Claude Code Subagents Guide

OpenCode is the canonical control plane for this repo.

This guide describes the secondary Claude Code-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 Claude Code-native pieces are:

• CLAUDE.md
• .claude/settings.json
• .claude/agents/autolab.md
• .claude/agents/planner.md
• .claude/agents/experiment-worker.md
• .claude/agents/memory-keeper.md
• .claude/agents/reviewer.md
• .claude/agents/reporter.md
• .claude/agents/researcher.md

Recommended Start

From the pre-training project root:

claude --agent autolab

Treat claude --agent autolab as the canonical launch path for this repo. Some Claude Code builds may not list project agents consistently in claude agents or /agents, even when direct launch works.

For a longer prompt skeleton:

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

Workflow

The Claude-native flow mirrors the OpenCode flow:

1. read AGENTS.md
2. refresh from local promoted master
3. use planner for fresh queues
4. use reviewer for rule checks
5. use experiment-worker for exactly one managed HF Jobs run
6. use memory-keeper to persist durable notes in the main checkout
7. use reporter for Trackio and HF Jobs summaries

The worker is configured with background: true and isolation: worktree, so it uses Claude Code's native worktree integration rather than a repo-specific worker launcher.

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 Claude-specific template tree.