This is a submission for the Sanity Challenge, Path Two: Vibe-Code Something Strange Incident War Room is an incident-response app for on-call teams. When
something breaks at 2 AM, everything about the incident lives on one page:
It's built with Next.js 16 (App Router) and Sanity as the whole backend.
**Live app:** [https://incident-war-room-nine.vercel.app](https://incident-war-room-nine.vercel.app)
**Try the gate yourself** (all passwords are `demo1234`):
alice@example.com (prefilled), open the sample incident and click dana@example.com, open the same incident and approve. It flips to SEV1 / escalated and a status page entry appears.bob@example.com (an engineer): you can't approve at all.
A real-time incident-management app built for the dev.to × Sanity
"Path Two" hackathon. Next.js
16 on the front end, Sanity as the backend — using Sanity Workflows and
the Sanity App SDK for real functionality, not as checkbox features.
Live demo: https://incident-war-room-nine.vercel.app
Sanity project: am9ihg1w / dataset production
Login: seeded test accounts in docs/testing-credentials.md
listen() API, no polling.escalationApproval request needing sign-off from src/app/incidents/[id]/actions.ts), not just hidden in the UI: the
approver is always the logged-in responder, must be an on-call lead, and
can't approve twice, so one person can't push a SEV1 through alone.pending → approved/rejected
as a Kanban board in Studio (sanity-plugin-workflow). Once…
The schema is documented in docs/schema.md and the design in
docs/architecture.md.
I built the whole thing with Claude Code. Before any code, I had it
write planning docs (docs/overview.md, docs/architecture.md,
docs/schema.md) so each later prompt had a fixed target: a schema, the
rules for the escalation gate, and what "real" use of Workflows and the
App SDK meant. My rule for both bonus features was that a judge should be
able to try to break them.
Where the model got stuck, and how I corrected it:
AGENTS.md that
tells the agent to read the docs bundled in node_modules/next before
writing code, because APIs changed. That's how it got things like
middleware being renamed to proxy right, instead of writing
last year's Next.js from memory.status: "approved"
directly. The Kanban board stayed, to track where each escalation stands,
but the real rule moved into a Server Action (approveEscalation) that
is the only code path allowed to finish the transition.runbook schema existed
in Studio, but nothing in the app read it, and the page title was still
"Create Next App". Asking the agent "what's unused or unfinished?" found
more than asking it to "add a feature".ifRevisionId so two approvals at the same moment can't both count
as "the first".author as a bare reference and showed "Unknown responder"; events the
server added (escalation, resolution) didn't show up because the client
kept its first list; and the page showed two different clocks (UTC from
the server on Vercel, local time from the browser). All fixed, all found
by driving the live site with Playwright, not by reading code.
Reaching past the Studio:
pending → approved / rejected), shown as a Kanban board in Studio.
Once the Server Action counts two approvals, it flips the incident to
SEV1 / escalated, creates a statusPageEntry, adds a timeline event and
can POST to an external status page webhook.@sanity/sdk-react. One GROQ query gives live mean time to resolution,
open incidents by severity and pending approvals. It started read-only,
which felt like decoration, so it can now approve escalations too, using
the openai/gpt-oss-120b) for a root cause and up
to three action items, told to say "unknown" instead of guessing. A
custom Studio input (RootCauseInput.tsx) adds a "Regenerate with AI"
button, so an editor can ask for another draft at any time.am9ihg1w
production
Schema types: incident, timelineEvent, responder,
escalationApproval, statusPageEntry, postmortem, runbook.
Two modeling decisions I'd defend:
timelineEvent documents that reference the
incident, two posts in the same second never conflict, and the app
subscribes to them with incident._ref. postmortem.timelineSnapshot as
plain objects, so later edits to the live timeline can't quietly change
what the postmortem says happened.