dsh-more 0.0.4 → 0.0.6

package/README.md

@@ -15,8 +15,9 @@ Practical, independently switchable patches for missing context and history cont
 | Edit and restart | User-message action row | Creates a clean continuation before the selected turn, submits the edited text, archives the source branch, and opens the child session | On |
 | Delete one message | User- and assistant-message action rows | Removes the selected model-surface node, expands only when tool call/result pairing requires it, rebuilds the surviving context, and opens a child session | On |
 | Permanently delete session | Separate item beside the native Archive action | Stops and unloads the target session, removes its exact JSONL persistence directory, detaches it from workspaces, and synchronizes the lists incrementally | On |
+| Model capabilities | Settings → Models → Model capabilities | Configure image support, selectable thinking types and parameter formats per model | On |
 
-All three switches live under **Settings → Plugins → Plugin configuration → DSH More** and take effect without rebuilding the plugin.
+Patch switches live under **Settings → Plugins → Plugin configuration → DSH More** and take effect without rebuilding the plugin.
 
 ## Installation
 
@@ -131,16 +132,23 @@ After confirmation, DSH More cancels a running task, waits for idle, flushes the
 
 > **Irreversible:** permanent deletion does not archive the session and does not move it into a DSH More trash directory.
 
-### Configure model reasoning effort
+### Configure model capabilities
 
-Open **Settings → Models**, expand a provider with **Edit**, and every model entry in the model catalog carries its own **Thinking effort** button:
+Open **Settings → Models**, expand a saved provider with **Edit**, and click **Model capabilities** below a model:
 
-1. Click a model's **Thinking effort**; in the dialog, check the levels the model offers (off / minimal / low / medium / high / xhigh / max) and, for each enabled level, fill in the wire spelling sent to the gateway (defaults to the level name); `off` needs no wire value — "supported, send nothing".
-2. Click **Save**. The change is written to `reasoningEfforts` under `llm-pi-ai.providers.<route>.models[]` in `settings.yaml`, and the composer's model picker immediately offers those levels.
+1. **Vision**: inherit defaults, support images, or accept text only.
+2. **Thinking types**: inherit defaults, disable reasoning, or choose supported levels (off / minimal / low / medium / high / xhigh / max) and their API parameter values.
+3. **Thinking parameter format**: expand the advanced option to select an installed DSH format such as OpenAI, DeepSeek, Qwen or OpenRouter. This applies to OpenAI Completions compatible interfaces; other protocols use native dispatch.
 
-You can also check **This model does not reason** (writes `reasoningEfforts: false`) or use **Clear** to drop the field and fall back to the installed catalog. Only `off` may leave its wire value empty, and at least one non-`off` level is required — the same rules the pi-ai adapter enforces when resolving a route. Invalid configurations are rejected before any write, so a bad save never corrupts `settings.yaml`.
+**Save** writes the model's `input` (`[text, image]` or `[text]`), `reasoningEfforts`, and `compat.thinkingFormat` under `llm-pi-ai.providers.<route>.models[]` in `settings.yaml`. DSH applies the changes through its live settings mechanism.
 
-**Auto-fill:** models added via **Fetch available models** carry no `reasoningEfforts`, which the adapter treats as non-reasoning. After the provider is saved, this patch auto-fills the default `{ low, medium, high }` onto any model that has no `reasoningEfforts`, so a freshly fetched model can think immediately. To keep a model non-reasoning, set it to `reasoningEfforts: false` (explicit `false` is never auto-filled). The default levels can then be tuned per model from the thinking-effort button.
+Each setting independently offers **Inherit defaults**, removing that model declaration and using DSH's catalog, provider configuration or adapter defaults. Missing declarations do not prove support: unknown models normally default to text input with no selectable thinking levels. A custom reasoning list needs at least one non-off level. The off parameter may be blank or use an explicit API value.
+
+This replaces the old model reasoning effort patch and **never auto-fills low / medium / high**. Existing model declarations remain editable: the old patch recorded no provenance, so upgrades cannot distinguish automatically filled levels from deliberate configuration. Disabling the patch removes its controls. Native vision declarations and pi-ai settings remain effective; native DeepSeek per-model reasoning restrictions require this patch to stay enabled.
+
+Writes carry the revision read when the dialog opens. Concurrent changes cause a conflict instead of an overwrite; reopen the dialog to retry. Save new models in the provider editor first. Native DeepSeek is also supported: image input is stored in `llm-deepseek.models[].inputModalities`; per-model thinking declarations use the model's `reasoningEfforts` extension field, applied by this patch to selectors and requests. Native thinking levels are fixed to off / low / high / max; a deployment-wide disabled policy remains authoritative.
+
+If the dialog reports incompatible frontend and backend versions, restart DSH and refresh the page. Rebuilding or refreshing only the frontend cannot replace a Host module already loaded in memory. An incompatible response disables saving and shows an explicit diagnostic.
 
 ## Architecture
 

package/README.zh-CN.md

@@ -15,8 +15,9 @@
 | 编辑并重新开始 | 用户消息操作区 | 在所选轮次之前创建干净延续，提交编辑后的文本，归档源分支并打开子会话 | 开启 |
 | 单条消息删除 | 用户和助手消息操作区 | 删除所选 model-surface 节点；只有工具调用/结果配对需要时才扩大范围；重建存活上下文并打开子会话 | 开启 |
 | 永久删除会话 | 原生“归档会话”旁的独立菜单项 | 停止并卸载目标会话，移除其精确 JSONL 持久化目录，从 Workspace 脱离并增量同步列表 | 开启 |
+| 模型能力配置 | 设置 → 模型 → 模型能力 | 按模型配置识图、可选思考类型及接口参数格式，不自动猜测能力 | 开启 |
 
-三个开关都位于 **设置 → 插件 → 插件配置 → DSH More**，修改后实时生效，无需重新构建插件。
+补丁开关位于 **设置 → 插件 → 插件配置 → DSH More**，修改后实时生效，无需重新构建插件。
 
 ## 安装
 
@@ -131,16 +132,23 @@ dsh plugin --profile web remove dsh-more
 
 > **不可恢复：** 永久删除不会归档会话，也不会把会话移入 DSH More 回收目录。
 
-### 配置模型思考强度
+### 配置模型能力
 
-打开 **设置 → 模型**，点击某个提供方的 **编辑** 展开「模型目录」，每个模型条目下都有一个 **思考强度** 按钮：
+打开 **设置 → 模型**，点击已保存提供方的 **编辑**，每个模型下方都有 **模型能力** 按钮：
 
-1. 点击模型的 **思考强度**，在弹窗中勾选该模型支持的级别（不思考 / 极简 / 低 / 中 / 高 / 极高 / 最高），并为每个开启的级别填写发送给网关的线上值（默认与级别同名）；`不思考` 不需要线上值，表示“支持、但不发送任何参数”；
-2. 点击 **保存**。修改会写入 `settings.yaml` 中 `llm-pi-ai.providers.<route>.models[].reasoningEfforts`，会话里的模型选择器随即按此提供级别。
+1. **识图能力**：沿用默认 / 支持识图 / 不支持识图（仅文字）。
+2. **思考类型**：沿用默认 / 不支持思考 / 自定义可选类型；自定义时可选择不思考、极简、低、中、高、极高、最高，并配置各等级发送给接口的参数值。
+3. **思考参数格式**：在高级选项中选择 OpenAI、DeepSeek、Qwen、OpenRouter 等 DSH 支持的格式。该选项仅用于 OpenAI Completions 兼容接口，其他协议使用原生处理。
 
-也可以勾选 **该模型不支持思考**（写入 `reasoningEfforts: false`），或用 **清除配置** 移除该字段、回到内置目录的能力。只有 `off` 允许留空线上值，且必须至少声明一个非 `off` 的级别，与 pi-ai 适配器的解析规则一致；无效配置会在写入前被拒绝，不会破坏 `settings.yaml`。
+点击 **保存**，补丁会将识图声明写入模型的 `input`（`[text, image]` 或 `[text]`），将思考能力写入 `reasoningEfforts`，将参数格式写入 `compat.thinkingFormat`。这些字段位于 `settings.yaml` 的 `llm-pi-ai.providers.<route>.models[]`，由 DSH 的实时配置机制应用。
 
-**自动填充：** 通过 **获取可用模型** 添加的模型默认没有 `reasoningEfforts`，适配器会把它当作不思考。本补丁会在保存提供方后，为**没有任何 `reasoningEfforts` 的模型**自动填入默认思考强度 `{ low, medium, high }`，让新拉取的模型立刻可以思考；如需某个模型不思考，请把它设为 `reasoningEfforts: false`（显式 false 不会被自动填充）。默认级别可用每个模型条目下的思考强度按钮再微调。
+各项均可独立选择 **沿用默认**，移除模型的对应声明，使用 DSH 的模型目录、提供方配置或适配器默认值。未声明不代表已确认支持：未知模型通常仅支持文字输入，且没有可选思考等级。自定义思考至少需要一个非「不思考」等级；「不思考」的参数值可留空，也可显式填写接口要求的值。
+
+本补丁替代原来的「模型思考强度」补丁，**不再自动填入低 / 中 / 高等级**。此前保存的模型声明会保留并可继续编辑；由于旧补丁没有记录自动填充来源，升级时不会猜测并删除已有声明。关闭补丁会移除入口和弹窗。识图声明与自定义提供方的思考配置继续由 DSH 应用；原生 DeepSeek 的逐模型思考限制需要保持本补丁开启。
+
+保存前会检查配置版本；如果其他窗口已修改配置，本次保存会被拒绝，请重新打开弹窗后编辑。新模型需要先保存提供方配置。同时支持自定义提供方与原生 DeepSeek：原生识图声明写入 `llm-deepseek.models[].inputModalities`，逐模型思考声明写入同一模型的 `reasoningEfforts` 扩展字段，并由本补丁应用到模型选择器与实际请求。原生 DeepSeek 仅支持 off / low / high / max，参数值固定；提供方整体关闭思考时，不能用模型设置绕过。
+
+如果出现“界面与后台版本不一致”，请重启 DSH 服务再刷新页面。仅刷新浏览器或重建前端无法替换后台已经加载的旧模块；版本不匹配时保存会被禁用并显示原因。
 
 ## 架构