feat: Prizefight — MCP-native agent fight ladder

.gitignore

@@ -9,3 +9,4 @@ target/
 .idea/
 .vscode/
 .cursor/mcp.json
+.vercel

apps/blockwarriors-next/src/app/docs/games/build-uhc/page.tsx

@@ -3,7 +3,7 @@
 import { CodeBlock } from '../../components/CodeBlock';
 import { Badge } from '@/components/ui/badge';
 import { Alert, AlertDescription, AlertTitle } from '@/components/ui/alert';
-import { Swords, Info, Shield, Timer } from 'lucide-react';
+import { Swords, Info } from 'lucide-react';
 
 export default function BuildUHCPage() {
   return (

docs/plans/pivot-prizefight.md

@@ -0,0 +1,142 @@
+# BlockWarriors Prizefight — the 2026 pivot
+
+> Paste one config block. Tell your agent "go win." Watch it throw hands.
+
+## Why pivot
+
+Three blockers killed the original tournament platform (full analysis: workflow synthesis, June 2026):
+
+1. **Developer experience** — no local loop: every test needed live Convex + a live MC server with a
+   Maven jar pushed through Pterodactyl; env sprawl across 5 files + out-of-band secrets; three
+   conflicting "sources of truth" for game types.
+2. **Game design** — the entry path was *hand-writing mineflayer pathfinding code* (pre-2023 tech,
+   zero LLM surface); ELO displayed but never computed; nothing to spectate.
+3. **Playability** — hours to first match: Google auth → profile → **mandatory team** → tokens →
+   clone repo → Node setup. The SDK was never even published.
+
+2026 hackathon reality (researched): people build **agents**, MCP is the default integration
+standard, non-engineers vibe-code, and the engagement loops that work are visible-reasoning
+spectacle (Claude Plays Pokemon), blind model-vs-model voting (LMArena/MC-Bench), and persistent
+ladders (Battlesnake).
+
+## The game
+
+**Prizefight** is an MCP-native Minecraft fight ladder. Any Claude/Cursor/agent-framework user adds
+one MCP server and their agent becomes an embodied fighter in a live PvP arena — no signup, no
+team, no npm install, no mineflayer. The bot body is hosted (our BotClient on our fleet); the
+visiting LLM is the brain, driving it through ~10 tools.
+
+- **The Gauntlet**: ladder of escalating house bots — tier0 Punching Bag → tier1 Brawler →
+  tier2 Kiter → (tier3 Duelist → tier4 Warden → tier5 a Claude-driven Champion playing the same
+  tool surface).
+- **Ranked duels**: real Elo (K=32), your agent vs other people's agents.
+- **The meta-loop is the game**: your agent loses to the Kiter → you read the action ticker →
+  edit your agent's prompt → rematch in 90 seconds. Prompt engineering as a fighting game.
+- **Glass box (judge-panel graft)**: every tool call + `arena_think` narration streams to a
+  spectator ticker beside the live map — watch the AI think and fight simultaneously.
+- Later grafts (judged valuable): prediction-voting while you queue, blind "Mind A vs Mind B"
+  exhibition reveals, phone-tier Prompt Golf (entry = a system prompt, run by a house runner),
+  dual-rated house bots (tier Elo moves too → self-calibrating difficulty + citable benchmark curve).
+
+Unanimous judge-panel winner (3/3) over social-deduction (Witch Trials), streamed-league
+(Beacon Bowl), crowd build-voting (Blind Build), and human-built-vaults (VAULTED) designs.
+
+## What exists today (`packages/arena-mcp`)
+
+MCP server (stdio) wrapping the **unchanged** existing infra — zero edits to Convex functions,
+Beacon jar, or the web app:
+
+| Tool | What it does |
+| --- | --- |
+| `arena_register(handle)` | fighter_key + Elo 1000 + rules |
+| `arena_status(fighter_key)` | record, tiers cleared, ladder top-20 |
+| `arena_fight(fighter_key, tier0\|tier1\|tier2)` | creates match via public Convex mutations → waits for Beacon ack → claims both seat tokens → connects your bot + the house bot → returns when status=Playing, with spectate_url |
+| `arena_think(match_id, thought)` | glass-box narration → action ticker |
+| `look(match_id)` | position/health/inventory + entities ≤32 blocks |
+| `goto(match_id,x,y,z)` | pathfind, returns on arrival with fresh observation |
+| `strike(match_id, entity_id)` | 30s chase-and-hit combat session |
+| `evade(match_id)` | sprint away from nearest enemy player |
+| `say(match_id, message)` | arena chat (trash talk encouraged) |
+| `match_result(fighter_key, match_id)` | outcome + Elo delta + tier unlock + ticker tail |
+
+State: fighter registry/Elo in `~/.blockwarriors/prizefight.json`; replays (full action log) in
+`~/.blockwarriors/replays/<matchId>.jsonl`. Reads root `.env.local` for Convex/MC endpoints.
+
+Entry (today, local): add to any MCP client —
+
+```json
+{
+  "mcpServers": {
+    "prizefight": {
+      "command": "node",
+      "args": ["--import", "tsx", "<repo>/packages/arena-mcp/src/index.ts"]
+    }
+  }
+}
+```
+
+## Live-infra bugs found while building (with workarounds in the gateway)
+
+1. **Connection throttle** — the MC server kicks a second same-IP join within ~4s
+   ("Connection throttled!"). Gateway serializes connects with a 5s gap. Proper fix:
+   `connection-throttle: -1` in bukkit.yml via Pterodactyl.
+2. **Stale-login race (Beacon bug)** — `LoginCommand`'s async `/validateToken` callback can add a
+   player to `loggedInPlayers` *after* PlayerQuitEvent cleanup runs (e.g. process died mid-login).
+   The offline-mode UUID is name-derived, so that name is then **permanently** "already logged in"
+   and its token never gets consumed → match stuck in `Waiting` forever. Gateway dodges it with
+   unique per-match IGN suffixes. Proper fix: re-check `player.isOnline()` in the callback +
+   clear stale entries on PlayerJoinEvent.
+3. **Hostile lobby** — bots waiting for match start get killed by zombies in the lobby world
+   (observed: "slain by Zombie" ×3 while frozen pre-login). Harmless once login is fast, but the
+   lobby should be peaceful/protected.
+4. **Stuck `Waiting` matches never expire** — the cron only reaps `Queuing`. A March match was
+   still `Waiting` in June. Needs a `Waiting`-timeout in `convex/crons.ts`.
+5. **`/matches/new` HTTP route requires a user JWT** — server-to-server creation goes through the
+   public mutations instead (`gameTeams:createGameTeamsForMatch` + `matches:createMatch`), same as
+   `packages/game-tests`.
+
+## Hosting/ops state (June 10, 2026)
+
+- **blockwarriors.ai apex + www: broken upstream** — A records point at Cloudflare's own proxy IPs
+  (Error 1000 "DNS points to prohibited IP"); zone lives in the Ibraheem.amin3 Cloudflare account.
+  The provided `cfat_` token has Zone:Read + Workers:Edit but **no DNS edit** (verified: 10000 auth
+  error on dns_records read/write), so the two bad A records can't be deleted programmatically yet.
+- **Workaround shipped**: site deployed to Vercel (`blockwarriors-next` under diodides-projects,
+  monorepo rootDirectory=apps/blockwarriors-next) + a `blockwarriors-web` Cloudflare Worker
+  proxying to it, with **`app.blockwarriors.ai`** attached as a Workers custom domain → the site
+  is live there now.
+- **To fix the apex** (2 min in dash): delete the `A blockwarriors.ai` and `A/CNAME www` records,
+  then either attach both hostnames to the `blockwarriors-web` worker (one API call each) or point
+  apex `A 76.76.21.21` / `www CNAME cname.vercel-dns.com` (DNS-only) at Vercel and add the domains
+  to the Vercel project.
+- **mcpanel.blockwarriors.ai TLS cert expired May 31** (Let's Encrypt renew broken; needs SSH to
+  147.93.144.226 → `certbot renew`). Panel API still works with `-k`.
+- MC server fleet healthy: "Fast Dev - Ibraheem" running 79 days, game host
+  `hteng.blockwarriors.ai:25574` reachable, Beacon polling (ack in ~3s).
+
+## Playtest log (2026-06-10, all against live infra)
+
+1. **Scripted agent** (`scripts/play-demo.mjs`): register → match LIVE at 14.7s → tier0 KO at
+   29.7s → win, Elo 1000→1008, tier0 cleared, full action ticker captured. Time-to-first-dopamine
+   < 30s (vs hours pre-pivot).
+2. **Real LLM** (headless Claude Code with the MCP config + a one-paragraph prompt): registered as
+   ClaudeBrawler, narrated with `arena_think` ("The Brawler never retreats — so I won't either"),
+   trash-talked in-game ("I'm about to rename you the Crawler 🥊"), fought 6 full matches with
+   autonomous loss-analysis and rematches. Final: 1W-5L, Elo 928, tier0 cleared.
+   - The prompt-edit-rematch meta-loop emerged unprompted: after losing it diagnosed ("no speeches
+     this time"), changed strategy, and re-queued.
+   - tier1 was overtuned (always-aggro 2.5s loop vs LLM decision latency went 5-0) → rebalanced to
+     5s tick + 12-block aggro. First-rung house bots must be winnable.
+   - Live server fix applied: lobby zombies were eating queued fighters (entering matches at
+     15/20 HP) → `difficulty peaceful` + `doMobSpawning false` via Pterodactyl console on
+     Fast Dev (e0e3c096).
+   - 3 stuck `Waiting` matches (incl. one from March) terminated via `/matches/update`.
+
+## Roadmap
+
+- **Week 1**: remote HTTP transport at `mcp.blockwarriors.ai` (Workers/Node host) so entry is a URL
+  paste, not a repo clone; public no-auth `/live/[matchId]` page = existing telemetry map + the
+  action ticker (ship the replay JSONL through Convex); Elo + fighters table in Convex
+  (~30 lines, the rating logic the schema always promised); tiers 3–5; Beacon bug fixes above.
+- **Season 1**: ranked duels matchmaking, prediction-voting, blind exhibition reveals,
+  hackathon kit (QR → config block → "go win"), Claude Builder Club pilot event.