release: bump version to 0.74.5

docs/CHANGELOG.md

@@ -8,6 +8,23 @@ This project follows the repository versioning rule:
 - `improvement` -> `+0.0.1`（bugfix + 小功能混合）
 - `bugfix` -> `+0.0.1`
 
+## [0.74.5] — 2026-09-13
+
+Released: 2026-09-13
+
+### Highlights
+- 恢复模型供应商预设的加载与持久化，模型刷新和轻量连接检查统一使用供应商官方目录接口。
+- 修复 MiniMax M3 思考模式和 OpenCode Go 模型协议路由，补齐 Responses 适配。
+
+### Added / Fixed / Improved
+- 修复 `llm_provider_preset:*` 动态设置分类触发 KeyError 和列表 500，使用真实数据库读写回归验证保存与重载。
+- 统一九家供应商的目录 URL、鉴权、响应解析与分页，保留服务地域和自定义地址；明确处理限流、空目录和失败，移除 404 白名单误判成功。
+- 模型刷新读取表单草稿地址和凭证，保留默认模型、未保存修改及失败前目录；连接提示区分目录可用与实际生成成功。
+- 增加 Responses 请求与响应解析、M3 adaptive 思考适配和 OpenCode 会话标识；已知模型选择同步协议，自定义 OpenAI 网关保留手动协议。
+- 同步中英文接口说明、操作手册与快速开始，并覆盖模型目录失败、草稿保留和协议适配回归。
+
+---
+
 ## [0.74.4] — 2026-09-13
 
 Released: 2026-09-13

docs/technical/en/agents-aiprovider.md

@@ -51,6 +51,7 @@ That split makes MiniMax, Claude-compatible gateways, and self-hosted OpenAI-com
 Supported request adapters:
 
 - `openai-completions`
+- `openai-responses`
 - `anthropic-messages`
 - `ollama-generate`
 
@@ -102,7 +103,29 @@ The AI settings page uses:
 
 These endpoints require an authenticated user. The `secrets` endpoint is only used when the settings page reveals a key or token; hiding the field restores the masked preview.
 
-`backend/app/services/llm_provider_catalog.py` refreshes catalogs from models.dev for most providers and from OpenCode Go's own models endpoint for that provider. Entries from models.dev are sorted by release date, newest first, with undated entries last. Successful results and `refreshed_at` are stored per provider in the `system_settings` category `llm_provider_preset:<provider>`. Listing prefers saved catalogs and uses bundled presets for providers that have not been refreshed. Refresh does not write the active model, protocol, URLs, or credentials in `external_integrations`. Failure returns 502 and preserves the previous catalog without exposing raw upstream errors. The frontend reloads catalog state while retaining the form draft.
+Model refresh and lightweight checks share `backend/app/services/llm_model_catalog.py` and query provider endpoints directly, without models.dev. Refresh uses the draft base URL, protocol and credentials, or saved configuration when no draft is supplied. Regional hosts and custom gateway paths are preserved. Missing required credentials return 400; upstream failures return safe errors and retain the previous catalog.
+
+Successful catalogs and `refreshed_at` are stored under `llm_provider_preset:<provider>` in `system_settings`. Dynamic categories use the provider preset as their defaults and are validated before committing. Listing prefers saved catalogs and otherwise shows bundled suggestions. Refresh never modifies the active model, protocol, URL or credentials in `external_integrations`. The frontend retains the draft and previously loaded models when listing fails, with an explicit error message.
+
+### Model Catalog Endpoints
+
+These endpoints were checked against official documentation on 2026-09-13. The table lists discovery URLs; the form still takes the generation base URL. Credentials must belong to the configured service region.
+
+| Provider | Model list GET endpoint | Authentication and parsing |
+| --- | --- | --- |
+| [MiniMax](https://platform.minimax.io/docs/api-reference/models/anthropic/list-models) | `https://api.minimaxi.com/anthropic/v1/models`; international: `api.minimax.io` | `x-api-key`; `data[].id`; `has_more/last_id` pagination |
+| [OpenAI](https://developers.openai.com/api/reference/resources/models/methods/list) | `https://api.openai.com/v1/models` | Bearer; `data[].id` |
+| [Anthropic](https://platform.claude.com/docs/en/api/models/list) | `https://api.anthropic.com/v1/models` | `x-api-key`, `anthropic-version`; paginate with `after_id` |
+| [DeepSeek](https://api-docs.deepseek.com/api/list-models) | `https://api.deepseek.com/v1/models` | Bearer; `data[].id` |
+| [Alibaba Model Studio](https://help.aliyun.com/zh/model-studio/list-models) | `/api/v1/models` on the configured regional host | Bearer; `output.models[].model`; `page_no/page_size/output.total`; filter `capabilities=TG` |
+| [Moonshot / Kimi](https://platform.kimi.ai/docs/api/list-models) | `https://api.moonshot.ai/v1/models`; China: `api.moonshot.cn` | Bearer; `data[].id` |
+| [OpenRouter](https://openrouter.ai/docs/api/api-reference/models/list-all-models-and-their-properties) | `https://openrouter.ai/api/v1/models` | Public catalog; Bearer when supplied; `data[].id` |
+| [OpenCode Go](https://opencode.ai/docs/go/#models) | `https://opencode.ai/zen/go/v1/models` | Public catalog; Bearer when supplied; `data[].id` |
+| [Ollama](https://docs.ollama.com/api/tags) | `/api/tags` on the configured server | Local servers need no key; `models[].model/name`; empty means no installed models |
+
+New Model Studio endpoints for Beijing, Tokyo, Frankfurt and Virginia require the actual workspace host, such as `<WorkspaceId>.cn-beijing.maas.aliyuncs.com`. Singapore uses `dashscope-intl.aliyuncs.com`; Hong Kong uses `cn-hongkong.dashscope.aliyuncs.com`. Discovery changes only the path, never guesses a workspace or switches credential regions. If a legacy Beijing host no longer accepts the account, update the base URL from the provider console.
+
+Pagination and retries have a 30-second overall deadline. Network failures and 502/503/504 allow up to two attempts; authentication failures are not retried. Models are sorted by supplied creation/release dates, preserving upstream order on ties, without arbitrary list truncation. Public catalog membership does not establish account-specific generation permission.
 
 Admin keeps the AI page aligned with the legacy information architecture:
 
@@ -286,22 +309,11 @@ Tool keys follow the same rule. WebSearch and OCR must match the current tool an
 
 ### Lightweight Connectivity Testing
 
-The Admin plug button performs a lightweight connectivity check and does not save configuration. Common API-platform practice is two-tiered:
-
-- Check a provider catalog or low-cost endpoint to validate base URL, authentication, and model reachability.
-- Send full model requests only when the user explicitly runs Playground or a business task.
-
-Connectivity results should be explicit:
-
-- `ok`: authentication, route, and model catalog are usable.
-- `warning`: service is reachable, but the current model is missing from the catalog or capability metadata is incomplete.
-- `error`: authentication, network, protocol, or model lookup failed.
-
-Toast titles must match the result; failures must not be titled as a successful connection.
+The Admin plug button checks proxy configuration and queries the current provider catalog. It returns `success=true` only when discovery succeeds and contains the selected model. A 404, authentication error, invalid response or missing model never becomes success based on bundled suggestions. The message distinguishes discovery from generation; Playground or a business task verifies actual generation. Testing never saves the draft or switches the default provider.
 
 ### OpenCode Go Routing Model
 
-Subscription channels such as OpenCode Go should not be handled by hard-coded frontend model sets. Prefer provider catalog or backend capability discovery that records per-model capabilities such as `chat_completions`, `anthropic_messages`, `models_endpoint`, and whether a subscription key is required. The frontend should display capabilities; the backend should map provider, base URL, model, and adapter into the real request.
+OpenCode Go protocol mappings are owned by the backend catalog: MiniMax M3/M2.7/M2.5 and the verified Qwen3.6/3.7/3.8 models use Anthropic Messages; GPT-5.6 Luna, Grok 4.6 and Muse Spark Contributor use Responses; other supported models use Chat Completions. The official `/models` response currently supplies IDs only, so new models still require checking the documented endpoint table. Selecting a model updates its protocol in the draft, and backend mappings supersede known stale mappings. Responses uses `input`, `max_output_tokens` and `store=false`, and parses text and reasoning summaries. OpenCode calls carry an application User-Agent and a stable Playground conversation identifier. Generic `thinking=enabled` is translated to MiniMax M3’s `adaptive` format.
 
 #### Settings Page Behavior
 

docs/technical/en/manual.md

@@ -163,10 +163,12 @@ Providers and models accept presets or arbitrary custom IDs. Common fields:
 - Max Tokens, Anthropic Version: keep defaults if unsure
 - Timeout / Retry: timeout and retry attempts
 
-The plug icon at the end of the Base URL input runs a connection test. A passing test echoes the model's short reply.
+The plug icon at the end of the Base URL input runs a lightweight connection check: it checks the proxy and model catalog, then confirms the selected model is listed. Use Playground afterward to verify generation access and the model's reply; the catalog check itself does not generate a response.
 
 Use the refresh icon at the top right to update the available models. A successful refresh saves the catalog for later visits; catalogs with release dates show newer models first. Refresh preserves the current model, credentials, URLs, and unsaved edits. Select a model and click Save to change the model used by the application. A failed refresh displays an error and keeps the previous catalog.
 
+Refresh uses the base URL, protocol, and API key currently entered in the form, so you can fill new configuration and refresh before saving. Credentialed providers require a key for the matching service region; an empty Ollama catalog means no models are installed. Selecting a model with a known protocol also updates the protocol field. Custom OpenAI gateways retain the manually selected protocol.
+
 ### Tools
 
 - **WebSearch**: provider, API key, base URL, max results, timeout, advanced provider parameters. While disabled, all fields except the enable switch are greyed out

docs/technical/en/quickstart.md

@@ -33,7 +33,7 @@ The default role is `viewer`: you can sign in but only see public pages. For col
 After landing on the `/admin` dashboard, here's a recommended walk-through:
 
 1. `/collection-management?section=collector_credentials`: pick a collector and click the plug icon to test connectivity. Free collectors (e.g. open BGP) usually work right away; credential-bearing ones like `AISStream` or `BarentsWatch` need an API key / client secret first
-2. `/ai?section=integrations`: fill an LLM provider (e.g. `minimax` / `openai`), model, base URL, API key, and click the plug at the end of the base URL to test. The refresh icon updates and saves the available model catalog while preserving your form; select a model and click Save to apply it. WebSearch / OCR tools are optional
+2. `/ai?section=integrations`: fill an LLM provider (e.g. `minimax` / `openai`), model, base URL, API key, and click the plug at the end of the base URL to check the model catalog; verify generation in Playground. The refresh icon updates and saves the available model catalog while preserving your form; select a model and click Save to apply it. WebSearch / OCR tools are optional
 3. `/earth-content?section=brand`: maintain the Earth HUD logo and title image in Branding. The Upload button inside each URL field opens a file picker, and image files can also be dropped directly onto the matching field. Save the brand configuration after upload
 4. `/datasources` or `/data`: check whether collectors have produced data. Use `/datasources -> Built-in Sources` for finite collectors: with no rows selected, click `Trigger All`; after selecting rows, the primary button becomes `Trigger Selected N`. The top-right queue button shows progress. Use `/datasources -> Realtime Sources` for AISStream / WebSocket health and counters
 5. `/alerts/system`: verify system alerts look right

docs/technical/zh/agents-aiprovider.md

@@ -51,6 +51,7 @@
 支持的请求适配器：
 
 - `openai-completions`
+- `openai-responses`
 - `anthropic-messages`
 - `ollama-generate`
 
@@ -102,7 +103,29 @@ AI 配置页使用的接口：
 
 这些接口都需要用户登录。`secrets` 接口只用于配置页点击显示 key/token 时取回明文，隐藏时前端恢复为脱敏预览。
 
-模型目录刷新由 `backend/app/services/llm_provider_catalog.py` 获取上游目录；多数供应商读取 models.dev，OpenCode Go 使用自己的模型接口。models.dev 条目按发布日期倒序排列，缺少日期的条目排在后面。成功结果和 `refreshed_at` 按供应商保存在 `system_settings` 的 `llm_provider_preset:<provider>` 分类中，列表接口优先返回已保存目录，未刷新过的供应商使用内置预设。刷新只更新目录，不写入 `external_integrations` 中的当前模型、协议、地址或凭证。失败返回 502 并保留上次目录；上游异常原文不返回客户端。前端重新加载目录，可选模型读取目录状态，表单草稿独立保留。
+模型目录刷新和轻量连通性测试共用 `backend/app/services/llm_model_catalog.py`，直接请求当前供应商的模型接口，不再依赖 models.dev。使用当前表单草稿中的基础地址、协议和凭证；不传草稿时读取已保存配置。国内、国际和自定义网关地址保持各自的地域与路径。缺少必需凭证时返回 400 并提示配置 API Key；上游超时、鉴权或响应错误返回安全提示，并保留上次目录。
+
+成功目录和 `refreshed_at` 按供应商保存到 `system_settings` 的 `llm_provider_preset:<provider>` 分类。该动态分类使用对应供应商预设作为默认值，提交前校验分类，避免已提交却返回失败。列表优先返回已保存目录，未刷新过的供应商显示内置建议。刷新不修改 `external_integrations` 的当前模型、协议、地址或凭证；页面保留草稿，目录加载失败时保留已显示模型并明确提示失败。
+
+### 模型目录接口
+
+以下接口于 2026-09-13 对照官方文档核对。表中的 URL 是模型列表地址，配置表单仍填写生成接口的基础地址。需要凭证的目录必须使用对应服务地域的 API Key。
+
+| 供应商 | 模型列表 GET 接口 | 鉴权与解析 |
+| --- | --- | --- |
+| [MiniMax](https://platform.minimax.io/docs/api-reference/models/anthropic/list-models) | `https://api.minimaxi.com/anthropic/v1/models`；国际站使用 `api.minimax.io` | `x-api-key`；`data[].id`；处理 `has_more/last_id` |
+| [OpenAI](https://developers.openai.com/api/reference/resources/models/methods/list) | `https://api.openai.com/v1/models` | Bearer；`data[].id` |
+| [Anthropic](https://platform.claude.com/docs/en/api/models/list) | `https://api.anthropic.com/v1/models` | `x-api-key`、`anthropic-version`；按 `after_id` 翻页 |
+| [DeepSeek](https://api-docs.deepseek.com/api/list-models) | `https://api.deepseek.com/v1/models` | Bearer；`data[].id` |
+| [阿里百炼](https://help.aliyun.com/zh/model-studio/list-models) | 同地域主机的 `/api/v1/models` | Bearer；`output.models[].model`；按 `page_no/page_size/output.total` 翻页，筛选 `capabilities=TG` |
+| [Moonshot / Kimi](https://platform.kimi.ai/docs/api/list-models) | `https://api.moonshot.ai/v1/models`；国内站使用 `api.moonshot.cn` | Bearer；`data[].id` |
+| [OpenRouter](https://openrouter.ai/docs/api/api-reference/models/list-all-models-and-their-properties) | `https://openrouter.ai/api/v1/models` | 支持公共目录；有凭证时附带 Bearer；`data[].id` |
+| [OpenCode Go](https://opencode.ai/docs/go/#models) | `https://opencode.ai/zen/go/v1/models` | 支持公共目录；有凭证时附带 Bearer；`data[].id` |
+| [Ollama](https://docs.ollama.com/api/tags) | 本机或配置服务器的 `/api/tags` | 本地无需 Key；`models[].model/name`；空目录表示尚未安装模型 |
+
+百炼的北京、东京、法兰克福、弗吉尼亚新接口要求实际业务空间主机，例如 `<WorkspaceId>.cn-beijing.maas.aliyuncs.com`；新加坡使用 `dashscope-intl.aliyuncs.com`，香港使用 `cn-hongkong.dashscope.aliyuncs.com`。系统只替换目录路径，不猜测业务空间或跨地域切换凭证。旧北京域名若不再接受账号凭证，应按控制台给出的业务空间地址更新基础地址。
+
+完整分页和重试受 30 秒总超时保护；网络故障及 502/503/504 最多尝试两次，鉴权失败不重试。按接口提供的创建/发布日期倒序排列；日期相同时保留上游顺序，不截断为固定数量。公共目录只表示供应商公开的模型集合，不代表账号已获得每个模型的生成权限。
 
 Admin 的 AI 页面按业务信息架构组织为：
 
@@ -286,22 +309,11 @@ Admin 的密钥状态必须按 provider / tool 精确判断：
 
 ### 轻量连通性测试
 
-Admin 的插头按钮走轻量连通性测试，不承担保存职责。业界常见做法是分两层：
-
-- 快速检查 provider 目录或低成本 endpoint，确认 base URL、鉴权和当前模型是否可达。
-- 只有在用户明确运行 Playground 或业务任务时才发完整模型请求。
-
-因此，连通性测试应尽量使用低成本请求，并返回明确状态：
-
-- `ok`: 鉴权、路由和模型目录可用。
-- `warning`: 服务可达，但当前模型不在目录或能力声明不完整。
-- `error`: 鉴权失败、网络失败、协议错误或模型不可用。
-
-错误 toast 标题必须和结果一致，不能在失败时显示“连通性正常”。
+Admin 插头按钮检查代理服务配置，再查询当前供应商目录；只有目录请求成功且包含所选模型时才返回 `success=true`。404、鉴权失败、无效响应和缺失模型都不会因命中内置建议而被改判成功。提示明确区分“找到模型”和“实际生成成功”；完整调用由 Playground 或业务任务验证。测试不保存表单或切换默认供应商。
 
 ### OpenCode Go 路由模型
 
-OpenCode Go 这类订阅通道不要靠前端硬编码模型集合判断协议。推荐在 provider catalog 或后端能力发现中记录每个模型的协议能力，例如 `chat_completions`、`anthropic_messages`、`models_endpoint` 和是否需要订阅 key。前端只展示能力结果；后端负责把 provider、base URL、model 和协议适配映射为实际请求。
+OpenCode Go 的模型协议映射由后端目录集中维护：MiniMax M3/M2.7/M2.5 和已核对的 Qwen3.6/3.7/3.8 模型使用 Anthropic Messages；GPT-5.6 Luna、Grok 4.6 和 Muse Spark Contributor 使用 Responses；其余已支持模型使用 Chat Completions。官方 `/models` 当前只提供模型 ID，新增模型的协议仍需对照官方端点表核对。前端选择模型时同步对应协议，后端覆盖旧的已知错误映射。Responses 请求使用 `input`、`max_output_tokens` 和 `store=false`；回复解析文本及推理摘要。OpenCode 请求携带应用 User-Agent 和会话标识；Playground 同一会话保持标识稳定。MiniMax M3 的通用 `thinking=enabled` 转为官方 `adaptive` 格式。
 
 #### 配置页行为
 

docs/technical/zh/manual.md

@@ -166,10 +166,12 @@ provider 和模型既可选预设也可直接输入自定义 id/name。常用字
 - Max Tokens、Anthropic Version：可保持默认
 - Timeout / Retry：超时和重试次数
 
-Base URL 输入框尾端的插头图标会触发连接测试。测试通过会显示当前模型返回的简短回复。
+Base URL 输入框尾端的插头图标会触发轻量连接测试：检查代理和模型目录，并确认当前模型出现在目录中。通过后可到 Playground 发起实际生成，核对账号权限和模型回复；目录检查本身不执行生成调用。
 
 点击右上角的刷新图标可更新“可选模型”。成功后目录会保存，重新打开页面仍可使用；有发布日期的目录按新到旧排列。刷新保留当前模型、密钥、地址和未保存的修改。点击一个可选模型后，再点“保存”才会改变实际使用的模型。刷新失败时会显示错误并保留上次目录。
 
+刷新使用当前表单中的基础地址、协议和 API Key，因此可以先填写新配置再刷新，无需提前保存。需要凭证的供应商必须填写对应地域的 Key；Ollama 的空目录表示尚未安装模型。选择已知协议的模型时，协议适配会同步到表单；自定义 OpenAI 网关保留手动选择的协议。
+
 ### 工具
 
 - **WebSearch**：provider、API Key、Base URL、最大结果数、超时、高级 provider 参数。未启用时除"启用"开关外其它配置项和连接测试都会置灰

docs/technical/zh/quickstart.md

@@ -33,7 +33,7 @@
 进入 `/admin` 仪表盘后，建议按这个顺序熟悉控制台：
 
 1. `/collection-management?section=collector_credentials`：选一个采集器，点插头图标做连接测试。免费 collector（开源 BGP 等）通常直接可用；像 `AISStream`、`BarentsWatch` 这类需要凭证的，需要先填 API Key/Client Secret
-2. `/ai?section=integrations`：填一个 LLM provider（例如 `minimax` / `openai`）、模型名、Base URL、API Key，点 Base URL 末端的插头测试连接。右上角刷新图标会更新并保存可选模型目录，同时保留当前表单；选中新模型后点“保存”生效。WebSearch / OCR 工具可选
+2. `/ai?section=integrations`：填一个 LLM provider（例如 `minimax` / `openai`）、模型名、Base URL、API Key，点 Base URL 末端的插头检查模型目录；实际生成在 Playground 验证。右上角刷新图标会更新并保存可选模型目录，同时保留当前表单；选中新模型后点“保存”生效。WebSearch / OCR 工具可选
 3. `/earth-content?section=brand`：在“品牌标识”里维护智能星球的 Logo 和标题图；对应地址字段内的“上传”按钮支持选择文件，也支持把图片直接拖到字段上，保存后会应用到智能星球 HUD
 4. `/datasources` 或 `/data`：看采集器是否已经产出数据。有限采集器看 `/datasources -> 内置源`，不勾选时点“触发全部”，勾选后主按钮会变成“触发已选 N”；右上角队列按钮可查看进度。AISStream / WebSocket 长连接看 `/datasources -> 实时源` 的健康状态和计数
 5. `/alerts/system`：看系统告警是否正常

docs/version-history.md

@@ -16,12 +16,13 @@
 ## Current Version
 
 - `main` 当前主线历史推导到：`0.16.5`
-- `dev` 当前开发分支历史推导到：`0.74.4`
+- `dev` 当前开发分支历史推导到：`0.74.5`
 
 ## Timeline
 
 | Version | Type | Branch | Commit | Summary |
 | --- | --- | --- | --- | --- |
+| `0.74.5` | improvement | `dev` | `v0.74.5` | 恢复模型预设持久化，统一官方目录刷新与连通性检查，修复 M3 思考模式和 OpenCode 协议路由 |
 | `0.74.4` | improvement | `dev` | `v0.74.4` | Earth 全量渲染与船舶增量更新优化，直播目录搜索和分页，定位队列状态恢复、模型目录保存及启动提速 |
 | `0.74.3` | improvement | `dev` | `v0.74.3` | Ubuntu / WSL 初始化自动准备 Docker 及用户权限，修正启动诊断，并在建表前核对数据库端口、实际连接和认证 |
 | `0.74.2` | bugfix | `dev` | `pending` | 收敛 agent harness 到根规则和 Codex skills，删除旧 Claude command 重复入口，并强化视觉证据路径解析与 OCR fallback 规则 |