Plugins

Long-form novel writing plugin: each book is a git repository of frontmatter files, with 39 DSH domain tools and a web writing UI. Deep import drafts candidates, AI adoption into canonical assets requires DSH approval, and deterministic continuity checks ship with a report-only benchmark. Independent open source (MIT), not an official DeepSeek product; requires Node >= 24.11.0 and currently targets DSH 0.1.2-rc.1.

One MCP tool instead of every schema you own — a lazy capability broker that cuts resident tool context by a measured 94.4%. Search, inspect, enable and call MCP servers and skills on demand.

Compares existing DSH sessions or runs controlled A/B trials between agent presets, showing aligned timelines, paired metric deltas, explicit success checks, and exportable reports.

Piano performance plugin: ask the agent to play a piece and it renders on a Canvas2D grand piano with real Salamander Grand samples, an immersive stage, and an interactive 88-key keyboard.

Experience validation layer for DSH: zero-token collection of tool-failure/retry/search signals, AI refinement into verified skill books (three-layer verification), looked up at task start. Ships 10 trial skills and benchmark data (complex-task success 100% vs 60% bare).

Runs N serial speedtest rounds in one unchanged environment from the plugin detail page, then reports per-round tables with median, min, max and MAD, and keeps each batch as an archive that can be renamed or deleted.

Runs a four-task, 150-point model intelligence benchmark inside DeepSeek Harness (Qin 60, Pompeii 45, Pelican 22.5, Candy 22.5): each selected model × reasoning-effort pair executes every task in its own isolated subagent session, Qin and Pelican are graded by the selected judge model against rubrics, Pompeii and Candy programmatically, and batches keep scores, artifacts and cost. Adds a 0–1000 real-world performance score per model × effort computed from local session logs across seven dimensions (output speed, first-token latency, reasoning speed, tool cadence, tool latency, tool success rate, cache hit), with baseline deltas and controlled speed runs, plus a sidebar page with ranking, performance, history and settings views. Each batch spends real tokens of the candidate and judge models, and candidate subagents get file and bash tools inside an isolated run directory.

DeepSeek Harness plugin: AI research toolkit — model release radar, model benchmark & community report, and AI news digest (4 tools + 3 skills).

A local mirrored A/B benchmark for context compression, not another compressor. It runs one fixed read-only scenario twice with the same model and temperature — compression off, then on — and reports input tokens, turns, tool-call success, fact coverage, answer similarity and a per-line diff, with JSON, a self-contained SVG and raw warnings. 18 node --test tests, zero dependencies. The first measured run: 63,029 → 6,927 locally estimated input tokens (−89.0%), 3 → 3 turns, facts 3/3 → 3/3, answer similarity 1.00 on one sample.

Device inspector and control panel for macOS and Windows that shows hardware, system, software, network, security and processes on one sidebar page and applies 19 whitelisted actions, with a confirmable cleanup scan, a repeatable benchmark and device_info, device_control and device_benchmark tools for the agent.

Blocks the end of a turn while temporary files from that turn are still lying around: it registers a listener on the serial agent/turn-stopping hook and, if anything is left, hands the model the file list plus two ways out — delete them, or justify each one and move it somewhere deliberate. Candidates come from two narrow temp-file sources, plus a backup-retention channel that is on by default and flags excess .bak copies of files touched this turn. The first is the .agent-tmp/ convention directory under the session cwd: whatever is left in it is reported outright, because that directory exists only for throwaway files. The second is the paths this turn actually wrote, captured through the fs/write-intent, fs/edit-intent and fs/delete-intent hooks; those are reported only when they sit under that cwd (compared after realpath normalisation), fall outside .git, node_modules, out, dist, build, target, vendor, .venv and the Python bytecode cache directory, match a temp-file pattern that requires a separator after the prefix so checkout.js and benchmark.ts are not taken for scratch files, and the file is still present — which keeps junk already sitting in the workspace from being resurrected as a complaint about this turn. One gate per turn, subagent sessions skipped, nothing deleted unless the backup mode is explicitly set to prune, and the injected wording follows the host locale instead of assuming a Chinese one.

Channel test page for any configured provider/model: multi-round prompt-cache hit rate, streaming TTFT and tok/s, and a pelican SVG test.

Public/synthetic eval-pack lint and deterministic JSON grader; not a validated benchmark.

Skill bundle for tiered multi-agent science campaigns in math, physics, astronomy, engineering, and biomedical evidence: a shared blackboard, independent verifier and red team, wolf-pack and flanking tactics, token and time budgets, known-answer evals, a four-agent pendulum pilot, and a six-agent fourth-order pilot that caught and corrected a benchmark answer-key error. Does not claim open-problem solutions.

Plays a character entrance card beside the composer when you switch model: identity name, six scored axes (intelligence, coding, speed, stability, cost, context), quota hint and a dialogue line. 23 curated cards across 15 providers; models without a curated card fall back to a data card derived from public benchmarks, or a generic card when no benchmark data exists. A web client plugin with no dependencies and no hooks.

Diagnostic evaluation plugin for DeepSeek Harness: freezes the confirmed Harness configuration, runs a benchmark, keeps the full session trace, and attributes failures to model, Harness, benchmark or environment.

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.

AI research toolkit for DSH: a model release radar, a per-model benchmark and community report from 13 structured leaderboards plus Chinese and English community discussion, and an AI news radar over 46 feeds and 22 official X accounts; registers four tools and three skills and renders every report to Markdown with a generated table of contents.

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.

Cross-session memory for your dsh agent: remembered, recalled, and re-verified instead of assumed.

Retrospective agent-session evaluation — grades the sessions that already happened: deterministic grade cards, regression diffs, /eval-history trends, and /eval --report straight from exported artifacts; no benchmark runs, no LLM judges.

Run the same prompt through every model you have, and see the difference without squinting.

DSH web plugin: a LiveBench tab in the Trajectory view (right of 对话/轨迹). Run LiveBench evaluations against every model configured in the DeepSeek Harness — pick provider/model, category, task, release and question range from dropdowns, watch progress, and

Control panel for local LLM inference servers in the DSH web view: service registry with health checks and stop controls, named parameter profiles for llama.cpp, SGLang and vLLM with pre-save KV/VRAM validation, GPU detection, and tok/s benchmarks.