AI 网关切换 OpenAI 兼容面并移除 chat 端点,新增模型黑白名单

CHANGELOG.md

@@ -2,19 +2,28 @@
 
 格式遵循 [Keep a Changelog](https://keepachangelog.com/zh-CN/1.1.0/)，版本号遵循语义化版本。
 
-## [0.2.0] - 2026-07-10
+## [0.2.0] - 2026-07-12
 
 ### Added
 
-- AI 网关模型池自愈：同步与每日探测后自动逐个试调验证模型可用性（maxTokens=1），不可按需调用的模型（微调基座 / 区域未开放按需供给）自动标记并从模型列表、路由与探测候选中剔除；已标记模型每 20 小时复检，OCI 恢复供给后自动解除标记。`ai_model_caches` 新增 `unusable` / `unusable_reason` / `checked_at`（启动自动迁移），渠道详情接口保留标记行与原因便于排查
+- AI 网关模型黑名单：面板集中拉黑模型，全渠道剔除（模型列表 / 路由 / 探测候选不再出现，同步不入库），移除条目并重新同步后恢复；新表 `ai_model_blacklists`（启动自动迁移），管理接口 `/api/v1/ai-blacklist`
+- AI 密钥模型白名单：密钥可限定放行的模型列表（空 = 不限），白名单外调用 404，模型列表只返回交集
+- 推理力度（effort）直通：Responses `reasoning.effort` 与 Anthropic Messages `output_config.effort` 原样透传上游、不做档位校验；grok-4.3 五档（`none` 完全关闭思考）、gpt-oss `minimal`-`high`、multi-agent `xhigh`（控制并行 agent 数）实测生效
+- xAI Grok 服务端工具：`/ai/v1/responses` 的 `tools` 支持 `{"type":"web_search"}` / `{"type":"x_search"}`（官方 Agent Tools 格式，可与 `function` 混用，仅非流式），响应保留 `web_search_call` 输出项与引用
 
 ### Changed
 
+- **AI 网关对话链路整体切换到 OCI OpenAI 兼容面**（`/actions/v1/responses`），typed chat 面全链路移除：`/ai/v1/responses` 请求体原样直通（仅强制 `store:false`），流式为上游原生事件流，gpt-oss 推理增量直达客户端；`/ai/v1/messages` 转换为 Responses 请求直通，流式经事件桥转回标准 Anthropic 事件序列，`stop_sequences` / `top_k` 等无兼容面对应物的字段改为静默忽略，思考输出不转 `thinking` 块；渠道探测改用兼容面最小请求；流式建立后绑定渠道，中断不重试不计熔断
 - 后台任务「立即执行」改异步触发：接口立即返回 202，执行结果经任务日志轮询呈现；同一任务在途时重复触发返回 409「任务正在执行中」，与 cron 重叠触发静默跳过（此前同步阻塞数十秒且同一任务可并发重复执行）
 - 渠道探测候选跨厂商分散（上限 8，voice 等不可对话形态排除）：单一厂商在区域内全为微调基座时不再拖垮整个渠道的探测结论
-- 网关调用（对话 / 流式 / 向量化）遇「模型不可按需调用」类错误（微调基座 400 / 实体不存在 404）自动标记该模型并换渠道重试，且不计入渠道熔断；仅鉴权类错误才判定租户无配额
+- 网关调用遇「模型不可按需调用」类错误（微调基座 400 / 实体不存在 404）自动换渠道重试且不计入渠道熔断，错误信息带模型名便于加入黑名单；仅鉴权类错误才判定租户无配额
 - 同名模型多条目去重优先保留非微调基座条目，降低缓存到不可调用 OCID 的概率
 
+### Removed
+
+- **`/ai/v1/chat/completions` 端点移除**（404）：OpenAI 官方新能力已集中到 Responses 且本面板真实流量为零，请迁移 `/ai/v1/responses` 或 `/ai/v1/messages`
+- google / cohere 对话模型不再供给（OCI OpenAI 兼容面不承接，上游 400；cohere embed 向量模型不受影响），模型目录按 `xai.` / `meta.` / `openai.` 厂商前缀过滤
+
 ### Fixed
 
 - 渠道「同步模型成功但探测报错不可用」：法兰克福等区域的微调基座模型占满探测候选导致的误报（探测状态与真实可用性不符的根因）

DASH_VERSION

@@ -1 +1 @@
-v0.1.0
+v0.2.0

README.md

@@ -34,7 +34,7 @@
 - **资源管理**：实例（含创建/电源操作/更换公网 IP/IPv6/VNIC/引导卷）、VCN/子网/安全列表、块存储、限额与成本查询，多区域/多区间支持
 - **抢机任务**：cron 周期尝试创建实例直到成功，支持熔断与通知
 - **网页控制台**：实例串行控制台（xterm）与 VNC（noVNC），两跳 SSH 隧道
-- **AI 网关**：OpenAI / Anthropic 兼容端点转发 OCI GenAI（对话/Responses/Embeddings），号池渠道加权负载均衡、熔断探测、密钥管理与用量日志
+- **AI 网关**：OpenAI Responses / Anthropic Messages / Embeddings 兼容端点转发 OCI GenAI，号池渠道加权负载均衡、熔断探测、密钥管理与用量日志；对话统一走 OCI OpenAI 兼容面（xai / meta / openai 厂商模型，流式事件与推理增量原样保真），支持 xAI Grok 服务端工具 `web_search` / `x_search`（格式同 xAI 官方 Agent Tools，仅非流式）——兼容面属实测可用但未见 Oracle 文档承诺的能力，行为可能随上游调整
 - **日志回传**：OCI Audit 事件经 Connector Hub → Notifications HTTPS 订阅回传入库，一键创建链路；自定义告警规则（事件类型/来源 IP 白名单/资源/频率阈值）命中即推送
 - **租户治理**：IAM 用户/MFA/API Key 管理、密码策略、身份提供商（SAML）、通知收件人；多 Identity Domain 租户可按域切换管理
 - **安全**：JWT + bcrypt（凭据变更旧令牌立即失效，可一键撤销全部会话）、TOTP 两步验证、OIDC/GitHub 外部登录、登录锁定、IP 限速、真实 IP 头可配、请求体/超时防护、系统操作审计
@@ -60,7 +60,7 @@ mkdir -p ./data && sudo chown 65532:65532 ./data
 DATA_KEY=$(openssl rand -hex 32) JWT_SECRET=$(openssl rand -hex 32) ADMIN_PASSWORD=<初始密码> docker compose up -d
 ```
 
-默认只监听 `127.0.0.1:18888`;公网访问须置于 TLS 反向代理之后,见下文「反向代理」。
+默认只监听 `127.0.0.1:18888`；公网访问须置于 TLS 反向代理之后，见下文「反向代理」。
 
 ### 源码构建（单文件，含前端）
 
@@ -148,6 +148,87 @@ services:
 - 反代默认追加的 `X-Forwarded-For` 用于还原真实客户端 IP，系统日志留痕、登录锁定与 IP 限速都依赖它
 - 请求体上限建议与后端一致：`/api/*` 1MB、`/ai/*` 10MB（Caddy 用 `request_body` 按前缀分层）
 
+## AI 网关接口
+
+对外提供 OpenAI / Anthropic 兼容端点，转发 OCI GenAI On-Demand 推理。鉴权用面板创建的 AI 密钥，`Authorization: Bearer sk-...` 与 `x-api-key: sk-...` 双头均可；密钥可绑定渠道分组实现路由隔离，也可配置模型白名单（白名单外调用 404，模型列表只返回交集）。请求体上限 10MB。
+
+| 端点 | 协议 | 流式 |
+| --- | --- | --- |
+| `POST /ai/v1/responses` | OpenAI Responses（无状态子集） | SSE 支持（服务端工具除外） |
+| `POST /ai/v1/messages` | Anthropic Messages | SSE 支持 |
+| `POST /ai/v1/embeddings` | OpenAI Embeddings | — |
+| `GET /ai/v1/models` | OpenAI 模型列表（来自渠道模型目录） | — |
+
+对话端点统一转发 OCI OpenAI 兼容面（`/actions/v1/responses`），因此仅提供该面支持的 `xai.` / `meta.` / `openai.` 前缀模型；google / cohere 对话模型不在兼容面供给（上游 400），不再提供（cohere embed 模型不受影响）。兼容面属实测可用但无 Oracle 文档承诺的能力，行为可能随上游调整。**Chat Completions（`/ai/v1/chat/completions`）端点已移除**，请迁移到 Responses 或 Messages。
+
+以下各表对照网关行为：✅ 转发上游；⚠️ 接受但忽略（静默丢弃，不影响请求）；❌ 拒绝（400，不发上游）。
+
+### Responses 字段
+
+请求体除下列例外**原样直通上游**（未列字段一并转发，效果由上游决定）：
+
+| 字段 | 支持 | 说明 |
+| --- | --- | --- |
+| `model`、`input` | ✅ | `input` 接受 string 或 item 数组 |
+| `stream` | ✅ | SSE，上游原生事件流（含 gpt-oss 推理增量）；与服务端工具互斥 |
+| `reasoning.effort` | ✅ | 直达上游不限档位，见下「推理力度」 |
+| `store` | ⚠️ | 网关无状态，强制改写为 `false` |
+| `previous_response_id`、`conversation`、`background` | ❌ | 网关不保存历史，请求需自带完整上下文 |
+| `web_search` / `x_search` / `function` 之外的工具类型 | ❌ | 如 `file_search`、`code_interpreter`、`mcp` |
+
+**服务端工具**：`tools` 支持 `{"type":"web_search"}` 与 `{"type":"x_search"}`（xAI Grok 官方 Agent Tools 格式，可与 `function` 混用），响应保留 `web_search_call` 输出项与引用；**仅非流式**。响应 `usage.input_tokens_details.cached_tokens` 透传缓存命中量。
+
+### Anthropic Messages 字段
+
+网关把 Messages 请求转换为 Responses 请求送上游，响应（含流式事件序列）转回 Anthropic 形态：
+
+| 字段 | 支持 | 说明 |
+| --- | --- | --- |
+| `model`、`max_tokens`、`messages` | ✅ | `max_tokens` 必填；content 块支持 `text` / `image` / `tool_use` / `tool_result`，其余块类型 400 |
+| `system` | ✅ | string 或块数组，映射为 `instructions` |
+| `temperature`、`top_p` | ✅ | |
+| `stream` | ✅ | SSE，标准 Anthropic 事件序列（`message_start` → `content_block_*` → `message_delta` → `message_stop`） |
+| `tools`、`tool_choice` | ✅ | 映射为 Responses `function` 工具；`tool_choice` 支持 `auto` / `any` / `none` / `tool` |
+| `output_config.effort` | ✅ | 直达上游不限档位，见下「推理力度」 |
+| `stop_sequences`、`top_k`、`metadata`、`thinking`、`service_tier` | ⚠️ | Responses 面无对应物，静默忽略 |
+
+模型的思考输出（reasoning）不转为 `thinking` 块，静默丢弃；`usage.cache_read_input_tokens` 透传缓存命中量。
+
+### Embeddings 字段
+
+| 字段 | 支持 | 说明 |
+| --- | --- | --- |
+| `model`、`input` | ✅ | `input` 接受 string 或数组 |
+| `dimensions` | ✅ | |
+| `encoding_format`、`user` | ⚠️ | 输出恒为 float 数组 |
+
+### 推理力度（effort）
+
+两个对话协议各以原生字段控制推理深度，取值直达上游不做档位校验：
+
+| 协议 | 字段 |
+| --- | --- |
+| Responses | `reasoning.effort` |
+| Anthropic Messages | `output_config.effort`（转小写后透传） |
+
+是否支持、可用档位与实际效果由模型决定（2026-07 兼容面实测）：
+
+| 模型 | 支持档位 | 实测行为 |
+| --- | --- | --- |
+| `xai.grok-4.3` | `none` / `minimal` / `low` / `medium` / `high` | 全部生效：`none` 完全关闭推理，默认 `low`，`high` 推理量最大 |
+| `openai.gpt-oss-*` | `minimal` / `low` / `medium` / `high` | `none` 被上游 400（Harmony 格式不支持）；思考计入 completion tokens |
+| `xai.grok-3-mini(-fast)` | 名义 `low` / `high` | 各档均被接受，但 `none` 并不关闭推理（仍产生思考） |
+| `meta.llama-*` | 接受任意档 | 非推理模型，参数无实际效果 |
+| `xai.grok-4.20-multi-agent*` | `low` / `medium` / `high` / `xhigh` | effort 控制并行 agent 数（4 / 16）而非推理深度 |
+| 其余 grok（`grok-3`、`grok-4`、`grok-4-fast-*`、`grok-4-1-fast-*`、`grok-code-fast-1`、`grok-4.20-*` 单/双型号全系） | 不支持 | 携带即被上游 400（`does not support parameter reasoningEffort`） |
+
+不携带该字段时网关不下发，行为由模型默认档位决定。
+
+### 通用行为
+
+- 未知模型 404；无可用渠道 503；单次请求最多尝试 3 个渠道（429/5xx/网络错误自动换渠道并计入熔断，模型级 400/404 换渠道不计熔断，其余 4xx 原样透传）；流式建立后绑定渠道，中断不重试
+- 调用计量（渠道、token 用量、缓存命中、时延、重试数）写入面板「调用日志」；密钥可按需开启限时内容抓取
+
 ## 环境变量
 
 | 变量 | 必填 | 默认值 | 说明 |

docs/docs.go

@@ -15,34 +15,6 @@ const docTemplate = `{
     "host": "{{.Host}}",
     "basePath": "{{.BasePath}}",
     "paths": {
-        "/ai/v1/chat/completions": {
-            "post": {
-                "tags": [
-                    "AI 网关"
-                ],
-                "summary": "OpenAI 兼容对话补全",
-                "parameters": [
-                    {
-                        "description": "OpenAI chat/completions 请求体(支持 stream)",
-                        "name": "body",
-                        "in": "body",
-                        "required": true,
-                        "schema": {
-                            "type": "object"
-                        }
-                    }
-                ],
-                "responses": {
-                    "200": {
-                        "description": "OpenAI 兼容响应(流式为 SSE)",
-                        "schema": {
-                            "type": "object",
-                            "additionalProperties": true
-                        }
-                    }
-                }
-            }
-        },
         "/ai/v1/embeddings": {
             "post": {
                 "tags": [
@@ -79,7 +51,7 @@ const docTemplate = `{
                 "summary": "Anthropic Messages 兼容端点",
                 "parameters": [
                     {
-                        "description": "Anthropic messages 请求体(支持 stream)",
+                        "description": "Anthropic messages 请求体(支持 stream;经 OCI OpenAI 兼容面直通)",
                         "name": "body",
                         "in": "body",
                         "required": true,
@@ -124,7 +96,7 @@ const docTemplate = `{
                 "summary": "OpenAI Responses 兼容端点",
                 "parameters": [
                     {
-                        "description": "OpenAI responses 请求体(支持 stream)",
+                        "description": "OpenAI responses 请求体(支持 stream;web_search/x_search 服务端工具仅非流式)",
                         "name": "body",
                         "in": "body",
                         "required": true,
@@ -166,6 +138,86 @@ const docTemplate = `{
                 }
             }
         },
+        "/api/v1/ai-blacklist": {
+            "get": {
+                "security": [
+                    {
+                        "BearerAuth": []
+                    }
+                ],
+                "tags": [
+                    "AI 管理"
+                ],
+                "summary": "模型黑名单列表",
+                "responses": {
+                    "200": {
+                        "description": "OK",
+                        "schema": {
+                            "type": "object",
+                            "additionalProperties": true
+                        }
+                    }
+                }
+            },
+            "post": {
+                "security": [
+                    {
+                        "BearerAuth": []
+                    }
+                ],
+                "tags": [
+                    "AI 管理"
+                ],
+                "summary": "添加模型黑名单",
+                "parameters": [
+                    {
+                        "description": "请求体:{name: 模型名}",
+                        "name": "body",
+                        "in": "body",
+                        "required": true,
+                        "schema": {
+                            "type": "object"
+                        }
+                    }
+                ],
+                "responses": {
+                    "201": {
+                        "description": "Created",
+                        "schema": {
+                            "type": "object",
+                            "additionalProperties": true
+                        }
+                    }
+                }
+            }
+        },
+        "/api/v1/ai-blacklist/{id}": {
+            "delete": {
+                "security": [
+                    {
+                        "BearerAuth": []
+                    }
+                ],
+                "tags": [
+                    "AI 管理"
+                ],
+                "summary": "移除模型黑名单(重新同步后模型恢复入池)",
+                "parameters": [
+                    {
+                        "type": "integer",
+                        "description": "黑名单条目 ID",
+                        "name": "id",
+                        "in": "path",
+                        "required": true
+                    }
+                ],
+                "responses": {
+                    "204": {
+                        "description": "无内容"
+                    }
+                }
+            }
+        },
         "/api/v1/ai-channels": {
             "get": {
                 "security": [

docs/swagger.json

@@ -8,34 +8,6 @@
     },
     "basePath": "/",
     "paths": {
-        "/ai/v1/chat/completions": {
-            "post": {
-                "tags": [
-                    "AI 网关"
-                ],
-                "summary": "OpenAI 兼容对话补全",
-                "parameters": [
-                    {
-                        "description": "OpenAI chat/completions 请求体(支持 stream)",
-                        "name": "body",
-                        "in": "body",
-                        "required": true,
-                        "schema": {
-                            "type": "object"
-                        }
-                    }
-                ],
-                "responses": {
-                    "200": {
-                        "description": "OpenAI 兼容响应(流式为 SSE)",
-                        "schema": {
-                            "type": "object",
-                            "additionalProperties": true
-                        }
-                    }
-                }
-            }
-        },
         "/ai/v1/embeddings": {
             "post": {
                 "tags": [
@@ -72,7 +44,7 @@
                 "summary": "Anthropic Messages 兼容端点",
                 "parameters": [
                     {
-                        "description": "Anthropic messages 请求体(支持 stream)",
+                        "description": "Anthropic messages 请求体(支持 stream;经 OCI OpenAI 兼容面直通)",
                         "name": "body",
                         "in": "body",
                         "required": true,
@@ -117,7 +89,7 @@
                 "summary": "OpenAI Responses 兼容端点",
                 "parameters": [
                     {
-                        "description": "OpenAI responses 请求体(支持 stream)",
+                        "description": "OpenAI responses 请求体(支持 stream;web_search/x_search 服务端工具仅非流式)",
                         "name": "body",
                         "in": "body",
                         "required": true,
@@ -159,6 +131,86 @@
                 }
             }
         },
+        "/api/v1/ai-blacklist": {
+            "get": {
+                "security": [
+                    {
+                        "BearerAuth": []
+                    }
+                ],
+                "tags": [
+                    "AI 管理"
+                ],
+                "summary": "模型黑名单列表",
+                "responses": {
+                    "200": {
+                        "description": "OK",
+                        "schema": {
+                            "type": "object",
+                            "additionalProperties": true
+                        }
+                    }
+                }
+            },
+            "post": {
+                "security": [
+                    {
+                        "BearerAuth": []
+                    }
+                ],
+                "tags": [
+                    "AI 管理"
+                ],
+                "summary": "添加模型黑名单",
+                "parameters": [
+                    {
+                        "description": "请求体:{name: 模型名}",
+                        "name": "body",
+                        "in": "body",
+                        "required": true,
+                        "schema": {
+                            "type": "object"
+                        }
+                    }
+                ],
+                "responses": {
+                    "201": {
+                        "description": "Created",
+                        "schema": {
+                            "type": "object",
+                            "additionalProperties": true
+                        }
+                    }
+                }
+            }
+        },
+        "/api/v1/ai-blacklist/{id}": {
+            "delete": {
+                "security": [
+                    {
+                        "BearerAuth": []
+                    }
+                ],
+                "tags": [
+                    "AI 管理"
+                ],
+                "summary": "移除模型黑名单(重新同步后模型恢复入池)",
+                "parameters": [
+                    {
+                        "type": "integer",
+                        "description": "黑名单条目 ID",
+                        "name": "id",
+                        "in": "path",
+                        "required": true
+                    }
+                ],
+                "responses": {
+                    "204": {
+                        "description": "无内容"
+                    }
+                }
+            }
+        },
         "/api/v1/ai-channels": {
             "get": {
                 "security": [

docs/swagger.yaml

@@ -599,24 +599,6 @@ info:
   title: OCI Portal API
   version: 0.0.1
 paths:
-  /ai/v1/chat/completions:
-    post:
-      parameters:
-      - description: OpenAI chat/completions 请求体(支持 stream)
-        in: body
-        name: body
-        required: true
-        schema:
-          type: object
-      responses:
-        "200":
-          description: OpenAI 兼容响应(流式为 SSE)
-          schema:
-            additionalProperties: true
-            type: object
-      summary: OpenAI 兼容对话补全
-      tags:
-      - AI 网关
   /ai/v1/embeddings:
     post:
       parameters:
@@ -638,7 +620,7 @@ paths:
   /ai/v1/messages:
     post:
       parameters:
-      - description: Anthropic messages 请求体(支持 stream)
+      - description: Anthropic messages 请求体(支持 stream;经 OCI OpenAI 兼容面直通)
         in: body
         name: body
         required: true
@@ -667,7 +649,7 @@ paths:
   /ai/v1/responses:
     post:
       parameters:
-      - description: OpenAI responses 请求体(支持 stream)
+      - description: OpenAI responses 请求体(支持 stream;web_search/x_search 服务端工具仅非流式)
         in: body
         name: body
         required: true
@@ -695,6 +677,54 @@ paths:
       summary: 返回构建与运行环境信息,「设置 · 关于」页展示
       tags:
       - 设置
+  /api/v1/ai-blacklist:
+    get:
+      responses:
+        "200":
+          description: OK
+          schema:
+            additionalProperties: true
+            type: object
+      security:
+      - BearerAuth: []
+      summary: 模型黑名单列表
+      tags:
+      - AI 管理
+    post:
+      parameters:
+      - description: '请求体:{name: 模型名}'
+        in: body
+        name: body
+        required: true
+        schema:
+          type: object
+      responses:
+        "201":
+          description: Created
+          schema:
+            additionalProperties: true
+            type: object
+      security:
+      - BearerAuth: []
+      summary: 添加模型黑名单
+      tags:
+      - AI 管理
+  /api/v1/ai-blacklist/{id}:
+    delete:
+      parameters:
+      - description: 黑名单条目 ID
+        in: path
+        name: id
+        required: true
+        type: integer
+      responses:
+        "204":
+          description: 无内容
+      security:
+      - BearerAuth: []
+      summary: 移除模型黑名单(重新同步后模型恢复入池)
+      tags:
+      - AI 管理
   /api/v1/ai-channels:
     get:
       responses:

internal/aiwire/anthropic.go

@@ -20,6 +20,13 @@ type MessagesRequest struct {
 	ToolChoice    json.RawMessage `json:"tool_choice,omitempty"`
 	Metadata      json.RawMessage `json:"metadata,omitempty"`
 	Thinking      json.RawMessage `json:"thinking,omitempty"`
+	// OutputConfig 是 Anthropic 官方输出控制(effort 等);网关映射 effort 到 OCI reasoningEffort。
+	OutputConfig *AnthOutputConfig `json:"output_config,omitempty"`
+}
+
+// AnthOutputConfig 是 Anthropic Messages 顶层 output_config 对象。
+type AnthOutputConfig struct {
+	Effort string `json:"effort,omitempty"`
 }
 
 // SystemText 把顶层 system(string 或块数组)拍平为文本。

internal/aiwire/openai.go

@@ -1,35 +1,11 @@
-// Package aiwire 定义 AI 网关的线格式类型:OpenAI Chat Completions(兼作网关内部
+// Package aiwire 定义 AI 网关的线格式类型:OpenAI Responses / Anthropic Messages /
-// 中间表示 IR,因 OCI GENERIC 格式即其镜像)与 Anthropic Messages。零内部依赖,
+// Embeddings 与通用记账、错误、模型列表结构。上游统一为 OCI OpenAI 兼容面。
-// oci 层与 service 层共同引用。
 package aiwire
 
 import (
 	"encoding/json"
-	"strings"
 )
 
-// ChatRequest 是 OpenAI /v1/chat/completions 请求体;TopK 为 OCI/Anthropic 扩展字段。
-type ChatRequest struct {
-	Model               string          `json:"model"`
-	Messages            []ChatMessage   `json:"messages"`
-	MaxTokens           *int            `json:"max_tokens,omitempty"`
-	MaxCompletionTokens *int            `json:"max_completion_tokens,omitempty"`
-	Temperature         *float64        `json:"temperature,omitempty"`
-	TopP                *float64        `json:"top_p,omitempty"`
-	TopK                *int            `json:"top_k,omitempty"`
-	FrequencyPenalty    *float64        `json:"frequency_penalty,omitempty"`
-	PresencePenalty     *float64        `json:"presence_penalty,omitempty"`
-	Stop                StringList      `json:"stop,omitempty"`
-	Seed                *int            `json:"seed,omitempty"`
-	N                   *int            `json:"n,omitempty"`
-	Stream              bool            `json:"stream,omitempty"`
-	StreamOptions       *StreamOptions  `json:"stream_options,omitempty"`
-	Tools               []Tool          `json:"tools,omitempty"`
-	ToolChoice          json.RawMessage `json:"tool_choice,omitempty"`
-	ResponseFormat      *ResponseFormat `json:"response_format,omitempty"`
-	ReasoningEffort     string          `json:"reasoning_effort,omitempty"`
-}
-
 // StringList 兼容 OpenAI stop 的 string 与 []string 两种线格式。
 type StringList []string
 
@@ -50,151 +26,6 @@ func (s *StringList) UnmarshalJSON(b []byte) error {
 	return nil
 }
 
-// StreamOptions 控制流式末尾是否附带 usage 事件。
-type StreamOptions struct {
-	IncludeUsage bool `json:"include_usage"`
-}
-
-// ChatMessage 是请求与非流式响应共用的消息结构。
-type ChatMessage struct {
-	Role       string     `json:"role"`
-	Content    Content    `json:"content"`
-	Name       string     `json:"name,omitempty"`
-	ToolCalls  []ToolCall `json:"tool_calls,omitempty"`
-	ToolCallID string     `json:"tool_call_id,omitempty"`
-}
-
-// Content 兼容 string 与多模态块数组两种线格式;一期仅透传文本。
-type Content struct {
-	// Text 在原始值为字符串时填充
-	Text string
-	// Parts 在原始值为数组时填充
-	Parts []ContentPart
-	// isArray 记录原始形态,序列化时保形
-	isArray bool
-}
-
-// ContentPart 是块数组中的一段;text 与 image_url 被网关处理,其余类型拒绝。
-type ContentPart struct {
-	Type     string    `json:"type"`
-	Text     string    `json:"text,omitempty"`
-	ImageURL *ImageURL `json:"image_url,omitempty"`
-}
-
-// ImageURL 是 image_url 块负载;url 为 data URI(base64)或公网地址,与 OCI ImageUrl 同构。
-type ImageURL struct {
-	URL    string `json:"url"`
-	Detail string `json:"detail,omitempty"`
-}
-
-func (c *Content) UnmarshalJSON(b []byte) error {
-	trimmed := strings.TrimSpace(string(b))
-	if trimmed == "null" {
-		return nil
-	}
-	if len(trimmed) > 0 && trimmed[0] == '[' {
-		c.isArray = true
-		return json.Unmarshal(b, &c.Parts)
-	}
-	return json.Unmarshal(b, &c.Text)
-}
-
-func (c Content) MarshalJSON() ([]byte, error) {
-	if c.isArray {
-		return json.Marshal(c.Parts)
-	}
-	return json.Marshal(c.Text)
-}
-
-// NewTextContent 构造字符串形态的内容。
-func NewTextContent(text string) Content {
-	return Content{Text: text}
-}
-
-// NewPartsContent 构造块数组形态的内容(多模态消息用)。
-func NewPartsContent(parts []ContentPart) Content {
-	return Content{Parts: parts, isArray: true}
-}
-
-// JoinText 返回全部文本内容(数组形态拼接 text 块)。
-func (c Content) JoinText() string {
-	if !c.isArray {
-		return c.Text
-	}
-	var sb strings.Builder
-	for _, p := range c.Parts {
-		if p.Type == "" || p.Type == "text" {
-			sb.WriteString(p.Text)
-		}
-	}
-	return sb.String()
-}
-
-// HasUnsupported 报告是否含网关无法承接的内容块(text / image_url 之外的类型)。
-func (c Content) HasUnsupported() bool {
-	for _, p := range c.Parts {
-		switch p.Type {
-		case "", "text":
-		case "image_url":
-			if p.ImageURL == nil || p.ImageURL.URL == "" {
-				return true
-			}
-		default:
-			return true
-		}
-	}
-	return false
-}
-
-// Tool 是 function 工具定义。
-type Tool struct {
-	Type     string      `json:"type"`
-	Function FunctionDef `json:"function"`
-}
-
-// FunctionDef 描述工具的名称与 JSON Schema 参数。
-type FunctionDef struct {
-	Name        string          `json:"name"`
-	Description string          `json:"description,omitempty"`
-	Parameters  json.RawMessage `json:"parameters,omitempty"`
-}
-
-// ToolCall 是助手消息中的工具调用。
-type ToolCall struct {
-	ID       string       `json:"id"`
-	Type     string       `json:"type"`
-	Function FunctionCall `json:"function"`
-}
-
-// FunctionCall 携带调用名与 JSON 编码的实参。
-type FunctionCall struct {
-	Name      string `json:"name"`
-	Arguments string `json:"arguments"`
-}
-
-// ResponseFormat 对应 response_format(json_object / json_schema)。
-type ResponseFormat struct {
-	Type       string          `json:"type"`
-	JSONSchema json.RawMessage `json:"json_schema,omitempty"`
-}
-
-// ChatResponse 是非流式响应体。
-type ChatResponse struct {
-	ID      string   `json:"id"`
-	Object  string   `json:"object"`
-	Created int64    `json:"created"`
-	Model   string   `json:"model"`
-	Choices []Choice `json:"choices"`
-	Usage   *Usage   `json:"usage,omitempty"`
-}
-
-// Choice 是一条候选回复。
-type Choice struct {
-	Index        int         `json:"index"`
-	Message      ChatMessage `json:"message"`
-	FinishReason string      `json:"finish_reason"`
-}
-
 // Usage 是 token 用量。
 type Usage struct {
 	PromptTokens     int `json:"prompt_tokens"`
@@ -217,44 +48,6 @@ func (u *Usage) CachedTokens() int {
 	return u.PromptTokensDetails.CachedTokens
 }
 
-// ChatChunk 是流式增量事件体(chat.completion.chunk)。
-type ChatChunk struct {
-	ID      string        `json:"id"`
-	Object  string        `json:"object"`
-	Created int64         `json:"created"`
-	Model   string        `json:"model"`
-	Choices []ChunkChoice `json:"choices"`
-	Usage   *Usage        `json:"usage,omitempty"`
-}
-
-// ChunkChoice 是流式事件中的增量候选。
-type ChunkChoice struct {
-	Index        int     `json:"index"`
-	Delta        Delta   `json:"delta"`
-	FinishReason *string `json:"finish_reason"`
-}
-
-// Delta 是流式增量内容;Content 为纯文本片段。
-type Delta struct {
-	Role      string          `json:"role,omitempty"`
-	Content   string          `json:"content,omitempty"`
-	ToolCalls []ToolCallDelta `json:"tool_calls,omitempty"`
-}
-
-// ToolCallDelta 是工具调用的流式片段;Index 标识聚合目标。
-type ToolCallDelta struct {
-	Index    int               `json:"index"`
-	ID       string            `json:"id,omitempty"`
-	Type     string            `json:"type,omitempty"`
-	Function FunctionCallDelta `json:"function"`
-}
-
-// FunctionCallDelta 是函数名 / 实参的增量片段。
-type FunctionCallDelta struct {
-	Name      string `json:"name,omitempty"`
-	Arguments string `json:"arguments,omitempty"`
-}
-
 // Model 是 /v1/models 列表项。
 type Model struct {
 	ID      string `json:"id"`

internal/aiwire/responses.go

@@ -156,39 +156,6 @@ type RespReasoning struct {
 	Effort string `json:"effort,omitempty"`
 }
 
-// Response 是 /responses 响应对象。
-type Response struct {
-	ID                string          `json:"id"`
-	Object            string          `json:"object"`
-	CreatedAt         int64           `json:"created_at"`
-	Status            string          `json:"status"`
-	Model             string          `json:"model"`
-	Output            []RespOutItem   `json:"output"`
-	Usage             *RespUsage      `json:"usage,omitempty"`
-	Error             *RespError      `json:"error"`
-	IncompleteDetails *RespIncomplete `json:"incomplete_details"`
-	Store             bool            `json:"store"`
-}
-
-// RespOutItem 是 output 数组项:message 或 function_call。
-type RespOutItem struct {
-	Type      string        `json:"type"`
-	ID        string        `json:"id"`
-	Status    string        `json:"status,omitempty"`
-	Role      string        `json:"role,omitempty"`
-	Content   []RespOutPart `json:"content,omitempty"`
-	CallID    string        `json:"call_id,omitempty"`
-	Name      string        `json:"name,omitempty"`
-	Arguments string        `json:"arguments,omitempty"`
-}
-
-// RespOutPart 是 message item 的内容部件(output_text)。
-type RespOutPart struct {
-	Type        string `json:"type"`
-	Text        string `json:"text"`
-	Annotations []any  `json:"annotations"`
-}
-
 // RespUsage 是 Responses 口径的用量(input/output 命名)。
 type RespUsage struct {
 	InputTokens         int            `json:"input_tokens"`

internal/api/aiadmin.go

@@ -43,12 +43,13 @@ func (h *aiAdminHandler) createKey(c *gin.Context) {
 		Name   string   `json:"name" binding:"required"`
 		Value  string   `json:"value"`
 		Group  string   `json:"group"`
+		Models []string `json:"models"`
 	}
 	if err := c.ShouldBindJSON(&req); err != nil {
 		c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
 		return
 	}
-	plaintext, key, err := h.gw.CreateKey(c.Request.Context(), req.Name, req.Value, req.Group)
+	plaintext, key, err := h.gw.CreateKey(c.Request.Context(), req.Name, req.Value, req.Group, req.Models)
 	if err != nil {
 		c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
 		return
@@ -72,12 +73,13 @@ func (h *aiAdminHandler) updateKey(c *gin.Context) {
 		Name    string    `json:"name"`
 		Enabled *bool     `json:"enabled"`
 		Group   *string   `json:"group"`
+		Models  *[]string `json:"models"`
 	}
 	if err := c.ShouldBindJSON(&req); err != nil {
 		c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
 		return
 	}
-	if err := h.gw.UpdateKey(c.Request.Context(), id, req.Name, req.Enabled, req.Group); err != nil {
+	if err := h.gw.UpdateKey(c.Request.Context(), id, req.Name, req.Enabled, req.Group, req.Models); err != nil {
 		respondError(c, err)
 		return
 	}
@@ -285,6 +287,64 @@ func (h *aiAdminHandler) gatewayModels(c *gin.Context) {
 	c.JSON(http.StatusOK, gin.H{"items": list.Data})
 }
 
+// ---- 模型黑名单 ----
+
+// @Summary  模型黑名单列表
+// @Tags     AI 管理
+// @Success  200 {object} map[string]any
+// @Security BearerAuth
+// @Router   /api/v1/ai-blacklist [get]
+func (h *aiAdminHandler) listBlacklist(c *gin.Context) {
+	items, err := h.gw.Blacklist(c.Request.Context())
+	if err != nil {
+		respondError(c, err)
+		return
+	}
+	c.JSON(http.StatusOK, gin.H{"items": items})
+}
+
+// addBlacklist 拉黑模型:删除全部渠道缓存中的同名条目,后续同步 / 探测均过滤。
+//
+// @Summary  添加模型黑名单
+// @Tags     AI 管理
+// @Param    body body object true "请求体:{name: 模型名}"
+// @Success  201 {object} map[string]any
+// @Security BearerAuth
+// @Router   /api/v1/ai-blacklist [post]
+func (h *aiAdminHandler) addBlacklist(c *gin.Context) {
+	var req struct {
+		Name string `json:"name" binding:"required"`
+	}
+	if err := c.ShouldBindJSON(&req); err != nil {
+		c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+		return
+	}
+	item, err := h.gw.AddBlacklist(c.Request.Context(), req.Name)
+	if err != nil {
+		c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+		return
+	}
+	c.JSON(http.StatusCreated, gin.H{"item": item})
+}
+
+// @Summary  移除模型黑名单(重新同步后模型恢复入池)
+// @Tags     AI 管理
+// @Param    id path int true "黑名单条目 ID"
+// @Success  204 "无内容"
+// @Security BearerAuth
+// @Router   /api/v1/ai-blacklist/{id} [delete]
+func (h *aiAdminHandler) removeBlacklist(c *gin.Context) {
+	id, ok := aiPathID(c)
+	if !ok {
+		return
+	}
+	if err := h.gw.RemoveBlacklist(c.Request.Context(), id); err != nil {
+		c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+		return
+	}
+	c.Status(http.StatusNoContent)
+}
+
 // @Summary  AI 调用日志列表
 // @Tags     AI 管理
 // @Success  200 {object} map[string]any

internal/api/aigateway.go

@@ -1,13 +1,15 @@
 package api
 
 import (
+	"bufio"
+	"bytes"
 	"crypto/rand"
 	"encoding/hex"
 	"encoding/json"
 	"errors"
-	"fmt"
 	"io"
 	"net/http"
+	"slices"
 	"strings"
 	"time"
 
@@ -86,6 +88,25 @@ func keyGroup(c *gin.Context) string {
 	return ""
 }
 
+// keyModels 取当前请求密钥的模型白名单(空 = 不限模型)。
+func keyModels(c *gin.Context) []string {
+	if key, ok := c.Get(aiKeyCtx); ok {
+		return key.(*model.AiKey).Models
+	}
+	return nil
+}
+
+// checkKeyModel 校验请求模型是否在密钥白名单内;拒绝时按端点协议返回
+// 404 model_not_found(与未知模型同口径,不泄露密钥配置细节)并返回 false。
+func checkKeyModel(c *gin.Context, modelName string) bool {
+	allowed := keyModels(c)
+	if len(allowed) == 0 || slices.Contains(allowed, modelName) {
+		return true
+	}
+	aiError(c, http.StatusNotFound, "model_not_found", "模型不存在或无权访问: "+modelName)
+	return false
+}
+
 // logEntry 组装调用日志骨架;usage / status 由调用方补齐。
 func (h *aiGatewayHandler) logEntry(c *gin.Context, endpoint, modelName string, stream bool, meta service.ChatMeta, start time.Time) model.AiCallLog {
 	entry := model.AiCallLog{
@@ -102,21 +123,6 @@ func (h *aiGatewayHandler) logEntry(c *gin.Context, endpoint, modelName string,
 }
 
 // validateIR 做协议无关的入参检查(模型名、消息、不支持的内容块)。
-func validateIR(ir aiwire.ChatRequest) error {
-	if strings.TrimSpace(ir.Model) == "" {
-		return fmt.Errorf("model 不能为空")
-	}
-	if len(ir.Messages) == 0 {
-		return fmt.Errorf("messages 不能为空")
-	}
-	for _, m := range ir.Messages {
-		if m.Content.HasUnsupported() {
-			return service.ErrAiUnsupportedBlock
-		}
-	}
-	return nil
-}
-
 // maybeLogContent 在调用密钥显式开启内容日志且未过期时写正文(红线例外);
 // respBody 为 nil 时只记请求(流式响应与向量结果不记录);callLogID 关联同次调用日志。
 func (h *aiGatewayHandler) maybeLogContent(c *gin.Context, callLogID uint, endpoint, modelName string, stream bool, reqBody, respBody any) {
@@ -147,47 +153,6 @@ func (h *aiGatewayHandler) logFailure(c *gin.Context, entry model.AiCallLog, req
 	h.maybeLogContent(c, callID, entry.Endpoint, entry.Model, entry.Stream, reqBody, nil)
 }
 
-// chatCompletions 是 OpenAI /ai/v1/chat/completions 端点。
-//
-// @Summary  OpenAI 兼容对话补全
-// @Tags     AI 网关
-// @Param    body body object true "OpenAI chat/completions 请求体(支持 stream)"
-// @Success  200 {object} map[string]any "OpenAI 兼容响应(流式为 SSE)"
-// @Router   /ai/v1/chat/completions [post]
-func (h *aiGatewayHandler) chatCompletions(c *gin.Context) {
-	var ir aiwire.ChatRequest
-	if err := c.ShouldBindJSON(&ir); err != nil {
-		aiError(c, http.StatusBadRequest, "invalid_request_error", err.Error())
-		return
-	}
-	if err := validateIR(ir); err != nil {
-		aiError(c, http.StatusBadRequest, "invalid_request_error", err.Error())
-		return
-	}
-	if ir.Stream {
-		h.streamOpenAI(c, ir)
-		return
-	}
-	start := time.Now()
-	resp, meta, err := h.gw.Chat(c.Request.Context(), ir, keyGroup(c))
-	entry := h.logEntry(c, "openai", ir.Model, false, meta, start)
-	if err != nil {
-		upstreamError(c, err)
-		entry.ErrMsg = err.Error()
-		h.logFailure(c, entry, ir)
-		return
-	}
-	resp.ID = aiRandID("chatcmpl-")
-	if resp.Created == 0 {
-		resp.Created = time.Now().Unix()
-	}
-	entry.Status = http.StatusOK
-	fillUsage(&entry, resp.Usage)
-	callID := h.gw.LogCall(entry)
-	h.maybeLogContent(c, callID, "openai", ir.Model, false, ir, resp)
-	c.JSON(http.StatusOK, resp)
-}
-
 func fillUsage(entry *model.AiCallLog, u *aiwire.Usage) {
 	if u == nil {
 		return
@@ -217,6 +182,9 @@ func (h *aiGatewayHandler) embeddings(c *gin.Context) {
 		aiError(c, http.StatusBadRequest, "invalid_request_error", "仅支持 encoding_format=float")
 		return
 	}
+	if !checkKeyModel(c, req.Model) {
+		return
+	}
 	start := time.Now()
 	resp, meta, err := h.gw.Embeddings(c.Request.Context(), req, keyGroup(c))
 	entry := h.logEntry(c, "embeddings", req.Model, false, meta, start)
@@ -235,44 +203,6 @@ func (h *aiGatewayHandler) embeddings(c *gin.Context) {
 	c.JSON(http.StatusOK, resp)
 }
 
-// streamOpenAI 以 OpenAI SSE 直通流式响应,末尾发 [DONE]。
-func (h *aiGatewayHandler) streamOpenAI(c *gin.Context, ir aiwire.ChatRequest) {
-	start := time.Now()
-	stream, meta, err := h.gw.OpenStream(c.Request.Context(), ir, keyGroup(c))
-	entry := h.logEntry(c, "openai", ir.Model, true, meta, start)
-	if err != nil {
-		upstreamError(c, err)
-		entry.ErrMsg = err.Error()
-		h.logFailure(c, entry, ir)
-		return
-	}
-	defer stream.Close()
-	sseHeaders(c)
-	id, created := aiRandID("chatcmpl-"), time.Now().Unix()
-	var usage *aiwire.Usage
-	for {
-		chunk, err := stream.Next()
-		if err != nil {
-			if !errors.Is(err, io.EOF) {
-				entry.ErrMsg = err.Error()
-			}
-			break
-		}
-		chunk.ID, chunk.Created = id, created
-		if chunk.Usage != nil {
-			usage = chunk.Usage
-		}
-		writeSSEData(c, chunk)
-	}
-	c.Writer.WriteString("data: [DONE]\n\n")
-	c.Writer.Flush()
-	entry.Status = http.StatusOK
-	entry.LatencyMs = time.Since(start).Milliseconds()
-	fillUsage(&entry, usage)
-	callID := h.gw.LogCall(entry)
-	h.maybeLogContent(c, callID, "openai", ir.Model, true, ir, nil)
-}
-
 func sseHeaders(c *gin.Context) {
 	c.Header("Content-Type", "text/event-stream")
 	c.Header("Cache-Control", "no-cache")
@@ -280,22 +210,11 @@ func sseHeaders(c *gin.Context) {
 	c.Writer.Flush()
 }
 
-func writeSSEData(c *gin.Context, v any) {
-	b, err := json.Marshal(v)
-	if err != nil {
-		return
-	}
-	c.Writer.WriteString("data: ")
-	c.Writer.Write(b)
-	c.Writer.WriteString("\n\n")
-	c.Writer.Flush()
-}
-
 // messages 是 Anthropic /ai/v1/messages 端点。
 //
 // @Summary  Anthropic Messages 兼容端点
 // @Tags     AI 网关
-// @Param    body body object true "Anthropic messages 请求体(支持 stream)"
+// @Param    body body object true "Anthropic messages 请求体(支持 stream;经 OCI OpenAI 兼容面直通)"
 // @Success  200 {object} map[string]any "Anthropic 兼容响应(流式为 SSE)"
 // @Router   /ai/v1/messages [post]
 func (h *aiGatewayHandler) messages(c *gin.Context) {
@@ -308,69 +227,92 @@ func (h *aiGatewayHandler) messages(c *gin.Context) {
 		aiError(c, http.StatusBadRequest, "invalid_request_error", "max_tokens 必填且需大于 0")
 		return
 	}
-	ir, err := service.AnthropicToIR(req)
+	if len(req.Messages) == 0 {
+		aiError(c, http.StatusBadRequest, "invalid_request_error", "messages 不能为空")
+		return
+	}
+	if !checkKeyModel(c, req.Model) {
+		return
+	}
+	body, err := service.AnthropicToResponsesBody(req)
 	if err != nil {
 		aiError(c, http.StatusBadRequest, "invalid_request_error", err.Error())
 		return
 	}
-	if err := validateIR(ir); err != nil {
-		aiError(c, http.StatusBadRequest, "invalid_request_error", err.Error())
-		return
-	}
 	if req.Stream {
-		h.streamAnthropic(c, ir)
+		h.streamAnthropic(c, body, req)
 		return
 	}
 	start := time.Now()
-	resp, meta, err := h.gw.Chat(c.Request.Context(), ir, keyGroup(c))
+	payload, meta, err := h.gw.RespPassthrough(c.Request.Context(), body, req.Model, keyGroup(c))
-	entry := h.logEntry(c, "anthropic", ir.Model, false, meta, start)
+	entry := h.logEntry(c, "anthropic", req.Model, false, meta, start)
 	if err != nil {
 		upstreamError(c, err)
 		entry.ErrMsg = err.Error()
 		h.logFailure(c, entry, req)
 		return
 	}
+	out, err := service.ResponsesToAnthropic(payload, aiRandID("msg_"))
+	if err != nil {
+		aiError(c, http.StatusBadGateway, "api_error", err.Error())
+		entry.ErrMsg = err.Error()
+		h.logFailure(c, entry, req)
+		return
+	}
 	entry.Status = http.StatusOK
-	fillUsage(&entry, resp.Usage)
+	fillUsage(&entry, service.RespPassthroughUsage(payload))
 	callID := h.gw.LogCall(entry)
-	out := service.IRRespToAnthropic(resp, aiRandID("msg_"))
+	h.maybeLogContent(c, callID, "anthropic", req.Model, false, req, out)
-	h.maybeLogContent(c, callID, "anthropic", ir.Model, false, req, out)
 	c.JSON(http.StatusOK, out)
 }
 
-// streamAnthropic 经状态机把 IR chunk 流聚合为 Anthropic SSE 事件流。
+// streamAnthropic 流式直通并桥接:上游 Responses SSE 事件经 AnthRespBridge
-func (h *aiGatewayHandler) streamAnthropic(c *gin.Context, ir aiwire.ChatRequest) {
+// 转为 Anthropic 事件序列逐块写出;usage 从 completed 事件记账。
+func (h *aiGatewayHandler) streamAnthropic(c *gin.Context, body []byte, req aiwire.MessagesRequest) {
 	start := time.Now()
-	stream, meta, err := h.gw.OpenStream(c.Request.Context(), ir, keyGroup(c))
+	upstream, meta, err := h.gw.RespPassthroughStream(c.Request.Context(), body, req.Model, keyGroup(c))
-	entry := h.logEntry(c, "anthropic", ir.Model, true, meta, start)
+	entry := h.logEntry(c, "anthropic", req.Model, true, meta, start)
 	if err != nil {
 		upstreamError(c, err)
 		entry.ErrMsg = err.Error()
-		h.logFailure(c, entry, ir)
+		h.logFailure(c, entry, req)
 		return
 	}
-	defer stream.Close()
+	defer upstream.Close()
 	sseHeaders(c)
-	st := service.NewAnthStream(aiRandID("msg_"), ir.Model)
+	bridge := service.NewAnthRespBridge(aiRandID("msg_"), req.Model)
-	for {
+	if err := forwardAnthSSE(c, upstream, bridge); err != nil {
-		chunk, err := stream.Next()
-		if err != nil {
-			if !errors.Is(err, io.EOF) {
 		entry.ErrMsg = err.Error()
 	}
-			break
+	writeAnthEvents(c, bridge.Finish())
-		}
-		writeAnthEvents(c, st.Feed(chunk))
-	}
-	writeAnthEvents(c, st.Finish())
 	entry.Status = http.StatusOK
 	entry.LatencyMs = time.Since(start).Milliseconds()
-	usage := st.Usage()
+	usage := bridge.Usage()
 	entry.PromptTokens, entry.CompletionTokens = usage.InputTokens, usage.OutputTokens
 	entry.TotalTokens = usage.InputTokens + usage.OutputTokens
 	entry.CachedTokens = usage.CacheReadInputTokens
 	callID := h.gw.LogCall(entry)
-	h.maybeLogContent(c, callID, "anthropic", ir.Model, true, ir, nil)
+	h.maybeLogContent(c, callID, "anthropic", req.Model, true, req, nil)
+}
+
+// forwardAnthSSE 逐行读上游 SSE,把 data 行喂给桥并即时写出转换后的事件。
+func forwardAnthSSE(c *gin.Context, upstream io.Reader, bridge *service.AnthRespBridge) error {
+	reader := bufio.NewReader(upstream)
+	for {
+		line, err := reader.ReadBytes('\n')
+		if len(line) > 0 {
+			trimmed := bytes.TrimSpace(line)
+			if data, ok := bytes.CutPrefix(trimmed, []byte("data: ")); ok {
+				writeAnthEvents(c, bridge.Feed(data))
+			}
+		}
+		if err != nil {
+			if errors.Is(err, io.EOF) {
+				return nil
+			}
+			return err
+		}
+	}
 }
 
 func writeAnthEvents(c *gin.Context, events []service.AnthEvent) {
@@ -398,6 +340,15 @@ func (h *aiGatewayHandler) listModels(c *gin.Context) {
 		aiError(c, http.StatusInternalServerError, "api_error", "查询模型列表失败")
 		return
 	}
+	if allowed := keyModels(c); len(allowed) > 0 {
+		kept := make([]aiwire.Model, 0, len(list.Data))
+		for _, m := range list.Data {
+			if slices.Contains(allowed, m.ID) {
+				kept = append(kept, m)
+			}
+		}
+		list.Data = kept
+	}
 	c.JSON(http.StatusOK, list)
 }
 
@@ -405,31 +356,45 @@ func (h *aiGatewayHandler) listModels(c *gin.Context) {
 //
 // @Summary  OpenAI Responses 兼容端点
 // @Tags     AI 网关
-// @Param    body body object true "OpenAI responses 请求体(支持 stream)"
+// @Param    body body object true "OpenAI responses 请求体(支持 stream;web_search/x_search 服务端工具仅非流式)"
 // @Success  200 {object} map[string]any "OpenAI 兼容响应(流式为 SSE)"
 // @Router   /ai/v1/responses [post]
 func (h *aiGatewayHandler) responses(c *gin.Context) {
-	var req aiwire.RespRequest
+	raw, err := c.GetRawData()
-	if err := c.ShouldBindJSON(&req); err != nil {
-		aiError(c, http.StatusBadRequest, "invalid_request_error", err.Error())
-		return
-	}
-	ir, err := service.ResponsesToIR(req)
 	if err != nil {
 		aiError(c, http.StatusBadRequest, "invalid_request_error", err.Error())
 		return
 	}
-	if err := validateIR(ir); err != nil {
+	var req aiwire.RespRequest
+	if err := json.Unmarshal(raw, &req); err != nil {
+		aiError(c, http.StatusBadRequest, "invalid_request_error", err.Error())
+		return
+	}
+	if !checkKeyModel(c, req.Model) {
+		return
+	}
+	h.responsesPassthrough(c, raw, req)
+}
+
+// responsesPassthrough 直通链路(唯一上游):原始 body 改写后直达
+// OCI /actions/v1/responses,响应原样透传。
+func (h *aiGatewayHandler) responsesPassthrough(c *gin.Context, raw []byte, req aiwire.RespRequest) {
+	if err := service.RespPassthroughValidate(req); err != nil {
+		aiError(c, http.StatusBadRequest, "invalid_request_error", err.Error())
+		return
+	}
+	body, err := service.RespPassthroughBody(raw)
+	if err != nil {
 		aiError(c, http.StatusBadRequest, "invalid_request_error", err.Error())
 		return
 	}
 	if req.Stream {
-		h.streamResponses(c, ir)
+		h.responsesPassthroughStream(c, body, req)
 		return
 	}
 	start := time.Now()
-	resp, meta, err := h.gw.Chat(c.Request.Context(), ir, keyGroup(c))
+	payload, meta, err := h.gw.RespPassthrough(c.Request.Context(), body, req.Model, keyGroup(c))
-	entry := h.logEntry(c, "responses", ir.Model, false, meta, start)
+	entry := h.logEntry(c, "responses", req.Model, false, meta, start)
 	if err != nil {
 		upstreamError(c, err)
 		entry.ErrMsg = err.Error()
@@ -437,54 +402,61 @@ func (h *aiGatewayHandler) responses(c *gin.Context) {
 		return
 	}
 	entry.Status = http.StatusOK
-	fillUsage(&entry, resp.Usage)
+	fillUsage(&entry, service.RespPassthroughUsage(payload))
 	callID := h.gw.LogCall(entry)
-	out := service.IRRespToResponses(resp, aiRandID("resp_"), time.Now().Unix())
+	h.maybeLogContent(c, callID, "responses", req.Model, false, req, json.RawMessage(payload))
-	h.maybeLogContent(c, callID, "responses", ir.Model, false, req, out)
+	c.Data(http.StatusOK, "application/json; charset=utf-8", payload)
-	c.JSON(http.StatusOK, out)
 }
 
-// streamResponses 经状态机把 IR chunk 流聚合为 Responses 语义事件流。
+// responsesPassthroughStream 流式直通:SSE 事件原样转发(推理增量等直达客户端),
-func (h *aiGatewayHandler) streamResponses(c *gin.Context, ir aiwire.ChatRequest) {
+// 逐行扫描 completed 事件提取 usage 记账。
+func (h *aiGatewayHandler) responsesPassthroughStream(c *gin.Context, body []byte, req aiwire.RespRequest) {
 	start := time.Now()
-	stream, meta, err := h.gw.OpenStream(c.Request.Context(), ir, keyGroup(c))
+	upstream, meta, err := h.gw.RespPassthroughStream(c.Request.Context(), body, req.Model, keyGroup(c))
-	entry := h.logEntry(c, "responses", ir.Model, true, meta, start)
+	entry := h.logEntry(c, "responses", req.Model, true, meta, start)
 	if err != nil {
 		upstreamError(c, err)
 		entry.ErrMsg = err.Error()
-		h.logFailure(c, entry, ir)
+		h.logFailure(c, entry, req)
 		return
 	}
-	defer stream.Close()
+	defer upstream.Close()
 	sseHeaders(c)
-	st := service.NewRespStream(aiRandID("resp_"), ir.Model, time.Now().Unix())
+	usage, err := forwardSSE(c, upstream)
-	for {
-		chunk, err := stream.Next()
 	if err != nil {
-			if !errors.Is(err, io.EOF) {
 		entry.ErrMsg = err.Error()
 	}
-			break
-		}
-		writeRespEvents(c, st.Feed(chunk))
-	}
-	writeRespEvents(c, st.Finish())
 	entry.Status = http.StatusOK
 	entry.LatencyMs = time.Since(start).Milliseconds()
-	fillUsage(&entry, st.Usage())
+	fillUsage(&entry, usage)
 	callID := h.gw.LogCall(entry)
-	h.maybeLogContent(c, callID, "responses", ir.Model, true, ir, nil)
+	h.maybeLogContent(c, callID, "responses", req.Model, true, req, nil)
 }
 
-func writeRespEvents(c *gin.Context, events []service.RespEvent) {
+// forwardSSE 把上游 SSE 逐行转发给客户端,空行(事件边界)即 flush;
-	for _, ev := range events {
+// 顺带从 data 行提取 response.completed 的 usage。
-		b, err := json.Marshal(ev.Data)
+func forwardSSE(c *gin.Context, upstream io.Reader) (*aiwire.Usage, error) {
-		if err != nil {
+	reader := bufio.NewReader(upstream)
-			continue
+	var usage *aiwire.Usage
-		}
+	for {
-		c.Writer.WriteString("event: " + ev.Event + "\ndata: ")
+		line, err := reader.ReadBytes('\n')
-		c.Writer.Write(b)
+		if len(line) > 0 {
-		c.Writer.WriteString("\n\n")
+			c.Writer.Write(line)
-	}
+			trimmed := bytes.TrimSpace(line)
+			if len(trimmed) == 0 {
 				c.Writer.Flush()
+			} else if data, ok := bytes.CutPrefix(trimmed, []byte("data: ")); ok {
+				if u := service.RespStreamCompletedUsage(data); u != nil {
+					usage = u
+				}
+			}
+		}
+		if err != nil {
+			c.Writer.Flush()
+			if errors.Is(err, io.EOF) {
+				return usage, nil
+			}
+			return usage, err
+		}
+	}
 }

internal/api/aigateway_test.go

@@ -0,0 +1,206 @@
+package api
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"strconv"
+	"strings"
+	"testing"
+	"time"
+
+	"gorm.io/gorm"
+
+	"oci-portal/internal/model"
+)
+
+// seedGatewayModel 直插启用渠道与模型缓存,绕过云同步。
+func seedGatewayModel(t *testing.T, db *gorm.DB, name string) {
+	t.Helper()
+	ch := &model.AiChannel{Name: "t-" + name, OciConfigID: 1, Region: "r-" + name, Enabled: true, Priority: 1, Weight: 1}
+	if err := db.Create(ch).Error; err != nil {
+		t.Fatalf("seed channel: %v", err)
+	}
+	cache := &model.AiModelCache{ChannelID: ch.ID, ModelOcid: "ocid1.." + name, Name: name, Vendor: "v", SyncedAt: time.Now()}
+	if err := db.Create(cache).Error; err != nil {
+		t.Fatalf("seed model cache: %v", err)
+	}
+}
+
+func TestAiGatewayKeyModelRestrict(t *testing.T) {
+	r, auth, _, db := newTestRouterDB(t)
+	token, _, err := auth.Login(context.Background(), "admin", "pass123", "127.0.0.1", "")
+	if err != nil {
+		t.Fatalf("login: %v", err)
+	}
+
+	// 受限密钥:白名单含脏输入,入库应规范化为 2 项(cohere + ghost)
+	w := doRequest(t, r, http.MethodPost, "/api/v1/ai-keys", token,
+		`{"name":"limited","value":"limited-key-1234","models":[" cohere.command-a-03-2025 ","cohere.command-a-03-2025","","ghost-model"]}`)
+	if w.Code != http.StatusCreated {
+		t.Fatalf("创建受限密钥 status = %d, body %s", w.Code, w.Body.String())
+	}
+	var created struct {
+		Item model.AiKey `json:"item"`
+	}
+	if err := json.Unmarshal(w.Body.Bytes(), &created); err != nil {
+		t.Fatalf("decode: %v", err)
+	}
+	if len(created.Item.Models) != 2 || created.Item.Models[0] != "cohere.command-a-03-2025" {
+		t.Fatalf("创建后 models 未规范化: %v", created.Item.Models)
+	}
+	// 不限密钥对照
+	w = doRequest(t, r, http.MethodPost, "/api/v1/ai-keys", token, `{"name":"open","value":"open-key-12345"}`)
+	if w.Code != http.StatusCreated {
+		t.Fatalf("创建不限密钥 status = %d, body %s", w.Code, w.Body.String())
+	}
+
+	seedGatewayModel(t, db, "cohere.command-a-03-2025")
+	seedGatewayModel(t, db, "meta.llama-3.3-70b-instruct")
+
+	deny := []string{"model_not_found", "无权访问"}
+	pass := []string{"未知模型"} // 穿过白名单闸门,由编排层以「未知模型」拒绝(池内无 ghost-model)
+	tests := []struct {
+		name     string
+		key      string
+		path     string
+		body     string
+		wantCode int
+		wantSub  []string
+	}{
+		{"responses 流式同样拦截", "limited-key-1234", "/ai/v1/responses",
+			`{"model":"meta.llama-3.3-70b-instruct","stream":true,"input":"hi"}`, 404, deny},
+		{"responses 白名单内穿透", "limited-key-1234", "/ai/v1/responses",
+			`{"model":"ghost-model","input":"hi"}`, 404, pass},
+		{"embeddings 白名单外拦截", "limited-key-1234", "/ai/v1/embeddings",
+			`{"model":"meta.llama-3.3-70b-instruct","input":["hi"]}`, 404, deny},
+		{"messages 白名单外拦截为 Anthropic 体", "limited-key-1234", "/ai/v1/messages",
+			`{"model":"meta.llama-3.3-70b-instruct","max_tokens":16,"messages":[{"role":"user","content":"hi"}]}`, 404,
+			[]string{`"type":"error"`, "model_not_found", "无权访问"}},
+		{"responses 白名单外拦截", "limited-key-1234", "/ai/v1/responses",
+			`{"model":"meta.llama-3.3-70b-instruct","input":"hi"}`, 404, deny},
+		{"不限密钥穿透闸门", "open-key-12345", "/ai/v1/responses",
+			`{"model":"ghost-model","input":"hi"}`, 404, pass},
+		{"chat completions 端点已删除", "open-key-12345", "/ai/v1/chat/completions",
+			`{"model":"ghost-model","messages":[{"role":"user","content":"hi"}]}`, 404, []string{"not found"}},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			w := doRequest(t, r, http.MethodPost, tt.path, tt.key, tt.body)
+			if w.Code != tt.wantCode {
+				t.Fatalf("status = %d, want %d, body %s", w.Code, tt.wantCode, w.Body.String())
+			}
+			for _, sub := range tt.wantSub {
+				if !strings.Contains(w.Body.String(), sub) {
+					t.Errorf("body 缺少 %q: %s", sub, w.Body.String())
+				}
+			}
+		})
+	}
+}
+
+func TestAiGatewayModelsFilteredByKey(t *testing.T) {
+	r, auth, _, db := newTestRouterDB(t)
+	token, _, err := auth.Login(context.Background(), "admin", "pass123", "127.0.0.1", "")
+	if err != nil {
+		t.Fatalf("login: %v", err)
+	}
+	seedGatewayModel(t, db, "cohere.command-a-03-2025")
+	seedGatewayModel(t, db, "meta.llama-3.3-70b-instruct")
+	w := doRequest(t, r, http.MethodPost, "/api/v1/ai-keys", token,
+		`{"name":"limited","value":"limited-key-1234","models":["cohere.command-a-03-2025"]}`)
+	if w.Code != http.StatusCreated {
+		t.Fatalf("创建密钥 status = %d", w.Code)
+	}
+	w = doRequest(t, r, http.MethodPost, "/api/v1/ai-keys", token, `{"name":"open","value":"open-key-12345"}`)
+	if w.Code != http.StatusCreated {
+		t.Fatalf("创建密钥 status = %d", w.Code)
+	}
+
+	tests := []struct {
+		name    string
+		key     string
+		want    []string
+		notWant []string
+	}{
+		{"受限密钥仅见交集", "limited-key-1234", []string{"cohere.command-a-03-2025"}, []string{"meta.llama-3.3-70b-instruct"}},
+		{"不限密钥全量可见", "open-key-12345", []string{"cohere.command-a-03-2025", "meta.llama-3.3-70b-instruct"}, nil},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			w := doRequest(t, r, http.MethodGet, "/ai/v1/models", tt.key, "")
+			if w.Code != http.StatusOK {
+				t.Fatalf("status = %d, body %s", w.Code, w.Body.String())
+			}
+			for _, sub := range tt.want {
+				if !strings.Contains(w.Body.String(), sub) {
+					t.Errorf("应包含 %q: %s", sub, w.Body.String())
+				}
+			}
+			for _, sub := range tt.notWant {
+				if strings.Contains(w.Body.String(), sub) {
+					t.Errorf("不应包含 %q: %s", sub, w.Body.String())
+				}
+			}
+		})
+	}
+}
+
+func TestAiKeyModelsUpdateRoundTrip(t *testing.T) {
+	r, auth, _, _ := newTestRouterDB(t)
+	token, _, err := auth.Login(context.Background(), "admin", "pass123", "127.0.0.1", "")
+	if err != nil {
+		t.Fatalf("login: %v", err)
+	}
+	w := doRequest(t, r, http.MethodPost, "/api/v1/ai-keys", token, `{"name":"k","value":"round-key-1234"}`)
+	if w.Code != http.StatusCreated {
+		t.Fatalf("创建 status = %d", w.Code)
+	}
+	var created struct {
+		Item model.AiKey `json:"item"`
+	}
+	_ = json.Unmarshal(w.Body.Bytes(), &created)
+	id := created.Item.ID
+
+	listModels := func() []string {
+		w := doRequest(t, r, http.MethodGet, "/api/v1/ai-keys", token, "")
+		var resp struct {
+			Items []model.AiKey `json:"items"`
+		}
+		_ = json.Unmarshal(w.Body.Bytes(), &resp)
+		for _, k := range resp.Items {
+			if k.ID == id {
+				return k.Models
+			}
+		}
+		t.Fatalf("密钥 %d 不在列表中", id)
+		return nil
+	}
+
+	tests := []struct {
+		name string
+		body string
+		want []string
+	}{
+		{"设置白名单", `{"models":[" a-model ","a-model","b-model"]}`, []string{"a-model", "b-model"}},
+		{"不传 models 保持不变", `{"name":"k2"}`, []string{"a-model", "b-model"}},
+		{"空数组清空恢复不限", `{"models":[]}`, nil},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			w := doRequest(t, r, http.MethodPut, "/api/v1/ai-keys/"+strconv.Itoa(int(id)), token, tt.body)
+			if w.Code != http.StatusNoContent {
+				t.Fatalf("update status = %d, body %s", w.Code, w.Body.String())
+			}
+			got := listModels()
+			if len(got) != len(tt.want) {
+				t.Fatalf("models = %v, want %v", got, tt.want)
+			}
+			for i := range tt.want {
+				if got[i] != tt.want[i] {
+					t.Fatalf("models = %v, want %v", got, tt.want)
+				}
+			}
+		})
+	}
+}

internal/api/router_test.go

@@ -57,6 +57,13 @@ func (nullClient) GetSubscription(context.Context, oci.Credentials, string) (oci
 }
 
 func newTestRouter(t *testing.T) (*gin.Engine, *service.AuthService, *service.SystemLogService) {
+	t.Helper()
+	r, auth, systemLogs, _ := newTestRouterDB(t)
+	return r, auth, systemLogs
+}
+
+// newTestRouterDB 额外回传 db 句柄,供需要直插数据(如网关模型缓存)的测试使用。
+func newTestRouterDB(t *testing.T) (*gin.Engine, *service.AuthService, *service.SystemLogService, *gorm.DB) {
 	t.Helper()
 	gin.SetMode(gin.TestMode)
 	db, err := gorm.Open(sqlite.Open(":memory:"), &gorm.Config{
@@ -90,7 +97,7 @@ func newTestRouter(t *testing.T) (*gin.Engine, *service.AuthService, *service.Sy
 	logEvents := service.NewLogEventService(db)
 	oauth := service.NewOAuthService(db, settings, auth)
 	r := NewRouter(auth, oauth, ociConfigs, tasks, service.NewConsoleService(ociConfigs), settings, notifier, systemLogs, logEvents, service.NewProxyService(db, cipher), service.NewAiGatewayService(db, ociConfigs, nullClient{}))
-	return r, auth, systemLogs
+	return r, auth, systemLogs, db
 }
 
 func doRequest(t *testing.T, r *gin.Engine, method, path, token, body string) *httptest.ResponseRecorder {

internal/api/routes_ai.go

@@ -11,7 +11,6 @@ import (
 func registerAiGateway(r *gin.Engine, aiGateway *service.AiGatewayService) {
 	aih := &aiGatewayHandler{gw: aiGateway}
 	ai := r.Group("/ai/v1", bodyLimit(10<<20), aih.auth)
-	ai.POST("/chat/completions", aih.chatCompletions)
 	ai.POST("/responses", aih.responses)
 	ai.POST("/messages", aih.messages)
 	ai.POST("/embeddings", aih.embeddings)
@@ -33,6 +32,9 @@ func registerAiAdmin(secured *gin.RouterGroup, aiGateway *service.AiGatewayServi
 	secured.POST("/ai-channels/:id/probe", aiadmin.probeChannel)
 	secured.POST("/ai-channels/:id/sync-models", aiadmin.syncChannelModels)
 	secured.GET("/ai-models", aiadmin.gatewayModels)
+	secured.GET("/ai-blacklist", aiadmin.listBlacklist)
+	secured.POST("/ai-blacklist", aiadmin.addBlacklist)
+	secured.DELETE("/ai-blacklist/:id", aiadmin.removeBlacklist)
 	secured.GET("/ai-logs", aiadmin.listLogs)
 	secured.GET("/ai-content-logs", aiadmin.listContentLogs)
 }

internal/database/database.go

@@ -68,7 +68,7 @@ func autoMigrate(db *gorm.DB) error {
 		&model.CheckSnapshot{}, &model.CostSnapshot{},
 		&model.RegionCache{}, &model.CompartmentCache{}, &model.Setting{},
 		&model.SystemLog{}, &model.LogEvent{}, &model.AlertRule{}, &model.AlertRuleHit{}, &model.Proxy{},
-		&model.AiKey{}, &model.AiChannel{}, &model.AiModelCache{}, &model.AiCallLog{},
+		&model.AiKey{}, &model.AiChannel{}, &model.AiModelCache{}, &model.AiModelBlacklist{}, &model.AiCallLog{},
 		&model.AiContentLog{},
 	)
 }

internal/model/models.go

@@ -266,6 +266,8 @@ type AiKey struct {
 	KeyHash string `gorm:"uniqueIndex;size:64" json:"-"`
 	Tail    string `gorm:"size:4" json:"tail"`
 	Group   string `gorm:"column:key_group;size:32" json:"group"`
+	// Models 是模型白名单(精确匹配模型名);空 = 不限,非空时网关仅放行列表内模型
+	Models     []string   `gorm:"serializer:json;type:text" json:"models"`
 	Enabled    bool       `json:"enabled"`
 	LastUsedAt *time.Time `json:"lastUsedAt"`
 	// ContentLogUntil 是内容日志开启截止时间;nil = 永关(默认,红线),
@@ -309,13 +311,14 @@ type AiModelCache struct {
 	// 弃用后仍可调用,直到 RetiredAt(按需推理退役,同步层已剔除过期项)。
 	DeprecatedAt *time.Time `json:"deprecatedAt"`
 	RetiredAt    *time.Time `json:"retiredAt"`
-	// Unusable 标记该模型在此渠道不可按需调用(微调基座/区域未开放按需),
+}
-	// 由探测、真实调用或同步后的后台验证习得:不参与网关列表/路由/探测候选;
+
-	// 任何同步都按 OCID 结转标记,已标记模型定期复检、恢复供给自动解除。
+// AiModelBlacklist 是全局模型黑名单(按模型名):加入即从全部渠道缓存删除该模型,
-	Unusable       bool   `json:"unusable"`
+// 同步与探测过滤不再入库;适用于微调基座等实际不可按需调用的模型,由用户手动维护。
-	UnusableReason string `gorm:"size:200" json:"unusableReason,omitempty"`
+type AiModelBlacklist struct {
-	// CheckedAt 是最近一次可用性验证时间;NULL 表示从未验证(同步后的后台验证以此为队列)。
+	ID        uint      `gorm:"primaryKey" json:"id"`
-	CheckedAt *time.Time `json:"checkedAt"`
+	Name      string    `gorm:"size:96;uniqueIndex" json:"name"`
+	CreatedAt time.Time `json:"createdAt"`
 }
 
 // AiCallLog 是网关调用日志:仅元数据与 token 用量,绝不记录 prompt / 响应正文。

internal/oci/client.go

@@ -6,6 +6,7 @@ import (
 	"context"
 	"errors"
 	"fmt"
+	"io"
 	"strings"
 	"sync"
 	"time"
@@ -99,10 +100,12 @@ type Client interface {
 	AddVnicIpv6(ctx context.Context, cred Credentials, region, vnicID, address string) (string, error)
 	// GenAI 网关:区域模型列表、聊天(非流式/流式)、配额探测、文本向量化。
 	ListGenAiModels(ctx context.Context, cred Credentials, region string) ([]GenAiModel, error)
-	GenAiChat(ctx context.Context, cred Credentials, region, modelOcid string, ir aiwire.ChatRequest) (*aiwire.ChatResponse, error)
-	GenAiChatStream(ctx context.Context, cred Credentials, region, modelOcid string, ir aiwire.ChatRequest) (GenAiStream, error)
 	GenAiProbeChat(ctx context.Context, cred Credentials, region, modelOcid, modelName string) (int, error)
 	GenAiEmbed(ctx context.Context, cred Credentials, region, modelOcid string, inputs []string, dimensions *int) ([][]float32, *aiwire.Usage, error)
+	// GenAiCompatResponses 直通 OpenAI Responses 请求体到 /actions/v1/responses(xAI 服务端工具通路)。
+	GenAiCompatResponses(ctx context.Context, cred Credentials, region string, body []byte) ([]byte, error)
+	// GenAiCompatResponsesStream 流式直通 /actions/v1/responses,建立成功返回 SSE body。
+	GenAiCompatResponsesStream(ctx context.Context, cred Credentials, region string, body []byte) (io.ReadCloser, error)
 	// 控制台连接：创建（VNC/串口连接串）、列出、删除。
 	CreateConsoleConnection(ctx context.Context, cred Credentials, region, instanceID, sshPublicKey string) (ConsoleConnection, error)
 	ListConsoleConnections(ctx context.Context, cred Credentials, region, instanceID string) ([]ConsoleConnection, error)

internal/oci/genai.go

@@ -2,13 +2,11 @@ package oci
 
 import (
 	"context"
+	"encoding/json"
 	"fmt"
-	"io"
 	"net/http"
-	"strings"
 	"time"
 
-	"github.com/oracle/oci-go-sdk/v65/common"
 	"github.com/oracle/oci-go-sdk/v65/generativeai"
 	"github.com/oracle/oci-go-sdk/v65/generativeaiinference"
 
@@ -30,12 +28,6 @@ type GenAiModel struct {
 	Retired *time.Time `json:"retired"`
 }
 
-// GenAiStream 逐事件读取流式聊天;Next 在流结束时返回 io.EOF。
-type GenAiStream interface {
-	Next() (aiwire.ChatChunk, error)
-	Close() error
-}
-
 func (c *RealClient) genAiClient(cred Credentials, region string) (generativeai.GenerativeAiClient, error) {
 	gc, err := generativeai.NewGenerativeAiClientWithConfigurationProvider(provider(cred))
 	if err != nil {
@@ -152,121 +144,15 @@ func capStrings(caps []generativeai.ModelCapabilityEnum) []string {
 	return out
 }
 
-// GenAiChat 实现 Client:非流式聊天;modelOcid 走 on-demand serving,
+// GenAiProbeChat 实现 Client:经 OpenAI 兼容面(直通同链路)发一次极小请求探测渠道;配额探测专用的最小聊天(maxTokens=1),返回 HTTP 状态码;
-// cohere.* 模型走 COHERE 请求格式,其余走 GENERIC。
-func (c *RealClient) GenAiChat(ctx context.Context, cred Credentials, region, modelOcid string, ir aiwire.ChatRequest) (*aiwire.ChatResponse, error) {
-	ic, err := c.genAiInferenceClient(cred, region)
-	if err != nil {
-		return nil, err
-	}
-	req, err := buildChatRequest(ir, false)
-	if err != nil {
-		return nil, err
-	}
-	resp, err := ic.Chat(ctx, chatRequest(cred, modelOcid, req))
-	if err != nil {
-		return nil, fmt.Errorf("genai chat: %w", err)
-	}
-	return chatResponseToIR(resp.ChatResult.ChatResponse, ir.Model)
-}
-
-// buildChatRequest 按模型 vendor 组装底层请求体。
-func buildChatRequest(ir aiwire.ChatRequest, stream bool) (generativeaiinference.BaseChatRequest, error) {
-	if isCohereModel(ir.Model) {
-		return irToCohereSDK(ir, stream)
-	}
-	return irToSDK(ir, stream), nil
-}
-
-// chatResponseToIR 按响应实际形态(GENERIC / COHERE)转回 IR。
-func chatResponseToIR(resp generativeaiinference.BaseChatResponse, model string) (*aiwire.ChatResponse, error) {
-	switch r := resp.(type) {
-	case generativeaiinference.GenericChatResponse:
-		return sdkToIR(r, model), nil
-	case generativeaiinference.CohereChatResponse:
-		return cohereSDKToIR(r, model), nil
-	default:
-		return nil, fmt.Errorf("genai chat: unexpected response format %T", resp)
-	}
-}
-
-func chatRequest(cred Credentials, modelOcid string, req generativeaiinference.BaseChatRequest) generativeaiinference.ChatRequest {
-	return generativeaiinference.ChatRequest{
-		ChatDetails: generativeaiinference.ChatDetails{
-			CompartmentId: &cred.TenancyOCID,
-			ServingMode:   generativeaiinference.OnDemandServingMode{ModelId: &modelOcid},
-			ChatRequest:   req,
-		},
-	}
-}
-
-// GenAiChatStream 实现 Client:流式聊天,返回逐事件读取器。
-// SDK 对 text/event-stream 跳过 unmarshal,原始流经 RawResponse 交由 SSEReader 消费。
-func (c *RealClient) GenAiChatStream(ctx context.Context, cred Credentials, region, modelOcid string, ir aiwire.ChatRequest) (GenAiStream, error) {
-	ic, err := c.genAiInferenceClient(cred, region)
-	if err != nil {
-		return nil, err
-	}
-	req, err := buildChatRequest(ir, true)
-	if err != nil {
-		return nil, err
-	}
-	resp, err := ic.Chat(ctx, chatRequest(cred, modelOcid, req))
-	if err != nil {
-		return nil, fmt.Errorf("genai chat stream: %w", err)
-	}
-	reader, err := common.NewSSEReader(resp.RawResponse)
-	if err != nil {
-		return nil, fmt.Errorf("genai chat stream: sse reader: %w", err)
-	}
-	return &genAiSSEStream{reader: reader, body: resp.RawResponse.Body, model: ir.Model}, nil
-}
-
-// genAiSSEStream 把 OCI SSE 事件流适配为 IR chunk 流。
-type genAiSSEStream struct {
-	reader *common.SseReader
-	body   io.ReadCloser
-	model  string
-}
-
-// Next 读取下一条有效事件;空事件与 [DONE] 哨兵跳过,流尽返回 io.EOF。
-func (s *genAiSSEStream) Next() (aiwire.ChatChunk, error) {
-	for {
-		data, err := s.reader.ReadNextEvent()
-		if err != nil {
-			return aiwire.ChatChunk{}, err
-		}
-		text := strings.TrimSpace(string(data))
-		if text == "" {
-			continue
-		}
-		if text == "[DONE]" {
-			return aiwire.ChatChunk{}, io.EOF
-		}
-		if chunk, ok := parseGenAiEvent([]byte(text), s.model); ok {
-			return chunk, nil
-		}
-	}
-}
-
-func (s *genAiSSEStream) Close() error {
-	if s.body != nil {
-		return s.body.Close()
-	}
-	return nil
-}
-
-// GenAiProbeChat 实现 Client:配额探测专用的最小聊天(maxTokens=1),返回 HTTP 状态码;
 // modelName 决定请求格式(cohere.* 走 COHERE)。
 func (c *RealClient) GenAiProbeChat(ctx context.Context, cred Credentials, region, modelOcid, modelName string) (int, error) {
-	one := 1
+	body, err := json.Marshal(map[string]any{"model": modelName, "input": "hi",
-	ir := aiwire.ChatRequest{
+		"max_output_tokens": 1, "store": false})
-		Model:     modelName,
+	if err != nil {
-		Messages:  []aiwire.ChatMessage{{Role: "user", Content: aiwire.NewTextContent("hi")}},
+		return 0, err
-		MaxTokens: &one,
 	}
-	_, err := c.GenAiChat(ctx, cred, region, modelOcid, ir)
+	if _, err = c.GenAiCompatResponses(ctx, cred, region, body); err == nil {
-	if err == nil {
 		return http.StatusOK, nil
 	}
 	if status, ok := ServiceStatus(err); ok {
@@ -275,6 +161,27 @@ func (c *RealClient) GenAiProbeChat(ctx context.Context, cred Credentials, regio
 	return 0, err
 }
 
+// sdkUsageToIR 把 SDK 用量转为内部记账结构(缓存命中挂 details,仅命中时出现)。
+func sdkUsageToIR(u *generativeaiinference.Usage) *aiwire.Usage {
+	if u == nil {
+		return nil
+	}
+	out := &aiwire.Usage{}
+	if u.PromptTokens != nil {
+		out.PromptTokens = *u.PromptTokens
+	}
+	if u.CompletionTokens != nil {
+		out.CompletionTokens = *u.CompletionTokens
+	}
+	if u.TotalTokens != nil {
+		out.TotalTokens = *u.TotalTokens
+	}
+	if u.PromptTokensDetails != nil && u.PromptTokensDetails.CachedTokens != nil && *u.PromptTokensDetails.CachedTokens > 0 {
+		out.PromptTokensDetails = &aiwire.PromptTokensDetails{CachedTokens: *u.PromptTokensDetails.CachedTokens}
+	}
+	return out
+}
+
 // GenAiEmbed 实现 Client:文本向量化(on-demand serving);dimensions 透传 OutputDimensions。
 func (c *RealClient) GenAiEmbed(ctx context.Context, cred Credentials, region, modelOcid string, inputs []string, dimensions *int) ([][]float32, *aiwire.Usage, error) {
 	ic, err := c.genAiInferenceClient(cred, region)

internal/oci/genai_cohere.go

@@ -1,266 +0,0 @@
-package oci
-
-import (
-	"encoding/json"
-	"fmt"
-	"strings"
-
-	"github.com/oracle/oci-go-sdk/v65/generativeaiinference"
-
-	"oci-portal/internal/aiwire"
-)
-
-// isCohereModel 依据模型名前缀判定 COHERE 系(走专属请求格式)。
-func isCohereModel(name string) bool {
-	return strings.HasPrefix(strings.ToLower(name), "cohere.")
-}
-
-// irToCohereSDK 把 IR 转为 COHERE 聊天请求;工具与结构化输出做有损降级
-// (参数仅取 JSON Schema 顶层 properties,tool_choice / json_schema 的 name、strict 无对应)。
-func irToCohereSDK(ir aiwire.ChatRequest, stream bool) (generativeaiinference.CohereChatRequest, error) {
-	var req generativeaiinference.CohereChatRequest
-	if err := cohereRejectUnsupported(ir); err != nil {
-		return req, err
-	}
-	p, err := cohereSplitMessages(ir.Messages)
-	if err != nil {
-		return req, err
-	}
-	req = generativeaiinference.CohereChatRequest{
-		Message:          &p.message,
-		ChatHistory:      p.history,
-		ToolResults:      p.toolResults,
-		Tools:            cohereTools(ir.Tools),
-		ResponseFormat:   cohereResponseFormat(ir.ResponseFormat),
-		MaxTokens:        ir.MaxTokens,
-		Temperature:      ir.Temperature,
-		TopP:             ir.TopP,
-		TopK:             ir.TopK,
-		FrequencyPenalty: ir.FrequencyPenalty,
-		PresencePenalty:  ir.PresencePenalty,
-		Seed:             ir.Seed,
-	}
-	if p.preamble != "" {
-		req.PreambleOverride = &p.preamble
-	}
-	if len(ir.Stop) > 0 {
-		req.StopSequences = ir.Stop
-	}
-	if stream {
-		req.IsStream = &stream
-		includeUsage := ir.StreamOptions == nil || ir.StreamOptions.IncludeUsage
-		req.StreamOptions = &generativeaiinference.StreamOptions{IsIncludeUsage: &includeUsage}
-	}
-	return req, nil
-}
-
-// cohereRejectUnsupported 拒绝 COHERE 无法承接的能力(多模态图片)。
-func cohereRejectUnsupported(ir aiwire.ChatRequest) error {
-	for _, m := range ir.Messages {
-		for _, p := range m.Content.Parts {
-			if p.Type == "image_url" {
-				return fmt.Errorf("cohere 系模型暂不支持图片输入,请改用 meta/google 等多模态模型")
-			}
-		}
-	}
-	return nil
-}
-
-// cohereParts 是 IR 消息拆装为 COHERE 请求的中间结果。
-type cohereParts struct {
-	message     string
-	history     []generativeaiinference.CohereMessage
-	preamble    string
-	toolResults []generativeaiinference.CohereToolResult
-}
-
-// cohereSplitMessages 拆装消息:system 拼 preamble,末尾 user 抽为 message,
-// assistant(含 toolCalls)与其余 user 进 history,tool 结果转顶层 toolResults
-// (工具结果在末尾时 message 留空,COHERE 以 toolResults 续跑)。
-func cohereSplitMessages(msgs []aiwire.ChatMessage) (cohereParts, error) {
-	lastUser, lastTool := -1, -1
-	for i, m := range msgs {
-		switch m.Role {
-		case "user":
-			lastUser = i
-		case "tool":
-			lastTool = i
-		}
-	}
-	if lastUser == -1 && lastTool == -1 {
-		return cohereParts{}, fmt.Errorf("cohere 系模型至少需要一条 user 消息")
-	}
-	var p cohereParts
-	var preamble []string
-	calls := map[string]generativeaiinference.CohereToolCall{}
-	for i, m := range msgs {
-		text := m.Content.JoinText()
-		switch m.Role {
-		case "system", "developer":
-			preamble = append(preamble, text)
-		case "assistant":
-			p.history = append(p.history, cohereBotMessage(text, m.ToolCalls, calls))
-		case "tool":
-			p.toolResults = append(p.toolResults, cohereToolResult(m, calls))
-		default: // user
-			if i == lastUser && lastUser > lastTool {
-				p.message = text
-				continue
-			}
-			p.history = append(p.history, generativeaiinference.CohereUserMessage{Message: &text})
-		}
-	}
-	p.preamble = strings.Join(preamble, "\n")
-	return p, nil
-}
-
-// cohereBotMessage 转 assistant 消息;toolCalls 同步登记进 id→call 映射供 tool 结果回查。
-func cohereBotMessage(text string, tcs []aiwire.ToolCall, calls map[string]generativeaiinference.CohereToolCall) generativeaiinference.CohereChatBotMessage {
-	msg := generativeaiinference.CohereChatBotMessage{}
-	if text != "" {
-		msg.Message = &text
-	}
-	for _, tc := range tcs {
-		name := tc.Function.Name
-		var params interface{}
-		if json.Unmarshal([]byte(tc.Function.Arguments), &params) != nil || params == nil {
-			params = map[string]any{}
-		}
-		call := generativeaiinference.CohereToolCall{Name: &name, Parameters: &params}
-		calls[tc.ID] = call
-		msg.ToolCalls = append(msg.ToolCalls, call)
-	}
-	return msg
-}
-
-// cohereToolResult 把 IR tool 消息转顶层工具结果;COHERE 工具调用无 id,
-// 靠 assistant 历史登记的映射回查,缺失时以 tool_call_id 名义调用兜底。
-func cohereToolResult(m aiwire.ChatMessage, calls map[string]generativeaiinference.CohereToolCall) generativeaiinference.CohereToolResult {
-	call, ok := calls[m.ToolCallID]
-	if !ok {
-		name := m.ToolCallID
-		var params interface{} = map[string]any{}
-		call = generativeaiinference.CohereToolCall{Name: &name, Parameters: &params}
-	}
-	text := m.Content.JoinText()
-	var output interface{}
-	if json.Unmarshal([]byte(text), &output) != nil || output == nil {
-		output = map[string]any{"output": text}
-	}
-	if arr, isArr := output.([]interface{}); isArr {
-		return generativeaiinference.CohereToolResult{Call: &call, Outputs: arr}
-	}
-	if _, isMap := output.(map[string]interface{}); !isMap {
-		output = map[string]any{"output": output}
-	}
-	return generativeaiinference.CohereToolResult{Call: &call, Outputs: []interface{}{output}}
-}
-
-// cohereTools 把 JSON Schema 工具定义降级为 COHERE 扁平参数表(嵌套结构有损:仅取顶层)。
-func cohereTools(tools []aiwire.Tool) []generativeaiinference.CohereTool {
-	if len(tools) == 0 {
-		return nil
-	}
-	out := make([]generativeaiinference.CohereTool, 0, len(tools))
-	for _, t := range tools {
-		name, desc := t.Function.Name, t.Function.Description
-		if desc == "" {
-			desc = name // Description 为 COHERE 必填
-		}
-		out = append(out, generativeaiinference.CohereTool{
-			Name: &name, Description: &desc,
-			ParameterDefinitions: cohereParams(t.Function.Parameters),
-		})
-	}
-	return out
-}
-
-// cohereParams 取 JSON Schema 顶层 properties 转扁平参数定义;嵌套 schema 只保留类型名。
-func cohereParams(schema json.RawMessage) map[string]generativeaiinference.CohereParameterDefinition {
-	var s struct {
-		Properties map[string]struct {
-			Type        string `json:"type"`
-			Description string `json:"description"`
-		} `json:"properties"`
-		Required []string `json:"required"`
-	}
-	if json.Unmarshal(schema, &s) != nil || len(s.Properties) == 0 {
-		return nil
-	}
-	req := map[string]bool{}
-	for _, r := range s.Required {
-		req[r] = true
-	}
-	out := make(map[string]generativeaiinference.CohereParameterDefinition, len(s.Properties))
-	for k, v := range s.Properties {
-		typ, desc := v.Type, v.Description
-		if typ == "" {
-			typ = "string"
-		}
-		pd := generativeaiinference.CohereParameterDefinition{Type: &typ}
-		if desc != "" {
-			pd.Description = &desc
-		}
-		if req[k] {
-			t := true
-			pd.IsRequired = &t
-		}
-		out[k] = pd
-	}
-	return out
-}
-
-// cohereResponseFormat 映射 json_object / json_schema(name、strict 无对应,有损降级)。
-func cohereResponseFormat(rf *aiwire.ResponseFormat) generativeaiinference.CohereResponseFormat {
-	if rf == nil {
-		return nil
-	}
-	switch rf.Type {
-	case "json_object", "json_schema":
-		out := generativeaiinference.CohereResponseJsonFormat{}
-		var in struct {
-			Schema json.RawMessage `json:"schema"`
-		}
-		if json.Unmarshal(rf.JSONSchema, &in) == nil && len(in.Schema) > 0 {
-			var schema interface{}
-			if json.Unmarshal(in.Schema, &schema) == nil {
-				out.Schema = &schema
-			}
-		}
-		return out
-	}
-	return nil
-}
-
-// cohereSDKToIR 把 COHERE 非流式响应转回 IR;工具调用补派生 ID,存在时结束原因归 tool_calls。
-func cohereSDKToIR(resp generativeaiinference.CohereChatResponse, model string) *aiwire.ChatResponse {
-	msg := aiwire.ChatMessage{Role: "assistant", Content: aiwire.NewTextContent(deref(resp.Text))}
-	for i, tc := range resp.ToolCalls {
-		msg.ToolCalls = append(msg.ToolCalls, cohereCallToIR(deref(tc.Name), tc.Parameters, i))
-	}
-	finish := mapFinishReason(string(resp.FinishReason))
-	if len(msg.ToolCalls) > 0 {
-		finish = "tool_calls"
-	}
-	return &aiwire.ChatResponse{
-		Object:  "chat.completion",
-		Model:   model,
-		Choices: []aiwire.Choice{{Message: msg, FinishReason: finish}},
-		Usage:   sdkUsageToIR(resp.Usage),
-	}
-}
-
-// cohereCallToIR 生成 IR 工具调用;COHERE 无调用 id,按名称+序号派生(回传时按此回查)。
-func cohereCallToIR(name string, params *interface{}, idx int) aiwire.ToolCall {
-	args := "{}"
-	if params != nil && *params != nil {
-		if b, err := json.Marshal(*params); err == nil {
-			args = string(b)
-		}
-	}
-	return aiwire.ToolCall{
-		ID:       fmt.Sprintf("call_%s_%d", name, idx),
-		Type:     "function",
-		Function: aiwire.FunctionCall{Name: name, Arguments: args},
-	}
-}

internal/oci/genai_convert.go

@@ -1,367 +0,0 @@
-package oci
-
-import (
-	"encoding/json"
-	"strings"
-
-	"github.com/oracle/oci-go-sdk/v65/generativeaiinference"
-
-	"oci-portal/internal/aiwire"
-)
-
-// irToSDK 把 IR(OpenAI 线格式)转为 OCI GENERIC 聊天请求;字段近一一镜像。
-func irToSDK(ir aiwire.ChatRequest, stream bool) generativeaiinference.GenericChatRequest {
-	req := generativeaiinference.GenericChatRequest{
-		Messages:            irMessages(ir.Messages),
-		MaxTokens:           ir.MaxTokens,
-		MaxCompletionTokens: ir.MaxCompletionTokens,
-		Temperature:         ir.Temperature,
-		TopP:                ir.TopP,
-		TopK:                ir.TopK,
-		FrequencyPenalty:    ir.FrequencyPenalty,
-		PresencePenalty:     ir.PresencePenalty,
-		Seed:                ir.Seed,
-		NumGenerations:      ir.N,
-		Tools:               irTools(ir.Tools),
-		ToolChoice:          irToolChoice(ir.ToolChoice),
-		ResponseFormat:      irResponseFormat(ir.ResponseFormat),
-	}
-	if len(ir.Stop) > 0 {
-		req.Stop = ir.Stop
-	}
-	if ir.ReasoningEffort != "" {
-		req.ReasoningEffort = generativeaiinference.GenericChatRequestReasoningEffortEnum(strings.ToUpper(ir.ReasoningEffort))
-	}
-	if stream {
-		req.IsStream = &stream
-		includeUsage := ir.StreamOptions == nil || ir.StreamOptions.IncludeUsage
-		req.StreamOptions = &generativeaiinference.StreamOptions{IsIncludeUsage: &includeUsage}
-	}
-	return req
-}
-
-// irMessages 按 role 拆装消息;user 消息保留图文块,其余角色只取文本。
-func irMessages(msgs []aiwire.ChatMessage) []generativeaiinference.Message {
-	out := make([]generativeaiinference.Message, 0, len(msgs))
-	for _, m := range msgs {
-		switch m.Role {
-		case "system", "developer":
-			out = append(out, generativeaiinference.SystemMessage{Content: textContents(m.Content.JoinText())})
-		case "assistant":
-			out = append(out, generativeaiinference.AssistantMessage{
-				Content:   textContents(m.Content.JoinText()),
-				ToolCalls: irToolCalls(m.ToolCalls),
-			})
-		case "tool":
-			tm := generativeaiinference.ToolMessage{Content: textContents(m.Content.JoinText())}
-			if m.ToolCallID != "" {
-				id := m.ToolCallID
-				tm.ToolCallId = &id
-			}
-			out = append(out, tm)
-		default: // user
-			out = append(out, generativeaiinference.UserMessage{Content: chatContents(m.Content)})
-		}
-	}
-	return out
-}
-
-// chatContents 把 IR 内容转 SDK 内容块;文本直通,image_url 转 IMAGE 块(data URI / 公网地址)。
-func chatContents(c aiwire.Content) []generativeaiinference.ChatContent {
-	if len(c.Parts) == 0 {
-		return textContents(c.Text)
-	}
-	var out []generativeaiinference.ChatContent
-	for _, p := range c.Parts {
-		switch {
-		case p.Type == "image_url" && p.ImageURL != nil:
-			img := generativeaiinference.ImageUrl{Url: &p.ImageURL.URL}
-			if d, ok := generativeaiinference.GetMappingImageUrlDetailEnum(p.ImageURL.Detail); ok {
-				img.Detail = d
-			}
-			out = append(out, generativeaiinference.ImageContent{ImageUrl: &img})
-		case p.Text != "":
-			t := p.Text
-			out = append(out, generativeaiinference.TextContent{Text: &t})
-		}
-	}
-	return out
-}
-
-// textContents 把纯文本包装为单元素 TextContent 数组;空文本返回 nil。
-func textContents(text string) []generativeaiinference.ChatContent {
-	if text == "" {
-		return nil
-	}
-	return []generativeaiinference.ChatContent{generativeaiinference.TextContent{Text: &text}}
-}
-
-func irToolCalls(calls []aiwire.ToolCall) []generativeaiinference.ToolCall {
-	if len(calls) == 0 {
-		return nil
-	}
-	out := make([]generativeaiinference.ToolCall, 0, len(calls))
-	for _, c := range calls {
-		id, name, args := c.ID, c.Function.Name, c.Function.Arguments
-		out = append(out, generativeaiinference.FunctionCall{Id: &id, Name: &name, Arguments: &args})
-	}
-	return out
-}
-
-func irTools(tools []aiwire.Tool) []generativeaiinference.ToolDefinition {
-	if len(tools) == 0 {
-		return nil
-	}
-	out := make([]generativeaiinference.ToolDefinition, 0, len(tools))
-	for _, t := range tools {
-		name, desc := t.Function.Name, t.Function.Description
-		fd := generativeaiinference.FunctionDefinition{Name: &name}
-		if desc != "" {
-			fd.Description = &desc
-		}
-		if len(t.Function.Parameters) > 0 {
-			var params interface{}
-			if json.Unmarshal(t.Function.Parameters, &params) == nil {
-				fd.Parameters = &params
-			}
-		}
-		out = append(out, fd)
-	}
-	return out
-}
-
-// irToolChoice 解析 "auto"/"none"/"required" 或 {type:function,function:{name}}。
-func irToolChoice(raw json.RawMessage) generativeaiinference.ToolChoice {
-	if len(raw) == 0 {
-		return nil
-	}
-	var s string
-	if json.Unmarshal(raw, &s) == nil {
-		switch s {
-		case "none":
-			return generativeaiinference.ToolChoiceNone{}
-		case "required":
-			return generativeaiinference.ToolChoiceRequired{}
-		case "auto":
-			return generativeaiinference.ToolChoiceAuto{}
-		}
-		return nil
-	}
-	var obj struct {
-		Function struct {
-			Name string `json:"name"`
-		} `json:"function"`
-	}
-	if json.Unmarshal(raw, &obj) == nil && obj.Function.Name != "" {
-		return generativeaiinference.ToolChoiceFunction{Name: &obj.Function.Name}
-	}
-	return nil
-}
-
-func irResponseFormat(rf *aiwire.ResponseFormat) generativeaiinference.ResponseFormat {
-	if rf == nil {
-		return nil
-	}
-	switch rf.Type {
-	case "json_object":
-		return generativeaiinference.JsonObjectResponseFormat{}
-	case "json_schema":
-		return irJSONSchemaFormat(rf.JSONSchema)
-	default:
-		return nil
-	}
-}
-
-// irJSONSchemaFormat 映射 OpenAI json_schema{name,description,schema,strict}。
-func irJSONSchemaFormat(raw json.RawMessage) generativeaiinference.ResponseFormat {
-	var in struct {
-		Name        string          `json:"name"`
-		Description string          `json:"description"`
-		Schema      json.RawMessage `json:"schema"`
-		Strict      *bool           `json:"strict"`
-	}
-	if json.Unmarshal(raw, &in) != nil || in.Name == "" {
-		return generativeaiinference.JsonObjectResponseFormat{}
-	}
-	js := generativeaiinference.ResponseJsonSchema{Name: &in.Name, IsStrict: in.Strict}
-	if in.Description != "" {
-		js.Description = &in.Description
-	}
-	if len(in.Schema) > 0 {
-		var schema interface{}
-		if json.Unmarshal(in.Schema, &schema) == nil {
-			js.Schema = &schema
-		}
-	}
-	return generativeaiinference.JsonSchemaResponseFormat{JsonSchema: &js}
-}
-
-// sdkToIR 把非流式 GenericChatResponse 转回 IR 响应;id/created 由调用方补。
-func sdkToIR(resp generativeaiinference.GenericChatResponse, model string) *aiwire.ChatResponse {
-	out := &aiwire.ChatResponse{Object: "chat.completion", Model: model}
-	if resp.TimeCreated != nil {
-		out.Created = resp.TimeCreated.Unix()
-	}
-	for i, ch := range resp.Choices {
-		idx := i
-		if ch.Index != nil {
-			idx = *ch.Index
-		}
-		out.Choices = append(out.Choices, aiwire.Choice{
-			Index:        idx,
-			Message:      sdkMessageToIR(ch.Message),
-			FinishReason: mapFinishReason(deref(ch.FinishReason)),
-		})
-	}
-	out.Usage = sdkUsageToIR(resp.Usage)
-	return out
-}
-
-func sdkUsageToIR(u *generativeaiinference.Usage) *aiwire.Usage {
-	if u == nil {
-		return nil
-	}
-	out := &aiwire.Usage{}
-	if u.PromptTokens != nil {
-		out.PromptTokens = *u.PromptTokens
-	}
-	if u.CompletionTokens != nil {
-		out.CompletionTokens = *u.CompletionTokens
-	}
-	if u.TotalTokens != nil {
-		out.TotalTokens = *u.TotalTokens
-	}
-	if d := u.PromptTokensDetails; d != nil && d.CachedTokens != nil && *d.CachedTokens > 0 {
-		out.PromptTokensDetails = &aiwire.PromptTokensDetails{CachedTokens: *d.CachedTokens}
-	}
-	return out
-}
-
-// sdkMessageToIR 提取助手消息文本与工具调用。
-func sdkMessageToIR(m generativeaiinference.Message) aiwire.ChatMessage {
-	out := aiwire.ChatMessage{Role: "assistant"}
-	am, ok := m.(generativeaiinference.AssistantMessage)
-	if !ok {
-		return out
-	}
-	var sb strings.Builder
-	for _, c := range am.Content {
-		if tc, ok := c.(generativeaiinference.TextContent); ok && tc.Text != nil {
-			sb.WriteString(*tc.Text)
-		}
-	}
-	out.Content = aiwire.NewTextContent(sb.String())
-	for _, call := range am.ToolCalls {
-		if fc, ok := call.(generativeaiinference.FunctionCall); ok {
-			out.ToolCalls = append(out.ToolCalls, aiwire.ToolCall{
-				ID:   deref(fc.Id),
-				Type: "function",
-				Function: aiwire.FunctionCall{
-					Name:      deref(fc.Name),
-					Arguments: deref(fc.Arguments),
-				},
-			})
-		}
-	}
-	return out
-}
-
-// mapFinishReason 归一化结束原因;OCI 与 OpenAI 取值同名,做小写与别名兜底。
-func mapFinishReason(reason string) string {
-	switch strings.ToLower(reason) {
-	case "", "null":
-		return ""
-	case "stop", "completed", "end_turn":
-		return "stop"
-	case "length", "max_tokens":
-		return "length"
-	case "tool_calls", "tool_call", "tool_use":
-		return "tool_calls"
-	case "complete":
-		return "stop"
-	case "error_toxic":
-		return "content_filter"
-	default:
-		return strings.ToLower(reason)
-	}
-}
-
-// genAiStreamEvent 是 GENERIC 流式事件的宽容解析结构(字段按增量 choice 形状)。
-type genAiStreamEvent struct {
-	Index   *int `json:"index"`
-	Message struct {
-		Role    string `json:"role"`
-		Content []struct {
-			Type string `json:"type"`
-			Text string `json:"text"`
-		} `json:"content"`
-		ToolCalls []struct {
-			ID        string `json:"id"`
-			Name      string `json:"name"`
-			Arguments string `json:"arguments"`
-		} `json:"toolCalls"`
-	} `json:"message"`
-	FinishReason *string                      `json:"finishReason"`
-	Usage        *generativeaiinference.Usage `json:"usage"`
-	Choices      []json.RawMessage            `json:"choices"` // 兼容整包 choices 形态
-	Text         string                       `json:"text"`    // COHERE 流事件的顶层增量文本
-	// CohereCalls 是 COHERE 流事件顶层的工具调用(与 GENERIC 的 message.toolCalls 不同层级)
-	CohereCalls []struct {
-		Name       string      `json:"name"`
-		Parameters interface{} `json:"parameters"`
-	} `json:"toolCalls"`
-}
-
-// parseGenAiEvent 把一条 SSE 事件 JSON 解析为 IR chunk;无有效负载时 ok=false。
-func parseGenAiEvent(data []byte, model string) (aiwire.ChatChunk, bool) {
-	var ev genAiStreamEvent
-	if err := json.Unmarshal(data, &ev); err != nil {
-		return aiwire.ChatChunk{}, false
-	}
-	if len(ev.Choices) > 0 { // choices 包裹形态:取首个展开重解析
-		var inner genAiStreamEvent
-		if json.Unmarshal(ev.Choices[0], &inner) == nil {
-			inner.Usage = ev.Usage
-			ev = inner
-		}
-	}
-	chunk := aiwire.ChatChunk{Object: "chat.completion.chunk", Model: model}
-	choice := aiwire.ChunkChoice{}
-	if ev.Index != nil {
-		choice.Index = *ev.Index
-	}
-	choice.Delta.Role = strings.ToLower(ev.Message.Role)
-	var sb strings.Builder
-	for _, c := range ev.Message.Content {
-		sb.WriteString(c.Text)
-	}
-	if sb.Len() == 0 && ev.Text != "" { // COHERE 事件形态
-		sb.WriteString(ev.Text)
-	}
-	choice.Delta.Content = sb.String()
-	for i, tc := range ev.Message.ToolCalls {
-		d := aiwire.ToolCallDelta{Index: i, ID: tc.ID}
-		if tc.ID != "" {
-			d.Type = "function"
-		}
-		d.Function.Name = tc.Name
-		d.Function.Arguments = tc.Arguments
-		choice.Delta.ToolCalls = append(choice.Delta.ToolCalls, d)
-	}
-	for i, tc := range ev.CohereCalls { // COHERE 顶层工具调用:整包出现,补派生 ID
-		call := cohereCallToIR(tc.Name, &tc.Parameters, i)
-		d := aiwire.ToolCallDelta{Index: len(choice.Delta.ToolCalls) + i, ID: call.ID, Type: "function"}
-		d.Function.Name = call.Function.Name
-		d.Function.Arguments = call.Function.Arguments
-		choice.Delta.ToolCalls = append(choice.Delta.ToolCalls, d)
-	}
-	if r := mapFinishReason(deref(ev.FinishReason)); r != "" {
-		choice.FinishReason = &r
-	}
-	hasPayload := choice.Delta.Content != "" || len(choice.Delta.ToolCalls) > 0 || choice.FinishReason != nil
-	if hasPayload {
-		chunk.Choices = []aiwire.ChunkChoice{choice}
-	}
-	chunk.Usage = sdkUsageToIR(ev.Usage)
-	return chunk, hasPayload || chunk.Usage != nil
-}

internal/oci/genai_effort_integration_test.go

@@ -0,0 +1,496 @@
+package oci
+
+// 真实 OCI GenAI effort 探针(重建版)。
+//
+// 前身 genai_cache_integration_test.go 承载缓存/亲和/服务端工具等历轮调研模式,
+// 调研结论均已归档(.trellis/tasks/archive/2026-07/*/research/),该文件已删除;
+// 本文件重建共享基础设施,保留 effort 支持矩阵探测,并新增 responses 端点探测
+// (multi-agent 模型仅允许走 Responses 面)。
+//
+// 运行(默认跳过,不触网):
+//
+//	OCI_GENAI_CACHE_PROBE=effort-matrix OCI_CACHE_DB=<db 绝对路径> DATA_KEY=<主密钥> \
+//	OCI_CACHE_RUN_ID=<唯一标识> go test ./internal/oci/ -run TestRealGenAiEffortMatrix -v
+//
+// 可选:OCI_CACHE_CONFIG_ID(默认 6)、OCI_CACHE_REGION(默认 us-chicago-1)、
+// OCI_EFFORT_CASES 自定义用例("model=effort" 逗号分隔;effort 后缀 "@responses"
+// 表示走 /actions/v1/responses 端点,如 "xai.grok-4.20-multi-agent=low@responses")。
+//
+// 纪律:数据库只读打开;OCI_GO_SDK_DEBUG 必须为空;日志不落原始请求体与凭据,
+// 错误消息经 OCID 遮掩后截断。
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"os"
+	"regexp"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/glebarez/sqlite"
+	"github.com/oracle/oci-go-sdk/v65/common"
+	"github.com/oracle/oci-go-sdk/v65/generativeaiinference"
+	"gorm.io/gorm"
+	"gorm.io/gorm/logger"
+
+	projectcrypto "oci-portal/internal/crypto"
+	"oci-portal/internal/model"
+)
+
+const (
+	effortProbeRegion = "us-chicago-1"
+	effortProbeLimit  = int64(4 << 20)
+)
+
+var effortRunIDPattern = regexp.MustCompile(`^[A-Za-z0-9._-]{1,64}$`)
+
+type effortProbeCase struct {
+	model     string
+	effort    string
+	responses bool // true 走 /actions/v1/responses,否则 typed /actions/chat
+	stream    bool // 仅 responses 面:请求 SSE 流,观测事件类型序列
+}
+
+type effortProbeResources struct {
+	cred      Credentials
+	region    string
+	modelOCID map[string]string
+	inference generativeaiinference.GenerativeAiInferenceClient
+}
+
+type effortProbeResult struct {
+	status           int
+	requestRef       string
+	duration         time.Duration
+	errorCode        string
+	errorMessage     string
+	completionTokens int
+	reasoningTokens  int
+	answerLen        int
+	outputTypes      []string
+}
+
+// effortMatrixCases 默认批为 grok-4.3 五档基线;全模型矩阵结论见任务归档,
+// 复测用 OCI_EFFORT_CASES 自定义。
+func effortMatrixCases() []effortProbeCase {
+	if raw := os.Getenv("OCI_EFFORT_CASES"); raw != "" {
+		return parseEffortCases(raw)
+	}
+	return []effortProbeCase{
+		{model: "xai.grok-4.3", effort: "none", responses: true},
+		{model: "xai.grok-4.3", effort: "low", responses: true},
+		{model: "xai.grok-4.3", effort: "high", responses: true},
+	}
+}
+
+// parseEffortCases 解析 "model=effort[@responses|@responses-stream]" 逗号分隔用例。
+func parseEffortCases(raw string) []effortProbeCase {
+	var out []effortProbeCase
+	for _, item := range strings.Split(raw, ",") {
+		parts := strings.SplitN(strings.TrimSpace(item), "=", 2)
+		if len(parts) != 2 || parts[0] == "" || parts[1] == "" {
+			continue
+		}
+		effort, streaming := strings.CutSuffix(parts[1], "@responses-stream")
+		if streaming {
+			out = append(out, effortProbeCase{model: parts[0], effort: effort, responses: true, stream: true})
+			continue
+		}
+		effort, _ = strings.CutSuffix(effort, "@responses") // 后缀兼容保留;所有用例均走 responses 面
+		out = append(out, effortProbeCase{model: parts[0], effort: effort, responses: true})
+	}
+	return out
+}
+
+func TestRealGenAiEffortMatrix(t *testing.T) {
+	requireEffortProbeMode(t, "effort-matrix")
+	ctx, cancel := context.WithTimeout(context.Background(), 20*time.Minute)
+	defer cancel()
+	resources := newEffortProbeResources(t, ctx)
+	for _, tc := range effortMatrixCases() {
+		runEffortCase(t, ctx, resources, tc)
+	}
+}
+
+func requireEffortProbeMode(t *testing.T, wanted string) {
+	t.Helper()
+	if os.Getenv("OCI_GENAI_CACHE_PROBE") != wanted {
+		t.Skipf("set OCI_GENAI_CACHE_PROBE=%s to run", wanted)
+	}
+	if os.Getenv("OCI_GO_SDK_DEBUG") != "" {
+		t.Fatal("OCI_GO_SDK_DEBUG must be unset to protect signed requests")
+	}
+	if !effortRunIDPattern.MatchString(os.Getenv("OCI_CACHE_RUN_ID")) {
+		t.Fatal("set a unique OCI_CACHE_RUN_ID (1-64 safe characters)")
+	}
+}
+
+func runEffortCase(t *testing.T, ctx context.Context, resources effortProbeResources, tc effortProbeCase) {
+	t.Helper()
+	ocid, ok := resources.modelOCID[tc.model]
+	if !ok {
+		t.Logf("model=%s effort=%s resolve-failed: model not in on-demand catalog", tc.model, tc.effort)
+		return
+	}
+	_ = ocid
+	body, path, err := effortProbeBody(tc)
+	if err != nil {
+		t.Fatalf("build effort body: %v", err)
+	}
+	result := executeEffortProbe(ctx, resources, path, body)
+	logEffortResult(t, tc, result)
+}
+
+// effortProbeBody 构造 OpenAI Responses 直通请求体(typed chat 面已随网关剔除,
+// 探针仅保留 responses 端点;历史 typed 结论见任务归档)。
+func effortProbeBody(tc effortProbeCase) ([]byte, string, error) {
+	question := "How many positive divisors does 360 have? Answer with the number only."
+	body := map[string]interface{}{"model": tc.model, "input": question,
+		"max_output_tokens": 2048, "store": false}
+	if tc.effort != "" {
+		body["reasoning"] = map[string]string{"effort": tc.effort}
+	}
+	if tc.stream {
+		body["stream"] = true
+	}
+	payload, err := json.Marshal(body)
+	return payload, "/actions/v1/responses", err
+}
+
+// executeEffortProbe 以 IAM 签名裸 POST 指定路径;responses 面按生产实现补 compartment 头。
+func executeEffortProbe(ctx context.Context, resources effortProbeResources, path string, body []byte) effortProbeResult {
+	client := resources.inference.BaseClient
+	client.Configuration.CircuitBreaker = nil
+	common.UpdateEndpointTemplateForOptions(&client)
+	common.SetMissingTemplateParams(&client)
+	request, err := http.NewRequestWithContext(ctx, http.MethodPost, path, strings.NewReader(string(body)))
+	if err != nil {
+		return effortProbeResult{errorCode: "BuildRequest", errorMessage: err.Error()}
+	}
+	request.Header.Set("Content-Type", "application/json")
+	if path != "/actions/chat" {
+		request.Header.Set("CompartmentId", resources.cred.TenancyOCID)
+		request.Header.Set("opc-compartment-id", resources.cred.TenancyOCID)
+	}
+	return effortCallProbe(ctx, client, request)
+}
+
+func effortCallProbe(ctx context.Context, client common.BaseClient, request *http.Request) effortProbeResult {
+	started := time.Now()
+	response, err := client.Call(ctx, request)
+	result := effortProbeResult{duration: time.Since(started)}
+	if response != nil {
+		result.status = response.StatusCode
+		result.requestRef = shortEffortRef(response.Header.Get("opc-request-id"))
+		defer response.Body.Close()
+	}
+	if err != nil {
+		result.errorCode, result.errorMessage = effortProbeError(err)
+		return result
+	}
+	payload, readErr := io.ReadAll(io.LimitReader(response.Body, effortProbeLimit))
+	if readErr != nil {
+		result.errorCode, result.errorMessage = "ReadResponse", readErr.Error()
+		return result
+	}
+	parseEffortPayload(&result, payload)
+	return result
+}
+
+func effortProbeError(err error) (string, string) {
+	if serviceErr, ok := common.IsServiceError(err); ok {
+		return fmt.Sprintf("%d", serviceErr.GetHTTPStatusCode()), sanitizeEffortText(serviceErr.GetMessage())
+	}
+	return "CallError", sanitizeEffortText(err.Error())
+}
+
+// parseEffortPayload 兼容三种 usage 形态:typed 驼峰(chatResponse.usage.completionTokens)、
+// chat 兼容面下划线(usage.completion_tokens)、responses 面(usage.output_tokens)。
+func parseEffortPayload(result *effortProbeResult, payload []byte) {
+	if text := strings.TrimSpace(string(payload)); strings.HasPrefix(text, "event:") || strings.HasPrefix(text, "data:") {
+		parseEffortSSE(result, text)
+		return
+	}
+	var root map[string]interface{}
+	if json.Unmarshal(payload, &root) != nil {
+		return
+	}
+	result.answerLen = len(parseEffortAnswer(root))
+	result.outputTypes = parseEffortOutputTypes(root)
+	if typed, ok := root["chatResponse"].(map[string]interface{}); ok {
+		root = typed
+	}
+	usage, _ := root["usage"].(map[string]interface{})
+	for _, key := range []string{"completionTokens", "completion_tokens", "output_tokens"} {
+		if v, ok := effortInt(usage, key); ok {
+			result.completionTokens = v
+			break
+		}
+	}
+	for _, detailsKey := range []string{"completionTokensDetails", "completion_tokens_details", "output_tokens_details"} {
+		details, _ := usage[detailsKey].(map[string]interface{})
+		for _, key := range []string{"reasoningTokens", "reasoning_tokens"} {
+			if v, ok := effortInt(details, key); ok {
+				result.reasoningTokens = v
+			}
+		}
+	}
+}
+
+// parseEffortAnswer 提取正文文本:typed 取 chatResponse.choices[].message.content[].text,
+// responses 取 output[] 里 message 项的 content[].text。仅在内存中量长度,不落日志。
+func parseEffortAnswer(root map[string]interface{}) string {
+	var sb strings.Builder
+	if typed, ok := root["chatResponse"].(map[string]interface{}); ok {
+		choices, _ := typed["choices"].([]interface{})
+		for _, c := range choices {
+			cm, _ := c.(map[string]interface{})
+			msg, _ := cm["message"].(map[string]interface{})
+			collectEffortText(&sb, msg["content"])
+		}
+		return sb.String()
+	}
+	output, _ := root["output"].([]interface{})
+	for _, item := range output {
+		im, _ := item.(map[string]interface{})
+		if im["type"] == "message" {
+			collectEffortText(&sb, im["content"])
+		}
+	}
+	return sb.String()
+}
+
+func collectEffortText(sb *strings.Builder, content interface{}) {
+	parts, _ := content.([]interface{})
+	for _, p := range parts {
+		pm, _ := p.(map[string]interface{})
+		if text, ok := pm["text"].(string); ok {
+			sb.WriteString(text)
+		}
+	}
+}
+
+// parseEffortSSE 汇总 SSE 流:事件类型去重序列进 outputTypes,正文增量长度进 answerLen,
+// 末尾 completed 事件里的 usage 进 tokens 字段。
+func parseEffortSSE(result *effortProbeResult, text string) {
+	seen := map[string]bool{}
+	for _, line := range strings.Split(text, "\n") {
+		if name, ok := strings.CutPrefix(line, "event: "); ok {
+			if name = strings.TrimSpace(name); !seen[name] {
+				seen[name] = true
+				result.outputTypes = append(result.outputTypes, name)
+			}
+			continue
+		}
+		data, ok := strings.CutPrefix(line, "data: ")
+		if !ok {
+			continue
+		}
+		var ev map[string]interface{}
+		if json.Unmarshal([]byte(data), &ev) != nil {
+			continue
+		}
+		if name, ok := ev["type"].(string); ok && !seen[name] { // 纯 data: 行流(无 event: 行)从事件体取类型
+			seen[name] = true
+			result.outputTypes = append(result.outputTypes, name)
+		}
+		if delta, ok := ev["delta"].(string); ok && strings.Contains(fmt.Sprint(ev["type"]), "output_text") {
+			result.answerLen += len(delta)
+		}
+		if resp, ok := ev["response"].(map[string]interface{}); ok {
+			usage, _ := resp["usage"].(map[string]interface{})
+			if v, ok := effortInt(usage, "output_tokens"); ok {
+				result.completionTokens = v
+			}
+		}
+	}
+}
+
+// parseEffortOutputTypes 收集 responses 面 output 项类型序列(multi-agent 行为观测)。
+func parseEffortOutputTypes(root map[string]interface{}) []string {
+	output, _ := root["output"].([]interface{})
+	var types []string
+	for _, item := range output {
+		im, _ := item.(map[string]interface{})
+		if s, ok := im["type"].(string); ok {
+			types = append(types, s)
+		}
+	}
+	return types
+}
+
+func effortInt(root map[string]interface{}, key string) (int, bool) {
+	if root == nil {
+		return 0, false
+	}
+	if value, ok := root[key].(float64); ok {
+		return int(value), true
+	}
+	return 0, false
+}
+
+func logEffortResult(t *testing.T, tc effortProbeCase, result effortProbeResult) {
+	t.Helper()
+	endpoint := "typed-chat"
+	if tc.responses {
+		endpoint = "compat-responses"
+	}
+	t.Logf("endpoint=%s model=%s effort=%s status=%d completion=%d reasoning=%d answer_len=%d output_types=%v request=%s duration_ms=%d error=%s message=%q",
+		endpoint, tc.model, tc.effort, result.status, result.completionTokens, result.reasoningTokens,
+		result.answerLen, result.outputTypes, result.requestRef, result.duration.Milliseconds(),
+		result.errorCode, result.errorMessage)
+}
+
+// sanitizeEffortText 遮掩 OCID 并截断,避免错误消息携带租户可定位信息。
+func sanitizeEffortText(text string) string {
+	text = regexp.MustCompile(`ocid1\.[a-z0-9._-]+`).ReplaceAllString(text, "ocid1.***")
+	if len(text) > 300 {
+		text = text[:300] + "..."
+	}
+	return text
+}
+
+func shortEffortRef(ref string) string {
+	if len(ref) > 12 {
+		return ref[:12]
+	}
+	return ref
+}
+
+func effortEnvOr(key, fallback string) string {
+	if value := os.Getenv(key); value != "" {
+		return value
+	}
+	return fallback
+}
+
+// newEffortProbeResources 装配凭据、区域、inference 客户端与按需模型目录。
+func newEffortProbeResources(t *testing.T, ctx context.Context) effortProbeResources {
+	t.Helper()
+	cred := effortProbeCredentials(t)
+	region := effortEnvOr("OCI_CACHE_REGION", effortProbeRegion)
+	cred.Region = region
+	real := &RealClient{}
+	ic, err := real.genAiInferenceClient(cred, region)
+	if err != nil {
+		t.Fatalf("new inference client: %v", err)
+	}
+	models, err := real.ListGenAiModels(ctx, cred, region)
+	if err != nil {
+		t.Fatalf("list models: %s", sanitizeEffortText(err.Error()))
+	}
+	index := make(map[string]string, len(models))
+	for _, m := range models {
+		index[m.Name] = m.Ocid
+	}
+	return effortProbeResources{cred: cred, region: region, modelOCID: index, inference: ic}
+}
+
+// effortProbeCredentials 优先从面板数据库(只读)取渠道凭据,否则退回本地 ini 测试凭据。
+func effortProbeCredentials(t *testing.T) Credentials {
+	t.Helper()
+	if dbPath := os.Getenv("OCI_CACHE_DB"); dbPath != "" {
+		return effortDBCredentials(t, dbPath)
+	}
+	return loadTestCredentials(t, effortEnvOr("OCI_TEST_KEY", "试用期"))
+}
+
+func effortDBCredentials(t *testing.T, dbPath string) Credentials {
+	t.Helper()
+	dsn := fmt.Sprintf("file:%s?mode=ro", dbPath)
+	db, err := gorm.Open(sqlite.Open(dsn), &gorm.Config{Logger: logger.Default.LogMode(logger.Silent)})
+	if err != nil {
+		t.Fatalf("open read-only probe database: %v", err)
+	}
+	cipher, err := projectcrypto.NewCipher(os.Getenv("DATA_KEY"))
+	if err != nil {
+		t.Fatalf("create data cipher: %v", err)
+	}
+	var config model.OciConfig
+	if err := db.First(&config, effortEnvOr("OCI_CACHE_CONFIG_ID", "6")).Error; err != nil {
+		t.Fatalf("load OCI config: %v", err)
+	}
+	return decryptEffortCredentials(t, db, cipher, config)
+}
+
+func decryptEffortCredentials(t *testing.T, db *gorm.DB, cipher *projectcrypto.Cipher, config model.OciConfig) Credentials {
+	t.Helper()
+	privateKey, err := cipher.DecryptString(config.PrivateKeyEnc)
+	if err != nil {
+		t.Fatalf("decrypt OCI config %d private key: %v", config.ID, err)
+	}
+	passphrase := ""
+	if config.PassphraseEnc != "" {
+		if passphrase, err = cipher.DecryptString(config.PassphraseEnc); err != nil {
+			t.Fatalf("decrypt OCI config %d passphrase: %v", config.ID, err)
+		}
+	}
+	return Credentials{TenancyOCID: config.TenancyOCID, UserOCID: config.UserOCID,
+		Fingerprint: config.Fingerprint, Region: config.Region, PrivateKey: privateKey,
+		Passphrase: passphrase, Proxy: loadEffortProxy(t, db, cipher, config.ProxyID)}
+}
+
+func loadEffortProxy(t *testing.T, db *gorm.DB, cipher *projectcrypto.Cipher, proxyID *uint) *ProxySpec {
+	t.Helper()
+	if proxyID == nil {
+		return nil
+	}
+	var proxy model.Proxy
+	if err := db.First(&proxy, *proxyID).Error; err != nil {
+		t.Fatalf("load proxy %d: %v", *proxyID, err)
+	}
+	password := ""
+	if proxy.PasswordEnc != "" {
+		decrypted, err := cipher.DecryptString(proxy.PasswordEnc)
+		if err != nil {
+			t.Fatalf("decrypt proxy %d password: %v", *proxyID, err)
+		}
+		password = decrypted
+	}
+	return &ProxySpec{Type: proxy.Type, Host: proxy.Host, Port: proxy.Port,
+		Username: proxy.Username, Password: password}
+}
+
+// TestParseEffortCases 断言自定义用例解析:端点后缀、空项与畸形项跳过。
+func TestParseEffortCases(t *testing.T) {
+	got := parseEffortCases("a=low, b=high@responses ,bad,=x,c=")
+	want := []effortProbeCase{{model: "a", effort: "low", responses: true}, {model: "b", effort: "high", responses: true}}
+	if len(got) != len(want) {
+		t.Fatalf("parseEffortCases = %+v, want %+v", got, want)
+	}
+	for i := range want {
+		if got[i] != want[i] {
+			t.Fatalf("case %d = %+v, want %+v", i, got[i], want[i])
+		}
+	}
+}
+
+// TestParseEffortPayload 断言三种 usage 形态与 output 类型序列的解析。
+func TestParseEffortPayload(t *testing.T) {
+	cases := []struct {
+		name       string
+		payload    string
+		completion int
+		reasoning  int
+		types      int
+	}{
+		{"typed 驼峰", `{"chatResponse":{"choices":[{"message":{"content":[{"type":"TEXT","text":"24"}]}}],"usage":{"completionTokens":5,"completionTokensDetails":{"reasoningTokens":9}}}}`, 5, 9, 0},
+		{"responses 面", `{"output":[{"type":"reasoning"},{"type":"message","content":[{"type":"output_text","text":"24"}]}],"usage":{"output_tokens":7,"output_tokens_details":{"reasoning_tokens":3}}}`, 7, 3, 2},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			var result effortProbeResult
+			parseEffortPayload(&result, []byte(tc.payload))
+			if result.completionTokens != tc.completion || result.reasoningTokens != tc.reasoning || len(result.outputTypes) != tc.types {
+				t.Fatalf("parse = %+v, want completion=%d reasoning=%d types=%d", result, tc.completion, tc.reasoning, tc.types)
+			}
+			if result.answerLen != 2 {
+				t.Fatalf("answerLen = %d, want 2", result.answerLen)
+			}
+		})
+	}
+}

internal/oci/genai_responses.go

@@ -0,0 +1,73 @@
+package oci
+
+import (
+	"bytes"
+	"context"
+	"fmt"
+	"io"
+	"net/http"
+
+	"github.com/oracle/oci-go-sdk/v65/common"
+)
+
+// compatResponsesLimit 限制直通响应体大小;web_search 输出含多段引用,给足余量。
+const compatResponsesLimit = int64(8 << 20)
+
+// GenAiCompatResponses 实现 Client:把 OpenAI Responses 请求体直通到 OCI
+// `/20231130/actions/v1/responses`(IAM 签名)。该端点实测可执行 xAI 服务端工具
+// (web_search / x_search),但不在 Oracle 文档化工具白名单内,行为可能随服务
+// 版本、模型或区域变化;调用方须自行校验并改写请求体(store/stream)。
+func (c *RealClient) GenAiCompatResponses(ctx context.Context, cred Credentials, region string, body []byte) ([]byte, error) {
+	ic, err := c.genAiInferenceClient(cred, region)
+	if err != nil {
+		return nil, err
+	}
+	client := ic.BaseClient
+	common.UpdateEndpointTemplateForOptions(&client)
+	common.SetMissingTemplateParams(&client)
+	request, err := http.NewRequestWithContext(ctx, http.MethodPost, "/actions/v1/responses", bytes.NewReader(body))
+	if err != nil {
+		return nil, fmt.Errorf("build compat responses request: %w", err)
+	}
+	request.Header.Set("Content-Type", "application/json")
+	request.Header.Set("CompartmentId", cred.TenancyOCID)
+	request.Header.Set("opc-compartment-id", cred.TenancyOCID)
+	response, err := client.Call(ctx, request)
+	if err != nil {
+		return nil, err
+	}
+	defer response.Body.Close()
+	payload, err := io.ReadAll(io.LimitReader(response.Body, compatResponsesLimit))
+	if err != nil {
+		return nil, fmt.Errorf("read compat responses body: %w", err)
+	}
+	return payload, nil
+}
+
+// GenAiCompatResponsesStream 实现 Client:以流式直通 OCI `/actions/v1/responses`,
+// 建立成功(2xx)返回 SSE body(调用方负责 Close);建立失败返回 SDK ServiceError,
+// 与既有渠道切换/熔断错误分类兼容。请求体须由调用方置 stream:true。
+func (c *RealClient) GenAiCompatResponsesStream(ctx context.Context, cred Credentials, region string, body []byte) (io.ReadCloser, error) {
+	ic, err := c.genAiInferenceClient(cred, region)
+	if err != nil {
+		return nil, err
+	}
+	client := ic.BaseClient
+	common.UpdateEndpointTemplateForOptions(&client)
+	common.SetMissingTemplateParams(&client)
+	request, err := http.NewRequestWithContext(ctx, http.MethodPost, "/actions/v1/responses", bytes.NewReader(body))
+	if err != nil {
+		return nil, fmt.Errorf("build compat responses stream request: %w", err)
+	}
+	request.Header.Set("Content-Type", "application/json")
+	request.Header.Set("CompartmentId", cred.TenancyOCID)
+	request.Header.Set("opc-compartment-id", cred.TenancyOCID)
+	response, err := client.Call(ctx, request)
+	if err != nil {
+		if response != nil && response.Body != nil {
+			response.Body.Close()
+		}
+		return nil, err
+	}
+	return response.Body, nil
+}

internal/oci/genai_test.go

@@ -1,193 +1,27 @@
 package oci
 
 import (
-	"encoding/json"
 	"testing"
 
 	"github.com/oracle/oci-go-sdk/v65/generativeaiinference"
-
-	"oci-portal/internal/aiwire"
 )
 
-// TestIrToSDK 断言 IR→GENERIC 的角色拆装、采样参数与工具映射。
+// TestSdkUsageCachedTokens 断言 embed 用量映射:缓存命中挂 details,仅命中 >0 时出现。
-func TestIrToSDK(t *testing.T) {
-	temp, mt := 0.7, 100
-	ir := aiwire.ChatRequest{
-		Model: "meta.llama-3.3-70b-instruct",
-		Messages: []aiwire.ChatMessage{
-			{Role: "system", Content: aiwire.NewTextContent("你是助手")},
-			{Role: "user", Content: aiwire.NewTextContent("你好")},
-			{Role: "assistant", ToolCalls: []aiwire.ToolCall{{ID: "c1", Type: "function", Function: aiwire.FunctionCall{Name: "get_weather", Arguments: `{"city":"东京"}`}}}},
-			{Role: "tool", ToolCallID: "c1", Content: aiwire.NewTextContent("晴")},
-		},
-		Temperature: &temp,
-		MaxTokens:   &mt,
-		Stop:        aiwire.StringList{"END"},
-		Tools:       []aiwire.Tool{{Type: "function", Function: aiwire.FunctionDef{Name: "get_weather", Parameters: json.RawMessage(`{"type":"object"}`)}}},
-		ToolChoice:  json.RawMessage(`"auto"`),
-	}
-	req := irToSDK(ir, true)
-	if len(req.Messages) != 4 {
-		t.Fatalf("messages = %d, want 4", len(req.Messages))
-	}
-	if _, ok := req.Messages[0].(generativeaiinference.SystemMessage); !ok {
-		t.Fatalf("messages[0] = %T, want SystemMessage", req.Messages[0])
-	}
-	am, ok := req.Messages[2].(generativeaiinference.AssistantMessage)
-	if !ok || len(am.ToolCalls) != 1 {
-		t.Fatalf("assistant tool calls 未映射: %T %+v", req.Messages[2], am)
-	}
-	tm, ok := req.Messages[3].(generativeaiinference.ToolMessage)
-	if !ok || deref(tm.ToolCallId) != "c1" {
-		t.Fatalf("tool message 未映射 toolCallId: %+v", tm)
-	}
-	if req.Temperature == nil || *req.Temperature != 0.7 {
-		t.Fatalf("temperature 未直通")
-	}
-	if req.MaxTokens == nil || *req.MaxTokens != 100 || len(req.Stop) != 1 {
-		t.Fatalf("maxTokens/stop 未直通")
-	}
-	if _, ok := req.ToolChoice.(generativeaiinference.ToolChoiceAuto); !ok {
-		t.Fatalf("toolChoice = %T, want auto", req.ToolChoice)
-	}
-	if len(req.Tools) != 1 || req.IsStream == nil || !*req.IsStream || req.StreamOptions == nil {
-		t.Fatalf("tools/stream 选项未映射")
-	}
-}
-
-// TestSdkToIR 断言响应文本、工具调用与用量的反向映射。
-func TestSdkToIR(t *testing.T) {
-	text, fr := "东京晴", "tool_calls"
-	idx, pt, ct, tt := 0, 10, 5, 15
-	id, name, args := "c1", "get_weather", `{"city":"东京"}`
-	resp := generativeaiinference.GenericChatResponse{
-		Choices: []generativeaiinference.ChatChoice{{
-			Index: &idx,
-			Message: generativeaiinference.AssistantMessage{
-				Content:   []generativeaiinference.ChatContent{generativeaiinference.TextContent{Text: &text}},
-				ToolCalls: []generativeaiinference.ToolCall{generativeaiinference.FunctionCall{Id: &id, Name: &name, Arguments: &args}},
-			},
-			FinishReason: &fr,
-		}},
-		Usage: &generativeaiinference.Usage{PromptTokens: &pt, CompletionTokens: &ct, TotalTokens: &tt},
-	}
-	out := sdkToIR(resp, "m1")
-	if len(out.Choices) != 1 || out.Choices[0].Message.Content.JoinText() != "东京晴" {
-		t.Fatalf("文本未映射: %+v", out)
-	}
-	if out.Choices[0].FinishReason != "tool_calls" || len(out.Choices[0].Message.ToolCalls) != 1 {
-		t.Fatalf("finishReason/toolCalls 未映射: %+v", out.Choices[0])
-	}
-	if out.Usage == nil || out.Usage.TotalTokens != 15 {
-		t.Fatalf("usage 未映射: %+v", out.Usage)
-	}
-}
-
-// TestParseGenAiEvent 断言流事件宽容解析:文本增量、结束原因与 usage 事件。
-func TestParseGenAiEvent(t *testing.T) {
-	chunk, ok := parseGenAiEvent([]byte(`{"index":0,"message":{"role":"ASSISTANT","content":[{"type":"TEXT","text":"你"}]}}`), "m1")
-	if !ok || len(chunk.Choices) != 1 || chunk.Choices[0].Delta.Content != "你" {
-		t.Fatalf("文本增量解析失败: %+v", chunk)
-	}
-	chunk, ok = parseGenAiEvent([]byte(`{"finishReason":"stop","usage":{"promptTokens":3,"completionTokens":2,"totalTokens":5}}`), "m1")
-	if !ok || chunk.Choices[0].FinishReason == nil || *chunk.Choices[0].FinishReason != "stop" {
-		t.Fatalf("finishReason 解析失败: %+v", chunk)
-	}
-	if chunk.Usage == nil || chunk.Usage.TotalTokens != 5 {
-		t.Fatalf("usage 解析失败: %+v", chunk.Usage)
-	}
-	if _, ok := parseGenAiEvent([]byte(`{}`), "m1"); ok {
-		t.Fatal("空事件应返回 ok=false")
-	}
-}
-
 func TestSdkUsageCachedTokens(t *testing.T) {
 	p, c, tot, cached := 10, 5, 15, 8
 	u := sdkUsageToIR(&generativeaiinference.Usage{PromptTokens: &p, CompletionTokens: &c, TotalTokens: &tot,
 		PromptTokensDetails: &generativeaiinference.PromptTokensDetails{CachedTokens: &cached}})
-	if u.CachedTokens() != 8 {
+	if u.PromptTokens != 10 || u.CompletionTokens != 5 || u.TotalTokens != 15 {
-		t.Errorf("CachedTokens = %d, want 8", u.CachedTokens())
+		t.Fatalf("usage 映射错误: %+v", u)
 	}
-	if sdkUsageToIR(&generativeaiinference.Usage{PromptTokens: &p}).CachedTokens() != 0 {
+	if u.PromptTokensDetails == nil || u.PromptTokensDetails.CachedTokens != 8 {
-		t.Error("无细分时 CachedTokens 应为 0")
+		t.Fatalf("cachedTokens 未映射: %+v", u.PromptTokensDetails)
 	}
+	zero := 0
+	if got := sdkUsageToIR(&generativeaiinference.Usage{PromptTokens: &p, PromptTokensDetails: &generativeaiinference.PromptTokensDetails{CachedTokens: &zero}}); got.PromptTokensDetails != nil {
+		t.Fatalf("零命中不应带 details: %+v", got.PromptTokensDetails)
 	}
-
+	if sdkUsageToIR(nil) != nil {
-func TestIrToCohereSDK(t *testing.T) {
+		t.Fatal("nil usage 应返回 nil")
-	ir := aiwire.ChatRequest{
-		Model: "cohere.command-r-plus",
-		Messages: []aiwire.ChatMessage{
-			{Role: "system", Content: aiwire.NewTextContent("你是助手")},
-			{Role: "user", Content: aiwire.NewTextContent("第一问")},
-			{Role: "assistant", Content: aiwire.NewTextContent("第一答")},
-			{Role: "user", Content: aiwire.NewTextContent("第二问")},
-		},
-	}
-	req, err := irToCohereSDK(ir, true)
-	if err != nil {
-		t.Fatalf("irToCohereSDK: %v", err)
-	}
-	if *req.Message != "第二问" || *req.PreambleOverride != "你是助手" || len(req.ChatHistory) != 2 {
-		t.Errorf("拆装错误: msg=%q preamble=%v history=%d", *req.Message, req.PreambleOverride, len(req.ChatHistory))
-	}
-	if _, ok := req.ChatHistory[0].(generativeaiinference.CohereUserMessage); !ok {
-		t.Errorf("history[0] 应为 user: %T", req.ChatHistory[0])
-	}
-	if _, ok := req.ChatHistory[1].(generativeaiinference.CohereChatBotMessage); !ok {
-		t.Errorf("history[1] 应为 chatbot: %T", req.ChatHistory[1])
-	}
-	if req.IsStream == nil || !*req.IsStream {
-		t.Error("IsStream 未设置")
-	}
-	if req.StreamOptions == nil || req.StreamOptions.IsIncludeUsage == nil || !*req.StreamOptions.IsIncludeUsage {
-		t.Error("流式应默认开启 usage 回传(StreamOptions.IsIncludeUsage)")
-	}
-	// 工具定义降级为扁平参数表(仅取 JSON Schema 顶层 properties)
-	ir.Tools = []aiwire.Tool{{Type: "function", Function: aiwire.FunctionDef{
-		Name: "get_weather", Parameters: json.RawMessage(`{"type":"object","properties":{"city":{"type":"string","description":"城市"}},"required":["city"]}`)}}}
-	req2, err := irToCohereSDK(ir, false)
-	if err != nil {
-		t.Fatalf("工具请求应支持: %v", err)
-	}
-	tool := req2.Tools[0]
-	if *tool.Name != "get_weather" || *tool.Description != "get_weather" {
-		t.Errorf("tool = %+v", tool)
-	}
-	if pd, ok := tool.ParameterDefinitions["city"]; !ok || *pd.Type != "string" || pd.IsRequired == nil || !*pd.IsRequired {
-		t.Errorf("param city = %+v", pd)
-	}
-	// 图片输入仍拒绝
-	ir.Tools = nil
-	ir.Messages = append(ir.Messages, aiwire.ChatMessage{Role: "user", Content: aiwire.NewPartsContent([]aiwire.ContentPart{
-		{Type: "image_url", ImageURL: &aiwire.ImageURL{URL: "data:image/png;base64,xx"}}})})
-	if _, err := irToCohereSDK(ir, false); err == nil {
-		t.Error("图片输入应被拒绝")
-	}
-	// 无 user 消息被拒
-	if _, err := irToCohereSDK(aiwire.ChatRequest{Model: "cohere.x", Messages: []aiwire.ChatMessage{{Role: "system", Content: aiwire.NewTextContent("s")}}}, false); err == nil {
-		t.Error("无 user 消息应被拒绝")
-	}
-}
-
-func TestCohereSDKToIR(t *testing.T) {
-	text := "回答"
-	resp := generativeaiinference.CohereChatResponse{
-		Text:         &text,
-		FinishReason: generativeaiinference.CohereChatResponseFinishReasonComplete,
-	}
-	out := cohereSDKToIR(resp, "cohere.command-r-plus")
-	if out.Choices[0].Message.Content.JoinText() != "回答" || out.Choices[0].FinishReason != "stop" {
-		t.Errorf("cohereSDKToIR = %+v", out.Choices[0])
-	}
-}
-
-func TestParseGenAiEventCohere(t *testing.T) {
-	chunk, ok := parseGenAiEvent([]byte(`{"apiFormat":"COHERE","text":"你好"}`), "cohere.command-r")
-	if !ok || chunk.Choices[0].Delta.Content != "你好" {
-		t.Errorf("cohere text 事件解析 = %+v, %v", chunk, ok)
-	}
-	chunk, ok = parseGenAiEvent([]byte(`{"apiFormat":"COHERE","finishReason":"COMPLETE","usage":{"promptTokens":3,"completionTokens":5,"totalTokens":8}}`), "cohere.command-r")
-	if !ok || *chunk.Choices[0].FinishReason != "stop" || chunk.Usage.TotalTokens != 8 {
-		t.Errorf("cohere 终帧解析 = %+v, %v", chunk, ok)
 	}
 }

internal/service/aiconvert.go

@@ -1,332 +0,0 @@
-package service
-
-import (
-	"encoding/json"
-	"fmt"
-	"strings"
-
-	"oci-portal/internal/aiwire"
-)
-
-// ErrAiUnsupportedBlock 表示请求含网关无法承接的内容块(仅支持文本与图片)。
-var ErrAiUnsupportedBlock = fmt.Errorf("暂不支持文本与图片以外的内容块")
-
-// AnthropicToIR 把 Anthropic Messages 请求转为 IR(OpenAI 线格式)。
-// tool_result 块拆为独立 tool 角色消息(一拆多),顶层 system 变首条 system 消息。
-func AnthropicToIR(req aiwire.MessagesRequest) (aiwire.ChatRequest, error) {
-	ir := aiwire.ChatRequest{
-		Model:       req.Model,
-		Temperature: req.Temperature,
-		TopP:        req.TopP,
-		TopK:        req.TopK,
-		Stop:        req.StopSequences,
-		Stream:      req.Stream,
-		Tools:       anthTools(req.Tools),
-		ToolChoice:  anthToolChoice(req.ToolChoice),
-	}
-	if req.MaxTokens > 0 {
-		mt := req.MaxTokens
-		ir.MaxTokens = &mt
-	}
-	if sys := req.SystemText(); sys != "" {
-		ir.Messages = append(ir.Messages, aiwire.ChatMessage{Role: "system", Content: aiwire.NewTextContent(sys)})
-	}
-	for _, m := range req.Messages {
-		msgs, err := anthMessageToIR(m)
-		if err != nil {
-			return ir, err
-		}
-		ir.Messages = append(ir.Messages, msgs...)
-	}
-	return ir, nil
-}
-
-// anthMessageToIR 拆解单条 Anthropic 消息;user 消息里的 tool_result 前置为独立 tool 消息,
-// image 块转为 IR image_url 部件。
-func anthMessageToIR(m aiwire.AnthMessage) ([]aiwire.ChatMessage, error) {
-	var out []aiwire.ChatMessage
-	var parts []aiwire.ContentPart
-	var toolCalls []aiwire.ToolCall
-	hasImage := false
-	for _, b := range m.Content.AllBlocks() {
-		switch b.Type {
-		case "text":
-			parts = append(parts, aiwire.ContentPart{Type: "text", Text: b.Text})
-		case "image":
-			p, err := anthImagePart(b.Source)
-			if err != nil {
-				return nil, err
-			}
-			parts, hasImage = append(parts, p), true
-		case "tool_use":
-			toolCalls = append(toolCalls, aiwire.ToolCall{ID: b.ID, Type: "function",
-				Function: aiwire.FunctionCall{Name: b.Name, Arguments: string(b.Input)}})
-		case "tool_result":
-			out = append(out, aiwire.ChatMessage{Role: "tool", ToolCallID: b.ToolUseID,
-				Content: aiwire.NewTextContent(b.ResultText())})
-		case "thinking", "redacted_thinking":
-			// 一期忽略 thinking 块
-		default:
-			return nil, ErrAiUnsupportedBlock
-		}
-	}
-	content := partsContent(parts, hasImage)
-	if hasImage || content.JoinText() != "" || len(toolCalls) > 0 {
-		out = append(out, aiwire.ChatMessage{Role: m.Role, Content: content, ToolCalls: toolCalls})
-	}
-	return out, nil
-}
-
-// anthImagePart 把 Anthropic image 块转为 IR image_url 部件(base64 → data URI)。
-func anthImagePart(source json.RawMessage) (aiwire.ContentPart, error) {
-	var src struct {
-		Type      string `json:"type"`
-		MediaType string `json:"media_type"`
-		Data      string `json:"data"`
-		URL       string `json:"url"`
-	}
-	if err := json.Unmarshal(source, &src); err != nil {
-		return aiwire.ContentPart{}, fmt.Errorf("image source 解析失败: %w", err)
-	}
-	switch src.Type {
-	case "base64":
-		if src.MediaType == "" || src.Data == "" {
-			return aiwire.ContentPart{}, fmt.Errorf("image source 缺少 media_type 或 data")
-		}
-		url := "data:" + src.MediaType + ";base64," + src.Data
-		return aiwire.ContentPart{Type: "image_url", ImageURL: &aiwire.ImageURL{URL: url}}, nil
-	case "url":
-		if src.URL == "" {
-			return aiwire.ContentPart{}, fmt.Errorf("image source 缺少 url")
-		}
-		return aiwire.ContentPart{Type: "image_url", ImageURL: &aiwire.ImageURL{URL: src.URL}}, nil
-	}
-	return aiwire.ContentPart{}, fmt.Errorf("不支持的 image source 类型 %q", src.Type)
-}
-
-// partsContent 无图片时退回字符串形态(与上游线格式习惯一致),含图片时保留块数组。
-func partsContent(parts []aiwire.ContentPart, hasImage bool) aiwire.Content {
-	if !hasImage {
-		var sb strings.Builder
-		for _, p := range parts {
-			sb.WriteString(p.Text)
-		}
-		return aiwire.NewTextContent(sb.String())
-	}
-	return aiwire.NewPartsContent(parts)
-}
-
-func anthTools(tools []aiwire.AnthTool) []aiwire.Tool {
-	if len(tools) == 0 {
-		return nil
-	}
-	out := make([]aiwire.Tool, 0, len(tools))
-	for _, t := range tools {
-		out = append(out, aiwire.Tool{Type: "function", Function: aiwire.FunctionDef{
-			Name: t.Name, Description: t.Description, Parameters: t.InputSchema}})
-	}
-	return out
-}
-
-// anthToolChoice 映射 {type:auto|any|tool,name} → OpenAI 形态。
-func anthToolChoice(raw json.RawMessage) json.RawMessage {
-	if len(raw) == 0 {
-		return nil
-	}
-	var tc struct {
-		Type string `json:"type"`
-		Name string `json:"name"`
-	}
-	if json.Unmarshal(raw, &tc) != nil {
-		return nil
-	}
-	switch tc.Type {
-	case "auto":
-		return json.RawMessage(`"auto"`)
-	case "any":
-		return json.RawMessage(`"required"`)
-	case "tool":
-		b, _ := json.Marshal(map[string]any{"type": "function", "function": map[string]string{"name": tc.Name}})
-		return b
-	case "none":
-		return json.RawMessage(`"none"`)
-	}
-	return nil
-}
-
-// IRRespToAnthropic 把 IR 非流式响应转为 Anthropic Messages 响应。
-func IRRespToAnthropic(resp *aiwire.ChatResponse, id string) aiwire.MessagesResponse {
-	out := aiwire.MessagesResponse{ID: id, Type: "message", Role: "assistant", Model: resp.Model, Content: []aiwire.AnthBlock{}}
-	if len(resp.Choices) > 0 {
-		choice := resp.Choices[0]
-		if text := choice.Message.Content.JoinText(); text != "" {
-			out.Content = append(out.Content, aiwire.AnthBlock{Type: "text", Text: text})
-		}
-		for _, tc := range choice.Message.ToolCalls {
-			out.Content = append(out.Content, aiwire.AnthBlock{Type: "tool_use", ID: tc.ID,
-				Name: tc.Function.Name, Input: argsToJSON(tc.Function.Arguments)})
-		}
-		out.StopReason = anthStopReason(choice.FinishReason)
-	}
-	if resp.Usage != nil {
-		out.Usage = aiwire.AnthUsage{InputTokens: resp.Usage.PromptTokens, OutputTokens: resp.Usage.CompletionTokens,
-			CacheReadInputTokens: resp.Usage.CachedTokens()}
-	}
-	return out
-}
-
-// argsToJSON 保证 tool_use.input 是合法 JSON 对象(模型可能产出非法片段)。
-func argsToJSON(args string) json.RawMessage {
-	trimmed := strings.TrimSpace(args)
-	if trimmed == "" {
-		return json.RawMessage(`{}`)
-	}
-	if json.Valid([]byte(trimmed)) {
-		return json.RawMessage(trimmed)
-	}
-	b, _ := json.Marshal(map[string]string{"_raw": args})
-	return b
-}
-
-func anthStopReason(finish string) string {
-	switch finish {
-	case "length":
-		return "max_tokens"
-	case "tool_calls":
-		return "tool_use"
-	default:
-		return "end_turn"
-	}
-}
-
-// ---- Anthropic 流式状态机 ----
-
-// AnthEvent 是一条待写出的 Anthropic SSE 事件。
-type AnthEvent struct {
-	Event string
-	Data  any
-}
-
-// AnthStream 把 IR chunk 流聚合为 Anthropic 事件序列:
-// message_start → content_block_start/delta/stop(text 与 tool_use 分块)→ message_delta → message_stop。
-type AnthStream struct {
-	id, model   string
-	started     bool
-	blockOpen   bool
-	blockIsTool bool
-	toolID      string
-	blockIndex  int
-	stopReason  string
-	usage       aiwire.AnthUsage
-}
-
-// NewAnthStream 构造状态机;id 为响应消息 ID。
-func NewAnthStream(id, model string) *AnthStream {
-	return &AnthStream{id: id, model: model, blockIndex: -1, stopReason: "end_turn"}
-}
-
-// Feed 消费一个 IR chunk,返回应立即写出的事件。
-func (st *AnthStream) Feed(chunk aiwire.ChatChunk) []AnthEvent {
-	var events []AnthEvent
-	if !st.started {
-		st.started = true
-		events = append(events, st.startEvent())
-	}
-	if chunk.Usage != nil {
-		st.usage = aiwire.AnthUsage{InputTokens: chunk.Usage.PromptTokens, OutputTokens: chunk.Usage.CompletionTokens,
-			CacheReadInputTokens: chunk.Usage.CachedTokens()}
-	}
-	for _, choice := range chunk.Choices {
-		events = append(events, st.feedDelta(choice.Delta)...)
-		if choice.FinishReason != nil && *choice.FinishReason != "" {
-			st.stopReason = anthStopReason(*choice.FinishReason)
-		}
-	}
-	return events
-}
-
-func (st *AnthStream) startEvent() AnthEvent {
-	return AnthEvent{Event: "message_start", Data: map[string]any{
-		"type": "message_start",
-		"message": map[string]any{
-			"id": st.id, "type": "message", "role": "assistant", "model": st.model,
-			"content": []any{}, "stop_reason": nil,
-			"usage": map[string]int{"input_tokens": 0, "output_tokens": 0},
-		},
-	}}
-}
-
-// feedDelta 处理文本与工具调用增量,必要时切块。
-func (st *AnthStream) feedDelta(d aiwire.Delta) []AnthEvent {
-	var events []AnthEvent
-	if d.Content != "" {
-		if !st.blockOpen || st.blockIsTool {
-			events = append(events, st.openBlock(false, "", "")...)
-		}
-		events = append(events, AnthEvent{Event: "content_block_delta", Data: map[string]any{
-			"type": "content_block_delta", "index": st.blockIndex,
-			"delta": map[string]string{"type": "text_delta", "text": d.Content},
-		}})
-	}
-	for _, tc := range d.ToolCalls {
-		if tc.ID != "" && (!st.blockOpen || !st.blockIsTool || st.toolID != tc.ID) {
-			events = append(events, st.openBlock(true, tc.ID, tc.Function.Name)...)
-		}
-		if tc.Function.Arguments != "" && st.blockOpen && st.blockIsTool {
-			events = append(events, AnthEvent{Event: "content_block_delta", Data: map[string]any{
-				"type": "content_block_delta", "index": st.blockIndex,
-				"delta": map[string]string{"type": "input_json_delta", "partial_json": tc.Function.Arguments},
-			}})
-		}
-	}
-	return events
-}
-
-// openBlock 关闭当前块并打开新块(text 或 tool_use)。
-func (st *AnthStream) openBlock(isTool bool, toolID, toolName string) []AnthEvent {
-	var events []AnthEvent
-	if st.blockOpen {
-		events = append(events, st.closeBlockEvent())
-	}
-	st.blockOpen, st.blockIsTool, st.toolID = true, isTool, toolID
-	st.blockIndex++
-	block := map[string]any{"type": "text", "text": ""}
-	if isTool {
-		block = map[string]any{"type": "tool_use", "id": toolID, "name": toolName, "input": map[string]any{}}
-	}
-	events = append(events, AnthEvent{Event: "content_block_start", Data: map[string]any{
-		"type": "content_block_start", "index": st.blockIndex, "content_block": block,
-	}})
-	return events
-}
-
-func (st *AnthStream) closeBlockEvent() AnthEvent {
-	return AnthEvent{Event: "content_block_stop", Data: map[string]any{
-		"type": "content_block_stop", "index": st.blockIndex,
-	}}
-}
-
-// Finish 在上游流结束后收尾:关块 → message_delta(stop_reason+usage)→ message_stop。
-func (st *AnthStream) Finish() []AnthEvent {
-	var events []AnthEvent
-	if !st.started {
-		events = append(events, st.startEvent())
-		st.started = true
-	}
-	if st.blockOpen {
-		events = append(events, st.closeBlockEvent())
-		st.blockOpen = false
-	}
-	events = append(events,
-		AnthEvent{Event: "message_delta", Data: map[string]any{
-			"type":  "message_delta",
-			"delta": map[string]any{"stop_reason": st.stopReason, "stop_sequence": nil},
-			"usage": map[string]int{"output_tokens": st.usage.OutputTokens},
-		}},
-		AnthEvent{Event: "message_stop", Data: map[string]any{"type": "message_stop"}},
-	)
-	return events
-}
-
-// Usage 返回聚合到的用量(供调用日志)。
-func (st *AnthStream) Usage() aiwire.AnthUsage { return st.usage }

internal/service/aiconvert_test.go

@@ -1,136 +0,0 @@
-package service
-
-import (
-	"encoding/json"
-	"strings"
-	"testing"
-
-	"oci-portal/internal/aiwire"
-)
-
-func mustAnthReq(t *testing.T, body string) aiwire.MessagesRequest {
-	t.Helper()
-	var req aiwire.MessagesRequest
-	if err := json.Unmarshal([]byte(body), &req); err != nil {
-		t.Fatalf("unmarshal anthropic request: %v", err)
-	}
-	return req
-}
-
-func TestAnthropicToIR(t *testing.T) {
-	req := mustAnthReq(t, `{
-		"model": "meta.llama-3.3-70b-instruct", "max_tokens": 128, "system": "你是助手",
-		"messages": [
-			{"role": "user", "content": "东京天气?"},
-			{"role": "assistant", "content": [
-				{"type": "text", "text": "查询中"},
-				{"type": "tool_use", "id": "t1", "name": "get_weather", "input": {"city": "东京"}}
-			]},
-			{"role": "user", "content": [
-				{"type": "tool_result", "tool_use_id": "t1", "content": "晴 25 度"},
-				{"type": "text", "text": "继续"}
-			]}
-		],
-		"tools": [{"name": "get_weather", "input_schema": {"type": "object"}}],
-		"tool_choice": {"type": "any"}
-	}`)
-	ir, err := AnthropicToIR(req)
-	if err != nil {
-		t.Fatalf("AnthropicToIR: %v", err)
-	}
-	roles := make([]string, 0, len(ir.Messages))
-	for _, m := range ir.Messages {
-		roles = append(roles, m.Role)
-	}
-	want := []string{"system", "user", "assistant", "tool", "user"}
-	if strings.Join(roles, ",") != strings.Join(want, ",") {
-		t.Fatalf("roles = %v, want %v", roles, want)
-	}
-	if ir.Messages[2].ToolCalls[0].Function.Name != "get_weather" {
-		t.Errorf("tool_use 未映射: %+v", ir.Messages[2])
-	}
-	if ir.Messages[3].ToolCallID != "t1" || ir.Messages[3].Content.JoinText() != "晴 25 度" {
-		t.Errorf("tool_result 未拆为 tool 消息: %+v", ir.Messages[3])
-	}
-	if ir.MaxTokens == nil || *ir.MaxTokens != 128 || len(ir.Tools) != 1 {
-		t.Errorf("max_tokens/tools 未直通")
-	}
-	if string(ir.ToolChoice) != `"required"` {
-		t.Errorf("tool_choice any → %s, want required", ir.ToolChoice)
-	}
-	// image 块拒绝
-	bad := mustAnthReq(t, `{"model":"m","max_tokens":1,"messages":[{"role":"user","content":[{"type":"image","source":{}}]}]}`)
-	if _, err := AnthropicToIR(bad); err == nil {
-		t.Error("image 块应被拒绝")
-	}
-}
-
-func TestIRRespToAnthropic(t *testing.T) {
-	resp := &aiwire.ChatResponse{
-		Model: "m1",
-		Choices: []aiwire.Choice{{
-			Message: aiwire.ChatMessage{
-				Role:    "assistant",
-				Content: aiwire.NewTextContent("查到了"),
-				ToolCalls: []aiwire.ToolCall{{ID: "t1", Type: "function",
-					Function: aiwire.FunctionCall{Name: "get_weather", Arguments: `{"city":"东京"}`}}},
-			},
-			FinishReason: "tool_calls",
-		}},
-		Usage: &aiwire.Usage{PromptTokens: 10, CompletionTokens: 5},
-	}
-	out := IRRespToAnthropic(resp, "msg_1")
-	if len(out.Content) != 2 || out.Content[0].Type != "text" || out.Content[1].Type != "tool_use" {
-		t.Fatalf("content 块 = %+v", out.Content)
-	}
-	if out.StopReason != "tool_use" || out.Usage.InputTokens != 10 || out.Usage.OutputTokens != 5 {
-		t.Errorf("stop_reason/usage 未映射: %+v", out)
-	}
-	if string(out.Content[1].Input) != `{"city":"东京"}` {
-		t.Errorf("tool input = %s", out.Content[1].Input)
-	}
-}
-
-// eventTypes 提取事件类型序列便于断言。
-func eventTypes(events []AnthEvent) string {
-	types := make([]string, 0, len(events))
-	for _, e := range events {
-		types = append(types, e.Event)
-	}
-	return strings.Join(types, ",")
-}
-
-func TestAnthStreamTextAndTool(t *testing.T) {
-	st := NewAnthStream("msg_1", "m1")
-	fr := "tool_calls"
-	var events []AnthEvent
-	events = append(events, st.Feed(aiwire.ChatChunk{Choices: []aiwire.ChunkChoice{{Delta: aiwire.Delta{Content: "你"}}}})...)
-	events = append(events, st.Feed(aiwire.ChatChunk{Choices: []aiwire.ChunkChoice{{Delta: aiwire.Delta{Content: "好"}}}})...)
-	events = append(events, st.Feed(aiwire.ChatChunk{Choices: []aiwire.ChunkChoice{{Delta: aiwire.Delta{
-		ToolCalls: []aiwire.ToolCallDelta{{ID: "t1", Function: aiwire.FunctionCallDelta{Name: "get_weather", Arguments: `{"ci`}}},
-	}}}})...)
-	events = append(events, st.Feed(aiwire.ChatChunk{
-		Choices: []aiwire.ChunkChoice{{Delta: aiwire.Delta{ToolCalls: []aiwire.ToolCallDelta{{ID: "t1", Function: aiwire.FunctionCallDelta{Arguments: `ty":"东京"}`}}}}, FinishReason: &fr}},
-		Usage:   &aiwire.Usage{PromptTokens: 8, CompletionTokens: 4},
-	})...)
-	events = append(events, st.Finish()...)
-
-	got := eventTypes(events)
-	want := "message_start,content_block_start,content_block_delta,content_block_delta," +
-		"content_block_stop,content_block_start,content_block_delta,content_block_delta," +
-		"content_block_stop,message_delta,message_stop"
-	if got != want {
-		t.Fatalf("事件序列:\n got %s\nwant %s", got, want)
-	}
-	if st.Usage().OutputTokens != 4 {
-		t.Errorf("usage 未聚合: %+v", st.Usage())
-	}
-}
-
-func TestAnthStreamEmpty(t *testing.T) {
-	st := NewAnthStream("msg_1", "m1")
-	events := st.Finish()
-	if got := eventTypes(events); got != "message_start,message_delta,message_stop" {
-		t.Errorf("空流事件序列 = %s", got)
-	}
-}

internal/service/aigateway.go

@@ -5,6 +5,7 @@ import (
 	"crypto/rand"
 	"crypto/sha256"
 	"encoding/hex"
+	"encoding/json"
 	"errors"
 	"fmt"
 	"log"
@@ -30,10 +31,6 @@ const (
 	aiBackoffCap    = 30 * time.Minute
 	// aiKeyTouchGap 是 LastUsedAt 的最小写库间隔,避免高频调用刷库
 	aiKeyTouchGap = time.Minute
-	// modelRecheckGap 是已标记不可用模型的复检间隔(恢复供给自动解除标记);
-	// validateBatchCap 限制单渠道单轮验证的试调次数
-	modelRecheckGap  = 20 * time.Hour
-	validateBatchCap = 32
 	// 内容日志(红线例外)约束:开启必须限时(上限 7 天),正文截断,短保留
 	aiContentLogMaxHours  = 168
 	aiContentLogRetention = 7 * 24 * time.Hour
@@ -59,17 +56,13 @@ type AiGatewayService struct {
 	// touchMu 保护各密钥的最近触达时间(内存节流,不追求跨实例精确)
 	touchMu   sync.Mutex
 	lastTouch map[uint]time.Time
-	// validateMu 保护 validating:同一渠道的模型验证不并发
-	validateMu sync.Mutex
-	validating map[uint]bool
 	// onChannelsChanged 在渠道增删后触发,由 main 装配为探测任务同步钩子
 	onChannelsChanged func(context.Context)
 }
 
 // NewAiGatewayService 组装依赖;调用 StartCleanup 后开始调用日志周期清理。
 func NewAiGatewayService(db *gorm.DB, configs *OciConfigService, client oci.Client) *AiGatewayService {
-	return &AiGatewayService{db: db, configs: configs, client: client,
+	return &AiGatewayService{db: db, configs: configs, client: client, lastTouch: map[uint]time.Time{}}
-		lastTouch: map[uint]time.Time{}, validating: map[uint]bool{}}
 }
 
 // SetOnChannelsChanged 注册渠道数量变化钩子(渠道创建/删除成功后调用)。
@@ -87,7 +80,7 @@ func (s *AiGatewayService) fireChannelsChanged(ctx context.Context) {
 
 // CreateKey 生成网关密钥并返回明文(仅此一次);customValue 非空时使用给定值,
 // group 非空时该密钥只在同分组渠道内路由。
-func (s *AiGatewayService) CreateKey(ctx context.Context, name, customValue, group string) (string, *model.AiKey, error) {
+func (s *AiGatewayService) CreateKey(ctx context.Context, name, customValue, group string, models []string) (string, *model.AiKey, error) {
 	name = strings.TrimSpace(name)
 	if name == "" {
 		return "", nil, fmt.Errorf("密钥名称不能为空")
@@ -103,13 +96,28 @@ func (s *AiGatewayService) CreateKey(ctx context.Context, name, customValue, gro
 	if len(raw) < 8 {
 		return "", nil, fmt.Errorf("自定义密钥至少 8 个字符")
 	}
-	key := &model.AiKey{Name: name, KeyHash: hashKey(raw), Tail: raw[len(raw)-4:], Group: strings.TrimSpace(group), Enabled: true}
+	key := &model.AiKey{Name: name, KeyHash: hashKey(raw), Tail: raw[len(raw)-4:], Group: strings.TrimSpace(group), Models: normalizeKeyModels(models), Enabled: true}
 	if err := s.db.WithContext(ctx).Create(key).Error; err != nil {
 		return "", nil, fmt.Errorf("密钥名称或取值与现有密钥重复")
 	}
 	return raw, key, nil
 }
 
+// normalizeKeyModels 规范化模型白名单:trim、剔空串、保序去重;结果为空返回 nil(= 不限)。
+func normalizeKeyModels(models []string) []string {
+	var out []string
+	seen := map[string]bool{}
+	for _, m := range models {
+		m = strings.TrimSpace(m)
+		if m == "" || seen[m] {
+			continue
+		}
+		seen[m] = true
+		out = append(out, m)
+	}
+	return out
+}
+
 func hashKey(raw string) string {
 	sum := sha256.Sum256([]byte(raw))
 	return hex.EncodeToString(sum[:])
@@ -122,8 +130,8 @@ func (s *AiGatewayService) Keys(ctx context.Context) ([]model.AiKey, error) {
 	return keys, err
 }
 
-// UpdateKey 修改密钥名称 / 启用状态 / 分组(group 指针非空即覆盖,可置空)。
+// UpdateKey 修改密钥名称 / 启用状态 / 分组 / 模型白名单(指针非空即覆盖,可置空)。
-func (s *AiGatewayService) UpdateKey(ctx context.Context, id uint, name string, enabled *bool, group *string) error {
+func (s *AiGatewayService) UpdateKey(ctx context.Context, id uint, name string, enabled *bool, group *string, models *[]string) error {
 	updates := map[string]any{}
 	if name = strings.TrimSpace(name); name != "" {
 		updates["name"] = name
@@ -134,6 +142,14 @@ func (s *AiGatewayService) UpdateKey(ctx context.Context, id uint, name string,
 	if group != nil {
 		updates["key_group"] = strings.TrimSpace(*group)
 	}
+	if models != nil {
+		// 手动序列化走 map 更新,不依赖 GORM map 路径对 serializer 的支持
+		b, err := json.Marshal(normalizeKeyModels(*models))
+		if err != nil {
+			return fmt.Errorf("serialize models: %w", err)
+		}
+		updates["models"] = string(b)
+	}
 	if len(updates) == 0 {
 		return nil
 	}
@@ -298,12 +314,17 @@ func (s *AiGatewayService) ProbeChannel(ctx context.Context, id uint) (*model.Ai
 	return &fresh, nil
 }
 
-// probe 执行探测并返回 (状态, 错误摘要);同时完成模型缓存同步(保留不可用标记)。
+// probe 执行探测并返回 (状态, 错误摘要);同时完成模型缓存同步(黑名单模型不入库)。
 func (s *AiGatewayService) probe(ctx context.Context, cred oci.Credentials, ch *model.AiChannel) (string, string) {
 	models, err := s.client.ListGenAiModels(ctx, cred, ch.Region)
 	if err != nil {
 		return classifyProbeErr(err), truncateErr(oci.CompactError(err))
 	}
+	models = supportedGatewayModels(models)
+	models, err = s.withoutBlacklisted(ctx, models)
+	if err != nil {
+		return "error", truncateErr(err.Error())
+	}
 	if len(models) == 0 {
 		_ = s.replaceModels(ctx, ch.ID, nil)
 		return "no_service", "区域无可用模型(GenAI 服务不可用或未开放)"
@@ -311,31 +332,12 @@ func (s *AiGatewayService) probe(ctx context.Context, cred oci.Credentials, ch *
 	if err := s.replaceModels(ctx, ch.ID, models); err != nil {
 		return "error", truncateErr(err.Error())
 	}
-	usable, err := s.usableModels(ctx, ch.ID, models)
+	return s.probeChat(ctx, cred, ch, models)
-	if err != nil {
-		return "error", truncateErr(err.Error())
-	}
-	return s.probeChat(ctx, cred, ch, usable)
 }
 
-// usableModels 过滤掉缓存中已标记不可按需调用的模型,探测候选不再反复踩坑。
+// probeChat 按候选顺序试调(上限 8):遇「模型不可按需调用」(微调基座 400 / 实体
-func (s *AiGatewayService) usableModels(ctx context.Context, channelID uint, models []oci.GenAiModel) ([]oci.GenAiModel, error) {
+// 不存在 404)换下一个候选,错误信息带模型名供用户加入黑名单;401/403 与鉴权类 404
-	marks, err := s.loadModelMarks(ctx, channelID)
+// 属租户级直接定论 no_quota;其余错误(元数据标 CHAT 但实际不可对话等)累计 3 次止损。
-	if err != nil {
-		return nil, err
-	}
-	out := make([]oci.GenAiModel, 0, len(models))
-	for _, m := range models {
-		if !marks[m.Ocid].Unusable {
-			out = append(out, m)
-		}
-	}
-	return out, nil
-}
-
-// probeChat 按候选顺序试调(上限 8):遇「模型不可按需调用」标记剔除并换下一个;
-// 401/403 与鉴权类 404 属租户级直接定论 no_quota;其余错误(部分模型元数据标 CHAT
-// 但实际不可对话,如 voice agent)累计 3 次止损。
 func (s *AiGatewayService) probeChat(ctx context.Context, cred oci.Credentials, ch *model.AiChannel, models []oci.GenAiModel) (string, string) {
 	status, detail := "error", "无可试调对话模型"
 	errBudget := 3
@@ -345,8 +347,7 @@ func (s *AiGatewayService) probeChat(ctx context.Context, cred oci.Credentials,
 		case code == 200 || code == 429:
 			return "ok", ""
 		case oci.IsModelUnavailable(err):
-			s.markModelUnusable(ctx, ch.ID, m.Ocid, oci.CompactError(err))
+			status, detail = "error", truncateErr(fmt.Sprintf("%s: 不可按需调用,建议加入模型黑名单", m.Name))
-			status, detail = "error", truncateErr(fmt.Sprintf("%s: 不可按需调用,已从模型池剔除", m.Name))
 		case code == 401 || code == 403 || code == 404:
 			return "no_quota", truncateErr(oci.CompactError(err))
 		default:
@@ -359,106 +360,8 @@ func (s *AiGatewayService) probeChat(ctx context.Context, cred oci.Credentials,
 	return status, detail
 }
 
-// markModelUnusable 把 (渠道, 模型OCID) 标记为不可按需调用;失败仅记日志不阻断主流程。
+// probeCandidateCap 是单次探测的候选上限:放宽到 8 让一次探测有机会越过整批
-func (s *AiGatewayService) markModelUnusable(ctx context.Context, channelID uint, ocid, reason string) {
+// 不可按需调用的坏模型找到可用者;其他错误另有 3 次止损预算。
-	err := s.db.WithContext(ctx).Model(&model.AiModelCache{}).
-		Where("channel_id = ? AND model_ocid = ?", channelID, ocid).
-		Updates(map[string]any{"unusable": true, "unusable_reason": shortReason(reason), "checked_at": time.Now()}).Error
-	if err != nil {
-		log.Printf("mark model unusable: %v", err)
-	}
-}
-
-func shortReason(reason string) string {
-	if len(reason) > 200 {
-		return reason[:200]
-	}
-	return reason
-}
-
-// beginValidate 抢占渠道的验证执行权,同渠道同时只跑一轮。
-func (s *AiGatewayService) beginValidate(id uint) bool {
-	s.validateMu.Lock()
-	defer s.validateMu.Unlock()
-	if s.validating[id] {
-		return false
-	}
-	s.validating[id] = true
-	return true
-}
-
-func (s *AiGatewayService) endValidate(id uint) {
-	s.validateMu.Lock()
-	delete(s.validating, id)
-	s.validateMu.Unlock()
-}
-
-// validateTargets 取待验证行:从未验证的新模型,以及标记超过复检间隔的模型(仅对话能力,
-// 兼容存量空串);单轮上限 validateBatchCap 控制调用量。
-func (s *AiGatewayService) validateTargets(ctx context.Context, channelID uint) ([]model.AiModelCache, error) {
-	stale := time.Now().Add(-modelRecheckGap)
-	var rows []model.AiModelCache
-	err := s.db.WithContext(ctx).
-		Where("channel_id = ? AND capability IN ?", channelID, []string{"CHAT", ""}).
-		Where("checked_at IS NULL OR (unusable = ? AND checked_at < ?)", true, stale).
-		Order("id ASC").Limit(validateBatchCap).Find(&rows).Error
-	return rows, err
-}
-
-// validateChannelModels 逐个试调渠道内待验证模型并落结论,把不可按需调用的模型
-// 从池中剔除、把恢复供给的解除标记;ListModels 无字段可事先判别,只能试调习得。
-func (s *AiGatewayService) validateChannelModels(ctx context.Context, channelID uint) {
-	if !s.beginValidate(channelID) {
-		return
-	}
-	defer s.endValidate(channelID)
-	var ch model.AiChannel
-	if err := s.db.WithContext(ctx).First(&ch, channelID).Error; err != nil {
-		return
-	}
-	cred, err := s.configs.credentialsByID(ctx, ch.OciConfigID)
-	if err != nil {
-		return
-	}
-	rows, err := s.validateTargets(ctx, channelID)
-	if err != nil {
-		log.Printf("validate channel %d models: %v", channelID, err)
-		return
-	}
-	for _, r := range rows {
-		s.validateOne(ctx, cred, &ch, r)
-	}
-}
-
-// validateOne 试调单个模型并落库:可用(200/429)清除标记;模型级不可用标记剔除;
-// 其余 4xx 记录已检不改状态;5xx/网络类瞬态不落 checked_at,下轮再验。
-func (s *AiGatewayService) validateOne(ctx context.Context, cred oci.Credentials, ch *model.AiChannel, r model.AiModelCache) {
-	code, err := s.client.GenAiProbeChat(ctx, cred, ch.Region, r.ModelOcid, r.Name)
-	updates := map[string]any{"checked_at": time.Now()}
-	switch {
-	case code == 200 || code == 429:
-		updates["unusable"], updates["unusable_reason"] = false, ""
-	case oci.IsModelUnavailable(err):
-		updates["unusable"], updates["unusable_reason"] = true, shortReason(oci.CompactError(err))
-	case code == 0 || code >= 500:
-		return
-	}
-	s.db.WithContext(ctx).Model(&model.AiModelCache{}).Where("id = ?", r.ID).Updates(updates)
-}
-
-// validateModelsAsync 后台验证渠道模型池(手动同步后触发,不阻塞接口响应)。
-func (s *AiGatewayService) validateModelsAsync(ctx context.Context, channelID uint) {
-	s.wg.Add(1)
-	go func() {
-		defer s.wg.Done()
-		vctx, cancel := context.WithTimeout(context.WithoutCancel(ctx), 3*time.Minute)
-		defer cancel()
-		s.validateChannelModels(vctx, channelID)
-	}()
-}
-
-// probeCandidateCap 是单次探测的候选上限:模型级不可用会被标记剔除不再重试,
-// 放宽到 8 让一次探测有机会越过整批坏模型找到可用者;其他错误另有 3 次止损预算。
 const probeCandidateCap = 8
 
 // probeCandidates 只取对话模型,按可靠度排序后跨厂商取候选:主流文本模型优先;
@@ -551,8 +454,7 @@ func truncateErr(msg string) string {
 	return msg
 }
 
-// SyncModels 重新拉取渠道区域的模型列表并覆盖缓存(标记按 OCID 结转),
+// SyncModels 重新拉取渠道区域的模型列表并覆盖缓存,黑名单中的模型不入库。
-// 随后触发后台验证:新模型逐个试调,不可按需调用的数十秒内从池中剔除。
 func (s *AiGatewayService) SyncModels(ctx context.Context, id uint) ([]model.AiModelCache, error) {
 	var ch model.AiChannel
 	if err := s.db.WithContext(ctx).First(&ch, id).Error; err != nil {
@@ -566,29 +468,23 @@ func (s *AiGatewayService) SyncModels(ctx context.Context, id uint) ([]model.AiM
 	if err != nil {
 		return nil, fmt.Errorf("同步模型失败:%s", oci.CompactError(err))
 	}
+	models = supportedGatewayModels(models)
+	if models, err = s.withoutBlacklisted(ctx, models); err != nil {
+		return nil, err
+	}
 	if err := s.replaceModels(ctx, id, models); err != nil {
 		return nil, err
 	}
-	s.validateModelsAsync(ctx, id)
 	return s.channelModels(ctx, id)
 }
 
-// replaceModels 以事务整组覆盖渠道模型缓存,按 OCID 结转不可用标记与验证时间
+// replaceModels 以事务整组覆盖渠道模型缓存。
-// (OCID 变化视为新条目,自然回到待验证状态)。
 func (s *AiGatewayService) replaceModels(ctx context.Context, channelID uint, models []oci.GenAiModel) error {
-	marks, err := s.loadModelMarks(ctx, channelID)
-	if err != nil {
-		return err
-	}
 	rows := make([]model.AiModelCache, 0, len(models))
 	now := time.Now()
 	for _, m := range models {
-		row := model.AiModelCache{ChannelID: channelID, ModelOcid: m.Ocid, Name: m.Name, Vendor: m.Vendor,
+		rows = append(rows, model.AiModelCache{ChannelID: channelID, ModelOcid: m.Ocid, Name: m.Name, Vendor: m.Vendor,
-			Capability: m.Capability, SyncedAt: now, DeprecatedAt: m.Deprecated, RetiredAt: m.Retired}
+			Capability: m.Capability, SyncedAt: now, DeprecatedAt: m.Deprecated, RetiredAt: m.Retired})
-		if prev, ok := marks[m.Ocid]; ok {
-			row.Unusable, row.UnusableReason, row.CheckedAt = prev.Unusable, prev.UnusableReason, prev.CheckedAt
-		}
-		rows = append(rows, row)
 	}
 	return s.db.WithContext(ctx).Transaction(func(tx *gorm.DB) error {
 		if err := tx.Where("channel_id = ?", channelID).Delete(&model.AiModelCache{}).Error; err != nil {
@@ -601,33 +497,17 @@ func (s *AiGatewayService) replaceModels(ctx context.Context, channelID uint, mo
 	})
 }
 
-// loadModelMarks 取渠道内带标记或已验证过的行(OCID → 行),供同步结转与候选过滤。
-func (s *AiGatewayService) loadModelMarks(ctx context.Context, channelID uint) (map[string]model.AiModelCache, error) {
-	var rows []model.AiModelCache
-	err := s.db.WithContext(ctx).Select("model_ocid", "unusable", "unusable_reason", "checked_at").
-		Where("channel_id = ? AND (unusable = ? OR checked_at IS NOT NULL)", channelID, true).Find(&rows).Error
-	if err != nil {
-		return nil, fmt.Errorf("load model marks: %w", err)
-	}
-	marks := make(map[string]model.AiModelCache, len(rows))
-	for _, r := range rows {
-		marks[r.ModelOcid] = r
-	}
-	return marks, nil
-}
-
 func (s *AiGatewayService) channelModels(ctx context.Context, channelID uint) ([]model.AiModelCache, error) {
 	var rows []model.AiModelCache
 	err := s.db.WithContext(ctx).Where("channel_id = ?", channelID).Order("name ASC").Find(&rows).Error
 	return rows, err
 }
 
-// GatewayModels 聚合启用渠道的可用模型(按名称去重,剔除不可按需调用标记),
+// GatewayModels 聚合启用渠道的可用模型(按名称去重),供 /ai/v1/models;
-// 供 /ai/v1/models;group 非空时仅聚合该分组渠道(与密钥分组路由口径一致)。
+// group 非空时仅聚合该分组渠道(与密钥分组路由口径一致)。
 func (s *AiGatewayService) GatewayModels(ctx context.Context, group string) (aiwire.ModelList, error) {
 	q := s.db.WithContext(ctx).
-		Joins("JOIN ai_channels ON ai_channels.id = ai_model_caches.channel_id AND ai_channels.enabled = ?", true).
+		Joins("JOIN ai_channels ON ai_channels.id = ai_model_caches.channel_id AND ai_channels.enabled = ?", true)
-		Where("(ai_model_caches.unusable = ? OR ai_model_caches.unusable IS NULL)", false)
 	if group != "" {
 		q = q.Where("ai_channels.channel_group = ?", group)
 	}
@@ -658,7 +538,6 @@ func (s *AiGatewayService) DeprecatingModels(ctx context.Context, within time.Du
 	err := s.db.WithContext(ctx).
 		Where("(retired_at IS NOT NULL AND retired_at > ? AND retired_at <= ?) OR (deprecated_at IS NOT NULL AND deprecated_at >= ? AND deprecated_at <= ?)",
 			now, deadline, now, deadline).
-		Where("(unusable = ? OR unusable IS NULL)", false).
 		Order("name ASC").Find(&rows).Error
 	if err != nil {
 		return nil, err
@@ -679,7 +558,7 @@ func (s *AiGatewayService) DeprecatingModels(ctx context.Context, within time.Du
 	return out, nil
 }
 
-// ProbeAll 逐个探测全部渠道并顺带验证模型池,返回状态汇总;供 AI 探测后台任务调用。
+// ProbeAll 逐个探测全部渠道,返回状态汇总;供 AI 探测后台任务调用。
 func (s *AiGatewayService) ProbeAll(ctx context.Context) (string, error) {
 	chs, err := s.Channels(ctx)
 	if err != nil {
@@ -695,8 +574,6 @@ func (s *AiGatewayService) ProbeAll(ctx context.Context) (string, error) {
 			continue
 		}
 		counts[fresh.ProbeStatus]++
-		// 后台任务里同步执行:新模型验证入池、坏模型定期复检,零人工收敛
-		s.validateChannelModels(ctx, ch.ID)
 	}
 	msg := fmt.Sprintf("probed %d: %d ok, %d no_service, %d no_quota, %d error",
 		len(chs), counts["ok"], counts["no_service"], counts["no_quota"], counts["error"])
@@ -706,6 +583,99 @@ func (s *AiGatewayService) ProbeAll(ctx context.Context) (string, error) {
 	return msg, nil
 }
 
+// ---- 模型黑名单 ----
+
+// Blacklist 列出全部黑名单模型(按名称排序)。
+func (s *AiGatewayService) Blacklist(ctx context.Context) ([]model.AiModelBlacklist, error) {
+	var rows []model.AiModelBlacklist
+	err := s.db.WithContext(ctx).Order("name ASC").Find(&rows).Error
+	return rows, err
+}
+
+// AddBlacklist 把模型名加入黑名单并删除全部渠道缓存中的同名条目;
+// 该模型此后同步 / 探测均被过滤,直到移出黑名单后重新同步。
+func (s *AiGatewayService) AddBlacklist(ctx context.Context, name string) (*model.AiModelBlacklist, error) {
+	name = strings.TrimSpace(name)
+	if name == "" {
+		return nil, fmt.Errorf("模型名不能为空")
+	}
+	row := model.AiModelBlacklist{Name: name}
+	err := s.db.WithContext(ctx).Transaction(func(tx *gorm.DB) error {
+		var n int64
+		if err := tx.Model(&model.AiModelBlacklist{}).Where("name = ?", name).Count(&n).Error; err != nil {
+			return err
+		}
+		if n > 0 {
+			return fmt.Errorf("模型已在黑名单中")
+		}
+		if err := tx.Create(&row).Error; err != nil {
+			return err
+		}
+		return tx.Where("name = ?", name).Delete(&model.AiModelCache{}).Error
+	})
+	if err != nil {
+		return nil, err
+	}
+	return &row, nil
+}
+
+// RemoveBlacklist 把模型移出黑名单;缓存不回填,下次同步 / 探测自然恢复入池。
+func (s *AiGatewayService) RemoveBlacklist(ctx context.Context, id uint) error {
+	res := s.db.WithContext(ctx).Delete(&model.AiModelBlacklist{}, id)
+	if res.Error != nil {
+		return res.Error
+	}
+	if res.RowsAffected == 0 {
+		return fmt.Errorf("黑名单条目不存在")
+	}
+	return nil
+}
+
+// withoutBlacklisted 过滤掉黑名单中的模型,同步与探测入库前统一经此收口。
+// supportedGatewayModels 过滤模型目录:对话模型仅保留实测支持 OpenAI 兼容面的
+// 厂商(xai / meta / openai)——typed chat 面已剔除,google / cohere 对话模型无
+// 上游通路,不入目录(不出现在列表、路由与探测候选);EMBEDDING 模型不受影响。
+func supportedGatewayModels(models []oci.GenAiModel) []oci.GenAiModel {
+	out := make([]oci.GenAiModel, 0, len(models))
+	for _, m := range models {
+		if m.Capability == "CHAT" && !compatChatVendor(m.Name) {
+			continue
+		}
+		out = append(out, m)
+	}
+	return out
+}
+
+func compatChatVendor(model string) bool {
+	for _, prefix := range []string{"xai.", "meta.", "openai."} {
+		if strings.HasPrefix(model, prefix) {
+			return true
+		}
+	}
+	return false
+}
+
+func (s *AiGatewayService) withoutBlacklisted(ctx context.Context, models []oci.GenAiModel) ([]oci.GenAiModel, error) {
+	var names []string
+	if err := s.db.WithContext(ctx).Model(&model.AiModelBlacklist{}).Pluck("name", &names).Error; err != nil {
+		return nil, fmt.Errorf("读取模型黑名单: %w", err)
+	}
+	if len(names) == 0 {
+		return models, nil
+	}
+	banned := make(map[string]bool, len(names))
+	for _, n := range names {
+		banned[n] = true
+	}
+	out := make([]oci.GenAiModel, 0, len(models))
+	for _, m := range models {
+		if !banned[m.Name] {
+			out = append(out, m)
+		}
+	}
+	return out, nil
+}
+
 // ---- 调用日志 ----
 
 // LogCall 落一条调用日志(仅元数据与用量,永不含请求 / 响应正文),返回落库 ID 供内容日志关联(失败为 0)。

internal/service/aigateway_chat.go

@@ -3,6 +3,7 @@ package service
 import (
 	"context"
 	"errors"
+	"io"
 	"math/rand"
 	"time"
 
@@ -26,24 +27,25 @@ type aiCandidate struct {
 	modelOcid string
 }
 
-// Chat 编排非流式调用:选渠道 → 调用 → 可重试错误换渠道(整请求上限 3 次)。
+// RespPassthrough 编排一次非流式直通调用:选渠道(priority→加权随机)→ 调用 →
-// group 非空时只在同分组渠道内路由(取自调用密钥)。
+// 可重试错误换渠道(整请求上限 3 次)并维护熔断;group 非空时只在同分组渠道内路由。
-func (s *AiGatewayService) Chat(ctx context.Context, ir aiwire.ChatRequest, group string) (*aiwire.ChatResponse, ChatMeta, error) {
+// 上游为 OpenAI-compatible /actions/v1/responses(实测可用,无 Oracle 文档合同)。
+func (s *AiGatewayService) RespPassthrough(ctx context.Context, raw []byte, modelName, group string) ([]byte, ChatMeta, error) {
 	meta := ChatMeta{}
 	excluded := map[uint]bool{}
 	var lastErr error
 	for attempt := 0; attempt < 3; attempt++ {
-		cand, err := s.pick(ctx, ir.Model, group, "CHAT", excluded)
+		cand, err := s.pick(ctx, modelName, group, "CHAT", excluded)
 		if err != nil {
 			return nil, meta, firstErr(lastErr, err)
 		}
 		meta.ChannelID, meta.ChannelName = cand.ch.ID, cand.ch.Name
-		resp, err := s.callOnce(ctx, cand, ir)
+		payload, err := s.passthroughOnce(ctx, cand, raw)
 		if err == nil {
 			s.markSuccess(ctx, cand.ch.ID)
-			return resp, meta, nil
+			return payload, meta, nil
 		}
-		retry, penalize := s.noteCallErr(ctx, cand, err)
+		retry, penalize := switchable(err)
 		if !retry {
 			return nil, meta, err
 		}
@@ -57,22 +59,22 @@ func (s *AiGatewayService) Chat(ctx context.Context, ir aiwire.ChatRequest, grou
 	return nil, meta, lastErr
 }
 
-func (s *AiGatewayService) callOnce(ctx context.Context, cand *aiCandidate, ir aiwire.ChatRequest) (*aiwire.ChatResponse, error) {
+func (s *AiGatewayService) passthroughOnce(ctx context.Context, cand *aiCandidate, raw []byte) ([]byte, error) {
 	cred, err := s.configs.credentialsByID(ctx, cand.ch.OciConfigID)
 	if err != nil {
 		return nil, err
 	}
-	return s.client.GenAiChat(ctx, cred, cand.ch.Region, cand.modelOcid, ir)
+	return s.client.GenAiCompatResponses(ctx, cred, cand.ch.Region, raw)
 }
 
-// OpenStream 编排流式调用:流建立成功后即绑定渠道,建立失败可换渠道重试。
+// RespPassthroughStream 编排流式直通:流建立成功即绑定渠道,建立失败按 switchable
-// group 语义与 Chat 相同。
+// 换渠道重试;建立后的中断不重试、不计熔断(与 OpenStream 语义一致)。
-func (s *AiGatewayService) OpenStream(ctx context.Context, ir aiwire.ChatRequest, group string) (oci.GenAiStream, ChatMeta, error) {
+func (s *AiGatewayService) RespPassthroughStream(ctx context.Context, raw []byte, modelName, group string) (io.ReadCloser, ChatMeta, error) {
 	meta := ChatMeta{}
 	excluded := map[uint]bool{}
 	var lastErr error
 	for attempt := 0; attempt < 3; attempt++ {
-		cand, err := s.pick(ctx, ir.Model, group, "CHAT", excluded)
+		cand, err := s.pick(ctx, modelName, group, "CHAT", excluded)
 		if err != nil {
 			return nil, meta, firstErr(lastErr, err)
 		}
@@ -81,12 +83,12 @@ func (s *AiGatewayService) OpenStream(ctx context.Context, ir aiwire.ChatRequest
 		if err != nil {
 			return nil, meta, err
 		}
-		stream, err := s.client.GenAiChatStream(ctx, cred, cand.ch.Region, cand.modelOcid, ir)
+		stream, err := s.client.GenAiCompatResponsesStream(ctx, cred, cand.ch.Region, raw)
 		if err == nil {
 			s.markSuccess(ctx, cand.ch.ID)
 			return stream, meta, nil
 		}
-		retry, penalize := s.noteCallErr(ctx, cand, err)
+		retry, penalize := switchable(err)
 		if !retry {
 			return nil, meta, err
 		}
@@ -108,12 +110,11 @@ func firstErr(lastErr, pickErr error) error {
 	return pickErr
 }
 
-// noteCallErr 汇总一次调用失败:模型级不可用(微调基座 400 / 实体不存在 404)先标记
+// switchable 判断调用失败是否换渠道重试、是否计入熔断:模型级不可用(微调基座
-// 剔除该 (渠道, 模型),换渠道重试且不计熔断——这是模型×区域供给问题而非渠道健康问题;
+// 400 / 实体不存在 404)换渠道但不计熔断——这是模型×区域供给问题而非渠道健康
-// 429 / 5xx / 网络错误换渠道并计熔断;其余 4xx 直接透传。
+// 问题,持久解决靠加入模型黑名单;429 / 5xx / 网络错误换渠道并计熔断;其余 4xx 直接透传。
-func (s *AiGatewayService) noteCallErr(ctx context.Context, cand *aiCandidate, err error) (retry, penalize bool) {
+func switchable(err error) (retry, penalize bool) {
 	if oci.IsModelUnavailable(err) {
-		s.markModelUnusable(ctx, cand.ch.ID, cand.modelOcid, oci.CompactError(err))
 		return true, false
 	}
 	if status, ok := oci.ServiceStatus(err); ok {
@@ -154,11 +155,10 @@ func (s *AiGatewayService) pick(ctx context.Context, modelName, group, capabilit
 	return &aiCandidate{ch: chosen, modelOcid: ocids[chosen.ID]}, nil
 }
 
-// modelChannels 查出提供该模型的渠道 ID 及各自的模型 OCID(剔除不可按需调用标记);
+// modelChannels 查出提供该模型的渠道 ID 及各自的模型 OCID;
-// capability=CHAT 时兼容存量空串(加列前只同步对话模型);unusable 需兼容 NULL
+// capability=CHAT 时兼容存量空串(加列前只同步对话模型)。
-// (AutoMigrate 加列后、首次重同步前的存量行)。
 func (s *AiGatewayService) modelChannels(ctx context.Context, modelName, capability string) (map[uint]string, []uint, error) {
-	q := s.db.WithContext(ctx).Where("name = ? AND (unusable = ? OR unusable IS NULL)", modelName, false)
+	q := s.db.WithContext(ctx).Where("name = ?", modelName)
 	if capability == "CHAT" {
 		q = q.Where("capability IN ?", []string{"CHAT", ""})
 	} else {
@@ -234,7 +234,7 @@ func (s *AiGatewayService) Embeddings(ctx context.Context, req aiwire.Embeddings
 			s.markSuccess(ctx, cand.ch.ID)
 			return resp, meta, nil
 		}
-		retry, penalize := s.noteCallErr(ctx, cand, err)
+		retry, penalize := switchable(err)
 		if !retry {
 			return nil, meta, err
 		}

internal/service/aigateway_models_test.go

@@ -0,0 +1,29 @@
+package service
+
+import (
+	"strings"
+	"testing"
+
+	"oci-portal/internal/oci"
+)
+
+// TestSupportedGatewayModels 断言 CHAT 目录仅保留兼容面厂商,EMBEDDING 不受影响。
+func TestSupportedGatewayModels(t *testing.T) {
+	in := []oci.GenAiModel{
+		{Name: "xai.grok-4.3", Capability: "CHAT"},
+		{Name: "meta.llama-4-maverick-17b-128e", Capability: "CHAT"},
+		{Name: "openai.gpt-oss-120b", Capability: "CHAT"},
+		{Name: "google.gemini-2.5-flash", Capability: "CHAT"},
+		{Name: "cohere.command-a-03-2025", Capability: "CHAT"},
+		{Name: "cohere.embed-v4.0", Capability: "EMBEDDING"},
+	}
+	out := supportedGatewayModels(in)
+	names := make([]string, 0, len(out))
+	for _, m := range out {
+		names = append(names, m.Name)
+	}
+	want := "xai.grok-4.3,meta.llama-4-maverick-17b-128e,openai.gpt-oss-120b,cohere.embed-v4.0"
+	if got := strings.Join(names, ","); got != want {
+		t.Fatalf("过滤结果 = %s, want %s", got, want)
+	}
+}

internal/service/aigateway_test.go

@@ -1,9 +1,11 @@
 package service
 
 import (
+	"bytes"
 	"context"
 	"errors"
 	"fmt"
+	"io"
 	"strings"
 	"testing"
 	"time"
@@ -46,14 +48,45 @@ type gatewayStubClient struct {
 	probeErr  error
 	// probeSeq 非空时逐次弹出,模拟按候选依次试调;弹尽后回落 probeCode/probeErr
 	probeSeq []probeResult
-	chatResp  *aiwire.ChatResponse
+	// probedNames 记录试调过的模型名,供断言候选过滤
-	chatErrs  []error
+	probedNames []string
 	chatCalls   int
 	regions     []string
 
 	embedVecs  [][]float32
 	embedUsage *aiwire.Usage
 	embedErr   error
+
+	passPayload []byte
+	passErrs    []error
+	passCalls   int
+	passRegions []string
+}
+
+func (f *gatewayStubClient) GenAiCompatResponses(ctx context.Context, cred oci.Credentials, region string, body []byte) ([]byte, error) {
+	f.passCalls++
+	f.passRegions = append(f.passRegions, region)
+	if len(f.passErrs) > 0 {
+		err := f.passErrs[0]
+		f.passErrs = f.passErrs[1:]
+		if err != nil {
+			return nil, err
+		}
+	}
+	return f.passPayload, nil
+}
+
+func (f *gatewayStubClient) GenAiCompatResponsesStream(ctx context.Context, cred oci.Credentials, region string, body []byte) (io.ReadCloser, error) {
+	f.passCalls++
+	f.passRegions = append(f.passRegions, region)
+	if len(f.passErrs) > 0 {
+		err := f.passErrs[0]
+		f.passErrs = f.passErrs[1:]
+		if err != nil {
+			return nil, err
+		}
+	}
+	return io.NopCloser(bytes.NewReader(f.passPayload)), nil
 }
 
 func (f *gatewayStubClient) GenAiEmbed(ctx context.Context, cred oci.Credentials, region, modelOcid string, inputs []string, dimensions *int) ([][]float32, *aiwire.Usage, error) {
@@ -71,6 +104,7 @@ type probeResult struct {
 }
 
 func (f *gatewayStubClient) GenAiProbeChat(ctx context.Context, cred oci.Credentials, region, modelOcid, modelName string) (int, error) {
+	f.probedNames = append(f.probedNames, modelName)
 	if len(f.probeSeq) > 0 {
 		r := f.probeSeq[0]
 		f.probeSeq = f.probeSeq[1:]
@@ -79,23 +113,10 @@ func (f *gatewayStubClient) GenAiProbeChat(ctx context.Context, cred oci.Credent
 	return f.probeCode, f.probeErr
 }
 
-func (f *gatewayStubClient) GenAiChat(ctx context.Context, cred oci.Credentials, region, modelOcid string, ir aiwire.ChatRequest) (*aiwire.ChatResponse, error) {
-	f.chatCalls++
-	f.regions = append(f.regions, region)
-	if len(f.chatErrs) > 0 {
-		err := f.chatErrs[0]
-		f.chatErrs = f.chatErrs[1:]
-		if err != nil {
-			return nil, err
-		}
-	}
-	return f.chatResp, nil
-}
-
 func newTestGateway(t *testing.T, client oci.Client) (*AiGatewayService, *OciConfigService) {
 	t.Helper()
 	svc := newTestService(t, client)
-	if err := svc.db.AutoMigrate(&model.AiKey{}, &model.AiChannel{}, &model.AiModelCache{}, &model.AiCallLog{}); err != nil {
+	if err := svc.db.AutoMigrate(&model.AiKey{}, &model.AiChannel{}, &model.AiModelCache{}, &model.AiModelBlacklist{}, &model.AiCallLog{}); err != nil {
 		t.Fatalf("auto migrate ai tables: %v", err)
 	}
 	return NewAiGatewayService(svc.db, svc, client), svc
@@ -105,14 +126,14 @@ func TestAiKeyLifecycle(t *testing.T) {
 	gw, _ := newTestGateway(t, &gatewayStubClient{fakeClient: &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}}})
 	ctx := context.Background()
 
-	raw, key, err := gw.CreateKey(ctx, "免费-ai-api-key", "test-api-key", "")
+	raw, key, err := gw.CreateKey(ctx, "免费-ai-api-key", "test-api-key", "", nil)
 	if err != nil || raw != "test-api-key" || key.Tail != "-key" {
 		t.Fatalf("CreateKey 自定义值 = %q %+v, %v", raw, key, err)
 	}
-	if _, _, err := gw.CreateKey(ctx, "short", "abc", ""); err == nil {
+	if _, _, err := gw.CreateKey(ctx, "short", "abc", "", nil); err == nil {
 		t.Error("过短自定义密钥应被拒绝")
 	}
-	auto, _, err := gw.CreateKey(ctx, "auto", "", "")
+	auto, _, err := gw.CreateKey(ctx, "auto", "", "", nil)
 	if err != nil || len(auto) < 40 || auto[:3] != "sk-" {
 		t.Fatalf("CreateKey 随机值 = %q, %v", auto, err)
 	}
@@ -124,7 +145,7 @@ func TestAiKeyLifecycle(t *testing.T) {
 		t.Errorf("错误密钥 err = %v", err)
 	}
 	off := false
-	_ = gw.UpdateKey(ctx, key.ID, "", &off, nil)
+	_ = gw.UpdateKey(ctx, key.ID, "", &off, nil, nil)
 	if _, err := gw.VerifyKey(ctx, "test-api-key"); !errors.Is(err, ErrAiKeyInvalid) {
 		t.Errorf("禁用后 VerifyKey err = %v", err)
 	}
@@ -187,51 +208,6 @@ func seedChannel(t *testing.T, gw *AiGatewayService, cfgID uint, region string,
 	return ch
 }
 
-func TestAiChatRetrySwitchesChannel(t *testing.T) {
-	client := &gatewayStubClient{
-		fakeClient: &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}},
-		chatResp:   &aiwire.ChatResponse{Model: "meta.llama-3.3-70b-instruct", Choices: []aiwire.Choice{{Message: aiwire.ChatMessage{Role: "assistant", Content: aiwire.NewTextContent("hi")}, FinishReason: "stop"}}},
-		chatErrs:   []error{stubServiceError{status: 429}},
-	}
-	gw, svc := newTestGateway(t, client)
-	cfg := importAliveConfig(t, svc)
-	// 两个同优先级渠道
-	seedChannel(t, gw, cfg.ID, "eu-frankfurt-1", 1, 1)
-	seedChannel(t, gw, cfg.ID, "us-chicago-1", 1, 1)
-	ctx := context.Background()
-
-	resp, meta, err := gw.Chat(ctx, aiwire.ChatRequest{Model: "meta.llama-3.3-70b-instruct", Messages: []aiwire.ChatMessage{{Role: "user", Content: aiwire.NewTextContent("你好")}}}, "")
-	if err != nil || resp == nil {
-		t.Fatalf("Chat = %v, %v", resp, err)
-	}
-	if meta.Retries != 1 || client.chatCalls != 2 {
-		t.Errorf("应换渠道重试一次: retries=%d calls=%d", meta.Retries, client.chatCalls)
-	}
-	if len(client.regions) != 2 && client.regions[0] == client.regions[1] {
-		t.Errorf("重试未换渠道: %v", client.regions)
-	}
-	// 未知模型
-	if _, _, err := gw.Chat(ctx, aiwire.ChatRequest{Model: "no-such-model"}, ""); !errors.Is(err, ErrAiUnknownModel) {
-		t.Errorf("未知模型 err = %v", err)
-	}
-}
-
-func TestAiChatNonRetryablePassThrough(t *testing.T) {
-	client := &gatewayStubClient{
-		fakeClient: &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}},
-		chatErrs:   []error{stubServiceError{status: 400}},
-	}
-	gw, svc := newTestGateway(t, client)
-	cfg := importAliveConfig(t, svc)
-	seedChannel(t, gw, cfg.ID, "eu-frankfurt-1", 1, 1)
-	seedChannel(t, gw, cfg.ID, "us-chicago-1", 1, 1)
-
-	_, meta, err := gw.Chat(context.Background(), aiwire.ChatRequest{Model: "meta.llama-3.3-70b-instruct"}, "")
-	if err == nil || meta.Retries != 0 || client.chatCalls != 1 {
-		t.Errorf("400 应直接透传不重试: err=%v retries=%d calls=%d", err, meta.Retries, client.chatCalls)
-	}
-}
-
 func TestPickPriorityAndBreaker(t *testing.T) {
 	gw, svc := newTestGateway(t, &gatewayStubClient{fakeClient: &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}}})
 	cfg := importAliveConfig(t, svc)
@@ -283,7 +259,7 @@ func TestMarkFailureBackoff(t *testing.T) {
 func TestAiGroupRouting(t *testing.T) {
 	client := &gatewayStubClient{
 		fakeClient:  &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}},
-		chatResp:   &aiwire.ChatResponse{Model: "meta.llama-3.3-70b-instruct", Choices: []aiwire.Choice{{Message: aiwire.ChatMessage{Role: "assistant", Content: aiwire.NewTextContent("hi")}, FinishReason: "stop"}}},
+		passPayload: []byte(`{"id":"resp_1","usage":{"input_tokens":3,"output_tokens":1}}`),
 	}
 	gw, svc := newTestGateway(t, client)
 	cfg := importAliveConfig(t, svc)
@@ -292,17 +268,17 @@ func TestAiGroupRouting(t *testing.T) {
 	gw.db.Model(vip).Update("channel_group", "vip")
 	gw.db.Create(&model.AiModelCache{ChannelID: other.ID, ModelOcid: "ocid1..cr", Name: "cohere.command-r", Vendor: "cohere", SyncedAt: time.Now()})
 	ctx := context.Background()
-	req := aiwire.ChatRequest{Model: "meta.llama-3.3-70b-instruct", Messages: []aiwire.ChatMessage{{Role: "user", Content: aiwire.NewTextContent("你好")}}}
+	body := []byte(`{"model":"meta.llama-3.3-70b-instruct","input":"你好"}`)
 
 	// 分组密钥只落同分组渠道
 	for i := 0; i < 5; i++ {
-		_, meta, err := gw.Chat(ctx, req, "vip")
+		_, meta, err := gw.RespPassthrough(ctx, body, "meta.llama-3.3-70b-instruct", "vip")
 		if err != nil || meta.ChannelID != vip.ID {
 			t.Fatalf("vip 组第 %d 次落点 = %d, err %v, want %d", i, meta.ChannelID, err, vip.ID)
 		}
 	}
 	// 分组内无渠道 → 无可用渠道
-	if _, _, err := gw.Chat(ctx, req, "nope"); !errors.Is(err, ErrAiNoChannel) {
+	if _, _, err := gw.RespPassthrough(ctx, body, "meta.llama-3.3-70b-instruct", "nope"); !errors.Is(err, ErrAiNoChannel) {
 		t.Errorf("空分组 err = %v, want ErrAiNoChannel", err)
 	}
 	// 模型列表按分组过滤
@@ -315,12 +291,12 @@ func TestAiGroupRouting(t *testing.T) {
 		t.Errorf("不限组模型数 = %d, want 2", len(all.Data))
 	}
 	// 密钥分组落库
-	_, key, err := gw.CreateKey(ctx, "vip-key", "vip-secret-1234", "vip")
+	_, key, err := gw.CreateKey(ctx, "vip-key", "vip-secret-1234", "vip", nil)
 	if err != nil || key.Group != "vip" {
 		t.Fatalf("CreateKey group = %+v, %v", key, err)
 	}
 	empty := ""
-	_ = gw.UpdateKey(ctx, key.ID, "", nil, &empty)
+	_ = gw.UpdateKey(ctx, key.ID, "", nil, &empty, nil)
 	var fresh model.AiKey
 	gw.db.First(&fresh, key.ID)
 	if fresh.Group != "" {
@@ -465,7 +441,8 @@ func entityNotFoundErr() stubServiceError {
 }
 
 func TestProbeSkipsUnavailableModels(t *testing.T) {
-	// 首选 llama 不可按需调用(基座 400 / 实体 404)→ 标记剔除 → 换候选成功 → 渠道判可用
+	// 首选 llama 不可按需调用(基座 400 / 实体 404)→ 换候选成功 → 渠道判可用;
+	// 模型全量入库不再自动标记,持久剔除交由用户手动拉黑
 	tests := []struct {
 		name string
 		bad  probeResult
@@ -496,112 +473,16 @@ func TestProbeSkipsUnavailableModels(t *testing.T) {
 			if err != nil || probed.ProbeStatus != "ok" {
 				t.Fatalf("坏候选后应换候选并判可用: %+v, %v", probed, err)
 			}
-			var row model.AiModelCache
+			rows, _ := gw.channelModels(ctx, ch.ID)
-			if err := gw.db.Where("channel_id = ? AND model_ocid = ?", ch.ID, "m1").First(&row).Error; err != nil || !row.Unusable {
+			if len(rows) != 3 {
-				t.Fatalf("首个坏候选应被标记不可用: %+v, %v", row, err)
+				t.Errorf("模型应全量入库, got %d", len(rows))
-			}
-			// 再次探测:同步保留标记,m1 不再进候选(probeSeq 只需一次 200)
-			client.probeSeq = []probeResult{{200, nil}}
-			probed, err = gw.ProbeChannel(ctx, ch.ID)
-			if err != nil || probed.ProbeStatus != "ok" {
-				t.Fatalf("复测应跳过已标记模型: %+v, %v", probed, err)
-			}
-			var again model.AiModelCache
-			gw.db.Where("channel_id = ? AND model_ocid = ?", ch.ID, "m1").First(&again)
-			if !again.Unusable {
-				t.Error("探测触发的同步应保留不可用标记")
-			}
-			// 手动同步同样保留标记(坏模型不随重新同步复活);近期已检的标记不被后台验证翻转
-			client.probeCode = 200
-			if _, err := gw.SyncModels(ctx, ch.ID); err != nil {
-				t.Fatalf("SyncModels: %v", err)
-			}
-			gw.Wait()
-			var kept model.AiModelCache
-			gw.db.Where("channel_id = ? AND model_ocid = ?", ch.ID, "m1").First(&kept)
-			if !kept.Unusable {
-				t.Error("手动同步不应清除不可用标记")
 			}
 		})
 	}
 }
 
-func TestValidateModelsAfterSync(t *testing.T) {
-	// 同步后后台验证:坏模型标记剔除、好模型记录已检、其他 4xx 不改状态、非对话模型不试调
-	client := &gatewayStubClient{
-		fakeClient: &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}},
-		models: []oci.GenAiModel{
-			{Ocid: "v1", Name: "meta.llama-4-maverick", Vendor: "meta", Capability: "CHAT"},
-			{Ocid: "v2", Name: "xai.grok-4", Vendor: "xai", Capability: "CHAT"},
-			{Ocid: "v3", Name: "xai.grok-voice-agent", Vendor: "xai", Capability: "CHAT"},
-			{Ocid: "v4", Name: "cohere.embed-v4.0", Vendor: "cohere", Capability: "EMBEDDING"},
-		},
-		probeSeq: []probeResult{{404, entityNotFoundErr()}, {200, nil}, {400, stubServiceError{status: 400}}},
-	}
-	gw, svc := newTestGateway(t, client)
-	cfg := importAliveConfig(t, svc)
-	ctx := context.Background()
-	ch, err := gw.CreateChannel(ctx, ChannelInput{OciConfigID: cfg.ID, Region: "us-ashburn-1"})
-	if err != nil {
-		t.Fatalf("CreateChannel: %v", err)
-	}
-	if _, err := gw.SyncModels(ctx, ch.ID); err != nil {
-		t.Fatalf("SyncModels: %v", err)
-	}
-	gw.Wait()
-
-	want := map[string]struct {
-		unusable bool
-		checked  bool
-	}{
-		"v1": {true, true},   // 实体 404 → 标记剔除
-		"v2": {false, true},  // 200 → 可用已检
-		"v3": {false, true},  // 普通 400 → 已检不标记
-		"v4": {false, false}, // EMBEDDING 不试调
-	}
-	rows, _ := gw.channelModels(ctx, ch.ID)
-	for _, r := range rows {
-		w := want[r.ModelOcid]
-		if r.Unusable != w.unusable || (r.CheckedAt != nil) != w.checked {
-			t.Errorf("%s: unusable=%v checked=%v, want %+v", r.ModelOcid, r.Unusable, r.CheckedAt != nil, w)
-		}
-	}
-	list, err := gw.GatewayModels(ctx, "")
-	if err != nil || len(list.Data) != 3 {
-		t.Errorf("网关列表应只剔除被标记的坏模型(余 v2/v3/v4), got %+v, %v", list.Data, err)
-	}
-}
-
-func TestValidateRecheckUnmarksRecovered(t *testing.T) {
-	// 已标记模型超过复检间隔后重验:恢复供给(200)自动解除标记
-	client := &gatewayStubClient{fakeClient: &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}}, probeCode: 200}
-	gw, svc := newTestGateway(t, client)
-	cfg := importAliveConfig(t, svc)
-	ch := seedChannel(t, gw, cfg.ID, "eu-frankfurt-1", 1, 1)
-	ctx := context.Background()
-	old := time.Now().Add(-25 * time.Hour)
-	gw.db.Model(&model.AiModelCache{}).Where("channel_id = ?", ch.ID).
-		Updates(map[string]any{"unusable": true, "unusable_reason": "x", "checked_at": old})
-
-	gw.validateChannelModels(ctx, ch.ID)
-	var row model.AiModelCache
-	gw.db.Where("channel_id = ?", ch.ID).First(&row)
-	if row.Unusable || row.UnusableReason != "" || row.CheckedAt == nil || !row.CheckedAt.After(old) {
-		t.Errorf("超期复检应解除标记并刷新已检时间: %+v", row)
-	}
-	// 未超期的标记不复检(probeSeq 为空、fallback 200 也不会被消费)
-	fresh := time.Now()
-	gw.db.Model(&model.AiModelCache{}).Where("channel_id = ?", ch.ID).
-		Updates(map[string]any{"unusable": true, "checked_at": fresh})
-	gw.validateChannelModels(ctx, ch.ID)
-	gw.db.Where("channel_id = ?", ch.ID).First(&row)
-	if !row.Unusable {
-		t.Error("未超期的标记不应被复检翻转")
-	}
-}
-
 func TestProbeAuth404StillNoQuota(t *testing.T) {
-	// 鉴权类 404(NotAuthorizedOrNotFound)仍属租户级,直接定论 no_quota 且不标记模型
+	// 鉴权类 404(NotAuthorizedOrNotFound)仍属租户级,直接定论 no_quota
 	client := &gatewayStubClient{
 		fakeClient: &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}},
 		models:     []oci.GenAiModel{{Ocid: "m1", Name: "meta.llama-3.3-70b-instruct", Vendor: "meta"}},
@@ -617,31 +498,26 @@ func TestProbeAuth404StillNoQuota(t *testing.T) {
 	if probed.ProbeStatus != "no_quota" {
 		t.Errorf("鉴权 404 status = %q, want no_quota", probed.ProbeStatus)
 	}
-	var marked int64
-	gw.db.Model(&model.AiModelCache{}).Where("unusable = ?", true).Count(&marked)
-	if marked != 0 {
-		t.Errorf("鉴权 404 不应标记模型, marked=%d", marked)
-	}
 }
 
 func TestAiChatFinetuneSwitchesChannelWithoutPenalty(t *testing.T) {
 	// 微调基座 400 换渠道重试成功,且不计入熔断失败
 	client := &gatewayStubClient{
 		fakeClient:  &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}},
-		chatResp:   &aiwire.ChatResponse{Model: "meta.llama-3.3-70b-instruct", Choices: []aiwire.Choice{{Message: aiwire.ChatMessage{Role: "assistant", Content: aiwire.NewTextContent("hi")}, FinishReason: "stop"}}},
+		passPayload: []byte(`{"id":"resp_1","usage":{"input_tokens":3,"output_tokens":1}}`),
-		chatErrs:   []error{finetuneBaseErr()},
+		passErrs:    []error{finetuneBaseErr()},
 	}
 	gw, svc := newTestGateway(t, client)
 	cfg := importAliveConfig(t, svc)
 	seedChannel(t, gw, cfg.ID, "eu-frankfurt-1", 1, 1)
 	seedChannel(t, gw, cfg.ID, "us-chicago-1", 1, 1)
 
-	resp, meta, err := gw.Chat(context.Background(), aiwire.ChatRequest{Model: "meta.llama-3.3-70b-instruct", Messages: []aiwire.ChatMessage{{Role: "user", Content: aiwire.NewTextContent("你好")}}}, "")
+	resp, meta, err := gw.RespPassthrough(context.Background(), []byte(`{}`), "meta.llama-3.3-70b-instruct", "")
 	if err != nil || resp == nil {
-		t.Fatalf("Chat = %v, %v", resp, err)
+		t.Fatalf("RespPassthrough = %v, %v", resp, err)
 	}
-	if meta.Retries != 1 || client.chatCalls != 2 {
+	if meta.Retries != 1 || client.passCalls != 2 {
-		t.Errorf("应换渠道重试一次: retries=%d calls=%d", meta.Retries, client.chatCalls)
+		t.Errorf("应换渠道重试一次: retries=%d calls=%d", meta.Retries, client.passCalls)
 	}
 	var chs []model.AiChannel
 	gw.db.Find(&chs)
@@ -650,38 +526,93 @@ func TestAiChatFinetuneSwitchesChannelWithoutPenalty(t *testing.T) {
 			t.Errorf("微调基座 400 不应计入熔断: 渠道 %s failCount=%d", ch.Name, ch.FailCount)
 		}
 	}
-	// 失败渠道的该模型被标记,不再参与路由;成功渠道不受影响
-	region := client.regions[0]
-	var row model.AiModelCache
-	gw.db.Where("model_ocid = ?", "ocid1..m-"+region).First(&row)
-	if !row.Unusable {
-		t.Errorf("失败渠道的模型应被标记不可用: %+v", row)
-	}
-	var usable int64
-	gw.db.Model(&model.AiModelCache{}).Where("unusable = ?", false).Count(&usable)
-	if usable != 1 {
-		t.Errorf("成功渠道模型应保持可用, usable=%d", usable)
-	}
 }
 
-func TestUnusableModelExcludedFromPoolAndRouting(t *testing.T) {
+func TestBlacklistLifecycle(t *testing.T) {
-	// 唯一渠道的模型被标记后:网关列表不再展示,路由按未知模型拒绝,渠道详情仍可见标记
+	// 拉黑:全渠道同名缓存删除、列表与路由立即不可见;重复/空名被拒;
-	gw, svc := newTestGateway(t, &gatewayStubClient{fakeClient: &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}}})
+	// 移出黑名单后重新同步恢复入池
+	client := &gatewayStubClient{
+		fakeClient: &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}},
+		models:     []oci.GenAiModel{{Ocid: "ocid1..m-eu-frankfurt-1", Name: "meta.llama-3.3-70b-instruct", Vendor: "meta"}},
+	}
+	gw, svc := newTestGateway(t, client)
 	cfg := importAliveConfig(t, svc)
 	ch := seedChannel(t, gw, cfg.ID, "eu-frankfurt-1", 1, 1)
+	seedChannel(t, gw, cfg.ID, "us-chicago-1", 1, 1)
 	ctx := context.Background()
 
-	gw.markModelUnusable(ctx, ch.ID, "ocid1..m-eu-frankfurt-1", "Entity with key … not found")
+	item, err := gw.AddBlacklist(ctx, " meta.llama-3.3-70b-instruct ")
-	list, err := gw.GatewayModels(ctx, "")
+	if err != nil || item.Name != "meta.llama-3.3-70b-instruct" {
-	if err != nil || len(list.Data) != 0 {
+		t.Fatalf("AddBlacklist = %+v, %v", item, err)
-		t.Errorf("已标记模型不应出现在网关列表: %+v, %v", list.Data, err)
 	}
-	if _, _, err := gw.Chat(ctx, aiwire.ChatRequest{Model: "meta.llama-3.3-70b-instruct"}, ""); !errors.Is(err, ErrAiUnknownModel) {
+	var left int64
-		t.Errorf("已标记模型路由应拒绝: %v", err)
+	gw.db.Model(&model.AiModelCache{}).Count(&left)
+	if left != 0 {
+		t.Errorf("拉黑应删除全部渠道的同名缓存, 剩 %d", left)
 	}
-	rows, err := gw.channelModels(ctx, ch.ID)
+	if _, _, err := gw.RespPassthrough(ctx, []byte(`{}`), "meta.llama-3.3-70b-instruct", ""); !errors.Is(err, ErrAiUnknownModel) {
-	if err != nil || len(rows) != 1 || !rows[0].Unusable || rows[0].UnusableReason == "" {
+		t.Errorf("拉黑后路由应按未知模型拒绝: %v", err)
-		t.Errorf("渠道详情应保留标记行便于排查: %+v, %v", rows, err)
+	}
+	if _, err := gw.AddBlacklist(ctx, "meta.llama-3.3-70b-instruct"); err == nil {
+		t.Error("重复拉黑应被拒绝")
+	}
+	if _, err := gw.AddBlacklist(ctx, "  "); err == nil {
+		t.Error("空模型名应被拒绝")
+	}
+	rows, err := gw.Blacklist(ctx)
+	if err != nil || len(rows) != 1 {
+		t.Fatalf("Blacklist = %+v, %v", rows, err)
+	}
+	if err := gw.RemoveBlacklist(ctx, rows[0].ID); err != nil {
+		t.Fatalf("RemoveBlacklist: %v", err)
+	}
+	if err := gw.RemoveBlacklist(ctx, rows[0].ID); err == nil {
+		t.Error("移除不存在的条目应报错")
+	}
+	// 移出后重新同步:模型恢复入池
+	if _, err := gw.SyncModels(ctx, ch.ID); err != nil {
+		t.Fatalf("SyncModels: %v", err)
+	}
+	list, _ := gw.GatewayModels(ctx, "")
+	if len(list.Data) != 1 {
+		t.Errorf("移出黑名单并同步后应恢复入池: %+v", list.Data)
+	}
+}
+
+func TestSyncAndProbeFilterBlacklisted(t *testing.T) {
+	// 同步与探测都过滤黑名单模型:不入库、不进试调候选
+	client := &gatewayStubClient{
+		fakeClient: &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}},
+		models: []oci.GenAiModel{
+			{Ocid: "m1", Name: "meta.llama-3-70b-instruct", Vendor: "meta"},
+			{Ocid: "m2", Name: "cohere.command-a-03-2025", Vendor: "cohere"},
+		},
+		probeCode: 200,
+	}
+	gw, svc := newTestGateway(t, client)
+	cfg := importAliveConfig(t, svc)
+	ctx := context.Background()
+	if _, err := gw.AddBlacklist(ctx, "meta.llama-3-70b-instruct"); err != nil {
+		t.Fatalf("AddBlacklist: %v", err)
+	}
+	ch, err := gw.CreateChannel(ctx, ChannelInput{OciConfigID: cfg.ID, Region: "eu-frankfurt-1"})
+	if err != nil {
+		t.Fatalf("CreateChannel: %v", err)
+	}
+	probed, err := gw.ProbeChannel(ctx, ch.ID)
+	if err != nil || probed.ProbeStatus != "ok" {
+		t.Fatalf("ProbeChannel = %+v, %v", probed, err)
+	}
+	rows, _ := gw.channelModels(ctx, ch.ID)
+	if len(rows) != 1 || rows[0].Name != "cohere.command-a-03-2025" {
+		t.Errorf("探测同步应过滤黑名单模型: %+v", rows)
+	}
+	if len(client.probedNames) != 1 || client.probedNames[0] != "cohere.command-a-03-2025" {
+		t.Errorf("试调候选不应包含黑名单模型: %v", client.probedNames)
+	}
+	models, err := gw.SyncModels(ctx, ch.ID)
+	if err != nil || len(models) != 1 || models[0].Name != "cohere.command-a-03-2025" {
+		t.Errorf("SyncModels 应过滤黑名单模型: %+v, %v", models, err)
 	}
 }
 
@@ -745,7 +676,7 @@ func TestAiEmbeddings(t *testing.T) {
 		t.Errorf("chat 模型走 embeddings err = %v, want ErrAiUnknownModel", err)
 	}
 	// embedding 模型名打 chat:同样未知模型
-	if _, _, err := gw.Chat(ctx, aiwire.ChatRequest{Model: "cohere.embed-v4.0", Messages: []aiwire.ChatMessage{{Role: "user", Content: aiwire.NewTextContent("x")}}}, ""); !errors.Is(err, ErrAiUnknownModel) {
+	if _, _, err := gw.RespPassthrough(ctx, []byte(`{}`), "cohere.embed-v4.0", ""); !errors.Is(err, ErrAiUnknownModel) {
 		t.Errorf("embedding 模型走 chat err = %v, want ErrAiUnknownModel", err)
 	}
 }
@@ -756,7 +687,7 @@ func TestAiContentLogSwitch(t *testing.T) {
 		t.Fatalf("migrate content log: %v", err)
 	}
 	ctx := context.Background()
-	_, key, err := gw.CreateKey(ctx, "k1", "content-key-1234", "")
+	_, key, err := gw.CreateKey(ctx, "k1", "content-key-1234", "", nil)
 	if err != nil {
 		t.Fatalf("CreateKey: %v", err)
 	}
@@ -793,23 +724,130 @@ func TestAiContentLogSwitch(t *testing.T) {
 	}
 }
 
-func TestLegacyNullUnusableRowsStayVisible(t *testing.T) {
+func TestRespPassthroughSwitchesChannel(t *testing.T) {
-	// 升级路径回归:AutoMigrate 加列后、首次重同步前,存量行 unusable 为 NULL,
+	client := &gatewayStubClient{
-	// 网关列表与路由必须照常包含这些行,不能因 unusable = false 过滤而整池消失
+		fakeClient:  &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}},
-	gw, svc := newTestGateway(t, &gatewayStubClient{fakeClient: &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}}})
+		passPayload: []byte(`{"id":"resp_1","usage":{"input_tokens":9,"output_tokens":3}}`),
+		passErrs:    []error{stubServiceError{status: 429}},
+	}
+	gw, svc := newTestGateway(t, client)
 	cfg := importAliveConfig(t, svc)
-	ch := seedChannel(t, gw, cfg.ID, "eu-frankfurt-1", 1, 1)
+	seedChannel(t, gw, cfg.ID, "eu-frankfurt-1", 1, 1)
-	ctx := context.Background()
+	seedChannel(t, gw, cfg.ID, "us-chicago-1", 1, 1)
-	if err := gw.db.Exec("UPDATE ai_model_caches SET unusable = NULL WHERE channel_id = ?", ch.ID).Error; err != nil {
+
-		t.Fatalf("set legacy null: %v", err)
+	payload, meta, err := gw.RespPassthrough(context.Background(), []byte(`{}`), "meta.llama-3.3-70b-instruct", "")
+	if err != nil || len(payload) == 0 {
+		t.Fatalf("RespPassthrough = %v, %v", payload, err)
+	}
+	if meta.Retries != 1 || client.passCalls != 2 {
+		t.Errorf("应换渠道重试一次: retries=%d calls=%d", meta.Retries, client.passCalls)
+	}
+	if len(client.passRegions) == 2 && client.passRegions[0] == client.passRegions[1] {
+		t.Errorf("重试未换渠道: %v", client.passRegions)
+	}
 }
 
-	list, err := gw.GatewayModels(ctx, "")
+func TestRespPassthroughNonRetryable(t *testing.T) {
-	if err != nil || len(list.Data) != 1 {
+	client := &gatewayStubClient{
-		t.Errorf("存量 NULL 行应仍在网关列表: %+v, %v", list.Data, err)
+		fakeClient: &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}},
+		passErrs:   []error{stubServiceError{status: 400}},
 	}
-	_, ids, err := gw.modelChannels(ctx, "meta.llama-3.3-70b-instruct", "CHAT")
+	gw, svc := newTestGateway(t, client)
-	if err != nil || len(ids) != 1 {
+	cfg := importAliveConfig(t, svc)
-		t.Errorf("存量 NULL 行应仍参与路由: %v, %v", ids, err)
+	seedChannel(t, gw, cfg.ID, "eu-frankfurt-1", 1, 1)
+	seedChannel(t, gw, cfg.ID, "us-chicago-1", 1, 1)
+
+	_, meta, err := gw.RespPassthrough(context.Background(), []byte(`{}`), "meta.llama-3.3-70b-instruct", "")
+	if err == nil || meta.Retries != 0 || client.passCalls != 1 {
+		t.Errorf("400 应直接透传不重试: err=%v retries=%d calls=%d", err, meta.Retries, client.passCalls)
+	}
+}
+
+func TestNormalizeKeyModels(t *testing.T) {
+	tests := []struct {
+		name string
+		in   []string
+		want []string
+	}{
+		{"nil 输入", nil, nil},
+		{"全空串", []string{"", "  "}, nil},
+		{"trim 与保序去重", []string{" a ", "b", "a", "", "b "}, []string{"a", "b"}},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got := normalizeKeyModels(tt.in)
+			if len(got) != len(tt.want) {
+				t.Fatalf("normalizeKeyModels(%v) = %v, want %v", tt.in, got, tt.want)
+			}
+			for i := range tt.want {
+				if got[i] != tt.want[i] {
+					t.Fatalf("normalizeKeyModels(%v) = %v, want %v", tt.in, got, tt.want)
+				}
+			}
+		})
+	}
+}
+
+// TestAiKeyModelsPersistence 钉住 serializer:json 字段经 Create 与 Updates(map) 两条路径的往返。
+func TestAiKeyModelsPersistence(t *testing.T) {
+	gw, _ := newTestGateway(t, &gatewayStubClient{fakeClient: &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}}})
+	ctx := context.Background()
+
+	_, key, err := gw.CreateKey(ctx, "m-key", "model-key-1234", "", []string{" x-model ", "x-model", "y-model", ""})
+	if err != nil {
+		t.Fatalf("CreateKey: %v", err)
+	}
+	fresh, err := gw.VerifyKey(ctx, "model-key-1234")
+	if err != nil || len(fresh.Models) != 2 || fresh.Models[0] != "x-model" || fresh.Models[1] != "y-model" {
+		t.Fatalf("创建后读回 models = %v, %v", fresh.Models, err)
+	}
+	set := []string{"z-model"}
+	if err := gw.UpdateKey(ctx, key.ID, "", nil, nil, &set); err != nil {
+		t.Fatalf("UpdateKey 覆盖: %v", err)
+	}
+	fresh, _ = gw.VerifyKey(ctx, "model-key-1234")
+	if len(fresh.Models) != 1 || fresh.Models[0] != "z-model" {
+		t.Fatalf("覆盖后 models = %v", fresh.Models)
+	}
+	if err := gw.UpdateKey(ctx, key.ID, "renamed", nil, nil, nil); err != nil {
+		t.Fatalf("UpdateKey 不传 models: %v", err)
+	}
+	fresh, _ = gw.VerifyKey(ctx, "model-key-1234")
+	if len(fresh.Models) != 1 {
+		t.Fatalf("未传 models 却被改动: %v", fresh.Models)
+	}
+	empty := []string{}
+	if err := gw.UpdateKey(ctx, key.ID, "", nil, nil, &empty); err != nil {
+		t.Fatalf("UpdateKey 清空: %v", err)
+	}
+	fresh, _ = gw.VerifyKey(ctx, "model-key-1234")
+	if len(fresh.Models) != 0 {
+		t.Fatalf("清空后 models = %v, want 空", fresh.Models)
+	}
+}
+
+// TestRespPassthroughStreamSwitchesChannel 断言流式直通建立失败按 switchable 换渠道。
+func TestRespPassthroughStreamSwitchesChannel(t *testing.T) {
+	client := &gatewayStubClient{
+		fakeClient:  &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}},
+		passPayload: []byte("data: {\"type\":\"response.completed\"}\n\n"),
+		passErrs:    []error{stubServiceError{status: 503}},
+	}
+	gw, svc := newTestGateway(t, client)
+	cfg := importAliveConfig(t, svc)
+	seedChannel(t, gw, cfg.ID, "eu-frankfurt-1", 1, 1)
+	seedChannel(t, gw, cfg.ID, "us-chicago-1", 1, 1)
+
+	stream, meta, err := gw.RespPassthroughStream(context.Background(), []byte(`{}`), "meta.llama-3.3-70b-instruct", "")
+	if err != nil || stream == nil {
+		t.Fatalf("RespPassthroughStream = %v", err)
+	}
+	defer stream.Close()
+	if meta.Retries != 1 || client.passCalls != 2 {
+		t.Errorf("应换渠道重试一次: retries=%d calls=%d", meta.Retries, client.passCalls)
+	}
+	payload, _ := io.ReadAll(stream)
+	if !bytes.Contains(payload, []byte("response.completed")) {
+		t.Errorf("流内容未透传: %s", payload)
 	}
 }

internal/service/airesponses.go

@@ -8,38 +8,8 @@ import (
 	"oci-portal/internal/aiwire"
 )
 
-// ResponsesToIR 把 Responses API 请求转换为网关 IR(OpenAI Chat 形态)。
+// respRejectStateful 拒绝有状态特性(网关无状态)。
-// 有状态特性与内置工具不支持,直接报错(API 层 400)。
+func respRejectStateful(req aiwire.RespRequest) error {
-func ResponsesToIR(req aiwire.RespRequest) (aiwire.ChatRequest, error) {
-	if err := respRejectUnsupported(req); err != nil {
-		return aiwire.ChatRequest{}, err
-	}
-	ir := aiwire.ChatRequest{
-		Model:       req.Model,
-		MaxTokens:   req.MaxOutputTokens,
-		Temperature: req.Temperature,
-		TopP:        req.TopP,
-		Stream:      req.Stream,
-		ToolChoice:  respToolChoice(req.ToolChoice),
-	}
-	if req.Reasoning != nil {
-		ir.ReasoningEffort = req.Reasoning.Effort
-	}
-	if req.Instructions != "" {
-		ir.Messages = append(ir.Messages, aiwire.ChatMessage{Role: "system", Content: aiwire.NewTextContent(req.Instructions)})
-	}
-	msgs, err := respInputToMessages(req.Input)
-	if err != nil {
-		return aiwire.ChatRequest{}, err
-	}
-	ir.Messages = append(ir.Messages, msgs...)
-	ir.Tools = respTools(req.Tools)
-	ir.ResponseFormat = respFormat(req.Text)
-	return ir, nil
-}
-
-// respRejectUnsupported 拒绝无状态网关无法承接的请求特性。
-func respRejectUnsupported(req aiwire.RespRequest) error {
 	if req.PreviousResponseID != "" {
 		return fmt.Errorf("previous_response_id 不支持:网关不保存历史响应,请在 input 中自带完整上下文(store:false 模式)")
 	}
@@ -49,195 +19,80 @@ func respRejectUnsupported(req aiwire.RespRequest) error {
 	if req.Background != nil && *req.Background {
 		return fmt.Errorf("background 模式不支持")
 	}
-	for _, t := range req.Tools {
-		if t.Type != "function" {
-			return fmt.Errorf("不支持的工具类型 %q:仅支持 function 工具", t.Type)
-		}
-	}
 	return nil
 }
 
-// respInputToMessages 把 input(string 或 item 数组)展开为 IR 消息序列。
+// RespServerTools 报告工具列表是否含 xAI 服务端工具(web_search / x_search)。
-func respInputToMessages(in aiwire.RespInput) ([]aiwire.ChatMessage, error) {
+func RespServerTools(tools []aiwire.RespTool) bool {
-	if !in.IsArray {
-		if strings.TrimSpace(in.Text) == "" {
-			return nil, fmt.Errorf("input 不能为空")
-		}
-		return []aiwire.ChatMessage{{Role: "user", Content: aiwire.NewTextContent(in.Text)}}, nil
-	}
-	var msgs []aiwire.ChatMessage
-	for _, it := range in.Items {
-		out, err := respItemToMessages(it, msgs)
-		if err != nil {
-			return nil, err
-		}
-		msgs = out
-	}
-	return msgs, nil
-}
-
-// respItemToMessages 追加一个 input item;连续 function_call 合并进同一 assistant 消息。
-func respItemToMessages(it aiwire.RespItem, msgs []aiwire.ChatMessage) ([]aiwire.ChatMessage, error) {
-	switch it.Type {
-	case "", "message":
-		if it.Content.HasUnsupported() {
-			return nil, ErrAiUnsupportedBlock
-		}
-		role := respRole(it.Role)
-		return append(msgs, aiwire.ChatMessage{Role: role, Content: respContentToIR(it.Content)}), nil
-	case "function_call":
-		call := aiwire.ToolCall{ID: it.CallID, Type: "function",
-			Function: aiwire.FunctionCall{Name: it.Name, Arguments: it.Arguments}}
-		if n := len(msgs); n > 0 && msgs[n-1].Role == "assistant" && len(msgs[n-1].ToolCalls) > 0 {
-			msgs[n-1].ToolCalls = append(msgs[n-1].ToolCalls, call)
-			return msgs, nil
-		}
-		return append(msgs, aiwire.ChatMessage{Role: "assistant", ToolCalls: []aiwire.ToolCall{call}}), nil
-	case "function_call_output":
-		return append(msgs, aiwire.ChatMessage{Role: "tool", ToolCallID: it.CallID,
-			Content: aiwire.NewTextContent(it.OutputText())}), nil
-	case "reasoning":
-		return msgs, nil // 推理块不回灌上游
-	default:
-		return nil, fmt.Errorf("不支持的 input item 类型 %q", it.Type)
-	}
-}
-
-// respRole 归一 item 角色;developer 视为 system,缺省 user。
-func respRole(role string) string {
-	switch role {
-	case "system", "developer":
-		return "system"
-	case "assistant":
-		return "assistant"
-	default:
-		return "user"
-	}
-}
-
-// respContentToIR 把 Responses 内容转 IR;纯文本拍平,含图片时保留部件顺序。
-func respContentToIR(c aiwire.RespContent) aiwire.Content {
-	var parts []aiwire.ContentPart
-	hasImage := false
-	for _, p := range c.Parts {
-		switch p.Type {
-		case "input_image":
-			parts = append(parts, aiwire.ContentPart{Type: "image_url",
-				ImageURL: &aiwire.ImageURL{URL: p.ImageURL, Detail: p.Detail}})
-			hasImage = true
-		default:
-			if p.Text != "" {
-				parts = append(parts, aiwire.ContentPart{Type: "text", Text: p.Text})
-			}
-		}
-	}
-	if !c.IsArray || !hasImage {
-		return aiwire.NewTextContent(c.JoinText())
-	}
-	return aiwire.NewPartsContent(parts)
-}
-
-// respTools 把扁平工具定义还原为嵌套 Chat 形态;strict 无 OCI 对应,忽略。
-func respTools(tools []aiwire.RespTool) []aiwire.Tool {
-	if len(tools) == 0 {
-		return nil
-	}
-	out := make([]aiwire.Tool, 0, len(tools))
 	for _, t := range tools {
-		out = append(out, aiwire.Tool{Type: "function",
+		if t.Type == "web_search" || t.Type == "x_search" {
-			Function: aiwire.FunctionDef{Name: t.Name, Description: t.Description, Parameters: t.Parameters}})
+			return true
 		}
-	return out
+	}
+	return false
 }
 
-// respToolChoice 转换 tool_choice:字符串直通;{type:function,name} 转嵌套;
+// RespPassthroughValidate 校验直通请求:模型必填,有状态特性不支持,工具类型
-// allowed_tools 等对象形态降级 auto。
+// 只放行 function 与已实测的 web_search / x_search;流式仅在含服务端工具时拒绝
-func respToolChoice(raw json.RawMessage) json.RawMessage {
+// (工具流式事件形态未实测,不放开)。
-	if len(raw) == 0 {
+func RespPassthroughValidate(req aiwire.RespRequest) error {
-		return nil
+	if strings.TrimSpace(req.Model) == "" {
+		return fmt.Errorf("model 不能为空")
 	}
-	t := strings.TrimSpace(string(raw))
+	if req.Stream && RespServerTools(req.Tools) {
-	if strings.HasPrefix(t, "\"") {
+		return fmt.Errorf("服务端工具暂不支持流式:请去掉 stream 或改用 function 工具")
-		return raw
 	}
-	var obj struct {
+	if err := respRejectStateful(req); err != nil {
-		Type string `json:"type"`
+		return err
-		Name string `json:"name"`
 	}
-	if err := json.Unmarshal(raw, &obj); err != nil || obj.Type != "function" || obj.Name == "" {
+	for _, t := range req.Tools {
-		return json.RawMessage(`"auto"`)
+		switch t.Type {
-	}
+		case "function", "web_search", "x_search":
-	out, _ := json.Marshal(map[string]any{"type": "function", "function": map[string]string{"name": obj.Name}})
-	return out
-}
-
-// respFormat 把 text.format 转成 response_format;text 形态无需显式指定。
-func respFormat(text *aiwire.RespText) *aiwire.ResponseFormat {
-	if text == nil || text.Format == nil {
-		return nil
-	}
-	switch text.Format.Type {
-	case "json_object":
-		return &aiwire.ResponseFormat{Type: "json_object"}
-	case "json_schema":
-		spec, _ := json.Marshal(map[string]any{
-			"name": text.Format.Name, "schema": json.RawMessage(text.Format.Schema), "strict": text.Format.Strict,
-		})
-		return &aiwire.ResponseFormat{Type: "json_schema", JSONSchema: spec}
 		default:
+			return fmt.Errorf("不支持的工具类型 %q:服务端工具仅支持 web_search / x_search", t.Type)
+		}
+	}
 	return nil
 }
+
+// RespPassthroughBody 以原始请求体为基构造上游 body:强制 store:false(禁上游
+// 存态),stream 原样保留(流式直通);用 json.Number 保真未知字段与数值。
+func RespPassthroughBody(raw []byte) ([]byte, error) {
+	dec := json.NewDecoder(strings.NewReader(string(raw)))
+	dec.UseNumber()
+	var body map[string]any
+	if err := dec.Decode(&body); err != nil {
+		return nil, fmt.Errorf("解析请求体: %w", err)
+	}
+	body["store"] = false
+	return json.Marshal(body)
 }
 
-// IRRespToResponses 把 IR 非流式响应装配为 Response 对象。
+// RespPassthroughUsage 从直通响应提取用量;缺失时返回 nil(日志记零)。
-func IRRespToResponses(resp *aiwire.ChatResponse, id string, created int64) aiwire.Response {
+func RespPassthroughUsage(payload []byte) *aiwire.Usage {
-	out := aiwire.Response{ID: id, Object: "response", CreatedAt: created,
+	var root struct {
-		Status: "completed", Model: resp.Model, Output: []aiwire.RespOutItem{}, Store: false}
+		Usage *aiwire.RespUsage `json:"usage"`
-	if len(resp.Choices) == 0 {
-		return out
 	}
-	choice := resp.Choices[0]
+	if json.Unmarshal(payload, &root) != nil || root.Usage == nil {
-	if text := choice.Message.Content.JoinText(); text != "" {
-		out.Output = append(out.Output, respMessageItem(id, 0, text))
-	}
-	for i, tc := range choice.Message.ToolCalls {
-		out.Output = append(out.Output, aiwire.RespOutItem{Type: "function_call",
-			ID: respItemID(id, "fc", len(out.Output)+i), Status: "completed",
-			CallID: tc.ID, Name: tc.Function.Name, Arguments: respArgs(tc.Function.Arguments)})
-	}
-	if choice.FinishReason == "length" {
-		out.Status = "incomplete"
-		out.IncompleteDetails = &aiwire.RespIncomplete{Reason: "max_output_tokens"}
-	}
-	out.Usage = respUsage(resp.Usage)
-	return out
-}
-
-func respMessageItem(respID string, idx int, text string) aiwire.RespOutItem {
-	return aiwire.RespOutItem{Type: "message", ID: respItemID(respID, "msg", idx),
-		Status: "completed", Role: "assistant",
-		Content: []aiwire.RespOutPart{{Type: "output_text", Text: text, Annotations: []any{}}}}
-}
-
-// respItemID 从响应 ID 派生确定性 item ID。
-func respItemID(respID, kind string, idx int) string {
-	return fmt.Sprintf("%s_%s_%d", kind, strings.TrimPrefix(respID, "resp_"), idx)
-}
-
-// respArgs 保证 arguments 是合法 JSON 字符串(空实参回退 {})。
-func respArgs(args string) string {
-	if strings.TrimSpace(args) == "" {
-		return "{}"
-	}
-	return args
-}
-
-// respUsage 把 IR 用量改写为 Responses 命名口径。
-func respUsage(u *aiwire.Usage) *aiwire.RespUsage {
-	if u == nil {
 		return nil
 	}
-	return &aiwire.RespUsage{InputTokens: u.PromptTokens, OutputTokens: u.CompletionTokens,
+	usage := &aiwire.Usage{PromptTokens: root.Usage.InputTokens,
-		TotalTokens:        u.TotalTokens,
+		CompletionTokens: root.Usage.OutputTokens, TotalTokens: root.Usage.TotalTokens}
-		InputTokensDetails: aiwire.RespInDetails{CachedTokens: u.CachedTokens()}}
+	if cached := root.Usage.InputTokensDetails.CachedTokens; cached > 0 {
+		usage.PromptTokensDetails = &aiwire.PromptTokensDetails{CachedTokens: cached}
+	}
+	return usage
+}
+
+// RespStreamCompletedUsage 从一行 SSE data JSON 中提取 response.completed 事件的
+// usage;非 completed 事件或解析失败返回 nil。流式直通逐行喂入,最后一次非 nil 生效。
+func RespStreamCompletedUsage(data []byte) *aiwire.Usage {
+	var ev struct {
+		Type     string          `json:"type"`
+		Response json.RawMessage `json:"response"`
+	}
+	if json.Unmarshal(data, &ev) != nil || ev.Type != "response.completed" || len(ev.Response) == 0 {
+		return nil
+	}
+	return RespPassthroughUsage(ev.Response)
 }

internal/service/airesponses_stream.go

@@ -1,199 +0,0 @@
-package service
-
-import (
-	"strings"
-
-	"oci-portal/internal/aiwire"
-)
-
-// RespEvent 是一条 Responses SSE 语义事件(event 名 + data 载荷)。
-type RespEvent struct {
-	Event string
-	Data  map[string]any
-}
-
-// RespStream 把 IR chunk 流聚合为 Responses 语义事件序列。
-// 与 AnthStream 同构,但终态事件须携带完整 Response 快照,故全程缓冲文本与实参。
-type RespStream struct {
-	id      string
-	model   string
-	created int64
-	seq     int
-	started bool
-	items   []aiwire.RespOutItem
-	msgOpen bool
-	text    strings.Builder
-	tool    aiwire.RespOutItem
-	toolIdx int
-	tOpen   bool
-	args    strings.Builder
-	finish  string
-	usage   *aiwire.Usage
-}
-
-// NewRespStream 构造状态机;id 形如 resp_*,created 为响应时间戳。
-func NewRespStream(id, model string, created int64) *RespStream {
-	return &RespStream{id: id, model: model, created: created}
-}
-
-// Feed 消化一个上游 chunk,返回应立即下发的事件。
-func (s *RespStream) Feed(chunk aiwire.ChatChunk) []RespEvent {
-	var evs []RespEvent
-	if !s.started {
-		s.started = true
-		evs = append(evs, s.respEvent("response.created", "in_progress"),
-			s.respEvent("response.in_progress", "in_progress"))
-	}
-	if chunk.Usage != nil {
-		s.usage = chunk.Usage
-	}
-	if len(chunk.Choices) == 0 {
-		return evs
-	}
-	choice := chunk.Choices[0]
-	if choice.Delta.Content != "" {
-		evs = append(evs, s.feedText(choice.Delta.Content)...)
-	}
-	for _, tc := range choice.Delta.ToolCalls {
-		evs = append(evs, s.feedTool(tc)...)
-	}
-	if choice.FinishReason != nil {
-		s.finish = *choice.FinishReason
-	}
-	return evs
-}
-
-// Finish 关闭未闭合的块并产出终态事件(带完整 Response 快照)。
-func (s *RespStream) Finish() []RespEvent {
-	var evs []RespEvent
-	if !s.started {
-		s.started = true
-		evs = append(evs, s.respEvent("response.created", "in_progress"))
-	}
-	evs = append(evs, s.closeMsg()...)
-	evs = append(evs, s.closeTool()...)
-	status := "completed"
-	if s.finish == "length" {
-		status = "incomplete"
-	}
-	return append(evs, s.respEvent("response."+status, status))
-}
-
-// Usage 返回聚合到的用量(可能为 nil),供调用日志。
-func (s *RespStream) Usage() *aiwire.Usage { return s.usage }
-
-// ev 构造带自增 sequence_number 的事件。
-func (s *RespStream) ev(typ string, kv map[string]any) RespEvent {
-	s.seq++
-	data := map[string]any{"type": typ, "sequence_number": s.seq}
-	for k, v := range kv {
-		data[k] = v
-	}
-	return RespEvent{Event: typ, Data: data}
-}
-
-// respEvent 构造携带 Response 快照的生命周期事件。
-func (s *RespStream) respEvent(typ, status string) RespEvent {
-	return s.ev(typ, map[string]any{"response": s.snapshot(status)})
-}
-
-// snapshot 组装当前累计状态的 Response 对象。
-func (s *RespStream) snapshot(status string) aiwire.Response {
-	resp := aiwire.Response{ID: s.id, Object: "response", CreatedAt: s.created,
-		Status: status, Model: s.model, Store: false,
-		Output: append([]aiwire.RespOutItem{}, s.items...)}
-	resp.Usage = respUsage(s.usage)
-	if status == "incomplete" {
-		resp.IncompleteDetails = &aiwire.RespIncomplete{Reason: "max_output_tokens"}
-	}
-	return resp
-}
-
-// feedText 处理文本增量:必要时开 message item 与 content part。
-func (s *RespStream) feedText(delta string) []RespEvent {
-	evs := s.closeTool()
-	if !s.msgOpen {
-		s.msgOpen = true
-		s.text.Reset()
-		item := aiwire.RespOutItem{Type: "message", ID: s.itemID("msg"), Status: "in_progress",
-			Role: "assistant", Content: []aiwire.RespOutPart{}}
-		evs = append(evs, s.ev("response.output_item.added", map[string]any{
-			"output_index": len(s.items), "item": item}))
-		evs = append(evs, s.ev("response.content_part.added", map[string]any{
-			"item_id": item.ID, "output_index": len(s.items), "content_index": 0,
-			"part": aiwire.RespOutPart{Type: "output_text", Text: "", Annotations: []any{}}}))
-	}
-	s.text.WriteString(delta)
-	return append(evs, s.ev("response.output_text.delta", map[string]any{
-		"item_id": s.itemID("msg"), "output_index": len(s.items), "content_index": 0, "delta": delta}))
-}
-
-// closeMsg 闭合当前 message item(text done → part done → item done)。
-func (s *RespStream) closeMsg() []RespEvent {
-	if !s.msgOpen {
-		return nil
-	}
-	s.msgOpen = false
-	id, idx, full := s.itemID("msg"), len(s.items), s.text.String()
-	part := aiwire.RespOutPart{Type: "output_text", Text: full, Annotations: []any{}}
-	item := aiwire.RespOutItem{Type: "message", ID: id, Status: "completed",
-		Role: "assistant", Content: []aiwire.RespOutPart{part}}
-	evs := []RespEvent{
-		s.ev("response.output_text.done", map[string]any{"item_id": id, "output_index": idx, "content_index": 0, "text": full}),
-		s.ev("response.content_part.done", map[string]any{"item_id": id, "output_index": idx, "content_index": 0, "part": part}),
-		s.ev("response.output_item.done", map[string]any{"output_index": idx, "item": item}),
-	}
-	s.items = append(s.items, item)
-	return evs
-}
-
-// feedTool 处理工具调用增量:新 Index 先闭合旧调用再开新 item。
-func (s *RespStream) feedTool(tc aiwire.ToolCallDelta) []RespEvent {
-	evs := s.closeMsg()
-	if s.tOpen && tc.Index != s.toolIdx {
-		evs = append(evs, s.closeTool()...)
-	}
-	if !s.tOpen {
-		s.tOpen, s.toolIdx = true, tc.Index
-		s.args.Reset()
-		s.tool = aiwire.RespOutItem{Type: "function_call", ID: s.itemID("fc"), Status: "in_progress",
-			CallID: tc.ID, Name: tc.Function.Name}
-		evs = append(evs, s.ev("response.output_item.added", map[string]any{
-			"output_index": len(s.items), "item": s.tool}))
-	}
-	if tc.ID != "" && s.tool.CallID == "" {
-		s.tool.CallID = tc.ID
-	}
-	if tc.Function.Name != "" && s.tool.Name == "" {
-		s.tool.Name = tc.Function.Name
-	}
-	if tc.Function.Arguments == "" {
-		return evs
-	}
-	s.args.WriteString(tc.Function.Arguments)
-	return append(evs, s.ev("response.function_call_arguments.delta", map[string]any{
-		"item_id": s.tool.ID, "output_index": len(s.items), "delta": tc.Function.Arguments}))
-}
-
-// closeTool 闭合当前 function_call item(arguments done → item done)。
-func (s *RespStream) closeTool() []RespEvent {
-	if !s.tOpen {
-		return nil
-	}
-	s.tOpen = false
-	s.tool.Arguments = respArgs(s.args.String())
-	s.tool.Status = "completed"
-	idx := len(s.items)
-	evs := []RespEvent{
-		s.ev("response.function_call_arguments.done", map[string]any{
-			"item_id": s.tool.ID, "output_index": idx, "arguments": s.tool.Arguments}),
-		s.ev("response.output_item.done", map[string]any{"output_index": idx, "item": s.tool}),
-	}
-	s.items = append(s.items, s.tool)
-	return evs
-}
-
-// itemID 按当前 output 序号派生确定性 item ID。
-func (s *RespStream) itemID(kind string) string {
-	return respItemID(s.id, kind, len(s.items))
-}

internal/service/airesponses_test.go

@@ -17,47 +17,6 @@ func respReq(t *testing.T, raw string) aiwire.RespRequest {
 	return req
 }
 
-func TestResponsesToIR(t *testing.T) {
-	req := respReq(t, `{
-		"model": "meta.llama-3.3-70b-instruct",
-		"instructions": "你是助手",
-		"max_output_tokens": 128,
-		"input": [
-			{"role": "user", "content": "查天气"},
-			{"type": "function_call", "call_id": "call_1", "name": "get_weather", "arguments": "{\"city\":\"上海\"}"},
-			{"type": "function_call_output", "call_id": "call_1", "output": "{\"temp\":31}"},
-			{"type": "message", "role": "assistant", "content": [{"type": "output_text", "text": "31 度"}]}
-		],
-		"tools": [{"type": "function", "name": "get_weather", "description": "查天气", "parameters": {"type": "object"}, "strict": true}],
-		"tool_choice": {"type": "function", "name": "get_weather"},
-		"text": {"format": {"type": "json_schema", "name": "out", "schema": {"type": "object"}}}
-	}`)
-	ir, err := ResponsesToIR(req)
-	if err != nil {
-		t.Fatalf("ResponsesToIR: %v", err)
-	}
-	roles := make([]string, 0, len(ir.Messages))
-	for _, m := range ir.Messages {
-		roles = append(roles, m.Role)
-	}
-	want := []string{"system", "user", "assistant", "tool", "assistant"}
-	if strings.Join(roles, ",") != strings.Join(want, ",") {
-		t.Errorf("roles = %v, want %v", roles, want)
-	}
-	if ir.Messages[2].ToolCalls[0].ID != "call_1" || ir.Messages[3].ToolCallID != "call_1" {
-		t.Errorf("tool call 链接错误: %+v", ir.Messages)
-	}
-	if deref(ir.MaxTokens) != 128 || len(ir.Tools) != 1 || ir.Tools[0].Function.Name != "get_weather" {
-		t.Errorf("参数映射错误: max=%v tools=%+v", ir.MaxTokens, ir.Tools)
-	}
-	if !strings.Contains(string(ir.ToolChoice), `"function"`) || !strings.Contains(string(ir.ToolChoice), "get_weather") {
-		t.Errorf("tool_choice = %s", ir.ToolChoice)
-	}
-	if ir.ResponseFormat == nil || ir.ResponseFormat.Type != "json_schema" {
-		t.Errorf("response_format = %+v", ir.ResponseFormat)
-	}
-}
-
 func deref(p *int) int {
 	if p == nil {
 		return 0
@@ -65,112 +24,124 @@ func deref(p *int) int {
 	return *p
 }
 
-func TestResponsesToIRRejects(t *testing.T) {
+func TestRespServerTools(t *testing.T) {
-	cases := []struct{ name, raw string }{
+	tests := []struct {
-		{"previous_response_id", `{"model":"m","input":"hi","previous_response_id":"resp_x"}`},
+		name  string
-		{"conversation", `{"model":"m","input":"hi","conversation":{"id":"conv_1"}}`},
+		tools []aiwire.RespTool
-		{"background", `{"model":"m","input":"hi","background":true}`},
+		want  bool
-		{"builtin tool", `{"model":"m","input":"hi","tools":[{"type":"web_search"}]}`},
+	}{
-		{"file_id image", `{"model":"m","input":[{"role":"user","content":[{"type":"input_image","file_id":"file_1"}]}]}`},
+		{"web_search", []aiwire.RespTool{{Type: "web_search"}}, true},
-		{"unknown item", `{"model":"m","input":[{"type":"item_reference","id":"x"}]}`},
+		{"x_search混用", []aiwire.RespTool{{Type: "function", Name: "f"}, {Type: "x_search"}}, true},
+		{"仅function", []aiwire.RespTool{{Type: "function", Name: "f"}}, false},
+		{"空", nil, false},
+		{"其他类型", []aiwire.RespTool{{Type: "code_interpreter"}}, false},
 	}
-	for _, c := range cases {
+	for _, test := range tests {
-		if _, err := ResponsesToIR(respReq(t, c.raw)); err == nil {
+		t.Run(test.name, func(t *testing.T) {
-			t.Errorf("%s 应被拒绝", c.name)
+			if got := RespServerTools(test.tools); got != test.want {
+				t.Fatalf("got %v, want %v", got, test.want)
+			}
+		})
 	}
 }
-	// input 字符串形态 + store/reasoning 忽略项不报错
+
-	req := respReq(t, `{"model":"m","input":"你好","store":false,"reasoning":{"effort":"low"},"parallel_tool_calls":true}`)
+func TestRespPassthroughValidate(t *testing.T) {
-	ir, err := ResponsesToIR(req)
+	prev := "resp_1"
-	if err != nil || len(ir.Messages) != 1 || ir.Messages[0].Role != "user" || ir.ReasoningEffort != "low" {
+	bg := true
-		t.Errorf("字符串 input = %+v, %v", ir.Messages, err)
+	tests := []struct {
+		name    string
+		req     aiwire.RespRequest
+		wantErr bool
+	}{
+		{"合法", aiwire.RespRequest{Model: "m", Tools: []aiwire.RespTool{{Type: "web_search"}}}, false},
+		{"混用function", aiwire.RespRequest{Model: "m", Tools: []aiwire.RespTool{{Type: "web_search"}, {Type: "function", Name: "f"}}}, false},
+		{"缺model", aiwire.RespRequest{Tools: []aiwire.RespTool{{Type: "web_search"}}}, true},
+		{"工具加流式拒绝", aiwire.RespRequest{Model: "m", Stream: true, Tools: []aiwire.RespTool{{Type: "web_search"}}}, true},
+		{"无工具流式放行", aiwire.RespRequest{Model: "m", Stream: true}, false},
+		{"function工具流式放行", aiwire.RespRequest{Model: "m", Stream: true, Tools: []aiwire.RespTool{{Type: "function", Name: "f"}}}, false},
+		{"有状态拒绝", aiwire.RespRequest{Model: "m", PreviousResponseID: prev}, true},
+		{"background拒绝", aiwire.RespRequest{Model: "m", Background: &bg}, true},
+		{"未知工具拒绝", aiwire.RespRequest{Model: "m", Tools: []aiwire.RespTool{{Type: "web_search"}, {Type: "mcp"}}}, true},
 	}
-	// input_image(url 形态)放行并保留图文顺序
+	for _, test := range tests {
-	req2 := respReq(t, `{"model":"m","input":[{"role":"user","content":[{"type":"input_text","text":"看图"},{"type":"input_image","image_url":"https://x/1.png","detail":"low"}]}]}`)
+		t.Run(test.name, func(t *testing.T) {
-	ir2, err := ResponsesToIR(req2)
+			err := RespPassthroughValidate(test.req)
+			if (err != nil) != test.wantErr {
+				t.Fatalf("err = %v, wantErr %v", err, test.wantErr)
+			}
+		})
+	}
+}
+
+func TestRespPassthroughBody(t *testing.T) {
+	raw := []byte(`{"model":"m","input":"hi","stream":true,"store":true,"max_output_tokens":128,"custom_field":{"a":1.5}}`)
+	out, err := RespPassthroughBody(raw)
 	if err != nil {
-		t.Fatalf("input_image 应放行: %v", err)
+		t.Fatalf("RespPassthroughBody: %v", err)
 	}
-	parts := ir2.Messages[0].Content.Parts
+	var body map[string]any
-	if len(parts) != 2 || parts[1].Type != "image_url" || parts[1].ImageURL == nil || parts[1].ImageURL.URL != "https://x/1.png" {
+	if err := json.Unmarshal(out, &body); err != nil {
-		t.Errorf("parts = %+v", parts)
+		t.Fatalf("unmarshal out: %v", err)
+	}
+	if body["store"] != false {
+		t.Errorf("store 应强制 false, got %v", body["store"])
+	}
+	if body["stream"] != true {
+		t.Errorf("stream 应原样保留, got %v", body["stream"])
+	}
+	if string(out) == "" || !strings.Contains(string(out), `"max_output_tokens":128`) {
+		t.Errorf("数值字段应保真: %s", out)
+	}
+	if !strings.Contains(string(out), `"custom_field"`) {
+		t.Errorf("未知字段应保留: %s", out)
+	}
+	if _, err := RespPassthroughBody([]byte("not-json")); err == nil {
+		t.Error("非法 JSON 应报错")
 	}
 }
 
-func TestIRRespToResponses(t *testing.T) {
+func TestRespPassthroughUsage(t *testing.T) {
-	resp := &aiwire.ChatResponse{Model: "m", Choices: []aiwire.Choice{{
+	tests := []struct {
-		Message: aiwire.ChatMessage{Role: "assistant", Content: aiwire.NewTextContent("你好"),
+		name           string
-			ToolCalls: []aiwire.ToolCall{{ID: "call_9", Type: "function",
+		payload        string
-				Function: aiwire.FunctionCall{Name: "f", Arguments: ""}}}},
+		wantNil        bool
-		FinishReason: "length",
+		prompt, cached int
-	}}, Usage: &aiwire.Usage{PromptTokens: 10, CompletionTokens: 5, TotalTokens: 15}}
+	}{
-	out := IRRespToResponses(resp, "resp_abc", 1751966400)
+		{"完整", `{"usage":{"input_tokens":9,"output_tokens":3,"total_tokens":12,"input_tokens_details":{"cached_tokens":5}}}`, false, 9, 5},
-	if out.Status != "incomplete" || out.IncompleteDetails == nil || out.IncompleteDetails.Reason != "max_output_tokens" {
+		{"无细分", `{"usage":{"input_tokens":9,"output_tokens":3,"total_tokens":12}}`, false, 9, 0},
-		t.Errorf("length 应映射 incomplete: %+v", out)
+		{"无usage", `{"id":"resp_1"}`, true, 0, 0},
+		{"非法JSON", `xx`, true, 0, 0},
 	}
-	if len(out.Output) != 2 || out.Output[0].Type != "message" || out.Output[1].Type != "function_call" {
+	for _, test := range tests {
-		t.Fatalf("output = %+v", out.Output)
+		t.Run(test.name, func(t *testing.T) {
+			usage := RespPassthroughUsage([]byte(test.payload))
+			if (usage == nil) != test.wantNil {
+				t.Fatalf("usage = %v, wantNil %v", usage, test.wantNil)
 			}
-	if out.Output[0].Content[0].Text != "你好" || out.Output[1].CallID != "call_9" || out.Output[1].Arguments != "{}" {
+			if usage == nil {
-		t.Errorf("item 装配错误: %+v", out.Output)
+				return
 			}
-	if out.Usage.InputTokens != 10 || out.Usage.OutputTokens != 5 || out.Store {
+			if usage.PromptTokens != test.prompt || usage.CachedTokens() != test.cached {
-		t.Errorf("usage/store 错误: %+v", out)
+				t.Fatalf("prompt=%d cached=%d, want %d/%d", usage.PromptTokens, usage.CachedTokens(), test.prompt, test.cached)
+			}
+		})
 	}
 }
 
-func chunkText(text string) aiwire.ChatChunk {
+// TestRespStreamCompletedUsage 断言流式 usage 只从 completed 事件提取。
-	return aiwire.ChatChunk{Choices: []aiwire.ChunkChoice{{Delta: aiwire.Delta{Content: text}}}}
+func TestRespStreamCompletedUsage(t *testing.T) {
+	completed := []byte(`{"type":"response.completed","response":{"usage":{"input_tokens":10,"output_tokens":5,"total_tokens":15,"input_tokens_details":{"cached_tokens":4}}}}`)
+	if u := RespStreamCompletedUsage(completed); u == nil || u.PromptTokens != 10 || u.CompletionTokens != 5 ||
+		u.PromptTokensDetails == nil || u.PromptTokensDetails.CachedTokens != 4 {
+		t.Fatalf("completed 事件 usage 解析失败: %+v", u)
 	}
-
+	for _, data := range []string{
-func TestRespStreamTextAndTool(t *testing.T) {
+		`{"type":"response.output_text.delta","delta":"hi"}`,
-	st := NewRespStream("resp_x", "m", 1751966400)
+		`{"type":"response.completed"}`,
-	var types []string
+		`not-json`,
-	collect := func(evs []RespEvent) {
+	} {
-		for _, ev := range evs {
+		if u := RespStreamCompletedUsage([]byte(data)); u != nil {
-			types = append(types, ev.Event)
+			t.Fatalf("非 completed 或缺 response 应返回 nil: %s", data)
 		}
 	}
-	collect(st.Feed(chunkText("你")))
-	collect(st.Feed(chunkText("好")))
-	collect(st.Feed(aiwire.ChatChunk{Choices: []aiwire.ChunkChoice{{Delta: aiwire.Delta{
-		ToolCalls: []aiwire.ToolCallDelta{{Index: 0, ID: "call_1", Function: aiwire.FunctionCallDelta{Name: "f", Arguments: "{\"a\""}}}}}}}))
-	collect(st.Feed(aiwire.ChatChunk{Choices: []aiwire.ChunkChoice{{Delta: aiwire.Delta{
-		ToolCalls: []aiwire.ToolCallDelta{{Index: 0, Function: aiwire.FunctionCallDelta{Arguments: ":1}"}}}}}},
-		Usage: &aiwire.Usage{PromptTokens: 3, CompletionTokens: 7, TotalTokens: 10}}))
-	finish := st.Finish()
-	collect(finish)
-	want := []string{
-		"response.created", "response.in_progress",
-		"response.output_item.added", "response.content_part.added", "response.output_text.delta",
-		"response.output_text.delta",
-		"response.output_text.done", "response.content_part.done", "response.output_item.done",
-		"response.output_item.added", "response.function_call_arguments.delta",
-		"response.function_call_arguments.delta",
-		"response.function_call_arguments.done", "response.output_item.done",
-		"response.completed",
-	}
-	if strings.Join(types, "\n") != strings.Join(want, "\n") {
-		t.Errorf("事件序列 =\n%s\nwant\n%s", strings.Join(types, "\n"), strings.Join(want, "\n"))
-	}
-	last := finish[len(finish)-1].Data
-	resp := last["response"].(aiwire.Response)
-	if len(resp.Output) != 2 || resp.Output[0].Content[0].Text != "你好" || resp.Output[1].Arguments != "{\"a\":1}" {
-		t.Errorf("终态快照 = %+v", resp.Output)
-	}
-	if resp.Usage == nil || resp.Usage.TotalTokens != 10 {
-		t.Errorf("终态 usage = %+v", resp.Usage)
-	}
-	// sequence_number 严格递增
-	if seq, ok := last["sequence_number"].(int); !ok || seq != len(want) {
-		t.Errorf("最终 sequence_number = %v, want %d", last["sequence_number"], len(want))
-	}
-}
-
-func TestRespStreamEmpty(t *testing.T) {
-	st := NewRespStream("resp_e", "m", 1)
-	evs := st.Finish()
-	if len(evs) != 2 || evs[0].Event != "response.created" || evs[1].Event != "response.completed" {
-		t.Errorf("空流事件 = %+v", evs)
-	}
 }

internal/service/anthresponses.go

@@ -0,0 +1,411 @@
+package service
+
+import (
+	"encoding/json"
+	"fmt"
+	"strings"
+
+	"oci-portal/internal/aiwire"
+)
+
+// Anthropic Messages ↔ OCI OpenAI 兼容面(/actions/v1/responses)直通转换。
+// typed chat 面剔除后 Messages 入口的唯一上游通路。语义损失(README 已披露):
+// stop_sequences / top_k / metadata / thinking 无对应字段,忽略;上游 reasoning
+// 输出项与增量事件丢弃(Anthropic thinking 块含签名语义,不伪造)。
+
+// ErrAiUnsupportedBlock 表示请求含网关无法承接的内容块(仅支持文本/图片/工具块)。
+var ErrAiUnsupportedBlock = fmt.Errorf("暂不支持文本与图片以外的内容块")
+
+// AnthropicToResponsesBody 把 Messages 请求转为直通 body(强制 store:false)。
+func AnthropicToResponsesBody(req aiwire.MessagesRequest) ([]byte, error) {
+	input, err := anthInputItems(req.Messages)
+	if err != nil {
+		return nil, err
+	}
+	body := map[string]any{"model": req.Model, "max_output_tokens": req.MaxTokens,
+		"input": input, "store": false}
+	if sys := req.SystemText(); sys != "" {
+		body["instructions"] = sys
+	}
+	if req.Temperature != nil {
+		body["temperature"] = *req.Temperature
+	}
+	if req.TopP != nil {
+		body["top_p"] = *req.TopP
+	}
+	if tools := anthRespTools(req.Tools); tools != nil {
+		body["tools"] = tools
+	}
+	if tc := anthRespToolChoice(req.ToolChoice); tc != nil {
+		body["tool_choice"] = tc
+	}
+	if req.OutputConfig != nil && req.OutputConfig.Effort != "" {
+		body["reasoning"] = map[string]string{"effort": strings.ToLower(req.OutputConfig.Effort)}
+	}
+	if req.Stream {
+		body["stream"] = true
+	}
+	return json.Marshal(body)
+}
+
+// anthInputItems 把消息序列展开为 Responses input 项:tool_use / tool_result 为
+// 独立 function_call / function_call_output 项,其余聚合为 message 项;遇独立项时
+// 先冲刷已聚合部件,保持块间相对顺序。
+func anthInputItems(messages []aiwire.AnthMessage) ([]any, error) {
+	var items []any
+	for _, m := range messages {
+		var parts []map[string]any
+		flush := func() {
+			if len(parts) > 0 {
+				items = append(items, map[string]any{"role": m.Role, "content": parts})
+				parts = nil
+			}
+		}
+		for _, b := range m.Content.AllBlocks() {
+			item, part, err := anthBlockItem(m.Role, b)
+			if err != nil {
+				return nil, err
+			}
+			if part != nil {
+				parts = append(parts, part)
+			}
+			if item != nil {
+				flush()
+				items = append(items, item)
+			}
+		}
+		flush()
+	}
+	return items, nil
+}
+
+// anthBlockItem 把单个内容块转为独立项或消息部件(thinking 忽略,未知块拒绝)。
+func anthBlockItem(role string, b aiwire.AnthBlock) (item, part map[string]any, err error) {
+	switch b.Type {
+	case "text":
+		return nil, anthTextPart(role, b.Text), nil
+	case "image":
+		part, err = anthImageInput(b.Source)
+		return nil, part, err
+	case "tool_use":
+		return map[string]any{"type": "function_call", "call_id": b.ID,
+			"name": b.Name, "arguments": string(b.Input)}, nil, nil
+	case "tool_result":
+		return map[string]any{"type": "function_call_output",
+			"call_id": b.ToolUseID, "output": b.ResultText()}, nil, nil
+	case "thinking", "redacted_thinking":
+		return nil, nil, nil
+	default:
+		return nil, nil, ErrAiUnsupportedBlock
+	}
+}
+
+// anthTextPart 按角色选择 Responses 文本部件类型(assistant 历史为 output_text)。
+func anthTextPart(role, text string) map[string]any {
+	if role == "assistant" {
+		return map[string]any{"type": "output_text", "text": text}
+	}
+	return map[string]any{"type": "input_text", "text": text}
+}
+
+// anthImageInput 把 Anthropic image source 转为 Responses input_image 部件。
+func anthImageInput(source json.RawMessage) (map[string]any, error) {
+	var src struct {
+		Type      string `json:"type"`
+		MediaType string `json:"media_type"`
+		Data      string `json:"data"`
+		URL       string `json:"url"`
+	}
+	if err := json.Unmarshal(source, &src); err != nil {
+		return nil, fmt.Errorf("image source 解析失败: %w", err)
+	}
+	switch src.Type {
+	case "base64":
+		if src.MediaType == "" || src.Data == "" {
+			return nil, fmt.Errorf("image source 缺少 media_type 或 data")
+		}
+		return map[string]any{"type": "input_image",
+			"image_url": "data:" + src.MediaType + ";base64," + src.Data}, nil
+	case "url":
+		if src.URL == "" {
+			return nil, fmt.Errorf("image source 缺少 url")
+		}
+		return map[string]any{"type": "input_image", "image_url": src.URL}, nil
+	}
+	return nil, fmt.Errorf("不支持的 image source 类型 %q", src.Type)
+}
+
+func anthRespTools(tools []aiwire.AnthTool) []map[string]any {
+	if len(tools) == 0 {
+		return nil
+	}
+	out := make([]map[string]any, 0, len(tools))
+	for _, t := range tools {
+		out = append(out, map[string]any{"type": "function", "name": t.Name,
+			"description": t.Description, "parameters": t.InputSchema})
+	}
+	return out
+}
+
+// anthRespToolChoice 映射 {type:auto|any|tool|none,name} → Responses 形态。
+func anthRespToolChoice(raw json.RawMessage) any {
+	if len(raw) == 0 {
+		return nil
+	}
+	var tc struct {
+		Type string `json:"type"`
+		Name string `json:"name"`
+	}
+	if json.Unmarshal(raw, &tc) != nil {
+		return nil
+	}
+	switch tc.Type {
+	case "auto":
+		return "auto"
+	case "any":
+		return "required"
+	case "none":
+		return "none"
+	case "tool":
+		return map[string]string{"type": "function", "name": tc.Name}
+	}
+	return nil
+}
+
+// respPayload 是直通响应中本转换关心的子集(未知字段忽略)。
+type respPayload struct {
+	Model             string             `json:"model"`
+	Status            string             `json:"status"`
+	IncompleteDetails *respIncomplete    `json:"incomplete_details"`
+	Output            []respOutputItem   `json:"output"`
+	Usage             *aiwire.RespUsage  `json:"usage"`
+	Error             *map[string]string `json:"error"`
+}
+
+type respIncomplete struct {
+	Reason string `json:"reason"`
+}
+
+type respOutputItem struct {
+	Type      string `json:"type"`
+	CallID    string `json:"call_id"`
+	Name      string `json:"name"`
+	Arguments string `json:"arguments"`
+	Content   []struct {
+		Type string `json:"type"`
+		Text string `json:"text"`
+	} `json:"content"`
+}
+
+// ResponsesToAnthropic 把直通非流式响应转为 Anthropic Messages 响应。
+func ResponsesToAnthropic(payload []byte, msgID string) (*aiwire.MessagesResponse, error) {
+	var resp respPayload
+	if err := json.Unmarshal(payload, &resp); err != nil {
+		return nil, fmt.Errorf("解析上游响应: %w", err)
+	}
+	out := &aiwire.MessagesResponse{ID: msgID, Type: "message", Role: "assistant",
+		Model: resp.Model, Content: []aiwire.AnthBlock{}, StopReason: "end_turn"}
+	for _, item := range resp.Output {
+		switch item.Type {
+		case "message":
+			for _, part := range item.Content {
+				if part.Type == "output_text" && part.Text != "" {
+					out.Content = append(out.Content, aiwire.AnthBlock{Type: "text", Text: part.Text})
+				}
+			}
+		case "function_call":
+			out.Content = append(out.Content, aiwire.AnthBlock{Type: "tool_use", ID: item.CallID,
+				Name: item.Name, Input: argsToJSON(item.Arguments)})
+			out.StopReason = "tool_use"
+		}
+	}
+	if resp.Status == "incomplete" && resp.IncompleteDetails != nil &&
+		resp.IncompleteDetails.Reason == "max_output_tokens" {
+		out.StopReason = "max_tokens"
+	}
+	out.Usage = anthUsageFromResp(resp.Usage)
+	return out, nil
+}
+
+func anthUsageFromResp(u *aiwire.RespUsage) aiwire.AnthUsage {
+	if u == nil {
+		return aiwire.AnthUsage{}
+	}
+	return aiwire.AnthUsage{InputTokens: u.InputTokens, OutputTokens: u.OutputTokens,
+		CacheReadInputTokens: u.InputTokensDetails.CachedTokens}
+}
+
+// argsToJSON 保证 tool_use.input 是合法 JSON 对象(模型可能产出非法片段)。
+func argsToJSON(args string) json.RawMessage {
+	trimmed := strings.TrimSpace(args)
+	if trimmed == "" {
+		return json.RawMessage(`{}`)
+	}
+	if json.Valid([]byte(trimmed)) {
+		return json.RawMessage(trimmed)
+	}
+	b, _ := json.Marshal(map[string]string{"_raw": args})
+	return b
+}
+
+// ---- Anthropic 流式桥:Responses SSE 事件 → Anthropic 事件序列 ----
+
+// AnthEvent 是一条待写出的 Anthropic SSE 事件。
+type AnthEvent struct {
+	Event string
+	Data  any
+}
+
+// AnthRespBridge 把直通 SSE 事件流桥接为 Anthropic 事件序列:
+// message_start → content_block_start/delta/stop(text 与 tool_use 分块)→ message_delta → message_stop。
+// reasoning 系列事件丢弃。
+type AnthRespBridge struct {
+	id, model   string
+	started     bool
+	blockOpen   bool
+	blockIsTool bool
+	blockIndex  int
+	stopReason  string
+	usage       aiwire.AnthUsage
+}
+
+// NewAnthRespBridge 构造桥;id 为响应消息 ID。
+func NewAnthRespBridge(id, model string) *AnthRespBridge {
+	return &AnthRespBridge{id: id, model: model, blockIndex: -1, stopReason: "end_turn"}
+}
+
+// respStreamEvent 是直通 SSE data JSON 中桥关心的子集。
+type respStreamEvent struct {
+	Type     string          `json:"type"`
+	Delta    string          `json:"delta"`
+	Item     *respOutputItem `json:"item"`
+	Response *respPayload    `json:"response"`
+}
+
+// Feed 消费一行 SSE data JSON,返回应立即写出的事件。
+func (st *AnthRespBridge) Feed(data []byte) []AnthEvent {
+	var ev respStreamEvent
+	if json.Unmarshal(data, &ev) != nil {
+		return nil
+	}
+	var events []AnthEvent
+	if !st.started {
+		st.started = true
+		events = append(events, st.startEvent())
+	}
+	switch ev.Type {
+	case "response.output_item.added":
+		if ev.Item != nil && ev.Item.Type == "function_call" {
+			events = append(events, st.openBlock(true, ev.Item.CallID, ev.Item.Name)...)
+			st.stopReason = "tool_use"
+		}
+	case "response.output_text.delta":
+		events = append(events, st.textDelta(ev.Delta)...)
+	case "response.function_call_arguments.delta":
+		events = append(events, st.argsDelta(ev.Delta)...)
+	case "response.completed", "response.incomplete", "response.failed":
+		st.finishFrom(ev.Response)
+	}
+	return events
+}
+
+func (st *AnthRespBridge) textDelta(delta string) []AnthEvent {
+	var events []AnthEvent
+	if !st.blockOpen || st.blockIsTool {
+		events = append(events, st.openBlock(false, "", "")...)
+	}
+	events = append(events, AnthEvent{Event: "content_block_delta", Data: map[string]any{
+		"type": "content_block_delta", "index": st.blockIndex,
+		"delta": map[string]string{"type": "text_delta", "text": delta},
+	}})
+	return events
+}
+
+func (st *AnthRespBridge) argsDelta(delta string) []AnthEvent {
+	if !st.blockOpen || !st.blockIsTool {
+		return nil
+	}
+	return []AnthEvent{{Event: "content_block_delta", Data: map[string]any{
+		"type": "content_block_delta", "index": st.blockIndex,
+		"delta": map[string]string{"type": "input_json_delta", "partial_json": delta},
+	}}}
+}
+
+// finishFrom 记录终态:usage 与 stop_reason(max_output_tokens 截断 → max_tokens)。
+func (st *AnthRespBridge) finishFrom(resp *respPayload) {
+	if resp == nil {
+		return
+	}
+	st.usage = anthUsageFromResp(resp.Usage)
+	if resp.Status == "incomplete" && resp.IncompleteDetails != nil &&
+		resp.IncompleteDetails.Reason == "max_output_tokens" {
+		st.stopReason = "max_tokens"
+	}
+}
+
+func (st *AnthRespBridge) startEvent() AnthEvent {
+	return AnthEvent{Event: "message_start", Data: map[string]any{
+		"type": "message_start",
+		"message": map[string]any{
+			"id": st.id, "type": "message", "role": "assistant", "model": st.model,
+			"content": []any{}, "stop_reason": nil,
+			"usage": map[string]int{"input_tokens": 0, "output_tokens": 0},
+		},
+	}}
+}
+
+// openBlock 关闭当前块并打开新块(text 或 tool_use)。
+func (st *AnthRespBridge) openBlock(isTool bool, toolID, toolName string) []AnthEvent {
+	var events []AnthEvent
+	if st.blockOpen {
+		events = append(events, st.closeBlockEvent())
+	}
+	st.blockOpen, st.blockIsTool = true, isTool
+	st.blockIndex++
+	block := map[string]any{"type": "text", "text": ""}
+	if isTool {
+		block = map[string]any{"type": "tool_use", "id": toolID, "name": toolName, "input": map[string]any{}}
+	}
+	events = append(events, AnthEvent{Event: "content_block_start", Data: map[string]any{
+		"type": "content_block_start", "index": st.blockIndex, "content_block": block,
+	}})
+	return events
+}
+
+func (st *AnthRespBridge) closeBlockEvent() AnthEvent {
+	return AnthEvent{Event: "content_block_stop", Data: map[string]any{
+		"type": "content_block_stop", "index": st.blockIndex,
+	}}
+}
+
+// Finish 在上游流结束后收尾:关块 → message_delta(stop_reason+usage)→ message_stop。
+func (st *AnthRespBridge) Finish() []AnthEvent {
+	var events []AnthEvent
+	if !st.started {
+		st.started = true
+		events = append(events, st.startEvent())
+	}
+	if st.blockOpen {
+		events = append(events, st.closeBlockEvent())
+		st.blockOpen = false
+	}
+	usage := map[string]int{"output_tokens": st.usage.OutputTokens}
+	if st.usage.InputTokens > 0 {
+		usage["input_tokens"] = st.usage.InputTokens
+	}
+	if st.usage.CacheReadInputTokens > 0 {
+		usage["cache_read_input_tokens"] = st.usage.CacheReadInputTokens
+	}
+	events = append(events,
+		AnthEvent{Event: "message_delta", Data: map[string]any{
+			"type":  "message_delta",
+			"delta": map[string]any{"stop_reason": st.stopReason, "stop_sequence": nil},
+			"usage": usage,
+		}},
+		AnthEvent{Event: "message_stop", Data: map[string]any{"type": "message_stop"}},
+	)
+	return events
+}
+
+// Usage 返回聚合到的用量(供调用日志)。
+func (st *AnthRespBridge) Usage() aiwire.AnthUsage { return st.usage }

internal/service/anthresponses_test.go

@@ -0,0 +1,173 @@
+package service
+
+import (
+	"encoding/json"
+	"strings"
+	"testing"
+
+	"oci-portal/internal/aiwire"
+)
+
+func mustAnthReq(t *testing.T, raw string) aiwire.MessagesRequest {
+	t.Helper()
+	var req aiwire.MessagesRequest
+	if err := json.Unmarshal([]byte(raw), &req); err != nil {
+		t.Fatalf("解析请求: %v", err)
+	}
+	return req
+}
+
+// TestAnthropicToResponsesBody 断言 system/消息/工具/effort 的直通装配与 store 强制。
+func TestAnthropicToResponsesBody(t *testing.T) {
+	req := mustAnthReq(t, `{
+		"model": "xai.grok-4.3", "max_tokens": 128, "system": "你是助手", "stream": true,
+		"temperature": 0.5,
+		"messages": [
+			{"role": "user", "content": "东京天气?"},
+			{"role": "assistant", "content": [
+				{"type": "text", "text": "查询中"},
+				{"type": "tool_use", "id": "t1", "name": "get_weather", "input": {"city": "东京"}}
+			]},
+			{"role": "user", "content": [
+				{"type": "tool_result", "tool_use_id": "t1", "content": "晴 25 度"},
+				{"type": "text", "text": "继续"}
+			]}
+		],
+		"tools": [{"name": "get_weather", "description": "查天气", "input_schema": {"type": "object"}}],
+		"tool_choice": {"type": "any"},
+		"output_config": {"effort": "HIGH"}
+	}`)
+	payload, err := AnthropicToResponsesBody(req)
+	if err != nil {
+		t.Fatalf("AnthropicToResponsesBody: %v", err)
+	}
+	var body map[string]any
+	if err := json.Unmarshal(payload, &body); err != nil {
+		t.Fatalf("unmarshal: %v", err)
+	}
+	if body["model"] != "xai.grok-4.3" || body["max_output_tokens"] != float64(128) ||
+		body["instructions"] != "你是助手" || body["store"] != false || body["stream"] != true {
+		t.Fatalf("顶层字段装配错误: %v", body)
+	}
+	if body["tool_choice"] != "required" {
+		t.Fatalf("tool_choice = %v, want required", body["tool_choice"])
+	}
+	reasoning, _ := body["reasoning"].(map[string]any)
+	if reasoning["effort"] != "high" {
+		t.Fatalf("effort = %v, want high(小写透传)", reasoning)
+	}
+	input, _ := body["input"].([]any)
+	// user 消息、assistant 文本消息、function_call、function_call_output、末条 user 文本
+	if len(input) != 5 {
+		t.Fatalf("input 项数 = %d, want 5: %s", len(input), payload)
+	}
+	kinds := make([]string, 0, len(input))
+	for _, it := range input {
+		m := it.(map[string]any)
+		if ty, ok := m["type"].(string); ok {
+			kinds = append(kinds, ty)
+		} else {
+			kinds = append(kinds, "message/"+m["role"].(string))
+		}
+	}
+	want := "message/user,message/assistant,function_call,function_call_output,message/user"
+	if got := strings.Join(kinds, ","); got != want {
+		t.Fatalf("input 顺序 = %s, want %s", got, want)
+	}
+	if !strings.Contains(string(payload), `"output_text"`) {
+		t.Fatalf("assistant 历史文本应为 output_text 部件: %s", payload)
+	}
+}
+
+// TestAnthropicToResponsesBodyRejects 断言不支持的内容块拒绝与图片装配。
+func TestAnthropicToResponsesBodyRejects(t *testing.T) {
+	bad := mustAnthReq(t, `{"model":"m","max_tokens":8,"messages":[
+		{"role":"user","content":[{"type":"document","source":{}}]}]}`)
+	if _, err := AnthropicToResponsesBody(bad); err == nil {
+		t.Fatal("document 块应拒绝")
+	}
+	img := mustAnthReq(t, `{"model":"m","max_tokens":8,"messages":[
+		{"role":"user","content":[{"type":"image","source":{"type":"base64","media_type":"image/png","data":"QUJD"}}]}]}`)
+	payload, err := AnthropicToResponsesBody(img)
+	if err != nil || !strings.Contains(string(payload), "data:image/png;base64,QUJD") {
+		t.Fatalf("图片应转 data URI: %v %s", err, payload)
+	}
+}
+
+// TestResponsesToAnthropic 断言输出块、stop_reason 与 usage 的回转。
+func TestResponsesToAnthropic(t *testing.T) {
+	payload := []byte(`{"model":"xai.grok-4.3","status":"completed","output":[
+		{"type":"reasoning","summary":[]},
+		{"type":"message","content":[{"type":"output_text","text":"你好"}]},
+		{"type":"function_call","call_id":"c1","name":"get_weather","arguments":"{\"city\":\"东京\"}"}],
+		"usage":{"input_tokens":10,"output_tokens":5,"total_tokens":15,"input_tokens_details":{"cached_tokens":4}}}`)
+	out, err := ResponsesToAnthropic(payload, "msg_1")
+	if err != nil {
+		t.Fatalf("ResponsesToAnthropic: %v", err)
+	}
+	if len(out.Content) != 2 || out.Content[0].Type != "text" || out.Content[0].Text != "你好" {
+		t.Fatalf("content 装配错误: %+v", out.Content)
+	}
+	if out.Content[1].Type != "tool_use" || out.Content[1].ID != "c1" || string(out.Content[1].Input) != `{"city":"东京"}` {
+		t.Fatalf("tool_use 装配错误: %+v", out.Content[1])
+	}
+	if out.StopReason != "tool_use" {
+		t.Fatalf("stop_reason = %s, want tool_use", out.StopReason)
+	}
+	if out.Usage.InputTokens != 10 || out.Usage.OutputTokens != 5 || out.Usage.CacheReadInputTokens != 4 {
+		t.Fatalf("usage = %+v", out.Usage)
+	}
+
+	trunc := []byte(`{"model":"m","status":"incomplete","incomplete_details":{"reason":"max_output_tokens"},
+		"output":[{"type":"message","content":[{"type":"output_text","text":"半"}]}]}`)
+	out2, err := ResponsesToAnthropic(trunc, "msg_2")
+	if err != nil || out2.StopReason != "max_tokens" {
+		t.Fatalf("截断 stop_reason = %v, %v", out2, err)
+	}
+}
+
+func bridgeEventTypes(events []AnthEvent) string {
+	kinds := make([]string, 0, len(events))
+	for _, ev := range events {
+		kinds = append(kinds, ev.Event)
+	}
+	return strings.Join(kinds, ",")
+}
+
+// TestAnthRespBridge 断言流桥:文本增量、工具调用切块、reasoning 丢弃与收尾事件。
+func TestAnthRespBridge(t *testing.T) {
+	st := NewAnthRespBridge("msg_1", "m1")
+	var events []AnthEvent
+	feed := func(lines ...string) {
+		for _, l := range lines {
+			events = append(events, st.Feed([]byte(l))...)
+		}
+	}
+	feed(`{"type":"response.created","response":{"model":"m1"}}`,
+		`{"type":"response.reasoning_text.delta","delta":"思考中"}`,
+		`{"type":"response.output_text.delta","delta":"你"}`,
+		`{"type":"response.output_text.delta","delta":"好"}`,
+		`{"type":"response.output_item.added","item":{"type":"function_call","call_id":"c1","name":"f"}}`,
+		`{"type":"response.function_call_arguments.delta","delta":"{\"a\":1}"}`,
+		`{"type":"response.completed","response":{"status":"completed","usage":{"input_tokens":8,"output_tokens":4}}}`)
+	events = append(events, st.Finish()...)
+
+	got := bridgeEventTypes(events)
+	want := "message_start,content_block_start,content_block_delta,content_block_delta," +
+		"content_block_stop,content_block_start,content_block_delta,content_block_stop," +
+		"message_delta,message_stop"
+	if got != want {
+		t.Fatalf("事件序列:\n got %s\nwant %s", got, want)
+	}
+	if st.Usage().InputTokens != 8 || st.Usage().OutputTokens != 4 {
+		t.Fatalf("usage = %+v", st.Usage())
+	}
+}
+
+// TestAnthRespBridgeEmpty 断言空流也产出完整事件骨架。
+func TestAnthRespBridgeEmpty(t *testing.T) {
+	st := NewAnthRespBridge("msg_1", "m1")
+	if got := bridgeEventTypes(st.Finish()); got != "message_start,message_delta,message_stop" {
+		t.Errorf("空流事件序列 = %s", got)
+	}
+}