FAQ

Choose the smallest operating boundary that can produce a real receipt, then add capability only when the work earns it. The answers below begin with that practice and use Railyard or Roundhouse mechanisms as evidence.

Railyard vs Roundhouse: do I need both?

Start with the outcome you need to prove. Railyard carries routing, review, merge, and post-merge proof; Roundhouse adds inventory, readiness, convergence, and remote administration when machines become part of that outcome. See the Start decision matrix.

Claude Code vs Codex: which harnesses are supported?

Keep one owning delivery even when execution surfaces change. Claude Code and Codex are both supported; same-harness execution is the default. Cross-harness dispatch is opt-in, requires the destination Codex CLI to be set up separately, and records the harness, model, and effort. The cross-harness practice shows that receipt.

Does this work on Windows and WSL?

Treat every operating-system boundary as its own proof surface. The documented support includes macOS, Linux, Windows, and WSL, with native Windows evidence kept distinct from WSL. The installation compatibility notes state the CLI and tooling assumptions.

Is a GitHub remote required? What if there is no test suite?

Match the receipt to the available delivery surface. A GitHub remote enables a published pull request; a local delivery can still run its checks and name the unavailable publish step. When no suite exists, the route chooses the smallest applicable command and records the verification boundary honestly.

What does it cost?

Railyard itself is free and open source (MIT); you pay only your own Claude/Codex usage, billed exactly as any other session in that harness. The control model cost guide explains how to choose models and efforts with cost in mind.

What else gets installed?

Railyard uses Compound Engineering (EveryInc) for selected workflow stages. Resolve CE when a stage needs it; startup does not bootstrap workflow dependencies. Ponytail (DietrichGebert) and Superpowers are not Railyard prerequisites.

Does every request need a delivery workflow?

Routine work runs natively, and useful CE workflows are selected automatically. Use compound-engineering:ce-commit-push-pr whenever creating a PR or pushing user-requested commits to an existing PR. CE owns one review settlement and CI loop. Fleet or account orchestration requires explicit scope; native child agents are the default for parallel work, and visible Codex tasks require an explicit user request.

What is a receipt, a run log, or Thermos?

• A receipt is the evidence chain from intent through merge and post-merge proof.
• A run log records decisions when an audit is useful; audit and retrospective are optional.
• Thermos supplies both review lenses for the whole-candidate gate before every push, alongside codex review on GPT-6.1 Sol at high effort. Its findings feed the existing CE owner.