Glossary

This page defines the core DeepSeek Harness terms for users and developers. The upstream repository's fuller glossary also covers scope, lineage, Ralph and other implementation details. For background see what is DeepSeek Harness, and for architecture see the Cordis primer.

harness

The runtime frame that hosts agents: it wires model adapters, tools, the filesystem, shell, sandbox, approvals, persistence and the agent loop together as plugins. DeepSeek Harness (dsh) is one such agent harness, built on the idea that everything is a plugin.

plugin

The unit of composition. Every capability — tools, skills, goals, subagents, UI blocks, model routing — is a plugin composed through Cordis. A Cordis plugin implements a Service: a function with optional inject and apply(ctx) fields, or a Service subclass.

profile

A named assembly under $DSH_HOME/profiles/<name>. It lists the bundles it stacks, stores its own out-of-tree plugins, and keeps the user's own cordis.patch.yml. Boot it with dsh --profile <name>; web and headless auto-initialize from shipped templates on first use.

preset (agent preset)

One session's agent composition, authored as one directory per preset with a cordis.yml file: the persona, prompt sections, tools, skills and restrictions. An agent is composed from a preset. See presets.

skill

A reusable, task-specific instruction package loaded on demand (for example astro, code-review), registered through the skill provider registry. See skills & goals.

goal

One durable completion objective attached to an existing session, with a revisioned active / paused / blocked / complete phase and a goal-round cap. A goal is state — not a scheduler or a separate conversation; the session log remains its source of truth.

subagent

A delegated agent that runs a self-contained task in its own context, implemented through the subagent capability seam and used to offload independent work. See subagents & workflows.

workflow

Multi-agent orchestration: a script fans work out across many subagents, running in phases with structured results, implemented through the workflow capability.

Cordis

The plugin framework underneath the harness, vendored in. It composes plugins, services, typed events and lifecycle (fibers); "everything is a plugin" is its model. The design is described in ecosystem.

seam

A swappable capability with three roles: a Service Definition (the Cordis Service that owns its ctx.<key> and vocabulary types), one or more Service Providers, and one or more Consumers that inject the service. The seam is the complete capability, never one role.

Service Provider

A naming role of a capability seam: it provides the implementation of a Service Definition. Singular "Service Provider", plural "Service Providers".

fiber

One loaded plugin instance: its lifecycle state, validated config, and registered effects. ctx.fiber is the current fiber, and ctx.effect() ties side effects to it so unloading unwinds them in reverse order.

agent loop

The core loop that drains a turn — model requests and the tool executions they cause — until the model and its tools stop or a terminal policy intervenes.

tool

A model-facing capability the agent can invoke, registered on the tools registry. See tools.

slot

A named, registrable position in the client architecture. Client plugins register UI into slots, and the browser UI is composed from them. See client UI & slots.

Host and Client

The two runtime planes of a dynamic plugin. Host runs in the Node.js process and owns files, networking, commands, services and model tools, plus JSON methods the Client can call. Client runs in the browser page and owns themes, layout, page state, tool cards and Slot UI. They communicate through package-private JSON methods (direction Client → Host).

dynamic plugin

A temporary runtime extension defined and run against the live process; its definitions are not written to source or disk and do not survive a restart. See dynamic plugins.

agent

The acting subject in a session: it reads and edits workspace files, runs commands, delegates work and maintains plans. A live agent is the key of its own scope.

turn and step

Two units of the loop hierarchy. A turn is one drain of admitted input in a session, ending after the model and its tools stop or a terminal policy intervenes. A step is one model request plus the tool executions caused by its response; a turn contains zero or more steps.