guava-router

One yes per step. Before each numbered step below, show its whole plan in your message text: every command, edit, install and setting change, word for word. Wait for one yes, then do the whole step and report what you did. Put the plan in your message text, never only inside a question or approval card, since some apps don't show what's in the card. Never ask for a yes to a change the owner hasn't seen.

Model Router: setup

These steps send each kind of work to the model that's best at it, through the Claude plan, ChatGPT plan or both that the owner already pays for: thinking and reviewing to a head-chef, and heavy lifting to a line-cook. With both plans, the head-chef runs in Claude Code and the line-cook in Codex; with one, both run in that plan's tool.

Start with the guava-hub skill. It builds the hub first and sends you here from Add a bot. The router needs a computer to run Claude Code and Codex, so it runs on a bot such as Grok Bot or Muse. In the Claude or ChatGPT app, tell the owner the router comes with adding a bot, and stop.

1. The models

Pick the row for the plans the owner pays for (hub setup asks; otherwise ask):

<!-- models:plans -->

<!-- /models -->

<!-- models:commands -->

<!-- /models -->

The picks come from models.yaml, the one place they're kept; don't edit the table by hand, since scripts/sync-models.py rewrites it. Add < /dev/null so neither CLI waits for input. The owner can pick their own for either slot: any Claude model through Claude Code or any OpenAI model through Codex, on a plan they have, with the same commands and that model's ID and effort, marked "(own pick)" in guava-stack settings so "check for newer models" leaves it alone. Guava supports no other setup yet: if the owner asks for another provider, tool or agent, say it isn't supported.

2. Install the command-line tools

Only the tools the owner's plans need: Claude Code for a Claude plan, Codex for a ChatGPT plan.

• Check first, without asking (these only read): claude --version and codex --version. Skip the install for any tool that answers with a version, and tell the owner it's already there.
• Install only what's missing: Claude Code npm i -g @anthropic-ai/claude-code and Codex npm i -g @openai/codex. Then bring both up to date, even when they were already installed: claude update and npm i -g @openai/codex@latest. An older Claude Code or Codex rejects the newest models (Codex 0.149 answered that GPT-6.1 Sol "is not supported when using Codex with a ChatGPT account"; 0.163 runs it). Check both versions again.
• Put the tools on the computer's PATH, so claude and codex also work in its desktop terminal if the owner ever signs in there: find their folder with dirname "$(command -v claude)" (or codex), often ~/.local/bin. If ~/.bashrc or ~/.profile doesn't already add that folder, back the file up (see Backups below) and add export PATH="<folder>:$PATH" at the end. Until a new terminal picks it up, give the owner the full path, such as ~/.local/bin/claude auth login.
• Check sign-in: claude auth status and codex login status. Skip any that's already signed in; otherwise run the sign-in yourself, on the owner's own subscription, so they only open a link and send you back what it shows:
• Codex: run codex login --device-auth in the background, send the owner its link and one-time code, and check codex login status once they say they've approved it.
• Claude Code: claude auth login waits for a code typed into its own terminal, so start it in a tmux session: tmux new-session -d -s claude-login 'claude auth login'. Read its link with tmux capture-pane -p -J -t claude-login, send it to the owner, and when they paste the code back to you, type it in with tmux send-keys -t claude-login '<code>' Enter. Check claude auth status, then tmux kill-session -t claude-login. Without tmux, give the owner the command with its full path to run in the computer's desktop terminal instead.
• Set each slot's model and effort as step 1 shows.
• On the bot's own computer: run every step here in Grok Bot's or Muse's own terminal, following the bot's adapter file. The owner opens each sign-in link and signs in with their own account, then pastes any code straight back into the bot.

3. Connect them to Notion

For each tool step 2 set up, check what's already connected first (claude mcp list, codex mcp list) and skip anything that already shows Notion.

• Claude Code: run claude mcp add --transport http --scope user notion https://mcp.notion.com/mcp, then have the owner type /mcp in Claude Code and approve Notion's sign-in page.
• Codex: run codex mcp add notion --url https://mcp.notion.com/mcp, then codex mcp login notion and have the owner approve Notion's sign-in page. On a bot's computer, the owner's browser then lands on an error page at 127.0.0.1:<port>, because that address belongs to the bot's computer: tell them so before they approve, and ask them to paste the whole address from the address bar back to you so you can finish the sign-in.

Test each tool with a quick non-interactive run that reads the context page from the hub, for example claude -p "..." < /dev/null and codex exec --skip-git-repo-check "..." < /dev/null.

The bot loads the hub and puts what each job needs into the tool's prompt (the hand-off rule in step 5), so neither ~/.claude/CLAUDE.md nor ~/.codex/AGENTS.md needs a Guava line.

<!-- guava-block: keep this paragraph the same in the guava-hub and guava-router skills --> Backups: whenever a step says to back up a file, copy it to <file>.bak-<local date and time>, such as settings.json.bak-2026-10-15-2122. The first time Guava changes a file, also copy it to <file>.pre-guava, or create that copy empty if the file doesn't exist yet, and never overwrite it: uninstalling puts the file back from it. Entries the plugin commands wrote before it, such as enabledPlugins, can already be in it; the uninstall commands remove those themselves. <!-- /guava-block -->

4. Let the tools read the hub

The bot runs Claude Code and Codex with nobody at the keyboard, and Claude Code then skips any tool that needs approval, so it can't read the hub. Pre-approve Notion's read-only tools, as part of the plan the owner says yes to:

• Back up ~/.claude/settings.json first (see Backups above).
• Under permissions.allow, add only notion-get-enabled-skills, notion-fetch, notion-search and notion-download-skill, with the exact prefix claude mcp list shows (for example mcp__claude_ai_Notion__notion-fetch). Never pre-approve write or delete tools.
• If Codex's test run in step 3 couldn't read the hub, approve the same four tools for Codex.
• Test it with a non-interactive run that reads the context page.

5. Model routing

• head-chef: orchestration, research, planning, brainstorming, copywriting, validation, review, verification, fuzzy and hard problems. For important work, it also checks the line-cook's result before anything is called done.
• line-cook: bulk execution, coding, sub-agent tasks, worker threads, long-running jobs.
• Browser and desktop tasks: keep using the assistant's own computer tools.
• Hand-off rule: if a tool can't reach the hub, the head-chef (or you) loads what the job needs from the hub and puts it in that tool's prompt.
• If either fails: do the task yourself and say so.
• If you are the head-chef or the line-cook (say, Claude Code on the head-chef's model), do that slot's work yourself and send only the other slot's work out. On one plan, the other slot's work goes to the same tool on the other slot's model.
• Grok Bot and Muse: run Claude Code, Codex or both on their own computers and hand each job to the CLI its slot names.
• Long jobs on Grok Bot: any hand-off that may take more than a few minutes runs in the background, as Long jobs below says.
• Skills with their own setup come first: when a skill defines how it runs (its own engine, scripts, models or setup), follow it as designed and use this routing only for work it leaves open. Never change a skill's own setup to fit the routing.

Save these routing rules as a standing instruction if hub setup hasn't already, pointing at the head-chef and line-cook in guava-stack settings instead of naming models, so the owner can swap a model by editing guava-stack settings. On Grok Bot, save them to shared user memory too, so every current and future bot follows them.

Long jobs

On Grok Bot, one command can wait at most 13 minutes, and the computer pauses between the bot's turns, so a CLI still running when the turn ends stops until the next one. For any hand-off that may take more than a few minutes:

1. Start it in the background, with its answer and log under /workspace/guava-runs/ (/workspace survives a computer recovery). Pick a short job name, and for Claude Code a new session ID (cat /proc/sys/kernel/random/uuid):

• Codex: mkdir -p /workspace/guava-runs && nohup codex exec -m <id> -c model_reasoning_effort=<effort> --skip-git-repo-check -o /workspace/guava-runs/<job>.md "<task>" < /dev/null > /workspace/guava-runs/<job>.log 2>&1 &
• Claude Code: mkdir -p /workspace/guava-runs && nohup claude -p --model <id> --effort <effort> --session-id <uuid> "<task>" < /dev/null > /workspace/guava-runs/<job>.md 2> /workspace/guava-runs/<job>.log &

1. Stay in your turn until it finishes. Check every few minutes with short commands, such as sleep 240; tail -n 20 /workspace/guava-runs/<job>.log, each well under the command limit. Tell the owner it's still running, but never end your turn while a job runs.
2. If it stopped without an answer, resume it instead of starting over, in the background the same way: Codex codex exec resume <session id> -o /workspace/guava-runs/<job>.md "Continue where you left off." < /dev/null, with the session ID from the top of <job>.log; Claude Code claude -p --resume <uuid> "Continue where you left off." < /dev/null. If it fails again, do the task yourself and say so.
3. Split big jobs. The head-chef breaks work that could run for hours into steps the line-cook can finish one at a time, and checks each before the next starts.

6. Stay inside each provider's terms

For Claude Code on a Claude plan:

• Use the official Claude Code CLI with your own login only.
• Never share tokens or route other people's requests through your plan.
• Keep Claude work tied to tasks you start. No unattended loops or 24/7 jobs on the plan; send those to the line-cook in Codex, or on Claude only, to an Anthropic API key.
• Stop when you hit usage limits.
• Never use outputs to train competing models.
• For heavy or always-on Claude automation, use an Anthropic API key.

For Codex, follow OpenAI's current terms for the owner's plan.

7. Final checks

• Check the current terms for each provider the owner uses.
• Test each slot with a quick prompt and tell the owner which model actually answered. Don't call setup done until both slots pass.
• Running this as part of hub setup? Go back to the hub setup skill's founder-mode step.

Changing the setup later

To swap models, or apply a hand-edited guava-stack settings section, follow Change your setup in the guava-hub skill. "Check for newer models" goes there too.

Update

Use this when the owner says "update guava-router", or when the guava-hub skill's Update step sends you here. Show the whole plan first, in your message text, as one numbered list (each command, and each change as old, then new); the owner can approve it with one yes or pick numbers. Then apply it in order.

1. Get the newest files (skip this when the guava-hub Update step already did it):

<!-- update-steps: keep this list the same in the guava-hub, guava-router and founder-mode skills --> Only for the Guava plugins the owner has (for "update guava-router", just guava-router):

• Claude Code: claude plugin marketplace update guava, then claude plugin update <plugin>@guava for each plugin. Updates apply in a new session.
• Codex: codex plugin marketplace upgrade guava, then codex plugin add <plugin>@guava for each plugin. Updates apply in a new chat.
• Grok Bot: npx skills update -g -y.
• Muse, and the Claude and ChatGPT apps: no skill files, so nothing to download.

<!-- /update-steps -->

1. Check for newer models: follow Change your setup in the guava-hub skill with "check for newer models".
2. Check the routing still runs: each tool the owner's plans need is installed and signed in on the bot's computer (step 2 of setup, checking first).
3. Test each slot with a quick prompt, as in step 7, and confirm in two or three lines: guava-router's version before and after, any model change, and anything left for the owner.

Uninstall

Use this when the owner says "uninstall guava-router", or when the guava-hub skill's Uninstall step sends you here. It stops the routing and puts back what router setup changed. Show the whole plan first, in your message text, as one numbered list (each command, each file change as old, then new, and each app step); the owner can approve it with one yes or pick numbers. Then apply it in order.

<!-- guava-restore: keep this paragraph the same in the guava-hub and guava-router skills --> Putting files back (~/.claude/settings.json and ~/.codex/config.toml): <file>.pre-guava next to each file shows how it was before Guava (see Backups in setup step 2 of the guava-hub skill). Run the plugin commands first, since they rewrite these files and remove their own entries. Then, for only the settings the Uninstall step lists, put each one back to its value in the .pre-guava copy, or remove it if that copy doesn't have it; if the owner has changed it since Guava set it, ask instead. Drop any section left empty that the copy doesn't have, such as an empty enabledPlugins. Never copy a backup over the file, since that would undo every change the owner made since, and never change anything else. With no .pre-guava copy, remove only what Guava added and ask about any value it may have replaced. Leave the .pre-guava and .bak- copies where they are and list them in the wrap-up, so the owner can delete them once they're happy. <!-- /guava-restore -->

1. Remove the plugin (skip this when the guava-hub Uninstall step already did it). When uninstalling guava-router on its own, first check what's there without asking, as step 1 of the guava-hub Uninstall step says:

<!-- uninstall-steps: keep this list the same in the guava-hub, guava-router and founder-mode skills --> Only for the Guava plugins the owner has (for "uninstall guava-router", just guava-router; for "uninstall guava-stack", every one installed):

• Claude Code: claude plugin uninstall <plugin>@guava for each plugin. Once no Guava plugin is left, claude plugin marketplace remove guava, then delete ~/.claude/plugins/cache/guava if it's still there (if your safety check won't let you delete a folder in these steps, move it into ~/guava-uninstalled/ instead and list it in the wrap-up). It applies in a new session.
• Codex: codex plugin remove <plugin>@guava for each plugin. Once no Guava plugin is left, codex plugin marketplace remove guava, then delete ~/.codex/plugins/cache/guava if it's still there, and each ~/.codex/.tmp/marketplaces/.staging/marketplace-upgrade-* folder holding a copy of guava-stack, which marketplace upgrades leave behind. It applies in a new chat.
• Grok Bot: npx skills remove -g -s <plugin> -y for each skill, and check with npx skills list -g.
• Muse, and the Claude and ChatGPT apps: no skill files, so nothing to delete.

<!-- /uninstall-steps -->

1. Put the computer back, following Putting files back above:

• ~/.claude/settings.json: model, if router setup set it. Say which model new sessions will use instead. The hub's read-only tools under permissions.allow stay while the hub does: guava-hub's Uninstall step removes them.
• ~/.codex/config.toml: model, if router setup set it, saying which model new chats will use instead.

1. Take routing out of Grok Bot and Muse: on Grok Bot, delete the routing memories, shared user memory included; on Muse, delete the routing line at the end of SOUL.md. Show the exact text to delete and wait for done.
2. Turn routing off in the hub (skip this when uninstalling the whole stack, since the hub then stays as it is): in guava-stack settings, remove the guava-router heading, with the head-chef and line-cook lines, their model lines and sub-bullets and, in Notion, its images. With no head-chef or line-cook named, the guava-hub start-up hook and the standing instructions do the work in one place, with no hand-offs. Add a dated line to the hub's changelog page.
3. Offer the rest as their own numbered items, left out of a plain yes, since they're the owner's own tools and other work may use them: the Notion MCP servers router setup added (claude mcp remove notion -s user, codex mcp remove notion), and the CLIs it installed, with their sign-ins (codex logout and npm rm -g @openai/codex; claude auth logout and npm rm -g @anthropic-ai/claude-code), together with the export PATH line router setup added to ~/.bashrc and ~/.profile. Never remove the tool this is running in.
4. Check and wrap up: claude plugin list and codex plugin list (or npx skills list -g) no longer show guava-router, and nothing from step 2 is left in those files. With the hub kept (not for the whole stack), run claude -p --permission-mode plan "Which head-chef and line-cook are in my guava-stack settings?" < /dev/null: it should find none. Confirm in two or three lines what was removed and what stayed, with the backups' paths.