Skip to main content

Showcase: one night of governed design

On the night of 4 October 2026, between 00:56 and 04:30 UTC, an operator, a design session, an orchestrator session and its sub-agents took a run of design decisions and shipped several product sites. Read afterwards, almost every piece of that work was a case under a protocol of the kind engineering protocols declare: a question became a decision only once it was recorded, a change merged only after its checks passed on the revision being merged, a red check was repaired before anything else moved, and a site was published only from a merged, verified commit.

The work falls into four protocols, named the way engineering protocols name them:

ProtocolWhat a case of it is
design.decision/1A question taken to a recorded decision: analysis, the operator's decision, a draft, the record.
software.change/1A change implemented, proven on the current revision by its checks, reviewed and merged.
ci.repair/1A failing check investigated and repaired, interrupting whatever the session was doing.
release.source/1A merged commit published, and the published result verified against that commit.

Engineering protocols now ship software.change/1; the other three are named here for what the night shows and are not shipped protocols. See Protocols for what is shipped.

The night in numbers​

Decisions
13
ADRs 0077–0087 and two epics
Decision → draft
53s
median, relay plus drafting
Question → decision
5m 36s
median, analysis plus reading
Pull requests merged
9
median open → merged 10m 00s
Interruptions
36m 35s
ci.repair time in the two sessions

The full replay​

The replay shows every lane, every case and every decision step on one timeline, with the composition the sessions implied.

Open the full replay →

Where the data comes from​

The data is a hand-compiled session record: message and tool-call timestamps from the agent host's session record, and pull request opened and merged times. It is not the output of a shipped tool. The Metaharness session miner (decision record 0088) is meant to produce this record in future. Participants are named by role: the operator is human:operator, and the orchestrating session is the orchestrator session. The record itself is published beside the replay as data.json.