Plugins

Registers system-prompt sections that make the model reason in Simplified or Traditional Chinese, with strict/prefer modes and an optional reply-language policy. Adds a /thinking-zh command for per-session switching. Zero runtime dependencies.

Syncs DSH model providers and models across computers through any WebDAV server: the backup is sealed with a passphrase before it leaves this machine, so a new machine opens with the same model list.

Inspects ONNX models (inputs/outputs, shapes, opset, operators, size) and benchmarks them with ONNX Runtime on the local machine or on a remote target over SSH, reporting average/P50/P95/P99 latency, FPS and the execution provider actually used, with PASS/FAIL gates for FPS, P95 latency and model size. Requires Python 3.10+ with numpy, onnx, onnxruntime, and psutil installed (pip install numpy onnx onnxruntime psutil); remote targets need python3, numpy, and onnxruntime.

Turns what a design, plan or document claims into a model and decides it: UML diagrams rendered, read and reviewed; state machines, protocols and data models checked as models; migrations tested for coverage and rollback symmetry; concurrency claims mined and routed to the checker that can decide them; dependency graphs audited; reports diffed.

Extends the composer model selector menu with a third Context window row, preset capacities (64K–2M), custom input, provider-wide sync, and adaptive compaction guard.

Generic monitor plugin for DSH Desktop: an HTTP polling rule engine. Poll any endpoint on schedule (multi-step, variable extraction, JS condition expressions); when conditions hit, wake a visible agent session to act — zero model calls otherwise. Configure everything in a sidebar GUI panel: create/edit/pause/test/fire-now/run history, same UX as the built-in scheduler plugin.

Expose a running DeepSeek Harness instance over MCP, so an external agent can list, create and continue that instance's own sessions, delegate a turn and collect its result, answer the questions the DSH agent asks, and set how future approval requests resolve.

Shows what each installed skill costs in the model context and hides a skill from the catalog without deleting it. Lists every skill found on disk with the exact catalog line the model receives (description normalized and truncated at the catalog cap), its body size and token estimate; previews the catalog message byte-for-byte as the agent will receive it, compares that with the copy already published in the current session, and reports skills on disk that never reach the catalog (invalid frontmatter, or shadowed by a higher-priority root). Hiding writes disable-model-invocation into frontmatter and keeps the file, so the skill still works via /name. Also imports skills from ClawHub, Skills.sh and GitHub URLs and uninstalls user-owned ones.

Skill manager page for dsh: set each skill to on, manual only, model only, or off per agent mode; count how often each skill is loaded per mode from the session history; search skills.sh and install from GitHub; import from Codex, Claude Code, or a folder; create, edit, and version skills, with a trash for deleted ones; modes keep working after an uninstall.

OCR for scanned, image-only PDFs into Markdown, built around the fact that a book takes about an hour: every page is checkpointed so an interrupted run resumes, a progress log reports live status, and a quality report names empty pages and low-confidence lines. Requires Python 3.10+ and pip install -r requirements; the first run downloads the OCR model.

Sorts every loaded skill into one or more user-owned workspaces in a right-sidebar tab: each workspace is its own category list plus where every skill sits, categories are created, renamed, reordered and deleted in place, and skills are dragged or picked between them. A first run can let the profile model build the first classification, or place skills into categories the user built first; clicking a row copies its /skill-name and hovering unfolds the description the skill ships.

Tag-based skill filtering: attach tags to DSH skills, then hand each workspace or conversation only the subset its boolean tag expression selects (for example `pdf + docx` or `aegis * plugin`). Excluded skills are deny-shadowed out of the model catalog rather than merely hidden in the settings UI; skills outside the expression can still be added by hand or force-disabled, and the bundled `tagging-skills` teaches the agent to derive a tag set from the machine it is running on.

DocChain knowledge base, document conversion and knowledge-graph toolkit — filename, content and metadata search, locate-by-doc_id/name with content reading, topic listing, bulk document upload, document-to-Markdown conversion, image-similarity search, per-document image descriptions saved to a sidecar markdown (flowcharts and architecture diagrams are turned into PlantUML sketches, other images into text descriptions; the original image is replaced by a clickable link), a deep-research workflow, and a knowledge-graph layer (ontology inspection, read-only Cypher, ontology/entity/relation writes). Every script-backed capability also ships as a native DSH tool executed in-process: doc search, content search, filename search, topic info, metadata search, upload, image search, image describe, doc2markdown, plus kg schema / kg cypher / kg upsert node / kg upsert relation / kg delete node / kg delete relation. All graph operations go through the DocChain API, so no Neo4j password is needed. The configuration page keeps several docchain environments (each with its own address and credential), switches between them in one click, and pre-selects topics per environment. Needs a docchain service URL and API key (except doc2markdown); image descriptions additionally need an OpenAI-compatible multimodal endpoint (defaults to an internal gateway model, configurable).

Windows x64 translator with a dedicated sidebar entry and Workbench / Chat interface, side-by-side text and document translation, screen-region translation, recent history, and host model selection, voice input and preview. CUDA document OCR with CPU fallback; screen OCR uses CPU. Do not enable the CPU and CUDA packages at the same time.

Windows x64 translator with a dedicated sidebar entry and Workbench / Chat interface, side-by-side text and document translation, screen-region translation, recent history, and host model selection, voice input and preview. CPU OCR without NVIDIA dependencies. Do not enable the CPU and CUDA packages at the same time.

Two-way voice for the DSH web app: dictate with the unified composer, hear replies with speaker buttons and auto-read, on any device. Needs DSH 0.1.7+, Node 22+, Python 3 with numpy and requests, a TTS model, and a microphone.

Meeting notes for dsh: one meeting per conversation, transcribed and speaker-separated on the machine itself, with points, assignments, handwritten notes, reviews and monthly scores in the same panel. Exports SRT/Markdown/Word and packs a meeting into a portable zip another dsh can import; confirmed assignments go to DingTalk/WeCom in one click, following your robot settings. Recognition models are swappable (a stronger local model can be imported from a file), and cloud recognition is off by default with no request made before consent. Word export needs Python and python-docx; the first run downloads a recognition model (28 or 77 MB).

Search public-domain libraries (Project Gutenberg, Wikisource, Internet Archive, Standard Ebooks) by title, author, character or half-remembered plot, or import EPUB / PDF / TXT / DOCX / Markdown, and listen in the sidebar with Microsoft Edge neural voices, offline Kokoro or the device voices; keeps your position, takes voice notes anchored to what you just heard, and discusses the book with your DSH model.

Structured image recognition: local sharp measurements plus a cloud vision model return description, OCR with coordinates, object locations and fine-grained attributes in one call; a fingerprint cache makes repeated images return in milliseconds without spending model quota.

Cloudflare Workers AI media tools for dsh - text-to-image, text-to-speech and speech-to-text. Per-model price badges, a native settings card, reuse of an existing Cloudflare account, and output files named after the model that made them.

Lets the model actually see a video: local ffmpeg samples timestamped stills and faster-whisper transcribes the audio track, so the agent reads real pixels frame by frame instead of trusting a text summary. Fully local, nothing uploaded; reports honestly when the audio has no speech. Windows only; needs Python 3.10+ and ffmpeg installed, and the first video_eyes_setup run downloads about 700 MB (faster-whisper and its model).

Tencent Yuanbao Search provider for the ctx.web seam: one direct Search API request per query returns portable title, URL, snippet and ISO-8601 date, with credentials resolved per search and no auxiliary model turn. Requires a Tencent Cloud WSA API key.

System One decision runtime for dsh. Ships fast non-generative tools (choice, score, screen, gate) on a local or hosted System One gateway, and — without the model calling anything — prunes oversized tool output segment by segment, refuses repeated identical calls once System One confirms stagnation, verifies completion claims against real tool evidence, and corrects failing retries. Published on npm as dsh-harness-reflex (plugin) and harness-reflex (core, also a stdio MCP server).

Per-workspace gate that hides the tool schemas and system prompt sections of selected plugins from the model in directories that do not need them.