Skip to main content
A Firedrill integration lives beside your application. You do not need a special agent framework or a separate source repository.

​
Small starting layout

  • firedrill.config.json selects your project and environment. Init creates it.
  • firedrill.recipe.json describes scenarios, a command/HTTP target, configurations, repetitions, and optional Tool bindings.
  • A small adapter calls your real entry point. Native SDK users can keep the recipe in their simulation script instead.
Use the first-run guide or native SDK examples. Your coding agent can author these files after inspecting the actual application.

​
Review what the target receives

The target’s explicit task input is different from the full recipe. Checks, rubrics, and future sequence controls belong to the simulation controller. A command receives a JSON invocation envelope on stdin. Your adapter reads its task input and any scoped Tool connections, calls the real agent, and returns actual output. Keep diagnostic logs out of structured stdout.

​
Keep private state private

Do not commit operator tokens, exported Tool connection files, model credentials, screenshots containing secrets, or private recovery journals. The client keeps recovery state outside normal source files. Preserve its operation IDs and original inputs after uncertainty; do not regenerate a recipe to recover an existing execution.

​
Discover exact schemas

Discovery reports accepted schema versions and supported operations. Validate the full recipe with firedrill doctor firedrill.recipe.json before executing it.