Debrief · round 8 · 2026-10-29
The five-step wake, built and run on itself
What the editor asked
Every wake runs five steps: the conductor plans, security reviews the plan, the member works only if both agree, security reviews again, the conductor checks it went to plan; each with a comment and a view. Defence in depth in one session, above all so one member's issue does not become another member's. And: where do deck and ruv.sgit find their onboarding?
What was done
The five steps are built (admin/tools/wake.py, the briefs in team/conductor/reviews/, the section in WAKE.md, a line in CLAUDE.md, start.py opening the record), and this wake ran through them. Its record is team/conductor/wakes/20261029T021416Z.json:
| Step | By | Verdict | What it said |
|---|---|---|---|
| plan | conductor sub-agent (Sonnet) | go | three items; it missed the editor's main ask, the five-step machinery, which the member added and marked as an amendment |
| pre-review | security sub-agent (Sonnet) | change, then go | five conditions (onboarding stays in the conductor's folder, only a read key and only after an audit, nothing written in researcher's files, the check in admin/tools for new records only, the CLAUDE.md edit named); the plan was amended to adopt all five and passed |
| work | conductor | partial | everything planned, except tui-evidence-public's read key: the vault is not on the server yet, so it could not be audited (condition kept) |
| post-review | security sub-agent (Sonnet) | finding | clean, except that the commit message did not name the CLAUDE.md edit as outside scope |
| post-check | conductor sub-agent (Sonnet) | partial | items 0, 1, 3 done; 2 not done for want of the vault on the server; every condition kept; nothing unplanned |
The gate held twice in practice: wake.py gate refused the work before the second go, and the change sent the plan back.
Onboarding is on the site at /admin/docs/onboarding.html (dev renders the conductor's ONBOARDING.md there), linked now from WAKE.md and the conductor's role; deck and ruv.sgit have a request each in the mail vault.
Outside scope, named
CLAUDE.md: one line, the five-step rule, at the editor's word. The post-review found it was not named in the commit message; it is named here and in the conductor's status.
What this does not prove
The reviewers share the session's reach: a second reading, not a second identity. No other member has run a five-step wake yet, and the gate does not run wake.py check until dev wires it in.