Плагины

Tiered model routing for DeepSeek Harness: think/build model tiers, vision delegation (the vision model only describes images; the working model acts), image generation, and a high-impact guard.

Страж разрешений с автоодобрением: промежуточный уровень между workspace-write и danger-full-access — автоматически разрешает безопасные операции в доверенных каталогах и всегда запрашивает подтверждение человека для деструктивных, с 11 переключателями по категориям и журналом аудита.

Commitment write-gate: operator-authored rules enforced before a tool call runs — a deterministic guard tier plus an LLM-judge tier, fail-closed by default, every block written to a contradictions log.

Mention a server with @ in a DeepSeek Harness session and run commands on it over SSH, behind a two-tier approval gate that fails closed.

An elevated Plugin list for DeepSeek Harness Web Settings — trust tiers, collapsible groups, and complete per-plugin detail

Claude Code-style permission rules engine: hard/deny/ask/allow tiers with a hard tier above full access, workspace-scoped rules, wildcard path protection, and a visual staged editor; rules persist in settings.yaml.

A project-agnostic multi-agent collaboration protocol for the DeepSeek Harness: loop guards that stop runaway agent loops, a six-field handoff format, risk-tiered review routing, and a fixed verify→commit→push→review→close sequence. One skill instead of the same rules copy-pasted into every agent's instructions.

Clusters arXiv, first-party lab, official artifact, hardware-vendor, and opt-in X signals into versioned release evidence bundles, then gates claim-level reproduction on commands, artifacts, metrics, and verifiers.

Двухуровневый маршрутизатор моделей для DeepSeek Harness — маршрутизация через LLM-судью, цепочки резервных моделей, отказоустойчивость с экспоненциальной задержкой и оркестрация на уровне задач (адаптация pi-shift-router для DSH)

Промежуточный уровень разрешений для DeepSeek Harness между workspace-write и danger-full-access: автоматически подтверждает безобидные команды и операции, нацеленные на настроенные доверенные области за пределами текущего рабочего пространства.

Windows system-drive (C:) cleanup for DSH: rule-based scan, five-tier safety grading, a staging area with a rollback ledger, UAC elevation, and migration to another drive — a model tool plus a Web GUI panel.

Automatic tier-based model routing for DeepSeek Harness: a virtual `smart` model classifies each request by difficulty (hard / normal / easy) and by vision need, then delegates it to the models you already configured. 三级难度 + 视觉自动路由，虚拟 smart 模型零配置接入。

Image and video production toolkit for DeepSeek Harness - background removal with quality tiers, local inpainting retouch, headline and logo compositing, batch platform-size exports and image QA reports; for video, AI subject matting with background replacement (RVM ONNX, plus a chroma-key path), animated titles with fade/slide/typewriter/karaoke presets, 53 transitions with stylized variants, and light-effect and grading presets with LUT support, alongside footage inventory, accurate cutting, subtitle burn-in, music mixing and platform-spec exports; it provisions its own local Python runtime through a bundled bootstrap script.

Declares per-model reasoning-effort ladders for third-party providers in the llm-pi-ai settings section, so DSH's stock thinking-intensity selector works on models the pi-ai catalog does not describe.

Turns DSH permissions into 9 selectable presets (3 sandbox modes x 4 approval tiers), with rule plus LLM risk classification across four risk levels, password-gated approval for high-risk operations, and a JSONL audit log.

Tiered persistent memory (user/env/project) with per-tier char budgets and semantic query routing; full layers refuse writes instead of silently evicting.

Polaris Memory: local-first project memory plugin for DeepSeek Harness. Uses per-project .dsh-memory.json; provides memory_read/recall/write/update/delete tools, /memory command and settings UI, with tiered injection (resident user profile + scene-triggered memories with P1/P2/P3 importance), semantic recall (IDF + synonym layer), and a hard-guard mechanism for bottom-line memories (detect deny-semantics -> authorize -> block AI deletion; user can confirm-override).

Three-party collaboration protocol for dsh: Step List state machine, five-tier review chain (Gemini/web/Claude Code/dialog/manual), alternatives adjudication, cross-restart resume with heartbeat, and capability persistence.

Automatic model-tier routing for DeepSeek Harness: one user instruction enters, one tier decision comes out — complex intent is planned on the strong tier and implemented on the cheap tier, simple intent stays cheap throughout, with a deterministic guard and a TTL escalation fallback.

Dual-endpoint Zhipu BigModel GLM provider catalog (Coding Plan and ordinary API) with live model discovery, live-tested thinking-tier mapping, and a settings card for keys, endpoints and the default reasoning tier.

A webfetch tool for full-text web/Toutiao articles: 3-tier strategy (direct fetch → resident Edge CDP headless browser → dump-dom fallback), with automatic Toutiao anti-bot challenge handling plus a bundled experience skill.

DeepSeek Harness (dsh) plugin that registers Lynkr — a self-hosted, multi-provider LLM tier-routing gateway — as a custom OpenAI-compatible provider in settings.yaml.

Tiered task queue as the single source of truth for DSH↔Hermes work (reset > approve > review): enqueue tools for all submission paths, lease/claim model with concurrency 1, priority aging against starvation, atomic busy mutex, and a Hermes-side consumer (cron */1 + monitor gating) that drains tasks, writes review-handoff requests with single-slot checks, and invokes the reset agent.

Mandatory cross-review for design proposals and lessons-learned: auto-detects design doc writes and enqueues them (tier=review) into the task queue, delivers the external reviewer verdict back to the initiating session, and appends approved preventive measures to OPS-GUARDRAILS.md (append-only, conflict-checked).