commands.md

Commands

Every host CLI has the same commands. The examples use Lox's lox-lanzer, run from packages/lanzer-lox; yours has your language's name.

Command	Calls an agent	What it does
validate <file>	no	Parses a campaign and checks every selector against your grammar.
types <file>	no	Lists the AST types a selector can use, with their properties.
plan <file>	no	Shows the jobs, and with --prompt the exact prompt the agent gets.
generate <files...>	yes	Runs campaigns through the agent and checks the result.
compare <a> <b>	no	Puts two reports side by side. See Evaluating a skill.

validate, types, plan and compare take --json to print data instead.

generate

npx lox-lanzer generate ./examples/hello.lanzer

Give it several files or folders and they run as one suite, with one report.

Flag	Effect
--model <m>, --command <bin>	Override the agent from .env.
--max-attempts <n>	Prompts per session, the first plus fix prompts.
--allow <kinds>, --allow-all	What the agent may do. See Agents and permissions.
--report <path>, --no-report	Where the JSON report goes (default .lanzer/reports/).
--verbose, --quiet	More or less live output.
--runs <n>, --parallel <k>, --min-pass <share>	Repeated and parallel runs.
--skill <dir>, --no-skill, --policy minimal, --no-tools, --isolated	Evaluating a skill.

Reading the result

A failure names the first stage that failed, so you fix the cause, not the loudest symptom:

✗ impossible — the files are valid but the campaign requirements are unmet
  failed at requirements — the files are valid but the campaign requirements are unmet
  2 attempt(s), 33.0s, stopped: end_turn
  lanzer tools: 2 call(s), 1 reporting problems
    - Forbidden selector matched 1 node(s) in mainFile: FunctionDeclaration

A pass can still have struggled. validate: lists the agent's own checks along the way:

✓ integerMath — generated and validated
  ok
  1 attempt(s), 125.3s, stopped: end_turn
  lanzer tools: 9 call(s), 8 reporting problems
  validate: ✗1 ✗1 ✗1 ✗1 ✗1 ✗1 ✗1 ✗1 ok

The JSON report has every diagnostic, the agent's configuration, tokens and cost, and what each failed check reported.

In code

The same run, from TypeScript:

import { runLoxCampaignFile } from 'lanzer-lox';
import { resolveAcpOptionsFromEnv } from 'lanzer';

const { runs } = await runLoxCampaignFile('campaign.lanzer', resolveAcpOptionsFromEnv());