autohelix-setup

Set up an AutoHelix run

Prepare a useful, reviewable improvement experiment for the user's project. Use the installed autohelix CLI; a source checkout is not required. Follow the requested mode: preparation, launch, or management through completion of the agreed execution. Preparing a configuration does not start a run.

AutoHelix gives each attempt a git worktree, runs configured checks and measurements, and merges accepted changes. Each iteration starts a fresh agent session, with persistent notes and measurements from earlier attempts.

Understand the experiment

Inspect the project's instructions, git state, relevant code, existing evaluators, and any autohelix.yaml. Preserve existing work and configuration. Check autohelix --help and relevant subcommand help against the installed version. The harness and the target project may use different environments.

Establish the desired improvement and propose useful loop designs from the project's available evidence, including missing evaluators that would help. Carry forward the user's budget and agent preferences; resolve missing information that changes the experiment without imposing a fixed questionnaire.

For help choosing measurements, constraints, feedback, and budgets, read loop design. For concrete patterns, read the relevant examples. A qualitative goal can use captured evidence and reviewer feedback without inventing a numeric score.

Prepare and verify the configuration

Read the configuration reference when writing or editing the configuration.

For a new setup, autohelix init creates the runtime directory and a config template; it requires a git repository. For an existing setup, preserve its custom prompt and runtime state.

Define a goal, editable scope, constraints, observations, optional metric gates or reviewer feedback, and budget settings or a commented template. Keep tests, benchmark definitions, and evaluation criteria outside the editable scope unless changing them is part of the user's task.