CLI & Startup

Official counterpart: the dsh CLI, --profile headless one-shot mode

Framework and plugins are in place. Now we give mini-dsh a "mouth": three run modes, mirroring the official's three usages.

Mode Overview

Command	What it does	Official counterpart
pnpm chat	Interactive REPL, prints as it generates	dsh interactive mode
pnpm run run "task"	One-shot task execution, prints the result	dsh --profile headless "task"
pnpm web	Starts the browser UI	dsh web

Implementation

/**
 * dsh CLI:chat(交互 REPL)与 run(一次性执行)两种形态。
 * 官方:`dsh --profile headless "task"` 是单次任务,`dsh web` 是浏览器应用。
 */

import { createInterface } from "node:readline/promises";
import { stdin as input, stdout as output } from "node:process";
import { buildAgent } from "../boot.ts";
import { web } from "./web.ts";

async function chat(): Promise<void> {
  const { ctx, agent } = await buildAgent();
  // 订阅流式增量,边生成边打印(演示事件驱动 UI)
  const off = ctx.on("assistant/chunk")(({ delta }: any) => {
    if (typeof delta.content === "string") process.stdout.write(delta.content);
  });
  const rl = createInterface({ input, output });
  console.log("mini-dsh chat — 输入 exit 退出\n");
  for (;;) {
    const line = await rl.question("你> ");
    if (line.trim() === "exit" || line.trim() === "") break;
    const reply = await agent.turn(line);
    console.log(`\n\nmini-dsh> ${reply.content}\n`);
  }
  off();
  await ctx.stop();
  rl.close();
}

async function run(task: string): Promise<void> {
  const { ctx, agent } = await buildAgent();
  const reply = await agent.turn(task);
  console.log(reply.content);
  await ctx.stop();
}

async function main(): Promise<void> {
  const [cmd, ...rest] = process.argv.slice(2);
  if (cmd === "chat") return chat();
  if (cmd === "run") return run(rest.join(" "));
  if (cmd === "web") return web(3080);
  console.error(`用法: dsh <chat|run "任务"|web>`);
  process.exit(1);
}

main().catch((err) => {
  console.error(err);
  process.exit(1);
});

chat: event-driven typewriter effect

async function chat(): Promise<void> {
  const { ctx, agent } = await buildAgent();
  // subscribe to streaming deltas and print as they arrive (event-driven UI)
  const off = ctx.on("assistant/chunk")(({ delta }: any) => {
    if (typeof delta.content === "string") process.stdout.write(delta.content);
  });
  const rl = createInterface({ input, output });
  console.log("mini-dsh chat — type exit to quit\n");
  for (;;) {
    const line = await rl.question("you> ");
    if (line.trim() === "exit" || line.trim() === "") break;
    const reply = await agent.turn(line);
    console.log(`\n\nmini-dsh> ${reply.content}\n`);
  }
  off();
  await ctx.stop();
  rl.close();
}

Note: ctx.on("assistant/chunk") subscribes to exactly the event broadcast by the Agent loop via ctx.emit("assistant/chunk"). The UI never touches the core loop — it only subscribes to events. That's the direct payoff of the event-driven architecture.

run: headless one-shot

async function run(task: string): Promise<void> {
  const { ctx, agent } = await buildAgent();
  const reply = await agent.turn(task);
  console.log(reply.content);
  await ctx.stop();
}

Two lines of core logic. Perfect for scripts, CI, and automation pipelines — exactly what the official headless profile does.

Real Run Output

You can run it without an API key — use the scripted demo:

npx tsx demo.ts

Output (real execution — run_bash really invoked echo 1+1 | bc):

让我先算一下。计算完成:1+1=2。

==== 最终回复: 计算完成:1+1=2。 ====

==== 会话日志(append-only SessionEvent)====
  turn/start           {"seq":0,"ts":...,"agent":"5a1b896f-beb"}
  user/message         {"seq":1,"content":"1+1 等于多少?用工具算一下"}
  step/start           {"seq":2,"step":0}
  assistant/message    {"seq":5,"content":"让我先算一下。","tool_calls":[...run_bash...]}
  tool/result          {"seq":6,"tool_call_id":"call_1","content":"2"}
  step/end             {"seq":7,"step":0,"tool_calls":1}
  step/start           {"seq":8,"step":1}
  assistant/message    {"seq":12,"content":"计算完成:1+1=2。","tool_calls":[]}
  step/end             {"seq":13,"step":1,"tool_calls":0}
  turn/end             {"seq":14}

Hooking Up the Real DeepSeek

export DEEPSEEK_API_KEY=sk-your-key
pnpm run run "write a bubble sort and save it to sort.py"

# or point at any OpenAI-compatible endpoint (local vLLM, gateway...)
export DEEPSEEK_BASE_URL=https://your-endpoint
# default model is deepseek-chat; switch via env (e.g. opencode endpoints)
export DEEPSEEK_MODEL=deepseek-v4-flash
pnpm chat

Troubleshooting

• 缺少 DEEPSEEK_API_KEY — you forgot to export, or the key is empty
• DeepSeek API 401 — invalid key
• Network issues — check your proxy/firewall; is DEEPSEEK_BASE_URL reachable?

Next: Testing & Verification →