跪拜 Guibai
← All articles
Backend

10 Plugins That Turn DeepSeek Harness from Bare Shell into a Full Workbench

By 苏三说技术 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Harness's plugin architecture means the tool you install is not the tool you're stuck with. A developer can start with a minimal chat interface and incrementally bolt on an IDE-grade workstation, terminal TUI, desktop app, or multi-agent orchestrator without leaving the ecosystem.

Summary

DeepSeek Harness defaults to a plain chat interface with no file tree, terminal, or task management. Its architecture treats every capability as a swappable plugin, and the community has responded with over 1,100 curated extensions. The most impactful ones fall into three layers: core UI overhauls like dsh-web-ui and dsh-better-sidebar that add task boards, Git panels, and Codex-style sidebars; interaction-mode plugins that bring Harness into the terminal (dsh-TUI) or package it as a desktop app requiring no Node.js; and capability extensions such as ModLens for image understanding, dsh-agent-teams for multi-agent collaboration, and dsh-context-doctor for auditing token consumption.

Installation is a single `dsh plugin --profile web add` command, though the `--patch` flag is required at startup for many plugins to work. The ecosystem is still in developer preview, with rapid version churn and occasional access requirements like locked version numbers or Git permissions. A recommended starting point is dsh-web-ui plus dsh-at-file for file referencing, then layering on interaction and capability plugins as needed.

Takeaways
Harness ships as a bare chat box; all UI panels, tools, and agent capabilities are plugins.
The ecosystem has grown to 1,117 curated plugins and 1,521 monitored repositories, totaling over 301,000 stars.
dsh-web-ui adds a task board with five-column Kanban, Git graph, token stats, and theme support in one install.
dsh-better-sidebar provides a Codex-style sidebar with file editing, terminal, Git, and sub-agent tabs.
dsh-TUI moves the entire Harness interaction into a full-screen terminal interface.
deepseek-harness-desktop packages Harness as an Electron app requiring no Node.js or command-line usage.
dsh-at-file enables @-mention file referencing in the chat input, matching Codex's workflow.
dsh-agent-teams drives multi-agent collaboration from a single natural-language command, with direct agent-to-agent messaging.
dsh-plan-execute routes planning to a reasoning model and execution to a cheaper model to cut token costs.
dsh-context-doctor audits the token cost of system prompts, skill directories, and tool schemas with pruning recommendations.
ModLens gives the text-only DeepSeek model the ability to process images by converting them to structured text evidence.
dsh-reverse-skill bundles 85 SKILL.md files for reverse engineering, penetration testing, and security research.
Startup requires the --patch flag, and some plugins need locked version numbers or Git access permissions.
Conclusions

The 'everything is a plugin' architecture means Harness's default austerity is a feature, not a bug: no capability is baked in, so no capability is forced on users who don't need it.

Separating planning and execution across different models (reasoning vs. economical) is a practical cost-optimization pattern that more agent frameworks should adopt as a first-class concept.

The community's speed in filling gaps the official team left open — desktop app, TUI, file referencing — suggests plugin ecosystems can outpace first-party development when the extension surface is genuinely open.

Context auditing as a plugin reveals how opaque token consumption is in agent workflows; most developers have no visibility into what their prompts actually cost per round.

Concepts & terms
Harness plugin architecture
DeepSeek Harness treats every capability — models, tools, skills, sessions, sandboxes, storage, loops, scheduling, UI — as a swappable plugin installed via `dsh plugin --profile web add` from a GitHub repository.
dsh.bundle
A declaration mechanism in Harness plugins that automatically enables the plugin once its code is pulled from GitHub, requiring only a service restart to take effect.
Multi-agent team (AgentTeams)
A plugin pattern where a captain agent spawns sub-agents, decomposes tasks with declared dependencies, and agents communicate directly via mailbox messages without a central relay.
Dual-model routing (plan-execute)
A cost-optimization strategy where a powerful reasoning model handles task planning and decomposition, while a cheaper model executes the resulting subtasks.
Context auditing
The process of quantifying the token cost of accumulated context — system prompts, skill directories, tool schemas — that a model carries with every request, often invisibly to the user.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗