼容网关可显式选择对应模式。`X_WATCHLIST_REQUEST_TIMEOUT_SECONDS` 控制 X 单账号请求超时，默认 `45` 秒。
 `*_CONTEXT_LENGTH` 仅表示模型上下文窗口，默认 `128000`；`*_MAX_TOKENS` 仅用于请求体里的 `max_tokens`，默认 `4096`，可按场景覆盖。
 消息面预检默认最多并发检查 5 只候选股；如果上游出现限流或 403/429，可将 `DASHBOARD_NEWS_CONCURRENCY` 降为 `2` 或 `1`。