ADR-0013: Suspend the program workflow; five ask-first gates; report after
Date: 2026-08-23 Status: accepted Decider: repository owner
Context
The governance model required owner approval for every commit, push, PR, merge, and program record, plus a milestone/package lifecycle with approval cards. Measured over the 2026-08-21/22 incident arc: ten PRs, about twelve owner approvals, all granted verbatim, none producing a change request. Every real defect that week was caught by a machine guardrail (argo-health, CI tripwires, netpols, quotas), none by a human gate. The owner reported the model made delivery slow, consumed the attention needed to understand what was implemented, and made the work joyless. The owner asked for the simplest model that keeps control where it matters.
Decision
- Ask first, always, for five things: production applies and dispatches; secrets and trust (vault writes, trust-root, RBAC, identity, federated credentials, edge auth); deletion of live resources, data, or history; new spend; posts or PRs to other orgs. When in doubt, ask.
- Everything else is autonomous: commit, PR, merge on green CI, verify, revert if wrong. Dev recovery of merged state included.
- Report after, instead of asking before: a short summary per finished unit, one ledger row per landed change, a weekly one-page shipped brief.
- The delivery-program workflow is suspended — milestones, work
packages, approval cards, and the
docs/programs/lifecycle gate nothing. The documents remain as history and reference. Production safety is untouched: two-phase saved-plan dispatch by the owner, W24/W25 blocking, break-glass ceremonies. - Records that remain mandatory: the ledger, ADRs for durable architecture decisions, runbooks. Asana becomes on-request.
Options considered
Option A — keep the model
- Benefit: maximal pre-control on paper.
- Cost: measured as pure latency; approvals granted unread; owner attention drained; owner explicitly unhappy.
Option B — risk-tiered model with milestone-level program approval
- Benefit: fewer asks, programs keep structured records.
- Cost: still a lifecycle to maintain and to learn; the owner judged it still too much. Rejected for now; its design is preserved in git history and can be revisited.
Option C — five gates + report-after, programs suspended (chosen)
- Benefit: one short rule set; velocity bounded by CI, not owner availability; owner reads summaries instead of granting approvals.
- Cost or risk: a bad autonomous merge reaches dev before a human sees it — mitigated by CI tripwires, argo-health, control-drift, netpols, quotas, and cheap revert. Program-grade record keeping stops; if staging/prod delivery later needs it, reactivation is a new owner decision in a new ADR.
Consequences
- CLAUDE.md carries the whole operating model in one short section; AGENTS.md points to it; GOVERNANCE.md carries a suspension notice.
- Agent memories and harness permission rules are updated to match after merge, and a weekly shipped brief is scheduled.
- The machine guardrails become the primary control surface and deserve continued investment.
- STP and PVW continue as normal work under the five gates: prod applies still wait for W24/W25 and the owner's dispatch; preview work proceeds autonomously except where it touches a gate.