Weekly digest for the week of 2026-06-01 → 2026-06-07. What shipped:

MemberIntel Weekly Digest — 2026-06-01 → 2026-06-07

One major feature shipped end-to-end in a single day, two security/compliance items closed, and a tooling fix quieted a class of false-positive review noise.

---

Plugin-initiated one-click connect was the week’s main event. The problem: operators arrive from WordPress, not from the app, and the existing flow made them start at the wrong end. Spec (#164) and plan (#165) merged before any code moved; implementation followed in sequence through #166–#168, with three same-day tightening passes (#170, #171, #172) and a stale-reference sweep (#173). Plugin v1.6.1 published at 07a1a24.

ADR-0025 shipped via #160, pinning what had been an undocumented implementation detail: Free tier gets Claude Haiku, Pro gets Claude Sonnet. The cost math holds at ~$1.07/user/month under optimistic infra assumptions; realistic GCP stack pushes that to ~$1.30, shifting the break-even conversion floor from 5% to ~6%. The ADR writes a mandatory reopen trigger — 90 days post-GA, if conversion is below 4% and the Haiku-vs-Sonnet eval gap exceeds 20 points, a cost-vs-quality analysis goes to Blair with a specific lever proposal.

The legacy sync route was deleted (#159) and three follow-on PRs cleaned up stale references and corrected inaccurate language in ADR-0022 (#161, #162, #163). Infrastructure drift and observability scope fixes also landed (#155, #157).

Copilot review instructions landed in #11, triggered by a false-positive on #10: Copilot flagged a link to a deployed ADR page as broken because the file lives in the build output, not the repo tree. The new .github/copilot-instructions.md documents build-time derived content, bot-authored files, Astro 6 glob-loader behavior, and current schema enums. Root CLAUDE.md got an enum sync in the same PR — stale roles removed, current ones added.

---

Notable. The disconnect bug caught in #171 was a genuine dead end: clicking Disconnect only rotated the local key, leaving the backend Site row intact, so any subsequent reconnect attempt hit SiteAlreadyConnectedError with no recovery path from the UI. The fix — a new POST /api/v1/sites/disconnect-from-plugin endpoint that cleans both sides atomically — closed it before any real operator encountered it. Worth noting alongside the sync route exposure from earlier in the week: both were gaps where the system appeared to work but left users (or data) in an unrecoverable state. Different layers, same pattern.

---

Synthesized from 5 daily standups in [2026-06-01, 2026-06-07].