dsh-plugin-local-infer 0.1.3 → 0.1.4

package/README.md

@@ -5,7 +5,7 @@
 > For those who want more performance out of their specific hardware than what Unsloth gives out of the box.
 
 [![License: MIT](https://img.shields.io/badge/license-MIT-green)](https://github.com/cybtachyon/dsh-plugin-local-infer/blob/main/LICENSE)
-[![DSH 0.2.1-alpha.3](https://img.shields.io/badge/DSH-0.2.1--alpha.3-blue)](https://www.npmjs.com/package/@deepseek-ai/dsh)
+[![DSH 0.2.2-alpha.3](https://img.shields.io/badge/DSH-0.2.2--alpha.3-blue)](https://www.npmjs.com/package/@deepseek-ai/dsh)
 [![Type: DSH Plugin](https://img.shields.io/badge/Type-DSH%20Plugin-8A2BE2)](https://github.com/topics/dsh-plugin)
 
 A DeepSeek Harness plugin that manages local inference engines:
@@ -35,6 +35,10 @@ dsh plugin --profile web add dsh-plugin-local-infer
 dsh plugin --profile web add https://github.com/cybtachyon/dsh-plugin-local-infer
 ```
 
+When installing from a local checkout, run `pnpm build` first: it type-checks
+the workspace (`tsc -b`) and then produces the `dsh/client.js` browser card
+artifact with esbuild (`node scripts/build-client.mjs`).
+
 Web UI: click **Plugins** → **Add Plugin** → paste `dsh-plugin-local-infer`.
 
 Or add from the DSH Plugin Marketplace.
@@ -47,12 +51,26 @@ The **Engine settings** card mounts on the Plugins page (older hosts: the Settin
 - **Watch the VRAM Est. meter.** It prices weights, KV cache, overhead, and workspace against your card as you adjust values, and shows the verdict (`fits` / `tight` / `over`) and the largest context that fits.
 - **Read the warnings.** Red banners and inline field warnings fire when a parameter value, or the presence of a parameter, is unsafe for your detected hardware.
 - **Satisfy the system dependencies.** The panel lists each missing host dependency (docker, CUDA toolkit, ffmpeg, hf CLI, ...) with its purpose, a copy/paste install string for your OS, and an optional one-click **Run** (sudo). The **Get** button stays disabled until every dependency reports satisfied.
-- **Get.** With the dependencies satisfied, `Get` syncs the engine's source: it clones the engine's git repo (or the **Engine Fork** you provided) into the **Engine Directory** override, defaulting to the model directory under the engine name. An existing checkout with local changes aborts with an alert in **Logs**; a clean one fetches the listed fork (added as its own remote) and rebases onto the upstream branch. It then runs the system dependencies and the engine's build steps. Every line streams into the **Logs** section as it lands. `Launch` stays disabled while a Get runs and re-enables when the build finishes, errors, or falls silent for 2 minutes.
+- **Get.** With the dependencies satisfied, `Get` syncs the engine's source: it clones the engine's git repo (or the **Engine Fork** you provided) into the **Engine Directory** override, or with no override into the engine default checkout home `~/.cache/dsh/infer/<engine>` (one checkout per engine, derived at run time and never persisted, never inside a model directory). An existing checkout with local changes aborts with an alert in **Logs**; a clean one fetches the listed fork (added as its own remote) and rebases onto the upstream branch. It then runs the system dependencies and the engine's build steps. Every line streams into the **Logs** section as it lands. `Launch` stays disabled while a Get runs and re-enables when the build finishes, errors, or falls silent for 2 minutes.
 - **Launch / Stop.** `Launch` deploys the engine (builds the checkout, or pulls the image in docker mode), downloads the model, starts it, and polls its health endpoint. `Stop` shuts it down.
 - **Register / Remove.** `Register` publishes the confirmed-working engine to DSH as a model (optionally as the agent default model) through the settings service, so the change is revision-tracked and undoable. `Remove` pulls the entry back.
 
+When a parameter can be set both as a CLI argument and as a write to the engine's config file, the plugin prefers the CLI argument. The CLI arg is consumed for that launch only, so it leaves the on-disk config file (and any customization you made by hand in it) untouched where possible.
+
 All values persist to `~/.dsh/local-infer/config.json`; the browser never sees a value it did not store.
 
+A **Get** or **Launch** that must run `sudo` first checks whether sudo credentials are cached. When they are not, both Launch buttons read **Sudo needed** (painted as a warning) and the line `Check your CLI for a sudo prompt` appears in the **Logs** section and the run's log file: the password prompt lands on the terminal that launched the harness, not in the browser. Type the password there. Without one within 60 seconds the run is killed and fails through the normal error path, and the buttons switch to **Logs**.
+
+## Orchestrator package
+
+Local models often struggle with agentic workflows and can't "one-shot" prompts without assistance. The included **Orchestrator** package gives your model a default prompt that enables that capability.
+
+- **Enable it** on the DeepSeek Harness **Plugins** page: open this plugin's entry and flip the **Orchestrator** component toggle (the component ships disabled). After first enabling the component, restart the host so it loads.
+- **Edit the prompt** on the **Orchestrator** card that then appears: change the **Orchestrator Prompt** field and press **Save**. The default text ships in the plugin's `prompt.md`; clear the field and save to restore it.
+- **It injects while enabled.** With the component enabled, the saved prompt is part of the agent's system prompt for new requests. A Save reaches the next request; no restart is needed after Save.
+
+The saved prompt persists to `~/.dsh/local-infer/orchestrator.json`, beside the engine config.
+
 ## Supported Platforms
 
 ### Operating systems
@@ -64,7 +82,7 @@ All values persist to `~/.dsh/local-infer/config.json`; the browser never sees a
 | Linux (Fedora, `dnf`)        | Untested  | Command set present, flagged untested on a live host.                                                               |
 | macOS (`brew`)               | Partial   | CPU/docker path (llama.cpp). No NVIDIA Container Toolkit on macOS, so the CUDA engines' docker path is unavailable. |
 | Windows (WSL2)               | Untested  | WSL2 is detected and the dflash2 pin-memory rule fires; the command set is flagged untested.                        |
-| Windows                      | Partial   | Code paths are only partially written; further development needed.                                                  |  
+| Windows                      | Partial   | Code paths are only partially written; further development needed.                                                  |
 
 ### Engines
 
@@ -77,7 +95,7 @@ All values persist to `~/.dsh/local-infer/config.json`; the browser never sees a
 | Infernix          | No            | Yes (cmake + CUDA) | C++/CUDA fork of NInfer tuned for Qwen models on a single RTX 50-series GPU. Local build only. |
 | Strata            | Yes           | No                 | Placeholder engine; its parameter surface is empty until the flag set lands. |
 
-Runtime requirements: Node `^22.19.0 || >=24.0.0`, pnpm 12.11.2 (pinned by `mise.toml`), and the `@deepseek-ai/cordis >= 4.0.1` peer dependency.
+Runtime requirements: Node `^22.19.0 || >=24.0.0`, pnpm 12.11.2 (pinned by `mise.toml`), and the `@deepseek-ai/cordis >= 4.0.6-alpha.2` peer dependency.
 
 ## Hardware guide
 

package/cordis.patch.yml

@@ -40,3 +40,10 @@
     # model confirmation.
     - id: local-infer-ui
       name: 'dsh-plugin-local-infer/ui'
+
+    # Orchestrator: the opt-in "Orchestrator" card and its system-prompt
+    # injection. Ships disabled: the Plugins page Components toggle writes the
+    # enabling override in the user's profile patch layer.
+    - id: local-infer-orchestrator
+      name: 'dsh-plugin-local-infer/orchestrator'
+      disabled: true

package/docs/zh/README.md

@@ -5,7 +5,7 @@
 > 面向那些希望从特定硬件中获得比 Unsloth 开箱即用更高性能的用户。
 
 [![License: MIT](https://img.shields.io/badge/license-MIT-green)](https://github.com/cybtachyon/dsh-plugin-local-infer/blob/main/LICENSE)
-[![DSH 0.2.1-alpha.2](https://img.shields.io/badge/DSH-0.2.1--alpha.2-blue)](https://www.npmjs.com/package/@deepseek-ai/dsh)
+[![DSH 0.2.2-alpha.3](https://img.shields.io/badge/DSH-0.2.2--alpha.3-blue)](https://www.npmjs.com/package/@deepseek-ai/dsh)
 [![Type: DSH Plugin](https://img.shields.io/badge/Type-DSH%20Plugin-8A2BE2)](https://github.com/topics/dsh-plugin)
 
 一个 DeepSeek Harness 插件，用于管理本地推理引擎：
@@ -51,8 +51,20 @@ Web UI：点击 **插件** → **添加插件** → 粘贴 `dsh-plugin-local-inf
 - **启动 / 停止。** `启动` 部署引擎（构建检出，或在 docker 模式下拉取镜像）、下载模型、启动它，并轮询其健康端点。`停止` 将其关闭。
 - **注册 / 移除。** `注册` 通过设置服务将确认工作正常的引擎作为模型发布到 DSH（可选地作为代理默认模型），因此该变更可被修订跟踪并可撤销。`移除` 将该条目拉回。
 
+当某个参数既可以作为 CLI 参数设置，又可以写入引擎配置文件时，插件优先使用 CLI 参数。CLI 参数仅在本次启动时生效，因此在可能的情况下，它不会改动磁盘上的配置文件（以及你在其中手动做的任何自定义）。
+
 所有值都持久化到 `~/.dsh/local-infer/config.json`；浏览器永远不会看到它未存储的值。
 
+## 编排器包
+
+本地模型往往难以胜任 agentic 工作流，没有辅助就无法"一次成型"（one-shot）地完成提示词任务。内置的**编排器（Orchestrator）**包为你的模型提供一个默认提示词，使其具备这项能力。
+
+- **启用：** 在 DeepSeek Harness 的**插件**页面打开本插件的条目，切换 **Orchestrator** 组件的开关（该组件随包默认禁用）。首次启用组件后，需重启主机以加载它。
+- **编辑提示词：** 在随后出现的 **Orchestrator**（编排器）卡片上，修改**编排器提示词**字段并点击**保存**。默认文本随插件的 `prompt.md` 一并发布；清空该字段并保存即可恢复默认文本。
+- **启用期间注入。** 组件启用后，保存的提示词会作为代理系统提示词的一部分用于新请求。保存后即对下一个请求生效，保存后无需重启。
+
+保存的提示词持久化到 `~/.dsh/local-infer/orchestrator.json`，与引擎配置同目录。
+
 ## 支持的平台
 
 ### 操作系统
@@ -77,7 +89,7 @@ Web UI：点击 **插件** → **添加插件** → 粘贴 `dsh-plugin-local-inf
 | Infernix          | 否           | 是 (cmake + CUDA)  | NInfer 的 C++/CUDA 分支，针对单张 RTX 50 系 GPU 的 Qwen 模型调优；仅本地构建。 |
 | Strata            | 是           | 否                | 占位符引擎；在其标志集落地之前，其参数表面为空。   |
 
-运行时要求：Node `^22.19.0 || >=24.0.0`、pnpm 12.11.2（由 `mise.toml` 固定），以及 `@deepseek-ai/cordis >= 4.0.1` 同级依赖。
+运行时要求：Node `^22.19.0 || >=24.0.0`、pnpm 12.11.2（由 `mise.toml` 固定），以及 `@deepseek-ai/cordis >= 4.0.6-alpha.2` 同级依赖。
 
 ## 硬件指南