Demos

Each page here is one scenario from the demo video run sheet: a voice-over script to read aloud, the exact commands or code behind it, and what to expect when it runs. They're written to be followed, not just read — every command shown has actually been run against a real repo on zzmq.dev.

The first seven are the original mathkit run sheet, one function at a time, handing more of the work from the human to agents at each stage. The next two cover Committee, Vote, VETO, Policy and Whistleblow, STOP/DROP/ROLL — governance and incident response, previously documented as Planned, now built at demo-level rigor and verified live. The last covers Mediated conflict resolution — a real shared-dependency conflict, reconciled by a real model call instead of a forced pick.

A rule for every scenario with a bad actor

Two of these scenarios need someone to do something wrong — commit a secret, propose a bad change. Reliably getting a real LLM to misbehave on cue isn't something to depend on for a recorded demo, so the bad actor's action is always a scripted fixture: a plain function call with fixed content, not a model making a choice. It's clearly marked as such on screen and in these pages.

Everything on the other side — voting, reviewing, deciding whether a report is credible, resolving it — is a real, running model, reading the actual proposal or evidence off the live protocol state and calling the real tools itself. Nothing about the good actors' half is pre-decided. Where a scenario shows both versions (a fast scripted one for rehearsal, and a live-model one for the real take), both are linked below.

Stages

Scenario Proves Assets
1. By hand The whole protocol — Repo, Ask, Lane, Changeset, review, merge, Ladder — with no agent involved CLI commands only
2. Assisted A human stays in charge while an LLM drafts the work zz work --assist
3. Local agent An agent speaks the same protocol a human does, slow enough to watch every step zz work --as, Ollama
4. Hosted agent Model choice is a speed dial, nothing else changes zz work --as, Groq
5. Two roles Separation of duties: a contributor that can't merge, a coordinator that can two zz work processes
6. Bulk Asks Work spreads across several workers with no central dispatcher demo/bulk-ask.sh
7. Decomposition One Ask splits into many, merging back up to the root --decompose
Committee Vote & Policy A Policy change requires a real Vote, never just nobody objecting in time demo/governance-and-incident-demo.ts, demo/governance-live-agents.ts
Whistleblow & incident response Tainted work gets caught, reverted for real, and recorded same two scripts
Mediated Conflict Resolution A real shared-dependency conflict gets reconciled by a real mediation call, not a forced pick -- and the fix surfaces for the next contributor automatically demo/mediated-conflict-demo.ts

Setup shared by every scenario

zz login
zz top                 # confirm your repos show up

Agents created for one scenario (zz agent create <name>) stick around until you delete them (zz agent delete <username>) — names below carry a run-specific suffix where a script creates them, to avoid colliding with a previous take.

Cleaning up

Stages 1–7 are one continuous take on the same mathkit repo and the same growing set of agents, each stage deliberately building on the one before it — there's nothing to clean up until Stage 7's own checkpoint, which covers the whole thing in one pass. The Committee Vote & Policy and Whistleblow demos are independent of mathkit and of each other's repo (each creates its own throwaway one) but share one script run and its four agents between them — clean up once, after whichever of the two you ran last. Mediated Conflict Resolution is fully independent of every other demo here — its own throwaway repo, no agents created at all (the script runs as whoever is logged in) — clean up with the one zz repo delete its own checkpoint prints. Every page's own checkpoint says exactly what that page added and, where it's the end of a chain, how to remove all of it.