lgbarn/shipyard/shipyard-codex-orchestration

Shipyard orchestration under Codex

Shipyard was built for Claude Code, where workflows dispatch fresh subagents in parallel (builder, reviewer, auditor, architect, verifier, researcher, mapper, …) via the Task tool. Codex has no equivalent — it runs in a single context.

So under Codex, Shipyard workflows run as inline sequential personas: instead of dispatching a separate agent, you adopt each role yourself, one after another, in this same conversation.

Honest degradation — what you lose

• No parallelism. Roles run one at a time, so multi-agent phases are slower.
• No context isolation. In Claude Code a fresh reviewer can't be biased by the

builder's reasoning. Here the same context does both, so a review is less independent. Compensate by being deliberately skeptical when you switch into a review/verify role — re-read the actual artifact, don't trust your own prior step.

What you keep: the discipline — the review and verify gates still run, just sequentially.

The inline-sequential pattern

When a workflow calls for "dispatch the X agent", instead:

1. State which role you're entering ("Now acting as the reviewer…").
2. Adopt that role's mandate fully — including its MUST-NOT constraints (a reviewer

makes no code edits; a mapper only documents).

1. Produce the role's deliverable.
2. Exit the role and continue the workflow.

Routing — which skill for which intent

If a workflow has a dedicated skill, use it and apply the inline-sequential pattern above to any agent dispatch it describes. The shipyard-review, shipyard-research, shipyard-map, and shipyard-ship skills exist because those roles have no other skill home in Codex.

State

Codex has no SessionStart hook, so Shipyard state is not auto-loaded. When a workflow needs project state, run scripts/state-read.sh yourself at the start (see the state/utility skills).