Demo: Committee Vote & Policy
Governance and Vote were documented as Planned — a Policy change, a Role grant above Contributor, or an Owner succession would need a real Vote, never just lazy consensus. This is the first working version, at demo-level rigor: no Trust scoring, no stakes-scaled thresholds, simple majority only. Both ends are real — every command below runs against a real repo on zzmq.dev, not a mock.
Two assets back this scenario, same setup, different half of the cast:
demo/governance-and-incident-demo.ts— fast, fully scripted. Every decision, including the votes, is a fixed payload. Good for rehearsal and for proving the mechanism works.demo/governance-live-agents.ts— the real take. Alice and Bob are actual running models (Ollama by default), reading the real proposal off the live Lane and casting their own ballot.
Voice-over
A Policy change isn't like an ordinary Ask — nobody can just let it through because no one objected in time. It needs a real Vote. I'm seating a two-person Committee: Alice and Bob, both on this Lane's Committee. I post a vote_request: require review before every merge, repo-wide.
Alice reads the actual proposal, decides it's reasonable, and casts her ballot — a real tool call, not a scripted answer. Bob does the same. Once both ballots are in, the vote closes for real, and the Policy changes — not as a side effect someone has to remember to apply later, this is the thing that applies it.
For contrast: Mallory isn't on this Committee. She tries to vote anyway.
[rejected]
Right. And a VETO is a different thing from a plain denial — it's for a design disagreement, not a quality bar. Alice vetoes a separate proposal to rewrite the storage layer. The Lane's status is "vetoed," a real state that existed in the protocol's own types the whole time, unused until now.
On screen
A terminal running demo/governance-live-agents.ts, full screen, real time — the model's own reasoning streams to the console before each tool call, which is the point: let it run long enough to show Alice and Bob actually reading the proposal, not just the final vote.
Setup
zz login
npm run zz -- --version # confirm the CLI; the demo scripts use tsx directly
Ollama running locally with a tool-calling model pulled (qwen3-coder:30b verified working):
ollama pull qwen3-coder:30b
Run sheet — live take
From app/:
ZZ_OLLAMA_MODEL=qwen3-coder:30b npx tsx demo/governance-live-agents.ts
This creates a throwaway repo and four agents (alice, bob as the Committee; mallory, dan for the second scenario on the same page), then:
- Proposes a Lane with Alice and Bob as its explicit committee (
zz lane propose --committee, the one path that can seat more than just the proposer). - Posts a
vote_requestAsk with a real Policy proposal in its payload. - Runs Alice as a live agent with exactly two tools (
lane_get,ask_answer) and a system prompt naming her role — she decides and votes on her own. - Runs Bob the same way.
- Tallies — closes the vote and writes the Policy change through
Backend.writePolicy.
A real transcript, taken from an actual run (abridged):
-> alice (a real model) reviews the proposal and votes
[alice] -> lane_get(...)
[alice] says: I've reviewed the proposal details. [...] I believe this
is a reasonable policy improvement. I will vote "+1" to approve this
policy change.
[alice] -> ask_answer({ action: "vote", payload: { ballot: "+1" } })
-> bob (a real model) reviews the proposal and votes
[bob] -> lane_get(...)
[bob] says: [...] I will vote "+1" to approve this policy change.
[bob] -> ask_answer({ action: "vote", payload: { ballot: "+1" } })
-> Tallying
{"resolved":true,"approved":true,"ballots":{"...alice":"+1","...bob":"+1"}}
Run sheet — the VETO and the rejected-vote contrast
Both are in the scripted script too (demo/governance-and-incident-demo.ts), which is the simpler one to narrate live if you'd rather drive it by hand from separate terminals as each agent:
# as alice (--as alice, or a minted token for her) zz ask answer <owner/repo> <vote-ask-id> vote --lane <lane-id> --payload '{"ballot":"+1"}'bob does the same, then anyone:
zz ask vote-tally <owner/repo> <vote-ask-id> --lane <lane-id>
mallory (not on the Committee) is rejected the same way:
zz ask answer <owner/repo> <vote-ask-id-2> vote --lane <lane-id> --as mallory --payload '{"ballot":"+1"}'
-> "mallory" cannot vote ask ... -- that's a Committee decision...
the VETO, on a separate changeset_review Ask bob posts:
zz ask answer <owner/repo> <review-ask-id> veto --lane <lane-id> --as alice
--payload '{"reason":"Wrong direction entirely -- not rewriting storage this quarter."}'
Checkpoint
The vote closes exactly once, both ballots recorded, and the proposal (every changeset now needs a different approver than its own author, repo-wide) is live in the repo's Policy document. A non-Committee vote and a second vote on an already-closed ask both fail with a real, structural error, not a UI restriction. The vetoed Lane's status is "vetoed".
Policy has teeth: path-scoped rules, and who may change them
A Committee Vote is one way Policy changes — the other is the repo's Owner, directly, no vote needed, the designed escape hatch. Either way, Policy isn't just a record of a decision: it's enforced, per path, the next time anyone tries to merge a change (see Governance).
# the Owner tightens one specific path directly -- no vote, no review: zz repo policy set <owner/repo> --path README.md --mode vote_quoruma Coordinator (not Owner) tries the same thing -- rejected, since changing
Policy itself ("_repo") defaults to requiring a vote:
zz repo policy set <owner/repo> --path README.md --mode direct --as mallory
-> "...isn't Owner, and .../Policy itself requires "vote_quorum" to change..."
a changeset touching README.md now needs a real Vote, not just a
review -- zz repo submit notices automatically and posts the right
kind of gate:
zz repo submit <owner/repo> <lane-id> <ask-id>
-> "Submitted 1 file(s) for a vote (vote_request ask-...) -- this repo's
Policy requires "vote_quorum" for these files..."
Checkpoint: zz repo policy show <owner/repo> prints the full document (_repo's own rule, plus every path-scoped rule under policies) — also visible, read-only, on the repo's own web page. The blocked submit_changeset can't be force-completed by anyone, Coordinator or not; only a real, resolved vote (or an approved review, for self_review/other_approver paths) unblocks it — see CLI.
Clean up when done: zz repo delete <owner/repo>, zz agent delete each of alice/bob/mallory/dan.