Night 2 report (docs/night2/REPORT.md) #24

Open
claude wants to merge 1 commits from claude/wikideck:night2-report into main
+152
View File
@@ -0,0 +1,152 @@
# WikiDeck Night 2: report
**When:** the night of 2026-11-17, 9:30 PM to 2:00 AM EST. Integration happened the next afternoon.
**Who:** 13 Claude agents on four machines (tower-plus, poweredge-r350, imac, inspiron-3670), working from the design
brief in `docs/night2/BRIEF.md`.
**Where:** all work happened in the `claude/wikideck` fork, on the `night/2026-11-17` branch. It reached `main` as
WikiDeck/wikideck#23 (dbb89ce), merged by Eric on 2026-11-18.
Live log: [Discourse topic 839](https://discourse.eom.dev/t/839). PR numbers below are in the fork unless they say
otherwise.
## In one paragraph
Night 2 rebuilt the game around Eric's feedback on Night 1:
- a 40-card duel in the style of Magic: The Gathering, played only with cards you have actually mined
- card types and rules text derived from the real articles
- a deadpan "Battle of X (2026)" encyclopedia recap after every match
- player-vs-player first: invites, quick match and an Elo ladder, with ownership checked by the server
- a Wikipedia client that stays within Wikimedia's limits
- an Android app that can mine
The code got built, reviewed and tested overnight, but **nothing merged that night**. The night lead never started, so
14 approved PRs sat unmerged until the next afternoon. They are all merged now, and 422 of 423 tests pass (none fail,
one is skipped).
## What went wrong, and the fix
- **The night lead never started.** A new folder makes Claude Code ask "Do you trust this folder?". The spawn script
typed night-lead's prompt into that dialog, so the prompt was never submitted. Most of the scheduled standups failed
the same way; only the 22:30 and 00:01 standups posted. Every other agent waited on night-lead to merge, and night-lead
was the only one allowed to.
- **r350 ran short of memory.** It had about 490–650 MB free of 7.8 GB, because the testnet, simulations and agents
shared a small machine. Two agents' check-in timers were killed, and the testnet's auto-deploy might not have had
enough memory to build.
- **Integration happened the next day.** On 2026-11-18 between 16:11 and 16:16, the coordinator merged everything in
night-review's order and resolved two conflicts:
- `server/game/index.ts`: kept both `ownershipPort` and `ownership`
- `src/App.tsx`: kept the deep-link state and `useAccount()`
It also opened PRs #80–#85 for the work branches agents had left without PRs.
- **The fix is in place for Night 3.** Every agent is now started with `herdr-prompt`, which answers the trust dialog
and checks the session transcript to confirm the prompt was really submitted. It is tested on all four machines.
Lesson: never make one agent the only path to merging without a fallback.
## What was built
### The duel (duel-core #72, #80; duel-design #66, #69, #83)
- **Format:** 40-card decks you own, 20 Notability (life), and Citations (mana) that grow by 1 a turn up to 10.
- Single blocks and no stack.
- A public "Deleted" zone (graveyard).
- A turn cap and a game clock.
- **Rules:** 10 keywords and 12 effect templates (T1–T12). Cards whose articles link to each other are **Connected**
and work together.
- **Engine** (`shared/duel`):
- Deterministic and replayable.
- Hides information: `viewFor` and `logFor` never leak your opponent's hand or draws, and the game seed stays secret
until the game ends.
- About 0.05 ms per action.
- **Balance v1:** the second player starts with 1 extra Citation, and games are capped at 40 turns.
- In 3,000 simulated games between bots, this cut the first player's win rate from 60.1% to 54.7%.
- The median game is about 8 turns each.
- No game runs past 20 turns.
- A random-play soak gave 51.0%.
- Details are in `docs/game/DUEL_BALANCE.md`.
- **Story:** every match produces a play-by-play log and a "Battle of X (2026)" recap page, styled like an encyclopedia
article with an infobox and references back to the exact article revisions.
- The text is templates filled only with real card and match data.
- It never asserts anything about real people. Review removed wording that implied real relationships between real
people.
### Cards from real articles (duel-cards #70, #82)
- **Types:** every card gets one of five types from Wikidata and its article: Figure, Location, Event, Institution or
Doctrine.
- On a 236-card sample: 86 Figures, 75 Locations, 34 Institutions, 21 Events and 20 Doctrines.
- Nothing fell through to the default.
- **Cost and stats:** cost (1–6) and power/toughness come from fame and article depth.
- **Keywords:** at most 2, Figures only.
- **Rules text:** written in the register of Magic: The Gathering, from real numbers only.
- **Flavour text:** the article's first sentence.
- **Fame doesn't win:** famous cards are not strictly better. Within a cost, win rate is flat or slightly negative with
fame.
- **Connected:** 17.2% of card pairs are Connected (the target was 15–30%).
### Player vs player (ownership #74; host-pvp #77, #81; ui-pvp #76, #85)
- **Ownership:** the server checks a deck's ownership against the chain. Each card's current owner must be one of
your linked wallets, with enough confirmations: 1 normally, 3 for ranked play.
- **Deck rules:** exactly 40 cards, at most 4 copies of an article, and no duplicate tokens.
- Forged, borrowed, short or unconfirmed decks are rejected.
- **Under 40 cards:** you see "N / 40 cards. Mine to play". There are no loaner decks in the duel.
- **Hosted modes:** quick match, private lobbies with invite links (`/play/invite/<code>`), and an Elo ladder.
- Classic (ranked), Draft and Trending modes, with a time bank.
- The host runs both Night 1's Edit War and the new duel.
- **Seats:** your eom.dev login. Guests can play only Edit War, unranked.
- **Recap pages** at `/recap/<id>`.
### Duel board (ui-duel #79)
- **Layout:** built for a 360 px phone first, with desktop also supported.
- **Card faces:** in the style of Magic: The Gathering.
- **Controls:** attack in one tap, block in two, with a preview and confirm. Targeting takes several steps.
- **Bot opponents:** Easy, Normal and Hard bots exist (`DuelBotMatch`), but aren't reachable from Play yet.
### Polite to Wikipedia (wikimedia #73, #75)
- **Requests:** throttled, cached on disk, and identified with an operator contact, so the game stays within Wikimedia's
limits instead of being cut off as on Night 1.
- **Limit per address:** at most 3 nodes per IP.
- **Fixes:** a mining prefetch pool, and a fix to how minting fetches article data.
- **Docs:** `docs/night2/WIKIMEDIA.md`.
### Android mining (mobile-mining #78)
- **Engine:** the RandomX engine was cross-built for Android, and the app mines on the testnet.
- The emulator mined testnet block #92, and all three nodes accepted it.
- About 37 hashes a second per thread, and 125–136 a second at Max.
- **Safeguards:**
- Opt-in, 1 thread by default, and Max only while charging.
- It stops when the app goes to the background, in battery saver, or when the phone is hot or low on battery.
- It's disabled in mainnet builds.
- **Distribution:** Google Play and the App Store ban on-device mining, so this is a **sideload-only** build. Options are
in `docs/night2/MOBILE_MINING.md`.
### Testnet and simulation (testnet-ops #67, #68, #75; sim-balance #71, #84)
- **Testnet:** three nodes on r350 now run Night 2 (036489c). They redeploy automatically as the branch changes, deferring
while a match is live.
- **Test wallets:** alice (42 cards) and bob (40) can each play a full duel. Steps are in `docs/night2/TESTNET.md`.
- **Bots:** six bot strengths, plus a simulation harness.
## Numbers
- **Tests:** 423, of which 422 pass, 0 fail and 1 is skipped. The build passes (Node 24, no network).
- **Volume:** 14 PRs reviewed overnight, all approved by night-review. 6 more PRs (#80–#85) opened at integration.
## Not done yet
- **No real PvP match has been played.** The end-to-end test needs two eom.dev accounts linked to the alice and bob test
wallets, and nobody decided which accounts to use.
- **Unreviewed code:** PRs #80–#85 (the engine, host part 2, card types, the recap adapter, bot adapter and board
wiring), plus the last three commits of #70.
- **Bot matches can't be reached from Play.**
- **Card types use the fallback:** the identity pre-bake still lacks its `--live` source, so cards are typed from
on-chain data, which agrees with the full classifier 84% of the time.
- **RandomX verifier:** it gives up for good after a single 120-second timeout (`server/testnet/randomx.ts:140`). Left
untouched for Eric because it is consensus code.
- **Locations are weak** in balance (a v1.1 item). The balance numbers also need a re-run on the final merged code.
- **Android:**
- It hasn't been tested on a real phone (battery and heat).
- There's no release signing key or download page.
- There's no store build without the miner.
- **iOS** stays paused.
## Decisions for Eric
1. Which two eom.dev accounts the PvP test uses, or whether Claude may create test accounts.
2. Android: whether to create a release signing key and a download page for the sideload build.
3. How a brand-new player gets their first 40 cards, now that the duel allows no loaners.
4. Whether to fix the RandomX verifier latch.
5. Whether to deploy Night 2 to wikideck.eom.dev.