What problem does it solve? Working with OpenAI Codex on shared repositories requires consistent rules for where plans are stored, how sandbox escalations are requested, and when subagents may be spawned. Without these conventions, plans end up in the wrong location, protected-path writes fail silently, and delegation happens without explicit authorization. ## Core Features & Use Cases - Plan Delivery and Storage Tiers: Delivers plans through a three-tier storage model (shared plans repository, gitignored .plans/ fallback, or conversation only) with commit-and-push escalation for tier 1. - Sandbox Escalation and Protected Paths: Requests require_escalated approval for writes to .git, .agents, and .codex paths, and reports refused escalations instead of bypassing the sandbox. - Subagent Tier and Role Resolution: Resolves deep/standard/mechanical/inherit capability tiers against the models available in the current session and spawns subagents only on explicit request. - Use Case: When implementing an approved plan in Codex on Windows, read the latest versioned plan from the shared plans repository, request an escalation before committing it, and link every document with an absolute forward-slash Markdown path. ## Quick Start Ask Codex to deliver the implementation plan for this repository task following the codex-conventions storage tiers and escalation rules.
