codex-claude-mcp: Instructions file for Claude Code

Noodle Seed Project Context

Agent target: Claude Code.

Build this project as a Noodle Seed MCP server or app authored in TypeScript. Every --json command uses the canonical envelope on stdout and keeps stderr empty: one-shot commands write exactly one envelope on stdout; streaming commands write one NDJSON envelope per line. Drive the loop by parsing machine state — not by reading source or scraping human prose.

Agent-native loop

• Applicability: when the request is unrelated to the Noodle MCP server or app, follow the project's normal instructions and run no Noodle lifecycle commands.
• Route first: read the noodle-seed skill SKILL.md, choose exactly one primary route for the requested outcome, and read that primary reference. Read supporting references only when the route or observed evidence requires them.
• Discover: noodle commands --json — every command, flag, and exit code (don't read source).
• Author, when the selected route requires it: edit the configured TypeScript entrypoint — follow that route and its capability references.
• Validate: noodle validate --json → on failure {ok:false,error:{code,message,fix,next,errors:[{code,path,message}]}}; per-field detail is in error.errors[].
• Repair: fix each error.errors[] entry at its path, re-run validate; noodle validate --fix-prompt gives ready repair prose. Never freeform re-edit.
• Smoke: noodle test --json.
• Continue only to the level the selected route requests: Apps/widgets use noodle check --json plus noodle devtools; hosted inspection stays read-only; hosted mutation runs only when the current user request explicitly authorizes the exact action and target.
• Stale skill? noodle agents doctor --json → noodle agents setup --write.

The exact --json/exit-code contract is in the noodle-seed skill and its references/agent-contract.md. Human-oriented command prose lives in the project README, not here.