Start here

Begin with one outcome you can prove. Install the smallest useful surface, make one delivery on the machine you already trust, and add fleet capabilities when you request placement or convergence. This sequence builds confidence from working evidence and keeps the operating model easy to understand.

If a command or gate stops, use the troubleshooting guide before changing the delivery path.

First delivery example

For a request to fix an issue and get it merged, begin with native execution and select useful CE workflows. CE owns review settlement and the CI loop through the requested merge. Show the actual checks and post-merge evidence so the operator can assess the result. Add Roundhouse for explicitly requested fleet work.

• Install — add Railyard to Claude Code, Codex, or both; Roundhouse is optional for fleet work.
• First delivery — go from one sentence to a verified merge on one machine.
• First machine — establish a readable fleet baseline and a readiness proof.
• Bootstrap a fleet — create the private store, signed genesis, verified remote, and first convergence.

What happens around your request

Routine work runs natively. Select CE workflows automatically when they help, and resolve CE only when a selected stage needs it. Startup loads routing guidance without bootstrapping workflow dependencies.

Before every push, complete the whole-candidate review gate: both Thermos lenses and codex review on GPT-6.1 Sol at high effort cover the complete cumulative change and affected lifecycle. Resolve findings and bind the completed evidence to the exact candidate. Use compound-engineering:ce-commit-push-pr to publish it; CE owns one review settlement and CI loop. Reuse valid evidence and recheck changed interactions before repair pushes. Audits or retrospectives remain optional.

Use native child agents for useful parallel work. Create visible Codex tasks only when the user explicitly requests them. Fleet or account orchestration also requires explicit scope.

Choose both model and reasoning effort deliberately. In Claude Code, Opus 5.5 at medium is the default for substantive subagent work; in Codex, GPT-6.1 Sol at medium is the baseline for implementation, debugging, planning, and ambiguous or multi-file work, even when small or bounded. Omitting the model means the child inherits. Disclose an unavailable selection instead of silently substituting one. The model routing guide explains the choice.

Choose the path

The delivery system and fleet system each produce value on their own. With both installed, their integration adds readiness-aware placement.

Railyard and Roundhouse provide the implementation evidence behind this way of working. Their delivery source and fleet source keep releases and review history visible beside the guide.