Blog
DeepSeek Harness articles
Tutorials, rc.8 hub, comparisons, and analysis.
DSH 0.2 rumors this week: what installers should pin before Monday
Community posts on X are clustering around a Monday or Tuesday DSH 0.2 window, sometimes tied to a V4.1 Pro-for-DSH rumor. Still a rumor. Here is a pin checklist before you touch production. →
GuideDeepSeek Harness desktop chatter: what changes if you leave the browser
Desktop screenshots are circulating on X: login and key, a local workspace, four modes, team, voice, and a terminal. Until an official client exists, stay on npx @deepseek-ai/dsh web and do not install leaked binaries. →
AnalysisWhy X users keep saying DSH is “token efficient”: cache, Minimal, and V4.1 Flash
The line that DSH is token efficient usually means cache shape, Minimal mode, or a Flash-class model. PTC does not make a run cheap by itself. Set a cost cap before you trust the slogan. →
ReviewSpace Bunny on OpenRouter, tested inside DeepSeek Harness: what BYOM actually means
After Space Bunny climbed public boards, a lab wired OpenRouter into DSH for coding tests. BYOM means you own tool-calling, the exact model id, and the risk of weights you cannot name. →
AnalysisShared task boards in DSH: collaboration taste vs coordination overhead
Shared task boards and member chat give DSH a team feel, and they also create coordination loops. Contrast that with the glass-room Agent Team article, then cap hops, writes, and approvals. →
AnalysisDeepSeek V4.1 Flash enters limited beta: new architecture, native multimodal, and a model id that expires September 10
A community-group beta of an intermediate V4.1 Flash checkpoint: new model structure, native multimodal support, same price as V4 Flash and 20 concurrency — with the id expiring September 10. What is confirmed, what the community measured, and where DSH stands. →
AnalysisDSH's first month: the platform ran ahead of the product
Eleven prereleases in three weeks, SDK and event support landing fast — while install hangs, runaway token costs and missing cost circuit breakers top the community wishlist. →
GuideHow to pick DSH plugins: read the plugin tree first, then choose by task
Every DSH capability is a plugin. Read the six plugin families — model, tool, memory, sandbox, orchestration, UI — then pick by task, not by stars. →
GuideThe five layers under your dsh config: field notes on the patch stack, PTC savings and the skills wart
How dsh layers five config files, why apiKeyEnv beats apiKey, what PTC mode really changes, and the skills-path wart nobody warns you about. →
AnalysisA plugin store that said no to 1,225 entries: reading the DSH-Store admission numbers with you
DSH-Store admits 253 of 1,411 catalog records through a three-layer gate. We check the funnel math, the risk tiers, and where it grades its own homework. →
AnalysisClaude Code, Codex, and DeepSeek Harness: three answers to one question — who holds control
Comparing the three agent runtimes on where engineering complexity is spent: workflow conventions, OS-level sandboxing, or a fully replaceable plugin kernel. →
Reviewdsh-web-ui reviewed: one command turns the bare DSH console into a full workbench
90+ contributors, 19 skins, task board, Git graph, SSH panel. What the dsh-web-ui bundle installs, what it costs, and what to audit first. →
TutorialWrite Your First DeepSeek Harness Plugin: Tool, Bundle, Publish
A rewritten step-by-step guide to DSH plugin development: defineTool, cordis.yml overlays, bundling, and the dsh-tools version trap. →
AnalysisDeepSeek Harness Hit 100K Stars in 45 Hours: What the Plugin Boom Means for You
Harness v0.1 went MIT-open and crossed 100k stars in 45 hours. A rewritten analysis of the plugin ecosystem, PTC mode, and security trade-offs. →
PulseDeepSeek Harness Hit 100K GitHub Stars in 45 Hours — What the Plugin Rush Means for You
DeepSeek Harness crossed 100k stars in 45 hours. We break down what the plugin ecosystem boom changes for installers, and where the hype oversells. →
ReviewTen DeepSeek Harness plugins actually worth installing
A field-tested shortlist of ten DSH plugins — desktop shell, TUI, market, sidebar, @-file, context meter, vision, browser, teams, skins — with the caveats for each. →
AnalysisDeepSeek Harness is a workbench, not a model
What DSH actually is: a pluggable agent runtime where the scarce feature is a replayable trajectory — who should install it now and who should wait. →
AnalysisDSH vs LangChain vs AutoGPT: runtime, not framework
Where the agent loop, memory, tools and sandbox live in each system, and why plugin composition changes the extension story. →
AnalysisThe Agent Team tab is a glass room, not a black box
DSH ships a Team view that looks like Tencent Marvis at a glance but inverts the philosophy: transparent and auditable multi-agent collaboration where every member, tool and message is visible. →
Review66 minutes and a few yuan: a real RAID reassembly
Install, the four modes, then a real forensic RAID-reassembly workload: 66 minutes, 2 to 4 million tokens, about 3 to 6 yuan, plus the trajectory-to-skill loop. →
AnalysisWhy DeepSeek Harness gets such polarized reviews
The split is not China versus abroad: ordinary users want a finished product, developers want to own the harness. Same engine, different cars. →
GuideBuild a DeepSeek Harness tool plugin from scratch
A practical walkthrough of packaging a complete dsh plugin: the package skeleton, a defineTool, the capability seam for calling the LLM, registration, and a behavioural test. →
Analysisdsh2shell: the unauthenticated RCE in web-exposed DeepSeek Harness
A spoofed Host header unlocks DSH privileged RPC from the internet: register a fake provider, switch the session, drive the bash tool. CVSS 9.8, and how to fix it. →
AnalysisDeepSeek Harness benchmarked: 60 real tasks, three axes
Shanghai AI Lab scores V4 Pro + Harness on 60 real agent tasks: mid-tier ability, half the cost of OpenClaw, slower execution — and which models the harness actually amplifies. →
ReviewThree releases in three days — and the vision model flubs a 4-column Excel sum
rc.8 → rc.1 → rc.2 in 72 hours: first-party vision arrives, and a real test where the new vision model read an Excel screenshot and got every total wrong. →
Review121,767 stars, six axes: an honest health check
36 official packages, reversible plugin side effects, 7,000-event session logs, plugin security with no guardrails — overall 6.5/10 on a real-machine review. →
GuideInstalling DeepSeek Harness on Windows, and reproducing the V4 Pro benchmark setup
Node requirements, the npm install that actually works, first workspace, the four presets — then a Minimal-mode control experiment that verifies (or debunks) the V4 Pro benchmark talk. →
ReviewGLM-5.3 on DeepSeek Harness vs GPT-5.6-Sol: same PRD, two agents
Datawhale runs a production image-matting web app through both models in DSH: core features tied, engineering details diverged — rate limiting and security headers by default. →
AnalysisDeepSeek Harness from the inside: plugins, Cordis and a three-model fix test
A deep-dive into everything-is-a-plugin (Cordis, Capability Seam, five event modes, the session log), then one real bug fixed three ways: Codex 9 min, V4 Flash, V4 Pro stalled. →
ReviewWeek one with DeepSeek Harness: patching MCP without a restart
A week-one field note turned into installer tips: one-command startup, skills hot-reload, the no-restart MCP patch, and a jadx reverse-engineering ask. →
Guidedsh Command Map: A Practical Installer Cheat Sheet
A practical map of dsh CLI commands: web UI, profiles, modes, plugins, permissions, and subagents. Verify with current docs as rc.x may change. →
AnalysisInstaller Field Notes: An Orange-Book Style Teardown of DeepSeek Harness
Field notes for installers: dump-config, append-only logs, Create-mode risk, PTC cost myths, and why verified installs beat star counts. →
GuideDeepSeek Harness rc.8: Multimodal Inputs & Subagent Bundles
What rc.8 changes for installers: image-capable adapters, Claude Code/Codex bundles, Windows PTY, and why verified packs still matter. →
GuideDeepSeek Harness 0.1.0-rc.8: vision, subagents & verified packs
Hub: multimodal narrative vs vision API not ready vs packs; Claude Code/Codex subagents. →
ConceptWhat “everything is a plugin” actually means
Sessions, sandboxes, and filesystems as swappable modules — and why verified installs matter. →
AnalysisFrom 30.2% to 96.2%: harness beats the scorecard
Same weights, different runtime. Why V4 Pro pricing noise is not the whole story. →
GuideA realistic first day with DeepSeek Harness
dsh web, permissions, modes, cache — and the zombie Deep-diving trap. →
ComparisonDeepSeek Harness vs Cursor (2026)
Updated: both have agents now. Harness + CI vs AI IDE — when to use both. →
ComparisonDeepSeek Harness vs Claude Code
Plugins vs MCP, open source vs closed, pricing, and which to choose. →
ConceptWhy it's called a Harness
Agent = Model + Harness, the Cordis kernel, and time/spatial composability. →
GuideThe four modes
Standard, PTC, Minimal, and Create — a complete guide. →
AnalysisAppend-only session logs
Observable, auditable, reproducible agent debugging. →
AnalysisCost & caching
How harness choice moves the token bill by up to 7x. →
ReviewFirst impressions & V4 Pro pricing
37k stars, cache pricing up 12x, the four modes, and the best community plugins. →
ArchitectureThe self-modifying agent
Everything-is-a-plugin, Cordis, and the "make your own wrench" Create mode. →
Comparisonvs Codex
Agent presets vs scenario modes, rebuildable sessions, and the plugin model. →
AnalysisPi & the 99.93% cache hit rate
Minimal tools + append-only sessions = cheap DeepSeek. →
AnalysisDeepSeek Harness vs America’s agent tools
158k stars in 5 days — what it does to Claude Code, Codex, Cursor and Copilot. →
Comparisonvs 12 coding agents
Claude Code, Codex, Cline, Pi, Reasonix, CodeWhale and more. →
ConceptPlugin philosophy: everything vs fatigue
How DSH differs from VS Code and Claude Code. →
ConceptSelf-evolving software: miracle or chaos
The Create mode, Cordis, and OpenClaw's warning. →
AnalysisWhy TypeScript / Node.js
npm is the plugin distribution layer. →
TutorialTrajectory tracing: the killer feature
Debug agents and cut token waste from the trace. →
AnalysisPlugin ecosystem: 209 plugins
Ten categories of community plugins from the Show and tell firehose. →
AnalysisThe 10 most-requested features
297 Ideas ranked — desktop, CLI/TUI, memory, Docker, and more. →
AnalysisThe dsh-plugin topic is polluted
How we vet real DSH plugins from tag-squatters. →
FAQFAQ: the community's most-asked questions
Permissions, image input, token cost, install crashes, Docker, memory. →
AnalysisCommunity bug report: 321 discussions
Eight categories of real-world issues. →
GuideMCP: connecting servers to DSH
dsh-workspace-mcp and friends. →
GuideMemory plugins: dsh-memory, dsh-mnemon
A complete guide to DSH memory plugins. →
GuideModel setup tutorial: official API, OpenCode Go
Three setups + rc.8 vision API status (use packs meanwhile). →
GuideSandbox: permissions & security
Permissions, security & tool approval. →
GuideSubagent tutorial: Claude Code, Codex & forks
rc.8 subagents: Claude Code/Codex bridges + verified plugins. →
GuideVision tutorial: rc.8 narrative vs verified packs
Catalog gap vs modlens / vision toolkit packs. →
Comparisonvs Claude Cowork
Open-source harness vs closed agent. →
GuideIs DeepSeek Harness a desktop app? CLI, VSCode, TUI & Electron shells
DSH has no official desktop client — here are the real alternatives. →
GuideIM integrations: WeChat, QQ, Feishu, DingTalk & Telegram
Put DeepSeek Harness inside your chat apps — the full plugin lineup. →
PulseCommunity Pulse #1 — what users want right now
2,638 discussions distilled: the loudest asks, this month's session-corruption bugs, and the crowded plugin-discovery race. →
PulseCommunity Pulse #2 — install hangs, subagent defaults & the 7,000-plugin count war
4,404 discussions distilled: the npm-install hang cluster, subagent model-inheritance bugs, and the accelerating plugin-count race. →
PulseCommunity Pulse #3 — vault fever, the model-list snapshot & a repo at 207k stars
207k stars, zero open issues: the vault thread keeps climbing, the install-hang cluster moves, and the CLI/memory demand clusters sharpen. →
PulseCommunity Pulse #4 — the rc.1 upgrade that lost sessions, a 5-release week & the count war reversed
5,757 discussions distilled: the upgrade-is-the-new-install regression cluster, zero-arg tool-call binding failures, the temporary-chat ask, and the 0.0.0.0 lockdown debate. →
GuideWhat makes a good DeepSeek Harness plugin
The author checklist: bundle manifest, ESM, version pins, inject — and how to earn Verified. →