Plugin

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.

All-media (video/audio/image) sibling of liustack/modlens: turns local media into structured evidence (timeline / transcript / layout regions) injected into the conversation so text-only models can cite concrete frames, seconds and speakers — Evidence, not imagination. One failover chain per media, recorded hop-by-hop in meta.attempts (a fallback is never silent): host-registered multimodal models first, then a configured OpenAI-compatible endpoint (comma-separated multi-key rotation), then local ffmpeg frame-sampling / a local ASR CLI, degrading to metadata-only as a last resort. Absorbs modlens battle-tested rules — cooldown keyed only on auth/quota failures and softened to chain-tail (never dropped), redact-before-clip error text, magic-byte sniff that never trusts file extensions. Optional vision-bridge registers a media-lens wrapper into the host model selector so a pure-text model gains image/audio/video sight. Registering / authenticating multimodal models and true inline audio-video upload lean on two companion plugins, declared as optional peers (absence auto-degrades, never errors): dsh-thinking-levels fronts the model config / registration surface, dsh-llm-openai-completions takes over the OpenAI-compatible transport that serializes file blocks into video_url / input_audio. Privacy: media goes to remote providers by default; strictLocalOnly keeps content on-machine. Prerequisites not shipped with the package: ffmpeg/ffprobe on PATH (frame sampling / reencode / duration probe; degrades to metadata when missing) and a user-configured local ASR CLI for the audio fallback. Hosts 0.1.7 / 0.2.0 / 0.2.1 (6 rc, exact engines.dsh enum). npm: dsh-media-lens.

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).

Registers one `valuation_service` tool that reads a workpaper tree of 01/02/04 JSON and returns one unified schema: six ratio groups, forecast ratios, DCF input slots with explicit missing markers, comparable PS/PE and IRR. Deterministic and local: missing WACC, growth, capex, D&A or working-capital inputs stay blank instead of being guessed, and assumption differences are flagged for review.

Requires Claude Code installed and signed in on this machine. Delegates coding tasks to the local Claude Code agent through the Claude Agent SDK, reports which files changed, supports one-command revert, and adds a Claude Mode toggle to the composer.

Run Python on the local NVIDIA CUDA GPU from DeepSeek Harness: device status, foreground and detached jobs, durable logs, and a measured peak-VRAM figure per run.

Requires Python 3.12+; Xiaohongshu reading needs the xiaohongshu-mcp backend and a QR sign-in (ffmpeg/OCR optional). Batch-read Xiaohongshu notes and WeChat public-account articles: per-note body, image text via local OCR, video frame text and comment samples, structured JSON/Markdown output, and the original source URL on every result. Read-only: it never publishes, likes, follows or comments.

Run the local Cursor CLI agent on a coding task in a workspace and continue the run by chat id.

Read Bilibili and Douyin from a session: WBI-signed Bilibili API calls behind one shared pacer (4 req/s, per-host cooldown after a 412) for video, feed, search, author, comments, danmaku and subtitles, plus Douyin hot list, video, feed and search rendered in a locally launched Edge over CDP.

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).

Built for DSH 0.2.1-rc.2 (engines.dsh pins it exactly). Technical support for Dragonfly 3D World using 2027.1 help documents and read-only searches of a detected local installation, with an explicit help-only fallback.

Frontend aesthetics engine as 17 native tools — a 118-palette evidence library matched by register, cross-register whole-page design directions, four mechanical gates (WCAG contrast, color-blind simulation, AI-slop detection, scorecard) with digest-bound verdicts, plus semantic design tokens, existing-page audit, CVD-checked chart palettes, CJK type scale, render-vs-tokens reconciliation and measured-numbers rationale. Fully local, zero network. Requires Python 3.8+.

Drives a local Hiddify installation from DeepSeek Harness without its GUI: seven tools to start and stop the core, switch split-tunnel modes (rebuilds the config and restarts the core), rotate exit nodes through the Clash control plane, and search or fetch pages through the proxy. Windows only.

Moves DSH plugins between machines and profiles: export a migration bundle where local plugins keep their source and remote ones reinstall by version, restore it elsewhere with the bundle's own restore.mjs, copy plugins to another profile, or pack the whole environment (config, memory, sessions, attachments, plugin sources) into checksummed volumes with a restore script.

Turns a learning request and local materials into a structured, locally stored AI course with outline, lessons, diagrams and learning progress.

Turn English audiobooks (MP3) and documents into listening lessons: transcribe every sentence locally with the speech model DSH voice input already downloaded (about 241 MB, fetched on first use if missing; ffmpeg required), proofread and explain each sentence with your DSH model, then drill with blind listening, intensive listening and spaced repetition, with Chinese-dubbing export and offline Kokoro voices; ships a sidebar panel and four AI tools.

Five production-grade DSH tools bundled into one plugin (8 tool slots total): git regression finder (find introducer, recent changes, bisect, blame age), stack-trace decoder with local code context, YAML/ENV/properties config validator with typo suggestions, cross-platform shell escape (POSIX/PowerShell/cmd), and MyBatis Plus SQL static auditor (N+1, missing @Transactional, LIKE leading %, SELECT *, missing WHERE). Zero npm dependencies, all offline; concurrency-safe except the bisect tool (dsh-when-broke). Each sub-plugin is also installable standalone (dsh-when-broke, dsh-stackdecoder, dsh-config-guard, dsh-cmd-shield, dsh-mp-auditor).

Ship a local DSH plugin project to GitHub and the plugin market in one call: preflight, per-file upload with byte verification, the market registry PR, then pull every published file back and compare bytes.

One screen for every work you have shipped: local project facts, GitHub stars and last push, whether the plugin market has picked it up and how many downloads it has, plus pending pull-request wait times against the upstream median.

Windows GUI automation reliability layer: probes whether a window is minimized, foreground or DPI-scaled, then clicks and types with the coordinate space and activation order handled for you. Zero dependencies, local only.

Hand a local file to any program's file picker: wait for the native Open dialog, paste the path through the clipboard, read it back to verify character by character, then confirm. The target is always explicit and every step leaves evidence. Windows only (EnumWindows / UI Automation).

Take a local video all the way to a Douyin creator post: open the upload page, trigger the picker, hand the file over with verification, fill the title, set visibility - then stop before publishing unless explicitly told to publish.

Local BM25 auto-recall, a fail-closed write guard, a static doctor and a memory panel for an Engramory markdown store. No network, no daemon and no model calls unless you opt in; recall quality is measured by a reproducible benchmark in the repository.

Local memory for DeepSeek Harness: what you write is filed under a topic and linked to related entries, recall ranks results by embedding similarity (word matching when no model is installed), and the data stays in a local SQLite database.