The toolchain is working.
Five class tool suites shipped in seven days. Each one — companion, bench, quiz bank — went from nothing to deployed in under 48 hours. The platform infrastructure from the last two months is now producing student-facing features at a pace that wasn’t possible before.
/internal/decks.Aug 21 → Aug 27
themultiverse.school — 89
multiversecampus — 1
Campus is post-ship hardening
Matrix agent runtime (1.15.0) shipped last week. This week: 30 open PRs, 15+ Groq rate-limit issues in GlitchTip, agent tools being tightened. One merge is the right pace for stabilising what just shipped.
Five tool suites, ready for class
Digital Identity Defense
Companion + bench. 400+ real phishing messages as a quiz bank. Home Network Defense bench. Streak tracking with confetti. First class in the Defender track.
Context Engineering
Live inference bench. 12-factor agent demos. Sampling, constitutional AI, embeddings, ICL, and chunking stations. Budget accrued from enrollment.
Using Large Language Models
Eight workbench tools in one PR (#1137, 25k lines). Full tool suite for the class.
Pricing Bench (rewrite)
Plain language. Split first screen. The floor that actually binds was below the fold — fixed. Arithmetic errors caught and corrected.
Bibliographies × 32
Verified sources added to 32 curriculum tools across three PRs. Adversarial review caught 28 citation errors.
Also shipped: Misinformation Cleanup Crew pathway fork (#1068). Value Chain Bench (#1071). Ceiling Bench (#1043). FF Program Health with persona-adjusted scoring (#1163).
Three things that keep the lights on
Smoke canary
Synthetic Stripe webhook → enrollment → email → Gmail inbox check. Alerts to #infra-alerts on Matrix. We know enrollment is broken before a student does.
Model registry
All hardcoded model IDs replaced. Picker, admin, and 3 tools read from model_benchmarks, synced daily from OpenRouter. One DB update changes every AI dropdown.
Exercise endpoints
Gated and routed. Lesson shelf on path dashboards. Bench packet submission. Quartz retired as a serving tier.
Who built what
Most code is agent-written. These are the people who directed and verified the work.
Class tools, smoke canary, model registry, pricing bench, bibliographies
All five tool suites. The enrollment canary end-to-end. Omni model benchmarks. FF program health. Adversarial test coverage for passes, PWYC floor, and profiles.
23 queued PRs landed — ops, a11y, docs, mutual aid
Inbox handbook + onboarding checklist. Theme contrast sweep on every admin template. GDPR delete fix. Support mailbox fix. Mutual aid local-pack guard. Scholarship backfill. Teacher pre-publish checklist. Dead link fixes. The review queue from last week’s deck is flowing.
Week 0 cohort on campus · CI red-test host fix
#1337 — eight-week cohort structure from cohort 4. Also stopped a red unit test from taking down the whole CI host (#1093).
Last week’s four, scored
| Decision | State |
|---|---|
| 1 — Docker socket for CI | Done — #1087, merged Aug 24 |
| 2 — Rotate the exposed secret | Week 3 — code shipped, rotation not done |
3 — 97 templates, asset_version() | Not done |
| 4 — Reviews as GitHub review objects | Not started |
Rotate the exposed secret or close #856
A: do the cutover. B: accept the exposure and close the issue. A third week of implying A while doing neither is the worst option. Recommend A.
Replace the hand-bumped literal in 97 templates
asset_version() exists. One template uses it. 97 don’t. It failed twice in the CDN breach. Same recommendation as last two weeks. Do it.
Cohort end is five days away
The shape changed
Last week: 124 PRs clearing the review queue. This week: 90 PRs building student tools. Infrastructure → features. Right direction heading into a cohort end.
Milestones, third observation
4.2.504 has 27 open issues and 4 closed. Same pattern as the last two releases. Work ships continuously to production. The milestones describe a parallel universe.
Nobody here is behind
These are merged PR counts. Most code is agent-written. The number measures what got through the queue, not who worked harder.
Merged, Aug 21–27, both repos
Megs went 16 → 23. Her bar still understates her week — inbox handbook, a11y sweep, and the GDPR fix are each worth more than a merge count.
Distribution
Bursty. Most of the week was spent building. Two days of merging. The pattern is sprint-then-land, not steady-state — which is fine when the tools are this size, but worth watching as a pacing signal.
The pace is the product.
Two months ago the question was whether we could build a complete class tool suite before the class ran. This week five of them shipped, each in under 48 hours. That’s the infrastructure paying off.
What’s different
Track passes, PWYC pricing, exercise endpoints, the bench framework, the model registry — none visible to students, all of them made this week’s tools possible at this pace.
Heading into cohort end
Sep 2 is five days out. The class tools are built. The smoke canary is watching. The question for this week is what we stop touching.