deepseek-harness/docs/user/develop/practice/index.md

Three-layer capability design

English | 中文

When a capability is general enough to need replaceable implementations, such as Bash execution, Harness splits it into three packages: an interface, an implementation, and a consumer. Each layer can evolve or be replaced independently.

Bash example

The Bash execution capability consists of:

• Interface (dsh-bash) — defines Bash request and result shapes
• Implementation (dsh-bash-local) — executes commands on the local machine

• Consumer (dsh-tool-bash) — exposes the capability as a model-callable tool

  ┌─────────────┐     ┌──────────────────┐     ┌──────────────┐
  │  dsh-bash   │────▶│  dsh-bash-local  │     │ dsh-tool-bash│
  │ (interface) │     │ (implementation) │     │(consumer/tool)│
  └─────────────┘     └──────────────────┘     └──────────────┘
     ▲                                            │
     └────────────────────────────────────────────┘
                  inject: ['bash']
  

Benefits of the split

Replace implementations

One interface can have multiple implementations selected through cordis.yml:

# Local execution
- name: '@deepseek-ai/dsh-bash-local'

# Or a future remote sandbox implementation
# - name: '@deepseek-ai/dsh-bash-remote'
#   config:
#     endpoint: 'https://sandbox.example.com'

The interface and tool remain unchanged while the implementation changes.

Evolve independently

• The interface changes rarely after its contract stabilizes.
• Implementations can improve performance and security independently.
• Consumers can change how they present the capability to the model.

Decouple dependencies

• The implementation depends on the interface.
• The consumer depends on the interface.
• The implementation and consumer do not depend on each other.

Built-in three-layer capabilities

Capability	Interface	Implementation	Consumer
Bash	dsh-bash	dsh-bash-local	dsh-tool-bash
Filesystem	dsh-fs	dsh-fs-local + dsh-fs-policy	dsh-tool-fs
Web	dsh-web	dsh-web-fetch-local / dsh-web-search-*	dsh-tool-web
Subagent	dsh-subagent	dsh-subagent-spawn / dsh-subagent-fork	dsh-tool-subagent
Compaction	dsh-compact	dsh-compact-basic	The implementation consumes agent-loop extension events

Develop a three-layer capability

Step 1: define the interface

// packages/my-cap/my-cap/src/index.ts
import { Service, type Context } from 'cordis'

declare module 'cordis' {
  interface Context {
    myCap: MyCapService
  }
}

export abstract class MyCapService extends Service {
  constructor(ctx: Context) {
    super(ctx, 'myCap')
  }

  /** Execute the capability. */
  abstract execute(request: MyCapRequest): Promise<MyCapResult>
}

export interface MyCapRequest {
  input: string
}

export interface MyCapResult {
  output: string
}

Step 2: write an implementation

// packages/my-cap/my-cap-local/src/index.ts
import type { Context } from 'cordis'
import { MyCapService, type MyCapRequest, type MyCapResult } from '@deepseek-ai/dsh-my-cap'

class MyCapLocal extends MyCapService {
  async execute(request: MyCapRequest): Promise<MyCapResult> {
    // Concrete implementation.
    return { output: request.input.toUpperCase() }
  }
}

export const name = 'my-cap-local'

export function apply(ctx: Context) {
  ctx.plugin(MyCapLocal)
}

Step 3: write a consumer

// packages/my-cap/tool-my-cap/src/index.ts
import type { Context } from 'cordis'
import { defineTool } from '@deepseek-ai/dsh-tools'

export const name = 'tool-my-cap'
export const inject = ['tools', 'myCap']

export function apply(ctx: Context) {
  ctx.tools.register(defineTool({
    name: 'my_cap',
    description: 'Execute my capability.',
    parameters: {
      input: { type: 'string', required: true },
    },
    async execute(args) {
      const result = await ctx.myCap.execute({ input: args.input })
      return [{ type: 'text', text: result.output }]
    },
  }))
}

Compose them in cordis.yml

- name: '@deepseek-ai/dsh-my-cap-local'
- name: '@deepseek-ai/dsh-tool-my-cap'

Design points

• Do not split preemptively — use three packages only when the capability needs replaceable implementations. A simple tool plugin does not.
• The interface owns Request/Result types — implementations and consumers depend only on the interface package.
• Explicit > implicit — resolve defaults in an explicit resolve(request): Spec step rather than hiding ?? default expressions inside run().

Next steps

• LLM adapter — implement an LLM backend, a common capability interface extension