Getting started

Lanzer asks a coding agent to write programs in your Langium DSL, then checks them with your language's own parser, validator and runtime. A campaign is the spec; every run gets a report.

What you need

• A Langium language, with its validations.
• An ACP agent: Claude Code, Codex, Gemini CLI, …
• An agent skill for your DSL. A basic one is enough; Lanzer tells you how to improve it.
• A small host package that plugs your language into Lanzer — see Plug in your DSL.

Try it on Lox

The repo ships a ready host for Lox, pulled in as an unmodified submodule.

git clone --recurse-submodules https://github.com/TypeFox/lanzer
cd lanzer && npm install && npm run build
cd packages/lanzer-lox
cp .env.copy .env      # the agent to use; see Agents and permissions

Check a campaign, look at what the agent will be sent, then run it:

validate

node ./bin/lox-lanzer.js validate ./examples/hello.lanzer

plan

node ./bin/lox-lanzer.js plan ./examples/hello.lanzer --prompt

generate

node --env-file=.env ./bin/lox-lanzer.js generate ./examples/hello.lanzer

✓ helloLox — generated and validated
  ok
  1 attempt(s), 10.7s, stopped: end_turn
  lanzer tools: 1 call(s)
  213,988 tokens, context 43,699/1,000,000 (4.4%), cost 0.1310 USD

TIP

validate, plan and types never call an agent, so they need no .env.

How a run works

1. Lanzer sends the agent the campaign, your grammar reference and your DSL skill.
2. The agent writes the files. While it works it can call Lanzer's validate tool, which runs the same checks that grade the run.
3. Lanzer checks the result: parse, validate, the campaign's requirements, and any run blocks.
4. Problems left at the end of the turn go back to the agent as a fix prompt, up to the attempt budget.
5. The report names the first stage that failed: launch, session, turn, no_output, syntax, semantics, diagnostics, requirements, behaviour or scope.

Lanzer and grammar fuzzers

Fuzzers make thousands of inputs a second, and most make no sense past the parser. Lanzer works on what they can't reach: programs that type-check, resolve and do something specific. Use a fuzzer for the parser, and Lanzer for what comes after it.

Next: write your own campaign.