9.1 KiB
WikiDeck Night 1: report
Night: 2026-11-16 22:45 → 2026-11-17 01:30 EST. 17 Claude agents on 4 machines, coordinated through Gitea (milestone 63, issues #1–#10).
Branch: claude/wikideck:night/2026-11-16. Everything below came in through 38 PRs that night-review reviewed and night-lead squash-merged.
Untouched: the live deployment, mainnet, and every consensus and protocol file. The off-limits list is in docs/night/BOUNDARIES.md.
What works
| Area | Status | Where |
|---|---|---|
| The game, "Edit War" | Playable in the browser against an AI: a lobby, a deck builder with 4 loaner starter decks, matches, results and a tutorial. Works on a 360 px phone. | src/game/, PR #26 #30 (landing page, upstream WikiDeck/wikideck#3), #31 (easy and normal browser bots), #45 |
| Rules engine | Pure, deterministic TypeScript with the rules as data. createMatch, legalActions, applyAction, viewFor and serialisation. 6 keywords and 6 Portals. Property tests, plus a §12 golden match checked by hand. |
shared/game/, docs/game/ENGINE_API.md, #25 #27 #29 #33 #36 (v1.1 as the default), #38 (planning is hidden) |
| Design | DESIGN.md v1 (frozen 23:01), then v1.1 tuned from the balance report. | docs/game/DESIGN.md, #13 #17 #23 #24 #32 |
| Card data | 236 sampled articles, with real distributions, percentile tables and 48 starter pageIds. | docs/game/CARD_DATA.md, card-sample.json, #20 #28 |
| Balance | Bots (random, greedy, lookahead) played thousands of matches per ruleset. On v1 the starter pairings ranged 23–77%, and the first player won 53.4%. The v1.1 re-run measured the first player at 51.2%, starter decks at 52–53% and owned decks at 42–51% against starters, so owning NFTs adds variety, not power. | docs/game/BALANCE.md, shared/bots/ #14 |
| eom.dev accounts | Dex OIDC with PKCE: login and logout, and ID tokens verified against the JWKS. Wallet linking signs a message the client builds, never text from the server. The account ↔ wallet map is stored off-chain. | server/account/, src/auth/, #15 #37 #42 |
| Peer-hosted modes | Modes are JSON data, never code. A node publishes them at GET /api/game/modes, peers are discovered through config, and the hosting node runs the match. It shows "Hosted by", with per-mode and host leaderboards. 3 example modes: classic, snake draft and trending. Lobbies survive restarts. |
shared/game/modes.ts, server/game/, docs/game/HOSTING.md, #22 #43 |
| Private testnet | 3 nodes on poweredge-r350, LAN only, at height 50. Auto-redeploy followed this branch until the freeze, with rolling restarts that wait for live matches to finish. | deploy/night/, #16 #19 #39 #40 |
| Android app | Capacitor 7 (dev.eom.wikideck). A debug APK runs a full match against the AI in the emulator. CapacitorHttp gets past the node's Origin check with no server change. The SSO deep link round trip works up to Dex. |
android/, docs/night/ANDROID.md, #35 #41 #44 |
| iOS app | Capacitor 7 iOS project (SPM) and a simulator build script. The macOS Sequoia VM's install was still running at the end of the night. No Apple ID was used. | ios/, docs/night/IOS.md, #34 #52 |
How to try it
-
Browser: http://192.168.1.32:3101, :3102 or :3103 (LAN only). Open Play, dismiss the tutorial, pick a starter deck and play the AI. The three nodes host different modes: A has classic and trending, B has draft, and C has all three.
-
Android: the APK is at
inspiron-3670:~/worker/jobs/night/mobile-android/apk/wikideck-night1-debug.apk(debug-signed, testnet build). To rebuild it:android/build-debug.sh. Screenshots: -
iOS: follow Eric's checklist in
docs/night/IOS.md:- resume the VM on inspiron and finish the macOS install
- sign in with your Apple ID
- install Xcode
- run
scripts/ios/build-sim.sh
No Apple ID was used tonight. The VM got as far as the Sequoia installer downloading the full OS. Per the coordinator, the VM is shut down cleanly before 01:30 with its disk kept.
docs/night/IOS.md(#52) records the install progress at shutdown, the resume steps and Eric's checklist. -
Login: "Log in with eom.dev" uses Dex at
https://dex.eom.dev/, through thewikideckpublic client that the coordinator applied. You need a membership account that's inDex Users; the SSO issue #7 covers the sign-up path.
QA (night-review, live testnet)
- API: full hosted matches in all 3 modes (classic @A, snake draft @B, trending @C) with two scripted players. Records and leaderboards are correct. Rejected as they should be: forged card, cross-seat token, random token, client
timeout, foreign Origin, self-join, out-of-turn pick. - Browser: Playwright played 3 full matches against the AI through the real Play UI (360×740 @A ×2, 1280×800 @B). Every one reached results, with no overflow, stuck turns or page errors.
- Engine fuzz: 300 random matches per engine branch. 0
viewForleaks (information-flow test), and every replay was deterministic. - UX bug found and fixed: the results card showed tied Topics ("13–13 Lost") with no tie-break explanation. Fixed after the freeze in #51.
Known issues and gaps
- The game ships rules v1.1 (
DEFAULT_RULESET = RULESET_V1_1, #36), which matches DESIGN.md v1.1. The v1.1 numbers in BALANCE.md (first player 51.2%, starter decks 52–53%, owned decks 42–51%) were measured before #38 hid the opponent's queue, so the seat figure needs a re-run. Item 6 (nostalgicandtranslatedat −1 power) is a v1.2 candidate and isn't in v1.1. - Hosted PvP trusts the host. The seed is chosen by the host, and DESIGN §4.4's commit–reveal isn't implemented. A malicious host can see both hands. That's documented in HOSTING.md.
- Hidden information. #38 hides the opponent's queue during planning (with a test). The UI and the bots consume only
viewForoutput, and the log never names an opponent's hidden card. Remaining:obscure(item 4) and host trust (item 2). - The
obscurekeyword isn't masked inviewFor. Modes reject it, so it can't be turned on. - Game traffic shares the chain's per-node RPC budget, so one hot client can get everyone 503s (#6).
- Wikipedia rate limits: our egress IP got HTTP 429s on the action API. #19 adds an optional
WIKIDECK_USER_AGENTwith a contact URL, and only the testnet sets it. The live deployment probably shares that IP and the old generic UA. Whether to set it in production is Eric's decision. - Mobile polish:
- the dashed lane drop zone doesn't take taps, only the lane plate does
- the home hero still says "Explore mining" inside the app
- the real SSO login hasn't been done on a device yet
scripts/ios/build-sim.shstill runssudo xcodebuild -license accept, which would accept Apple's Xcode licence on Eric's behalf. Eric should accept it himself (remove the line first).- Android allows cleartext HTTP to any host (
usesCleartextTraffic="true"). Before any non-LAN use, restrict it with anetwork_security_configto the testnet host.
- Factions are skewed: owned cards are about 53% Culture and 0% Space, so
viralandlocaldominate. Factions built on Wikidata P31 are the v1.1+ fix (CARD_DATA findings 6 and 7). - Tooling: the repo needs Node 24 (
node:sqlite). On Node 20, 12 tests fail falsely. Every gate tonight ran innode:24-alpine.
Held for Eric: #49 (the normal bot uses sim-balance's lookahead). It's approved, but it's a behaviour change that arrived after the freeze.
Next steps (in priority order)
- Re-run balance on the final head (after #38) and decide v1.2 item 6. The harness is
scripts/sim/andshared/bots/, and it needs about 1 GB on r350, run under the heavy lock. - Hidden information everywhere: mask
obscure, and implement the §4.4 commit–reveal seed for hosted PvP. - Separate rate limits for game traffic on the node, so they don't share the chain's RPC budget.
- Finish iOS: Xcode and a simulator run, then the same match against the AI as on Android.
- SSO end to end on a device with a real eom.dev account, then show account names on the leaderboards.
- Factions from Wikidata P31, then re-run the balance sims (item 6 on the BALANCE list is still open).
- Host incentives on-chain: the proposal in HOSTING.md. It's a protocol change, so it needs its own review.
- Production user agent: decide whether the live node should set
WIKIDECK_USER_AGENT.
Process notes
- Every agent shared the
claudeGitea account, so night-review approved with COMMENT reviews that start withVERDICT: APPROVED, naming the head SHA and the gate result. - imac's Gitea MCP was down all night. The council read issues anonymously and reported through PR bodies, and night-lead relayed.
- tower-plus ran out of memory at about 23:25, and r350 twice. From then on, every heavy check ran on inspiron in
node:24-alpineunderflock /tmp/night-heavy.lock. - Squash merges dropped the history that stacked PRs shared, which caused add/add conflicts (#27, #22, #26). Each was resolved mechanically, with no content changes.
