Claude Code에서 첫 번째 요청을 입력하면 가장 먼저 다음과 같은 JSON을 API로 보낸다. 이 요청은 실제 작업에 앞서 대화 주제를 파악하고 제목을 생성하기 위한 보조 요청이다.

{
  "model": "claude-haiku-4-5-20251001",
  "messages": [
    {
      "role": "user",
      "content": [
        {
          "type": "text",
          "text": "Request Body의 구조를 분석하고 분류별로 묶어서 표현한다. ultrathink"
        }
      ]
    },
    {
      "role": "assistant",
      "content": [
        {
          "type": "text",
          "text": "{"
        }
      ]
    }
  ],
  "system": [
    {
      "type": "text",
      "text": "You are Claude Code, Anthropic's official CLI for Claude."
    },
    {
      "type": "text",
      "text": "Analyze if this message indicates a new conversation topic. If it does, extract a 2-3 word title that captures the new topic. Format your response as a JSON object with two fields: 'isNewTopic' (boolean) and 'title' (string, or null if isNewTopic is false). Only include these fields, no other text. ONLY generate the JSON object, no other text (eg. no markdown)."
    }
  ],
  "tools": [],
  "metadata": {
    "user_id": "user-id"
  },
  "max_tokens": 32000,
  "stream": true
}

시스템 프롬프트를 보면 이 요청이 신규 대화인지 판단하고, 신규 대화라면 2-3 단어의 제목을 추출하여 isNewTopic과 title 필드로 구성된 JSON만 반환하라고 지시하고 있다.

여기서 내 눈에 띈 것은 첫 번째 요청임에도 불구하고 마치 멀티턴 대화가 진행된 것처럼 messages의 마지막 role이 assistant라는 점이었다. 게다가 Claude가 { 한 글자만 응답한 것처럼 구성되어 있다.

이 요청에 대한 응답은 다음과 같다.

{
  "id": "msg_id",
  "type": "message",
  "role": "assistant",
  "model": "claude-haiku-4-5-20251001",
  "content": [
    {
      "type": "text",
      "text": "\n  \"isNewTopic\": true,\n  \"title\": \"Request Body Formatting\"\n}"
    }
  ],
  "stop_reason": "end_turn",
  "stop_sequence": null,
  "usage": {
    "input_tokens": 187,
    "output_tokens": 26,
    "cache_creation_input_tokens": 0,
    "cache_read_input_tokens": 0
  }
}

content.text를 보기좋게 정리해서 적으면 다음과 같다.

  "isNewTopic": true,
  "title": "Request Body Formatting"
}

완전한 JSON에서 맨 앞의 {가 빠진 형태다. 알고 보니 이것은 prefill 기법이라 불리는 것으로, 모델이 응답의 앞부분을 이미 출력한 것처럼 설정하여 이어지는 응답을 원하는 형식으로 유도하는 방법이다.

Claude Code는 이 기법을 활용해 모델이 JSON 형식으로 응답하도록 강제하고 있다. 단순히 "JSON으로 응답해줘"라고 요청하는 것보다 훨씬 확실한 방법이다. 모델 입장에서는 이미 {로 시작했으니 자연스럽게 JSON을 완성할 수밖에 없기 때문이다.

Prefill은 JSON 외에도 다양하게 활용할 수 있다. 예를 들어 ```python으로 시작하면 모델이 파이썬 코드 블록을 완성하게 되고, <analysis>로 시작하면 XML 형식의 응답을 유도할 수 있다.

If you've built CLI tools, you've written code like this:

if (opts.reporter === "junit" && !opts.outputFile) {
  throw new Error("--output-file is required for junit reporter");
}
if (opts.reporter === "html" && !opts.outputFile) {
  throw new Error("--output-file is required for html reporter");
}
if (opts.reporter === "console" && opts.outputFile) {
  console.warn("--output-file is ignored for console reporter");
}

A few months ago, I wrote Stop writing CLI validation. Parse it right the first time. about parsing individual option values correctly. But it didn't cover the relationships between options.

In the code above, --output-file only makes sense when --reporter is junit or html. When it's console, the option shouldn't exist at all.

We're using TypeScript. We have a powerful type system. And yet, here we are, writing runtime checks that the compiler can't help with. Every time we add a new reporter type, we need to remember to update these checks. Every time we refactor, we hope we didn't miss one.

The state of TypeScript CLI parsers

The old guard—Commander, yargs, minimist—were built before TypeScript became mainstream. They give you bags of strings and leave type safety as an exercise for the reader.

But we've made progress. Modern TypeScript-first libraries like cmd-ts and Clipanion (the library powering Yarn Berry) take types seriously:

// cmd-ts
const app = command({
  args: {
    reporter: option({ type: string, long: 'reporter' }),
    outputFile: option({ type: string, long: 'output-file' }),
  },
  handler: (args) => {
    // args.reporter: string
    // args.outputFile: string
  },
});

// Clipanion
class TestCommand extends Command {
  reporter = Option.String('--reporter');
  outputFile = Option.String('--output-file');
}

These libraries infer types for individual options. --port is a number. --verbose is a boolean. That's real progress.

But here's what they can't do: express that --output-file is required when --reporter is junit, and forbidden when --reporter is console. The relationship between options isn't captured in the type system.

So you end up writing validation code anyway:

handler: (args) => {
  // Both cmd-ts and Clipanion need this
  if (args.reporter === "junit" && !args.outputFile) {
    throw new Error("--output-file required for junit");
  }
  // args.outputFile is still string | undefined
  // TypeScript doesn't know it's definitely string when reporter is "junit"
}

Rust's clap and Python's Click have requires and conflicts_with attributes, but those are runtime checks too. They don't change the result type.

If the parser configuration knows about option relationships, why doesn't that knowledge show up in the result type?

Modeling relationships with conditional()

Optique treats option relationships as a first-class concept. Here's the test reporter scenario:

import { conditional, object } from "@optique/core/constructs";
import { option } from "@optique/core/primitives";
import { choice, string } from "@optique/core/valueparser";
import { run } from "@optique/run";

const parser = conditional(
  option("--reporter", choice(["console", "junit", "html"])),
  {
    console: object({}),
    junit: object({
      outputFile: option("--output-file", string()),
    }),
    html: object({
      outputFile: option("--output-file", string()),
      openBrowser: option("--open-browser"),
    }),
  }
);

const [reporter, config] = run(parser);

The conditional() combinator takes a discriminator option (--reporter) and a map of branches. Each branch defines what other options are valid for that discriminator value.

TypeScript infers the result type automatically:

type Result =
  | ["console", {}]
  | ["junit", { outputFile: string }]
  | ["html", { outputFile: string; openBrowser: boolean }];

When reporter is "junit", outputFile is string—not string | undefined. The relationship is encoded in the type.

Now your business logic gets real type safety:

const [reporter, config] = run(parser);

switch (reporter) {
  case "console":
    runWithConsoleOutput();
    break;
  case "junit":
    // TypeScript knows config.outputFile is string
    writeJUnitReport(config.outputFile);
    break;
  case "html":
    // TypeScript knows config.outputFile and config.openBrowser exist
    writeHtmlReport(config.outputFile);
    if (config.openBrowser) openInBrowser(config.outputFile);
    break;
}

No validation code. No runtime checks. If you add a new reporter type and forget to handle it in the switch, the compiler tells you.

A more complex example: database connections

Test reporters are a nice example, but let's try something with more variation. Database connection strings:

myapp --db=sqlite --file=./data.db
myapp --db=postgres --host=localhost --port=5432 --user=admin
myapp --db=mysql --host=localhost --port=3306 --user=root --ssl

Each database type needs completely different options:

• SQLite just needs a file path
• PostgreSQL needs host, port, user, and optionally password
• MySQL needs host, port, user, and has an SSL flag

Here's how you model this:

import { conditional, object } from "@optique/core/constructs";
import { withDefault, optional } from "@optique/core/modifiers";
import { option } from "@optique/core/primitives";
import { choice, string, integer } from "@optique/core/valueparser";

const dbParser = conditional(
  option("--db", choice(["sqlite", "postgres", "mysql"])),
  {
    sqlite: object({
      file: option("--file", string()),
    }),
    postgres: object({
      host: option("--host", string()),
      port: withDefault(option("--port", integer()), 5432),
      user: option("--user", string()),
      password: optional(option("--password", string())),
    }),
    mysql: object({
      host: option("--host", string()),
      port: withDefault(option("--port", integer()), 3306),
      user: option("--user", string()),
      ssl: option("--ssl"),
    }),
  }
);

The inferred type:

type DbConfig =
  | ["sqlite", { file: string }]
  | ["postgres", { host: string; port: number; user: string; password?: string }]
  | ["mysql", { host: string; port: number; user: string; ssl: boolean }];

Notice the details: PostgreSQL defaults to port 5432, MySQL to 3306. PostgreSQL has an optional password, MySQL has an SSL flag. Each database type has exactly the options it needs—no more, no less.

With this structure, writing dbConfig.ssl when the mode is sqlite isn't a runtime error—it's a compile-time impossibility.

Try expressing this with requires_if attributes. You can't. The relationships are too rich.

The pattern is everywhere

Once you see it, you find this pattern in many CLI tools:

Authentication modes:

const authParser = conditional(
  option("--auth", choice(["none", "basic", "token", "oauth"])),
  {
    none: object({}),
    basic: object({
      username: option("--username", string()),
      password: option("--password", string()),
    }),
    token: object({
      token: option("--token", string()),
    }),
    oauth: object({
      clientId: option("--client-id", string()),
      clientSecret: option("--client-secret", string()),
      tokenUrl: option("--token-url", url()),
    }),
  }
);

Deployment targets, output formats, connection protocols—anywhere you have a mode selector that determines what other options are valid.

Why conditional() exists

Optique already has an or() combinator for mutually exclusive alternatives. Why do we need conditional()?

The or() combinator distinguishes branches based on structure—which options are present. It works well for subcommands like git commit vs git push, where the arguments differ completely.

But in the reporter example, the structure is identical: every branch has a --reporter flag. The difference lies in the flag's value, not its presence.

// This won't work as intended
const parser = or(
  object({ reporter: option("--reporter", choice(["console"])) }),
  object({ 
    reporter: option("--reporter", choice(["junit", "html"])),
    outputFile: option("--output-file", string())
  }),
);

When you pass --reporter junit, or() tries to pick a branch based on what options are present. Both branches have --reporter, so it can't distinguish them structurally.

conditional() solves this by reading the discriminator's value first, then selecting the appropriate branch. It bridges the gap between structural parsing and value-based decisions.

The structure is the constraint

Instead of parsing options into a loose type and then validating relationships, define a parser whose structure is the constraint.

Traditional approach	Optique approach
Parse → Validate → Use	Parse (with constraints) → Use
Types and validation logic maintained separately	Types reflect the constraints
Mismatches found at runtime	Mismatches found at compile time

The parser definition becomes the single source of truth. Add a new reporter type? The parser definition changes, the inferred type changes, and the compiler shows you everywhere that needs updating.

Try it

If this resonates with a CLI you're building:

• Documentation
• Tutorial
• conditional() reference
• GitHub

Next time you're about to write an if statement checking option relationships, ask: could the parser express this constraint instead?

The structure of your parser is the constraint. You might not need that validation code at all.