DeepSeek Harness プラグイン — 814 ページ

模型无关的沙箱升级修正 + 按模型定制的协议提示段：任何模型（GPT 系最常见）在工具调用中错误携带 sandbox_permissions/justification 导致 "not strictly wider" 死循环时，在工具执行前自动剥离；并按会话模型动态注入适配该模型的协议提示段。

ZCode-style model picker for the DSH web GUI — provider submenus that open upward, reasoning-effort selection, and keyboard navigation. · DSH 的 ZCode 式模型选择器：向上展开的提供方子菜单、推理等级选择与键盘导航。

把模型选择器按供应商折叠起来——每个供应商只占一行，点名称才展开它的模型，不用再一长条滚到底。

模型选择菜单按来源（provider）展开分类：默认菜单只列来源，点来源右侧滑出子面板选模型；双击来源标题切换为列表内展开（v0.2.0 行为，已备份）

DeepSeek Harness 输入框的模型挡位面板：模型选择、推理强度滑条、显示名，外加闪电编排三档（唯一的提示词注入）· Model gear for the DSH composer; the only prompt text it adds is the Lightning contract

Platform and manufacturer grouped model picker for DeepSeek Harness

DSH plugin that lists all configured models with availability/latency testing, by reading $DSH_HOME/settings.yaml.

DSH 模型健康检查：在会话视图新增「模型健康检查」页签，按「供应商(baseURL分组) → API类型 → 模型 → 路由」四行按钮选定目标，手动发送一条真实裸 HTTP 测试请求（支持 openai-completions / openai-responses / anthropic-messages 三协议，流式与非流式可切），一屏展示耗时、TTFT、HTTP 状态、token 用量、响应内容与三层诊断链（API请求 / 模型响应 / 严格校验），错误全部人话化。可把当前模型一键设为默认（下次打开即预选）。另提供用户显式启动、成功即停的有限自动重试会话（严格/连通两种成功判定、4 条停止条件、致命错误不短路）。不做后台监控、不轮询、不告警、不落盘。

Vendor marks in the DeepSeek Harness model picker: the host half mirrors the provider's model catalog and logo set to loopback, and the browser half draws one monochrome mark per model row, the way the provider's own dashboard draws it.

DSH model-menu enhancements: dots for models in use by other running sessions, per-model reasoning-effort memory, collapsed group state, and menu spacing fixes.

Show a hover hint with a model's full configuration (input types, context window, max tokens, protocol, base URL, credential env, …) over each model in the DeepSeek Harness web model picker, read from ~/.dsh/settings.yaml.

Structured cross-model peer review for DeepSeek Harness

Sync your DeepSeek Harness model configuration and third-party API keys across machines, over S3-compatible storage or WebDAV.

Model parameter editing, imports, and native USD/CNY pricing for DeepSeek Harness.

Model high/low ladder: force a high-tier model for a turn window (composer widget next to the native selector), then auto-fall back to the Session's own selection; native model UI untouched

OpenRouter free-models browser plugin (dual-face): fetch OpenRouter models, filter the free ones (price 0 or :free), list them newest-first in a Settings tab with local search and full model info, and create a ready-to-use DSH model config (settings.yaml route + stored credential) in one click. Bilingual UI (zh/en), no hardcoded paths, publishable to npm / GitHub

DSH plugin: manage local model inference servers (llama.cpp / vLLM / SGLang) — registry, safe start/stop, parameter profile versioning with validation. Red lines: never pkill, never stop protected port 11437, never auto-restart dsh.

Model tags, reasoning capabilities, image routing, and vision proxy support for DeepSeek Harness

Capability metadata for hand-declared gateway models, resolved at runtime by bare model name (context window, output cap, reasoning levels, vision) — matched against the official DeepSeek route's catalog, the installed pi-ai catalog, opencode and models.dev, without writing to your settings.

Declare whether a configured third-party model accepts image (multimodal) input; writes the modality into the owning provider settings and verifies it through runtime model resolution.

Per-model input type (text/image) checkboxes on the dsh Models page, plus a per-model 思考等级 field in the provider edit dialog. Writes the official llm-pi-ai `input` and `reasoningEfforts` fields.

内置 provider / 模型显示名的全量数据库：载入时按环境里实际启用的 provider 与模型覆盖显示名，只影响 UI，不改路由、请求与凭据。

A DeepSeek Harness plugin with a vertical model list and an animated Three.js particle reasoning-effort control.

Generic Model Orchestrator for DeepSeek Harness: discovers the live model pool, profiles real model capabilities, and routes each unit of work to the best available model. Auto and Guided modes, specialist subagents, multi-agent orchestration, a dynamic Captain, and a web control panel. Domain-agnostic.