#Install A3S Test

A3S Test has two separately versioned installable parts because they own different trust boundaries.

• Web Test Kit is a frontend development dependency. It provides post-render Page Context, source mapping, and the optional in-page review surface.
• CLI and Agent Skill run on the developer machine and in the coding agent's Skill directory. They own sessions, typed actions, ACL, evidence, repair coordination, and process cleanup.

Install Web Test Kit for in-page marking, capture, or sketching. Use the CLI and Agent Skill for terminal exploration and ACL. Install both when a coding agent must receive a page finding and verify the repair.

#Install Web Test Kit

Run this command in the frontend project:

npm install --save-dev @a3s-lab/testkit@0.6.2

Use the matching command when the project already uses another package manager:

@a3s-lab/testkit 0.6.2 is published on the official npm Registry with GitHub OIDC provenance. Pinning the version keeps installation reproducible, and the package manager stores its integrity in the project lockfile.

Verify the installed dependency:

npm ls @a3s-lab/testkit

The Test Kit package version advances independently from the A3S Test Release tag, so the two version numbers do not need to match. Continue with the three-step Web Test Kit setup. Test Kit 0.6.2 includes the simplified review panel, bounded SVG design board, in-page region capture, live CLI compatibility handshake, ranked rendered-node source mapping, and revision-scoped Page Context diffs without a browser extension, drawing SDK, screenshot plug-in, or screen-sharing permission.

#Install the CLI and Agent Skill

The installer detects the current platform, downloads the prebuilt CLI, verifies its SHA-256, and installs a matching Agent Skill. Run the same command again to upgrade.

#macOS and Linux

curl -fsSL https://github.com/A3S-Lab/Test/releases/latest/download/install.sh | sh

Target Codex explicitly:

curl -fsSL https://github.com/A3S-Lab/Test/releases/latest/download/install.sh |
  sh -s -- --agent codex

Pin a version:

curl -fsSL https://github.com/A3S-Lab/Test/releases/latest/download/install.sh |
  sh -s -- --version v1.0.1

#Windows PowerShell

& ([scriptblock]::Create((irm 'https://github.com/A3S-Lab/Test/releases/latest/download/install.ps1')))

Select an agent or version:

& ([scriptblock]::Create((irm 'https://github.com/A3S-Lab/Test/releases/latest/download/install.ps1'))) -Agent codex
& ([scriptblock]::Create((irm 'https://github.com/A3S-Lab/Test/releases/latest/download/install.ps1'))) -Version v1.0.1

#Supported agent targets

auto installs only for detected tools and falls back to the universal directory. all targets every supported directory. The installers also accept --skill-only, --cli-only, --install-dir, and --skill-dir, with corresponding PascalCase PowerShell parameters.

#Other CLI installation paths

Build the CLI from a Git tag:

cargo install --git https://github.com/A3S-Lab/Test \
  --tag v1.0.1 --locked a3s-test-cli

Each release also contains a3s-test.skill, platform archives, checksum files, the Test Kit package, and an immutable Linux runner image reference. Do not bypass checksums or replace the released image digest with a floating tag.

#Start one local review loop

The project-loop commands in this section are included in the v1.0.1 release and can be used directly after installation.

After mounting Test Kit in the frontend, run these commands from the project root:

a3s-test init
a3s-test doctor
a3s-test dev

init detects the package manager, dev or start script, framework port, and Test Kit declaration, then writes .a3s-test/project.acl. It does not install packages or start processes. When Test Kit is missing, run the exact install command printed by init, mount the provider and overlay, then run doctor again.

doctor separates a missing declaration from a declared but uninstalled package and rejects Test Kit versions outside the CLI compatibility range. dev reuses the configured URL when it already responds. Otherwise it owns the configured development process, waits for the URL, and opens one headed review browser. Ctrl+C closes the exact browser session and only the server that dev started.

For a project that intentionally uses browser accessibility without the in-page overlay, initialize with:

a3s-test init --testkit optional

Use --json on all three commands in agent automation. dev --json emits one compact a3s.test.dev/1 event per stdout line and keeps development-server logs on stderr. Once the live Test Kit handshake passes, its in-process a3s.test.local-repair-bridge/1 emits evidence-backed repair_batch events on that same stream. Each event contains the generated session ID, so agent automation does not launch a second manually coordinated repair-watch process.

#Add trusted verification checks

init deliberately does not infer a test command. Package scripts may start a watcher, wait for input, or perform effects. To let agent repair-verify run project checks automatically, add an explicit catalog inside the generated project block:

verification {
  check "component" {
    tier = "focused"
    executable = "npm"
    args = ["run", "test:component"]
    working_directory = "."
    file_prefixes = ["src/components"]
    timeout_ms = 120000
    cleanup_timeout_ms = 10000
  }

  check "workspace" {
    tier = "regression"
    executable = "npm"
    args = ["run", "test"]
    working_directory = "."
    file_prefixes = []
    timeout_ms = 300000
    cleanup_timeout_ms = 10000
  }
}

Focused checks map contained project-relative source prefixes to the smallest useful command. Regression checks have no prefixes and are selected only when source mapping, check coverage, browser-error deltas, or prior proof show that the repair may have broader impact. Commands run directly without a shell and inside an A3S Test-owned process tree with bounded execution and cleanup.

Omit --checks-json from agent repair-verify to use this catalog. Supplying --checks-json keeps caller-reported mode for an orchestrator that already ran its own checks. An expanded verification with no trusted project check fails closed instead of claiming broad coverage.