Test and ship

Your tests should not depend on which CLIs the CI machine has installed. Both entry points accept a fake probe, so a test simulates any machine without spawning a process. Once those tests pass, the checklist at the end of this page is what to confirm before release.

Simulate detection

detect({ probe }) takes two functions: which resolves a binary name to a path or undefined, and exec returns a command’s output:

import { detect } from "anyagent";

const probe = {
  which: async (bin: string) =>
    bin === "claude" ? "/home/sam/.local/bin/claude" : undefined,
  exec: async () => ({ stdout: "2.0.31", stderr: "", code: 0 }),
};

const installed = await detect({ probe });
// One result: Claude Code 2.0.31, without touching PATH.

Pass adapters alongside probe to scan a subset or include a custom adapter; see DetectOptions.

Simulate credentials and files

create(source, { probe }) takes the richer SystemProbe — env, home dir, and file reads — so authStatus() and models() run against a machine you define:

import { create } from "anyagent";
import { claudeCode } from "anyagent/claude-code";

const agent = create(claudeCode(), {
  probe: {
    env: {},
    exec: async () => ({
      stdout: '{"loggedIn": true}',
      stderr: "",
      code: 0,
    }),
    homedir: () => "/home/fake",
    readFile: async () => undefined,
    which: async () => "/home/fake/bin/claude",
  },
});

const auth = await agent.authStatus(); // answered from the fake, no process

Three machines to cover

Test the three machines your end users actually have. Each is a probe away, and your messages for the first two matter as much as the happy path:

• No supported agent installed: name a supported CLI and its setup steps
• Installed but signed out: report the sign-in state and tell the user how to fix it
• Installed and ready: run the prompt and stream progress

Before you ship

A shipped tool runs a coding agent on a stranger’s machine, under their login, at their expense. Confirm every item before release:

• Detect first, and handle the none-installed case with a message that names a supported CLI and its setup steps
• Check authStatus() where the agent supports it, and tell the user when the CLI is installed but not signed in
• Tell the user which agent runs and that it draws on their login and subscription
• Ask which agent to use when more than one is installed
• Scope the prompt: name the files, keep changes local, forbid commits and pushes
• Use readOnly: true for analysis features; a default run edits files and runs shell commands autonomously, so reserve it for tasks where the user asked for code changes
• Handle all five error codes with messages a stranger can act on
• Bound run time with AbortSignal.timeout so a hung CLI cannot hang your tool
• Pin nothing to agent versions; version is for display, and AnyAgent targets current CLIs
• Test with each supported CLI installed and with none, using a fake probe in CI
• Log argv and stderr on failure so users can file useful bug reports

Next steps