오케스트레이션 모드 만들기

오케스트레이션 모드 만들기 (Build an orchestration mode)

세션 단위 모드를 만들어서, 대화 중간 시스템 메시지로 켜고 끄며 다중 에이전트 팬아웃에 대한 상시 동의를 부여하는 방법을 다뤄요.

출처: 문서

본문

오케스트레이션 모드는 세션 단위 스위치예요. 켜져 있으면, 모델은 모든 실질적인 요청에 최대한의 철저함을 쏟고, 태스크를 스스로 정찰한 뒤 기본적으로 병렬 서브에이전트들에게 일을 나눠줘요. 꺼져 있으면, 같은 오케스트레이션 도구가 다시 요청별 옵트인 방식으로 돌아가요.

이 모드는 API 파라미터가 아니에요. 전부 문서화된 구성 요소로만 만들어져요:

  1. 노력 수준(effort level): 요청은 xhigh 같은 문서화된 Effort 값으로 실행돼요. 그 페이지에 있는 수준보다 더 높은 숨은 수준은 없어요. 이 예제는 각 요청의 최상위 레벨에서 effort를 설정하는데, 이는 베타 헤더가 필요 없어요.
  2. 모드 리마인더: 대화 중간 시스템 메시지가 모델에게 모드가 활성 상태임을 알려줘요. 몇 턴마다 한 줄짜리 리프레셔를 보내고, 모드가 꺼질 때는 종료 알림을 보내요. 최상위 system 필드는 절대 바뀌지 않으므로 캐시된 프리픽스가 유지돼요.
  3. 도구 설명 안의 상시 동의(standing consent): 오케스트레이션 도구의 설명은, 모드가 켜져 있는 동안 모델이 모든 실질적인 태스크에 대해 먼저 묻지 않고 워크플로를 작성·실행해야 한다고 명시해요.
이 예제는 대화 중간 시스템 메시지를 사용해요. 이를 지원하는 모델과 플랫폼은 [대화 중간 시스템 메시지](https://platform.claude.com/docs/en/build-with-claude/mid-conversation-system-messages)를 참고하세요. 팬아웃 자체가 토큰 사용량을 배가시켜요. 단일 요청이 많은 서브에이전트 대화를 만들어낼 수 있으니, 비용이 정당한 작업에만 이 모드를 사용하세요.

루프 설정하기

예제는 단일 파일이에요. 상수들은 노력 수준, 팬아웃 형태, 그리고 모드 리프레셔를 얼마나 자주 다시 보낼지를 제어해요. MAX_CONCURRENT는 동시에 실행되는 서브에이전트 수를 제한하고(PHP 포트는 순차 실행이라 이를 무시해요), MAX_TOTAL_SUBTASKS는 모델이 단일 Workflow 호출에서 큐에 넣을 수 있는 수를 제한해요. 둘을 분리하면 모델이 큰 백로그를 한꺼번에 실행하지 않고 계획할 수 있어요. DOC_TEST_MODE 확인은 그 환경 변수가 설정되어 있을 때 루프를 단일 턴으로 제한해서, 자동화된 문서 하네스가 전체 오케스트레이션을 실행하지 않고도 파일이 컴파일되고 빠르게 끝나는지 검증할 수 있게 해줘요. 직접 실행할 때는 이것을 설정하지 않은 채로 두세요.

```python Python import atexit import concurrent.futures import hashlib import json import os import shutil import subprocess import sys import tempfile import threading

import anthropic

client = anthropic.Anthropic()

MODEL = "claude-opus-5-5" EFFORT = "xhigh"

SYSTEM_PROMPT = "You are a helpful general-purpose agent. Answer the user's request directly."

REQUEST_TIMEOUT_SECONDS = 600 BASH_TIMEOUT_SECONDS = 60 TOOL_RESULT_MAX_CHARS = 8000 MAX_CONCURRENT = 10 DOC_TEST_MODE = bool(os.environ.get("DOC_TEST_MODE")) MAX_TOTAL_SUBTASKS = 2 if DOC_TEST_MODE else 200 MAX_SUBAGENT_TURNS = 1 if DOC_TEST_MODE else 15 MAX_MAIN_TURNS = 1 if DOC_TEST_MODE else 30 TURNS_BETWEEN_REFRESHERS = 10 JOURNAL_PATH = os.environ.get("ORCH_JOURNAL") or "orchestration_journal.json"


```typescript TypeScript
import { exec } from "node:child_process";
import { createHash } from "node:crypto";
import { rmSync } from "node:fs";
import { mkdtemp, readFile, rename, writeFile } from "node:fs/promises";
import { tmpdir } from "node:os";
import { join } from "node:path";
import { promisify } from "node:util";

import Anthropic from "@anthropic-ai/sdk";

const client = new Anthropic();

const MODEL = "claude-opus-5-5";
const EFFORT = "xhigh";

const SYSTEM_PROMPT =
  "You are a helpful general-purpose agent. Answer the user's request directly.";

const REQUEST_TIMEOUT_SECONDS = 600;
const BASH_TIMEOUT_SECONDS = 60;
const TOOL_RESULT_MAX_CHARS = 8000;
const MAX_CONCURRENT = 10;
const DOC_TEST_MODE = Boolean(process.env.DOC_TEST_MODE);
const MAX_TOTAL_SUBTASKS = DOC_TEST_MODE ? 2 : 200;
const MAX_SUBAGENT_TURNS = DOC_TEST_MODE ? 1 : 15;
const MAX_MAIN_TURNS = DOC_TEST_MODE ? 1 : 30;
const TURNS_BETWEEN_REFRESHERS = 10;
const JOURNAL_PATH = process.env.ORCH_JOURNAL || "orchestration_journal.json";
using System.Diagnostics;
using System.Security.Cryptography;
using System.Text;
using System.Text.Json;
using Anthropic;
using Anthropic.Models.Messages;

AnthropicClient client = new();

const Model model = Model.ClaudeOpus5_5;
var effort = Effort.Xhigh;

const string systemPrompt = "You are a helpful general-purpose agent. Answer the user's request directly.";

const int requestTimeoutSeconds = 600;
// The other ports stream with max_tokens 64000. This port uses non-streaming
// Messages.Create, and the API rejects non-streaming requests at that size.
// 8192 is the non-streaming ceiling for Opus 4.0 and 4.1 and a conservative
// choice for newer Opus models.
const int requestMaxTokens = 8192;
const int bashTimeoutSeconds = 60;
const int toolResultMaxChars = 8000;
const int maxConcurrent = 10;
var docTestMode = Environment.GetEnvironmentVariable("DOC_TEST_MODE") is { Length: > 0 };
int maxTotalSubtasks = docTestMode ? 2 : 200;
int maxSubagentTurns = docTestMode ? 1 : 15;
int maxMainTurns = docTestMode ? 1 : 30;
const int turnsBetweenRefreshers = 10;
var journalPath = Environment.GetEnvironmentVariable("ORCH_JOURNAL") is { Length: > 0 } p ? p : "orchestration_journal.json";
import (
	"bytes"
	"cmp"
	"context"
	"crypto/sha256"
	"encoding/hex"
	"encoding/json"
	"errors"
	"fmt"
	"log"
	"os"
	"os/exec"
	"path/filepath"
	"strings"
	"sync"
	"time"

	"github.com/anthropics/anthropic-sdk-go"
)

var client = anthropic.NewClient()

const (
	modelID = anthropic.ModelClaudeOpus5_5
	effort  = anthropic.OutputConfigEffortXhigh

	systemPrompt = "You are a helpful general-purpose agent. Answer the user's request directly."

	requestTimeoutSeconds  = 600
	bashTimeoutSeconds     = 60
	toolResultMaxChars     = 8000
	maxConcurrent          = 10
	turnsBetweenRefreshers = 10
)

var (
	docTestMode      = os.Getenv("DOC_TEST_MODE") != ""
	maxTotalSubtasks = ifTest(2, 200)
	maxSubagentTurns = ifTest(1, 15)
	maxMainTurns     = ifTest(1, 30)
	journalPath      = cmp.Or(os.Getenv("ORCH_JOURNAL"), "orchestration_journal.json")
)

func ifTest(test, normal int) int {
	if docTestMode {
		return test
	}
	return normal
}

import com.anthropic.client.AnthropicClient;
import com.anthropic.client.okhttp.AnthropicOkHttpClient;
import com.anthropic.core.JsonValue;
import com.anthropic.core.RequestOptions;
import com.anthropic.helpers.MessageAccumulator;
import com.anthropic.models.messages.ContentBlock;
import com.anthropic.models.messages.ContentBlockParam;
import com.anthropic.models.messages.Message;
import com.anthropic.models.messages.MessageCreateParams;
import com.anthropic.models.messages.MessageParam;
import com.anthropic.models.messages.Model;
import com.anthropic.models.messages.OutputConfig;
import com.anthropic.models.messages.StopReason;
import com.anthropic.models.messages.TextBlock;
import com.anthropic.models.messages.Tool;
import com.anthropic.models.messages.ToolBash20250124;
import com.anthropic.models.messages.ToolResultBlockParam;
import com.anthropic.models.messages.ToolUseBlock;
import com.fasterxml.jackson.core.JsonProcessingException;
import com.fasterxml.jackson.core.type.TypeReference;
import com.fasterxml.jackson.databind.JsonNode;
import com.fasterxml.jackson.databind.ObjectMapper;
import java.io.IOException;
import java.io.UncheckedIOException;
import java.nio.charset.StandardCharsets;
import java.nio.file.Files;
import java.nio.file.Path;
import java.nio.file.StandardCopyOption;
import java.security.MessageDigest;
import java.time.Duration;
import java.util.ArrayList;
import java.util.Comparator;
import java.util.HashMap;
import java.util.HexFormat;
import java.util.List;
import java.util.Map;
import java.util.Objects;
import java.util.Optional;
import java.util.concurrent.Callable;
import java.util.concurrent.CancellationException;
import java.util.concurrent.CompletableFuture;
import java.util.concurrent.ExecutionException;
import java.util.concurrent.ExecutorService;
import java.util.concurrent.Executors;
import java.util.concurrent.Future;
import java.util.concurrent.TimeUnit;
import java.util.concurrent.locks.ReentrantLock;
import java.util.stream.Collectors;
import java.util.stream.IntStream;

AnthropicClient client = AnthropicOkHttpClient.fromEnv();

static final Model MODEL = Model.CLAUDE_OPUS_5_5;
static final boolean DOC_TEST_MODE =
        !Objects.requireNonNullElse(System.getenv("DOC_TEST_MODE"), "").isEmpty();
static final OutputConfig.Effort EFFORT = OutputConfig.Effort.XHIGH;

static final String SYSTEM_PROMPT =
        "You are a helpful general-purpose agent. Answer the user's request directly.";

static final int REQUEST_TIMEOUT_SECONDS = 600;
static final RequestOptions REQUEST_OPTIONS =
        RequestOptions.builder().timeout(Duration.ofSeconds(REQUEST_TIMEOUT_SECONDS)).build();
static final int BASH_TIMEOUT_SECONDS = 60;
static final int TOOL_RESULT_MAX_CHARS = 8000;
static final int MAX_CONCURRENT = 10;
static final int MAX_TOTAL_SUBTASKS = DOC_TEST_MODE ? 2 : 200;
static final int MAX_SUBAGENT_TURNS = DOC_TEST_MODE ? 1 : 15;
static final int MAX_MAIN_TURNS = DOC_TEST_MODE ? 1 : 30;
static final int TURNS_BETWEEN_REFRESHERS = 10;
static final Path JOURNAL_PATH = Path.of(Optional.ofNullable(System.getenv("ORCH_JOURNAL"))
        .filter(s -> !s.isEmpty()).orElse("orchestration_journal.json"));
use Anthropic\Client;
use Anthropic\Messages\ToolUseBlock;

$client = new Client();

const MODEL = 'claude-opus-5-5';
define('DOC_TEST_MODE', (string) getenv('DOC_TEST_MODE') !== '');
const EFFORT = 'xhigh';

const SYSTEM_PROMPT = 'You are a helpful general-purpose agent. Answer the user\'s request directly.';

const REQUEST_TIMEOUT_SECONDS = 600;
const BASH_TIMEOUT_SECONDS = 60;
const TOOL_RESULT_MAX_CHARS = 8000;
const MAX_CONCURRENT = 10;
define('MAX_TOTAL_SUBTASKS', DOC_TEST_MODE ? 2 : 200);
define('MAX_SUBAGENT_TURNS', DOC_TEST_MODE ? 1 : 15);
define('MAX_MAIN_TURNS', DOC_TEST_MODE ? 1 : 30);
const TURNS_BETWEEN_REFRESHERS = 10;
define('JOURNAL_PATH', getenv('ORCH_JOURNAL') ?: 'orchestration_journal.json');
require "anthropic"
require "digest"
require "fileutils"
require "json"
require "open3"
require "tmpdir"

CLIENT = Anthropic::Client.new

MODEL = "claude-opus-5-5"
EFFORT = :xhigh

SYSTEM_PROMPT = "You are a helpful general-purpose agent. Answer the user's request directly."

REQUEST_TIMEOUT_SECONDS = 600
BASH_TIMEOUT_SECONDS = 60
TOOL_RESULT_MAX_CHARS = 8000
MAX_CONCURRENT = 10
DOC_TEST_MODE = !ENV["DOC_TEST_MODE"].to_s.empty?
MAX_TOTAL_SUBTASKS = DOC_TEST_MODE ? 2 : 200
MAX_SUBAGENT_TURNS = DOC_TEST_MODE ? 1 : 15
MAX_MAIN_TURNS = DOC_TEST_MODE ? 1 : 30
TURNS_BETWEEN_REFRESHERS = 10
JOURNAL_PATH = ENV["ORCH_JOURNAL"].to_s.empty? ? "orchestration_journal.json" : ENV["ORCH_JOURNAL"]

모드 리마인더 정의하기

리마인더는 의도적으로 짧게 만들어요. 모드를 전환하고, 무거운 지침이 담긴 도구 설명을 가리키기만 하면 되거든요. 전체 텍스트는 모드가 켜질 때 한 번 보내고, 리프레셔는 사용자 턴 몇 번 뒤에만 다시 보내며, 종료 알림은 모드가 꺼질 때 한 번 보내요.

```python Python MODE_ENTER = ( "Orchestration mode is on: optimize for the most exhaustive, correct answer rather than " "the fastest one. Use the Workflow tool on every substantive task, sized to the problem's " "natural decomposition rather than the maximum the tool allows. See the Workflow tool's " "description for standing consent, granularity guidance, and quality patterns. Work solo " "only on conversational or trivial turns." ) MODE_REFRESH = ( "Orchestration mode is still on. Use the Workflow tool; see its standing consent section." ) MODE_EXIT = ( "Orchestration mode is off. The Workflow tool's standard opt-in rule applies again." ) ```
const MODE_ENTER =
  "Orchestration mode is on: optimize for the most exhaustive, correct answer rather than " +
  "the fastest one. Use the Workflow tool on every substantive task, sized to the problem's " +
  "natural decomposition rather than the maximum the tool allows. See the Workflow tool's " +
  "description for standing consent, granularity guidance, and quality patterns. Work solo " +
  "only on conversational or trivial turns.";
const MODE_REFRESH =
  "Orchestration mode is still on. Use the Workflow tool; see its standing consent section.";
const MODE_EXIT =
  "Orchestration mode is off. The Workflow tool's standard opt-in rule applies again.";
const string modeEnter =
    "Orchestration mode is on: optimize for the most exhaustive, correct answer rather than "
    + "the fastest one. Use the Workflow tool on every substantive task, sized to the problem's "
    + "natural decomposition rather than the maximum the tool allows. See the Workflow tool's "
    + "description for standing consent, granularity guidance, and quality patterns. Work solo "
    + "only on conversational or trivial turns.";
const string modeRefresh =
    "Orchestration mode is still on. Use the Workflow tool; see its standing consent section.";
const string modeExit =
    "Orchestration mode is off. The Workflow tool's standard opt-in rule applies again.";
const (
	modeEnter = "Orchestration mode is on: optimize for the most exhaustive, correct answer rather than " +
		"the fastest one. Use the Workflow tool on every substantive task, sized to the problem's " +
		"natural decomposition rather than the maximum the tool allows. See the Workflow tool's " +
		"description for standing consent, granularity guidance, and quality patterns. Work solo " +
		"only on conversational or trivial turns."
	modeRefresh = "Orchestration mode is still on. Use the Workflow tool; see its standing consent section."
	modeExit    = "Orchestration mode is off. The Workflow tool's standard opt-in rule applies again."
)

static final String MODE_ENTER =
        "Orchestration mode is on: optimize for the most exhaustive, correct answer rather than "
                + "the fastest one. Use the Workflow tool on every substantive task, sized to the problem's "
                + "natural decomposition rather than the maximum the tool allows. See the Workflow tool's "
                + "description for standing consent, granularity guidance, and quality patterns. Work solo "
                + "only on conversational or trivial turns.";
static final String MODE_REFRESH =
        "Orchestration mode is still on. Use the Workflow tool; see its standing consent section.";
static final String MODE_EXIT =
        "Orchestration mode is off. The Workflow tool's standard opt-in rule applies again.";
const MODE_ENTER =
    'Orchestration mode is on: optimize for the most exhaustive, correct answer rather than '
    . 'the fastest one. Use the Workflow tool on every substantive task, sized to the problem\'s '
    . 'natural decomposition rather than the maximum the tool allows. See the Workflow tool\'s '
    . 'description for standing consent, granularity guidance, and quality patterns. Work solo '
    . 'only on conversational or trivial turns.';
const MODE_REFRESH =
    'Orchestration mode is still on. Use the Workflow tool; see its standing consent section.';
const MODE_EXIT =
    'Orchestration mode is off. The Workflow tool\'s standard opt-in rule applies again.';
MODE_ENTER =
  "Orchestration mode is on: optimize for the most exhaustive, correct answer rather than " \
  "the fastest one. Use the Workflow tool on every substantive task, sized to the problem's " \
  "natural decomposition rather than the maximum the tool allows. See the Workflow tool's " \
  "description for standing consent, granularity guidance, and quality patterns. Work solo " \
  "only on conversational or trivial turns."
MODE_REFRESH =
  "Orchestration mode is still on. Use the Workflow tool; see its standing consent section."
MODE_EXIT =
  "Orchestration mode is off. The Workflow tool's standard opt-in rule applies again."

도구 설명에 상시 동의 부여하기

Workflow 도구가 실제 행동 계약을 담고 있어요. 옵트인 규칙, 모드가 켜져 있는 동안 적용되는 상시 동의, 팬아웃 크기를 정하는 세분성(granularity) 가이드, 그리고 모델이 참고할 수 있는 품질 패턴(검증 웨이브, 완전성 비평가, 다단계 시퀀싱)이 그것이에요. 서브에이전트에게도 report_findings 도구가 주어져서, 결과가 산문이 아닌 구조화된 JSON으로 돌아와요. 그리고 bash 도구는 Anthropic이 정의한 bash_20250124 도구를 로컬에서 실행한 것이에요.

```python Python WORKFLOW_TOOL = { "name": "Workflow", "description": ( "Orchestrate a multiagent workflow: split a large task into independent subtasks " "and run them as parallel agents, then collect their results.\n\n" "Opt-in: only use this tool when the user explicitly asks for a workflow, or when a " "system message confirms that orchestration mode is on.\n\n" "Quality patterns: adversarial verification (a second wave of agents checks the first " "wave's findings against the source), a completeness critic (one agent hunts for what " "the others missed), and multiphase sequencing (understand, design, implement, and " "review as separate workflow calls, reading results between phases). A useful default " "is hybrid: scout inline first to discover the work-list, then fan out over it.\n\n" "Granularity: scope each subtask to a distinct concern, component, or question rather " "than per line or per file section. Scale the count to what the user asked for: a " "focused review of a module of a few hundred lines rarely needs more than about ten " "subtasks; a broad audit of a large codebase can justify more.\n\n" "Standing consent: while a system message confirms orchestration mode is on, that " "opt-in is standing. Author and run a workflow for every substantive task by default, " "and lean toward verifying findings adversarially. Work solo only on conversational " "turns or trivial mechanical edits. When a system message says the mode is off, " "revert to the opt-in rule above." ), "input_schema": { "type": "object", "properties": { "subtasks": { "type": "array", "items": {"type": "string"}, "description": "Independent subtask prompts to run as parallel agents", } }, "required": ["subtasks"], }, }

BASH_TOOL = {"type": "bash_20250124", "name": "bash"}

REPORT_TOOL = { "name": "report_findings", "description": ( "Report the final findings for your subtask. Call this exactly once, when you are " "done investigating; it ends your task." ), "input_schema": { "type": "object", "properties": { "summary": {"type": "string", "description": "Two or three sentences of synthesis"}, "findings": { "type": "array", "items": { "type": "object", "properties": { "claim": {"type": "string", "description": "The finding, one sentence"}, "evidence": { "type": "string", "description": "How it was verified (file, line, or command output)", }, "severity": {"type": "string", "enum": ["high", "medium", "low", "info"]}, }, "required": ["claim", "evidence", "severity"], }, }, }, "required": ["summary", "findings"], }, }


```typescript TypeScript
const WORKFLOW_TOOL: Anthropic.Tool = {
  name: "Workflow",
  description:
    "Orchestrate a multiagent workflow: split a large task into independent subtasks " +
    "and run them as parallel agents, then collect their results.\n\n" +
    "Opt-in: only use this tool when the user explicitly asks for a workflow, or when a " +
    "system message confirms that orchestration mode is on.\n\n" +
    "Quality patterns: adversarial verification (a second wave of agents checks the first " +
    "wave's findings against the source), a completeness critic (one agent hunts for what " +
    "the others missed), and multiphase sequencing (understand, design, implement, and " +
    "review as separate workflow calls, reading results between phases). A useful default " +
    "is hybrid: scout inline first to discover the work-list, then fan out over it.\n\n" +
    "Granularity: scope each subtask to a distinct concern, component, or question rather " +
    "than per line or per file section. Scale the count to what the user asked for: a " +
    "focused review of a module of a few hundred lines rarely needs more than about ten " +
    "subtasks; a broad audit of a large codebase can justify more.\n\n" +
    "Standing consent: while a system message confirms orchestration mode is on, that " +
    "opt-in is standing. Author and run a workflow for every substantive task by default, " +
    "and lean toward verifying findings adversarially. Work solo only on conversational " +
    "turns or trivial mechanical edits. When a system message says the mode is off, " +
    "revert to the opt-in rule above.",
  input_schema: {
    type: "object",
    properties: {
      subtasks: {
        type: "array",
        items: { type: "string" },
        description: "Independent subtask prompts to run as parallel agents",
      },
    },
    required: ["subtasks"],
  },
};

const BASH_TOOL: Anthropic.ToolBash20250124 = { type: "bash_20250124", name: "bash" };

const REPORT_TOOL: Anthropic.Tool = {
  name: "report_findings",
  description:
    "Report the final findings for your subtask. Call this exactly once, when you are " +
    "done investigating; it ends your task.",
  input_schema: {
    type: "object",
    properties: {
      summary: { type: "string", description: "Two or three sentences of synthesis" },
      findings: {
        type: "array",
        items: {
          type: "object",
          properties: {
            claim: { type: "string", description: "The finding, one sentence" },
            evidence: {
              type: "string",
              description: "How it was verified (file, line, or command output)",
            },
            severity: { type: "string", enum: ["high", "medium", "low", "info"] },
          },
          required: ["claim", "evidence", "severity"],
        },
      },
    },
    required: ["summary", "findings"],
  },
};
Tool workflowTool = new()
{
    Name = "Workflow",
    Description =
        "Orchestrate a multiagent workflow: split a large task into independent subtasks "
        + "and run them as parallel agents, then collect their results.\n\n"
        + "Opt-in: only use this tool when the user explicitly asks for a workflow, or when a "
        + "system message confirms that orchestration mode is on.\n\n"
        + "Quality patterns: adversarial verification (a second wave of agents checks the first "
        + "wave's findings against the source), a completeness critic (one agent hunts for what "
        + "the others missed), and multiphase sequencing (understand, design, implement, and "
        + "review as separate workflow calls, reading results between phases). A useful default "
        + "is hybrid: scout inline first to discover the work-list, then fan out over it.\n\n"
        + "Granularity: scope each subtask to a distinct concern, component, or question rather "
        + "than per line or per file section. Scale the count to what the user asked for: a "
        + "focused review of a module of a few hundred lines rarely needs more than about ten "
        + "subtasks; a broad audit of a large codebase can justify more.\n\n"
        + "Standing consent: while a system message confirms orchestration mode is on, that "
        + "opt-in is standing. Author and run a workflow for every substantive task by default, "
        + "and lean toward verifying findings adversarially. Work solo only on conversational "
        + "turns or trivial mechanical edits. When a system message says the mode is off, "
        + "revert to the opt-in rule above.",
    InputSchema = new InputSchema
    {
        Properties = new Dictionary<string, JsonElement>
        {
            ["subtasks"] = JsonSerializer.SerializeToElement(new
            {
                type = "array",
                items = new { type = "string" },
                description = "Independent subtask prompts to run as parallel agents",
            }),
        },
        Required = ["subtasks"],
    },
};

ToolBash20250124 bashTool = new();

Tool reportTool = new()
{
    Name = "report_findings",
    Description =
        "Report the final findings for your subtask. Call this exactly once, when you are "
        + "done investigating; it ends your task.",
    InputSchema = new InputSchema
    {
        Properties = new Dictionary<string, JsonElement>
        {
            ["summary"] = JsonSerializer.SerializeToElement(new
            {
                type = "string",
                description = "Two or three sentences of synthesis",
            }),
            ["findings"] = JsonSerializer.SerializeToElement(new
            {
                type = "array",
                items = new
                {
                    type = "object",
                    properties = new
                    {
                        claim = new { type = "string", description = "The finding, one sentence" },
                        evidence = new
                        {
                            type = "string",
                            description = "How it was verified (file, line, or command output)",
                        },
                        severity = new { type = "string", @enum = new[] { "high", "medium", "low", "info" } },
                    },
                    required = new[] { "claim", "evidence", "severity" },
                },
            }),
        },
        Required = ["summary", "findings"],
    },
};
var workflowTool = anthropic.ToolUnionParam{
	OfTool: &anthropic.ToolParam{
		Name: "Workflow",
		Description: anthropic.String("Orchestrate a multiagent workflow: split a large task into independent subtasks " +
			"and run them as parallel agents, then collect their results.\n\n" +
			"Opt-in: only use this tool when the user explicitly asks for a workflow, or when a " +
			"system message confirms that orchestration mode is on.\n\n" +
			"Quality patterns: adversarial verification (a second wave of agents checks the first " +
			"wave's findings against the source), a completeness critic (one agent hunts for what " +
			"the others missed), and multiphase sequencing (understand, design, implement, and " +
			"review as separate workflow calls, reading results between phases). A useful default " +
			"is hybrid: scout inline first to discover the work-list, then fan out over it.\n\n" +
			"Granularity: scope each subtask to a distinct concern, component, or question rather " +
			"than per line or per file section. Scale the count to what the user asked for: a " +
			"focused review of a module of a few hundred lines rarely needs more than about ten " +
			"subtasks; a broad audit of a large codebase can justify more.\n\n" +
			"Standing consent: while a system message confirms orchestration mode is on, that " +
			"opt-in is standing. Author and run a workflow for every substantive task by default, " +
			"and lean toward verifying findings adversarially. Work solo only on conversational " +
			"turns or trivial mechanical edits. When a system message says the mode is off, " +
			"revert to the opt-in rule above."),
		InputSchema: anthropic.ToolInputSchemaParam{
			Properties: map[string]any{
				"subtasks": map[string]any{
					"type":        "array",
					"items":       map[string]any{"type": "string"},
					"description": "Independent subtask prompts to run as parallel agents",
				},
			},
			Required: []string{"subtasks"},
		},
	},
}

var bashTool = anthropic.ToolUnionParam{
	OfBashTool20250124: &anthropic.ToolBash20250124Param{},
}

var reportTool = anthropic.ToolUnionParam{
	OfTool: &anthropic.ToolParam{
		Name: "report_findings",
		Description: anthropic.String("Report the final findings for your subtask. Call this exactly once, when you are " +
			"done investigating; it ends your task."),
		InputSchema: anthropic.ToolInputSchemaParam{
			Properties: map[string]any{
				"summary": map[string]any{"type": "string", "description": "Two or three sentences of synthesis"},
				"findings": map[string]any{
					"type": "array",
					"items": map[string]any{
						"type": "object",
						"properties": map[string]any{
							"claim": map[string]any{"type": "string", "description": "The finding, one sentence"},
							"evidence": map[string]any{
								"type":        "string",
								"description": "How it was verified (file, line, or command output)",
							},
							"severity": map[string]any{"type": "string", "enum": []string{"high", "medium", "low", "info"}},
						},
						"required": []string{"claim", "evidence", "severity"},
					},
				},
			},
			Required: []string{"summary", "findings"},
		},
	},
}

static final Tool WORKFLOW_TOOL = Tool.builder()
        .name("Workflow")
        .description("Orchestrate a multiagent workflow: split a large task into independent subtasks "
                + "and run them as parallel agents, then collect their results.\n\n"
                + "Opt-in: only use this tool when the user explicitly asks for a workflow, or when a "
                + "system message confirms that orchestration mode is on.\n\n"
                + "Quality patterns: adversarial verification (a second wave of agents checks the first "
                + "wave's findings against the source), a completeness critic (one agent hunts for what "
                + "the others missed), and multiphase sequencing (understand, design, implement, and "
                + "review as separate workflow calls, reading results between phases). A useful default "
                + "is hybrid: scout inline first to discover the work-list, then fan out over it.\n\n"
                + "Granularity: scope each subtask to a distinct concern, component, or question rather "
                + "than per line or per file section. Scale the count to what the user asked for: a "
                + "focused review of a module of a few hundred lines rarely needs more than about ten "
                + "subtasks; a broad audit of a large codebase can justify more.\n\n"
                + "Standing consent: while a system message confirms orchestration mode is on, that "
                + "opt-in is standing. Author and run a workflow for every substantive task by default, "
                + "and lean toward verifying findings adversarially. Work solo only on conversational "
                + "turns or trivial mechanical edits. When a system message says the mode is off, "
                + "revert to the opt-in rule above.")
        .inputSchema(Tool.InputSchema.builder()
                .properties(JsonValue.from(Map.of(
                        "subtasks", Map.of(
                                "type", "array",
                                "items", Map.of("type", "string"),
                                "description", "Independent subtask prompts to run as parallel agents"))))
                .putAdditionalProperty("required", JsonValue.from(List.of("subtasks")))
                .build())
        .build();

static final ToolBash20250124 BASH_TOOL = ToolBash20250124.builder().build();

static final Tool REPORT_TOOL = Tool.builder()
        .name("report_findings")
        .description("Report the final findings for your subtask. Call this exactly once, when you are "
                + "done investigating; it ends your task.")
        .inputSchema(Tool.InputSchema.builder()
                .properties(JsonValue.from(Map.of(
                        "summary", Map.of("type", "string", "description", "Two or three sentences of synthesis"),
                        "findings", Map.of(
                                "type", "array",
                                "items", Map.of(
                                        "type", "object",
                                        "properties", Map.of(
                                                "claim", Map.of(
                                                        "type", "string",
                                                        "description", "The finding, one sentence"),
                                                "evidence", Map.of(
                                                        "type", "string",
                                                        "description", "How it was verified (file, line, or command output)"),
                                                "severity", Map.of(
                                                        "type", "string",
                                                        "enum", List.of("high", "medium", "low", "info"))),
                                        "required", List.of("claim", "evidence", "severity"))))))
                .putAdditionalProperty("required", JsonValue.from(List.of("summary", "findings")))
                .build())
        .build();
const WORKFLOW_TOOL = [
    'name' => 'Workflow',
    'description' =>
        'Orchestrate a multiagent workflow: split a large task into independent subtasks '
        . "and run them as parallel agents, then collect their results.\n\n"
        . 'Opt-in: only use this tool when the user explicitly asks for a workflow, or when a '
        . "system message confirms that orchestration mode is on.\n\n"
        . 'Quality patterns: adversarial verification (a second wave of agents checks the first '
        . 'wave\'s findings against the source), a completeness critic (one agent hunts for what '
        . 'the others missed), and multiphase sequencing (understand, design, implement, and '
        . 'review as separate workflow calls, reading results between phases). A useful default '
        . "is hybrid: scout inline first to discover the work-list, then fan out over it.\n\n"
        . 'Granularity: scope each subtask to a distinct concern, component, or question rather '
        . 'than per line or per file section. Scale the count to what the user asked for: a '
        . 'focused review of a module of a few hundred lines rarely needs more than about ten '
        . "subtasks; a broad audit of a large codebase can justify more.\n\n"
        . 'Standing consent: while a system message confirms orchestration mode is on, that '
        . 'opt-in is standing. Author and run a workflow for every substantive task by default, '
        . 'and lean toward verifying findings adversarially. Work solo only on conversational '
        . 'turns or trivial mechanical edits. When a system message says the mode is off, '
        . 'revert to the opt-in rule above.',
    'input_schema' => [
        'type' => 'object',
        'properties' => [
            'subtasks' => [
                'type' => 'array',
                'items' => ['type' => 'string'],
                'description' => 'Independent subtask prompts to run as parallel agents',
            ],
        ],
        'required' => ['subtasks'],
    ],
];

const BASH_TOOL = ['type' => 'bash_20250124', 'name' => 'bash'];

const REPORT_TOOL = [
    'name' => 'report_findings',
    'description' =>
        'Report the final findings for your subtask. Call this exactly once, when you are '
        . 'done investigating; it ends your task.',
    'input_schema' => [
        'type' => 'object',
        'properties' => [
            'summary' => ['type' => 'string', 'description' => 'Two or three sentences of synthesis'],
            'findings' => [
                'type' => 'array',
                'items' => [
                    'type' => 'object',
                    'properties' => [
                        'claim' => ['type' => 'string', 'description' => 'The finding, one sentence'],
                        'evidence' => [
                            'type' => 'string',
                            'description' => 'How it was verified (file, line, or command output)',
                        ],
                        'severity' => ['type' => 'string', 'enum' => ['high', 'medium', 'low', 'info']],
                    ],
                    'required' => ['claim', 'evidence', 'severity'],
                ],
            ],
        ],
        'required' => ['summary', 'findings'],
    ],
];
WORKFLOW_TOOL = {
  name: "Workflow",
  description:
    "Orchestrate a multiagent workflow: split a large task into independent subtasks " \
    "and run them as parallel agents, then collect their results.\n\n" \
    "Opt-in: only use this tool when the user explicitly asks for a workflow, or when a " \
    "system message confirms that orchestration mode is on.\n\n" \
    "Quality patterns: adversarial verification (a second wave of agents checks the first " \
    "wave's findings against the source), a completeness critic (one agent hunts for what " \
    "the others missed), and multiphase sequencing (understand, design, implement, and " \
    "review as separate workflow calls, reading results between phases). A useful default " \
    "is hybrid: scout inline first to discover the work-list, then fan out over it.\n\n" \
    "Granularity: scope each subtask to a distinct concern, component, or question rather " \
    "than per line or per file section. Scale the count to what the user asked for: a " \
    "focused review of a module of a few hundred lines rarely needs more than about ten " \
    "subtasks; a broad audit of a large codebase can justify more.\n\n" \
    "Standing consent: while a system message confirms orchestration mode is on, that " \
    "opt-in is standing. Author and run a workflow for every substantive task by default, " \
    "and lean toward verifying findings adversarially. Work solo only on conversational " \
    "turns or trivial mechanical edits. When a system message says the mode is off, " \
    "revert to the opt-in rule above.",
  input_schema: {
    type: "object",
    properties: {
      subtasks: {
        type: "array",
        items: {type: "string"},
        description: "Independent subtask prompts to run as parallel agents"
      }
    },
    required: ["subtasks"]
  }
}.freeze

BASH_TOOL = {type: "bash_20250124", name: "bash"}.freeze

REPORT_TOOL = {
  name: "report_findings",
  description:
    "Report the final findings for your subtask. Call this exactly once, when you are " \
    "done investigating; it ends your task.",
  input_schema: {
    type: "object",
    properties: {
      summary: {type: "string", description: "Two or three sentences of synthesis"},
      findings: {
        type: "array",
        items: {
          type: "object",
          properties: {
            claim: {type: "string", description: "The finding, one sentence"},
            evidence: {
              type: "string",
              description: "How it was verified (file, line, or command output)"
            },
            severity: {type: "string", enum: ["high", "medium", "low", "info"]}
          },
          required: ["claim", "evidence", "severity"]
        }
      }
    },
    required: ["summary", "findings"]
  }
}.freeze

bash 도구를 로컬에서 실행하기

bash 핸들러는 요청된 명령을 타임아웃과 함께 실행하고, 결합된 stdout과 stderr를 캡처하며, 결과를 잘라서 폭주하는 명령이 컨텍스트 윈도우를 넘치게 하지 못하게 해요. 명령은 예제를 실행한 디렉터리에서 실행되므로, 프로젝트를 대상으로 하려면 거기서 시작해야 해요. DOC_TEST_MODE가 설정되면 하네스가 대신 bash에 버릴 수 있는 작은 임시 fixture 디렉터리를 주고, 종료 시 제거돼요. 여기에는 샌드박스가 없어요. 명령은 예제를 실행한 프로세스의 권한으로 실행돼요. 명확성을 위해 이 예제는 bash_20250124 계약이 설명하는 영구 세션을 유지하는 대신, 각 호출을 새 서브셸에서 실행해요. 프로덕션 에이전트는 도구를 장수명 셸로 뒷받침해서, 작업 디렉터리·환경·restart 액션이 문서화된 대로 동작하게 해야 해요.

```python Python # Run bash where the example was launched. In DOC_TEST_MODE the docs harness # points it at a throwaway fixture directory instead, removed on exit. if DOC_TEST_MODE: WORK_DIR = tempfile.mkdtemp(prefix="orchestration-") atexit.register(shutil.rmtree, WORK_DIR, ignore_errors=True) with open(os.path.join(WORK_DIR, "sample.py"), "w") as fixture: fixture.write( "def fib(n):\n" " return n if n < 2 else fib(n - 1) + fib(n - 2)\n\n" "print(fib(10))\n" ) else: WORK_DIR = os.getcwd()

def run_bash(command: str) -> tuple[str, bool]: """Run a shell command and return (output, is_error). No sandbox: example code only.""" print(f"[bash] {command}", file=sys.stderr) try: proc = subprocess.run( ["bash", "-c", command], cwd=WORK_DIR, capture_output=True, text=True, errors="replace", timeout=BASH_TIMEOUT_SECONDS, ) except subprocess.TimeoutExpired: return f"command timed out after {BASH_TIMEOUT_SECONDS}s", True output = (proc.stdout + proc.stderr).strip() or "(no output)" if len(output) > TOOL_RESULT_MAX_CHARS: output = output[:TOOL_RESULT_MAX_CHARS] + f"\n(truncated at {TOOL_RESULT_MAX_CHARS} chars)" if proc.returncode != 0: output = f"(exit code {proc.returncode})\n{output}" return output, proc.returncode != 0

def handle_bash_block(block) -> tuple[str, bool]: if block.input.get("restart") is True: return "Shell restarted.", False command = block.input.get("command") if not isinstance(command, str) or not command: return "bash error: no command was provided.", True return run_bash(command)


```typescript TypeScript
const execShell = promisify(exec);

// Run bash where the example was launched. In DOC_TEST_MODE the docs harness
// points it at a throwaway fixture directory instead, removed on exit.
const WORK_DIR = DOC_TEST_MODE
  ? await mkdtemp(join(tmpdir(), "orchestration-"))
  : process.cwd();
if (DOC_TEST_MODE) {
  await writeFile(
    join(WORK_DIR, "sample.py"),
    "def fib(n):\n" +
      "    return n if n < 2 else fib(n - 1) + fib(n - 2)\n\n" +
      "print(fib(10))\n",
  );
  process.on("exit", () => rmSync(WORK_DIR, { recursive: true, force: true }));
}

// Run a shell command and return its output. No sandbox: example code only.
async function runBash(command: string): Promise<{ output: string; isError: boolean }> {
  console.error(`[bash] ${command}`);
  let stdout = "";
  let stderr = "";
  let exitCode = 0;
  try {
    ({ stdout, stderr } = await execShell(command, {
      shell: "/bin/bash",
      cwd: WORK_DIR,
      timeout: BASH_TIMEOUT_SECONDS * 1000,
      maxBuffer: 16 * 1024 * 1024,
    }));
  } catch (error) {
    const failure = error as {
      stdout?: string;
      stderr?: string;
      code?: number | string;
      killed?: boolean;
    };
    if (failure.killed && failure.code !== "ERR_CHILD_PROCESS_STDIO_MAXBUFFER") {
      return { output: `command timed out after ${BASH_TIMEOUT_SECONDS}s`, isError: true };
    }
    stdout = failure.stdout ?? "";
    stderr = failure.stderr ?? "";
    exitCode = typeof failure.code === "number" ? failure.code : 1;
  }
  let output = (stdout + stderr).trim() || "(no output)";
  const codePoints = [...output];
  if (codePoints.length > TOOL_RESULT_MAX_CHARS) {
    output =
      codePoints.slice(0, TOOL_RESULT_MAX_CHARS).join("") +
      `\n(truncated at ${TOOL_RESULT_MAX_CHARS} chars)`;
  }
  if (exitCode !== 0) {
    output = `(exit code ${exitCode})\n${output}`;
  }
  return { output, isError: exitCode !== 0 };
}

async function handleBashBlock(
  block: Anthropic.ToolUseBlock,
): Promise<{ output: string; isError: boolean }> {
  const input = block.input as { command?: string; restart?: boolean };
  if (input.restart === true) {
    return { output: "Shell restarted.", isError: false };
  }
  if (!input.command) {
    return { output: "bash error: no command was provided.", isError: true };
  }
  return runBash(input.command);
}
// Run bash where the example was launched. In DOC_TEST_MODE the docs harness
// points it at a throwaway fixture directory instead, removed on exit.
var workDir = Environment.CurrentDirectory;
if (docTestMode)
{
    workDir = Directory.CreateTempSubdirectory("orchestration-").FullName;
    File.WriteAllText(Path.Combine(workDir, "sample.py"),
        "def fib(n):\n" +
        "    return n if n < 2 else fib(n - 1) + fib(n - 2)\n\n" +
        "print(fib(10))\n");
    var fixtureDir = workDir;
    AppDomain.CurrentDomain.ProcessExit += (_, _) =>
    {
        try { Directory.Delete(fixtureDir, recursive: true); }
        catch { /* Best-effort cleanup; the OS tmp sweeper handles leftovers. */ }
    };
}

// Run a shell command and return its output plus an error flag. No sandbox: example code only.
async Task<(string Output, bool IsError)> RunBash(string command)
{
    Console.Error.WriteLine($"[bash] {command}");
    using var process = Process.Start(new ProcessStartInfo("bash")
    {
        ArgumentList = { "-c", command },
        WorkingDirectory = workDir,
        RedirectStandardOutput = true,
        RedirectStandardError = true,
    });
    if (process is null)
    {
        return ("bash error: the shell process failed to start.", true);
    }
    var stdoutTask = process.StandardOutput.ReadToEndAsync();
    var stderrTask = process.StandardError.ReadToEndAsync();
    using var timeout = new CancellationTokenSource(TimeSpan.FromSeconds(bashTimeoutSeconds));
    try
    {
        await process.WaitForExitAsync(timeout.Token);
    }
    catch (OperationCanceledException)
    {
        process.Kill(entireProcessTree: true);
        // Let the reader tasks finish before the process is disposed.
        try
        {
            await Task.WhenAll(stdoutTask, stderrTask);
        }
        catch
        {
            // The output is discarded on timeout, so reader failures are ignored too.
        }
        return ($"command timed out after {bashTimeoutSeconds}s", true);
    }
    var output = (await stdoutTask + await stderrTask).Trim();
    if (output.Length == 0)
    {
        output = "(no output)";
    }
    if (output.Length > toolResultMaxChars)
    {
        output = output[..toolResultMaxChars] + $"\n(truncated at {toolResultMaxChars} chars)";
    }
    if (process.ExitCode != 0)
    {
        output = $"(exit code {process.ExitCode})\n{output}";
    }
    return (output, process.ExitCode != 0);
}

// Execute one bash tool call requested by the model.
async Task<(string Output, bool IsError)> HandleBashBlock(ToolUseBlock block)
{
    if (block.Input.TryGetValue("restart", out var restart) && restart.ValueKind == JsonValueKind.True)
    {
        return ("Shell restarted.", false);
    }
    var command = block.Input.TryGetValue("command", out var rawCommand) && rawCommand.ValueKind == JsonValueKind.String
        ? rawCommand.GetString()!
        : "";
    if (command.Length == 0)
    {
        return ("bash error: no command was provided.", true);
    }
    return await RunBash(command);
}
// Run bash where the example was launched. In DOC_TEST_MODE the docs harness
// points it at a throwaway fixture directory instead, removed on exit.
var workDir = func() string {
	if !docTestMode {
		dir, err := os.Getwd()
		if err != nil {
			log.Fatal(err)
		}
		return dir
	}
	dir, err := os.MkdirTemp("", "orchestration-")
	if err != nil {
		log.Fatal(err)
	}
	fixture := "def fib(n):\n" +
		"    return n if n < 2 else fib(n - 1) + fib(n - 2)\n\n" +
		"print(fib(10))\n"
	if err := os.WriteFile(filepath.Join(dir, "sample.py"), []byte(fixture), 0o644); err != nil {
		log.Fatal(err)
	}
	return dir
}()

// runBash runs a shell command and returns its output plus an error flag.
// No sandbox: example code only.
func runBash(ctx context.Context, command string) (string, bool) {
	fmt.Fprintf(os.Stderr, "[bash] %s\n", command)
	ctx, cancel := context.WithTimeout(ctx, bashTimeoutSeconds*time.Second)
	defer cancel()
	cmd := exec.CommandContext(ctx, "bash", "-c", command)
	cmd.Dir = workDir
	combined, err := cmd.CombinedOutput()
	if errors.Is(ctx.Err(), context.DeadlineExceeded) {
		return fmt.Sprintf("command timed out after %ds", bashTimeoutSeconds), true
	}
	output := strings.TrimSpace(string(combined))
	if output == "" {
		output = "(no output)"
	}
	if runes := []rune(output); len(runes) > toolResultMaxChars {
		output = string(runes[:toolResultMaxChars]) + fmt.Sprintf("\n(truncated at %d chars)", toolResultMaxChars)
	}
	if err == nil {
		return output, false
	}
	var exitErr *exec.ExitError
	if errors.As(err, &exitErr) {
		return fmt.Sprintf("(exit code %d)\n%s", exitErr.ExitCode(), output), true
	}
	return fmt.Sprintf("(%s)\n%s", err, output), true
}

// handleBashBlock executes one bash tool call requested by the model.
func handleBashBlock(ctx context.Context, block anthropic.ToolUseBlock) (string, bool) {
	var input struct {
		Command string `json:"command"`
		Restart bool   `json:"restart"`
	}
	if err := json.Unmarshal(block.Input, &input); err != nil {
		return fmt.Sprintf("bash error: could not parse input: %s", err), true
	}
	if input.Restart {
		return "Shell restarted.", false
	}
	if input.Command == "" {
		return "bash error: no command was provided.", true
	}
	return runBash(ctx, input.Command)
}

record ToolOutput(String output, boolean isError) {}

// Run bash where the example was launched. In DOC_TEST_MODE the docs harness
// points it at a throwaway fixture directory instead, removed on exit.
static final Path WORK_DIR = createWorkDir();

static Path createWorkDir() {
    if (!DOC_TEST_MODE) {
        return Path.of(System.getProperty("user.dir"));
    }
    try {
        var dir = Files.createTempDirectory("orchestration-");
        Files.writeString(dir.resolve("sample.py"), """
                def fib(n):
                    return n if n < 2 else fib(n - 1) + fib(n - 2)

                print(fib(10))
                """);
        Runtime.getRuntime().addShutdownHook(new Thread(() -> {
            try (var paths = Files.walk(dir)) {
                paths.sorted(Comparator.reverseOrder()).forEach(p -> {
                    try { Files.deleteIfExists(p); } catch (IOException ignored) {}
                });
            } catch (IOException ignored) {
                // Best-effort cleanup; the OS tmp sweeper handles leftovers.
            }
        }));
        return dir;
    } catch (IOException error) {
        throw new UncheckedIOException(error);
    }
}

// Run a shell command and return its output plus an error flag. No sandbox: example code only.
ToolOutput runBash(String command) throws InterruptedException {
    System.err.println("[bash] " + command);
    Process process;
    try {
        process = new ProcessBuilder("bash", "-c", command)
                .directory(WORK_DIR.toFile())
                .redirectErrorStream(true)
                .start();
    } catch (IOException error) {
        return new ToolOutput("(" + error + ")", true);
    }
    // Drain stdout on another thread so a filled pipe cannot stall the timeout wait below.
    CompletableFuture<String> outputReader = CompletableFuture.supplyAsync(() -> {
        try (var stdout = process.getInputStream()) {
            return new String(stdout.readAllBytes(), StandardCharsets.UTF_8);
        } catch (IOException error) {
            return "";
        }
    });
    if (!process.waitFor(BASH_TIMEOUT_SECONDS, TimeUnit.SECONDS)) {
        process.destroyForcibly();
        outputReader.cancel(true);
        return new ToolOutput("command timed out after " + BASH_TIMEOUT_SECONDS + "s", true);
    }
    String output = outputReader.join().trim();
    if (output.isEmpty()) {
        output = "(no output)";
    }
    if (output.length() > TOOL_RESULT_MAX_CHARS) {
        output = output.substring(0, TOOL_RESULT_MAX_CHARS)
                + "\n(truncated at " + TOOL_RESULT_MAX_CHARS + " chars)";
    }
    int exitCode = process.exitValue();
    if (exitCode != 0) {
        return new ToolOutput("(exit code " + exitCode + ")\n" + output, true);
    }
    return new ToolOutput(output, false);
}

// Execute one bash tool call requested by the model.
ToolOutput handleBashBlock(ToolUseBlock block) throws InterruptedException {
    Map<String, JsonValue> input = (Map<String, JsonValue>) block._input().asObject().orElse(Map.of());
    JsonValue restart = input.getOrDefault("restart", JsonValue.from(false));
    if (Boolean.TRUE.equals(restart.asBoolean().orElse(false))) {
        return new ToolOutput("Shell restarted.", false);
    }
    JsonValue raw = input.get("command");
    String command = raw != null && raw.asString().isPresent() ? raw.asStringOrThrow() : "";
    if (command.isEmpty()) {
        return new ToolOutput("bash error: no command was provided.", true);
    }
    return runBash(command);
}
// Run bash where the example was launched. In DOC_TEST_MODE the docs harness
// points it at a throwaway fixture directory instead, removed on exit.
if (DOC_TEST_MODE) {
    $workDir = sys_get_temp_dir() . '/orchestration-' . bin2hex(random_bytes(8));
    if (!mkdir($workDir, 0700)) {
        throw new RuntimeException("could not create working directory {$workDir}");
    }
    file_put_contents(
        $workDir . '/sample.py',
        "def fib(n):\n"
        . "    return n if n < 2 else fib(n - 1) + fib(n - 2)\n\n"
        . "print(fib(10))\n",
    );
    register_shutdown_function(function () use ($workDir): void {
        foreach (glob($workDir . '/*') ?: [] as $entry) {
            @unlink($entry);
        }
        @rmdir($workDir);
    });
} else {
    $workDir = getcwd() ?: '.';
}
define('WORK_DIR', $workDir);

/**
 * Run a shell command and return [output, isError]. The coreutils timeout command
 * enforces the time limit. No sandbox: example code only.
 */
function runBash(string $command): array
{
    fwrite(STDERR, "[bash] {$command}\n");
    // Requires GNU coreutils 'timeout'. On macOS: brew install coreutils, or replace with gtimeout.
    exec(
        'cd ' . escapeshellarg(WORK_DIR) . ' && timeout ' . BASH_TIMEOUT_SECONDS
            . ' bash -c ' . escapeshellarg($command) . ' 2>&1',
        $outputLines,
        $exitCode,
    );
    if ($exitCode === 124) {
        return ['command timed out after ' . BASH_TIMEOUT_SECONDS . 's', true];
    }
    $output = trim(implode("\n", $outputLines));
    if ($output === '') {
        $output = '(no output)';
    }
    if (mb_strlen($output) > TOOL_RESULT_MAX_CHARS) {
        $output = mb_substr($output, 0, TOOL_RESULT_MAX_CHARS)
            . "\n(truncated at " . TOOL_RESULT_MAX_CHARS . ' chars)';
    }
    if ($exitCode !== 0) {
        $output = "(exit code {$exitCode})\n{$output}";
    }
    return [$output, $exitCode !== 0];
}

/** Execute one bash tool call requested by the model. */
function handleBashBlock(ToolUseBlock $block): array
{
    if (($block->input['restart'] ?? null) === true) {
        return ['Shell restarted.', false];
    }
    $command = $block->input['command'] ?? '';
    if (!is_string($command) || $command === '') {
        return ['bash error: no command was provided.', true];
    }
    return runBash($command);
}
# Run bash where the example was launched. In DOC_TEST_MODE the docs harness
# points it at a throwaway fixture directory instead, removed on exit.
WORK_DIR =
  if DOC_TEST_MODE
    Dir.mktmpdir("orchestration-").tap do |dir|
      File.write(File.join(dir, "sample.py"), <<~PYTHON)
        def fib(n):
            return n if n < 2 else fib(n - 1) + fib(n - 2)

        print(fib(10))
      PYTHON
      at_exit { FileUtils.remove_entry(dir, true) }
    end
  else
    Dir.pwd
  end

# Tool input arrives as a Hash or as a raw JSON string from the streaming
# accumulator; normalize either shape to a string-keyed Hash.
def parse_tool_input(raw)
  return raw.transform_keys(&:to_s) if raw.is_a?(Hash)
  parsed = JSON.parse(raw.to_s) rescue nil
  parsed.is_a?(Hash) ? parsed : {}
end

# Run a shell command and return [output, is_error]. No sandbox: example code only.
def run_bash(command)
  warn "[bash] #{command}"
  begin
    stdin, stdout_and_stderr, wait_thr = Open3.popen2e("bash", "-c", command, pgroup: true, chdir: WORK_DIR)
    stdin.close
    reader = Thread.new { stdout_and_stderr.read.scrub }
    # Enforce the time limit with a monotonic-clock deadline so a timed-out command is
    # terminated rather than left running in the background.
    deadline = Process.clock_gettime(Process::CLOCK_MONOTONIC) + BASH_TIMEOUT_SECONDS
    until wait_thr.join(0.1)
      next if Process.clock_gettime(Process::CLOCK_MONOTONIC) < deadline

      begin
        Process.kill("-TERM", wait_thr.pid)
      rescue Errno::ESRCH
      end
      unless wait_thr.join(2)
        begin
          Process.kill("-KILL", wait_thr.pid)
        rescue Errno::ESRCH
        end
      end
      wait_thr.join(5)
      reader.join(1) || reader.kill
      stdout_and_stderr.close rescue nil
      return ["command timed out after #{BASH_TIMEOUT_SECONDS}s", true]
    end
    status = wait_thr.value
    output = reader.value.strip
    stdout_and_stderr.close
    output = "(no output)" if output.empty?
    if output.length > TOOL_RESULT_MAX_CHARS
      output = "#{output[0, TOOL_RESULT_MAX_CHARS]}\n(truncated at #{TOOL_RESULT_MAX_CHARS} chars)"
    end
    output = "(exit code #{status.exitstatus})\n#{output}" unless status.success?
    [output, !status.success?]
  rescue Errno::ENOENT => e
    return ["bash error: #{e.message}", true]
  end
end

# Execute one bash tool call requested by the model.
def handle_bash_block(block)
  input = parse_tool_input(block.input)
  return ["Shell restarted.", false] if input["restart"] == true

  command = input["command"]
  return ["bash error: no command was provided.", true] unless command.is_a?(String) && !command.empty?

  run_bash(command)
end

# Convert response content to request-shaped params. The streaming accumulator
# returns tool_use input as a raw JSON string and includes response-only fields,
# so reshape each block to the request schema before echoing it back.
def assistant_content_param(content)
  content.map do |block|
    case block
    when Anthropic::Models::ToolUseBlock
      input = parse_tool_input(block.input)
      {type: "tool_use", id: block.id, name: block.name, input: input}
    when Anthropic::Models::TextBlock
      {type: "text", text: block.text}
    when Anthropic::Models::ThinkingBlock
      {type: "thinking", thinking: block.thinking, signature: block.signature}
    when Anthropic::Models::RedactedThinkingBlock then {type: "redacted_thinking", data: block.data}
    else
      block.to_h
    end
  end
end

서브에이전트 한 개 실행하기

각 워크플로 서브태스크는 bash 도구를 가진 작은 에이전트 루프가 되어, 메인 루프와 같은 노력 수준으로 실행돼요. 요청별 타임아웃이 각 API 호출을 제한해서, 연결이 끊어져도 전체 실행이 멈추는 대신 서브에이전트 하나만 나빠져요.

```python Python def run_subagent(model: str, prompt: str) -> str: """One subagent: a small nested agent loop with the bash tool plus report_findings. Subagents inherit the main loop's effort level.""" subagent_system = ( "You are one agent in a larger parallel fan-out, assigned a single subtask. " "Investigate it directly, using bash to check facts rather than guessing, and finish " "by calling report_findings exactly once. Return findings, not narration." ) messages = [{"role": "user", "content": prompt}] for _ in range(MAX_SUBAGENT_TURNS): with client.messages.stream( model=model, max_tokens=64000, system=subagent_system, output_config={"effort": EFFORT}, tools=[BASH_TOOL, REPORT_TOOL], messages=messages, timeout=REQUEST_TIMEOUT_SECONDS, ) as stream: response = stream.get_final_message() messages.append({"role": "assistant", "content": response.content}) if response.stop_reason == "pause_turn": continue if response.stop_reason != "tool_use": text = "".join(block.text for block in response.content if block.type == "text") if response.stop_reason == "max_tokens": text += "\n\n(warning: subagent response was truncated at max_tokens)" return text tool_results = [] report = None for block in response.content: if block.type != "tool_use": continue match block.name: case "report_findings": report = json.dumps(block.input, indent=2) output, is_error = "Findings recorded.", False case "bash": output, is_error = handle_bash_block(block) case _: output, is_error = f"unknown tool: {block.name}", True tool_results.append( { "type": "tool_result", "tool_use_id": block.id, "content": output, "is_error": is_error, } ) if report is not None: return report messages.append({"role": "user", "content": tool_results}) return "(subagent hit the turn limit before finishing)" ```
// One subagent: a small nested agent loop with the bash tool plus report_findings.
// Subagents inherit the main loop's effort level.
async function runSubagent(model: string, prompt: string): Promise<string> {
  const subagentSystem =
    "You are one agent in a larger parallel fan-out, assigned a single subtask. " +
    "Investigate it directly, using bash to check facts rather than guessing, and finish " +
    "by calling report_findings exactly once. Return findings, not narration.";
  const messages: Anthropic.MessageParam[] = [{ role: "user", content: prompt }];
  for (let turn = 0; turn < MAX_SUBAGENT_TURNS; turn++) {
    const response = await client.messages
      .stream(
        {
          model,
          max_tokens: 64000,
          system: subagentSystem,
          output_config: { effort: EFFORT },
          tools: [BASH_TOOL, REPORT_TOOL],
          messages,
        },
        { signal: AbortSignal.timeout(REQUEST_TIMEOUT_SECONDS * 1000) },
      )
      .finalMessage();
    messages.push({ role: "assistant", content: response.content });
    if (response.stop_reason === "pause_turn") {
      continue;
    }
    if (response.stop_reason !== "tool_use") {
      let text = response.content
        .filter((block): block is Anthropic.TextBlock => block.type === "text")
        .map((block) => block.text)
        .join("");
      if (response.stop_reason === "max_tokens") {
        text += "\n\n(warning: subagent response was truncated at max_tokens)";
      }
      return text;
    }
    const toolResults: Anthropic.ToolResultBlockParam[] = [];
    let report: string | null = null;
    for (const block of response.content) {
      if (block.type !== "tool_use") {
        continue;
      }
      let output: string;
      let isError: boolean;
      switch (block.name) {
        case "report_findings":
          report = JSON.stringify(block.input, null, 2);
          output = "Findings recorded.";
          isError = false;
          break;
        case "bash":
          ({ output, isError } = await handleBashBlock(block));
          break;
        default:
          output = `unknown tool: ${block.name}`;
          isError = true;
      }
      toolResults.push({
        type: "tool_result",
        tool_use_id: block.id,
        content: output,
        is_error: isError,
      });
    }
    if (report !== null) {
      return report;
    }
    messages.push({ role: "user", content: toolResults });
  }
  return "(subagent hit the turn limit before finishing)";
}
// One subagent: a small nested agent loop with the bash tool plus report_findings.
// Subagents inherit the main loop's effort level.
async Task<string> RunSubagent(string prompt)
{
    const string subagentSystem =
        "You are one agent in a larger parallel fan-out, assigned a single subtask. "
        + "Investigate it directly, using bash to check facts rather than guessing, and finish "
        + "by calling report_findings exactly once. Return findings, not narration.";
    List<MessageParam> messages = [new() { Role = Role.User, Content = prompt }];
    for (var turn = 0; turn < maxSubagentTurns; turn++)
    {
        using var deadline = new CancellationTokenSource(TimeSpan.FromSeconds(requestTimeoutSeconds));
        var response = await client.Messages.Create(new MessageCreateParams
        {
            Model = model,
            MaxTokens = requestMaxTokens,
            System = subagentSystem,
            OutputConfig = new OutputConfig { Effort = effort },
            Tools = [bashTool, reportTool],
            Messages = messages,
        }, cancellationToken: deadline.Token);
        messages.Add(new()
        {
            Role = Role.Assistant,
            Content = response.Content.Select(block => new ContentBlockParam(block.Json)).ToList(),
        });
        if (response.StopReason == StopReason.PauseTurn)
        {
            continue;
        }
        if (response.StopReason != StopReason.ToolUse)
        {
            var text = string.Concat(
                response.Content.Select(block => block.TryPickText(out var textBlock) ? textBlock.Text : ""));
            if (response.StopReason == StopReason.MaxTokens)
            {
                text += "\n\n(warning: subagent response was truncated at max_tokens)";
            }
            return text;
        }
        List<ContentBlockParam> toolResults = [];
        string? report = null;
        foreach (var block in response.Content)
        {
            if (!block.TryPickToolUse(out var toolUse))
            {
                continue;
            }
            string output;
            bool isError;
            if (toolUse.Name == "report_findings")
            {
                report = JsonSerializer.Serialize(
                    toolUse.Input, new JsonSerializerOptions { WriteIndented = true });
                output = "Findings recorded.";
                isError = false;
            }
            else if (toolUse.Name == "bash")
            {
                (output, isError) = await HandleBashBlock(toolUse);
            }
            else
            {
                output = $"unknown tool: {toolUse.Name}";
                isError = true;
            }
            toolResults.Add(new ToolResultBlockParam(toolUse.ID) { Content = output, IsError = isError });
        }
        if (report is not null)
        {
            return report;
        }
        messages.Add(new() { Role = Role.User, Content = toolResults });
    }
    return "(subagent hit the turn limit before finishing)";
}
// runSubagent runs one subagent: a small nested agent loop with the bash tool plus
// report_findings. Subagents inherit the main loop's effort level.
func runSubagent(ctx context.Context, model string, prompt string) (string, error) {
	subagentSystem := "You are one agent in a larger parallel fan-out, assigned a single subtask. " +
		"Investigate it directly, using bash to check facts rather than guessing, and finish " +
		"by calling report_findings exactly once. Return findings, not narration."
	messages := []anthropic.MessageParam{anthropic.NewUserMessage(anthropic.NewTextBlock(prompt))}
	for range maxSubagentTurns {
		var response anthropic.Message
		err := func() error {
			ctx, cancel := context.WithTimeout(ctx, requestTimeoutSeconds*time.Second)
			defer cancel()
			stream := client.Messages.NewStreaming(ctx, anthropic.MessageNewParams{
				Model:        model,
				MaxTokens:    64000,
				System:       []anthropic.TextBlockParam{{Text: subagentSystem}},
				OutputConfig: anthropic.OutputConfigParam{Effort: effort},
				Tools:        []anthropic.ToolUnionParam{bashTool, reportTool},
				Messages:     messages,
			})
			defer stream.Close()
			for stream.Next() {
				if err := response.Accumulate(stream.Current()); err != nil {
					return err
				}
			}
			return stream.Err()
		}()
		if err != nil {
			return "", err
		}
		messages = append(messages, response.ToParam())
		if response.StopReason == anthropic.StopReasonPauseTurn {
			continue
		}
		if response.StopReason != anthropic.StopReasonToolUse {
			var text strings.Builder
			for _, block := range response.Content {
				if textBlock, ok := block.AsAny().(anthropic.TextBlock); ok {
					text.WriteString(textBlock.Text)
				}
			}
			if response.StopReason == anthropic.StopReasonMaxTokens {
				text.WriteString("\n\n(warning: subagent response was truncated at max_tokens)")
			}
			return text.String(), nil
		}
		var toolResults []anthropic.ContentBlockParamUnion
		var report string
		var reportRecorded bool
		for _, block := range response.Content {
			toolUse, ok := block.AsAny().(anthropic.ToolUseBlock)
			if !ok {
				continue
			}
			var output string
			var isError bool
			switch toolUse.Name {
			case "report_findings":
				report = string(toolUse.Input)
				var pretty bytes.Buffer
				if err := json.Indent(&pretty, toolUse.Input, "", "  "); err == nil {
					report = pretty.String()
				}
				reportRecorded = true
				output = "Findings recorded."
			case "bash":
				output, isError = handleBashBlock(ctx, toolUse)
			default:
				output, isError = fmt.Sprintf("unknown tool: %s", toolUse.Name), true
			}
			toolResults = append(toolResults, anthropic.NewToolResultBlock(toolUse.ID, output, isError))
		}
		if reportRecorded {
			return report, nil
		}
		messages = append(messages, anthropic.NewUserMessage(toolResults...))
	}
	return "(subagent hit the turn limit before finishing)", nil
}

// One subagent: a small nested agent loop with the bash tool plus report_findings.
// Subagents inherit the main loop's effort level.
String runSubagent(Model model, String prompt) throws InterruptedException {
    String subagentSystem = "You are one agent in a larger parallel fan-out, assigned a single subtask. "
            + "Investigate it directly, using bash to check facts rather than guessing, and finish "
            + "by calling report_findings exactly once. Return findings, not narration.";
    List<MessageParam> messages = new ArrayList<>();
    messages.add(MessageParam.builder().role(MessageParam.Role.USER).content(prompt).build());
    for (int turn = 0; turn < MAX_SUBAGENT_TURNS; turn++) {
        MessageCreateParams params = MessageCreateParams.builder()
                .model(model)
                .maxTokens(64000L)
                .system(subagentSystem)
                .outputConfig(OutputConfig.builder().effort(EFFORT).build())
                .addTool(BASH_TOOL)
                .addTool(REPORT_TOOL)
                .messages(messages)
                .build();
        MessageAccumulator accumulator = MessageAccumulator.create();
        try (var stream = client.messages().createStreaming(params, REQUEST_OPTIONS)) {
            stream.stream().forEach(accumulator::accumulate);
        }
        Message response = accumulator.message();
        messages.add(response.toParam());
        StopReason stopReason = response.stopReason().orElse(null);
        if (StopReason.PAUSE_TURN.equals(stopReason)) {
            continue;
        }
        if (!StopReason.TOOL_USE.equals(stopReason)) {
            String text = response.content().stream()
                    .flatMap(block -> block.text().stream())
                    .map(TextBlock::text)
                    .collect(Collectors.joining());
            if (StopReason.MAX_TOKENS.equals(stopReason)) {
                text += "\n\n(warning: subagent response was truncated at max_tokens)";
            }
            return text;
        }
        List<ContentBlockParam> toolResults = new ArrayList<>();
        String report = null;
        for (ContentBlock block : response.content()) {
            if (block.toolUse().isEmpty()) {
                continue;
            }
            ToolUseBlock toolUse = block.toolUse().get();
            ToolOutput result;
            if (toolUse.name().equals("report_findings")) {
                report = toolUse._input().convert(JsonNode.class).toPrettyString();
                result = new ToolOutput("Findings recorded.", false);
            } else if (toolUse.name().equals("bash")) {
                result = handleBashBlock(toolUse);
            } else {
                result = new ToolOutput("unknown tool: " + toolUse.name(), true);
            }
            toolResults.add(ContentBlockParam.ofToolResult(ToolResultBlockParam.builder()
                    .toolUseId(toolUse.id())
                    .content(result.output())
                    .isError(result.isError())
                    .build()));
        }
        if (report != null) {
            return report;
        }
        messages.add(MessageParam.builder()
                .role(MessageParam.Role.USER)
                .contentOfBlockParams(toolResults)
                .build());
    }
    return "(subagent hit the turn limit before finishing)";
}
/**
 * Consume a message stream and assemble the final assistant turn from its events:
 * the full content-block list plus the stop reason, equivalent to what a
 * non-streaming create call returns.
 */
function drainMessageStream(iterable $events): array
{
    $stringValue = fn ($value) => $value instanceof BackedEnum ? $value->value : $value;
    $blocks = [];
    $jsonBuffers = [];
    $stopReason = null;
    foreach ($events as $event) {
        $type = $stringValue($event->type);
        if ($type === 'content_block_start') {
            $blocks[$event->index] = $event->contentBlock;
            $jsonBuffers[$event->index] = '';
        } elseif ($type === 'content_block_delta') {
            $block = $blocks[$event->index];
            $delta = $event->delta;
            $deltaType = $stringValue($delta->type);
            if ($deltaType === 'text_delta') {
                $blocks[$event->index] = $block->withText($block->text . $delta->text);
            } elseif ($deltaType === 'input_json_delta') {
                $jsonBuffers[$event->index] .= $delta->partialJSON;
            } elseif ($deltaType === 'thinking_delta') {
                $blocks[$event->index] = $block->withThinking($block->thinking . $delta->thinking);
            } elseif ($deltaType === 'signature_delta') {
                $blocks[$event->index] = $block->withSignature($delta->signature);
            }
        } elseif ($type === 'message_delta') {
            $stopReason = $stringValue($event->delta->stopReason);
        }
    }
    foreach ($jsonBuffers as $index => $buffer) {
        if ($buffer !== '' && $blocks[$index] instanceof \Anthropic\Messages\ToolUseBlock) {
            $decoded = json_decode($buffer, true);
            $blocks[$index] = $blocks[$index]->withInput(is_array($decoded) ? $decoded : []);
        }
    }
    return [array_values($blocks), $stopReason];
}

/**
 * One subagent: a small nested agent loop with the bash tool plus report_findings.
 * Subagents inherit the main loop's effort level.
 */
function runSubagent(Client $client, string $model, string $prompt): string
{
    $subagentSystem =
        'You are one agent in a larger parallel fan-out, assigned a single subtask. '
        . 'Investigate it directly, using bash to check facts rather than guessing, and finish '
        . 'by calling report_findings exactly once. Return findings, not narration.';
    $messages = [['role' => 'user', 'content' => $prompt]];
    for ($turn = 0; $turn < MAX_SUBAGENT_TURNS; $turn++) {
        $stream = $client->messages->createStream(
            model: $model,
            maxTokens: 64000,
            system: $subagentSystem,
            outputConfig: ['effort' => EFFORT],
            tools: [BASH_TOOL, REPORT_TOOL],
            messages: $messages,
            requestOptions: ['timeout' => REQUEST_TIMEOUT_SECONDS],
        );
        [$content, $stopReason] = drainMessageStream($stream);
        $messages[] = ['role' => 'assistant', 'content' => $content];
        if ($stopReason === 'pause_turn') {
            continue;
        }
        if ($stopReason !== 'tool_use') {
            $text = '';
            foreach ($content as $block) {
                if ($block instanceof \Anthropic\Messages\TextBlock) {
                    $text .= $block->text;
                }
            }
            if ($stopReason === 'max_tokens') {
                $text .= "\n\n(warning: subagent response was truncated at max_tokens)";
            }
            return $text;
        }
        $report = null;
        $toolResults = [];
        foreach ($content as $block) {
            if (!$block instanceof \Anthropic\Messages\ToolUseBlock) {
                continue;
            }
            if ($block->name === 'report_findings') {
                $report = json_encode($block->input, JSON_PRETTY_PRINT);
                $output = 'Findings recorded.';
                $isError = false;
            } elseif ($block->name === 'bash') {
                [$output, $isError] = handleBashBlock($block);
            } else {
                $output = "unknown tool: {$block->name}";
                $isError = true;
            }
            $toolResults[] = [
                'type' => 'tool_result',
                'tool_use_id' => $block->id,
                'content' => $output,
                'is_error' => $isError,
            ];
        }
        if ($report !== null) {
            return $report;
        }
        $messages[] = ['role' => 'user', 'content' => $toolResults];
    }
    return '(subagent hit the turn limit before finishing)';
}
# One subagent: a small nested agent loop with the bash tool plus report_findings.
# Subagents inherit the main loop's effort level.
def run_subagent(model, prompt)
  subagent_system =
    "You are one agent in a larger parallel fan-out, assigned a single subtask. " \
    "Investigate it directly, using bash to check facts rather than guessing, and finish " \
    "by calling report_findings exactly once. Return findings, not narration."
  messages = [{role: "user", content: prompt}]
  MAX_SUBAGENT_TURNS.times do
    stream = CLIENT.messages.stream(
      model: model,
      max_tokens: 64_000,
      system_: subagent_system,
      output_config: {effort: EFFORT},
      tools: [BASH_TOOL, REPORT_TOOL],
      messages: messages,
      request_options: {timeout: REQUEST_TIMEOUT_SECONDS}
    )
    response = stream.accumulated_message
    messages << {role: "assistant", content: assistant_content_param(response.content)}
    next if response.stop_reason == :pause_turn

    unless response.stop_reason == :tool_use
      text = response.content.select { |block| block.type == :text }.map(&:text).join
      text += "\n\n(warning: subagent response was truncated at max_tokens)" if response.stop_reason == :max_tokens
      return text
    end

    report = nil
    tool_results = []
    response.content.each do |block|
      next unless block.type == :tool_use

      input = parse_tool_input(block.input)
      case block.name
      when "report_findings"
        report = JSON.pretty_generate(input)
        output, is_error = "Findings recorded.", false
      when "bash"
        output, is_error = handle_bash_block(block)
      else
        output, is_error = "unknown tool: #{block.name}", true
      end
      tool_results << {
        type: "tool_result",
        tool_use_id: block.id,
        content: output,
        is_error: is_error
      }
    end
    return report unless report.nil?

    messages << {role: "user", content: tool_results}
  end
  "(subagent hit the turn limit before finishing)"
end

결과를 저널링해서 재실행 시 이어가기

수십 개의 서브에이전트를 생성하는 팬아웃은 처음부터 다시 시작하기엔 비싸요. 작은 콘텐츠 주소 지정 저널(content-addressed journal)이 이것을 멱등하게 만들어줘요. 서브에이전트를 파견하기 전에, 그 프롬프트의 SHA-256을 로컬 JSON 파일에서 찾아보고 있으면 기록된 결과를 반환해요. 실행을 중단하고 다시 실행하면, 끝나지 못한 서브태스크만 다시 계산돼요. 저널은 단일 팬아웃 웨이브 내가 아니라 실행 간에 중복을 제거해요. 새로 시작하려면 저널 파일을 삭제하면 돼요.

```python Python _journal_lock = threading.Lock()

def _load_journal() -> dict: try: with open(JOURNAL_PATH) as file: return json.load(file) or {} except (OSError, json.JSONDecodeError): return {}

def journaled(prompt: str, compute) -> str: """Return a cached result for this exact prompt, or compute and persist it. This makes the fan-out resumable: interrupt the run, rerun it, and only the subtasks that never finished are recomputed. Delete the journal file to start fresh.""" key = hashlib.sha256(prompt.encode()).hexdigest() cached = _load_journal().get(key) if cached is not None: print(f"[journal] cache hit for {key[:12]}", file=sys.stderr) return cached result = compute() try: with _journal_lock: # fan-out writes from many threads journal = _load_journal() journal[key] = result temp = f"{JOURNAL_PATH}.tmp" with open(temp, "w") as file: json.dump(journal, file) os.replace(temp, JOURNAL_PATH) # atomic on POSIX and Windows except OSError as error: # the journal is best-effort; never discard a computed result print(f"[journal] write failed: {error}", file=sys.stderr) return result


```typescript TypeScript
let journalWriteChain = Promise.resolve();

async function loadJournal(): Promise<Record<string, string>> {
  try {
    return JSON.parse(await readFile(JOURNAL_PATH, "utf8")) ?? {};
  } catch (error) {
    if ((error as NodeJS.ErrnoException).code !== "ENOENT") {
      console.error(`[journal] discarding unreadable journal: ${error}`);
    }
    return {};
  }
}

// Return a cached result for this exact prompt, or compute and persist it. This
// makes the fan-out resumable: interrupt the run, rerun it, and only the subtasks
// that never finished are recomputed. Delete the journal file to start fresh.
async function journaled(prompt: string, compute: () => Promise<string>): Promise<string> {
  const key = createHash("sha256").update(prompt).digest("hex");
  const cached = (await loadJournal())[key];
  if (cached !== undefined) {
    console.error(`[journal] cache hit for ${key.slice(0, 12)}`);
    return cached;
  }
  const result = await compute();
  // Chain writes so concurrent subagents do not clobber each other's entries.
  // The chain is kept settled so one failed write does not poison later ones.
  await (journalWriteChain = journalWriteChain
    .then(async () => {
      const journal = await loadJournal();
      journal[key] = result;
      const temp = `${JOURNAL_PATH}.tmp`;
      await writeFile(temp, JSON.stringify(journal));
      await rename(temp, JOURNAL_PATH);
    })
    .catch((error) => console.error(`[journal] write failed: ${error}`)));
  return result;
}
SemaphoreSlim journalLock = new(1, 1);

async Task<Dictionary<string, string>> LoadJournal()
{
    try
    {
        return JsonSerializer.Deserialize<Dictionary<string, string>>(await File.ReadAllTextAsync(journalPath)) ?? [];
    }
    catch (Exception error) when (error is IOException or UnauthorizedAccessException or JsonException)
    {
        return [];
    }
}

// Return a cached result for this exact prompt, or compute and persist it. This
// makes the fan-out resumable: interrupt the run, rerun it, and only the subtasks
// that never finished are recomputed. Delete the journal file to start fresh.
async Task<string> Journaled(string prompt, Func<Task<string>> compute)
{
    var key = Convert.ToHexString(SHA256.HashData(Encoding.UTF8.GetBytes(prompt))).ToLowerInvariant();
    if ((await LoadJournal()).TryGetValue(key, out var cached))
    {
        Console.Error.WriteLine($"[journal] cache hit for {key[..12]}");
        return cached;
    }
    var result = await compute();
    await journalLock.WaitAsync(); // fan-out writes from many tasks
    try
    {
        var journal = await LoadJournal();
        journal[key] = result;
        var temp = journalPath + ".tmp";
        await File.WriteAllTextAsync(temp, JsonSerializer.Serialize(journal));
        File.Move(temp, journalPath, overwrite: true);
    }
    catch (Exception error) when (error is IOException or UnauthorizedAccessException or NotSupportedException)
    {
        // The journal is best-effort; never discard a computed result.
        Console.Error.WriteLine($"[journal] write failed: {error.Message}");
    }
    finally
    {
        journalLock.Release();
    }
    return result;
}
var journalMutex sync.Mutex

func loadJournal() map[string]string {
	data, err := os.ReadFile(journalPath)
	if err != nil {
		return map[string]string{}
	}
	var journal map[string]string
	if err := json.Unmarshal(data, &journal); err != nil || journal == nil {
		return map[string]string{}
	}
	return journal
}

// journaled returns a cached result for this exact prompt, or computes and persists
// it. This makes the fan-out resumable: interrupt the run, rerun it, and only the
// subtasks that never finished are recomputed. Delete the journal file to start fresh.
func journaled(prompt string, compute func() (string, error)) (string, error) {
	sum := sha256.Sum256([]byte(prompt))
	key := hex.EncodeToString(sum[:])
	if cached, ok := loadJournal()[key]; ok {
		fmt.Fprintf(os.Stderr, "[journal] cache hit for %s\n", key[:12])
		return cached, nil
	}
	result, err := compute()
	if err != nil {
		return "", err
	}
	journalMutex.Lock() // fan-out writes from many goroutines
	defer journalMutex.Unlock()
	journal := loadJournal()
	journal[key] = result
	data, _ := json.Marshal(journal)
	temp := journalPath + ".tmp"
	if err := os.WriteFile(temp, data, 0o644); err != nil {
		fmt.Fprintf(os.Stderr, "[journal] write failed: %s\n", err)
	} else if err := os.Rename(temp, journalPath); err != nil {
		fmt.Fprintf(os.Stderr, "[journal] write failed: %s\n", err)
		_ = os.Remove(temp)
	}
	return result, nil
}

static final ObjectMapper JOURNAL_MAPPER = new ObjectMapper();
static final ReentrantLock JOURNAL_LOCK = new ReentrantLock();

Map<String, String> loadJournal() {
    try {
        return Objects.requireNonNullElseGet(
                JOURNAL_MAPPER.readValue(Files.readString(JOURNAL_PATH), new TypeReference<HashMap<String, String>>() {}),
                HashMap::new);
    } catch (IOException error) {
        return new HashMap<>();
    }
}

// Return a cached result for this exact prompt, or compute and persist it. This
// makes the fan-out resumable: interrupt the run, rerun it, and only the subtasks
// that never finished are recomputed. Delete the journal file to start fresh.
String journaled(String prompt, Callable<String> compute) throws Exception {
    var digest = MessageDigest.getInstance("SHA-256").digest(prompt.getBytes(StandardCharsets.UTF_8));
    String key = HexFormat.of().formatHex(digest);
    String cached = loadJournal().get(key);
    if (cached != null) {
        System.err.println("[journal] cache hit for " + key.substring(0, 12));
        return cached;
    }
    String result = compute.call();
    JOURNAL_LOCK.lock(); // fan-out writes from many threads
    try {
        Map<String, String> journal = loadJournal();
        journal.put(key, result);
        Path temp = JOURNAL_PATH.resolveSibling(JOURNAL_PATH.getFileName() + ".tmp");
        Files.writeString(temp, JOURNAL_MAPPER.writeValueAsString(journal));
        Files.move(temp, JOURNAL_PATH, StandardCopyOption.REPLACE_EXISTING, StandardCopyOption.ATOMIC_MOVE);
    } catch (IOException error) {
        // The journal is best-effort; never discard a computed result.
        System.err.println("[journal] write failed: " + error);
    } finally {
        JOURNAL_LOCK.unlock();
    }
    return result;
}
function loadJournal(): array
{
    $raw = @file_get_contents(JOURNAL_PATH);
    if ($raw === false) {
        return [];
    }
    $decoded = json_decode($raw, true);
    return is_array($decoded) ? $decoded : [];
}

/**
 * Return a cached result for this exact prompt, or compute and persist it. This
 * makes the fan-out resumable: interrupt the run, rerun it, and only the subtasks
 * that never finished are recomputed. Delete the journal file to start fresh.
 */
function journaled(string $prompt, callable $compute): string
{
    $key = hash('sha256', $prompt);
    $journal = loadJournal();
    if (array_key_exists($key, $journal)) {
        fwrite(STDERR, '[journal] cache hit for ' . substr($key, 0, 12) . "\n");
        return $journal[$key];
    }
    $result = $compute();
    $journal = loadJournal();
    $journal[$key] = $result;
    $temp = JOURNAL_PATH . '.tmp';
    $encoded = json_encode($journal, JSON_INVALID_UTF8_SUBSTITUTE);
    if ($encoded === false || @file_put_contents($temp, $encoded) === false || !@rename($temp, JOURNAL_PATH)) {
        fwrite(STDERR, '[journal] write failed: ' . (error_get_last()['message'] ?? json_last_error_msg()) . "\n");
        @unlink($temp);
    }
    return $result;
}
JOURNAL_LOCK = Mutex.new

def load_journal
  JSON.parse(File.read(JOURNAL_PATH)) || {}
rescue SystemCallError, JSON::ParserError
  {}
end

# Return a cached result for this exact prompt, or compute and persist it. This
# makes the fan-out resumable: interrupt the run, rerun it, and only the subtasks
# that never finished are recomputed. Delete the journal file to start fresh.
def journaled(prompt)
  key = Digest::SHA256.hexdigest(prompt)
  cached = load_journal[key]
  unless cached.nil?
    warn "[journal] cache hit for #{key[0, 12]}"
    return cached
  end
  result = yield
  begin
    JOURNAL_LOCK.synchronize do # fan-out writes from many threads
      journal = load_journal
      journal[key] = result
      temp = "#{JOURNAL_PATH}.tmp"
      File.write(temp, JSON.generate(journal))
      File.rename(temp, JOURNAL_PATH)
    end
  rescue SystemCallError => error # the journal is best-effort; never discard a computed result
    warn "[journal] write failed: #{error}"
  end
  result
end

팬아웃한 뒤 검증하기

팬아웃은 최대 MAX_TOTAL_SUBTASKS개의 프롬프트를 받고, 최대 MAX_CONCURRENT개가 동시에 진행되도록 저널을 통해 실행하며(PHP 포트는 순차 실행), 실패를 격리해서 하나의 고장난 서브에이전트가 실행을 끝내는 대신 오류 문자열로 전락하게 해요. 첫 번째 웨이브가 끝나면, 두 번째 웨이브가 같은 서브에이전트 경로를 재사용해서 각 결과를 반박하려고 해요. 모든 검증자는 소스에서 주장을 다시 도출하고, 불확실할 때는 기본적으로 반박된 것으로 처리해요. 원래 결과와 그 판정 모두 오케스트레이터에게 반환되어 함께 저울질할 수 있어요.

```python Python def normalize_subtasks(raw) -> list[str]: """Accept the subtasks input in whatever shape the model emits: an array, the array JSON-encoded as a single string, or a newline-separated list.""" if isinstance(raw, str): try: raw = json.loads(raw) except json.JSONDecodeError: raw = raw.splitlines() if "\n" in raw else [raw] if not isinstance(raw, list): return [] return [task.strip() for task in raw if isinstance(task, str) and task.strip()]

def verify_prompt_for(subtask: str, result: str) -> str: return ( "Adversarially verify the subagent result below: try to REFUTE it. Re-derive the " "claims yourself with bash rather than trusting the result, and look for evidence " "that contradicts them. Default to refuted if uncertain. Call report_findings with " "summary 'refuted: ' or 'confirmed: ', citing the file:line or command " "output that decided it.\n\n" f"Subtask: {subtask}\n\nResult to verify:\n{result}" )

def run_workflow(model: str, raw_subtasks) -> tuple[str, bool]: """Run subtasks as parallel subagents, then run a second verification wave over the results, and return both. MAX_TOTAL_SUBTASKS bounds how many the model can queue; MAX_CONCURRENT bounds how many run at once.""" all_subtasks = normalize_subtasks(raw_subtasks) subtasks = all_subtasks[:MAX_TOTAL_SUBTASKS] dropped = len(all_subtasks) - len(subtasks) if not subtasks: return "Workflow error: no usable subtasks were provided.", True print(f"[workflow] fanning out {len(subtasks)} agents", file=sys.stderr)

  def run_one(prompt: str) -> str:
      try:
          return journaled(prompt, lambda: run_subagent(model, prompt))
      except Exception as error:  # isolation boundary: one bad subagent should not end the run
          return f"(subagent failed: {type(error).__name__}: {error})"

  with concurrent.futures.ThreadPoolExecutor(max_workers=MAX_CONCURRENT) as pool:
      results = list(pool.map(run_one, subtasks))
      print(f"[workflow] verifying {len(results)} results", file=sys.stderr)
      verify_prompts = [verify_prompt_for(task, result) for task, result in zip(subtasks, results)]
      verdicts = list(pool.map(run_one, verify_prompts))

  joined = "\n\n".join(
      f"[agent {index + 1}: {task}]\n{result}\n\n[verify {index + 1}]\n{verdict}"
      for index, (task, result, verdict) in enumerate(zip(subtasks, results, verdicts))
  )
  if dropped > 0:
      joined = (
          f"(note: {dropped} subtasks beyond MAX_TOTAL_SUBTASKS={MAX_TOTAL_SUBTASKS} were not "
          "run; rerun them in a follow-up Workflow call)\n\n" + joined
      )
  return joined, False

```typescript TypeScript
// Accept the subtasks input in whatever shape the model emits: an array, the array
// JSON-encoded as a single string, or a newline-separated list.
function normalizeSubtasks(raw: unknown): string[] {
  let value = raw;
  if (typeof raw === "string") {
    try {
      value = JSON.parse(raw);
    } catch {
      value = raw.includes("\n") ? raw.split("\n") : [raw];
    }
  }
  if (!Array.isArray(value)) {
    return [];
  }
  return value
    .filter((task): task is string => typeof task === "string")
    .map((task) => task.trim())
    .filter((task) => task.length > 0);
}

function verifyPromptFor(subtask: string, result: string): string {
  return (
    "Adversarially verify the subagent result below: try to REFUTE it. Re-derive the " +
    "claims yourself with bash rather than trusting the result, and look for evidence " +
    "that contradicts them. Default to refuted if uncertain. Call report_findings with " +
    "summary 'refuted: <why>' or 'confirmed: <why>', citing the file:line or command " +
    "output that decided it.\n\n" +
    `Subtask: ${subtask}\n\nResult to verify:\n${result}`
  );
}

// Map with a concurrency limit: at most `limit` tasks are in flight at once.
async function mapWithLimit<In, Out>(
  items: readonly In[],
  limit: number,
  task: (item: In) => Promise<Out>,
): Promise<Out[]> {
  const results = new Array<Out>(items.length);
  let cursor = 0;
  const workers = Array.from({ length: Math.min(limit, items.length) }, async () => {
    while (cursor < items.length) {
      const index = cursor++;
      results[index] = await task(items[index]);
    }
  });
  await Promise.all(workers);
  return results;
}

// Run subtasks as parallel subagents, then run a second verification wave over
// the results, and return both. MAX_TOTAL_SUBTASKS bounds how many the model can
// queue; MAX_CONCURRENT bounds how many run at once.
async function runWorkflow(
  model: string,
  rawSubtasks: unknown,
): Promise<{ output: string; isError: boolean }> {
  const allSubtasks = normalizeSubtasks(rawSubtasks);
  const subtasks = allSubtasks.slice(0, MAX_TOTAL_SUBTASKS);
  const dropped = allSubtasks.length - subtasks.length;
  if (subtasks.length === 0) {
    return { output: "Workflow error: no usable subtasks were provided.", isError: true };
  }
  console.error(`[workflow] fanning out ${subtasks.length} agents`);

  const runOne = async (prompt: string): Promise<string> => {
    try {
      return await journaled(prompt, () => runSubagent(model, prompt));
    } catch (error) {
      // Isolation boundary: one bad subagent should not end the run.
      const reason = error instanceof Error ? `${error.name}: ${error.message}` : String(error);
      return `(subagent failed: ${reason})`;
    }
  };

  const results = await mapWithLimit(subtasks, MAX_CONCURRENT, runOne);
  console.error(`[workflow] verifying ${results.length} results`);
  const verifyPrompts = subtasks.map((task, index) => verifyPromptFor(task, results[index]));
  const verdicts = await mapWithLimit(verifyPrompts, MAX_CONCURRENT, runOne);

  let joined = subtasks
    .map(
      (task, index) =>
        `[agent ${index + 1}: ${task}]\n${results[index]}\n\n[verify ${index + 1}]\n${verdicts[index]}`,
    )
    .join("\n\n");
  if (dropped > 0) {
    joined =
      `(note: ${dropped} subtasks beyond MAX_TOTAL_SUBTASKS=${MAX_TOTAL_SUBTASKS} were not ` +
      "run; rerun them in a follow-up Workflow call)\n\n" +
      joined;
  }
  return { output: joined, isError: false };
}
// Accept the subtasks input in whatever shape the model emits: an array, the array
// JSON-encoded as a single string, or a newline-separated list.
List<string> NormalizeSubtasks(JsonElement raw)
{
    List<string> tasks = [];
    if (raw.ValueKind == JsonValueKind.Array)
    {
        tasks = raw.EnumerateArray()
            .Where(item => item.ValueKind == JsonValueKind.String)
            .Select(item => item.GetString()!)
            .ToList();
    }
    else if (raw.ValueKind == JsonValueKind.String)
    {
        var single = raw.GetString()!;
        try
        {
            tasks = JsonSerializer.Deserialize<List<string>>(single) ?? [];
        }
        catch (JsonException)
        {
            tasks = [.. single.Split('\n')];
        }
    }
    return tasks.Where(task => task != null).Select(task => task.Trim()).Where(task => task.Length > 0).ToList();
}

string VerifyPromptFor(string subtask, string result) =>
    "Adversarially verify the subagent result below: try to REFUTE it. Re-derive the "
    + "claims yourself with bash rather than trusting the result, and look for evidence "
    + "that contradicts them. Default to refuted if uncertain. Call report_findings with "
    + "summary 'refuted: <why>' or 'confirmed: <why>', citing the file:line or command "
    + "output that decided it.\n\n"
    + $"Subtask: {subtask}\n\nResult to verify:\n{result}";

// Run subtasks as parallel subagents, then run a second verification wave over
// the results, and return both. maxTotalSubtasks bounds how many the model can
// queue; maxConcurrent bounds how many run at once.
async Task<(string Output, bool IsError)> RunWorkflow(JsonElement rawSubtasks)
{
    var allSubtasks = NormalizeSubtasks(rawSubtasks);
    var subtasks = allSubtasks.Take(maxTotalSubtasks).ToList();
    var dropped = allSubtasks.Count - subtasks.Count;
    if (subtasks.Count == 0)
    {
        return ("Workflow error: no usable subtasks were provided.", true);
    }
    Console.Error.WriteLine($"[workflow] fanning out {subtasks.Count} agents");

    using SemaphoreSlim gate = new(maxConcurrent);
    async Task<string> RunOne(string prompt)
    {
        await gate.WaitAsync();
        try
        {
            return await Journaled(prompt, () => RunSubagent(prompt));
        }
        catch (Exception error)
        {
            // Isolation boundary: one bad subagent should not end the run.
            return $"(subagent failed: {error.GetType().Name}: {error.Message})";
        }
        finally
        {
            gate.Release();
        }
    }

    var results = await Task.WhenAll(subtasks.Select(RunOne));
    Console.Error.WriteLine($"[workflow] verifying {results.Length} results");
    var verifyPrompts = subtasks.Select((task, index) => VerifyPromptFor(task, results[index])).ToList();
    var verdicts = await Task.WhenAll(verifyPrompts.Select(RunOne));

    var joined = string.Join(
        "\n\n",
        subtasks.Select((task, index) =>
            $"[agent {index + 1}: {task}]\n{results[index]}\n\n[verify {index + 1}]\n{verdicts[index]}"));
    if (dropped > 0)
    {
        joined = $"(note: {dropped} subtasks beyond maxTotalSubtasks={maxTotalSubtasks} were not run; "
            + "rerun them in a follow-up Workflow call)\n\n" + joined;
    }
    return (joined, false);
}
// normalizeSubtasks accepts the subtasks input in whatever shape the model emits: an
// array, the array JSON-encoded as a single string, or a newline-separated list.
func normalizeSubtasks(raw json.RawMessage) []string {
	var tasks []string
	if err := json.Unmarshal(raw, &tasks); err != nil {
		var single string
		if err := json.Unmarshal(raw, &single); err != nil {
			return nil
		}
		if err := json.Unmarshal([]byte(single), &tasks); err != nil {
			tasks = strings.Split(single, "\n")
		}
	}
	cleaned := make([]string, 0, len(tasks))
	for _, task := range tasks {
		if trimmed := strings.TrimSpace(task); trimmed != "" {
			cleaned = append(cleaned, trimmed)
		}
	}
	return cleaned
}

func verifyPromptFor(subtask, result string) string {
	return "Adversarially verify the subagent result below: try to REFUTE it. Re-derive the " +
		"claims yourself with bash rather than trusting the result, and look for evidence " +
		"that contradicts them. Default to refuted if uncertain. Call report_findings with " +
		"summary 'refuted: <why>' or 'confirmed: <why>', citing the file:line or command " +
		"output that decided it.\n\n" +
		"Subtask: " + subtask + "\n\nResult to verify:\n" + result
}

// mapWithLimit runs task over items with at most limit goroutines in flight.
func mapWithLimit(items []string, limit int, task func(string) string) []string {
	results := make([]string, len(items))
	semaphore := make(chan struct{}, limit)
	var waitGroup sync.WaitGroup
	for index, item := range items {
		waitGroup.Add(1)
		semaphore <- struct{}{}
		go func() {
			defer waitGroup.Done()
			defer func() { <-semaphore }()
			results[index] = task(item)
		}()
	}
	waitGroup.Wait()
	return results
}

// runWorkflow runs subtasks as parallel subagents, then runs a second verification wave
// over the results, and returns both. maxTotalSubtasks bounds how many the model can
// queue; maxConcurrent bounds how many run at once.
func runWorkflow(ctx context.Context, model string, rawSubtasks json.RawMessage) (string, bool) {
	allSubtasks := normalizeSubtasks(rawSubtasks)
	subtasks := allSubtasks
	if len(subtasks) > maxTotalSubtasks {
		subtasks = subtasks[:maxTotalSubtasks]
	}
	dropped := len(allSubtasks) - len(subtasks)
	if len(subtasks) == 0 {
		return "Workflow error: no usable subtasks were provided.", true
	}
	fmt.Fprintf(os.Stderr, "[workflow] fanning out %d agents\n", len(subtasks))

	runOne := func(prompt string) string {
		report, err := journaled(prompt, func() (string, error) { return runSubagent(ctx, model, prompt) })
		if err != nil {
			// Isolation boundary: one bad subagent should not end the run.
			return fmt.Sprintf("(subagent failed: %s)", err)
		}
		return report
	}

	results := mapWithLimit(subtasks, maxConcurrent, runOne)
	fmt.Fprintf(os.Stderr, "[workflow] verifying %d results\n", len(results))
	verifyPrompts := make([]string, len(subtasks))
	for index, task := range subtasks {
		verifyPrompts[index] = verifyPromptFor(task, results[index])
	}
	verdicts := mapWithLimit(verifyPrompts, maxConcurrent, runOne)

	sections := make([]string, len(subtasks))
	for index, task := range subtasks {
		sections[index] = fmt.Sprintf("[agent %d: %s]\n%s\n\n[verify %d]\n%s",
			index+1, task, results[index], index+1, verdicts[index])
	}
	joined := strings.Join(sections, "\n\n")
	if dropped > 0 {
		joined = fmt.Sprintf("(note: %d subtasks beyond maxTotalSubtasks=%d were not run; "+
			"rerun them in a follow-up Workflow call)\n\n", dropped, maxTotalSubtasks) + joined
	}
	return joined, false
}

// Accept the subtasks input in whatever shape the model emits: an array, the array
// JSON-encoded as a single string, or a newline-separated list.
List<String> normalizeSubtasks(JsonValue raw) {
    List<String> tasks = new ArrayList<>();
    if (raw.asArray().isPresent()) {
        for (JsonValue item : (List<JsonValue>) raw.asArray().get()) {
            tasks.add(item.asString().isPresent() ? item.asStringOrThrow() : item.toString());
        }
    } else if (raw.asString().isPresent()) {
        String single = raw.asStringOrThrow();
        try {
            String[] parsed = new ObjectMapper().readValue(single, String[].class);
            if (parsed != null) {
                for (String task : parsed) {
                    tasks.add(task);
                }
            }
        } catch (JsonProcessingException error) {
            for (String task : single.split("\n")) {
                tasks.add(task);
            }
        }
    }
    return tasks.stream()
            .filter(task -> task != null)
            .map(String::trim)
            .filter(task -> !task.isEmpty())
            .toList();
}

String verifyPromptFor(String subtask, String result) {
    return "Adversarially verify the subagent result below: try to REFUTE it. Re-derive the "
            + "claims yourself with bash rather than trusting the result, and look for evidence "
            + "that contradicts them. Default to refuted if uncertain. Call report_findings with "
            + "summary 'refuted: <why>' or 'confirmed: <why>', citing the file:line or command "
            + "output that decided it.\n\n"
            + "Subtask: " + subtask + "\n\nResult to verify:\n" + result;
}

List<String> runAll(ExecutorService pool, List<String> prompts, Model model) throws InterruptedException {
    List<Callable<String>> jobs = prompts.stream()
            .<Callable<String>>map(prompt -> () -> journaled(prompt, () -> runSubagent(model, prompt)))
            .toList();
    List<String> results = new ArrayList<>();
    for (Future<String> future : pool.invokeAll(jobs)) {
        try {
            results.add(future.get());
        } catch (ExecutionException | CancellationException error) {
            // Isolation boundary: one bad subagent should not end the run.
            Throwable cause = error.getCause() != null ? error.getCause() : error;
            results.add("(subagent failed: " + cause + ")");
        }
    }
    return results;
}

// Run subtasks as parallel subagents, then run a second verification wave over
// the results, and return both. MAX_TOTAL_SUBTASKS bounds how many the model can
// queue; MAX_CONCURRENT bounds how many run at once.
ToolOutput runWorkflow(Model model, JsonValue rawSubtasks) throws InterruptedException {
    List<String> allSubtasks = normalizeSubtasks(rawSubtasks);
    List<String> subtasks = allSubtasks.stream().limit(MAX_TOTAL_SUBTASKS).toList();
    int dropped = allSubtasks.size() - subtasks.size();
    if (subtasks.isEmpty()) {
        return new ToolOutput("Workflow error: no usable subtasks were provided.", true);
    }
    System.err.println("[workflow] fanning out " + subtasks.size() + " agents");

    List<String> results;
    List<String> verdicts;
    try (ExecutorService pool = Executors.newFixedThreadPool(MAX_CONCURRENT, Thread.ofVirtual().factory())) {
        results = runAll(pool, subtasks, model);
        System.err.println("[workflow] verifying " + results.size() + " results");
        List<String> verifyPrompts = IntStream.range(0, subtasks.size())
                .mapToObj(index -> verifyPromptFor(subtasks.get(index), results.get(index)))
                .toList();
        verdicts = runAll(pool, verifyPrompts, model);
    }
    String joined = IntStream.range(0, subtasks.size())
            .mapToObj(index -> "[agent " + (index + 1) + ": " + subtasks.get(index) + "]\n" + results.get(index)
                    + "\n\n[verify " + (index + 1) + "]\n" + verdicts.get(index))
            .collect(Collectors.joining("\n\n"));
    if (dropped > 0) {
        joined = "(note: " + dropped + " subtasks beyond MAX_TOTAL_SUBTASKS=" + MAX_TOTAL_SUBTASKS
                + " were not run; rerun them in a follow-up Workflow call)\n\n" + joined;
    }
    return new ToolOutput(joined, false);
}
/**
 * Accept the subtasks input in whatever shape the model emits: an array, the array
 * JSON-encoded as a single string, or a newline-separated list.
 */
function normalizeSubtasks(mixed $raw): array
{
    if (is_string($raw)) {
        try {
            $raw = json_decode($raw, true, flags: JSON_THROW_ON_ERROR);
        } catch (JsonException) {
            $raw = str_contains($raw, "\n") ? explode("\n", $raw) : [$raw];
        }
    }
    if (!is_array($raw)) {
        return [];
    }
    $tasks = array_map('trim', array_filter($raw, 'is_string'));
    return array_values(array_filter($tasks, fn ($task) => $task !== ''));
}

function verifyPromptFor(string $subtask, string $result): string
{
    return 'Adversarially verify the subagent result below: try to REFUTE it. Re-derive the '
        . 'claims yourself with bash rather than trusting the result, and look for evidence '
        . 'that contradicts them. Default to refuted if uncertain. Call report_findings with '
        . "summary 'refuted: <why>' or 'confirmed: <why>', citing the file:line or command "
        . "output that decided it.\n\n"
        . "Subtask: {$subtask}\n\nResult to verify:\n{$result}";
}

/**
 * Run subtasks through the journal, then run a second verification wave over the
 * results, and return both. PHP's standard runtime has no lightweight thread pool,
 * so both waves run sequentially here (MAX_CONCURRENT is unused); the SDK examples
 * in other languages fan them out in parallel.
 */
function runWorkflow(Client $client, string $model, mixed $rawSubtasks): array
{
    $allSubtasks = normalizeSubtasks($rawSubtasks);
    $subtasks = array_slice($allSubtasks, 0, MAX_TOTAL_SUBTASKS);
    $dropped = count($allSubtasks) - count($subtasks);
    if ($subtasks === []) {
        return ['Workflow error: no usable subtasks were provided.', true];
    }
    fwrite(STDERR, '[workflow] running ' . count($subtasks) . " agents\n");

    $runOne = function (string $prompt) use ($client, $model): string {
        try {
            return journaled($prompt, fn () => runSubagent($client, $model, $prompt));
        } catch (Throwable $error) {
            // Isolation boundary: one bad subagent should not end the run.
            return '(subagent failed: ' . $error::class . ': ' . $error->getMessage() . ')';
        }
    };

    $results = array_map($runOne, $subtasks);
    fwrite(STDERR, '[workflow] verifying ' . count($results) . " results\n");
    $verifyPrompts = array_map(verifyPromptFor(...), $subtasks, $results);
    $verdicts = array_map($runOne, $verifyPrompts);

    $sections = [];
    foreach ($subtasks as $index => $task) {
        $sections[] = '[agent ' . ($index + 1) . ": {$task}]\n{$results[$index]}"
            . "\n\n[verify " . ($index + 1) . "]\n{$verdicts[$index]}";
    }
    $joined = implode("\n\n", $sections);
    if ($dropped > 0) {
        $joined = '(note: ' . $dropped . ' subtasks beyond MAX_TOTAL_SUBTASKS=' . MAX_TOTAL_SUBTASKS
            . " were not run; rerun them in a follow-up Workflow call)\n\n" . $joined;
    }
    return [$joined, false];
}
# Accept the subtasks input in whatever shape the model emits: an array, the array
# JSON-encoded as a single string, or a newline-separated list.
def normalize_subtasks(raw)
  if raw.is_a?(String)
    begin
      raw = JSON.parse(raw)
    rescue JSON::ParserError
      raw = raw.include?("\n") ? raw.split("\n") : [raw]
    end
  end
  return [] unless raw.is_a?(Array)
  raw.select { |task| task.is_a?(String) }.map(&:strip).reject(&:empty?)
end

def verify_prompt_for(subtask, result)
  "Adversarially verify the subagent result below: try to REFUTE it. Re-derive the " \
    "claims yourself with bash rather than trusting the result, and look for evidence " \
    "that contradicts them. Default to refuted if uncertain. Call report_findings with " \
    "summary 'refuted: <why>' or 'confirmed: <why>', citing the file:line or command " \
    "output that decided it.\n\n" \
    "Subtask: #{subtask}\n\nResult to verify:\n#{result}"
end

# Map with a concurrency limit: at most `limit` threads are in flight at once.
def map_with_limit(items, limit)
  results = Array.new(items.length)
  queue = Queue.new
  items.each_with_index { |item, index| queue << [index, item] }
  workers = Array.new([limit, items.length].min) do
    Thread.new do
      until queue.empty?
        index, item = queue.pop(true) rescue break
        results[index] = yield item
      end
    end
  end
  workers.each(&:join)
  results
end

# Run subtasks as parallel subagents, then run a second verification wave over
# the results, and return both. MAX_TOTAL_SUBTASKS bounds how many the model can
# queue; MAX_CONCURRENT bounds how many run at once.
def run_workflow(model, raw_subtasks)
  all_subtasks = normalize_subtasks(raw_subtasks)
  subtasks = all_subtasks.first(MAX_TOTAL_SUBTASKS)
  dropped = all_subtasks.length - subtasks.length
  return ["Workflow error: no usable subtasks were provided.", true] if subtasks.empty?

  warn "[workflow] fanning out #{subtasks.length} agents"
  run_one = lambda do |prompt|
    journaled(prompt) { run_subagent(model, prompt) }
  rescue => error # isolation boundary: one bad subagent should not end the run
    "(subagent failed: #{error.class}: #{error.message})"
  end

  results = map_with_limit(subtasks, MAX_CONCURRENT, &run_one)
  warn "[workflow] verifying #{results.length} results"
  verify_prompts = subtasks.zip(results).map { |task, result| verify_prompt_for(task, result) }
  verdicts = map_with_limit(verify_prompts, MAX_CONCURRENT, &run_one)

  joined = subtasks.each_with_index.map do |task, index|
    "[agent #{index + 1}: #{task}]\n#{results[index]}\n\n[verify #{index + 1}]\n#{verdicts[index]}"
  end.join("\n\n")
  if dropped > 0
    joined =
      "(note: #{dropped} subtasks beyond MAX_TOTAL_SUBTASKS=#{MAX_TOTAL_SUBTASKS} were not " \
      "run; rerun them in a follow-up Workflow call)\n\n#{joined}"
  end
  [joined, false]
end

대화 중간 시스템 메시지로 모드 토글하기

에이전트는 사용자의 메시지를 먼저 추가한 뒤, 해야 할 시스템 메시지(종료 알림, 진입 시 전체 모드 텍스트, 주기적 리프레셔)를 추가해요. 시스템 메시지를 사용자 턴 뒤에 두면, 그 앞에 있는 모든 캐시된 바이트가 온전하게 유지되고, 시스템 메시지는 사용자 턴을 따른다는 배치 규칙도 충족해요.

```bash cURL # One orchestration-mode turn: the mode reminder rides in the messages array as a # {"role": "system"} entry placed after the user turn it applies to. The response # stops at the first tool call. The agent loop that executes tool calls and fans # out subagents is shown in the SDK tabs; the Workflow description is condensed # here, the SDK examples carry the full standing-consent text. curl --fail-with-body -sS https://api.anthropic.com/v1/messages \ -H "x-api-key: $ANTHROPIC_API_KEY" \ -H "anthropic-version: 2023-06-01" \ -H "content-type: application/json" \ -d @- <<'EOF' { "model": "claude-opus-5-5", "max_tokens": 64000, "system": "You are a helpful general-purpose agent. Answer the user's request directly.", "output_config": {"effort": "xhigh"}, "tools": [ { "name": "Workflow", "description": "Orchestrate a multiagent workflow: split a large task into independent subtasks and run them as parallel agents, then collect their results. Opt-in: only use this tool when the user explicitly asks for a workflow, or when a system message confirms that orchestration mode is on. Granularity: scope each subtask to a distinct concern rather than per line or file section; a review of a small module rarely needs more than about ten. Standing consent: while a system message confirms orchestration mode is on, author and run a workflow for every substantive task by default; work solo only on conversational turns or trivial mechanical edits.", "input_schema": { "type": "object", "properties": { "subtasks": { "type": "array", "items": {"type": "string"}, "description": "Independent subtask prompts to run as parallel agents" } }, "required": ["subtasks"] } }, {"type": "bash_20250124", "name": "bash"} ], "messages": [ { "role": "user", "content": "Explore the current directory, then give a thorough review: what it does, code-quality issues, and concrete improvements." }, { "role": "system", "content": "Orchestration mode is on: optimize for the most exhaustive, correct answer rather than the fastest one. Use the Workflow tool on every substantive task, sized to the problem's natural decomposition rather than the maximum the tool allows. See the Workflow tool's description for standing consent, granularity guidance, and quality patterns. Work solo only on conversational or trivial turns." } ] } EOF ```
# One orchestration-mode turn: the mode reminder rides in the messages array as a
# system-role entry placed after the user turn it applies to. The response stops
# at the first tool call. The agent loop that executes tool calls and fans out
# subagents is shown in the SDK tabs; the Workflow description is condensed here,
# the SDK examples carry the full standing-consent text.
ant messages create <<'YAML'
model: claude-opus-5-5
max_tokens: 64000
system: You are a helpful general-purpose agent. Answer the user's request directly.
output_config: {effort: xhigh}
tools:
  - name: Workflow
    description: >-
      Orchestrate a multiagent workflow: split a large task into independent
      subtasks and run them as parallel agents, then collect their results.
      Opt-in: only use this tool when the user explicitly asks for a workflow,
      or when a system message confirms that orchestration mode is on.
      Granularity: scope each subtask to a distinct concern rather than per
      line or file section; a review of a small module rarely needs more than
      about ten. Standing consent: while a system message confirms
      orchestration mode is on, author and run a workflow for every
      substantive task by default; work solo only on conversational turns or
      trivial mechanical edits.
    input_schema:
      type: object
      properties:
        subtasks:
          type: array
          items: {type: string}
          description: Independent subtask prompts to run as parallel agents
      required: [subtasks]
  - {type: bash_20250124, name: bash}
messages:
  - role: user
    content: >-
      Explore the current directory, then give a thorough review: what it
      does, code-quality issues, and concrete improvements.
  - role: system
    content: >-
      Orchestration mode is on: optimize for the most exhaustive, correct
      answer rather than the fastest one. Use the Workflow tool on every
      substantive task, sized to the problem's natural decomposition rather
      than the maximum the tool allows. See the Workflow tool's description
      for standing consent, granularity guidance, and quality patterns. Work
      solo only on conversational or trivial turns.
YAML
class ModeAgent:
    """An agent loop whose orchestration mode is toggled with mid-conversation system messages."""

    def __init__(self, model: str, mode_on: bool = True):
        self.model = model
        self.mode_on = mode_on
        self.messages: list[dict] = []
        self._mode_announced = False
        self._exit_pending = False
        self._turns_since_reminder = 0

    def set_mode(self, mode_on: bool) -> None:
        """Turn the mode on or off. The notice is delivered with the next user turn."""
        if mode_on == self.mode_on:
            return
        if not mode_on:
            if self._mode_announced:
                self._exit_pending = True
        else:
            self._exit_pending = False
        self.mode_on = mode_on

    def _due_system_messages(self) -> list[dict]:
        """System messages owed on this turn: an exit notice, the full mode text on entry,
        or a one-line refresher every TURNS_BETWEEN_REFRESHERS user turns."""
        due = []
        if self._exit_pending:
            self._exit_pending = False
            self._mode_announced = False
            due.append({"role": "system", "content": MODE_EXIT})
        if self.mode_on:
            if not self._mode_announced:
                self._mode_announced = True
                self._turns_since_reminder = 0
                due.append({"role": "system", "content": MODE_ENTER})
            elif self._turns_since_reminder >= TURNS_BETWEEN_REFRESHERS:
                self._turns_since_reminder = 0
                due.append({"role": "system", "content": MODE_REFRESH})
        return due

    def turn(self, user_input: str) -> str:
        # Mid-conversation system messages follow the user turn they apply to, which keeps
        # the cached prefix ahead of them untouched.
        self.messages.append({"role": "user", "content": user_input})
        self.messages.extend(self._due_system_messages())
        self._turns_since_reminder += 1

        for _ in range(MAX_MAIN_TURNS):
            with client.messages.stream(
                model=self.model,
                max_tokens=64000,
                system=SYSTEM_PROMPT,  # static for the whole session
                output_config={"effort": EFFORT},
                tools=[WORKFLOW_TOOL, BASH_TOOL],
                messages=self.messages,
                timeout=REQUEST_TIMEOUT_SECONDS,
            ) as stream:
                response = stream.get_final_message()
            self.messages.append({"role": "assistant", "content": response.content})

            if response.stop_reason == "pause_turn":
                continue
            if response.stop_reason != "tool_use":
                text = "".join(block.text for block in response.content if block.type == "text")
                if response.stop_reason == "max_tokens":
                    # Drop the truncated assistant message so later turns don't build on it.
                    self.messages.pop()
                    text += "\n\n(warning: response was truncated at max_tokens)"
                return text

            tool_results = []
            for block in response.content:
                if block.type != "tool_use":
                    continue
                match block.name:
                    case "Workflow":
                        output, is_error = run_workflow(self.model, block.input.get("subtasks", []))
                    case "bash":
                        output, is_error = handle_bash_block(block)
                    case _:
                        output, is_error = f"unknown tool: {block.name}", True
                tool_results.append(
                    {
                        "type": "tool_result",
                        "tool_use_id": block.id,
                        "content": output,
                        "is_error": is_error,
                    }
                )
            self.messages.append({"role": "user", "content": tool_results})
        return "(hit the main loop turn limit before finishing)"
// An agent loop whose orchestration mode is toggled with mid-conversation system messages.
class ModeAgent {
  private readonly model: string;
  private modeOn: boolean;
  private readonly messages: Anthropic.MessageParam[] = [];
  private modeAnnounced = false;
  private exitPending = false;
  private turnsSinceReminder = 0;

  constructor(model: string, modeOn = true) {
    this.model = model;
    this.modeOn = modeOn;
  }

  // Turn the mode on or off. The notice is delivered with the next user turn.
  setMode(modeOn: boolean): void {
    if (modeOn === this.modeOn) {
      return;
    }
    if (!modeOn) {
      if (this.modeAnnounced) {
        this.exitPending = true;
      }
    } else {
      this.exitPending = false;
    }
    this.modeOn = modeOn;
  }

  // System messages owed on this turn: an exit notice, the full mode text on entry,
  // or a one-line refresher every TURNS_BETWEEN_REFRESHERS user turns.
  private dueSystemMessages(): Anthropic.MessageParam[] {
    const due: Array<{ role: "system"; content: string }> = [];
    if (this.exitPending) {
      this.exitPending = false;
      this.modeAnnounced = false;
      due.push({ role: "system", content: MODE_EXIT });
    }
    if (this.modeOn) {
      if (!this.modeAnnounced) {
        this.modeAnnounced = true;
        this.turnsSinceReminder = 0;
        due.push({ role: "system", content: MODE_ENTER });
      } else if (this.turnsSinceReminder >= TURNS_BETWEEN_REFRESHERS) {
        this.turnsSinceReminder = 0;
        due.push({ role: "system", content: MODE_REFRESH });
      }
    }
    // The published SDK types message roles as "user" | "assistant"; typed support for
    // mid-conversation system messages ships with the SDK release that includes them.
    return due as unknown as Anthropic.MessageParam[];
  }

  async turn(userInput: string): Promise<string> {
    // Mid-conversation system messages follow the user turn they apply to, which keeps
    // the cached prefix ahead of them untouched.
    this.messages.push({ role: "user", content: userInput });
    this.messages.push(...this.dueSystemMessages());
    this.turnsSinceReminder += 1;

    for (let turn = 0; turn < MAX_MAIN_TURNS; turn++) {
      const response = await client.messages
        .stream(
          {
            model: this.model,
            max_tokens: 64000,
            system: SYSTEM_PROMPT, // static for the whole session
            output_config: { effort: EFFORT },
            tools: [WORKFLOW_TOOL, BASH_TOOL],
            messages: this.messages,
          },
          { signal: AbortSignal.timeout(REQUEST_TIMEOUT_SECONDS * 1000) },
        )
        .finalMessage();
      this.messages.push({ role: "assistant", content: response.content });

      if (response.stop_reason === "pause_turn") {
        continue;
      }
      if (response.stop_reason !== "tool_use") {
        let text = response.content
          .filter((block): block is Anthropic.TextBlock => block.type === "text")
          .map((block) => block.text)
          .join("");
        if (response.stop_reason === "max_tokens") {
          // Drop the truncated assistant message so later turns do not build on it.
          this.messages.pop();
          text += "\n\n(warning: response was truncated at max_tokens)";
        }
        return text;
      }

      const toolResults: Anthropic.ToolResultBlockParam[] = [];
      for (const block of response.content) {
        if (block.type !== "tool_use") {
          continue;
        }
        let output: string;
        let isError: boolean;
        switch (block.name) {
          case "Workflow": {
            const input = block.input as { subtasks?: unknown };
            ({ output, isError } = await runWorkflow(this.model, input.subtasks ?? []));
            break;
          }
          case "bash":
            ({ output, isError } = await handleBashBlock(block));
            break;
          default:
            output = `unknown tool: ${block.name}`;
            isError = true;
        }
        toolResults.push({
          type: "tool_result",
          tool_use_id: block.id,
          content: output,
          is_error: isError,
        });
      }
      this.messages.push({ role: "user", content: toolResults });
    }
    return "(hit the main loop turn limit before finishing)";
  }
}
// An agent loop whose orchestration mode is toggled with mid-conversation system messages.
List<MessageParam> messages = [];
var modeOn = true;
var modeAnnounced = false;
var exitPending = false;
var turnsSinceReminder = 0;

// Turn the mode on or off. The notice is delivered with the next user turn.
void SetMode(bool nextModeOn)
{
    if (nextModeOn == modeOn)
    {
        return;
    }
    if (!nextModeOn)
    {
        if (modeAnnounced)
        {
            exitPending = true;
        }
    }
    else
    {
        exitPending = false;
    }
    modeOn = nextModeOn;
}

// The Role property is an open enum, so the mid-conversation "system" role can be assigned
// as a raw string; a dedicated constant ships with the SDK release.
MessageParam SystemMessage(string content) => new() { Role = "system", Content = content };

// System messages owed on this turn: an exit notice, the full mode text on entry,
// or a one-line refresher every turnsBetweenRefreshers user turns.
List<MessageParam> DueSystemMessages()
{
    List<MessageParam> due = [];
    if (exitPending)
    {
        exitPending = false;
        modeAnnounced = false;
        due.Add(SystemMessage(modeExit));
    }
    if (modeOn)
    {
        if (!modeAnnounced)
        {
            modeAnnounced = true;
            turnsSinceReminder = 0;
            due.Add(SystemMessage(modeEnter));
        }
        else if (turnsSinceReminder >= turnsBetweenRefreshers)
        {
            turnsSinceReminder = 0;
            due.Add(SystemMessage(modeRefresh));
        }
    }
    return due;
}

// Send one user turn through the loop, executing tool calls until the model stops.
async Task<string> Turn(string userInput)
{
    // Mid-conversation system messages follow the user turn they apply to, which keeps
    // the cached prefix ahead of them untouched.
    messages.Add(new() { Role = Role.User, Content = userInput });
    messages.AddRange(DueSystemMessages());
    turnsSinceReminder++;

    for (var turn = 0; turn < maxMainTurns; turn++)
    {
        using var deadline = new CancellationTokenSource(TimeSpan.FromSeconds(requestTimeoutSeconds));
        var response = await client.Messages.Create(new MessageCreateParams
        {
            Model = model,
            MaxTokens = requestMaxTokens,
            System = systemPrompt, // static for the whole session
            OutputConfig = new OutputConfig { Effort = effort },
            Tools = [workflowTool, bashTool],
            Messages = messages,
        }, cancellationToken: deadline.Token);
        messages.Add(new()
        {
            Role = Role.Assistant,
            Content = response.Content.Select(block => new ContentBlockParam(block.Json)).ToList(),
        });

        if (response.StopReason == StopReason.PauseTurn)
        {
            continue;
        }
        if (response.StopReason != StopReason.ToolUse)
        {
            var text = string.Concat(
                response.Content.Select(block => block.TryPickText(out var textBlock) ? textBlock.Text : ""));
            if (response.StopReason == StopReason.MaxTokens)
            {
                // Drop the truncated assistant message so the next turn does not build on it.
                messages.RemoveAt(messages.Count - 1);
                text += "\n\n(warning: response was truncated at max_tokens)";
            }
            return text;
        }

        List<ContentBlockParam> toolResults = [];
        foreach (var block in response.Content)
        {
            if (!block.TryPickToolUse(out var toolUse))
            {
                continue;
            }
            string output;
            bool isError;
            if (toolUse.Name == "Workflow")
            {
                toolUse.Input.TryGetValue("subtasks", out var rawSubtasks);
                (output, isError) = await RunWorkflow(rawSubtasks);
            }
            else if (toolUse.Name == "bash")
            {
                (output, isError) = await HandleBashBlock(toolUse);
            }
            else
            {
                output = $"unknown tool: {toolUse.Name}";
                isError = true;
            }
            toolResults.Add(new ToolResultBlockParam(toolUse.ID) { Content = output, IsError = isError });
        }
        messages.Add(new() { Role = Role.User, Content = toolResults });
    }
    return "(hit the main loop turn limit before finishing)";
}
// modeAgent is an agent loop whose orchestration mode is toggled with mid-conversation
// system messages.
type modeAgent struct {
	model              string
	modeOn             bool
	messages           []anthropic.MessageParam
	modeAnnounced      bool
	exitPending        bool
	turnsSinceReminder int
}

func newModeAgent(model string) *modeAgent {
	return &modeAgent{model: model, modeOn: true}
}

// setMode turns the mode on or off. The notice is delivered with the next user turn.
func (agent *modeAgent) setMode(modeOn bool) {
	if modeOn == agent.modeOn {
		return
	}
	if !modeOn {
		if agent.modeAnnounced {
			agent.exitPending = true
		}
	} else {
		agent.exitPending = false
	}
	agent.modeOn = modeOn
}

// dueSystemMessages returns the system messages owed on this turn: an exit notice, the
// full mode text on entry, or a one-line refresher every turnsBetweenRefreshers user turns.
func (agent *modeAgent) dueSystemMessages() []anthropic.MessageParam {
	// MessageParamRole is an open string type, so the mid-conversation "system" role can
	// be expressed directly; a dedicated constant ships with the SDK release.
	systemMessage := func(content string) anthropic.MessageParam {
		return anthropic.MessageParam{
			Role:    anthropic.MessageParamRole("system"),
			Content: []anthropic.ContentBlockParamUnion{anthropic.NewTextBlock(content)},
		}
	}
	var due []anthropic.MessageParam
	if agent.exitPending {
		agent.exitPending = false
		agent.modeAnnounced = false
		due = append(due, systemMessage(modeExit))
	}
	if agent.modeOn {
		if !agent.modeAnnounced {
			agent.modeAnnounced = true
			agent.turnsSinceReminder = 0
			due = append(due, systemMessage(modeEnter))
		} else if agent.turnsSinceReminder >= turnsBetweenRefreshers {
			agent.turnsSinceReminder = 0
			due = append(due, systemMessage(modeRefresh))
		}
	}
	return due
}

// turn sends one user turn through the loop, executing tool calls until the model stops.
func (agent *modeAgent) turn(ctx context.Context, userInput string) (string, error) {
	// Mid-conversation system messages follow the user turn they apply to, which keeps
	// the cached prefix ahead of them untouched.
	agent.messages = append(agent.messages, anthropic.NewUserMessage(anthropic.NewTextBlock(userInput)))
	agent.messages = append(agent.messages, agent.dueSystemMessages()...)
	agent.turnsSinceReminder++

	for range maxMainTurns {
		var response anthropic.Message
		err := func() error {
			ctx, cancel := context.WithTimeout(ctx, requestTimeoutSeconds*time.Second)
			defer cancel()
			stream := client.Messages.NewStreaming(ctx, anthropic.MessageNewParams{
				Model:        agent.model,
				MaxTokens:    64000,
				System:       []anthropic.TextBlockParam{{Text: systemPrompt}}, // static for the whole session
				OutputConfig: anthropic.OutputConfigParam{Effort: effort},
				Tools:        []anthropic.ToolUnionParam{workflowTool, bashTool},
				Messages:     agent.messages,
			})
			defer stream.Close()
			for stream.Next() {
				if err := response.Accumulate(stream.Current()); err != nil {
					return err
				}
			}
			return stream.Err()
		}()
		if err != nil {
			return "", err
		}
		agent.messages = append(agent.messages, response.ToParam())

		if response.StopReason == anthropic.StopReasonPauseTurn {
			continue
		}
		if response.StopReason != anthropic.StopReasonToolUse {
			var text strings.Builder
			for _, block := range response.Content {
				if textBlock, ok := block.AsAny().(anthropic.TextBlock); ok {
					text.WriteString(textBlock.Text)
				}
			}
			if response.StopReason == anthropic.StopReasonMaxTokens {
				// Drop the truncated assistant message rather than leave a clipped turn in history.
				agent.messages = agent.messages[:len(agent.messages)-1]
				text.WriteString("\n\n(warning: response was truncated at max_tokens)")
			}
			return text.String(), nil
		}

		var toolResults []anthropic.ContentBlockParamUnion
		for _, block := range response.Content {
			toolUse, ok := block.AsAny().(anthropic.ToolUseBlock)
			if !ok {
				continue
			}
			var output string
			var isError bool
			switch toolUse.Name {
			case "Workflow":
				var input struct {
					Subtasks json.RawMessage `json:"subtasks"`
				}
				if err := json.Unmarshal(toolUse.Input, &input); err != nil {
					output, isError = fmt.Sprintf("Workflow error: could not parse input: %s", err), true
				} else {
					output, isError = runWorkflow(ctx, agent.model, input.Subtasks)
				}
			case "bash":
				output, isError = handleBashBlock(ctx, toolUse)
			default:
				output, isError = fmt.Sprintf("unknown tool: %s", toolUse.Name), true
			}
			toolResults = append(toolResults, anthropic.NewToolResultBlock(toolUse.ID, output, isError))
		}
		agent.messages = append(agent.messages, anthropic.NewUserMessage(toolResults...))
	}
	return "(hit the main loop turn limit before finishing)", nil
}

// An agent loop whose orchestration mode is toggled with mid-conversation system messages.
class ModeAgent {
    private final Model model;
    private boolean modeOn;
    private final List<MessageParam> messages = new ArrayList<>();
    private boolean modeAnnounced = false;
    private boolean exitPending = false;
    private int turnsSinceReminder = 0;

    ModeAgent(Model model) {
        this(model, true);
    }

    ModeAgent(Model model, boolean modeOn) {
        this.model = model;
        this.modeOn = modeOn;
    }

    // Turn the mode on or off. The notice is delivered with the next user turn.
    void setMode(boolean modeOn) {
        if (modeOn == this.modeOn) {
            return;
        }
        if (!modeOn) {
            if (modeAnnounced) {
                exitPending = true;
            }
        } else {
            exitPending = false;
        }
        this.modeOn = modeOn;
    }

    // System messages owed on this turn: an exit notice, the full mode text on entry,
    // or a one-line refresher every TURNS_BETWEEN_REFRESHERS user turns.
    private List<MessageParam> dueSystemMessages() {
        List<MessageParam> due = new ArrayList<>();
        if (exitPending) {
            exitPending = false;
            modeAnnounced = false;
            due.add(systemMessage(MODE_EXIT));
        }
        if (modeOn) {
            if (!modeAnnounced) {
                modeAnnounced = true;
                turnsSinceReminder = 0;
                due.add(systemMessage(MODE_ENTER));
            } else if (turnsSinceReminder >= TURNS_BETWEEN_REFRESHERS) {
                turnsSinceReminder = 0;
                due.add(systemMessage(MODE_REFRESH));
            }
        }
        return due;
    }

    // MessageParam.Role is an open enum, so the mid-conversation "system" role can be
    // expressed with Role.of; a dedicated constant ships with the SDK release.
    private MessageParam systemMessage(String content) {
        return MessageParam.builder()
                .role(MessageParam.Role.of("system"))
                .content(content)
                .build();
    }

    // Send one user turn through the loop, executing tool calls until the model stops.
    String turn(String userInput) throws InterruptedException {
        // Mid-conversation system messages follow the user turn they apply to, which keeps
        // the cached prefix ahead of them untouched.
        messages.add(MessageParam.builder().role(MessageParam.Role.USER).content(userInput).build());
        messages.addAll(dueSystemMessages());
        turnsSinceReminder++;

        for (int turn = 0; turn < MAX_MAIN_TURNS; turn++) {
            MessageCreateParams params = MessageCreateParams.builder()
                    .model(model)
                    .maxTokens(64000L)
                    .system(SYSTEM_PROMPT) // static for the whole session
                    .outputConfig(OutputConfig.builder().effort(EFFORT).build())
                    .addTool(WORKFLOW_TOOL)
                    .addTool(BASH_TOOL)
                    .messages(messages)
                    .build();
            MessageAccumulator accumulator = MessageAccumulator.create();
            try (var stream = client.messages().createStreaming(params, REQUEST_OPTIONS)) {
                stream.stream().forEach(accumulator::accumulate);
            }
            Message response = accumulator.message();
            messages.add(response.toParam());

            StopReason stopReason = response.stopReason().orElse(null);
            if (StopReason.PAUSE_TURN.equals(stopReason)) {
                continue;
            }
            if (!StopReason.TOOL_USE.equals(stopReason)) {
                String text = response.content().stream()
                        .flatMap(block -> block.text().stream())
                        .map(TextBlock::text)
                        .collect(Collectors.joining());
                if (StopReason.MAX_TOKENS.equals(stopReason)) {
                    // Drop the truncated assistant message so it does not poison later turns.
                    messages.removeLast();
                    text += "\n\n(warning: response was truncated at max_tokens)";
                }
                return text;
            }

            List<ContentBlockParam> toolResults = new ArrayList<>();
            for (ContentBlock block : response.content()) {
                if (block.toolUse().isEmpty()) {
                    continue;
                }
                ToolUseBlock toolUse = block.toolUse().get();
                ToolOutput result = switch (toolUse.name()) {
                    case "Workflow" -> {
                        Map<String, JsonValue> input =
                                (Map<String, JsonValue>) toolUse._input().asObject().orElse(Map.of());
                        JsonValue rawSubtasks = input.getOrDefault("subtasks", JsonValue.from(List.of()));
                        yield runWorkflow(model, rawSubtasks);
                    }
                    case "bash" -> handleBashBlock(toolUse);
                    default -> new ToolOutput("unknown tool: " + toolUse.name(), true);
                };
                toolResults.add(ContentBlockParam.ofToolResult(ToolResultBlockParam.builder()
                        .toolUseId(toolUse.id())
                        .content(result.output())
                        .isError(result.isError())
                        .build()));
            }
            messages.add(MessageParam.builder()
                    .role(MessageParam.Role.USER)
                    .contentOfBlockParams(toolResults)
                    .build());
        }
        return "(hit the main loop turn limit before finishing)";
    }
}
/** An agent loop whose orchestration mode is toggled with mid-conversation system messages. */
class ModeAgent
{
    private array $messages = [];
    private bool $modeAnnounced = false;
    private bool $exitPending = false;
    private int $turnsSinceReminder = 0;

    public function __construct(
        private readonly Client $client,
        private readonly string $model,
        private bool $modeOn = true,
    ) {
    }

    /** Turn the mode on or off. The notice is delivered with the next user turn. */
    public function setMode(bool $modeOn): void
    {
        if ($modeOn === $this->modeOn) {
            return;
        }
        if ($modeOn) {
            $this->exitPending = false;
        } elseif ($this->modeAnnounced) {
            $this->exitPending = true;
        }
        $this->modeOn = $modeOn;
    }

    public function turn(string $userInput): string
    {
        // Mid-conversation system messages follow the user turn they apply to, which keeps
        // the cached prefix ahead of them untouched.
        $this->messages[] = ['role' => 'user', 'content' => $userInput];
        array_push($this->messages, ...$this->dueSystemMessages());
        $this->turnsSinceReminder++;

        for ($turn = 0; $turn < MAX_MAIN_TURNS; $turn++) {
            $stream = $this->client->messages->createStream(
                model: $this->model,
                maxTokens: 64000,
                system: SYSTEM_PROMPT, // static for the whole session
                outputConfig: ['effort' => EFFORT],
                tools: [WORKFLOW_TOOL, BASH_TOOL],
                messages: $this->messages,
                requestOptions: ['timeout' => REQUEST_TIMEOUT_SECONDS],
            );
            [$content, $stopReason] = drainMessageStream($stream);
            $this->messages[] = ['role' => 'assistant', 'content' => $content];

            if ($stopReason === 'pause_turn') {
                continue;
            }
            if ($stopReason !== 'tool_use') {
                $text = '';
                foreach ($content as $block) {
                    if ($block instanceof \Anthropic\Messages\TextBlock) {
                        $text .= $block->text;
                    }
                }
                if ($stopReason === 'max_tokens') {
                    // Drop the truncated assistant message so the next turn does not build on it.
                    array_pop($this->messages);
                    $text .= "\n\n(warning: response was truncated at max_tokens)";
                }
                return $text;
            }

            $toolResults = [];
            foreach ($content as $block) {
                if (!$block instanceof \Anthropic\Messages\ToolUseBlock) {
                    continue;
                }
                if ($block->name === 'Workflow') {
                    [$output, $isError] =
                        runWorkflow($this->client, $this->model, $block->input['subtasks'] ?? []);
                } elseif ($block->name === 'bash') {
                    [$output, $isError] = handleBashBlock($block);
                } else {
                    $output = "unknown tool: {$block->name}";
                    $isError = true;
                }
                $toolResults[] = [
                    'type' => 'tool_result',
                    'tool_use_id' => $block->id,
                    'content' => $output,
                    'is_error' => $isError,
                ];
            }
            $this->messages[] = ['role' => 'user', 'content' => $toolResults];
        }
        return '(hit the main loop turn limit before finishing)';
    }

    /**
     * System messages owed on this turn: an exit notice, the full mode text on entry,
     * or a one-line refresher every TURNS_BETWEEN_REFRESHERS user turns.
     */
    private function dueSystemMessages(): array
    {
        $due = [];
        if ($this->exitPending) {
            $this->exitPending = false;
            $this->modeAnnounced = false;
            $due[] = ['role' => 'system', 'content' => MODE_EXIT];
        }
        if ($this->modeOn) {
            if (!$this->modeAnnounced) {
                $this->modeAnnounced = true;
                $this->turnsSinceReminder = 0;
                $due[] = ['role' => 'system', 'content' => MODE_ENTER];
            } elseif ($this->turnsSinceReminder >= TURNS_BETWEEN_REFRESHERS) {
                $this->turnsSinceReminder = 0;
                $due[] = ['role' => 'system', 'content' => MODE_REFRESH];
            }
        }
        return $due;
    }
}
# An agent loop whose orchestration mode is toggled with mid-conversation system messages.
class ModeAgent
  def initialize(model, mode_on: true)
    @model = model
    @mode_on = mode_on
    @messages = []
    @mode_announced = false
    @exit_pending = false
    @turns_since_reminder = 0
  end

  # Turn the mode on or off. The notice is delivered with the next user turn.
  def set_mode(mode_on)
    return if mode_on == @mode_on

    if mode_on
      @exit_pending = false
    else
      @exit_pending = true if @mode_announced
    end
    @mode_on = mode_on
  end

  def turn(user_input)
    # Mid-conversation system messages follow the user turn they apply to, which keeps
    # the cached prefix ahead of them untouched.
    @messages << {role: "user", content: user_input}
    @messages.concat(due_system_messages)
    @turns_since_reminder += 1

    MAX_MAIN_TURNS.times do
      stream = CLIENT.messages.stream(
        model: @model,
        max_tokens: 64_000,
        system_: SYSTEM_PROMPT, # static for the whole session
        output_config: {effort: EFFORT},
        tools: [WORKFLOW_TOOL, BASH_TOOL],
        messages: @messages,
        request_options: {timeout: REQUEST_TIMEOUT_SECONDS}
      )
      response = stream.accumulated_message
      @messages << {role: "assistant", content: assistant_content_param(response.content)}

      next if response.stop_reason == :pause_turn

      unless response.stop_reason == :tool_use
        text = response.content.select { |block| block.type == :text }.map(&:text).join
        if response.stop_reason == :max_tokens
          @messages.pop # drop the truncated assistant message from the history
          text += "\n\n(warning: response was truncated at max_tokens)"
        end
        return text
      end

      tool_results = []
      response.content.each do |block|
        next unless block.type == :tool_use

        input = parse_tool_input(block.input)
        case block.name
        when "Workflow"
          output, is_error = run_workflow(@model, input["subtasks"] || [])
        when "bash"
          output, is_error = handle_bash_block(block)
        else
          output, is_error = "unknown tool: #{block.name}", true
        end
        tool_results << {
          type: "tool_result",
          tool_use_id: block.id,
          content: output,
          is_error: is_error
        }
      end
      @messages << {role: "user", content: tool_results}
    end
    "(hit the main loop turn limit before finishing)"
  end

  private

  # System messages owed on this turn: an exit notice, the full mode text on entry,
  # or a one-line refresher every TURNS_BETWEEN_REFRESHERS user turns.
  def due_system_messages
    due = []
    if @exit_pending
      @exit_pending = false
      @mode_announced = false
      due << {role: "system", content: MODE_EXIT}
    end
    if @mode_on
      if !@mode_announced
        @mode_announced = true
        @turns_since_reminder = 0
        due << {role: "system", content: MODE_ENTER}
      elsif @turns_since_reminder >= TURNS_BETWEEN_REFRESHERS
        @turns_since_reminder = 0
        due << {role: "system", content: MODE_REFRESH}
      end
    end
    due
  end
end

실행하기

이 예제의 bash 도구는 모델이 작성한 명령을 샌드박스 없이 여러분의 머신에서 직접 실행하고, 팬아웃은 그런 에이전트 여러 개를 병렬로 실행해요. 노출해도 괜찮은 디렉터리와 환경에서 실행하고, 로컬 실험 이상으로 확장하기 전에 샌드박싱을 추가하세요. ```python Python if __name__ == "__main__": task = ( sys.argv[1] if len(sys.argv) > 1 else "Explore the current directory, then give a thorough review: what it does, " "code-quality issues, and concrete improvements." ) agent = ModeAgent(MODEL) print(agent.turn(task)) agent.set_mode(False) print(agent.turn("Briefly summarize what you found above, no fan-out needed.")) ```
const task =
  process.argv[2] ??
  "Explore the current directory, then give a thorough review: what it does, " +
    "code-quality issues, and concrete improvements.";
const agent = new ModeAgent(MODEL);
console.log(await agent.turn(task));
agent.setMode(false);
console.log(await agent.turn("Briefly summarize what you found above, no fan-out needed."));
var task = args.Length > 0
    ? args[0]
    : "Explore the current directory, then give a thorough review: what it does, "
        + "code-quality issues, and concrete improvements.";
Console.WriteLine(await Turn(task));
SetMode(false);
Console.WriteLine(await Turn("Briefly summarize what you found above, no fan-out needed."));
func main() {
	if err := run(context.Background()); err != nil {
		log.Fatal(err)
	}
}

func run(ctx context.Context) error {
	if docTestMode {
		defer os.RemoveAll(workDir)
	}
	task := "Explore the current directory, then give a thorough review: what it does, " +
		"code-quality issues, and concrete improvements."
	if len(os.Args) > 1 {
		task = os.Args[1]
	}
	agent := newModeAgent(modelID)
	answer, err := agent.turn(ctx, task)
	if err != nil {
		return err
	}
	fmt.Println(answer)

	agent.setMode(false)
	summary, err := agent.turn(ctx, "Briefly summarize what you found above, no fan-out needed.")
	if err != nil {
		return err
	}
	fmt.Println(summary)
	return nil
}

void main(String[] args) throws InterruptedException {
    String task = args.length > 0
            ? args[0]
            : "Explore the current directory, then give a thorough review: what it does, "
                    + "code-quality issues, and concrete improvements.";
    ModeAgent agent = new ModeAgent(MODEL);
    IO.println(agent.turn(task));
    agent.setMode(false);
    IO.println(agent.turn("Briefly summarize what you found above, no fan-out needed."));
}
$task = $argv[1] ??
    'Explore the current directory, then give a thorough review: what it does, '
    . 'code-quality issues, and concrete improvements.';
$agent = new ModeAgent($client, MODEL);
echo $agent->turn($task), PHP_EOL;
$agent->setMode(false);
echo $agent->turn('Briefly summarize what you found above, no fan-out needed.'), PHP_EOL;
task = ARGV[0] ||
  "Explore the current directory, then give a thorough review: what it does, " \
  "code-quality issues, and concrete improvements."
agent = ModeAgent.new(MODEL)
puts agent.turn(task)
agent.set_mode(false)
puts agent.turn("Briefly summarize what you found above, no fan-out needed.")

에이전트가 작업할 디렉터리에서 예제를 시작하세요. 예를 들어 검토할 리포지토리의 루트에서요:

python orchestration_mode.py "Review this repository for flaky tests and propose fixes."

모드가 켜져 있으면, 모델이 몇 개의 bash 명령으로 정찰하고, 요청 없이도 Workflow 도구를 파견하며, 서브에이전트 보고를 최종 답변으로 종합할 거예요. 사소하거나 대화적인 요청은 리마인더가 지시한 대로 단독으로 처리돼요.

프로덕션 하네스를 향해

이 예제는 의도적으로 작게 만들었어요. 실제 워크로드를 위한 하네스는 보통 다음을 추가할 거예요:

  • 샌드박스 처리된 오케스트레이션 스크립트: 평평한 서브태스크 문자열 목록만 받는 대신, 모델이 짧은 오케스트레이션 프로그램(분기, 루프, 리듀스 단계)을 만들고 이를 격리된 인터프리터 안에서 실행하게 해요.
  • 내구성 있는 저널링: 로컬 JSON 파일을 프로세스 재시작에도 살아남고 여러 머신의 동시 쓰기에서도 안전한 저장소로 바꿔요.
  • 예산 강제: Workflow 호출별이 아니라 전체 세션에 걸쳐 실행된 총 서브에이전트 수를 추적하고, 하드 상한을 초과하려 하면 거부해서 폭주하는 계획이 할당량을 소진하지 못하게 해요.

이 예제의 패턴(모드 리마인더, 도구 설명의 상시 동의, 저널링, 검증 웨이브)은 그대로 이어지고, 그 주변의 실행 기반만 더 견고해져요.

관련 자료

모드 리마인더가 사용하는 메커니즘과 프롬프트 캐싱과의 상호작용. API가 받아들이는 노력 수준과 하나를 고르는 방법. 도구 정의, 도구 호출 처리, 도구 결과. 이 예제가 로컬에서 실행하는, Anthropic이 정의한 bash 도구.

더 알아보기 (Learn more)