DeepSeek Harness glossary: scope, goals, turns, and seams

DeepSeek Harness uses familiar words such as scope, goal, turn, and step in precise ways. Understanding those definitions prevents configuration mistakes and misleading bug reports. This glossary explains the official terms in practical language, shows how related concepts differ, and links each term to the part of the runtime where it matters.

The canonical definitions live in the official Chinese DeepSeek Harness glossary. This page expands them with examples for readers searching the documentation.

Capability seam

A capability seam is a replaceable capability boundary. It contains:

1. a service definition that describes the contract;
2. one or more providers that implement the contract; and
3. one or more consumers that call it.

Shell execution is a useful mental model: a consumer asks for a command to run through a stable service, while a local or remote provider decides how that command is actually executed. The consumer does not need a direct dependency on every provider.

A seam is more than an interface name. Provider registration, plugin lifecycle, selection, and errors are part of the design. See DeepSeek Harness architecture for guidance on when a seam is appropriate.

Agent scope

An Agent scope is the scoped Cordis context created for one Agent. agent.ctx refers to that context. It is designed to hold Agent-local registrations and behavior.

The scope key uses object identity, not a matching string. The model is deliberately flat: a subagent receives its own scope rather than inheriting a deep parent scope tree. Parent and subagent relationships are recorded as lineage data, not treated as automatic visibility.

Global tool restrictions are applied before scope-local tools are merged. That means a scope can add its own relevant tools without silently undoing restrictions on the global catalog. The setup window creates registrations before the session is published, reducing races where consumers observe a half-configured Agent.

Goal

A Goal is the single persistent objective attached to a session. It can move through phases such as active, paused, blocked, and complete. A Goal Round provides a bounded autonomous work period for pursuing it.

A Goal is not a second conversation and is not a general-purpose scheduler. The session remains the conversation record. Scheduling a future action is a separate capability, and the tool catalog lists goal and schedule tools independently.

Goal activation can be armed or disarmed locally. That activation switch is control state and is not replayed as if it were a new persistent goal fact.

Human command

A human command is a slash-style command handled by the human adapter. It is not sent to the model as a normal user message and is not automatically a shell command.

This boundary prevents interface control from polluting the model conversation. When documenting a command, state whether it changes adapter behavior, invokes a tool, or adds a model-visible message; users should not have to infer the path from its leading slash.

Turn

A turn is the runtime response cycle triggered by input. It can contain zero or more model steps. The turn starts before a model request and finishes after the Agent has completed the resulting sequence of model and tool work, or exits through a terminal path.

Use “turn” for conversation-level traces and status. One user follow-up usually starts one turn, but the turn can include several calls to the model.

Step

A step is one model request plus the work associated with that response. If the response asks for tools, their results can lead to another step in the same turn.

Use “step” when measuring request count, diagnosing prompt assembly, tracing provider errors, or determining why tool output caused another model call. The Agent lifecycle guide shows the complete sequence.

Goal Round

A Goal Round is a bounded autonomous cycle used while an active Goal is being pursued. It is not interchangeable with a turn or step:

• a step is one model request;
• a turn is one input-driven execution cycle; and
• a Goal Round is a budgeted period of autonomous goal work that may span execution activity.

Keeping these units separate makes limits and telemetry meaningful. A “three-round” budget should not be reported as three model requests unless that is actually what happened.

Quick comparison table

Term	What it controls	Persistent by itself?
Capability seam	Replaceable service providers	Definition is code/config
Agent scope	Agent-local registrations	Scope is runtime state
Goal	Session objective and phase	Yes
Human command	Human-adapter control	Not a model message
Turn	One input-driven response cycle	Events may be persisted
Step	One model request cycle	Results may be persisted
Goal Round	Bounded autonomous goal work	Budget/progress depends on goal system

Official source

This expanded glossary was independently written from the official DeepSeek Harness glossary, reviewed on 2026-08-20. When an implementation or third-party article uses a term differently, prefer the current upstream definition.