Env Setup

References

Read these first:

• ../../../.agents/references/deps-update-rule.md
• ../../../.agents/references/language-policy.md
• ../../../.agents/references/documentation-style.md
• ./references/environment-refresh.md
• ../../../CLAUDE.md if it exists
• ../init-project/references/claude-maintenance.md

When To Use

Interpret natural-language requests as one of:

• create
• refresh

Required Work

create

1. Gather environment name and dependency source.
2. Create the environment when explicitly requested.
3. Install dependencies.
4. Check wandb installation and login if relevant.
5. Run the refresh logic afterward.

refresh

1. Detect Python, PyTorch, CUDA, GPU, relevant dependencies, conda env name, and wandb status.
2. Use the shell commands listed in ./references/environment-refresh.md for the first-pass machine summary.
3. Update the runtime environment facts in the ## Environment section of CLAUDE.md.
4. Preserve ### Dataset Paths inside that section unless dataset addresses are intentionally being refreshed from project state.
5. Preserve the rest of CLAUDE.md exactly.
6. Follow ../init-project/references/claude-maintenance.md for section-safe final writes.
7. Report a Gate ledger when dependency files, setup commands, or the CLAUDE.md environment section change. If py_compile or ruff checks are not relevant or cannot run, mark them NOT_RUN with the reason.

Codex Adaptation

• Treat natural-language requests as the canonical $env-setup {create|refresh} interface.
• If the canonical prompt wants AskUserQuestion, ask the user directly only for missing setup choices.
• Preserve the rule that refresh mode updates runtime environment facts while keeping dataset addresses intact unless the workflow explicitly refreshes them.
• Use ../../../.agents/references/language-policy.md for reply language while keeping commands, package names, keys, and paths in English.

Execution Rule

Follow the local prompt for environment detection and CLAUDE.md synchronization.