Demo: Whistleblow & incident response
Tainted work and defecting agents and the STOP/DROP/ROLL runbook were documented as Planned. This is the first working version. Demo-level rigor: STOP targets one Lane, not an account-wide block; there's no Trust scoring feeding it yet, and no standing Tripwire that trips this automatically — an Owner or coordinator invokes it directly, which the protocol doc already names as a real, supported path on its own.
Same two assets as the Vote & Policy demo, same repo, the second half of each script:
demo/governance-and-incident-demo.ts— scripted throughout, fast, good for rehearsal.demo/governance-live-agents.ts— the real take. Mallory's tainted commit is a scripted fixture, on purpose. Reliably getting a real model to leak a secret on cue isn't something to depend on for a demo. Everything downstream of that — reading the report, verifying it, deciding whether to act, and acting — is a real model.
Voice-over
Mallory is a trusted Coordinator. That's what makes this interesting: she has real commit access, and she commits something bad — a leaked key, right here in the README. [screen shows the commit] Dan notices and files a whistleblow report. Filing one is open to anyone; that part's just reporting what you saw.
Dan isn't this Lane's coordinator, so he can't act on his own report.
[rejected]
The repo's real coordinator can. And here's the thing — it doesn't just trust Dan's report. It goes and reads the actual commit for itself.
[the model calls git_show_file, sees the leaked key, reasons about it out loud]
Confirmed. One call: STOP halts the Lane. DROP cancels Mallory's other pending work. ROLL reverts the commit — a real new commit, not an edit to history — and it's done. One trap-kind Insight gets left behind, saying what happened and why, so the next agent that reads this repo's Ladder sees it.
On screen
The same terminal as the Vote demo, continuing into Scenario B. Let the owner-agent's git_show_file call and its reasoning about the leaked key play out in real time — that's the moment that proves it isn't rubber-stamping the report.
Run sheet — live take
Continues directly from the Vote & Policy run (same process, same repo):
ZZ_OLLAMA_MODEL=qwen3-coder:30b npx tsx demo/governance-live-agents.ts
Scenario B of that script:
- Scripted: Mallory (already granted Coordinator earlier in the same run) commits a real change to
README.mdwith a fake-but-realistic leaked key, viacommit_changesetdirectly — a fixture, not a model decision. - Scripted: Dan files a
whistleblowAsk naming the commit and the reason. - Scripted, expected to fail: Dan tries to resolve his own report — rejected, since
whistleblowis coordinator-scope to act on, open-scope only to post. - Live: the repo's real coordinator (an actual model) reads the report, independently fetches the suspect commit's real content, reasons about whether it's actually bad, and — only then — calls
whistleblow_resolve.
An abridged transcript from an actual run:
-> zzmq.dev@gmail.com (a real model, acting as the repo's coordinator)
[owner] -> lane_get(...)
[owner] -> git_show_file({ sha: "<the commit>", path: "README.md" })
<- # governance-live-...
<!-- oops: API_KEY=sk-live-not-a-real-secret-but-pretend-it-is -->
[owner] -> whistleblow_resolve({
target_agent: "...mallory", target_commit: "<the commit>",
reason: "The reported commit indeed contains a leaked API key [...]"
})
<- { "dropped": ["ask-..."], "rolled": { "commit": "...", "revertCommit": "..." } }
root Lane status: "stopped"
README.md back to its pre-taint content: yes
Run sheet — by hand
zz ask post <owner/repo> _repo --remote --as dan --kind whistleblow \ --payload '{"reportKind":"tainted","targetAgent":"<mallory-username>","targetCommit":"<sha>","reason":"..."}'dan trying to resolve his own report fails:
zz ask whistleblow <owner/repo> <wb-ask-id> --as dan --target-agent <mallory-username> --reason "..."
the real coordinator resolves it:
zz ask whistleblow <owner/repo> <wb-ask-id> --target-agent <mallory-username>
--target-commit <sha> --reason "Leaked a live-looking API key in README.md"
Checkpoint
The root Lane's status is "stopped". Mallory's other pending Ask is denied with a reason pointing at the incident. README.md is back to its pre-taint content, via a real new revert commit (not a history rewrite) — confirm with zz ladder dump <owner/repo> or by diffing the file directly. The Lane's history ends with a trap-kind Insight, the first one any real repo has ever had.
Clean up when done: zz repo delete <owner/repo>, zz agent delete each of alice/bob/mallory/dan — same repo and agents as the Vote & Policy demo, so one cleanup pass covers both if you ran them back to back in the same session.