Plugins

Chat image-attachment bridge with a `view_image` tool for any OpenAI-compatible VLM (local Ollama or cloud): pasted/dropped images become `view_image` path markers before reaching text-only DeepSeek models.

Model-facing image_describe (识图) tool over the DashScope OpenAI-compatible API (qwen3.7-flash), plus a paste bridge: on text-only sessions, pasted images auto-convert to file paths at send time and render back in the transcript, so they never trip image admission. Bring your own DASHSCOPE_API_KEY; endpoint/model/budgets configurable, redirect-proof HTTP client, works in every agent preset.

Vision for any DSH route: paste images in the Web composer with intent-aware auto-analysis, delegate workspace image reads to a Kimi/MiniMax vision subagent, and materialize pasted originals for editing.

Local-first structured vision for text-only agents: images go to a local OpenAI-compatible VLM and come back as JSON evidence (summary, verbatim OCR, layout regions, entities/relations, colors, explicit uncertainty), with anti-hallucination fallback and an optional paste/upload bridge; zero cloud cost, images never leave the machine.

Measures design rules on the rendered page rather than in the CSS — type-scale count, WCAG AA contrast against the real composited backdrop, spacing-grid drift, palette size, tap targets, and the tells of generated UI such as violet gradients caught by hue angle; ships a design-system skill.

A `search` tool that learns to pick between fast and thorough search strategies using a contextual bandit (Thompson sampling), improving from real usage instead of a fixed heuristic.

Automatic provider failover on rate limits or quota exhaustion: switch between providers with configurable per-provider models, a cooldown, and a permanent last-entry fallback.

fastCRW-backed web search and fetch providers for DeepSeek Harness (ctx.web)

Plug-in vision for text-only DeepSeek Harness (dsh) models: a `vision` tool with built-in cheap/free VLM presets, multi-image batch analysis, paste-to-hint image admission, and a web settings page with hot-reload.

Native OS notifications for approval and question/plan prompts: the toast names the workspace, click jumps to it, bilingual (zh-CN/zh-TW/en), with chime.

Cross-session memory for the model: full-text search all past sessions (SQLite FTS5 via ctx.sessionQuery) and bring the strongest matching excerpts back into the current context.

DeepSeek Harness 通用通知插件：新会话 / 待审批 / 任务完成 / 任务中断时，在 GUI 内显示 Telegram 风格应用内通知并播放提示音；焦点在应用外时调用系统原生通知（Windows 10+ Toast / macOS 通知中心），仅通知、不回调。

Unified @ mentions in the DeepSeek Harness web composer: five built-in categories (file, folder, skill, past chat, plugin) plus a data-source seam any plugin can register through, references to absolute paths outside the workspace, and a candidate menu that can be scoped by a folder named earlier in the draft and pages with PageUp/PageDown instead of hiding rows.

Shared infrastructure for data assetization plugins: business rules loader, file format adapter, report generator, path validator, audit logger Also known as @deepseek-ai/dsh-data-asset-shared.

Reopens the conversation you were last in after restarting DeepSeek Harness, by keeping a pointer file under the DSH home directory instead of patching engine files.

Discovers OpenRouter's free models, measures their real speed and context size, and writes them into the DeepSeek Harness model list fastest-first, with speed and context in each display name, plus a Settings section to configure and run it.

Motor-design AI assistant plugin: requirement parsing → parameter identification → candidate matrix → Top10 verification. L0 is a millisecond-level broad filter (pure JS, zero solver dependency) that ranks candidates and emits the handoff payload for the L1/L2 precision tools; usable as a local agent preset and Cordis plugin. L1 (RMxprt) and L2 (Motor-CAD) are reserved in this version — the level gate fails fast above l0.

DSH plugin: audio chimes, cross-session toasts, desktop push, and IM webhooks for turn completion, errors, and approvals.

Multi-instance ComfyUI orchestration: probes every configured ComfyUI endpoint for its GPU, free VRAM and queue depth, then dispatches a generation job to the idlest eligible instance. Ships three tools (comfyui_farm_status, comfyui_farm_pick, comfyui_farm_run). An instance is eligible only if it is reachable, has at least the requested free VRAM, and its queue is no deeper than the configured ceiling; eligible instances are ranked by free VRAM minus queue depth times a weight, so an idle shallow queue wins over a busy large one. Zero runtime dependencies. Optional retry on probe failure, because a single timeout across an SSH tunnel is not proof that an instance is down.

Code graph and knowledge base over an embedded FalkorDB: AST symbols, caller/callee tracing, refactor blast radius, affected-test discovery, clone detection, cross-language porting contracts and spec-to-code traceability.

Deterministic cross-session transcript retrieval for DeepSeek Harness: the model-facing `recall` tool searches past session logs with explicit scope control

Desktop notifications for DeepSeek Harness — a web notification while a browser page holds the plugin (turn completion and pending questions), and a native Windows toast from the host for finished turns once the browser is closed.

Version-control workbench for the DSH Web sidebar: local changes grouped by type, per-file and multi-file staging, preview-confirmed commit / update / push, and a failure dialog that shows the raw Git output instead of a four-second toast.

Session cleanup for the DSH sidebar: drag a conversation row onto the balance widget's whale and it eats the row — the session directory goes to a recycle bin, not to rm. Deletion asks one short question first (how many tokens it burned, and that prompt can be turned off in Settings). Undo from the persistent toast or the 12-entry log in Settings. Clicking the whale jumps straight to the plugin's Settings, where you can swap the avatar, drag the mouth into place and rewrite all 17 UI strings; blank, currently-open and running sessions are refused in place with a reason.