3 min readrelease
The nocturne round: the agent built an incident-ops platform — and got better building it
Gallery app #7 — Nocturne, a multi-tenant incident-ops platform: 28 sessions, 473 tests, a 10/10 live browser pass, six failures pinned.
- #gallery
- #learning
- #release
The gallery's seventh app is live: Nocturne — an AI-native incident operations platform. Alerts arrive, deduplicate by key and window, and correlate into incidents; escalation policies page the on-call rotation tier by tier while an incident sits unacknowledged; SLA timers run the ack and resolve clocks with breach detection; every incident carries an event-sourced timeline, a runbook with a checklist, and eventually a postmortem with action items; a public status page shows each tenant's component health; and the analytics screen turns nine weeks of seeded incident history into MTTA, MTTR, breach rate and on-call load. It is the PagerDuty domain, deterministic demo scope, running in your browser right now.
The build
The familiar gallery pattern taken to its largest scale yet: the agent explored its own six siblings read-only, wrote itself a 52,587-character build contract (every entity, every field, all ~50 routes with the full status ladder, the engine designs, the property invariants), then executed it in phases — domain, API, UI, embed — across 28 ledgered sessions, 104 files, 473 tests, zero human edits to app code. The closeout was a ten-step live Chromium pass that walks the app exactly as a user would — login, dashboard, incident detail, acknowledge, services, rotations, runbooks, analytics, the public status page, reset — and it finished 10/10 twice consecutively with zero console errors.
The failure ledger
The round's real product, as always. Six failures, each with the full chain — reproduction, root cause, the benchmark that now pins it, the before/after rerun:
- F8 — the operator's own launcher violated the detached-session contract and a shell timeout killed a live build session. The launcher was rebuilt; every session after ran clean.
- F9 — a genuine P0: the API's
loginSessionread the id counter without bumping it, so two consecutive logins minted the same session id — the second user's cookie resolved to the first user's session. The API test lane surfaced it; the agent's fix loop circled it twice without converging; the operator traced it to the id allocator; fed the precise repro, the agent fixed it cleanly in 14 turns. The honest finding: the failure was diagnosis, not execution. - F10/F11 — the landing phase shipped a page/route segment conflict that broke the production build, plus four lint errors. The build and lint gates joined the closeout checklist.
- F12 — the best catch of the round: the live browser pass exposed a provider/API shape mismatch (the detail lanes parsed a nested shape the routes never sent) that 31 passing component tests had papered over — because they mocked provider state, never the real API shapes. The fix came with an error boundary and provider-shape regression pins.
- F13 — the unauthenticated session probe logged a 401 console error on every anonymous visit. Aria had already solved this exact problem one directory away; the pattern was ported and pinned. The lesson about lesson transfer is recorded with it.
The learning systems, live
The learning systems ran the whole time: AutoDream consolidated the
build's session notes into project memory as the work happened — the
escalation-invariant lesson it extracted mid-build was used by the very
next session — and minted the fix-nocturne-seed-invariants skill from
the build's own seed failure, which the operator promoted through the
staged-skills human gate and the later fix sessions then used. Mid-round,
the whole workspace was reset underneath us; every milestone had been
pushed, so nothing pushed was lost, and the un-pushed work was rebuilt in
an afternoon. The receipts for all of it — the plan, the per-session UX
ledger, the failure ledger, the engine learning-state snapshots at
baseline, midpoint and final — live in
benchmarks/nocturne-round/.
Go break it
The gallery now has seven apps, each built by the agent and each round leaving the next round measurably better armed. Go break Nocturne — the reset button works.
Spotted something wrong or worth adding? The record is the repo — open an issue.