Plugins

3D Gaussian Splatting / NeRF / computer-graphics research skills — 18 AI skills (paper reader, method compare, experiment planner, code reviewer, training debugger, visualizer, MCP renderer…) over a curated 1011-method catalog, plus a stdio MCP server for agent-driven 3DGS rendering.

Anatomy panel for the Agent-Body organs: serves a live vitals dashboard at /anatomy inside the harness itself - organs, heartbeat, healing ledger, self-training counters and the real pulse stream - read straight from the runtime files rather than retyped. Host-only, no client bundle and no build step.

Heads-up Texas Hold'em training with an independent AI opponent, persistent public-action memory, and decision reviews.

Nightly self-improvement for your DeepSeek Harness agent: harvest past sessions offline, replay recurring tasks on your own API budget, and consolidate validated skills behind a held-out gate. Brings microsoft/SkillOpt's SkillOpt-Sleep engine to dsh as native tools plus a bundled skill - no weight training. Requires Python 3.10+ and the SkillOpt-Sleep engine (pip install skillopt).

English edition of Weiwen's Law (KISS's Law): the first and only domain-agnostic white-box causal adjudication middleware for AI agents. Unlike domain-specific causal-inference tooling (statistics/economics/ML) or evidence-gated agent firewalls, it adjudicates *every* tool call along a deterministic structural causal chain (R->S->D->H->M) before execution - across any domain, with no training or per-domain tuning. It blocks destructive and credential-file operations, escalates repeated boundary violations, cuts faulting links until verified-fixed, and gives risk verdicts on un-auditable execution; runs fully local with zero API cost.

Speaks your agent's reports in a voice you choose, compressing long reports before speaking. Character voices need no training - a 3-10 second reference clip clones one, and community-trained models work too - and no 6.4 GB GPT-SoVITS install: the plugin installs its own runtime and voice. An existing GPT-SoVITS can be used instead, and it is the same voice model.

Endurance-coaching skill bundle that defaults to Intervals.icu for training data (activities, intervals, streams, training load, HRV, resting heart rate, sleep in-bed hours, eFTP) and uses COROS only for training-plan read/write and for fields Intervals.icu lacks (sleep score, recovery score), with iGPSPORT as a fallback for ride details. It turns that data into readiness gating, VDOT pace and heart-rate zones, periodized weekly plans, session reviews, race-day pacing and fueling, and weather-adjusted workouts. The bundle registers the endurance-coach skill in the profile, ships deterministic helper scripts, reads an athlete profile created by its init script, and requires the agent to state when data is unavailable instead of inventing it. Data comes through separately installed MCP servers — the Intervals.icu MCP server for training data, the dsh-coros plugin for training-plan read/write and sleep/recovery scores, an iGPSPORT MCP server only as a ride-detail fallback — and its bundled scripts need Python 3.10+; with none of them it works only from numbers you type in.

Bridges the official COROS MCP server into DeepSeek Harness with a self-contained local stdio-to-HTTP bridge and on-device OAuth login, exposing 34 COROS tools on the current endpoint for sleep, HRV, resting heart rate, recovery, training load, activities and FIT files, plus workout and training-plan CRUD and scheduling.

Local LoRA and QLoRA fine-tuning in a panel: import JSONL, see a VRAM estimate before training starts, evaluate, and export an adapter. One-click Python venv setup, but you must supply your own Python 3.10+ and an NVIDIA GPU (6 GB or more). Verified end to end on Windows 11 with an RTX 3060 6 GB; the installer refuses non-Windows hosts, so Linux, macOS and other GPUs are unsupported for now.

COROS Sport panel for DeepSeek Harness: sign in to COROS inside the panel, read the account, bound devices, recovery, fitness, recent workouts, daily health and the training schedule in one dashboard, and get a plan from the model with /coros-plan. The plan page writes a fixed four-week, 16-session plan to the account only after a two-step confirmation.

Checks a site safety-qualification ledger for training hours, the four required record elements, and special-operation certificate validity against the dates the person worked.

Multi-agent Chinese debate simulator with eight debaters, live two-column discussion and draft notes, judges and scoreboard, single-side clocks, per-round ZIP archives, a training side with a configurable end condition (win-rate threshold or fixed rounds), and a Word prep package.

Training Lens opens a local TensorBoard scalar dashboard with selectable metrics, explicit improvement direction, optional EMA smoothing and Chinese/English UI; requires an existing Python environment with TensorBoard.

Zagent — unified three-role self-improvement loop with three operators. Data-RSI (daily cross-sectional cleaning: winsorize → cross-sectional standardization → median fill, plus experience extraction), Harness-RSI (evolves system prompt / memory / skills), and Model-RSI (evolves alpha training recipes via framework-native walk-forward with 2-day gap and multi-seed). Offline one-command run, GPU auto-detect. Ships one skill.

Legal training workbench: three-pane academic UI that reads the "Minglü" (lawagent preset) learning profile (case dossiers, flashcards, wrong-answer book, notes) and sends instructions into a Minglü session with one click.

AlphaRSI — three-role self-improvement loop that evolves the training recipe of alpha sequence models. Operator runs walk-forward, Evaluator grades with OOS RankIC, Refiner proposes hyperparameters via TPE Bayesian search; only strictly-better recipes are kept. Offline one-command run, data bundle shipped with sha1 reconciliation, GPU auto-detect. Ships one skill.

Training attestation — reproducible, attributable, auditable off-site standardized training for asset pricing. Hard-gates four failure modes (half-refresh, hash mismatch, unknown-read-as-pass, master pollution). Ships one skill.

Model training stage — end-to-end analysis, model training, and estimation for asset pricing and machine learning research. Ships one skill.

Algorithm learning toolkit syncing Codeforces and AtCoder submissions, profiling weaknesses, generating AI training plans, and tracking spaced-repetition reviews across a 114-lesson algorithm curriculum.