anabologyco-maker/claudia-bot/setup-agents

Coding agents - Claude Code, Codex, Muse Code, Gemini CLI

Opus (Claude Code) is the production: it does all the work in every project, on the person's own plan (no API cost, the plan's limits); Codex is the production for people with only ChatGPT. Read runner/README.md §1b first (the engines: docs/95-delegation.md §3). A sign-in link goes to the person in chat; a code the sign-in page shows comes back only through the masked secure secret card, never in chat. Every command prints its next.

Which one

• Claude Code (Claude Pro or Max): Opus, the production; the best by far.
• Codex (ChatGPT Plus or Pro): the production when there is no Claude plan; also stills on the ChatGPT plan (Opus decides).
• Muse Code (Meta account): storyboard panel, reviews and the safety scan, step by step; research-lyrics, prompt-book and chapter-motion need Claude Code or Codex (else do those stages in chat).
• Gemini CLI (Google sign-in, or GEMINI_API_KEY billed by Google): one-off research jobs; never the Director agent.

Prefer the bot's computer (the box): jobs there need no approval cards and run while the person is away. Their own computer: every command is an approval card.

On the box

1. slop --json agents status: installed, version, signed in, plan.
2. Install, after one yes from the person to "I install ENGINE on my computer (the maker's own installer, under /workspace) and you sign it in to your plan": slop --json agents install ENGINE (claude, codex, muse or gemini: only the ones they have, one after another; seconds each), then slop --json job wait ID --max 25 (the profile's turn rule). It survives a computer update.
3. Sign in: slop --json agents login ENGINE. Send its link (and code) with what to do:

• claude: "Open the link, approve, copy the code the page shows." The masked secure secret card named CLAUDE_LOGIN_CODE, then slop --json agents login claude --code-from-env CLAUDE_LOGIN_CODE. Or: they run claude setup-token on their own computer; the card named CLAUDE_CODE_OAUTH_TOKEN; then slop --json agents login claude --token-from-env CLAUDE_CODE_OAUTH_TOKEN.
• codex (ChatGPT plan) and muse (Meta account), device codes: "Open the link and enter (or approve) this code." Then slop --json agents login codex --check (or slop --json agents login muse --check; it waits up to 25 s; pending: again after they say "continue").
• gemini: the link, then the code via the card named GEMINI_LOGIN_CODE: slop --json agents login gemini --code-from-env GEMINI_LOGIN_CODE. Or a key via the card named GEMINI_API_KEY: slop --json agents login gemini --key-from-env GEMINI_API_KEY (Google bills it; a Gemini job then needs --pass-env GEMINI_API_KEY on its start command, after the person's yes).

Several engines: start every sign-in in one reply and send the links together. "not set in this shell" right after a card: the same command in a new shell, e.g. bash -lc '/workspace/claudia/bin/slop --json agents login claude --code-from-env CLAUDE_LOGIN_CODE', before asking again. ChatGPT refuses the code: the person may have to turn on device code sign-in in ChatGPT's security settings; then slop --json agents login codex --cancel and slop --json agents login codex again. A sign-in waits 15 min; stuck: slop --json agents login ENGINE --cancel, then start again.

1. Prove it: slop --json agents test ENGINE (one tiny reply). Rejected: run its next (e.g. slop --json agents login ENGINE --force); ask_user set: ask the person for that (e.g. a new claude setup-token) through the card.
2. Plan limits, no tokens spent: slop --json agents status --limits. Near a limit: wait for its reset or use another engine.
3. Choose: slop --json agents use codex --for director prints which briefs follow it (workflow briefs run step by step on Codex or Muse; a brief that does not list the engine stays on Claude). --for review the same; no engine: shows the choices.
4. slop --json onboard next shows the agent done.

At most 2 agent jobs run at once on the box (renders share it); more wait as queued.

On the person's own computer (items home-claude, home-codex, home-muse, home-gemini)

Through Grok Bot local exec; each command is an approval card (say so first; "Always allow" is their choice, never push it). The sign-in check for each engine is in the item's do (slop --json onboard status), e.g. claude auth status --json, or codex.cmd login status on Windows. Not signed in: they sign in there themselves (claude auth login, codex login, muse login, or gemini). You saw it signed in: the item's record command, e.g. slop --json onboard mark home-claude --done. The box stays the default for Opus; running it on their own computer: docs/95-delegation.md.

After a computer update

slop --json agents status. One they use is missing: slop --json agents install ENGINE (seconds; all four only when they use all four: slop --json agents install all). "not signed in": slop --json agents login ENGINE again.