3.3 Host as ACP Client — Architecture

3.2 showed the wire protocol. This section is about the machinery you build around it: subprocess lifecycle, stdio I/O loops, concurrency, permission-UI bridging, and the cross-language shape of it all. Any language that can spawn subprocesses and read/write stdio can implement this — examples below include Node/TypeScript and Go.

3.3.1 The six concerns of an ACP host

A production ACP client has to handle:

Concern	What you need to implement
Process lifecycle	Spawn, health-check, crash-restart, graceful shutdown
Framing	NDJSON decode on stdout, NDJSON encode on stdin, never mix stderr/stdout
Dispatch	Route incoming JSON by shape: response vs. notification vs. server→client request
Matching	Correlate responses back to the id of the request that sent them
UI bridge	Show session/request_permission prompts, route replies back
Observability	Log every RPC; tail the agent's agentao.log for deep debug

You can skip none of these in production.

3.3.2 The three threads (or coroutines)

Whichever language you pick, the minimal structure is three loops running concurrently:

┌──────────────────────────────────────────────────────────┐
│  Host process                                             │
│                                                            │
│  ┌───────────────┐   ┌───────────────┐   ┌─────────────┐ │
│  │ Reader thread │   │ Writer thread │   │ Main logic  │ │
│  │               │   │               │   │             │ │
│  │ parse stdout  │   │ send requests │   │ call(),     │ │
│  │ → dispatch    │   │ from queue    │   │ handle UI   │ │
│  └───────────────┘   └───────────────┘   └─────────────┘ │
│         │                    ▲                 │         │
│         │     response map   │                 │         │
│         └──────────┬─────────┘                 │         │
│                    │                           │         │
│               notifications ◄──────────────────┘         │
└──────────────────────────────────────────────────────────┘
                     ▲ stdio
                     │
           agentao --acp --stdio (subprocess)

Why three? Because you cannot block the reader while handling a message — the agent's next notification is already arriving. Decouple:

• Reader: single-threaded, does nothing but parse one NDJSON line and push it onto an in-memory dispatcher
• Writer: single-threaded, drains a send queue into stdin (single writer avoids interleaving)
• Main logic: calls call(method, params) which blocks on a future/promise, drives the UI, etc.

3.3.3 Dispatch by shape

JSON-RPC 2.0 messages come in three shapes; dispatch logic is identical in every language:

def dispatch(msg: dict) -> None:
    has_id     = "id" in msg
    has_method = "method" in msg

    if has_id and not has_method:
        # RESPONSE — complete the pending request future
        future = pending.pop(msg["id"])
        future.set_result(msg)

    elif has_method and not has_id:
        # NOTIFICATION — route to event handler, never reply
        handle_notification(msg)

    elif has_method and has_id:
        # SERVER → CLIENT REQUEST — respond with result or error
        handle_server_request(msg)    # must eventually send a response

The agent sends server→client requests only in two cases today:

• session/request_permission — tool approval
• _agentao.cn/ask_user — free-form question (extension)

Every one of these must be answered, or the agent blocks forever waiting. Enforce this with a timer.

3.3.4 TypeScript / Node reference implementation

// acp-client.ts
import { spawn, ChildProcessByStdio } from "node:child_process";
import { Readable, Writable } from "node:stream";
import readline from "node:readline";

type JsonRpcId = number | string;
type JsonObject = Record<string, any>;

export class ACPClient {
  private proc!: ChildProcessByStdio<Writable, Readable, Readable>;
  private nextId = 1;
  private pending = new Map<JsonRpcId, (msg: JsonObject) => void>();

  onNotification: (msg: JsonObject) => void = () => {};
  onServerRequest: (msg: JsonObject) => Promise<JsonObject> =
    async () => ({ result: { outcome: { outcome: "cancelled" } } });

  async start() {
    this.proc = spawn("agentao", ["--acp", "--stdio"], {
      stdio: ["pipe", "pipe", "pipe"],
    });

    const rl = readline.createInterface({ input: this.proc.stdout });
    rl.on("line", (line) => this.dispatch(line));

    this.proc.stderr.on("data", (chunk) => {
      // agentao writes logs here; forward to your logger
      process.stderr.write(`[agentao] ${chunk}`);
    });

    await this.call("initialize", {
      protocolVersion: 1,
      clientCapabilities: {},
      clientInfo: { name: "my-host", version: "0.1.0" },
    });
  }

  private async dispatch(line: string) {
    if (!line.trim()) return;
    let msg: JsonObject;
    try { msg = JSON.parse(line); }
    catch (e) { console.error("bad json from agent:", line); return; }

    if ("id" in msg && !("method" in msg)) {
      const resolve = this.pending.get(msg.id);
      if (resolve) { this.pending.delete(msg.id); resolve(msg); }
    } else if ("method" in msg && !("id" in msg)) {
      this.onNotification(msg);
    } else if ("method" in msg && "id" in msg) {
      const response = await this.onServerRequest(msg);
      this.send({ jsonrpc: "2.0", id: msg.id, ...response });
    }
  }

  call(method: string, params: JsonObject): Promise<JsonObject> {
    return new Promise((resolve, reject) => {
      const id = this.nextId++;
      this.pending.set(id, resolve);
      setTimeout(() => {
        if (this.pending.delete(id)) reject(new Error(`rpc timeout: ${method}`));
      }, 60_000);
      this.send({ jsonrpc: "2.0", id, method, params });
    });
  }

  private send(msg: JsonObject) {
    this.proc.stdin.write(JSON.stringify(msg) + "\n");
  }

  async close() {
    this.proc.stdin.end();
    await new Promise((r) => this.proc.on("exit", r));
  }
}

Usage in a VS Code extension:

const acp = new ACPClient();
acp.onNotification = (msg) => {
  if (msg.method === "session/update") {
    updateWebviewStream(msg.params);     // push to UI
  }
};
acp.onServerRequest = async (msg) => {
  if (msg.method === "session/request_permission") {
    const ok = await vscode.window.showWarningMessage(
      msg.params.toolCall.title, { modal: true }, "Allow", "Reject",
    );
    return { result: { outcome: { outcome: "selected", optionId: ok === "Allow" ? "allow_once" : "reject_once" } } };
  }
  return { result: { outcome: { outcome: "cancelled" } } };
};
await acp.start();
const { result: { sessionId } } = await acp.call("session/new", { cwd: workspaceFolder });
await acp.call("session/prompt", {
  sessionId, prompt: [{ type: "text", text: "list files" }],
});

A fuller VS Code integration appears in Blueprint B.

3.3.5 Go reference implementation (skeleton)

// acpclient.go
package acpclient

import (
    "bufio"
    "encoding/json"
    "io"
    "os/exec"
    "sync"
)

type Client struct {
    cmd     *exec.Cmd
    stdin   io.WriteCloser
    stdout  *bufio.Scanner
    nextID  int
    pending map[int]chan json.RawMessage
    mu      sync.Mutex

    OnNotification func(method string, params json.RawMessage)
    OnServerReq    func(method string, params json.RawMessage) json.RawMessage
}

func Start() (*Client, error) {
    cmd := exec.Command("agentao", "--acp", "--stdio")
    stdin, _ := cmd.StdinPipe()
    stdout, _ := cmd.StdoutPipe()
    if err := cmd.Start(); err != nil { return nil, err }

    c := &Client{
        cmd: cmd, stdin: stdin,
        stdout:  bufio.NewScanner(stdout),
        pending: make(map[int]chan json.RawMessage),
    }
    go c.readerLoop()
    return c, nil
}

func (c *Client) readerLoop() {
    for c.stdout.Scan() {
        line := c.stdout.Bytes()
        var head struct {
            ID     *int            `json:"id"`
            Method *string         `json:"method"`
            Params json.RawMessage `json:"params"`
        }
        if err := json.Unmarshal(line, &head); err != nil { continue }

        switch {
        case head.ID != nil && head.Method == nil: // response
            c.mu.Lock()
            ch := c.pending[*head.ID]
            delete(c.pending, *head.ID)
            c.mu.Unlock()
            if ch != nil { ch <- line }
        case head.Method != nil && head.ID == nil: // notification
            if c.OnNotification != nil { c.OnNotification(*head.Method, head.Params) }
        case head.Method != nil && head.ID != nil: // server→client request
            result := c.OnServerReq(*head.Method, head.Params)
            c.sendRaw(map[string]any{"jsonrpc": "2.0", "id": *head.ID, "result": result})
        }
    }
}

func (c *Client) Call(method string, params any) (json.RawMessage, error) {
    c.mu.Lock()
    id := c.nextID; c.nextID++
    ch := make(chan json.RawMessage, 1)
    c.pending[id] = ch
    c.mu.Unlock()
    c.sendRaw(map[string]any{"jsonrpc": "2.0", "id": id, "method": method, "params": params})
    return <-ch, nil
}

func (c *Client) sendRaw(msg any) {
    b, _ := json.Marshal(msg)
    c.stdin.Write(append(b, '\n'))
}

Plug any UI (TUI, web, native) into OnNotification and OnServerReq.

3.3.6 Permission UI bridge — the shape of it

session/request_permission is the trickiest part because it's synchronous from the agent's view: the agent blocks until you respond, yet you have to collect user input asynchronously through your UI.

Common pattern

agent ──req (id=X)──▶ reader thread ──enqueue──▶ UI thread
                                                     │
                                                     ▼
                                           show modal, await click
                                                     │
                                                     ▼
UI thread ──response (id=X)──▶ writer thread ──▶ agent

Must-haves

• Display the tool details: toolCall.title, toolCall.kind, rawInput so users know what they're approving
• Return quickly: the agent holds tool execution. Timeouts keep it from stalling
• Respect options: the agent tells you the allowed reply values (allow_once, reject_once, …). Don't invent your own
• If the user closes the window: respond {"outcome":{"outcome":"cancelled"}}, not silently

See Appendix C.6 for the full schema.

3.3.7 Error handling & reconnection

ACP servers can die. Your host must cope:

Failure	How you'll notice	What to do
Process exits with non-zero code	exit event on child process	Surface error, don't auto-restart unless you capture why
stdout EOF / stream closed	Reader loop sees EOF	Mark all pending requests as failed; don't leak their futures
Unresponsive (no reply in N seconds)	Per-request timer fires	Fail the specific request; decide whether to also session/cancel
initialize returns a lower version than you sent	Response payload	Agentao echoes protocolVersion=1; if you sent 2, accept the downgrade
JSON parse error	Reader catches exception	Log loud; drop the line (don't crash)

Restart strategy: avoid auto-restart during development (makes bugs invisible). In production, exponential backoff + circuit-breaker. Each restart spawns a fresh ACP process; existing sessionIds are lost unless you use session/load during reconnection.

3.3.8 Observability checklist

• Log every outgoing request with its id, method, and params (truncate large payloads)
• Log every incoming response with its id and a duration_ms
• Surface session/update streams in a debug panel
• Tail <session_cwd>/agentao.log for the agent side — that's where stack traces live
• Alert on: RPC timeout rate, process restart count, permission-prompt timeout

---

Next: 3.4 Reverse: calling external ACP agents →