DeepSeek Harness Tutorials

A plain-language tour of what DSH is, how it differs from the DeepSeek web chat, what it can do for you, and the words plugin / profile / workspace demystified.

From installing Node.js to launching with one command and picking a workspace — zero to a running DSH in ten minutes, plus where data lives and how to remove it all.

What an API key is, where to get one, how to plug it into DSH — plus attaching OpenAI / Anthropic models and a quick error decoder.

Add capabilities to AI like installing apps: read every segment of the install command, install the hugely popular open-design, and dodge the GitHub build-authorization trap.

List installed plugins, upgrade one or all, uninstall cleanly, upgrade DSH itself — and the deep-clean guide to resetting a profile or removing everything.

Go from user to contributor: a plugin is just a module that exports apply(ctx). Write a minimal plugin, register a model-callable tool, and package it as a bundle for npm or GitHub.

Every pitfall beginners hit, in one place: npx stalls, port conflicts, MISSING_CREDENTIAL, plugins that seem dead after install, migrating data — grouped for quick lookup.

Pick the right channel between npm / GitHub ref / tarball, the 6-item bundle self-check, and what to verify before publishing — the complete path from your repo to a user's install command.

What dsh.bundle and ~/.dsh/profiles/*.json each control, how to isolate web/tui/headless workspaces, and where the profile-switching boundaries actually live.

Pick the right sandbox mode (auto / full-auto / never), read an approval card, understand the fail-closed design, and how sandbox sits orthogonal to plugin permission boundaries.

Three risk surfaces (Browser-only / Host-side / Agent-facing), structured tool results, Cordis resource cleanup, and the credential three-no principles — write plugins users can actually trust.

When Python SDK beats TS, package layout, the @tool decorator, Python-type-to-JSON-Schema mapping, asyncio traps, and the pydantic v2-only limitation.

Why the DSH ecosystem needs postmortems, the 5-section structure (summary/timeline/root cause/mitigation/lessons), blameless principle, and how derived tasks land in issue trackers.

Choose plugins without blindly following stars: turn a need into a minimal, reversible install plan using evidence, maintenance signals, and workflow fit.

A systematic path through command, network, build-approval, wrong-profile, and post-install failures — collect evidence first and make the smallest safe repair.

A day-by-day first-week plan for model setup, one useful plugin, workflow collections, backups, and reflection — with a definition of done for each step.

A ten-minute check of source, command, permissions, credentials, and rollback path — distinguish missing evidence from explicit risk before granting a plugin access.