Agent 連携 · DeepSeek Harness プラグイン — 34 ページ

One place for every API credential DeepSeek Harness and its plugins use: every reference (env name) the model routes, media APIs, web search, desktop pet, memory embeddings and cost meter point at, whether it is configured and where it is stored, which features are bound to it, an owner-given alias/note, and set / replace / delete — through the core credential store, never a plain file.

Windows Credential Manager backend for DeepSeek Harness credentials: the stock local provider with its reference half stored in the system keyring instead of a plaintext file. Secrets migrate on write; every keyring failure falls back to the inherited behavior.

A desktop companion for DeepSeek Harness: its own persona, lorebook and multimodal API, a user profile distilled from the harness's own chat history, awareness of the running task, proactive chatter, reminders, voice, 8-24 frame sprite animations, screen reading and computer control behind a three-level permission, and a click-to-talk desktop window (WinForms sprite or Edge app window) next to the in-app floating pet.

Drop any file onto the DeepSeek Harness chat: non-image files are saved into the session workspace's .dsh-uploads/ folder and referenced from the composer as @.dsh-uploads/<name>, so the agent reads them with its file tools. Images keep going to the core's own attachment rail.

DeepSeek Harness bundle that mounts gemini-web-mcp as mcp__gemini-web__* tools

One card under Settings → Plugins for the community session importer (dsh-chat-import): what happens to a source transcript's system prompt on import — dropped by default, at most a context message inside the imported session when the importer's switch is on, never dsh's own system prompt or a plugin's section. The importer's own settings page belongs to that plugin (single-owner locale namespace, no slot inside it), so the sentence lives here.

Deny unpaired LAN devices the bare /api when the DeepSeek Harness web server binds all interfaces: removes the auto-derived LAN IP authorities from the connection trust list in place, without any core change, so pairing-fronted remote access is safe by default and a core upgrade never reopens the hole.

Per-model context window, max output and thinking levels for every DeepSeek Harness route (DeepSeek official, OpenRouter, local LM Studio / Ollama …); probes local backends and auto-teaches their levels. Managed from the DSH Launcher "Model parameters" page; OpenRouter :online web-search variants on the Model parameters page.

Extra media APIs for DeepSeek Harness: image generation, video generation, text-to-speech and transcription behind one settings panel, with adapters for OpenAI-compatible endpoints, Google Gemini/Veo, Replicate, fal, ElevenLabs, Fish Audio, MiniMax and any custom HTTP endpoint.

Lightweight memory for DeepSeek Harness: stores compaction summaries and extracted facts, recalls them lexically (BM25, CJK-aware) or with local embeddings, injects relevant items as a separate context message, offers memory_recall / memory_note tools, and lets the user view and edit every session's active compaction summary.

OOMOL Connector integration for DeepSeek Harness: connect apps and call Actions through progressive MCP discovery.

Reach the DeepSeek Harness on your dev machine from your phone — across networks, through no server at all (iroh P2P)

Hover price hints on the dsh model picker: shows input/output USD per 1M tokens (from the cost-meter price table) as a native tooltip on every model entry.

Keeps an OpenRouter route's model list in DeepSeek Harness current: pulls the public catalog on a schedule (and on demand), adds new models, refreshes names / context windows / output limits, and declares the reasoning-effort levels for every model that supports them — so the reasoning picker works instead of answering 'unsupported'.

Create a dsh workspace by absolute path over HTTP, so the DSH Launcher can add one without the native folder dialog (works around the ungrouped '+' being inert for the no-workspace group).

Deny destructive shell commands (rm -rf /, git push --force, raw-device writes, drive formats) at the dsh tools/pre-execute gate before they run; extra deny / ask-first regexes hot-load from ~/.dsh/safe-guard.json.

Load a user frontend skin ($DSH_HOME/frontend-skin.css) into the dsh Web UI as a token-override style tag. Managed by DSH Launcher.

Skin Studio for DeepSeek Harness: a model-facing skin_studio tool that lets the agent design and one-click install launcher or frontend skins on user request — with a guided flow (ask requirements first; request an image from the user when the model cannot generate one, or call an image-generation tool when it can), local image inlining as data URIs, and instant apply through the DSH Launcher API.

Live Agent Teams task, communication, subagent, and Workflow topology for DeepSeek Harness

Temporary DeepSeek Harness conversations that belong to no project: a scratch directory under $DSH_HOME/scratch, its own workspace, and a shipped chat-only agent preset with no file, shell, job, subagent or workflow tools.

Token-usage statistics for the DeepSeek Harness settings UI: summary cards + 3-month heatmap of real token spend converted to official peak/off-peak prices (CNY/USD), incremental index [fork: peak/off-peak price display removed]

Unofficial DeepSeek Harness integration for Unreal Engine 5.8 Unreal MCP

图片理解桥接(dsh-vision-bridge 的中文界面分支):聊天里的图片交给你选定的视觉模型描述,非视觉模型也能"看图"。Fork of @goodandready/dsh-vision-bridge 0.4.5 (MIT) with a Chinese UI; the wire, routes and config file are unchanged.

SillyTavern-style web search for DeepSeek Harness, configured once in dsh Settings: off / let the model decide (tool call served by Serper, SerpApi, Tavily, Brave, SearXNG or DeepSeek official) / search on trigger words / use the API provider's own search (OpenRouter :online, billed per search); a web_fetch page reader (public addresses only) built from the core's own tool definition with an address-pinned transport; optional page visiting that opens the top results and hands their article text to the model (built-in extraction, or a local trafilatura service); keys through the dsh credential store.