deepseek-harness/docs/cookbook/extension-cookbook.md

Cookbook: extension plugin shapes

English | 中文

FIXME: This important guide has not received sufficient human design review; complete that review before the first release.

The three plugin shapes you write against the harness extension surface, as illustrative snippets (elided imports and helper stubs — not copy-paste-complete). For the full step-by-step guides see adding a package, adding a tool, and adding an LLM adapter; for the seams these hook into see docs/architecture.md.

A tool plugin

A tool registers on ctx.tools. The annotated defineTool example (typed execute args, result shaping, the run_in_background pattern) lives in adding-a-tool.md — that guide is the source of truth for the tool shape. Raw JSON-Schema ToolDefinitions are also accepted by ctx.tools.register() directly (that is how MCP-sourced tools arrive); defineTool is the typed sugar for first-party tools.

A hook plugin (permission-gate example)

This permission gate is one example of a hook plugin. It returns a typed decision from the tools/pre-execute gate to allow or deny a call; sandbox, permission, and plan-mode plugins can use this seam. Hook plugins can intercept other seams and are not inherently permission gates. A "native hook" is an ordinary Cordis plugin on an interception seam; it needs no external protocol.

import type { Context } from 'cordis'
import type { PreToolDecision, ToolExecution } from '@deepseek-ai/dsh-tools'

declare function isAllowed(exec: ToolExecution): Promise<boolean>

export const name = 'permission-gate'

export function apply(ctx: Context) {
  ctx.on('tools/pre-execute', async (exec, next): Promise<PreToolDecision> => {
    if (!(await isAllowed(exec))) {
      return { kind: 'deny', reason: 'Denied by policy.' }
    }
    return next()
  })
}

This waterfall is the reorderable policy layer. Use ctx.tools.guard() when an invariant needs a monotonic final denial, tools/execute when a plugin must wrap the actual dispatch lifetime (timeouts/retries/metrics; only exec.signal is replaceable), tools/post-execute for explicit result transformation, and tools/result for contained observation of the immutable final outcome. The adding-a-tool guide gives the selection rule.

A UI plugin

A UI plugin renders from the session/event feed (the assistant token stream as assistant/chunk, plus turn/step boundaries and tool activity), and drives input back in via agent.followup() / agent.steer().

import type { Context } from 'cordis'
import { createUserMessage } from '@deepseek-ai/dsh-llm'
import { SessionId } from '@deepseek-ai/dsh-session'

declare function render(text: string): void
declare function onUserInput(handler: (text: string) => void): void

export const name = 'my-ui'
export const inject = ['agents']

export function apply(ctx: Context) {
  ctx.on('session/event', (_session, event) => {
    if (event.type === 'assistant/chunk' && event.data.chunk.type === 'text-delta') {
      render(event.data.chunk.text)
    }
  })
  onUserInput(text => ctx.agents.get(SessionId('client-session'))?.followup(createUserMessage({
    content: [{ type: 'text', text }],
    source: { kind: 'user' },
  })))
}

An external protocol driver

A protocol driver adapts a wire peer to ctx.agents; it may serve a UI or an automation client. A stdio driver owns stdout, creates or resumes agents through the factory, maps the protocol's requests to followup() or cancel(), and settles each request exactly once from durable turn/end. Tear agents down with AgentHandle.dispose() so disposal reaches quiescence.

packages/acp/acp is the automation-only worked example: it exposes fresh text sessions over Agent Client Protocol JSON-RPC stdio, emits committed assistant text, and registers a one-shot machine permission answerer for agents it owns. Its README owns the exact method and lifecycle contract.

import type { Context } from 'cordis'

export const name = 'my-protocol-bridge'
export const inject = ['agents', 'sessions', 'sessionPersistence']

export function apply(ctx: Context) {
  // Stream every logged assistant text/reasoning delta out to the client.
  ctx.on('session/event', (_session, event) => {
    if (event.type === 'assistant/chunk') {
      const chunk = event.data.chunk
      if (chunk.type === 'text-delta') {
        // sendToClient({ kind: 'message_chunk', text: chunk.text })
      }
    }
  })
  // Inbound "prompt": create/resume an agent and feed it; settle on turn end.
  // Teardown reaches quiescence via AgentHandle.dispose() (stop + await exit).
}

Runnable wirings

Runnable leaves load their plugin trees from examples/*/cordis.yml; the root demo:* scripts and those leaf directories are the authoritative inventory. Interactive leaves use @deepseek-ai/dsh-tui, non-interactive leaves use @deepseek-ai/dsh-cli-demo, ACP leaves use @deepseek-ai/dsh-acp-demo, and the app packages share @deepseek-ai/dsh-agent-spine-demo.

The feature → mechanism map

Every product feature maps to a listener on a documented extension seam — the microkernel claim made checkable (microkernel Agent Note). No row modifies the loop.

system-prompt/assemble is an expert cooperative whole-assembly transform: its returned assembly is authoritative, so listener authors own preserving active Code Mode and structured-output protocol contributions. Prefer ctx.tools.restrict() for tool filtering that must stay aligned across presentation, lookup, and execution.

Product feature	Plugin mechanism
Hook system (user + project level)	listeners on agent/session-start, agent/prompt-submit, agent/request, tools/pre-execute, tools/post-execute, and agent/turn-stopping; the waterfall seams return typed decisions, while agent/turn-stopping may steer another step; the dsh-hooks-claude / dsh-hooks-codex bridges map hook config files onto these seams
/goal	ctx.goals owns durable state, dsh-goal-session schedules same-session rounds through the public Agent, and separate command/tool producers expose human/model control
/loop	on the turn/end session event, followup() the next iteration; or force-continue
Dynamic workflow	ctx.workflows + the worker-thread engine + the workflow tool; structured in-process children enforce output with scoped prompt/tool registrations, a monotonic tool guard, final tools/result commit (including enclosing run_code), and the structured-output execution's monotonic concludeTurn() marker
Queued + steering messages	core Agent.followup() / Agent.steer()
Context compaction (auto + manual)	the ctx.compact seam + dsh-compact-basic; automatic pressure runs on serial agent/step, canonical overflow recovery runs on agent/request-error, and manual callers use the same compact service (compaction Agent Note — the model-facing /compact consumer tool is deferred)
System prompt configurability	ctx.systemPrompt.section() with ordering and scope-local shadowing
AGENTS.md (root)	a section provider reading the file
AGENTS.md (subdir, on-touch) + file-change notices	agent.inject() from a watcher / tool-result listener
Built-in tools	ctx.tools.register(); schemas flow into the assembly automatically — the dsh-tool-* families (bash, fs, web, subagent, todo) are the shipped examples
ToolSearch / progressive disclosure	replace a scoped ctx.tools.restrict() registration as the visible set changes; the registry keeps presentation, lookup, and execution aligned
Tool deadline / retry / metrics	wrap core dispatch with tools/execute; a wrapper may replace exec.signal, delegate, and inspect the normalized result in one lexical lifetime
Final tool-result metrics / audit / capture	observe immutable authoritative outcomes with tools/result; use tools/post-execute instead only when the plugin must transform the result or attach context
Monotonic terminal turn policy	call ToolExecution.concludeTurn() from the successful terminal tool; later tool calls in the same response remain guardable, and the loop stops after the step
Subprocess sandbox (landlock / sandbox-exec)	use a ctx.sandbox backend through dsh-bash-sandbox; use tools/pre-execute for capability-level denial
Permission system / AskUserQuestion	return ask from tools/pre-execute and answer through ctx.approval; register a separate model-facing ask tool for ordinary user questions
Plan mode	Shipped: @deepseek-ai/dsh-plan-mode — logged plan/mode state, the plan:policy guidance section, /plan [message] entry, /plan off direct exit, and the user-reviewed exit_plan_mode exit; enforcement stays on the independent sandbox/approval axes
Sub-agent delegation	the ctx.subagents provider registry (dsh-subagent-spawn/-fork/-acp) + dsh-tool-subagent exposing one configured provider to the model
MCP	one plugin per server: discover tools → ctx.tools.register()
Skills	section + tool registration; inject() skill content on invocation
Memory	section provider + tool
Scheduled tasks (cron)	a plugin registers model-callable scheduling tools; timer fires → followup(…, {source: {kind: 'cron', …}}) when idle / inject() notification when busy
UI (GUI; CLI emits JSONL)	listen session/event (assistant chunks, boundaries, tool activity); input → followup()
Telemetry / replayable trace	session/event → JSONL; replay = sessions.create(id, { seed })
Model adapters	LlmAdapter subclass via registerAdapter (dsh-llm-deepseek, dsh-llm-pi-ai)
Plugin hot-reload	every registration is a ctx.effect → vendored HMR just works