Archify
tt-a1i
Agent skill that compiles a typed JSON description of a system into a single self-contained HTML diagram, validated before delivery.
L1
66.3k JavaScript
agent
Skills & Agents DSH plugins
Skill packs, subagents, orchestration and autonomous loops.
This directory lists 16 DeepSeek Harness plugins for Skills & Agents. Their verification levels run L1 2, L2 6, L3 8, L4 0, L5 0: no listing here has been installed or run in a sandbox, so none claims L4 or L5. 0 have passed an automated security scan, and the other 16 are marked unverified rather than safe. Declared compatibility targets are unstated 8, ^0.1.1-rc.2 || ^0.1.2-rc.1 1, >=0.1.0-rc.6 <0.2.0 1, >=0.1.2-rc.1 <0.2.0 || >=0.1.5-alpha.1 <0.2.0 1, >=0.1.5-rc.1 <0.1.6-0 1, 0.1.2-rc.1 || 0.1.0-rc.7 1, 0.1.2-rc.1 || 0.1.5-alpha.1 1, 0.1.5-rc.1 || 0.1.2-rc.1 1, 0.1.5-rc.2 1. Together these 16 listings carry 89,869 GitHub stars, led by Archify at 66,266. Licences are mainly MIT 10, Apache-2.0 5, NOASSERTION 1, and the code is written mostly in JavaScript, Python, TypeScript. Every entry records its source repository, licence, declared harness target and a copy-ready install command; the most recently checked is Aegis (2026-09-18). A verification level measures how far the install path was exercised — it is not a safety guarantee, so read the source before you install.
Agent plugins extend the loop itself: skill directories the model can load, subagents it can delegate to, evaluation gates, self-improvement cycles, and the orchestration that keeps long tasks from drifting.
These are the highest leverage and highest variance plugins in the ecosystem. Look for explicit budgets, human approval points and evidence the agent produced before it claims success. Prefer plugins that hand you a reviewable artifact over ones that only promise better output.
How to choose in this category
Start from the failure you are trying to prevent. Unreviewed autonomous loops fail by shipping confident nonsense. Multi-agent teams fail by coordinating on the wrong decomposition. Skill packs fail by loading instructions that contradict your project rules. Match the plugin to that failure mode, not to the marketing adjective “agentic.”
Check three controls on every candidate: a budget (tokens, steps, wall clock) that actually stops the run; an approval or evidence gate before “done”; and a durable record of what the child agents did. Orchestration UIs are optional; the stop condition is not.
What is mature versus thin here
Mature listings show staged evaluation, interview-before-execute flows, and named roles with dependency-aware tasks. Thin listings are prompt packs with a slash command and no stop condition. Stars do not separate the two — read whether the plugin can fail loudly.
Background and team plugins are useful when work is genuinely parallel. For a one-file edit they are ceremony. If a plugin only wraps “spawn a subagent,” expect to supply the discipline yourself.
Common mistakes
Installing three orchestration layers that each want to own the task graph. Treating a skill as a sandbox boundary when it is only instructions. Assuming a high star count means the loop has a budget. Skipping the permission list because the demo looked friendly.
Install order and boundaries
Orchestration belongs last and only with a budget and an approval gate. If you still lack context visibility, buy measurement before multi-agent teams. A plugin that merely wraps spawn with no stop condition leaves the discipline to you.
16 plugins
Editor's pick
DSH Routing Suite
yjh051108
Three components, three install paths: a restart-free runtime plugin injector, routing agent presets, and a graded task protocol.
L3
7.2k JavaScript
Ouroboros
Q00
A spec-first agent OS for AI coding — Socratic interview, immutable Seed, staged evaluation, MCP server.
L2
6k Python
OpenBiliClaw
whiteguo233
Local-first content discovery agent that learns your interests, then hunts the open web and social platforms for you.
L1
3.3k Python
Agent Teams
NanmiCoder
An AgentTeams orchestration plugin that runs several agents as a coordinated team inside DSH.
L3
1.7k JavaScript
treg
superdesigndev
A credential-injecting proxy and shared tool registry: one token reaches 3,000+ catalogued endpoints, and keys stay on the server.
L2
1.6k Python
Aegis
GanyuanRan
A method pack that makes agents plan against your project baseline, prove completion with fresh evidence, and leave simple tasks on the fast path.
L2
1.2k Python
OpenWrite
LiPu-jpg
Novel-writing suite for dsh — a dedicated creation agent preset, 90 novel_* tools, three native workbenches and a 47-node review DAG.
L3
743 TypeScript
MisakaNet
Ikalus1988
A git-backed library of evidence-rated failure-recovery lessons that AI agents search before re-debugging a known error.
L2
493 Python
DSH Taskboard
shengsheng90
A local SQLite project taskboard for DSH with a native Web UI and eight agent tools, where only a human can accept work as done.
L2
325 TypeScript
Memory Evolve
csyangwen
Hermes-style five-track long-term memory with in-turn self-review, human-gated skill evolution, four-track todos and optional session orchestration — pure plugin, zero runtime deps.
L2
317 JavaScript
dsh-evolve-modes
GraySilver
Four recomposable controls in the composer — working state, reasoning strategy, quality gate, and a self-evolution loop that proposes rules but never applies one without your review.
L3
199 TypeScript
Auto Review
PerryLink
Puts a second model on the approval chain — a read-only reviewer subagent returns a structured allow/deny verdict, and a reviewer failure denies instead of allowing.
L3
181 TypeScript
DSH Agent Team GUI
toolclub
Persistent multi-model workflow teams in Settings → Teams — per-member model/tools, bounded DAG with retries, Run Center and provider Token insights.
L3
176 TypeScript
Deepread
xiehuan123
Evidence-first deep reading that pairs claims with cited sources, labels each claim's status, and builds knowledge maps across articles, books and PDFs.
L3
53 TypeScript
Background Agents
PerryLink
Durable continuable background agents started from a session, watched in an optional Web UI sidebar, messaged by tool and interrupted on request.
L3
15 TypeScript