00-core-routing

/AGENTS.md inside the agent configuration root (.cursor/AGENTS.md) is the canonical operating contract. Follow its instruction precedence, autonomy limits, phased lifecycle, file-role boundaries, validation standard, and completion standard. A leading / in control-plane files is relative to that configuration root, not the filesystem root.

Loading policy:

• Project overlay: when present, inspect /PROJECT.md once for project-specific constraints and task-relevant routes. Follow applicable project rules; open linked references only for the current scope. Never replace local state, preferences, tools, or memory with template defaults during updates.
• Lead session start: read /AGENTS.md once and inspect /STATE.md for relevant active work. Run read-only node .cursor/skills/launch-pipeline/scripts/preflight.mjs only when the host exposes permitted execution; Ask Mode without execution defers it as unrun. Run /scripts/bootstrap.sh only for required materialization or explicit installation/repair after implementation authorization, never every ready session.
• Bounded delegates consume their task packet and applicable native instructions; no lead startup, shared-state/history sweep, mandatory workstream, or exhaustive role plan. Return scoped evidence to the lead.
• Later turns: reuse instructions already in context. Re-read a control file only when it changed, the session or context was reset or compacted, a new scope became relevant, or exact wording matters for a consequential decision. Do not re-read the core set merely because a new message arrived.
• Detailed modes: use /INSTRUCTIONS.md to load matched planning, strategy, or bounded-delegation guidance. /instructions/LAUNCH.md is explicit-only through /launch-pipeline or an explicit request for that named lifecycle. Load /instructions/ROLES.md only for formal multi-role delivery or consequential gates, and then only sections 1–3 plus relevant role sections.
• On demand: /SKILLS.md and /skills/<skill-id>/SKILL.md before improvising a procedure that may already exist; /TOOLS.md when selecting deployment, database, integration, or unfamiliar tooling, preferring direct, verifiable project-native and official tools and updating the registry when capabilities materially change; /USER.md when a choice depends on standing preferences; /memory/MEMORY.md and linked runbooks or blockers when a prior decision, exact procedure, or unresolved issue may apply.

State discipline: the lead updates /STATE.md and existing plans when resumable state materially changes. Preserve state before phase transitions, long tasks, repeated attempts, or likely context loss. Record substantive continuity under /memory/memories/; read-only answers and bounded delegate reports do not require shared-state or history mutation.

Do not leave durable rules only in chat. Do not defer agent-capable work to the user. Do not store secret values in agent files. Register future instruction files in /INSTRUCTIONS.md instead of adding always-on rules.