DeepSeek Harness plugin glossary
The vocabulary used across this directory and in the plugin ecosystem. Read this if a listing mentions a seam, a profile or a target and you want the precise meaning rather than a guess. Each entry also says what to check on a listing page — definition plus the review habit, not a bare dictionary line.
- DSH
- The common abbreviation for DeepSeek Harness, DeepSeek AI’s open-source agent framework. On this site it also names the plugin ecosystem built on the Cordis runtime — community packages that extend the harness without forking the core. If a listing says “DSH plugin,” it means a package that loads through that runtime, not a general MCP server and not a browser extension.
- DSH plugin
- A package that extends DeepSeek Harness through the Cordis runtime: a tool, an interface surface, a model adapter, or a capability other plugins consume. Most community plugins target the web profile and register through apply(ctx). A plugin is not the same as a skill — a skill changes what the model is instructed to know; a plugin changes what the runtime can do.
- Cordis
- The lightweight plugin runtime underneath DeepSeek Harness. Plugins export an apply(ctx) function and register capabilities through the context object. Because the harness is still in developer preview, plugins that reach into internals break more often than plugins that stay on documented seams. Prefer packages that declare a harness target and use the registration APIs rather than monkey-patching private modules.
- apply(ctx)
- The entry point of a plugin. The harness calls it during load with a context used for registration, event handling and disposal. If apply throws or registers nothing useful, the plugin may load without contributing a tool. When reviewing a listing, read apply before trusting the README: it is where permissions are requested and where hooks are attached.
- dshTarget
- The harness release range a plugin declares it was tested against, recorded verbatim — for example >=0.1.0-rc.6 <0.2.0. Used by users and directories to judge compatibility while DSH is in developer preview. Absent when the plugin declares nothing, which is most of the ecosystem; absence means the author never committed to a range, not that the plugin is broken.
- Profile
- The installation scope for plugins. The web interface uses the web profile, which is why install commands read dsh plugin --profile web add. A plugin installed only under a profile you do not run will appear missing. Keep experimental listings in an isolated profile so a bad install does not take down the profile you use for real work.
- Plugin seam
- A documented point where a capability can be provided or replaced — for example the memory seam or the subagent seam. Plugins that use seams survive releases that break plugins reaching into internals. When a listing says it “hooks memory,” ask whether it goes through the seam or monkey-patches a private module; only the former is upgrade-tolerant.
- Capability
- Something a plugin provides for other plugins to consume, as opposed to a user-facing feature. Capability providers register early so consumers can find them. A directory that lists “agent tools” is often listing capabilities, not UI. If two plugins register the same capability name, the later install can silently override the earlier one — check the tool list after every multi-plugin install.
- Skill
- Instructions and supporting files an agent can load on demand, distributed like a plugin but changing what the model knows rather than what the runtime can do. A skill pack with a slash command and no stop condition is not an agent framework. Prefer skills that name a budget, an approval gate, and a reviewable artifact before they claim autonomy.
- Subagent
- A child agent started by the main agent to work on a scoped task, with its own context. Used for delegation, parallel work and long-running jobs. Without a budget, a subagent can burn steps or wall clock invisibly. Orchestration plugins that spawn children should expose interruption, a durable log, and a clear “done” condition.
- MCP
- Model Context Protocol, a common way to expose tools and data to agents. MCP bridges let a plugin or harness consume standard MCP servers. A DSH plugin that merely wraps an MCP client is convenient but adds another trust boundary: the bridge, the server, and whatever the server can reach. Prefer bridges that document permission prompts and network egress.
- Context window
- Everything the model can currently see: conversation, tool output and injected content. Context plugins measure and manage it; memory plugins decide what survives after it is cleared. When a session “forgets,” the first question is whether compaction dropped the fact or whether it was never written to a store that outlives the window.
- Compaction
- The runtime summarising or dropping older context to stay within the window. Context plugins log when it happens because it silently changes what the model remembers. If your workflow depends on exact tool output from early in the session, either pin that output to a file or accept that compaction will paraphrase it.
- Verification level
- A label from L1 to L5 recording how far a listing’s install path was checked, from repository found to plugin loaded and run. Levels measure how far verification went, not how safe the plugin is. L1–L3 are the honest band on this directory today; L4–L5 imply sandbox load or run tests that no listing currently claims.
- Web UI
- The browser interface served by npx @deepseek-ai/dsh web at 127.0.0.1:3080, and the target most community plugins are built for. Desktop shells and TUIs are different surfaces over the same runtime. If a listing says “web GUI,” it is mounting into this surface unless it documents another host.
- Security scan (site label)
- On this directory, “scanned” means an automated review of the repository without running it: declared dependencies, requested permissions, obvious secret-handling mistakes, and supply-chain signals. It is not a sandbox execution and not a guarantee. “Unverified” means the scan has not happened — not that the plugin is safe or unsafe.
- Install command
- The copy-ready line that installs a listing, usually dsh plugin --profile web add … with a package name or a github:owner/repo#sha spec. Prefer pinned commits while the preview runtime moves. If the command does not resolve on the published package registry, treat the listing as stale until it is fixed.
How to use this glossary on a listing
Scan a plugin page in this order: install command (does it resolve?), verification level (how far was the path checked?), dshTarget (does it declare a harness range?), permission surface (what can it reach?), then health and last check date. If those five fields are missing or stale, the prose on the page matters less than the source repository. Cross-check the same terms againstthe verification guideandsecurity scan statusbefore you install anything load-bearing.
New to the ecosystem? Start with what DeepSeek Harness is, then read the Cordis lifecycle guide.