Skip to main content
Browse documentation
REPOTOIRE / DOCSCODE. CONTEXT. EXECUTION.

Quick Start

Start with one goal and the repository it belongs to. The five-step walkthrough explains how that goal becomes checked work.

Current boundary: v1 execution is serial. The homepage's parallel agents illustrate the future architecture.

Set up the private beta

RepoToire is a commercial private beta. Request access for installation and host setup guidance for Codex or Claude Code.

With RepoToire installed, open the repository you want to work on and check its supported capabilities:

cd your-project
repotoire doctor .

Start with the outcome

Give your agent a goal in your own words. For example:

Make this website easier to use on a phone. Readers should be able to find the docs and request access without horizontal scrolling.

This is example intent, not an executable program. Decide what should improve, what must stay unchanged, and how you will check the result.

Keep that outcome visible as the agent breaks the work into tasks. A task list is useful only when its results serve your goal.

Gather code evidence

Your agent can choose context, impact, and direct file inspection as needed. There is no required tool sequence.

# Show code structure around a real symbol in your repository.
repotoire context . --focus SymbolName

# Find supported in-repository dependents of that symbol.
repotoire impact SymbolName .

# Read the repository's maintained verification commands.
repotoire verification-profiles --format markdown .

Replace SymbolName with the symbol you are investigating. Code facts help the agent choose where to look; they do not prove behavior or replace tests.

Review before execution

When using the APU workflow, review two separate things:

  1. The Goal Program: the outcome, non-goals, success criteria, and intended work match what you asked for.
  2. The compiled executable: its scopes and checks match the reviewed intent. Execution approval must identify that exact image.

The compiler uses current CodebaseView evidence and repository-owned verification recipes. Missing or unsupported inputs produce diagnostics; do not replace them with guessed commands or broader permissions.

Approving intent is not permission to execute. Machine owns admitted execution and checked writeback. The host or coding agent cannot grant itself that authority.

Inspect the run

For an initialized Machine run, inspect its state:

repotoire machine report .

A report describes runtime state; reading it does not start work or prove the goal is complete. If progress stops, review the reported reason and missing evidence before authorizing a changed executable.

Check what finished

Compare the result with the original success criteria. For the phone example, check that readers can reach the docs and access form, and that the page fits the screen.

Require the matching checks and Machine's retirement evidence for an APU completion claim. Worker completion alone is insufficient. A failed check must not be reported as success.

Retirement records checked execution outcomes. Committing, pushing, merging, and deploying are separate delivery actions with their own authorization and evidence.

Next steps