Verify AI-generated code with an agent

Lingua already gives shell-capable agents the important part: a local CLI with stable JSON, deterministic exit codes, bounded output, and explicit runtime failures. The portable lingua-verify skill teaches an agent to use that contract instead of guessing whether generated code worked.

The integration is instruction-only. It does not add a background service, MCP server, hook, credential, or automatic installer.

Check the prerequisites

The skill contains instructions, not the CLI or language runtimes. You need:

• an agent client that can run local terminal commands;
• HTTPS access to GitHub while installing the skill;
• Lingua CLI 1.3.0 or later on PATH;
• the command-line runtime used by the code you want to execute.

Client	Installation requirements
VS Code + GitHub Copilot	Current VS Code, Copilot access, Git, and chat.plugins.enabled: true
Codex	Built-in $skill-installer, GitHub network access, and a writable personal skills folder
Claude Code	Current Claude Code with claude plugin, Git, and HTTPS access to GitHub

Install the CLI first

On macOS, Homebrew installs the CLI and Node.js 24:

brew install johnny4young/tap/lingua-cli
lingua --version

On Windows, Linux, or macOS with Node.js 24 and npm:

npm install -g @linguacode/cli
lingua --version

The Homebrew formula installs Node.js 24. The npm option requires Node.js 24.x, npm, and a user-writable global package location whose bin directory is on PATH. Do not use an unreviewed sudo npm install to hide a permissions problem; fix the npm prefix or use a user-scoped Node installation.

Standalone Windows x64 and Linux x64 archives can run utilities and validate Capsules without a separate Node installation. Running or replaying JavaScript or TypeScript still requires Node.js 24 on PATH. The skill requires Lingua CLI 1.3.0 or later, checks the version, and never installs software on your behalf.

Execution also needs the target runtime:

Target	Runtime on PATH
JavaScript / TypeScript	Node.js 24.x
Python	Project .venv, PYTHON, python3, python, or py
Go	go
Rust file / project	rustc / cargo
Ruby	ruby
Lua	lua

Install the Agent Skill

VS Code and GitHub Copilot

Run Chat: Install Plugin From Source from the Command Palette, then enter:

https://github.com/johnny4young/lingua

Review and trust the repository before enabling the plugin. VS Code discovers the root plugin.json and its skills/lingua-verify directory. If the command is missing, update VS Code, confirm GitHub Copilot access, and enable chat.plugins.enabled.

Codex

Invoke the built-in skill installer and ask it to install the public skill directory:

$skill-installer
Install skills/lingua-verify from https://github.com/johnny4young/lingua

The installed skill becomes available on your next Codex turn. Then mention $lingua-verify, or ask Codex to verify a trusted local script with Lingua.

Claude Code

Add Lingua’s marketplace, then install the native plugin. These commands are the same in macOS, Linux, and Windows terminals:

claude plugin marketplace add https://raw.githubusercontent.com/johnny4young/lingua/main/.claude-plugin/marketplace.json --scope user
claude plugin install lingua@linguacode --scope user

Start a new Claude Code session after a shell installation, or run /reload-plugins inside an existing session. Invoke the namespaced skill as /lingua:lingua-verify, or describe a matching verification task. Claude manages the plugin cache, updates, and removal:

claude plugin update lingua@linguacode
claude plugin uninstall lingua@linguacode

To preview the plugin from a local checkout, without installing or changing your personal config:

claude --plugin-dir ./skills

Then invoke /lingua:lingua-verify. The public marketplace resolves the repository’s default branch, so a pull-request head is not installable through the public URL until it is merged. In a managed environment that blocks third-party marketplaces, an administrator can instead review and copy skills/lingua-verify into the approved Claude skills directory; that fallback gives up the native update and uninstall lifecycle.

Give the agent a bounded task

Good prompts name the trusted target and the evidence you want:

Use Lingua to run ./scripts/normalize.ts with --check as its program argument.
Report the exact runtime, exit code, stdout, stderr, duration, and any truncation.

Validate ./artifacts/run.capsule.json without executing it. If it is valid, stop before replay.

Use Lingua's json-format utility on ./fixtures/payload.json and report structured failures.

The skill chooses the narrowest command:

Goal	Command
Run a trusted file or project	lingua run <target> --json
Validate without execution	lingua capsule validate <file> --json
Reproduce a trusted capture	lingua capsule replay <file> --json
Discover transformations	lingua list utilities --json
Apply one transformation	lingua utility <id> --input <file> --json

It does not replace the repository’s own lint, typecheck, or test gates.

Read the evidence, not just the exit

The agent checks the process exit code and the JSON body:

Exit	Meaning
0	Command completed; inspect the structured result before claiming verification.
1	Invalid arguments or input.
2	Runtime failure, timeout, stop, or non-zero program exit.
3	Unsupported capability or missing runtime.
4	Unclassified internal failure.

A Capsule replay can exit successfully while comparison.matches is false. That is reproducible output drift, not a matching result.

Keep the execution boundary explicit

lingua run and lingua capsule replay execute with your operating-system permissions. Lingua avoids shell interpolation, filters inherited environment values, caps output, and owns the timeout, but it is not an OS sandbox.

• Run only a target you trust inside the active workspace.
• Do not let an agent install Lingua or a missing runtime silently.
• Do not forward secrets through --env.
• Prefer Capsule validation when execution is unnecessary.
• Put untrusted code in a sandbox or container you control.

Why this is not another MCP server

The CLI is already the shortest path for coding agents that have a terminal. The skill adds workflow knowledge and honest evidence without another process or permission surface. Lingua Desktop’s existing local MCP remains the read-only option for exposing one explicitly approved project to a trusted MCP client.

A future stdio MCP adapter should exist only if structured tool discovery proves more useful than this CLI-first path, and execution would require a separate permission design.