Project contract for Codex and any other AI coding assistant.
The canonical contract is CLAUDE.md. This project keeps a single source of truth to avoid drift — read CLAUDE.md for §1–§14 (project, stack, layout, setup & hard rules, env vars, patterns, Docker, make targets, langgraph.json, conventions, testing, workflow, out-of-scope, references) and PLAN.md for the architecture, locked decisions, and build phases.
Terminology note (for any tool reading this): review knowledge ships as portable Agent Skills in the open SKILL.md format (originated by Claude). The agent loads them through its own provider-agnostic loader — default provider OpenAI gpt-5-mini, switchable to Anthropic/Google — and does not use Claude's native code-execution Skills runtime. Skills are prompt-only.
Do not maintain a separate full copy here. If your tooling needs tool-specific notes, add them below this line rather than duplicating
CLAUDE.md(a prior mechanical "Claude→Codex" copy corrupted factual content — e.g. doc URLs — and is what this pointer prevents).