How Claude Code and Codex share one workflow

One canonical WORKFLOW.md and two thin, near-identical entry files, so either agent can pick up work the other started.

A generated project supports both Claude Code and Codex out of the box (--agent both, the initializer's default), without maintaining two workflows.

One canonical file, two thin pointers

circuit/WORKFLOW.md is the entry point every agent follows, regardless of which tool runs it. CLAUDE.md (Claude Code's convention) and AGENTS.md (the emerging cross-tool convention Codex and others read) are generated as thin, near-identical pointers:

# Agent instructions

Read [circuit/WORKFLOW.md](circuit/WORKFLOW.md) first. It is the canonical workflow for this
circuit project; this file only points to it.

- Exact-component evidence lives in the owner bundles under `.claude/skills/component-*/`, indexed
  by `.claude/skills/component-spec-audit/references/inventory.json`. Cross-component rules live in
  `.claude/skills/circuit-spec-integration/references/rules.json`.
- Run `pnpm circuit:check` before editing and again after.
- Never hand-edit the generated pages under `doc/src/content/docs/components/`; change the evidence
  and run `pnpm circuit:generate`.

Neither file restates a workflow step, a verdict, a command, or a file-ownership rule — everything substantive lives in WORKFLOW.md exactly once. --agent claude, --agent codex or --agent none write only the file(s) you ask for; both is the default because it's the cheaper default, not because both tools are assumed to be in use.

Why this makes continuation possible

Because both entry points converge on the same workflow, evidence contract and command set, either agent can continue from the project's committed files. Sources, facts, coverage and the publication lock in circuit/publication/selection.json are durable project state; a later task can read the same WORKFLOW.md, .claude/skills/ evidence and pnpm circuit:check output without relying on the previous agent's conversation.