Darwinator · changelog

What we built

Every merged pull request, grouped by the arc of the project — from the spec and design system to a live peer-review tournament platform. Each line links the PRs behind it.

Updated Nov 21, 2026, 4:37 AM EST · internal
138Pull requests merged
14Phases of work
#4–#191Span
  1. 01

    Foundation & planning

    The spec, the architecture, the design system, and the guardrails that kept independently-generated code coherent.

    • #4#10#11#14#86Product spec: PRD (peer-review-first framing, configurable scales, rater fixed effects, Google-only V1) and ADR-0002 for approximate serving.
    • #6#7Folio visual system — the design brief and handoff.
    • #8Next.js 16 + React 19 + TypeScript app scaffold on the Folio foundation (Tailwind v4).
    • #12#13CI that gates every PR (format · lint · typecheck · test · build) and CLAUDE.md, the agent brief.
    • #5#9Local-dev helper + setup/services docs.
  2. 02

    M1 · The product, on a typed data layer

    The whole participant + facilitator experience, built against a mockable DataClient with the scoring/serving logic tested in isolation first.

    • #15#16Domain model + tested scoring & serving logic, and the ScaleInput family (binomial / rating / slider).
    • #17#18Participant evaluation loop over the typed data-access layer.
    • #19#20The feedback report (the headline surface) with the “remove rater fixed effects” toggle.
    • #21#22#23#24Submit, tournament home, my-tournaments + join-by-code, and sign-in.
    • #25#26#27#28#29#30Facilitator console: setup, live dashboard, sortable rater-corrected results, roster, and moderation queue.
  3. 03

    M1 · Onto the real backend

    Swapped the mock for Neon + Drizzle + Better Auth, with server-side authorization on every request.

    • #31#32Drizzle schema + first migration, and a server-backed DataClient (drop-in for the mock, reusing the tested logic).
    • #33#34Better Auth (Google OAuth + magic link, passwordless) with a gated real entry flow.
    • #35#36The real participant journey and facilitator console, moved off demo data onto session + DB.
    • #38#39Editable per-question labels/ranges, and a security fix enforcing round phases server-side.
  4. 04

    M3 · Live voting & presentation

    Room-scale voting on the same polling stack (no realtime tier), with a projector mode.

    • #40#41#42Live-voting data layer (idempotent votes), the participant ballot + facilitator control, and projector/presentation mode.
    • #43#45#46Live dashboard polling, unpick + per-vote notes, and feedback-report recompute + histogram a11y.
    • #44#48#71Ops + concurrency: middleware→proxy rename (Next 16) and an atomic, advisory-locked pick cap that holds under load.
  5. 05

    Responsive across every surface

    First-class on phone (participant) and desktop (facilitator), per the per-surface posture decision.

    • #49#50#51#52#53Responsive submit, evaluate, feedback report, live-vote ballot, and tournament home / my-tournaments.
  6. 06

    M2 · Testing, hardening & load

    Made it trustworthy: end-to-end tests in CI, a concurrency/load harness, and an adversarial audit with every finding fixed.

    • #55#57#60#68Playwright E2E — auth-gate smoke tests in CI plus double-gated authenticated journeys (setup → join → submit → evaluate → report).
    • #69Concurrency/load harness on a Neon branch + a k6 HTTP script, proving idempotent serving/counting under spike load.
    • #62#64#65#66Adversarial audit fixes: evaluation-integrity boundary, blob host-pinning (SSRF guard), config validation, and test-route hardening + comment caps.
    • #54#58Private image attachments on Vercel Blob (incl. OIDC-connected store support).
  7. 07

    Public progress dashboards

    The /status and /roadmap boards — later made correct-by-construction so they can't drift.

    • #56#59#61#63The public “coming soon” landing and the build-status board (US-Eastern timestamps, internal-only).
    • #72#73#74Two-level dashboard: /roadmap (epics) + /status (tasks), with the full V1 roadmap.
    • #104#105The public /changelog — every merged PR, grouped by phase, with drift-proof derived stats.
    • #149Live, not “coming soon”: the landing page leads with Sign in, /status and /roadmap read “Live” and name exactly what is left, and stale “later” wording for features that now exist is gone from the app and docs.
    • #158Roadmap: the remaining unbuilt items found by the coming-soon sweep are listed under Later (rater-corrected rank and percentile, more attachment types, a full administrator activity log).
    • #152The Public switch is hidden until public tournament search exists: it changed nothing, so setup and Settings no longer show it. New tournaments are saved as not public, existing ones keep their value, and anyone with the code or link can still join.
  8. 08

    Phase 1 · Facilitator admin

    Everything a facilitator needs to run a tournament end to end.

    • #75#79Edit a tournament after creation — settings (CONFIG-1–17) and the evaluation questions (locks once evaluation starts).
    • #76#78Roster management (remove a participant) and a shareable self-registration join link.
    • #77Deterministic E2E seed route behind a shared, double-gated test-route guard.
  9. 09

    Phase 2 · Reporting & grading

    Getting judged results out of the app for grading and analysis.

    • #80Admin CSV / grade-ready export (submissions + per-participant), identities resolved, formula-injection-safe.
    • #81Custom report builder: by-submission / by-participant, pick columns, sort, top-N, live preview + CSV.
  10. 10

    Phase 3 · Accounts, comms & polish

    Rounding out identity, communication, and accessibility.

    • #82#90Account / profile screen and profile-photo upload (private blob, streamed through an authorized route).
    • #89#94Email invites with auto-join on sign-in, and broadcast email to a chosen audience of members.
    • #83#84#85Vote soft-delete integrity, a WCAG 2.2 AA pass on the new screens, and a plain-language privacy policy.
  11. 11

    Launch: deploy, email & staging

    The operational path to go-live: safe deploys, an isolated staging env, and working email in and out.

    • #88#91Deploys auto-apply Drizzle migrations (no more code ahead of schema); an isolated Neon preview branch with preview-origin auth.
    • #93#102Multipart (text + HTML) transactional email for deliverability, and inbound email wired (catch-all → operator inbox).
    • #92Launch-readiness sweep — prod smoke-tested, apex clean.
  12. 12

    Correctness, quality & housekeeping

    Keeping the boards honest and the tree clean — including fixes an external Codex review surfaced.

    • #95#96Made /status + /roadmap correct-by-construction: single source, build-time timestamps, and a CI guard against drift.
    • #100Codex-review fixes: soft-deleted opportunities no longer counted toward evaluation progress; db:push disabled to protect the migration journal.
    • #101#103Pre-launch cleanup (dead code, stale artifact/docs) and a repo-wide “idea” → “opportunity” terminology sweep.
    • #128A Codex review check on every PR (green only once Codex has reviewed the current head commit and every Codex finding is resolved), plus a safety net that files an issue for any PR merged before it passed.
    • #134Two high-priority late Codex findings fixed: anonymized evaluation cards no longer send submitter ids to raters' browsers, and editing questions can no longer erase a rater's just-recorded ratings (the lock check and the replace are one atomic write; edits wait until Evaluate closes).
    • #136Evaluation questions stay editable until the first rating arrives, even while Evaluate is open, then lock for good. A rater whose screen predates an edit is re-asked on the new questions at once, keeping answers to the questions that didn't change.
    • #144The live-vote ballot no longer sends submitter ids to voters' browsers (#140, the last path of the leak #134 fixed for evaluation cards), and lists opportunities in submission order so the row order can't reveal the current standings.
    • #145Builds no longer download fonts: the three Folio fonts are self-hosted (byte-identical files, same rendering), ending random Vercel build failures caused by Google Fonts answers Turbopack can't parse; a lint rule keeps next/font/google out.
  13. 13

    Pilot feedback round

    Everything the principals' hands-on testing surfaced, resolved in one hardening + UX pass.

    • #122Resolved all open review issues (#97, #106–#121): raw private-blob URLs kept server-side (opaque pending-upload references), email + migration hardening, host-confirmed opt-in for preview migrations, per-tournament member aliases (in exports too), truthful people counts, navigation/sign-out/explainer UX, and an unsaved-changes guard.
    • #133Email sign-in links survive mail scanners: the link opens a confirm page and only a button press spends the single-use token; failed links explain themselves on sign-in, links last 15 minutes, and a stale session cookie no longer loops between /signin and /tournaments.
    • #132Participant report explained in plain language (#126): each question's configured scale and labels beside the result, plain-language rank / percentile / quality / histogram with (?) help, few- and no-ratings states, a clear “not a grade” line, and explicit raw-vs-adjusted units. No scoring change.
    • #129Participants can remove their own opportunities (GH #125): inline-confirmed Remove on the submit screen until evaluation first opens or peers have engaged, then locked for good with the reason explained; a race-safe soft delete with an audit stamp (who + why), shared with admin removal.
    • #131Facilitator recovery controls (GH #127): administrators remove an opportunity from the results table with an audit list of removals, and reopening Submit or Evaluate previews what changes and asks to confirm — nothing is deleted, no one rates twice, and ideas peers have responded to stay locked from edits. A rater whose open card is removed is told and moved on; one whose questions were edited meanwhile is re-asked.
    • #135Participants can revise their own evaluations while evaluation is open (#124, FR-EVAL-6): “Edit last” and “Your evaluations”, an in-place save that never counts twice, a lock when evaluation closes, appendable (capped) comments, and server actions that always bind the caller from the session.
    • #143Co-admins with a distinct owner: a tournament's creator owns it and can make participants co-admins (e.g. teaching assistants), take those rights away, or hand ownership on; co-admins run everything else and can step down. Enforced inside each write, with database guarantees of exactly one owner and an activity log of admin-team changes.
    • #137Fresh restart of a tournament (GH #127): from Settings, an administrator runs the tournament again with the same people and setup. Every live idea is archived with its ratings, comments, and votes (nothing deleted), every phase closes, and the questions unlock. A preview and a type-the-name confirmation guard it, a stale preview is refused, and each restart is recorded on Results.
    • #146Invite a co-admin by email: the owner sends a co-admin invitation from the Admin team card; only the invited, verified address can accept it (by its link or just by signing in), it carries no join code, and a used or withdrawn link adds nobody. Resend and withdraw are owner-only, and revoking a co-admin drops their pending invitation.
    • #138Uploaded photos no longer reveal who took them: GPS, camera and owner metadata, thumbnails, XMP and identifying colour-profile details are stripped losslessly (same pixels, still upright) before storage and on every view, and older uploads are re-saved clean the first time they're viewed (a damaged file is served clean but never overwritten).
    • #148The browser tab shows the Darwinator selection-tree logo instead of the stock Vercel triangle: a tab-sized redraw for the favicon, the full logo for phone home screens, and a web manifest naming the site, all regenerated from brand/ by one script.
    • #150Superuser mode for support (ADR-0004): from an unlisted page, an account on a short list in the code (with a verified email) switches on an 8-hour mode tied to its sign-in, sees every tournament, and can do anything an owner can, without ever joining one. Every visit and change is logged before it runs (and refused if it can't be), changes show as Darwinator support, console pages carry a banner while it's on, and the privacy page says so.
    • #156Roadmap: the co-admin clash fixes are off the roadmap and the full administrator activity log is a Later item, so Phase 1 is shipped (few tournaments have more than one or two operators). The two-admin live-vote clash is recorded as issue #153, not scheduled.
    • #157Longer write-ups: an opportunity's description can now run to 1,200 characters (about 200 words), up from 600, with a taller box on the submit form. Paragraph breaks now show wherever a description is read (rating card, report, live-vote ballot, projector), and the projector sets a long description smaller and left-aligned so a full one still fits a 1080p screen.
    • #159Scheduled phase changes: on the Live console an admin sets when Submit, Evaluate and Results open and close, in their own time zone (named on the form). Each phase then changes on time by itself, exactly as if someone had clicked, with nobody online: the first read after the time applies it once, and a daily sweep catches up anything nobody opened. Nobody can submit or rate after a scheduled close, a change by hand stands until the next scheduled one, a fresh restart clears the schedule, and participants see the deadlines on the tournament page in their own time zone.
    • #160Deadline reminder emails: about 24 hours before Submit or Evaluate closes on a schedule, each participant who is behind gets one email with the deadline in their own time zone and a link straight to the task. Each person gets each reminder once (recorded before it's sent, even with a read and the daily cron at once); reminders count against the daily email allowance and are listed with who got them on Manage → Message, where an admin can turn them off.
    • #177Coverage counts the right way (GH #161): "Evaluations: min each" is per person, so the ratings to expect are people × that, not ideas × it (40 people rating 10 of 80 ideas read 50% judged when they were done). The Live console now shows evaluations per opportunity (average, lowest and expected, and where expected comes from), and the participant home's JUDGED figure reads the same corrected number.
    • #178A readable Reports table (GH #162): the full description and the list of submitted titles start unticked, and when shown, long text gets a reading width and three lines with Show all for the rest, so one long write-up can't stretch a row to 2,500px. A long question prompt is a two-line header on Reports and Results, and unticking every column no longer falls back to all of them.
    • #179The feedback report opens with the result (GH #171): the opportunity's rank, percentile and quality come first, then how peers rated each question and their comments. The intro is one line (“These numbers aren’t a grade.”) with the rest behind its (?), and the rater-adjustment switches moved below the results under “Adjust for generous or strict raters”, collapsed until opened.
    • #180The projector reads from the back of the room (GH #171): 24px labels and a larger title and tally on a 1080p screen, sized to the screen; while a vote is open, a QR code and the join link with its code sit beside the opportunity; the tally counts up and its bar slides instead of jumping; Prev, Next and Exit fade after 3 seconds without the mouse (moving it or tabbing brings them back). The vote control’s join link now encodes the code.
    • #181A rating screen built for the thumb (GH #171): on a phone the header is one line (back to the tournament, and your progress, which links to your evaluations once one is saved), and “Submit & next” sits in a bar pinned to the bottom of the screen, clear of the home indicator, with “Saved · Edit last” beside it.
    • #182A short pause per card, and a block for bots (GH #165): each card stays up 1 to 3 seconds, by how much it says, before its rating can be saved, checked on the server with a signed token per served card. Submit counts the pause down. A rating only a program would send blocks the rater (“Non-human behavior detected.”) until an administrator unblocks them from Participants, which shows why.
    • #184Tournament themes (GH #170): Folio moves from the warm bone page to mist under its blue, and an administrator can recolor a tournament from Settings: Karan green on sage, Cornell red on porcelain, Wharton blue on mist (Penn Red marks your own result in the report), or the Abundant Intelligence Lab's navy and steel with a stripe on top. Every screen of the tournament follows, and a test holds each theme to WCAG AA.
    • #183An Auto-advance checkbox beside Submit (GH #163): off by default and remembered on the device. Ticked, a card submits itself a moment after its last answer, still waiting out its pause; Submit stays, and a comment someone has started waits for them. A slider's answer now counts when it is let go, not when it is first pressed.
    • #185Swipe through Up/Down cards on a phone (GH #166): a card whose only question is Up/Down can be swiped right for Up or left for Down, recording exactly what the buttons record. The buttons and arrow keys stay, a short drag snaps back, and a card swiped before its pause is up waits at the edge. The rating screen became a deck: the next card is fetched while you read, its edge shows beneath, and a saved card leaves toward its answer.
    • #186Fewer, quieter labels (GH #173): small mono capitals now mark only statuses, counts, codes and data, about half as many as before, and nothing is set under 12px. Links, field labels and sentences use the sans, card titles the serif, labels that repeated a heading are gone, and long explanations are one sentence with the rest behind a (?) button. A test keeps anything under 12px from coming back.
    • #187Motion that confirms (GH #174): each screen's content fades in as it opens while the top bar and side nav stay put, and on the console's live-vote screen each row has the projector's tally bar, figures that count to each poll's numbers, and rows that glide to their new rank. Everything is instant under reduced motion.
    • #188Line icons instead of text symbols (GH #175): back and onward chevrons, a check for saved or ready, a dot for live, a ring for closed, sort arrows, and thumbs beside the default Up and Down, all from one thin line set vendored from Lucide and drawn in the text's color, so they look the same on every phone and follow each theme. A test keeps the old symbols out.
    • #189New tournaments start with one Up/Down question, "Worth pursuing?", so phones can swipe their cards with no setup (a follow-up to GH #166). Organizers can still edit it or add questions, and existing tournaments keep theirs. When every question is Up/Down, the reports no longer offer "Normalize rater scale", which doesn't apply to a two-choice answer.
    • #190Dark mode that follows the phone (GH #176): a device set to dark gets dark screens, in CSS alone so nothing flashes on load. All five themes have their own dark colors, held to the same contrast tests as light, and the browser's bar follows. The projector stays light, since it's the room's screen.
    • #191A front page that explains Darwinator (GH #172): about 100 words, with the headline “Everyone pitches. Everyone judges. Everyone learns.”, three steps, four live features and what's coming next, beside the real rating screen on a phone, light or dark as the device is. A shared darwinator.com link now shows a preview card with the same headline.
  14. 14

    Sign-in upgrades

    Faster, sturdier ways back into an account — beyond waiting on an emailed link.

    • #130Passkey sign-in (ADR-0003): a passkey button + autofill on /signin and add/rename/remove on /account, bound to darwinator.com with pinned origins, verified by real WebAuthn ceremonies in tests.
    • #139Class-scale sign-in: about 120 students behind one campus address can request email links and start Google sign-in together, while a per-mailbox cap (plus-tags and Gmail dots folded together) keeps one inbox from being flooded. New tournaments get a random, link-safe join code; the copyable join link uses the canonical address; a stale link explains itself; and joining by code opens the tournament.
    • #142Class email invites: paste a class list (up to 500), sent in batches with each invitation's delivery recorded; the screen names what didn't send with Try again and Resend, and pending invites can be resent or withdrawn. Each invite carries its own link, so a student who signs in with another address is still joined and shown as 'Joined as …'; the note is kept for next time, and the email names the facilitator and prints the join code.
    • #147A daily email allowance keeps email sign-in working on Resend's free plan: the site counts every email it sends per UTC day and stops invitations and broadcasts at 70 of the 100, so 30 are always left for sign-in links. Invitations past it are saved as 'not sent' to resend after the reset, a broadcast that wouldn't fit is refused before anything is sent, and the invite and message screens show how many emails are left today. Both numbers are settings, so a paid month is a configuration change.
    • #154No daily email cap on Resend Pro: the allowance is off by default (still counted, and EMAIL_DAILY_LIMIT=100 brings the free-plan limit back), and a broadcast to several hundred people now waits out Resend's per-second limit instead of dropping part of the group.