Plugins

Multi-threaded downloader plugin for DeepSeek Harness with bundled aria2: probes the target for Range support, then picks aria2 (4 or 8 connections by file size) or falls back to curl.

Skill that types the request before it searches: it splits a question into premises, definitions and analysis units, decides which report skeleton fits (verdict, explainer, causal mechanism, comparison, survey map, metric card), confirms the direction and runs a feasibility check, then emits that skeleton instead of forcing every question into a true-false verdict — in Chinese or English. Evidence gathering never trusts model memory: it searches domestic and international ecosystems in parallel, routes each point to the primary-source ladder for its domain (policy, academic, medical, financial, disaster, food safety, science, education, sports, entertainment, judicial, and more), fetches primary pages, grades sources by type, re-searches with debunking keywords, and keeps every statement bound to a full URL and citation position. Three factual judgements can be routed to a TypeSafe Jev (System One) decision model when a key is configured, falling back to the skill's deterministic rules when it is not; a model verdict can only lower a confidence tier.

Cristiano Ronaldo desktop pet for DeepSeek Harness: a chibi Portugal number 7 mascot in a bottom-right overlay whose animation follows the agent state — dribbling while tools run, thinking during a turn, waiting at approvals, diving on errors — and jumping in a SIU celebration with a host-process sound when a turn completes; it can be dragged, looks at the cursor on hover, dives on a triple click, and a settings panel manages several pets and imports custom spritesheets. It also runs as a borderless, always-on-top native Windows window that keeps living after the browser is closed: hovering shows the current workspace and conversation title, double-clicking opens the Harness page in Edge, and dragging, mouse-wheel resizing and a right-click menu work there too, sharing position, size and behaviour with the web panel. Pet sheets may use any cell resolution and scale preserving aspect ratio (atlas.scaling: smooth | pixelated). v2.2 adds a video route — feed it a clip (green screen or ordinary footage) and its real motion becomes the pet animations, via automatic backdrop sampling, chroma keying with spill suppression, motion-based segment splitting and union-bbox alignment; already-extracted PNG frames work too, with no video decoder installed, and the clip audio can be cut into per-action sounds. Pets can also be published to the community gallery under the DPSL-1.0 sharing declaration, and installed from it in one click. v2.3 adapts it to DSH 0.2 desktop clients — the celebration sound now falls back to spawning PowerShell when the host no longer provides the shell service — and makes a single click play a short, configurable taunt animation (`interactions.clickAnim`).

Local OCR fallback for text-only routes: when the session model declares it cannot accept images, the attached image is cached locally and its path injected so the model can call ocr_image — PP-OCRv5 + ONNX Runtime on CPU, no API key, images never leave the machine. Silent when the model can see images.

Evidence-driven goal verification plugin for DeepSeek Harness. Locks goals via /ultragoal and verifies completion against command exit codes and generated artifacts, preventing the model from falsely reporting task completion.

Selectable SearXNG, Brave, Tavily, Gemini Search Grounding, or Wikipedia backend for the native web_search tool, with Web settings and no implicit fallback.

Speaks agent replies in the DeepSeek Harness web UI through a provider fallback chain (OpenAI, ElevenLabs, Google, Azure, Groq, Deepgram, OpenRouter, Edge, Piper, eSpeak), so a failing or rate-limited provider falls through to the next instead of going silent.

Self-evolving pitfall immunity for coding agents (Claude Code, Codex, Cursor, ZCode, DSH): capture failed-then-fixed tool episodes as SQLite+FTS5 antibodies and inject them before the agent repeats the mistake. Zero runtime dependencies.

One-line sidebar badge for OpenCode Go quota (5-hour, weekly and monthly windows), switching to the DeepSeek official balance when that model source is selected. Read-only, no telemetry. The OpenCode Go quota needs OPENCODE_GO_API_KEY configured in DSH, or an OpenCode Console login (its local credential file is read as a fallback), and the DeepSeek balance needs DEEPSEEK_API_KEY configured in DSH.

Game asset production pipeline: generate images, video and audio, turn a video into a coherent sprite animation, derive recoloured variants, adapt assets to an engine schema, install them into the project and verify the code actually references them. Requires ffmpeg for the video-to-sprite path, and an image/video API key (Agnes AI / OpenAI / fal) for generation.

Model failover and turn keep-alive for DeepSeek Harness: a request that fails with a model error before any visible output falls through the models of the configured provider groups (then the next provider) until one streams; transient failures retry with backoff under a liveness watchdog, a dead turn self-resumes with templated 「继续」 wake-ups, and an optional full-auto mode auto-answers permission, question and plan approvals with every decision appended to an AUTO-MODE.md audit log.

Reads and writes a self-hosted Bitwarden/Vaultwarden vault from any session: eleven bitwarden_* tools, real-time WebSocket sync with a polling fallback, a settings panel with readonly, ask and auto write modes, and optional local masking — with masking on, passwords surface as stable placeholders unless revealed one field at a time, and on-disk session logs are swept for bitwarden secrets. Masking and log sweeping are separate switches.

Plan-usage pet for the DSH web GUI: a floating skeletal-animated character reports quota in a speech bubble, auto-matching the configured provider (OpenCode Go, Command Code, DeepSeek) to show a plan percentage or an account balance, and falling back to local token stats when the provider API is unavailable; draggable, wheel-scalable (90-260px), synthesized click sounds, loopback-only routes.

Large-image recognition for vision-exp models: lossless 800×800 tile recognition (smart/pipeline/full), local OCR with preprocessing & handwriting routing, optional multi-vendor GPU (DirectML/CUDA/OpenVINO) with auto CPU fallback.

Free web search for DSH: anonymous Parallel default plus Exa fallback, a Settings toggle, and an MCP server compatible with Claude Code and Codex.

Version management for dsh: inspect the installed version against every npm channel or version line, install or downgrade in one click, stream the install log live (rollback-snapshot copy progress, npm extraction progress, elapsed time), roll back instantly from local snapshots, and automate the cycle with a policy engine (silent auto-update, execution windows, daily checks). Never stops npm mid-install — a wedged run is only stopped after an hour, and a half-committed global tree is detected and repaired from the pre-install snapshot automatically. Confirms before installing a found update, auto-restarts the host after install, and falls back to a mirror when the primary registry is unreachable.

Four GitHub workflow skills for PR triage, review feedback, GitHub Actions diagnosis, and safe draft-PR publishing, reusing existing DSH capabilities with gh/git fallbacks.

Renders pasted/uploaded images inline in the DSH Web chat and gives text-only models vision: the model-invokable visual_review tool calls any OpenAI-compatible multimodal API first, falling back to a local Qwen3-VL worker.

Two-level model circuit breaker with failover: trip a model or a whole provider after repeated request failures and route the next request to a configured fallback.

Search, list and read past and current sessions through the harness's own `sessionQuery` service: uses the SQLite FTS5 index where a deployment enables it, and falls back to a bounded newest-first scan where it does not.

Self-evolving pitfall immunity for DeepSeek Harness: capture every failed-then-fixed tool episode, distil it into a reusable antibody stored in SQLite+FTS5, and inject the relevant ones before the agent repeats the mistake.

A fifth DeepSeek Harness work mode that requires a falsification design (at least two pieces of evidence that could disprove the plan) in every plan and tags each factual claim with its source (\[文件]/\[检索]/\[记忆]) or \[未知]; evidence grading follows the project's AGENTS.md.

Windows shell takeover: replace the default pwsh tool with a configurable multi-client shell tool (git bash, WSL, cmd, pwsh, custom paths), with a deny blacklist and automatic fallback.

Android in the sidebar, driven by text instead of screenshots. 30 tools read the real UI tree, tap by text/id/ref, measure animation motion, and mirror the screen into the sidebar where you circle a spot and send it back with one sentence — the host ships the marked-up frame plus the controls inside the circle to the chat. Token cost is the point: a screen check is a 200-600 token text block, not a 1k+ token screenshot, and a warm check costs 100-200 ms of wall clock. Video is raw H.264 straight into WebCodecs — no transcoding — with p50 2 ms host-write to decoded frame over a 26 s run, one decoder, zero dropped frames; it falls back to MSE on its own and ?avc=0 forces the old path. Sound is one switch, off by default and remembered per serial, so the default path pays nothing; when on, audio rides its own socket (48 kHz stereo PCM, measured 22 s / 1047 packets / 0 loss) and closing the window reclaims it too. Nothing runs behind your back: closing the window stops capture at once, the viewer-less host decoder, recorder and device-side server are gone within 3 s, an idle channel is dropped after 90 s, and a request you closed is never reopened by itself. `adb_proc audit` settles who leaked what in one call. Bundled helpers, no extra services; needs adb and ffmpeg on the host, and the adb_ocr tool plus the OCR fallback for empty UI trees are macOS-only.