release: Buddy2api v2.1.1 reasoning effort compatibility

README.md

@@ -4,7 +4,7 @@
 
 > 把本机已经登录的消费级 AI 客户端，接成 OpenAI 兼容接口，给 Codex、OpenCode、Cherry Studio、NextChat 等用。默认打开 Work Buddy / CodeBuddy、QClaw、千问办公（QwenWork）、TraeWork 四个通道；管理页下拉选其中一个。一次请求只走一个通道。
 
-当前版本 **2.1.0**。这个项目只适合本机自用，不要公开部署，也不要把登录凭据、API Key、数据库文件发给别人。
+当前版本 **2.1.1**。这个项目只适合本机自用，不要公开部署，也不要把登录凭据、API Key、数据库文件发给别人。
 
 ## 这是什么？
 
@@ -155,6 +155,27 @@ python server.py
 
 接口：`/v1/chat/completions`、`/v1/responses`、`/v1/models`。没加前缀的 `auto` 走这把 Key 绑定的通道。Codex 用 Responses 接口；管理页选 Codex 类型的 Key 会按 Codex 特征 prompt 做清洗（其它客户端借用这把 Key、但没有 Codex 特征时不改写）。
 
+### 思考强度
+
+智能体可以在 Chat Completions 中发送顶层 `reasoning_effort`，在 Responses 中发送标准的 `reasoning: {"effort": "high"}`。网关也兼容 OpenCode、DSH、Cherry 和 Claude 风格的 `reasoning.effort`、`reasoningEffort`、`thinking.type`、`thinking.effort`、`output_config.effort`、`enable_thinking` 等写法。可用档位为 `none`、`minimal`、`low`、`medium`、`high`、`xhigh`、`max`、`ultra`；`off` 等同于 `none`。
+
+```json
+{
+  "model": "deepseek-v4-pro",
+  "messages": [{"role": "user", "content": "分析这个问题"}],
+  "reasoning_effort": "high"
+}
+```
+
+| 通道 | 实际能力 |
+|---|---|
+| WorkBuddy | DeepSeek V4 Pro/Flash 支持 `low` / `high` / `max`，标准档位会投影到这三档；未指定时默认 `high`，可用 `CB_GATEWAY_DEFAULT_REASONING_EFFORT=off` 关闭默认 |
+| QClaw | 统一转换成 `reasoning_effort` 后透传；具体档位是否生效由所选上游模型决定，不额外注入默认值 |
+| QwenWork | 协议只有 `is_reasoning` 开关；`none` 关闭，其它显式档位开启，无法区分多档强度 |
+| TraeWork | 当前会话协议没有可验证的思考控制字段，因此暂不支持调档 |
+
+Chat 流会保留 `reasoning_content`。Responses 流会转换成标准的 `response.reasoning_summary_*` 事件，仅有推理、没有最终正文的有效响应也会正常完成。
+
 OpenCode 示例（WorkBuddy Key）：
 
 ```json
@@ -204,6 +225,7 @@ QwenWork、QClaw、TraeWork 各用自己那把 Key，不要混用。
 | `CB_GATEWAY_PROVIDERS` | 启用哪些通道，逗号分隔。默认 `workbuddy,qclaw,qwenwork,traework`。只想留一家时再改 |
 | `CB_GATEWAY_AUTO_IMPORT` | 设 `1` 则启动时自动导入。默认 `0` |
 | `CB_GATEWAY_CHECKIN_GAP_MS` | 一键领取间隔，默认 `800` |
+| `CB_GATEWAY_DEFAULT_REASONING_EFFORT` | WorkBuddy DeepSeek V4 Pro/Flash 的默认思考强度，支持 `low` / `high` / `max`，默认 `high`；设为 `off` 可关闭默认值。Responses 的 `reasoning.effort` 或 Chat Completions 的 `reasoning_effort` 会覆盖它 |
 | `CB_AUTH_DIR` | WorkBuddy 登录目录 |
 | `CB_QCLAW_AUTH_DIR` | QClaw 登录目录 |
 | `CB_QWENWORK_AUTH_DIR` | QwenWork 登录目录 |

README_EN.md

@@ -4,7 +4,7 @@
 
 > Local consumer AI clients → one OpenAI-compatible API for Codex, OpenCode, Cherry Studio, NextChat, and similar agents. Work Buddy / CodeBuddy, QClaw, QwenWork, and TraeWork are on by default; pick one in the UI dropdown. Each request stays on one channel.
 
-Release **2.1.0**. Local use only. Do not expose this on the public internet, and do not share credentials, API keys, or the database.
+Release **2.1.1**. Local use only. Do not expose this on the public internet, and do not share credentials, API keys, or the database.
 
 ## What is this?
 
@@ -80,6 +80,19 @@ The database migrates on startup. Existing keys stay on `workbuddy`. Startup no
 
 Unprefixed `auto` follows the key’s channel. Use a separate key per channel.
 
+### Reasoning effort
+
+Agent clients can send top-level `reasoning_effort` to Chat Completions and the standard `reasoning: {"effort": "high"}` object to Responses. Compatibility forms used by OpenCode, DSH, Cherry, and Claude-style clients are also accepted: `reasoning.effort`, `reasoningEffort`, `thinking.type`, `thinking.effort`, `output_config.effort`, and `enable_thinking`. Accepted levels are `none`, `minimal`, `low`, `medium`, `high`, `xhigh`, `max`, and `ultra`; `off` is an alias for `none`.
+
+| Channel | Effective capability |
+|---|---|
+| WorkBuddy | DeepSeek V4 Pro/Flash supports `low` / `high` / `max`; standard levels are projected onto those tiers. The default is `high` and can be disabled with `CB_GATEWAY_DEFAULT_REASONING_EFFORT=off` |
+| QClaw | The control is normalized to `reasoning_effort` and forwarded. Whether a tier takes effect depends on the selected upstream model; no gateway default is injected |
+| QwenWork | The protocol exposes only an `is_reasoning` switch. `none` disables it and any other explicit tier enables it; distinct effort levels are unavailable |
+| TraeWork | The current session protocol has no verified reasoning control field, so effort selection is not supported |
+
+Chat streams preserve `reasoning_content`. Responses streams expose standard `response.reasoning_summary_*` events and accept valid reasoning-only completions.
+
 ```bash
 curl http://127.0.0.1:8787/v1/chat/completions \
   -H "Content-Type: application/json" \
@@ -91,6 +104,8 @@ curl http://127.0.0.1:8787/v1/chat/completions \
 
 `CB_GATEWAY_PROVIDERS` (default `workbuddy,qclaw,qwenwork,traework`), `CB_GATEWAY_AUTO_IMPORT` (default `0`), `CB_AUTH_DIR` / `CB_QCLAW_AUTH_DIR` / `CB_QWENWORK_AUTH_DIR` / `CB_TRAEWORK_AUTH_DIR`, `CB_GATEWAY_ADMIN_TOKEN`, `CB_GATEWAY_MASTER_KEY`.
 
+`CB_GATEWAY_DEFAULT_REASONING_EFFORT` controls the default reasoning effort for WorkBuddy DeepSeek V4 Pro/Flash. It accepts `low`, `high`, or `max`, defaults to `high`, and can be disabled with `off`. A Responses `reasoning.effort` or Chat Completions `reasoning_effort` value overrides the default.
+
 Keep `--host 127.0.0.1`. Do not share the database, auth folders, or key screenshots.
 
 ## License

docs/releases/v2.1.1.md

@@ -0,0 +1,31 @@
+# Buddy2api v2.1.1
+
+发布日期：2026-08-28
+
+本版本补齐智能体思考强度兼容，并修复 Responses API 在多通道路由和推理输出上的行为。
+
+## 思考强度
+
+- Chat Completions 支持顶层 `reasoning_effort`，Responses 支持标准的 `reasoning: {"effort": "high"}`。
+- 同时兼容 `reasoning.effort`、`reasoningEffort`、`thinking.type`、`thinking.effort`、`output_config.effort`、`enable_thinking` 等常见智能体写法。
+- 接受 `none`、`minimal`、`low`、`medium`、`high`、`xhigh`、`max`、`ultra`；`off` 等同于 `none`。无效或冲突的控制返回 `400 invalid_reasoning_control`。
+- WorkBuddy DeepSeek V4 Pro/Flash 映射到 `low` / `high` / `max`，未指定时默认 `high`；可用 `CB_GATEWAY_DEFAULT_REASONING_EFFORT=off` 关闭默认值。
+- QClaw 规范化为 `reasoning_effort` 后透传；QwenWork 只能映射为开关；TraeWork 当前不支持调档。
+
+## Responses 兼容
+
+- `/v1/responses` 现在遵循 API Key 绑定的通道，不再固定走 WorkBuddy。
+- Chat 的 `reasoning_content` 会转换为标准 `response.reasoning_summary_*` 事件。
+- 支持只有推理内容、没有最终正文的有效完成，并避免兼容上游将相同推理文本复制到正文时重复显示。
+- 补齐 Responses 的请求回显、usage 明细、错误结构和流式事件收尾。
+
+## 升级说明
+
+- 无数据库迁移。
+- Docker 用户需要重新构建镜像并重启服务，才能加载新代码。
+- 思考强度选择器是否显示仍由客户端自己的模型能力配置决定。例如 DSH 自定义模型需要在其本地配置中声明 `reasoningEfforts`；Buddy2api 不会通过 `/v1/models` 强制客户端显示控件。
+
+## 验证
+
+- 完整测试集：`238 passed`。
+- 已实测 Chat 默认档位、`none`、`max`，以及 Responses `high` 的真实上游响应。

providers/qclaw/chat.py

@@ -10,6 +10,7 @@
 
 import auth_manager
 import database as db
+from reasoning_controls import normalize_chat_reasoning
 from providers.qclaw.constants import AIZONE_BASE, ALIASES, CHANNEL_ID, RETRYABLE_STATUS
 from providers.qclaw.sign import aizone_headers
 
@@ -44,6 +45,8 @@ def fill_empty_content(payload: dict) -> dict:
     for key in ("reasoning_content", "reasoning"):
         text = _alt_text(out.get(key))
         if text:
+            if key == "reasoning":
+                out["reasoning_content"] = text
             out["content"] = text
             return out
     return out
@@ -105,7 +108,7 @@ def _log(api_key_info, account, model_name, stream, prompt_t, completion_t, tota
 
 
 def _build_body(payload: dict) -> tuple[dict, str]:
-    body = dict(payload)
+    body = normalize_chat_reasoning(payload)
     body["model"] = translate_model(str(body.get("model") or "default"))
     raw = json.dumps(body, ensure_ascii=False, separators=(",", ":"))
     return body, raw

providers/qwenwork/chat.py

@@ -11,6 +11,7 @@
 
 import auth_manager
 import database as db
+from reasoning_controls import resolve_reasoning_control
 from providers.qwenwork import cosy
 from providers.qwenwork.constants import (
     ALIASES,
@@ -118,6 +119,8 @@ def build_body(payload: dict) -> tuple[dict, str, str]:
     if system:
         messages = [{"role": "system", "content": system}, *messages]
     last_user = _last_user_text(messages)
+    reasoning_control = resolve_reasoning_control(payload)
+    is_reasoning = reasoning_control.enabled is True
     parameters = {}
     for key in ("temperature", "top_p", "max_tokens", "presence_penalty", "frequency_penalty"):
         if key in payload and payload[key] is not None:
@@ -136,7 +139,7 @@ def build_body(payload: dict) -> tuple[dict, str, str]:
             "features": [],
             "extra": {
                 "context": [],
-                "modelConfig": {"key": model, "is_reasoning": False},
+                "modelConfig": {"key": model, "is_reasoning": is_reasoning},
                 "originalContent": last_user,
             },
             "chatPrompt": "",
@@ -156,7 +159,7 @@ def build_body(payload: dict) -> tuple[dict, str, str]:
             "model": "",
             "format": "openai",
             "is_vl": model == "qwork-advanced",
-            "is_reasoning": False,
+            "is_reasoning": is_reasoning,
             "api_key": "",
             "url": "",
             "source": "system",

proxy.py

@@ -20,6 +20,11 @@
 
 import database as db
 import auth_manager
+from reasoning_controls import (
+    chat_reasoning_effort,
+    resolve_reasoning_control,
+    workbuddy_reasoning_effort,
+)
 
 BACKEND = "https://copilot.tencent.com"
 RETRYABLE_STATUS_CODES = {408, 409, 425, 429, 500, 502, 503, 504}
@@ -126,6 +131,7 @@ async def _retry_delay(attempt: int):
     "deepseek-v4-pro",
     "deepseek-v4-flash",
 })
+_DEFAULT_REASONING_EFFORT = "high"
 _VALID_REASONING_DEFAULTS = frozenset({"low", "high", "max"})
 _BACKEND_ROLE_ALIASES = {
     "developer": "system",
@@ -195,14 +201,18 @@ def resolve_model_alias(model: str) -> str:
 
 
 def _configured_reasoning_default(model: str) -> str | None:
-    """Return the opt-in reasoning default for supported DeepSeek V4 models."""
+    """Return the configured reasoning default for supported DeepSeek V4 models."""
     if model not in _REASONING_DEFAULT_MODEL_IDS:
         return None
-    value = os.environ.get("CB_GATEWAY_DEFAULT_REASONING_EFFORT", "").strip().lower()
+    value = os.environ.get(
+        "CB_GATEWAY_DEFAULT_REASONING_EFFORT",
+        _DEFAULT_REASONING_EFFORT,
+    ).strip().lower()
     return value if value in _VALID_REASONING_DEFAULTS else None
 
 
 def build_backend_body(payload: dict) -> dict:
+    reasoning_control = resolve_reasoning_control(payload)
     body = {k: payload[k] for k in PASSTHROUGH_BODY_KEYS if k in payload}
     messages = body.get("messages")
     if isinstance(messages, list):
@@ -215,14 +225,21 @@ def build_backend_body(payload: dict) -> dict:
             else message
             for message in messages
         ]
-    has_explicit_thinking = "thinking" in payload
     # Resolve model alias before forwarding
     raw_model = body.get("model", "auto")
     body["model"] = resolve_model_alias(raw_model)
-    if "reasoning_effort" not in body and not has_explicit_thinking:
+    body.pop("reasoning_effort", None)
+    if reasoning_control.mode == "default":
         default_reasoning = _configured_reasoning_default(body["model"])
         if default_reasoning:
             body["reasoning_effort"] = default_reasoning
+    else:
+        if body["model"] in _REASONING_DEFAULT_MODEL_IDS:
+            reasoning_effort = workbuddy_reasoning_effort(reasoning_control)
+        else:
+            reasoning_effort = chat_reasoning_effort(reasoning_control)
+        if reasoning_effort:
+            body["reasoning_effort"] = reasoning_effort
     body["stream"] = True
     if "stream_options" not in body:
         body["stream_options"] = {"include_usage": True}
@@ -283,6 +300,7 @@ def __init__(self, fallback_model: str, expected_choices: int = 1):
         self.finish_reasons: dict[int, str | None] = {}
         self.closed_choices: set[int] = set()
         self.content_choices: set[int] = set()
+        self.reasoning_choices: set[int] = set()
         self.tool_call_choices: set[int] = set()
         self.tool_calls: dict[tuple[int, int], dict] = {}
         self.malformed_data_event = False
@@ -423,6 +441,8 @@ def observe_event(self, data: bytes) -> dict | None:
             if content:
                 self.content_parts.append(content)
                 self.content_choices.add(index)
+            if delta.get("reasoning_content"):
+                self.reasoning_choices.add(index)
             tool_deltas = delta.get("tool_calls")
             if tool_deltas is None:
                 continue
@@ -505,9 +525,10 @@ def eof_error(self) -> str | None:
             if (
                 reason not in {"length", "content_filter"}
                 and choice_index not in self.content_choices
+                and choice_index not in self.reasoning_choices
                 and choice_index not in self.tool_call_choices
             ):
-                return "The upstream choice ended without content or a tool call."
+                return "The upstream choice ended without content, reasoning, or a tool call."
         return None
 
     def terminal_event(self, choice_indices: list[int]) -> bytes:
@@ -1141,14 +1162,15 @@ async def _collect_stream(
 
     if (
         not content_parts
+        and not reasoning_parts
         and not tool_calls
         and finish_reason not in {"length", "content_filter"}
     ):
         return (
             "error",
             (502, {
                 "error": {
-                    "message": "The upstream choice ended without content or a tool call.",
+                    "message": "The upstream choice ended without content, reasoning, or a tool call.",
                     "type": "upstream_error",
                 },
             }),