Feature Request: Native Subagent System (spawn_subagents / chain_subagents)

Summary

I've implemented a native subagent system for Codex CLI that allows the main agent to delegate tasks to parallel workers with isolated context windows.

Problem

Currently, when Codex needs to explore a codebase, it fills its own context window with raw file contents. This leads to:

• Context exhaustion during exploration
• Need for aggressive auto-compaction
• Loss of important details when compacting

Issue #2604 (219+ upvotes) requests this feature.

Solution Implemented

I built a working implementation with:

• spawn_subagents: Parallel execution (up to 24 agents)
• chain_subagents: Sequential execution with {{previous_output}} templating
• Context isolation per subagent
• Read-only "explore" mode with restricted tools
• Hardcoded gpt-5.1-codex-mini for subagents
• Transcript persistence for resume capability

Implementation

• Fork with working code: https://github.com/elchapita43/codex/tree/subagentsv2
• New files: core/src/subagents.rs, core/src/tools/handlers/subagents.rs, core/src/tools/policy.rs
• Docs: docs/subagents.md

Request

Would the team be interested in reviewing this for potential merge? Happy to adapt the implementation to fit your roadmap and coding standards.