Codex Subagents Guide

OpenCode is the canonical control plane for this repo.

This guide describes the secondary Codex-native integration. It follows the same local workflow:

• no hosted Autolab backend
• no Gastown assets or terminology
• local promoted master in train_orig.py, research/live/master.json, and research/results.tsv
• managed benchmark runs through Hugging Face Jobs
• observability through Trackio

The Codex-native pieces are:

• .codex/config.toml
• .codex/agents/planner.toml
• .codex/agents/experiment-worker.toml
• .codex/agents/memory-keeper.toml
• .codex/agents/reviewer.toml
• .codex/agents/reporter.toml
• .codex/agents/researcher.toml

Recommended Start

Start Codex in the pre-training project root and let the main session spawn the checked-in subagents.

For a longer prompt skeleton:

uv run scripts/print_codex_kickoff.py --gpu-slots 1

Workflow

The Codex-native flow mirrors the OpenCode flow:

1. read AGENTS.md
2. refresh from local promoted master
3. spawn planner for fresh queues
4. spawn reviewer for rule checks
5. spawn experiment_worker for exactly one managed HF Jobs run
6. spawn memory_keeper to persist durable notes
7. spawn reporter for Trackio and HF Jobs summaries

Concurrency

Codex custom agents in this repo use native project-scoped agent files and the same local scripts as OpenCode, but this guide does not assume a built-in Codex worktree-isolation layer for benchmark workers.

Treat one active experiment_worker per checkout as the safe default. If you want more parallel benchmark runs, start additional top-level Codex sessions in separate git worktrees.

Durable State

Use the same repo-local notebook and ledger as OpenCode:

• research/notes.md
• research/do-not-repeat.md
• research/campaigns/
• research/experiments/
• research/results.tsv

Use research/templates/ as the canonical template source. Do not recreate a parallel Codex-specific template tree.