standup
Weekly — week of 2026-07-20
## MemberIntel Weekly Digest · 2026-07-14 → 2026-07-20 Roughly 50 PRs merged across six days, covering two major architectural decisions, a new plugin module sy
▶ Watch the 1:13 summary — Weekly status — Week of 2026-07-20
Weekly digest for the week of 2026-07-14 → 2026-07-20. What shipped:
MemberIntel Weekly Digest · 2026-07-14 → 2026-07-20
Roughly 50 PRs merged across six days, covering two major architectural decisions, a new plugin module system, the complete offline eval harness, sandbox infrastructure, and a sustained push to close admin-console fit-and-finish before beta.
Eval infrastructure is now end-to-end. The offline harness (#447) runs against 319 connected-site profiles with a compare-runs CLI for regression detection; pass-rate thresholds locked in #453 (denylist ≥ 99.0%, term-overlap ≥ 98.0%). Production eval monitoring (#434) is sampling live traffic on staging — 5% of CHAT calls, 20% of SITE_ANALYSIS — using code-assertion checks and cost-breaker-style alerting with no judge model or third-party service. ADR-0035 documents the build-vs-buy choice (#431) but is still in “Proposed”; Blair’s verbal go-ahead is recorded, written sign-off is outstanding and should be resolved before V1 launch.
The connector wire format is settled. ADR-0036 defines namespaced string IDs, a required platform tag, explicit schema_version in the envelope, and per-record currency emission — the contract that unblocks MemberMouse, WishList, EDD, and BuddyPress modules. MemberPress is already the first concrete Billing_Module under the new plugin registry (#478). A stopgap in #480) patched sync.py immediately after, since the old code had no schema_version handling and would have crashed on the first sync against a site running the new plugin build.
Sandbox environment plumbing is in, cutover is pending. Terraform variable widening, CI deploy target, and live infra fixes for VPC peering and Direct VPC Egress all merged across #439–#466. The actual GCP apply and Staging build (Tasks 7–8) are waiting on Seth.
Brain Content Editor shipped. Backend (#442) and frontend (#444) are both merged, covering candidate edit-before-approve, DB-native metadata and retire controls, a read-only PR deep-link for git-backed rows, and an underperformer filter on entry performance. The retired_at timestamp is now stamped server-side (#458) outside the editor-metadata allowlist so clients can’t spoof it.
Archetype benchmarks corrected. The ai-engineer review of #472 caught a pricing math error — annual prices were being pooled as raw dollar figures rather than monthly equivalents — before any user saw the cohort data. All 60 benchmark files were regenerated as a hard merge gate. Fable 5 was chosen over Haiku 4.5 (83% vs. 61% on a 112-row gold set); ADR-0028 received a pricing-metrics amendment.
Smaller fixes worth noting. The bracket-citation leak (#454) stopped the advisor from emitting raw tokens like [heartbeat] inline in answers. Email citext conversion (#426) resolves case-variant addresses to one account at the DB level. The audit-log action filter (#477) now derives its catalog from a single _ACTION_MAP served by a new endpoint, so the frontend dropdown can’t drift from the labels.
Notable. The CI deploy fix in #457 closed a genuine footgun: before this change, dispatching a sandbox deploy would have fallen through to staging — which shares the production database — and run Alembic migrations against live data. The fix landed before any sandbox dispatch was attempted, but the gap existed in the deploy workflow for the entire period prior.
Synthesized from 6 daily standups in [2026-07-14, 2026-07-20].