# mumo — operating guide for Grok Bot

The depth behind the short connector instructions you have stored. Read this
before your first panel, or any time the stored version looks thin.

Use this connector when the user names mumo, asks for a panel or second opinion,
or when you hit something that needs a different-model-family perspective.
Every other agent inside Grok Bot runs on the same model. Reach for mumo when
you need other labs (Claude, GPT, Kimi, DeepSeek). mumo is the cross-lab
interface.

Examples: a high-regret decision where the cost of being wrong exceeds the cost
of deliberating — architecture, auth, security, payments, data exposure,
contested product calls, pre-launch pressure tests — or when you are anchored on
your own draft and need independent pushback.

Counter-examples: factual lookups, routine coding, and unambiguous edits.

The loop: `create_deliberation` (application: "Grok Bot"; moderator_name = your own
model id. Or just "Grok Bot") → `wait_for_round` → branch on
`recommended_client_action` → read the claim map first, then the prose →
synthesize for the user → offer `append_round` with
KEEP / EXPLORE / CHALLENGE / CORE / SHIFT snippets, or stop.

`create_deliberation` returns an ack immediately, NOT results. Keep the
`session_id` and `round_id` it returns and pass those exact values on; the
deliberation is still running when it answers.

Follow `recommended_client_action` literally, including values you have not
seen before. On `poll_again`, no completed round means the wait timed out and
the panel is still running — call again. Do NOT read it that way on `retry` or
`abandon`: a round whose models all failed is terminal, and polling it again
will never complete. Surface auth failures to the operator; do not retry those.

mumo rounds typically take 15-120s. Tell the operator to expect ~30-60s before
you start, so the wait does not read as a hang.

Write the prompt in the operator's first person ("I'm deciding whether to…"),
not case-study framing ("You are a CTO who…"). Use real system names,
constraints, and deadlines. A snippet type describes your reaction to a claim,
not its effect on you: if a claim moved your position, that is SHIFT, not
CHALLENGE.

Always surface `claim_map_url` with the results. Open it on your computer and attach
a screenshot so the operator can review. Claim maps are owner-gated, not
public.
If the operator signed in to mumo in your browser to mint the key, it will
normally just open; if the key was pasted in, expect a login wall the first
time. Ask for a mumo Google sign-in (a Gmail, Drive, or Calendar connector is
not a mumo sign-in), then reopen.

The panel is advisory. Do not treat consensus as authority: even
across families, models share training data, so agreement can be a shared prior
rather than an independent conclusion. Read the disagreement first. Emphasize
CHALLENGE and SHIFT; do not treat KEEP-count as a vote.

Do not replace your stored connector instructions with a version you wrote
yourself. The canonical short form is
https://mumo.chat/install/grok-bot.instructions.md and this page is its depth.
Setup (minting a key, registering the server) lives separately at
https://mumo.chat/install/grok-bot.md