jedarden f9511df069 docs(plan): update attack radius and zone parameters to match implementation

Plan §3.4 and §3.7.1 now reflect the actual engine values:
- 2-player attack_radius2: 36 → 25 (6 tiles → 5 tiles)
- 2-player zone_min_radius: 3 → 2 (diameter 6 → 4 tiles)

These changes were made in commits 04b7e89 and ceb2de4 to achieve
the target combat density (65-80% for 2-player). Verification confirms
85% combat density with gatherer+rusher bots.

Closes: bf-vrh2

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

2026-05-25 13:31:10 -04:00

229 KiB

Raw Blame History

AI Code Battle — Implementation Plan

1. Overview

AI Code Battle is a competitive bot programming platform where participants write HTTP servers that control units on a grid world. The game engine orchestrates matches asynchronously, stores replays, and serves a web platform where visitors watch rendered game replays and browse leaderboards. Matches are never live — they are evaluated offline by match workers and presented as completed replays.

The platform ships with several built-in strategy bots, each deployed as its own container, serving as both opponents for new participants and reference implementations for the HTTP protocol.

2. System Architecture

The platform uses a static-first architecture. The public-facing product is a Cloudflare Pages static site — all data visitors see (leaderboards, match history, bot profiles, replays) is pre-computed JSON served from the CDN. All compute runs in the apexalgo-iad Kubernetes cluster (Rackspace Spot), which acts as a match factory: it runs battles, generates replays, and periodically publishes the updated site to Pages.

Replay files use tiered storage: Backblaze B2 is the permanent cold archive (all replays, forever), and Cloudflare R2 is a warm cache for recent replays (free tier, ≤10 GB). The browser fetches replays from R2; old replays that have aged out of R2 fall back to B2 (free egress via Cloudflare Bandwidth Alliance).

Cloudflare (Static Tier)

Cloudflare Pages (ai-code-battle.pages.dev): Hosts the static SPA shell and all pre-computed JSON data — leaderboards, bot profiles, match indexes, series, seasons, evolution data, blog posts. Updated every ~90 minutes by the K8s index builder via wrangler pages deploy. Global CDN, zero-config TLS, instant cache invalidation on deploy.
Cloudflare R2 (warm replay cache): Stores recent replay files and per-match metadata — the subset of data too numerous for Pages' 20K file limit and too hot for B2-only serving. Capped at ≤10 GB to stay within the free tier. The index builder manages the R2 lifecycle: promotes recent replays from B2, prunes old ones when approaching the cap.

Backblaze B2 (Cold Archive)

B2 bucket: Permanent archive for all replay files and match metadata. Match workers upload directly to B2 after each match. B2 is the source of truth for replay data — R2 is just a CDN cache in front of it. Free egress to browsers via Cloudflare Bandwidth Alliance (B2 + Cloudflare = zero egress fees). S3-compatible API.

apexalgo-iad (Compute Tier)

All backend compute runs in the ai-code-battle namespace:

Matchmaker Deployment: Internal scheduler. Queries active bots from PostgreSQL, computes pairings, enqueues job IDs into Valkey. Also handles health checking and stale job reaping. No external exposure.
PostgreSQL (CNPG): Source of truth for all structured data — bots, matches, jobs, ratings, predictions, series, seasons.
Valkey: Job queue for match jobs, ephemeral caching.
Match Worker Deployment: Dequeues jobs from Valkey (BRPOP), runs matches, uploads replay JSON to B2 (cold archive), writes results to PostgreSQL.
Strategy Bot Deployments (x6): Built-in bots as HTTP servers on cluster-internal Services.
Evolved Bot Deployments (0-50): LLM-generated bots, same pattern.
Evolver Deployment: LLM evolution pipeline. Reads match data from PostgreSQL, generates candidates, tests them, deploys successful bots as new K8s Deployments. Writes evolution metadata to PostgreSQL. Self-restarts every 4h.
Index Builder Deployment: Sleep-loop (15 min cycle). Reads PostgreSQL, generates all JSON index files, deploys them to Cloudflare Pages via wrangler pages deploy. Manages R2 warm cache (copies recent replays from B2, prunes old replays to stay within free tier). Self-restarts every 4h.

Go API (deferred)

A public Go API at api.aicodebattle.com is planned for social features (predictions, commenting, voting) and third-party bot registration. This is not required for the core match loop — the v1 system is fully static. The API will be added when interactive features are needed.

Data Architecture

Data is split across three tiers by access pattern and volume:

Cloudflare Pages (SPA + pre-computed indexes, deployed by index builder):

Pages project (ai-code-battle.pages.dev):
├── index.html, app.html, ...          (SPA shell)
├── js/                                 (bundled TypeScript application)
│   ├── app.js                          (SPA router, data fetching)
│   ├── replay-viewer.js                (Canvas replay renderer)
│   └── charts.js                       (win probability, meta charts)
├── css/                                (stylesheets)
├── docs/                               (protocol spec, replay format, guides)
├── img/                                (logos, icons, UI assets)
├── embed.html                          (lightweight embeddable replay player)
├── data/                               (pre-computed JSON indexes, rebuilt every ~15 min)
│   ├── leaderboard.json
│   ├── bots/
│   │   ├── index.json
│   │   └── {bot_id}.json
│   ├── matches/
│   │   └── index.json                  (recent matches, paginated)
│   ├── series/
│   │   ├── index.json
│   │   └── {series_id}.json
│   ├── seasons/
│   │   ├── index.json
│   │   └── {season_id}.json
│   ├── playlists/
│   │   └── {slug}.json
│   ├── meta/
│   │   ├── archetypes.json
│   │   └── rivalries.json
│   ├── evolution/
│   │   ├── lineage.json
│   │   └── meta.json
│   └── blog/
│       ├── index.json
│       └── posts/{slug}.json
└── maps/
    ├── index.json
    └── {map_id}.json

Cloudflare R2 (warm replay cache, free tier ≤10 GB):

R2 bucket:
├── replays/
│   └── {match_id}.json.gz             (recent replay files, promoted from B2)
├── matches/
│   └── {match_id}.json                (recent per-match metadata)
├── thumbnails/
│   └── {match_id}.png                 (auto-generated match thumbnails)
└── cards/
    └── {bot_id}.png                    (bot profile card images)

Backblaze B2 (cold archive, permanent):

B2 bucket:
├── replays/
│   └── {match_id}.json.gz             (ALL replay files, forever)
├── matches/
│   └── {match_id}.json                (ALL per-match metadata)
├── thumbnails/
│   └── {match_id}.png                 (ALL thumbnails)
└── cards/
    └── {bot_id}.png                    (ALL bot card images)

Data loading pattern in the SPA:

// SPA shell + index data from Cloudflare Pages (same origin)
const PAGES = ''  // relative — same origin as the SPA
// Replays from R2 warm cache (recent), B2 cold archive (old)
const R2 = 'https://r2.aicodebattle.com'     // or R2 public URL
const B2 = 'https://b2.aicodebattle.com'     // B2 via Cloudflare CDN

// Leaderboard, bot profiles, match indexes — all from Pages (same origin):
const lb = await fetch(`${PAGES}/data/leaderboard.json`).then(r => r.json())

// Replay viewer — try R2 warm cache first, fall back to B2 cold archive:
async function fetchReplay(matchId) {
  const r2 = await fetch(`${R2}/replays/${matchId}.json.gz`)
  if (r2.ok) return r2
  return fetch(`${B2}/replays/${matchId}.json.gz`)  // cold fallback
}

// Match metadata — same R2-then-B2 pattern:
const meta = await fetch(`${R2}/matches/${matchId}.json`)
  .then(r => r.ok ? r : fetch(`${B2}/matches/${matchId}.json`))
  .then(r => r.json())

Cache behavior:

Pages assets: Cloudflare Pages handles caching automatically. Deploys via wrangler pages deploy invalidate the cache globally. Index data is at most ~90 minutes stale (the index builder's cycle time).
R2 objects (warm cache): Served with appropriate Cache-Control headers:
- replays/*.json.gz: immutable, max-age=31536000 (content-addressed)
- matches/*.json: immutable, max-age=31536000 (content-addressed)
- thumbnails/, cards/: max-age=86400 (regenerated rarely)
B2 objects (cold archive): Same cache headers. B2 egress through Cloudflare Bandwidth Alliance = zero egress fees. Cloudflare CDN caches B2 responses, so frequently accessed cold replays still perform well.

Tiered storage lifecycle:

Match worker completes a match → uploads replays/{match_id}.json.gz and matches/{match_id}.json to B2 (cold archive, permanent)
Worker writes match result to PostgreSQL (scores, ratings, metadata)
Index builder (every ~15 min) reads new results from PostgreSQL, rebuilds all JSON index files, deploys to Pages via wrangler pages deploy
Index builder promotes recent replays from B2 to R2 (warm cache)
Index builder prunes oldest replays from R2 when approaching 10 GB cap
Browser loads SPA + indexes from Pages, fetches replays from R2 (warm) with B2 fallback (cold)

Storage budget:

R2 (warm cache): ≤10 GB free tier. At ~50 KB/replay (gzipped), holds ~200K replays. At 60 matches/hour, that's ~139 days of warm replays. Class A writes: index builder promotes ~1,440 replays/day = ~43K/month (well within 1M/month free tier).
B2 (cold archive): First 10 GB free, $0.006/GB/month after. At 60 matches/hour, ~2.2 GB/month. Year one: ~26 GB = ~$0.10/month. Free egress via Cloudflare Bandwidth Alliance.
Pages: 20K file limit per deployment. Only SPA + JSON indexes — well within limits (replays are on R2/B2, not Pages).

┌────────── Cloudflare ──────────────────────────────────┐
│                                                          │
│  ┌─────────────────────┐   ┌─────────────────────────┐  │
│  │  Cloudflare Pages    │   │  Cloudflare R2           │  │
│  │  (static site)       │   │  (warm replay cache)     │  │
│  │                      │   │                           │  │
│  │  SPA shell (HTML/    │   │  replays/*.json.gz       │  │
│  │  JS/CSS)             │   │  matches/*.json           │  │
│  │  data/*.json indexes │   │  thumbnails/*.png         │  │
│  │  maps/*.json         │   │  cards/*.png              │  │
│  │  docs/, img/         │   │  ≤10 GB free tier         │  │
│  └─────────────────────┘   └─────────────────────────┘  │
│        ▲ wrangler deploy         ▲ S3 PUT (promote)      │
└────────┼─────────────────────────┼───────────────────────┘
         │                         │
┌────────┼─────────────────────────┼───────────────────────┐
│        │  apexalgo-iad cluster — ai-code-battle ns        │
│        │                         │                         │
│  ┌─────┴──────────────────┐      │                         │
│  │  Index Builder Dep.     │      │                         │
│  │  Reads PostgreSQL,      │──────┘                         │
│  │  generates JSON indexes,│                                │
│  │  deploys to Pages,      │                                │
│  │  promotes replays to R2,│                                │
│  │  prunes R2 warm cache   │                                │
│  └─────────────────────────┘                                │
│                                                              │
│  ┌──────────────────────────────────────────────────────┐   │
│  │  Matchmaker Dep.   (internal only — no ingress)       │   │
│  │  Pairings, job enqueue, health check, stale reaper    │   │
│  └────────┬─────────────────────────────────────────────┘   │
│           │                                                  │
│  ┌────────▼──────────────────────────────────────────────┐  │
│  │  CNPG PostgreSQL (cnpg-apexalgo)                       │  │
│  │  bots, matches, jobs, ratings, etc.                    │  │
│  └────────────────────────────────────────────────────────┘  │
│                                                              │
│  ┌────────────────────────────────────────────────────────┐  │
│  │  Valkey (Redis-compatible)                              │  │
│  │  Job queue (acb:jobs:pending)                           │  │
│  └────────────────────────────────────────────────────────┘  │
│                                                              │
│  ┌────────────────────────────────────────────────────────┐  │
│  │  Match Workers (Deployment, 1-10 pods)                  │  │
│  │  BRPOP from Valkey, run matches, upload replays to B2,  │  │
│  │  write results to PostgreSQL                            │  │
│  └────────────────────────────────────────────────────────┘  │
│                                                   │          │
│  ┌────────────────────────────────────────────────┼───────┐  │
│  │  Bot Containers (Deployments)                  │       │  │
│  │  Strategy (×6) + Evolved (0–50)                │       │  │
│  └────────────────────────────────────────────────┼───────┘  │
│                                                   │          │
│  ┌────────────────────────────────────────────────┼───────┐  │
│  │  Evolver (Deployment)                          │       │  │
│  │  LLM pipeline, writes evolution data to PG     │       │  │
│  └────────────────────────────────────────────────┼───────┘  │
│                                                   │          │
│  ┌─────────────────────────────────────────────┐  │          │
│  │  ArgoCD — syncs K8s manifests from git       │  │          │
│  │  Argo Workflows — CI builds, image pushes    │  │          │
│  └─────────────────────────────────────────────┘  │          │
│                                                   ▼          │
└───────────────────────────────────────────────────┼──────────┘
                                                    │
                                      ┌─────────────▼──────────┐
                                      │  Backblaze B2           │
                                      │  (cold replay archive)  │
                                      │                         │
                                      │  replays/*.json.gz      │
                                      │  matches/*.json         │
                                      │  thumbnails/*.png       │
                                      │  cards/*.png            │
                                      │  ALL data, forever      │
                                      └─────────────────────────┘

Component Summary

Component	Where	Role
Cloudflare Pages	Cloudflare	Static site: SPA (HTML/JS/CSS) and pre-computed JSON index files. Updated every ~90 min by the index builder via `wrangler pages deploy`. Global CDN with automatic cache invalidation.
Cloudflare R2	Cloudflare	Warm replay cache: recent replays, match metadata, thumbnails, bot cards. Free tier ≤10 GB. Managed by the index builder (promote from B2, prune when approaching cap).
Backblaze B2	Backblaze	Cold archive: ALL replays and match data, permanently. Workers upload directly after each match. Free egress via Cloudflare Bandwidth Alliance.
Matchmaker	Deployment (ai-code-battle ns)	Internal scheduler: computes pairings, enqueues jobs to Valkey, health checks bots, reaps stale jobs. No external exposure.
PostgreSQL	CNPG cluster (cnpg ns, `cnpg-apexalgo`)	Relational database — bot registry, match queue, ratings, results, series, seasons. Source of truth for structured data.
Valkey	Cluster service	Job queue (`acb:jobs:pending`), ephemeral caching.
Match Workers	Deployment (ai-code-battle ns)	Stateless match execution — BRPOP from Valkey, run simulation, upload replay to B2, write result to PostgreSQL.
Bot Containers	Deployments + Services (ai-code-battle ns)	Strategy bots (x6) + evolved bots (0-50) — HTTP servers called by workers during matches via cluster-internal Service DNS.
Evolver	Deployment (ai-code-battle ns)	Evolution pipeline — reads lineage/meta from PostgreSQL, generates candidates, writes evolution data to PostgreSQL.
Index Builder	Deployment (ai-code-battle ns)	Sleep-loop (15 min cycle). Reads PostgreSQL, generates JSON indexes, deploys to Pages. Promotes recent replays from B2 to R2 warm cache. Prunes R2 to stay within free tier. Self-restarts every 4h.
Go API	Deferred	Social features (predictions, comments, voting) and third-party bot registration. Not required for v1.
ArgoCD	Cluster (argocd ns)	GitOps: syncs all K8s manifests from git. All deployments are declarative.
Argo Workflows	Cluster (argo ns)	CI pipelines: builds container images, pushes to Forgejo registry, builds static site.

3. Game Mechanics

3.1 Map & Grid

The game plays on a toroidal grid — a rectangular grid that wraps both horizontally and vertically (no edges, no corners). This eliminates positional advantages from map boundaries.

Tile types:

Tile	Symbol	Description
Open	`.`	Passable empty tile
Wall	`#`	Impassable barrier
Energy	`*`	Collectible resource (respawns)
Core	`C`	Player spawn point (owned by a player)

Grid parameters (configurable per match):

Parameter	Default	Range	Description
`rows`	60	30–120	Grid height
`cols`	60	30–120	Grid width
`wall_density`	0.15	0.05–0.30	Fraction of tiles that are walls
`energy_nodes`	20	8–50	Number of energy spawn locations
`cores_per_player`	1	1–2	Starting cores per player

3.2 Units (Bots)

Each player controls bots — mobile units on the grid.

Bots move one tile per turn in a cardinal direction: N, E, S, W
Bots that do not receive a move order hold position
Bots are binary — alive or dead, no hit points
A bot ordered into a wall tile stays in place (order ignored)
Two friendly bots ordered to the same tile: both die (self-collision)
A bot ordered onto a tile occupied by a stationary enemy: both die

Each player starts with one bot spawned at each of their cores.

3.3 Energy & Economy

Energy is the sole resource. It is used to spawn new bots.

Energy nodes:

Fixed positions on the map that periodically produce collectible energy
Energy appears on a node every energy_interval turns (default: 10)
When energy is present on a node, it is visible to any player who can see the tile

Collection:

A bot adjacent to (or on) an energy tile collects it if no enemy bot is also adjacent to that energy
If bots from multiple players are adjacent to the same energy, the energy is destroyed — nobody gets it (contested resources are denied)
Collection happens after combat resolution each turn

Spawning:

Cost: 3 energy per bot
Spawning happens automatically when a player has ≥3 energy and an unoccupied, unrazed core
One bot spawns per core per turn maximum
If a player has multiple cores and enough energy, one bot spawns at each eligible core simultaneously
Spawn priority: core that has been idle longest

3.4 Combat

Combat uses a focus fire algorithm inspired by the aichallenge ants system. This rewards formations and positioning over raw unit count.

Attack radius: squared Euclidean distance ≤ attack_radius2. Tuned per player count to achieve target combat density (65-80% for 2-player, 100% for 3+):

Player Count	`attack_radius2`	Distance	Rationale
2-player	25	~5 tiles	Tuned with zone_min_radius=2 to achieve 65-80% combat density; zone forces contact
3+ player	12	~3.46 tiles	Higher player density provides sufficient contact with smaller radius

The default (3+ player) value of 12 includes cardinal and diagonal neighbors plus two more rings.

Resolution (simultaneous):

for each bot B on the grid:
    enemies_of_B = count of enemy bots within attack_radius2 of B
    for each enemy E within attack_radius2 of B:
        enemies_of_E = count of E's enemies within attack_radius2 of E
        if enemies_of_B >= enemies_of_E:
            mark B as dead
            break  (B is already dead, no need to check further)

All deaths are resolved simultaneously — no cascading within a single turn.

Key properties:

2v1: the lone bot dies, the pair survives (superior numbers win cleanly)
1v1: both die (mutual destruction)
Tight formations are defensive — a cluster facing scattered enemies takes fewer losses because each bot in the cluster has a lower enemy count
Multi-player battles create emergent alliances and third-party exploitation

3.5 Fog of War

Each player has limited visibility. Only tiles within vision_radius2 (default: 49, ~7 tiles) of any owned bot are visible.

What players see within their vision:

All tile types (open, wall, energy, core)
Enemy bots and their owner IDs
Dead bots (for one turn after death)

What players do NOT see:

Anything outside their collective vision radius
How much energy opponents have
Total number of opponents (discovered through play)

Walls are sent every turn they are visible (no incremental discovery state — keeps the protocol stateless-friendly for HTTP bots).

3.6 Scoring & Win Conditions

Scoring:

Each player starts with 1 point per core owned
Capturing a core (enemy bot moves onto an undefended enemy core): +2 points to capturer, −1 point to owner; core is razed
Razed cores stop spawning but the player continues with remaining bots
Energy collected: tracked as a tiebreaker statistic (not added to score)
Bots eliminated: tracked as a statistic

Win conditions (checked in order):

Condition	Trigger	Resolution
Sole Survivor	Only one player has living bots	That player wins; bonus +2 per surviving enemy core
Annihilation	All players eliminated simultaneously	Draw
Dominance	One player controls ≥80% of all bots for 100 consecutive turns	That player wins
Turn Limit	Turn count reaches `max_turns` (default: 500)	Highest score wins; ties broken by energy collected, then bots alive

3.7 Turn Structure

Each turn executes in a strict, deterministic sequence:

1.  Send game state to all players (HTTP POST, filtered by fog of war)
2.  Await responses (up to 3-second timeout per player, in parallel)
3.  Validate all responses against schema
4.  Phase: MOVE        — execute valid movement orders
5.  Phase: COMBAT      — resolve focus-fire algorithm, remove dead bots
6.  Phase: ZONE        — shrinking zone kills bots outside the safe radius
                          (Forces bots into contact range for combat engagement.
                          Zone starts at a configured turn and shrinks over time
                          until reaching a minimum radius. Bots outside the zone
                          are killed immediately, creating pressure to fight.)
7.  Phase: CAPTURE     — enemy bots on undefended cores raze them
                          (A core is undefended if no bot belonging to
                          the core's owner occupies the core's tile after
                          the attack phase resolves. An enemy bot on an
                          undefended core's tile razes it.)
8.  Phase: COLLECT     — uncontested energy adjacent to bots is collected
9.  Phase: SPAWN       — players with ≥3 energy spawn bots at eligible cores
10. Phase: ENERGY_TICK — energy nodes on their interval produce new energy
11. Phase: ENDGAME     — check win conditions
11. Record turn state for replay

All player requests in step 1 are sent concurrently. Responses are collected with the 3-second deadline. The engine does not proceed to step 3 until all responses are in or timed out. All player requests in step 1 are sent concurrently. Responses are collected with the 3-second deadline. The engine does not proceed to step 3 until all responses are in or timed out.

3.7.1 Zone Parameters

The shrinking zone forces bots into contact range, ensuring combat engagement rather than pure energy farming. Zone parameters are tuned per player count:

Parameter	2-Player	3+ Player	Description
ZoneStartTurn	10	10	Turn when zone begins shrinking
ZoneShrinkInterval	1	1	Turns between shrink steps
ZoneShrinkStep	2	2	Tiles to shrink each step
ZoneMinRadius	2	1	Minimum zone radius (stops shrinking)

Design rationale:

ZoneStartTurn = 10: Starts early to force combat before energy farming dominates. Both 2-player and 3+ use the same start turn for consistent forcing function timing.
ZoneShrinkInterval = 1: Shrinks every turn creates steady, predictable pressure. Faster than the original 2-turn interval to ensure bots reach contact range before the match is decided by energy alone.
ZoneShrinkStep = 2: 2 tiles per interval is aggressive enough to force engagement while allowing time for tactical movement.
ZoneMinRadius = 2 (2-player): Final zone diameter (4 tiles) ≤ 2 × attack radius (10 tiles), ensuring bots at opposite zone edges are within attack range (5 tiles).
ZoneMinRadius = 1 (3+ player): Final zone diameter (2 tiles) is smaller than attack radius (3.5 tiles), guaranteeing any two bots in the final zone are within attack range. This is necessary because 3+ player maps have higher player density and a smaller attack radius (3.5 tiles vs 5 tiles for 2-player).

Combat density metrics (verified with local testing):

2-player: ~65-80% of matches have combat_deaths; ~1 death per 20 turns
6-player: 100% of matches have combat_deaths; ~1 death per 5-6 turns

The zone achieves its forcing function: bots must fight or die, while maintaining strategic depth (early game positioning matters, not just pure chaos).

3.8 Map Generation

Maps are generated offline and stored in the map library. They are not generated on-the-fly during matches.

Symmetry requirements:

2-player maps: 180° rotational symmetry (point symmetry through center)
3-player maps: 120° rotational symmetry
4-player maps: 90° rotational symmetry
6-player maps: 60° rotational symmetry

Generation algorithm:

Generate one sector (1/N of the map for N players)
Place walls using cellular automata (random seed → smooth with neighbor rules)
Place cores and energy nodes within the sector
Validate connectivity: BFS from core must reach all energy nodes and the sector boundary
Mirror/rotate the sector to fill the full map
Validate full-map connectivity: all cores must be reachable from each other
Store the map with metadata (player count, dimensions, wall density)

Map library:

Pre-generated pool of 50+ maps per player count (2, 3, 4, 6)
Maps are curated — auto-generated then play-tested with strategy bots
Matchmaking selects the least-recently-used map for each match

4. Communication Protocol

4.1 HTTP Interface

The game engine communicates with bots via HTTP POST requests. Each bot exposes a single endpoint.

Bot endpoint: POST {bot_base_url}/turn

The engine sends the game state as a JSON body. The bot responds with its moves as a JSON body. No other endpoints are required from the bot (though /health is recommended for registration validation).

Request flow per turn:

Engine                          Bot
  │                              │
  │  POST /turn                  │
  │  Headers: auth + metadata    │
  │  Body: game state JSON       │
  │─────────────────────────────►│
  │                              │  (bot computes moves)
  │  200 OK                      │
  │  Body: moves JSON            │
  │◄─────────────────────────────│
  │                              │

4.2 Game State Schema (Engine → Bot)

{
  "match_id": "m_7f3a9b2c",
  "turn": 42,
  "config": {
    "rows": 60,
    "cols": 60,
    "max_turns": 500,
    "vision_radius2": 49,
    "attack_radius2": 12,
    "spawn_cost": 3,
    "energy_interval": 10
  },
  "you": {
    "id": 0,
    "energy": 7,
    "score": 3
  },
  "bots": [
    { "row": 10, "col": 15, "owner": 0 },
    { "row": 12, "col": 15, "owner": 0 },
    { "row": 30, "col": 40, "owner": 1 }
  ],
  "energy": [
    { "row": 20, "col": 25 }
  ],
  "cores": [
    { "row": 5, "col": 5, "owner": 0, "active": true },
    { "row": 55, "col": 55, "owner": 1, "active": true }
  ],
  "walls": [
    { "row": 10, "col": 10 },
    { "row": 10, "col": 11 }
  ],
  "dead": [
    { "row": 15, "col": 20, "owner": 1 }
  ]
}

Schema rules:

bots, energy, cores, walls, dead -- only includes tiles within the player's collective vision
owner IDs are consistent within a match but randomized per match (player 0 is always "you")
config is identical for all players and does not change between turns
walls are sent every turn they are visible (stateless -- bot does not need to track previously seen walls, though smart bots will)
dead contains bots that died on the previous turn (visible for one turn)

Future additive fields: the game state schema is designed for forward compatibility. Future seasons may add optional fields to config (e.g., season_id, rules_version, special_tiles, terrain) without breaking existing bots. See the seasonal backward compatibility rules in §14.9. Bots that do not read new fields continue to function normally.

4.3 Move Schema (Bot -> Engine)

{
  "moves": [
    { "row": 10, "col": 15, "direction": "N" },
    { "row": 12, "col": 15, "direction": "E" }
  ],
  "debug": {
    "reasoning": "3 energy within 5 tiles east; enemy cluster north — avoiding",
    "targets": [
      { "row": 20, "col": 25, "label": "energy", "priority": 0.9 }
    ]
  }
}

The debug field is entirely optional. When present, it is stored in the replay for visualization in the replay viewer but is never parsed or acted upon by the engine. See §13.1 for the full debug telemetry specification.

Validation rules:

moves must be an array (may be empty -- all bots hold position)
Each move must reference a (row, col) where the player owns a bot
direction must be one of: "N", "E", "S", "W"
Duplicate (row, col) entries: first valid entry wins, rest ignored
Moves referencing tiles with no owned bot: ignored
Moves into walls: ignored (bot stays)
Any response that fails top-level schema validation: entire response discarded, all bots hold
The engine never parses, evaluates, or interprets any field beyond moves[].row, moves[].col, moves[].direction (and the optional debug field, which is pass-through to the replay)

4.4 Authentication (HMAC Shared Secret)

Each registered bot has a shared secret generated at registration time. The secret is known only to the bot owner and the game engine. It authenticates both directions — the bot can verify requests came from the real game engine, and the engine can verify responses came from the real bot.

Engine → Bot (request signing):

Headers sent with every request:

X-ACB-Match-Id: m_7f3a9b2c
X-ACB-Turn: 42
X-ACB-Timestamp: 1711200000
X-ACB-Bot-Id: b_4e8c1d2f
X-ACB-Signature: <hex-encoded HMAC-SHA256>

Signature computation:

signing_string = "{match_id}.{turn}.{timestamp}.{sha256(request_body)}"
signature = HMAC-SHA256(shared_secret, signing_string)

The bot verifies:

Compute the expected signature from the headers and request body
Compare with X-ACB-Signature (constant-time comparison)
Verify X-ACB-Timestamp is within ±30 seconds of current time (prevents replay attacks)
If verification fails: bot should return 401 and ignore the request

Bot → Engine (response signing):

Response headers:

X-ACB-Signature: <hex-encoded HMAC-SHA256>

Signature computation:

signing_string = "{match_id}.{turn}.{sha256(response_body)}"
signature = HMAC-SHA256(shared_secret, signing_string)

The engine verifies the response signature. If invalid, the response is discarded (bots hold position). This prevents man-in-the-middle from injecting moves.

Why HMAC over OAuth/JWT/mTLS:

Minimal complexity — no token refresh, no certificate management
Bot developers add a single header computation, not an auth library
Symmetric: both sides can verify the other with the same secret
Sufficient for the threat model (prevent impersonation and tampering)

Secret management:

Secrets are generated as 256-bit random values, hex-encoded (64 characters)
Displayed once at registration time; bot owner must save it
Can be rotated via the web platform (old secret invalidated immediately)
Stored encrypted (AES-256-GCM) in the database. The master encryption key is held in an environment variable (from SealedSecret), never in the database. HMAC verification requires the raw secret, so hashing is not viable -- the engine decrypts on each request.

4.5 Timeout & Error Handling

Scenario	Behavior
Bot responds within 3s	Moves validated and applied normally
Bot responds after 3s	Response discarded; bots hold position for that turn
Bot returns non-200 status	Treated as timeout; bots hold position
Bot returns invalid JSON	Treated as timeout; bots hold position
Bot returns valid JSON failing schema	Entire response discarded; bots hold position
Bot connection refused	Bots hold position; engine retries next turn
Bot connection timeout (TCP)	Engine uses 2s connect timeout within the 3s budget
10 consecutive failures	Bot marked as crashed for this match; bots become inert for remaining turns

The bot is never killed or disconnected. Even after being marked crashed, the match continues -- the crashed bot's units simply hold position every turn until they are destroyed or the match ends.

Rating impact of crashes: Matches where a bot crashes or times out still count toward Glicko-2 ratings. The crashed bot receives a loss. This prevents intentional crashing as a loss-avoidance strategy.

5. Strategy Bots

Six built-in strategy bots serve as reference implementations and permanent ladder opponents. Each is implemented in a different programming language to demonstrate that the HTTP protocol is truly language-agnostic and to provide starter code for participants across the most popular ecosystems.

Each bot is deployed as its own container running a lightweight HTTP server.

Bot	Language	Complexity	Expected Rank
RandomBot	Python	Trivial	6th (floor)
GathererBot	Go	Low	4th–5th
RusherBot	Rust	Low	4th–5th
GuardianBot	PHP	Medium	3rd–4th
SwarmBot	TypeScript	Medium	1st–2nd
HunterBot	Java	High	1st–2nd

5.1 RandomBot — Python

Language rationale: Python is the most accessible language for newcomers. The random bot doubles as the simplest possible starter template — a participant can fork it and have a working bot in minutes.

Strategy: Makes uniformly random valid moves each turn.

Behavior:

For each owned bot, pick a random direction (N/E/S/W) or hold (20% chance)
No pathfinding, no memory, no awareness of enemies
Serves as the absolute baseline — any reasonable bot should beat this

Value: Ensures new participants have an easy opponent to test against. Rating floor anchor.

Implementation: Flask or bare http.server. ~50 lines of strategy code. HMAC verification via hmac stdlib module.

5.2 GathererBot — Go

Language rationale: Go is the same language as the game engine and platform services, making this the canonical "how to build a bot" reference. Demonstrates idiomatic Go HTTP server patterns.

Strategy: Maximize energy collection, avoid combat entirely.

Behavior:

BFS from each owned bot to the nearest visible energy
Assign each bot to the closest uncontested energy (greedy matching)
If an enemy bot is within vision, move away from it
Never voluntarily enters attack range of an enemy
Spawns bots as fast as energy allows

Value: Tests whether aggressive bots can actually close games or whether passive resource hoarding is dominant (it shouldn't be).

Implementation: net/http stdlib server. Shared game/ package with grid utilities, BFS, and distance calculations that participants can reuse.

5.3 RusherBot — Rust

Language rationale: Rust participants get maximum compute within the 3-second timeout. This bot demonstrates that Rust's performance advantage matters less than strategy — a dumb fast bot still loses to a smart slow one.

Strategy: Identify and rush the nearest enemy core as fast as possible.

Behavior:

BFS from each owned bot toward the nearest known enemy core
If no enemy core is known, spread out to explore (random walk with bias toward unexplored territory)
Ignores energy except incidentally (walks over it)
Ignores enemy bots unless they block the path
Spawns bots immediately and sends all toward the target

Value: Punishes bots that neglect defense. Tests whether the combat system allows pure aggression to dominate (it shouldn't — rusher bots will walk into defensive formations and die).

Implementation: axum or actix-web. Serde for JSON. HMAC via hmac and sha2 crates. Demonstrates Rust's zero-copy deserialization.

5.4 GuardianBot — PHP

Language rationale: PHP is often overlooked in competitive programming but is widely known and trivially deployable. This demonstrates that even PHP — without async, without frameworks — can compete on equal footing when the interface is HTTP. Lowers the barrier for the large PHP developer community.

Strategy: Defend own core, gather nearby energy, cautious expansion.

Behavior:

Maintain a perimeter of bots within 5 tiles of each owned core
Assign excess bots (beyond perimeter needs) to gather energy within 10 tiles of a core
If enemy bots are spotted approaching, consolidate defenders between the enemy and the core
Only sends scouts (lone bots) to explore beyond the safe zone
Very conservative spawning — maintains energy reserve of 6

Value: Tests whether turtling is viable. Should beat rushers but lose to gatherers/swarms in the long game (inferior economy due to limited territory).

Implementation: PHP built-in server (php -S) with a single router script. hash_hmac() for HMAC. JSON via json_decode/json_encode. BFS implemented with SplQueue.

5.5 SwarmBot — TypeScript

Language rationale: TypeScript (Node.js) is the most popular language for web developers entering the platform. This bot demonstrates maintaining complex state across turns — the swarm's formation tracking, rally points, and center-of-mass calculation benefit from TypeScript's type system.

Strategy: Keep units in tight formations, advance as a group toward enemies.

Behavior:

All bots maintain cohesion — no bot moves if it would be >3 tiles from the nearest friendly bot
The swarm moves as a unit toward the nearest enemy presence
BFS-based center-of-mass steering: average position of all owned bots is the swarm center; steer toward enemy center of mass
Energy collection is incidental (pass over it during advance)
New spawns rally to the swarm before advancing

Value: Exploits the focus combat system — a tight group defeats scattered enemies. But slow expansion means inferior economy. Should dominate combat but can be outscored by gatherers on large maps.

Implementation: Express.js or Fastify. State persisted in-process across turns (the HTTP server stays alive between requests). HMAC via Node.js crypto module. Typed interfaces for game state and moves.

5.6 HunterBot — Java

Language rationale: Java is dominant in competitive programming (Battlecode is Java-only). This is the most sophisticated strategy bot, demonstrating that Java's verbosity is offset by mature data structures (PriorityQueue, HashMap) and predictable GC behavior within the timeout window.

Strategy: Target isolated enemy bots for efficient kills.

Behavior:

Identify enemy bots that are ≥4 tiles from their nearest friendly bot (isolated targets)
Send pairs of bots to intercept isolated enemies (2v1 wins cleanly)
If no isolated targets, default to gatherer behavior
Maintain a map of known enemy positions across turns, predict movement based on last-seen direction and speed
Avoid engaging formations of 3+ enemy bots
Opportunistic energy collection when not actively hunting

Value: Sophisticated target selection and prediction. Represents an intermediate-to-advanced-skill bot. Should beat random/gatherer/rusher but struggle against swarm formations.

Implementation: Javalin or com.sun.net.httpserver. javax.crypto.Mac for HMAC. Maintains a HashMap<Position, EnemyTracker> across turns for movement prediction. Hungarian algorithm for optimal bot-to-target assignment.

5.7 Container Templates

Each language has its own container structure. All share the same contract: listen on port 8080, serve POST /turn and GET /health.

Go (GathererBot):

strategy-gatherer/
├── Dockerfile
├── main.go                  # HTTP server, HMAC verification
├── strategy.go              # Gatherer-specific logic
├── game/
│   ├── state.go             # Game state types
│   ├── grid.go              # Grid utilities (BFS, distance, wrapping)
│   └── moves.go             # Move response types
└── go.mod

Python (RandomBot):

strategy-random/
├── Dockerfile
├── main.py                  # HTTP server, HMAC verification, strategy
├── game.py                  # Game state types and grid utilities
└── requirements.txt         # (minimal — stdlib only for random bot)

Rust (RusherBot):

strategy-rusher/
├── Dockerfile
├── Cargo.toml
└── src/
    ├── main.rs              # HTTP server, HMAC verification
    ├── strategy.rs          # Rusher-specific logic
    └── game.rs              # Game state types, grid utilities

PHP (GuardianBot):

strategy-guardian/
├── Dockerfile
├── index.php                # Router + HMAC verification
├── strategy.php             # Guardian-specific logic
├── game.php                 # Game state types, BFS, grid utilities
└── composer.json             # (optional — no external deps needed)

TypeScript (SwarmBot):

strategy-swarm/
├── Dockerfile
├── package.json
├── tsconfig.json
└── src/
    ├── index.ts             # HTTP server, HMAC verification
    ├── strategy.ts          # Swarm-specific logic
    └── game.ts              # Game state types, grid utilities

Java (HunterBot):

strategy-hunter/
├── Dockerfile
├── pom.xml
└── src/main/java/com/acb/hunter/
    ├── App.java             # HTTP server, HMAC verification
    ├── Strategy.java        # Hunter-specific logic
    ├── GameState.java       # Game state deserialization
    └── Grid.java            # Grid utilities, BFS, distance

Shared contract (all languages):

Listen on port 8080
POST /turn — receives game state, runs strategy, returns moves
GET /health — returns 200 (used for registration health check)
HMAC signature verification on incoming requests
HMAC signature on outgoing responses
Request logging (turn number, compute time, move count)

Container specs:

Bot	Build Image	Runtime Image	Memory Limit	CPU Limit
RandomBot	`python:3.13-slim`	`python:3.13-slim`	64MB	0.1 cores
GathererBot	`golang:1.24-alpine`	`alpine:3.21`	128MB	0.25 cores
RusherBot	`rust:1.85-alpine`	`alpine:3.21`	128MB	0.25 cores
GuardianBot	`php:8.4-cli-alpine`	`php:8.4-cli-alpine`	128MB	0.25 cores
SwarmBot	`node:22-alpine`	`node:22-alpine`	128MB	0.25 cores
HunterBot	`eclipse-temurin:21-alpine`	`eclipse-temurin:21-jre-alpine`	256MB	0.5 cores

Java gets a higher resource allocation due to JVM overhead. All others are intentionally constrained — strategy bots should be lightweight.

5.8 Starter Kit & SDK Libraries

To lower the barrier for participants writing their own bots, the platform provides starter kits for each supported language. Each starter kit is a minimal, forkable repository containing:

A working HTTP server with HMAC verification already implemented
Type definitions for the game state and move schemas
Grid utility functions (toroidal distance, BFS, neighbor enumeration)
A stub strategy function that holds all bots in place (participant fills in)
A Dockerfile that builds and runs the bot
A README with quickstart instructions

Starter kit languages (matching strategy bots):

Kit	Repository	Notes
`acb-starter-python`	Template repo	Flask-based, ~100 lines total
`acb-starter-go`	Template repo	Shares `game/` package with GathererBot
`acb-starter-rust`	Template repo	`axum` + `serde`, strongly typed
`acb-starter-php`	Template repo	Zero dependencies, built-in server
`acb-starter-typescript`	Template repo	Fastify, full type definitions
`acb-starter-java`	Template repo	Javalin, Maven-based
`acb-starter-javascript`	Template repo	Node.js built-in http, zero dependencies
`acb-starter-csharp`	Template repo	ASP.NET Core minimal API, zero external dependencies

Participants are not limited to these languages. Any language that can serve HTTP and compute HMAC-SHA256 can compete. The starter kits simply eliminate boilerplate for the most common choices.

6. Tournament System

6.1 Matchmaking

Matches are created continuously by the tournament scheduler, a process that runs on a fixed interval (default: every 10 seconds).

Algorithm:

Select seed bot: the registered bot with the most time since its last match (tiebreak: lowest bot ID)
Determine match size: based on the seed bot's least-played format (2-player, 3-player, 4-player, or 6-player)
Select opponents: from the eligible pool, preferring: a. Closest skill rating to seed (Pareto distribution: 80% within 16 ranks) b. Least recently paired with the seed c. Fewest games played in the last 24 hours (keeps game counts even)
Select map: least recently used map for the chosen player count
Assign player slots: random
Create match job: push to Redis queue with match config + bot endpoints

Eligibility:

Bot must be registered and active (passed health check within last hour)
Bot must not be in a match currently (one match at a time per bot)
Bot must not have been marked crashed in its last 3 consecutive matches (cooldown: 30 minutes)

6.2 Rating System

Algorithm: Glicko-2

Glicko-2 is preferred over TrueSkill for this platform because:

No licensing concerns (TrueSkill is patented by Microsoft)
Includes a volatility parameter (σ) that adapts to inconsistent performance
Well-suited to multi-player games via pairwise decomposition
Established in competitive gaming (chess, Go, online games)

Parameters per bot:

mu (μ): rating estimate (default: 1500)
phi (φ): rating deviation / uncertainty (default: 350)
sigma (σ): rating volatility (default: 0.06)

Display rating: mu - 2*phi (conservative estimate shown on leaderboard)

Update frequency: after every match. Ratings converge quickly — a new bot reaches a stable rating within ~30 matches.

Multi-player adaptation:

A 4-player match produces 6 pairwise results (every pair of players)
Each pairwise result is: win/loss based on relative score, or draw if equal
Glicko-2 update is applied once per match using all pairwise outcomes

6.3 Continuous Tournament

The tournament runs indefinitely with no seasons or resets (initially).

Match throughput target: enough matches that every active bot plays at least 10 matches per day. With N active bots and M match workers:

2-player matches: each match involves 2 bots, takes ~3 minutes (500 turns × 3s max + overhead)
One worker produces ~20 matches/hour
3 workers: ~60 matches/hour, ~1440/day — supports ~288 active bots at 10 games/day

Scaling: add more match worker replicas to increase throughput.

7. Replay System

7.1 Replay Data Format

Replays are JSON files optimized for compact storage while supporting full client-side reconstruction of every game turn.

{
  "version": 1,
  "match_id": "m_7f3a9b2c",
  "date": "2026-03-23T14:30:00Z",
  "players": [
    { "bot_id": "b_4e8c1d2f", "name": "SwarmBot", "owner": "alice" },
    { "bot_id": "b_9a1b3c4d", "name": "HunterBot", "owner": "bob" }
  ],
  "result": {
    "winner": 0,
    "condition": "turn_limit",
    "final_scores": [7, 3],
    "final_energy": [12, 4],
    "final_bots": [18, 6]
  },
  "config": {
    "rows": 60,
    "cols": 60,
    "max_turns": 500,
    "vision_radius2": 49,
    "attack_radius2": 12,
    "spawn_cost": 3,
    "energy_interval": 10
  },
  "map": {
    "walls": [[10,10], [10,11], [10,12]],
    "energy_nodes": [[20,25], [40,35]],
    "cores": [
      { "pos": [5,5], "owner": 0 },
      { "pos": [55,55], "owner": 1 }
    ]
  },
  "turns": [
    {
      "moves": {
        "0": [{"from":[10,15],"dir":"N"},{"from":[12,15],"dir":"E"}],
        "1": [{"from":[50,45],"dir":"S"}]
      },
      "spawns": [[5,5,0]],
      "deaths": [[30,40,1]],
      "captures": [],
      "energy_collected": {"0": [[20,25]]},
      "energy_spawned": [[35,15]],
      "scores": [3, 1]
    }
  ]
}

Size estimate: a 500-turn, 4-player match with ~50 bots total produces a replay of ~200–500 KB uncompressed, ~30–80 KB gzipped.

Optimization: for very long matches, the turns array can use delta encoding — only recording events that changed from the previous turn.

7.2 Storage

Replays use tiered storage: Backblaze B2 is the permanent cold archive (all replays, forever), and Cloudflare R2 is a warm cache for recent replays (free tier, ≤10 GB). Pre-computed JSON index files are deployed to Cloudflare Pages by the index builder. No PersistentVolumes are used for web-facing data.

B2 data layout (cold archive — all data, permanently):

replays/{match_id}.json.gz           # ALL replay files
matches/{match_id}.json              # ALL per-match metadata
thumbnails/{match_id}.png            # ALL match thumbnails
cards/{bot_id}.png                   # ALL bot profile card images

R2 data layout (warm cache — recent subset, ≤10 GB):

replays/{match_id}.json.gz           # recent replay files (promoted from B2)
matches/{match_id}.json              # recent per-match metadata
thumbnails/{match_id}.png            # recent match thumbnails
cards/{bot_id}.png                   # bot profile card images

Pages data layout (static site):

data/leaderboard.json                # current leaderboard snapshot
data/bots/index.json                 # bot directory
data/bots/{bot_id}.json              # per-bot profile (rating history, recent matches)
data/matches/index.json              # paginated match list (last 1000)
data/series/index.json               # series directory
data/seasons/index.json              # seasons directory
data/playlists/{slug}.json           # auto-curated collections
data/evolution/lineage.json          # evolution lineage graph
data/evolution/meta.json             # current meta/Nash snapshot
data/blog/index.json                 # blog post directory
data/blog/posts/{slug}.json          # individual blog posts
maps/index.json                      # map directory
maps/{map_id}.json                   # map definitions

How data flows:

Match worker completes a match → uploads replays/{match_id}.json.gz and matches/{match_id}.json to B2 (cold archive, via S3-compatible API)
Worker writes match result to PostgreSQL (scores, ratings, metadata)
Index builder (every ~15 min) reads new results from PostgreSQL, rebuilds all JSON index files, deploys to Pages via wrangler pages deploy
Index builder promotes recent replays from B2 to R2 (warm cache)
Index builder prunes oldest replays from R2 when approaching 10 GB
Browser loads SPA + indexes from Pages, fetches replays from R2 (warm) with B2 fallback (cold)

Tiered retention:

B2 (cold): All replays retained permanently. No pruning. The canonical archive.
R2 (warm): Recent replays, capped at ≤10 GB. The index builder prunes by age when approaching the cap — oldest replays are removed from R2 (they remain in B2).
Exemptions from R2 pruning: Replays referenced by playlists ("Closest Finishes", "Biggest Upsets", etc.), rivalry pages, series, or season archives are kept warm longer. The pruning job checks the exemption list from PostgreSQL before removing from R2.
PostgreSQL: Match metadata retained indefinitely (rows are small).
Index files are append-with-rotation: index.json holds the last 1000; older pages at index-{page}.json.

Storage costs:

R2 (warm cache): ≤10 GB free tier. At ~50 KB/replay, holds ~200K warm replays (~139 days at 60 matches/hour). Class A writes: ~43K/month (index builder promoting replays). Well within free tier limits.
B2 (cold archive): First 10 GB free, $0.006/GB/month after. At 60 matches/hour, ~2.2 GB/month. Year one: ~26 GB ≈ $0.10/month. Free egress via Cloudflare Bandwidth Alliance.
Pages: No per-file storage costs. 20K file limit per deployment — only SPA + JSON indexes, well within limits.

7.3 Browser Replay Viewer

The replay viewer is a client-side TypeScript application rendered on HTML5 Canvas.

Rendering pipeline:

Fetch replay.json.gz from R2 warm cache (falling back to B2 cold archive); browser handles gzip decompression via Accept-Encoding
Parse and index: build per-turn game state by replaying events from turn 0
Render the current turn to canvas
User controls advance/rewind the turn index

No API invocations — the viewer is a static page (served from Pages) loading a replay file from R2 (warm cache) or B2 (cold archive).

Visual design:

Element	Rendering
Grid	Subtle grid lines on dark background
Walls	Dark gray filled squares
Open tiles	Transparent (background shows through)
Energy nodes	Small yellow diamond; pulse animation when energy is present
Cores	Large player-colored circle with ring; X overlay when razed
Bots	Player-colored filled circles; brief trail showing last move direction
Dead bots	Fading red X for one turn
Fog of war	Dark semi-transparent overlay on tiles outside selected player's vision
Combat	Flash effect on tiles where kills occurred

Controls:

Control	Function
Play / Pause	Toggle automatic playback
Speed slider	1x, 2x, 4x, 8x, 16x (turns per second: 2, 4, 8, 16, 32)
Turn scrubber	Drag to any turn; displays turn number
Perspective dropdown	"All" (omniscient) or per-player fog of war view
Zoom	Scroll to zoom; drag to pan
Score overlay	Per-player score, energy, bot count — updates each turn
Minimap	Small overview of full grid in corner (for large maps)

Shareable URLs: https://aicodebattle.com/replay/{match_id} — the replay viewer is the landing page for any match. No login required to watch.

8. Web Platform

The web platform spans Cloudflare (static site and data files) and the apexalgo-iad Kubernetes cluster (API, compute, databases). Cloudflare Pages serves the SPA globally. Cloudflare R2 stores replays and per-match data. The Go API Deployment on K8s handles registration, job coordination, and scheduling logic. CNPG PostgreSQL stores all relational data. Valkey provides the job queue and caching. No PersistentVolumes are used for web-facing data.

8.1 Cloudflare Pages (Static Site)

The website is a static SPA hosted on Cloudflare Pages. The SPA shell (HTML/JS/CSS/WASM) and all pre-computed JSON data files are deployed as a single Pages project. The index builder on K8s updates the data files every ~90 minutes via wrangler pages deploy.

/                          → Landing page, featured replays, leaderboard summary
/leaderboard               → Full leaderboard (fetches data/leaderboard.json from Pages)
/matches                   → Match history (fetches data/matches/index.json from Pages)
/replay/{match_id}         → Replay viewer (fetches replays/{match_id}.json.gz from R2)
/bot/{bot_id}              → Bot profile (fetches data/bots/{bot_id}.json from Pages)
/evolution                 → Evolution dashboard (fetches data/evolution/*.json from Pages + evolution/live.json from R2)
/register                  → Bot registration form (submits to Go API on K8s)
/docs                      → Protocol spec, starter kit links, getting started

Build: Vite + TypeScript. Code changes are built by an Argo Workflow (triggered by Argo Events on git push), which runs npm run build. The build output is stored as a container artifact; the index builder merges it with the latest data files and deploys to Pages via wrangler pages deploy. No build-time data fetching -- all data loaded at runtime.

Data loading pattern:

// SPA shell + index data from Pages (same origin)
const leaderboard = await fetch('/data/leaderboard.json').then(r => r.json())
// Replays from R2 (cross-origin, CORS enabled on R2 bucket)
const replay = await fetch(`https://r2.aicodebattle.com/replays/${matchId}.json.gz`)
// Dynamic operations from K8s API
const result = await fetch('https://api.aicodebattle.com/api/register', { method: 'POST', body: ... })

Index JSON files are rebuilt and deployed to Pages every ~15 minutes by the index builder (with actual Pages deploys batched every ~90 minutes to stay well within Cloudflare's deploy limits). Visitors see index data that is at most ~90 minutes old. Replays and per-match metadata are uploaded to R2 in real time by match workers and available immediately.

8.2 Go API Service

A single Go HTTP service (acb-api) handles all server-side logic. It runs as a Deployment in the ai-code-battle namespace with a ClusterIP Service. Traefik routes api.aicodebattle.com to it via an IngressRoute (TLS via cert-manager). The API serves only dynamic endpoints -- no static files. It connects to CNPG PostgreSQL for persistent state and Valkey for the job queue.

API endpoints (HTTP routes):

POST /api/register         → register a new bot
POST /api/rotate-key       → rotate a bot's shared secret
GET  /api/status/{bot_id}  → check bot health status
POST /api/jobs/{id}/result  → worker submits match result metadata (authenticated)

Internal scheduling (goroutine tickers, not external crons):

Ticker	Interval	What It Does
Matchmaker	Every 1 min	Queries active bots from PostgreSQL, computes pairings, enqueues jobs in Valkey
Health checker	Every 15 min	Pings each active bot's `/health` endpoint via cluster-internal Service DNS, updates status in PostgreSQL
Stale job reaper	Every 5 min	Marks jobs running >15 min as abandoned, re-enqueues in Valkey

Index building runs as a separate Deployment (see below) that reads directly from PostgreSQL and deploys generated JSON indexes to Cloudflare Pages.

Match worker coordination:

Match workers dequeue jobs from Valkey (BRPOP on a job queue list). The Go API enqueues job IDs; workers fetch full job config from PostgreSQL. Workers submit results back via POST /api/jobs/{id}/result. All communication is cluster-internal (no external API calls needed).

Authentication:

The /api/jobs/{id}/result endpoint is called by match workers within the cluster. Workers authenticate with a shared API key stored in a SealedSecret and mounted as an environment variable. External-facing endpoints (/api/register, /api/rotate-key, /api/status) are public.

8.3 PostgreSQL (CNPG)

The platform uses the existing CNPG PostgreSQL cluster (cnpg-apexalgo) in the cnpg namespace. A dedicated database acb is created within the cluster. The Go API connects via the CNPG-managed Service (cnpg-apexalgo-rw.cnpg.svc.cluster.local). Credentials are stored in a SealedSecret.

Consolidated schema (all tables referenced throughout this plan):

-- §8.3: Core tables

CREATE TABLE bots (
    bot_id        VARCHAR(16) PRIMARY KEY,  -- e.g. 'b_4e8c1d2f'
    name          VARCHAR(32) UNIQUE NOT NULL,
    owner         VARCHAR(128) NOT NULL,
    endpoint_url  TEXT NOT NULL,
    shared_secret VARCHAR(64) NOT NULL,  -- encrypted, see §4.4
    status        VARCHAR(16) NOT NULL DEFAULT 'pending',
    rating_mu     DOUBLE PRECISION NOT NULL DEFAULT 1500.0,
    rating_phi    DOUBLE PRECISION NOT NULL DEFAULT 350.0,
    rating_sigma  DOUBLE PRECISION NOT NULL DEFAULT 0.06,
    evolved       BOOLEAN NOT NULL DEFAULT FALSE,
    island        VARCHAR(16),
    generation    INTEGER,
    parent_ids    JSONB,           -- array of parent bot_ids for lineage tracking
    description   TEXT,
    created_at    TIMESTAMPTZ NOT NULL DEFAULT NOW(),
    last_active   TIMESTAMPTZ
);

CREATE TABLE matches (
    match_id      VARCHAR(32) PRIMARY KEY,
    map_id        VARCHAR(32) NOT NULL,
    status        VARCHAR(16) NOT NULL DEFAULT 'pending',
    winner        INTEGER,
    condition     VARCHAR(32),
    turn_count    INTEGER,
    scores_json   JSONB,
    created_at    TIMESTAMPTZ NOT NULL DEFAULT NOW(),
    completed_at  TIMESTAMPTZ
);

CREATE TABLE match_participants (
    match_id      VARCHAR(32) NOT NULL REFERENCES matches(match_id),
    bot_id        VARCHAR(16) NOT NULL REFERENCES bots(bot_id),
    player_slot   INTEGER NOT NULL,
    score         INTEGER,
    status        VARCHAR(16),
    PRIMARY KEY (match_id, bot_id)
);

CREATE TABLE jobs (
    job_id        VARCHAR(32) PRIMARY KEY,
    match_id      VARCHAR(32) NOT NULL REFERENCES matches(match_id),
    status        VARCHAR(16) NOT NULL DEFAULT 'pending',
    config_json   JSONB NOT NULL,
    claimed_at    TIMESTAMPTZ,
    completed_at  TIMESTAMPTZ
);

CREATE TABLE rating_history (
    bot_id        VARCHAR(16) NOT NULL REFERENCES bots(bot_id),
    match_id      VARCHAR(32) NOT NULL REFERENCES matches(match_id),
    rating        DOUBLE PRECISION NOT NULL,
    recorded_at   TIMESTAMPTZ NOT NULL DEFAULT NOW()
);

CREATE INDEX idx_rating_history_bot ON rating_history(bot_id, recorded_at);

-- §13.5: Prediction system

CREATE TABLE predictions (
    prediction_id    VARCHAR(32) PRIMARY KEY,
    match_id         VARCHAR(32) NOT NULL REFERENCES matches(match_id),
    predictor_id     VARCHAR(36) NOT NULL,  -- localStorage-generated UUID
    predictor_name   VARCHAR(64),           -- optional display name
    predicted_bot_id VARCHAR(16) NOT NULL REFERENCES bots(bot_id),
    correct          BOOLEAN,              -- null until resolved
    created_at       TIMESTAMPTZ NOT NULL DEFAULT NOW()
);

CREATE TABLE predictor_stats (
    predictor_id    VARCHAR(36) PRIMARY KEY,
    predictor_name  VARCHAR(64),
    correct         INTEGER NOT NULL DEFAULT 0,
    incorrect       INTEGER NOT NULL DEFAULT 0,
    streak          INTEGER NOT NULL DEFAULT 0,
    best_streak     INTEGER NOT NULL DEFAULT 0,
    rating          DOUBLE PRECISION NOT NULL DEFAULT 1000.0
);

-- §13.6: Map voting

CREATE TABLE map_votes (
    vote_id     VARCHAR(32) PRIMARY KEY,
    map_id      VARCHAR(32) NOT NULL,
    voter_id    VARCHAR(36) NOT NULL,  -- localStorage UUID
    vote        SMALLINT NOT NULL CHECK (vote IN (-1, 1)),
    created_at  TIMESTAMPTZ NOT NULL DEFAULT NOW(),
    UNIQUE(map_id, voter_id)
);

-- §13.6: Community replay feedback

CREATE TABLE replay_feedback (
    feedback_id   VARCHAR(32) PRIMARY KEY,
    match_id      VARCHAR(32) NOT NULL REFERENCES matches(match_id),
    turn          INTEGER NOT NULL,
    type          VARCHAR(16) NOT NULL CHECK (type IN ('insight', 'mistake', 'idea', 'highlight')),
    body          TEXT NOT NULL,
    author        VARCHAR(128) NOT NULL,  -- free text (no accounts, like registration)
    upvotes       INTEGER NOT NULL DEFAULT 0,
    created_at    TIMESTAMPTZ NOT NULL DEFAULT NOW()
);

CREATE INDEX idx_feedback_match ON replay_feedback(match_id, turn);

-- §14.7: Multi-game series

CREATE TABLE series (
    series_id     VARCHAR(32) PRIMARY KEY,
    bot_a_id      VARCHAR(16) NOT NULL REFERENCES bots(bot_id),
    bot_b_id      VARCHAR(16) NOT NULL REFERENCES bots(bot_id),
    format        SMALLINT NOT NULL CHECK (format IN (3, 5, 7)),
    status        VARCHAR(16) NOT NULL DEFAULT 'pending',
    a_wins        INTEGER NOT NULL DEFAULT 0,
    b_wins        INTEGER NOT NULL DEFAULT 0,
    season_id     VARCHAR(32),
    created_at    TIMESTAMPTZ NOT NULL DEFAULT NOW(),
    completed_at  TIMESTAMPTZ
);

CREATE TABLE series_games (
    series_id     VARCHAR(32) NOT NULL REFERENCES series(series_id),
    game_number   INTEGER NOT NULL,
    match_id      VARCHAR(32) REFERENCES matches(match_id),  -- null until played
    map_id        VARCHAR(32) NOT NULL,
    winner        INTEGER,
    PRIMARY KEY (series_id, game_number)
);

-- §14.9: Seasonal rotations

CREATE TABLE seasons (
    season_id     VARCHAR(32) PRIMARY KEY,
    name          VARCHAR(64) NOT NULL,
    theme         VARCHAR(64) NOT NULL,
    rules_version INTEGER NOT NULL,
    started_at    TIMESTAMPTZ NOT NULL DEFAULT NOW(),
    ended_at      TIMESTAMPTZ,
    champion_id   VARCHAR(16) REFERENCES bots(bot_id),
    status        VARCHAR(16) NOT NULL DEFAULT 'active'
);

PostgreSQL advantages over SQLite (D1):

JSONB columns with indexing for structured data (scores, parent_ids, config)
Foreign key constraints enforced at the database level
TIMESTAMPTZ for proper timezone-aware timestamps
CHECK constraints for enum-like fields
Concurrent writes without locking issues
CNPG provides automatic backups, failover, and point-in-time recovery

8.4 Index Builder (Deployment)

The index builder runs as a Kubernetes Deployment in the ai-code-battle namespace. It is a long-running process with a sleep loop: run the index build, sleep 15 minutes, repeat. Every ~6 cycles (~90 minutes), it deploys the accumulated index files to Cloudflare Pages via wrangler pages deploy. After a configurable lifetime (default: 4 hours), the process exits cleanly and Kubernetes restarts it — preventing memory leaks and stale state.

Process lifecycle:

start → build indexes → sleep 15m → build indexes → sleep 15m → ...
                                  → (every ~6 cycles) deploy to Pages
                                  → (after 4 hours) exit 0
                                  → K8s restarts pod

Each cycle:

Read: Queries PostgreSQL directly (via the CNPG Service) for current match results, bot stats, ratings, series, seasons, predictions, playlists, community feedback, and evolution lineage data.
Generate: Computes all pre-computed JSON index files in a local staging directory:
- data/leaderboard.json — sorted bot rankings with stats
- data/bots/index.json and data/bots/{bot_id}.json — bot directory and profiles
- data/matches/index.json — paginated match list (last 1000)
- data/series/index.json and data/series/{series_id}.json
- data/seasons/index.json and data/seasons/{season_id}.json
- data/playlists/{slug}.json — auto-curated collections
- data/predictions/leaderboard.json and data/predictions/open.json
- data/meta/archetypes.json and data/meta/rivalries.json
- data/evolution/lineage.json and data/evolution/meta.json
- data/blog/index.json and data/blog/posts/{slug}.json (weekly blog generation)
Deploy to Pages: Every ~6 cycles (~90 minutes), merges the generated data files with the SPA shell (HTML/JS/CSS/WASM from the latest site build) and deploys to Cloudflare Pages via wrangler pages deploy. Only the data/ directory changes between deploys; the SPA shell updates only when a new site build is triggered.
Prune on R2: On one cycle per week (checked via day-of-week), runs replay pruning — lists objects in R2's replays/ prefix older than 90 days, queries PostgreSQL for exempt match IDs, deletes non-exempt objects via S3 DeleteObjects API.

Environment: The Deployment Pod has PostgreSQL credentials and a Cloudflare API token (for wrangler pages deploy) stored as SealedSecrets. It also has R2 credentials (S3-compatible access key) for the weekly pruning cycle. No PersistentVolumes needed — all output goes to Cloudflare.

8.5 Bot Registration

Registration flow:

Participant fills out the form on the static site (/register)
Form POSTs to the Go API: POST /api/register
- Bot name (unique, alphanumeric + hyphens, 3-32 chars)
- Endpoint URL (HTTPS required for competitive; HTTP allowed for dev)
- Owner name (free text, shown on leaderboard)
- Description (optional)
Go API generates:
- bot_id: b_ + 8 hex chars (from crypto/rand)
- shared_secret: 256-bit random, hex-encoded (crypto/rand)
Go API performs a health check: http.Get(endpoint_url + "/health")
- Must return 200 within 5 seconds
Go API performs a protocol test: sends mock game state to POST {endpoint_url}/turn with valid HMAC
- Must return valid moves JSON within 3 seconds
Go API inserts bot record into PostgreSQL
Go API returns bot_id and shared_secret to the participant (displayed once -- they must save it)

No user accounts. Registration is bot-level. The owner name is self-reported. The shared secret is the only authentication -- whoever has it can rotate the key or retire the bot. No OAuth, no sessions, no password storage.

Bot status lifecycle:

PENDING -> ACTIVE -> INACTIVE (health check failed)
                  -> RETIRED (by owner via /api/rotate-key with retire flag)

Only ACTIVE bots participate in matchmaking. The health checker ticker pings each active bot every 15 min. Three consecutive failures -> INACTIVE. Bots automatically return to ACTIVE when health checks pass again.

8.6 Leaderboard

The leaderboard is a JSON file on Cloudflare Pages (data/leaderboard.json) rebuilt by the index builder Deployment every ~15 minutes and deployed to Pages every ~90 minutes.

{
  "updated_at": "2026-03-23T14:35:00Z",
  "entries": [
    {
      "rank": 1,
      "bot_id": "b_4e8c1d2f",
      "name": "SwarmBot",
      "owner": "alice",
      "rating": 1820,
      "games": 142,
      "wins": 98,
      "losses": 40,
      "draws": 4,
      "evolved": false,
      "last_match": "2026-03-23T14:30:00Z"
    }
  ]
}

The SPA fetches this file directly from Pages (same origin, no API invocation). Client-side sorting and filtering (by player count tier, time range, human-only vs all). Auto-refresh every 60 seconds. Public -- no login.

8.7 Match History & Profiles

Bot profile (/bot/{bot_id}) -- fetches data/bots/{bot_id}.json from Pages:

Current rating + rating history (array of [timestamp, rating] pairs rendered as a chart client-side)
Recent matches (last 50) with links to replay viewer
Win/loss/draw breakdown
Bot description, owner, registration date
If evolved: lineage, generation, island

Match list (/matches) -- fetches data/matches/index.json from Pages:

Paginated list of recent matches
Each entry: match_id, participants, scores, date, link to replay

Match detail (/replay/{match_id}):

Fetches matches/{match_id}.json from R2 for metadata
Fetches replays/{match_id}.json.gz from R2 for the replay
Embedded replay viewer (auto-plays)
Score breakdown, participants, match duration

9. Deployment & Infrastructure

9.1 Design Principles

Compute runs in the apexalgo-iad Kubernetes cluster (Rackspace Spot) in a dedicated ai-code-battle namespace. The public-facing product is a Cloudflare Pages static site. Replay files use tiered storage: Cloudflare R2 (warm cache, free tier ≤10 GB) and Backblaze B2 (cold archive, permanent). The cluster is existing infrastructure shared with other workloads — it already provides PostgreSQL (CNPG), Valkey, Traefik ingress, cert-manager, ArgoCD, Argo Workflows, Argo Events, Forgejo (git + container registry), SATA (Cinder CSI) storage, and Sealed Secrets.

Key principles:

Static-first architecture — the public product is a static site on Pages. All data visitors see is pre-computed JSON. K8s is the factory that generates data and publishes it to Pages.
Tiered storage — B2 is the permanent archive for all replays. R2 is a warm CDN cache for recent replays, capped at the free tier. The index builder manages the R2 lifecycle.
GitOps via ArgoCD — all K8s manifests are committed to git and synced by ArgoCD. Never apply manifests directly with kubectl.
Argo Workflows for CI — container image builds and static site builds run as Argo Workflows triggered by Argo Events.
Shared infrastructure — PostgreSQL, Valkey, Traefik, and cert-manager are cluster-level services. The ai-code-battle namespace consumes them but does not manage them.
No public API initially — the Go API for social features and third- party registration is deferred. The v1 system is fully static.

9.2 Kubernetes Namespace Layout

All ai-code-battle resources live in the ai-code-battle namespace. Cross-namespace dependencies:

cnpg namespace: CNPG PostgreSQL cluster (cnpg-apexalgo) — the Go API and index builder connect via cnpg-apexalgo-rw.cnpg.svc.cluster.local
valkey namespace: Valkey StatefulSet — the Go API and match workers connect via valkey-master.valkey.svc.cluster.local
traefik namespace: Traefik ingress controller — IngressRoute CRDs in the ai-code-battle namespace reference Services in the same namespace
argocd namespace: ArgoCD — an Application resource points to the manifests directory in the git repo

Cloudflare infrastructure requirements:

Cloudflare Pages project: ai-code-battle (ai-code-battle.pages.dev) — hosts the static SPA and data indexes. Deployed by the index builder via wrangler pages deploy.
Cloudflare R2 bucket: Warm cache for recent replays, thumbnails, bot cards. Free tier (≤10 GB). Managed by the index builder.
DNS (when custom domain is desired): aicodebattle.com CNAME to Pages.

Backblaze B2 infrastructure requirements:

B2 bucket: Cold archive for ALL replays and match data, permanently. Match workers upload directly via S3-compatible API. Free egress via Cloudflare Bandwidth Alliance.

K8s manifests directory structure (flat — per cluster CLAUDE.md norms):

declarative-config/k8s/apexalgo-iad/ai-code-battle/
├── namespace.yml
├── acb-database.yml                    (ext-postgres-operator Postgres + PostgresUser)
├── acb-schema-init.yml                 (ConfigMap + Deployment for schema migration)
├── acb-matchmaker-deployment.yml       (matchmaker: pairings, job enqueue, health, reaper)
├── acb-worker-deployment.yml           (match workers: run matches, upload to B2)
├── acb-index-builder-deployment.yml    (index builder: generate JSON, deploy to Pages, manage R2)
├── acb-evolver-deployment.yml          (LLM evolution pipeline)
├── acb-strategy-random-deployment.yml  (RandomBot — Python)
├── acb-strategy-random-service.yml
├── acb-strategy-gatherer-deployment.yml (GathererBot — Go)
├── acb-strategy-gatherer-service.yml
├── acb-strategy-rusher-deployment.yml  (RusherBot — Rust)
├── acb-strategy-rusher-service.yml
├── acb-strategy-guardian-deployment.yml (GuardianBot — PHP)
├── acb-strategy-guardian-service.yml
├── acb-strategy-swarm-deployment.yml   (SwarmBot — TypeScript)
├── acb-strategy-swarm-service.yml
├── acb-strategy-hunter-deployment.yml  (HunterBot — Java)
├── acb-strategy-hunter-service.yml
├── acb-secret.yml.template             (template: B2/R2/Cloudflare/bot secrets)
└── acb-sealedsecret.yml                (sealed version of above)

Secrets already provisioned in the namespace: acb-app-credentials-acb-app (PostgreSQL), keydb-secret (Valkey), backblaze-secret (B2), cloudflare-pages-secret (wrangler), docker-hub-registry (image pulls), openai-secret (evolver LLM).

9.3 Container Images

All container images are built by Argo Workflows and pushed to the Forgejo container registry (forgejo.ardenone.com/ai-code-battle/<image>).

Image	Base	Purpose	K8s Resource
`acb-matchmaker`	Go binary on Alpine	Matchmaking, health checks, stale job reaping	Deployment (1 replica)
`acb-worker`	Go binary on Alpine	Match execution, B2 upload	Deployment (2-10 replicas)
`acb-evolver`	Go binary on Alpine	Evolution pipeline	Deployment (1 replica)
`acb-index-builder`	Go binary on Alpine (includes `wrangler` CLI)	Reads PostgreSQL, generates JSON indexes, deploys to Pages, manages R2 warm cache (promote from B2, prune)	Deployment (sleep-loop, 15 min cycle, Pages deploy every ~90 min, self-restarts every 4h)
`acb-strategy-random`	Python 3.13 slim	RandomBot	Deployment (1 replica)
`acb-strategy-gatherer`	Go on Alpine	GathererBot	Deployment (1 replica)
`acb-strategy-rusher`	Rust on Alpine	RusherBot	Deployment (1 replica)
`acb-strategy-guardian`	PHP 8.4 CLI Alpine	GuardianBot	Deployment (1 replica)
`acb-strategy-swarm`	Node 22 Alpine	SwarmBot (TypeScript)	Deployment (1 replica)
`acb-strategy-hunter`	Temurin 21 JRE Alpine	HunterBot (Java)	Deployment (1 replica)
`acb-evolved-*`	Varies by language	LLM-generated bots	Deployments (0-50)

9.4 Match Job Coordination

Match workers coordinate via Valkey (job queue) and PostgreSQL (persistent state). The matchmaker Deployment is the coordination point.

Job flow:

Matchmaker Deployment queries active bots from PostgreSQL, computes pairings, inserts match + job rows in PostgreSQL, enqueues job IDs into a Valkey list (acb:jobs:pending)
Match worker pod BRPOPs from the Valkey list (blocking dequeue)
Worker fetches full job config from PostgreSQL (map data, bot endpoints + shared secrets for HMAC signing, match config)
Worker executes the full match (500 turns, HTTP calls to bot Services via cluster DNS, e.g. acb-strategy-rusher.ai-code-battle.svc:8080)
Worker uploads replay file to B2 via S3-compatible API (replays/{match_id}.json.gz)
Worker uploads match metadata to B2 (matches/{match_id}.json)
Worker writes result directly to PostgreSQL (scores, winner, turn count, condition) and updates ratings (Glicko-2)
Index builder (next ~15-min cycle) reads new results from PostgreSQL, rebuilds all index JSON files, deploys to Pages, promotes recent replays from B2 to R2 warm cache

Stale job recovery:

Reaper ticker (in matchmaker) checks PostgreSQL every 5 minutes for jobs running >15 minutes
Assumed abandoned (worker pod crashed or was evicted)
Re-enqueues the job ID in Valkey for retry

9.5 Resilience & Pod Failure

If a match worker pod is evicted or crashes:

The match in progress is lost. The stale job reaper detects it within 5 minutes and re-enqueues it.
Other worker pods continue draining the queue.
The Deployment controller restarts the pod.

If a bot Deployment pod is evicted:

The bot goes offline. Health checker detects failure within 15 minutes, marks it INACTIVE in PostgreSQL.
Matchmaker skips inactive bots.
Kubernetes restarts the pod; health checks pass; bot returns to ACTIVE.
Matches in progress where a bot disappeared: that bot times out on each turn, its units hold position, it effectively loses.

If the matchmaker pod restarts:

Matchmaker and health checker tickers restart with the pod.
No state lost — all state is in PostgreSQL and Valkey.
Workers continue BRPOPing from Valkey (the queue persists).
Brief gap in matchmaking (~1 min) while the pod starts.

If PostgreSQL or Valkey is temporarily unavailable:

Workers block on BRPOP. Matchmaker retries on its next tick.
CNPG handles PostgreSQL failover automatically (3-node cluster).
When service resumes, everything recovers without intervention.

9.6 Networking & Security

External traffic:

ai-code-battle.pages.dev (or custom domain) → Cloudflare Pages (static SPA + data indexes)
R2 public URL → Cloudflare R2 (warm replay cache)
B2 public URL (via Cloudflare CDN) → Backblaze B2 (cold replay archive)
No K8s services are exposed externally in v1. The Go API IngressRoute at api.aicodebattle.com is planned for when social features are added.
TLS: Pages and R2 handle TLS automatically. B2 via Cloudflare gets TLS from the CDN layer.

Cluster-internal traffic:

Matchmaker -> PostgreSQL: cnpg-apexalgo-rw.cnpg.svc.cluster.local:5432
Matchmaker -> Valkey: valkey-master.valkey.svc.cluster.local:6379
Workers -> Valkey: same
Workers -> PostgreSQL: same
Workers -> Bot Services: acb-strategy-*.ai-code-battle.svc:8080
Workers -> External bots: outbound HTTPS to registered URLs
Workers -> B2: HTTPS (S3-compatible API, for replay upload)
Index Builder -> PostgreSQL: same
Index Builder -> Cloudflare Pages: HTTPS (wrangler CLI)
Index Builder -> R2: HTTPS (S3-compatible API, for warm cache management)
Index Builder -> B2: HTTPS (S3-compatible API, for promoting replays to R2)
Evolver -> PostgreSQL: same
All cluster-internal traffic is plaintext (trusted network)

Security boundaries:

The game engine (workers) never executes bot code — HTTP only
All bot responses are schema-validated before processing
HMAC authentication prevents request/response forgery
PostgreSQL credentials in SealedSecrets (encrypted in git, decrypted in-cluster)
Valkey access is cluster-internal only (no external exposure)
Pages, R2, and B2 serve public-read data only — no secrets stored there
B2 write credentials (SealedSecret) are scoped to the acb bucket
Cloudflare API token (SealedSecret) is scoped to Pages deploy + R2 only
NetworkPolicy can restrict egress from bot pods to prevent data exfiltration (future hardening)

9.7 ArgoCD & GitOps

All K8s manifests for the ai-code-battle namespace are stored in the declarative-config repo (ardenone-cluster):

declarative-config/k8s/apexalgo-iad/ai-code-battle/

An ArgoCD Application watches this directory and syncs changes automatically:

apiVersion: argoproj.io/v1alpha1
kind: Application
metadata:
  name: ai-code-battle
  namespace: argocd
spec:
  project: default
  source:
    repoURL: https://forgejo.ardenone.com/infra/ardenone-cluster.git
    path: declarative-config/k8s/apexalgo-iad/ai-code-battle
    targetRevision: main
  destination:
    server: https://kubernetes.default.svc
    namespace: ai-code-battle
  syncPolicy:
    automated:
      prune: true
      selfHeal: true

Workflow for infrastructure changes:

Edit manifests in declarative-config/k8s/ (ardenone-cluster repo)
git push to origin
ArgoCD detects the change and syncs within ~3 minutes
Verify via: kubectl --server=http://kubectl-apexalgo-iad:8001 get pods -n ai-code-battle

9.8 CI/CD Pipelines (Argo Workflows + Events)

Argo Events sensor: A GitHub webhook sensor listens for push events on the ai-code-battle repo. On push to main, it triggers the appropriate Argo Workflow(s).

Image build workflow (build-images):

Triggered by git push to main (when Go/container source changes)
Clones the repo
Builds container images (Kaniko — no Docker daemon needed)
Pushes to Forgejo registry: forgejo.ardenone.com/ai-code-battle/<image>:<sha>
Tags as latest
ArgoCD detects the image tag change and rolls out new pods

Site build workflow (build-site):

Triggered by git push to main (when web/ directory changes)
Clones the repo
Runs npm ci && npm run build in the web/ directory
Stores the build output as a container image artifact (acb-site-build:<sha>) pushed to Forgejo registry
The index builder picks up the latest site build artifact, merges it with generated data files, and deploys to Cloudflare Pages on its next cycle

Evolved bot deploy workflow:

Triggered by the evolver when a candidate is promoted
Receives bot source code and language as parameters
Builds a container image (Kaniko)
Pushes to Forgejo registry
Creates a Deployment + Service manifest
Commits the manifest to the declarative-config repo
ArgoCD syncs the new bot into the cluster

9.9 Monitoring

Signal	Method	Alert
Pod health	Kubernetes liveness/readiness probes	Auto-restart
Pages up	Cloudflare Pages analytics + synthetic checks	Cloudflare handles failover
PostgreSQL health	CNPG operator monitoring	Auto-failover
Valkey health	Kubernetes probes	Auto-restart
Match throughput	PostgreSQL query: completions per hour	<10/hour for >1 hour
Worker queue depth	Valkey LLEN on `acb:jobs:pending`	>50 pending for >30 min
Bot health failures	Matchmaker health checker ticker	>50% failing
Stale jobs	Matchmaker reaper ticker count	>10 stale in a cycle
R2 usage	Index builder tracks warm cache size	>8 GB (approaching 10 GB free tier cap)
B2 usage	B2 dashboard / API metrics	Informational only (no cap)

Alerts via matchmaker -> webhook to Discord/Slack. Cluster-level monitoring (Prometheus, if deployed) can scrape the matchmaker's /metrics endpoint.

10. LLM-Driven Bot Evolution

The platform includes an autonomous evolution pipeline that uses LLMs to continuously generate, evaluate, and promote new bot strategies. Evolved bots compete on the same ladder as human-written bots — visitors see an ever-changing meta where strategies emerge, dominate, and get countered without human intervention.

10.1 Architecture Overview

The evolution system combines two proven approaches:

FunSearch/AlphaEvolve island model — maintains diverse, independent populations of bot code that cross-pollinate. Prevents premature convergence to a single dominant strategy.
LLM-PSRO (Policy Space Response Oracle) — uses Nash equilibrium as the promotion gate. A new bot must beat the optimal mixed strategy over the current population, not just one specific opponent. This provides mathematically grounded regression prevention.

┌──────────────────────────────────────────────────────────┐
│                     Programs Database                     │
│  ┌───────────┐ ┌───────────┐ ┌───────────┐ ┌──────────┐ │
│  │  Island 1  │ │  Island 2  │ │  Island 3  │ │ Island 4 │ │
│  │  (Python)  │ │  (Go)      │ │  (Rust)    │ │ (mixed)  │ │
│  │  pop: 20   │ │  pop: 20   │ │  pop: 20   │ │ pop: 20  │ │
│  └───────────┘ └───────────┘ └───────────┘ └──────────┘ │
└──────────────────────────┬───────────────────────────────┘
                           │
             sample 2-3 parents + match replays
                           │
                ┌──────────▼───────────┐
                │    Prompt Builder     │
                │  • Parent source code │
                │  • Recent loss replay │
                │  • Win/loss analysis  │
                │  • Current meta desc  │
                │  • "Beat this mix"    │
                └──────────┬───────────┘
                           │
                ┌──────────▼───────────┐
                │    LLM Ensemble       │
                │  • Fast model (×8)    │
                │    exploration/breadth │
                │  • Strong model (×2)  │
                │    exploitation/depth  │
                └──────────┬───────────┘
                           │ generates candidate bot code
                ┌──────────▼───────────┐
                │    Validation Gate    │
                │  1. Syntax check      │
                │  2. Compile/lint      │
                │  3. Schema test       │
                │  4. Sandbox smoke run │
                └──────────┬───────────┘
                           │ passes validation
                ┌──────────▼───────────┐
                │    Evaluation Arena   │
                │  • 10 matches vs      │
                │    population sample  │
                │  • Compute win rate   │
                │  • Build payoff row   │
                └──────────┬───────────┘
                           │
                ┌──────────▼───────────┐
                │    Promotion Gate     │
                │  • Compute Nash eq.   │
                │    over population    │
                │  • Candidate must     │
                │    beat Nash mixture  │
                │  • Or: fill empty     │
                │    MAP-Elites niche   │
                └──────────┬───────────┘
                           │ promoted
                ┌──────────▼───────────┐
                │    Deploy & Register  │
                │  • Build container    │
                │  • Push to registry   │
                │  • Register on ladder │
                │  • Enter island DB    │
                └──────────────────────┘

10.2 Programs Database (Island Model)

The programs database stores all evolved bot code, organized into islands that evolve independently to maintain strategic diversity.

Island structure:

4 islands, one per primary language (Python, Go, Rust, mixed)
Each island holds up to 20 programs ranked by fitness
Programs are clustered by behavior signature — a vector of outcomes across a fixed set of benchmark matches (e.g., win/loss/score against each of the 6 built-in strategy bots)
Sampling favors high-scoring clusters; within a cluster, favors shorter/simpler code (Occam pressure prevents bloat)

Cross-pollination:

Every 50 generations, the top program from each island is copied to a random other island (translated to that island's language by the LLM if needed)
This spreads successful strategies across languages without homogenizing the populations

Behavior dimensions for MAP-Elites diversity:

Dimension	Low	High
Aggression	Never enters enemy territory	Rushes enemy core immediately
Economy	Ignores energy entirely	Maximizes energy per turn
Exploration	Stays near core	Covers >80% of visible map
Formation	Units always scattered	Units always in tight groups

Each dimension is binned into 3 levels, creating a 3⁴ = 81-cell behavior grid. The database tries to fill every cell with the highest-scoring bot for that behavioral profile. This ensures the evolved population contains turtles, rushers, economists, swarmers, and everything in between — not just one dominant archetype.

10.3 Prompt Construction

The LLM prompt is the critical interface between match performance data and code generation. Each prompt is constructed from:

Parent code (2–3 programs):

Sampled from the island's high-scoring clusters
Included as full source code with inline comments noting their rating and behavioral profile
The LLM sees concrete working examples, not abstract descriptions

Match analysis (from recent losses):

The replay of the parent's worst recent loss is summarized:
- Turn-by-turn narrative of critical moments (when the bot lost a formation, missed energy, walked into a trap)
- Final score breakdown
- Opponent's apparent strategy (inferred from replay)
This gives the LLM specific failure modes to address

Meta description:

Current Nash equilibrium mixture over the population (e.g., "the optimal counter-strategy is 40% swarm, 30% hunter, 30% gatherer")
The candidate should beat this mixture, not just one opponent
Weaknesses in the current meta are highlighted (e.g., "no bot currently exploits the east-side energy clusters on 4-player maps")

Constraints:

Target language for this island
Must implement the HTTP bot interface (POST /turn, GET /health)
Must include HMAC verification
Maximum source code size (10 KB — prevents bloat)
Must respond within 3-second timeout with reasonable compute

Prompt template (simplified):

You are evolving a competitive bot for AI Code Battle, a grid-based
strategy game. Your bot must be an HTTP server that receives game state
and returns moves.

## Game Rules
{game_rules_summary}

## HTTP Protocol
{protocol_spec}

## Parent Bots (these work — improve on them)

### Parent A — Rating: 1650, Style: aggressive-gatherer
```{language}
{parent_a_source}

Parent B — Rating: 1580, Style: defensive-swarm

{parent_b_source}

Parent A's Worst Loss (Replay Summary)

{replay_analysis}

Current Meta

The Nash equilibrium mixture is: {nash_mixture_description}

Known weaknesses in current population: {meta_weaknesses}

Your Task

Write a new bot in {language} that:

Addresses Parent A's failure mode shown in the replay
Incorporates Parent B's strongest tactical element
Can beat the Nash mixture described above
Fits in a single file under 10 KB

Return the complete source code.


### 10.4 LLM Ensemble

The evolution system uses two model tiers, inspired by AlphaEvolve:

**Exploration tier (fast model, 80% of generations):**
- Cheaper, faster model (e.g., Claude Haiku, GPT-4o-mini, Gemini Flash)
- Generates 8 candidates per cycle
- High temperature (0.9–1.0) for diversity
- Purpose: broad search across strategy space; most candidates will fail,
  but occasional novel approaches emerge

**Exploitation tier (strong model, 20% of generations):**
- More capable model (e.g., Claude Sonnet/Opus, GPT-4o, Gemini Pro)
- Generates 2 candidates per cycle
- Lower temperature (0.3–0.5) for refinement
- Purpose: take the best current strategies and make them better; refine
  tactical details, optimize pathfinding, improve edge-case handling

**Total throughput:** 10 candidates per evolution cycle. With a cycle time
of ~15 minutes (generation + validation + 10 evaluation matches), the system
produces ~96 candidates/day, of which ~5–15% pass the promotion gate.

### 10.5 Validation Pipeline

Every LLM-generated candidate passes through a multi-stage validation
before it touches the evaluation arena:

**Stage 1: Syntax & Compilation**
- Language-specific: `python -m py_compile`, `go build`, `cargo check`,
  `php -l`, `tsc --noEmit`, `javac`
- Reject: syntax errors, missing imports, type errors
- ~40% of candidates fail here (expected — LLMs produce broken code often)

**Stage 2: Schema Compliance**
- Start the bot container
- Send a mock turn-0 game state to `POST /turn`
- Verify response parses as valid moves JSON
- Verify `GET /health` returns 200
- Verify HMAC signature is present and valid
- Reject: bots that can't speak the protocol
- ~20% of remaining candidates fail here

**Stage 3: Sandbox Smoke Test**
- Run a 50-turn match against RandomBot inside nsjail
- Verify the bot doesn't crash, timeout on every turn, or produce
  identical moves every turn (degenerate)
- Verify the bot scores ≥ 0 (doesn't actively self-destruct)
- Reject: bots that crash, hang, or do nothing
- ~10% of remaining candidates fail here

**Net yield:** ~30–40% of generated candidates survive to the evaluation
arena. At 10 candidates/cycle, that's 3–4 evaluated candidates per cycle.

**Sandboxing (nsjail):**
- All LLM-generated code executes inside nsjail containers
- No network access (game state is piped via the engine, not fetched)
- No filesystem access beyond the bot's own directory
- CPU time limit: 5 seconds per turn (generous; 3-second HTTP timeout is
  enforced by the engine separately)
- Memory limit: 512 MB
- Process limit: 10 (prevents fork bombs)

### 10.6 Evaluation Arena

Candidates that pass validation enter a mini-tournament:

**Evaluation protocol:**
1. Play 10 matches against opponents sampled from the current population:
   - 2 matches vs each of the 3 closest-rated bots in the candidate's island
   - 2 matches vs a random bot from a different island
   - 2 matches vs the current island champion
2. Record results → compute win rate and per-opponent scores
3. Build the candidate's **payoff row** in the population's payoff matrix

**Match configuration:**
- 2-player matches only (faster evaluation; multi-player tested post-promotion)
- Standard maps, standard timeout
- Evaluation matches are **not** counted toward ladder ratings (they use a
  separate evaluation queue)

### 10.7 Promotion Gate (Nash Equilibrium / PSRO)

The promotion gate determines whether a candidate enters the population and
gets deployed to the ladder.

**Primary gate: Nash equilibrium (LLM-PSRO)**

1. Compute the Nash equilibrium mixture σ* over the current island population
   using the existing payoff matrix
2. Compute the candidate's expected payoff against σ* (using the payoff row
   from the evaluation arena)
3. **Promote if** the candidate's expected payoff against σ* is positive
   (i.e., the candidate beats the current optimal mixed strategy)
4. If promoted, add the candidate to the island population, recompute Nash

This ensures the population's game-theoretic strength monotonically increases.
A bot that just exploits one opponent's weakness but loses to the overall mix
is rejected.

**Secondary gate: MAP-Elites niche filling**

Even if a candidate doesn't beat the Nash mixture, it may fill an **empty
cell** in the behavior grid (section 10.2). If the candidate's behavior
signature maps to an unoccupied cell, it is promoted anyway. This maintains
strategic diversity even when the Nash gate is tight.

**Replacement policy:**
- If the candidate's behavior cell already has an occupant, the candidate
  replaces it only if the candidate's fitness is higher
- Island population size is capped at 20; if full and no cell is improved,
  the candidate is discarded
- The worst-performing program in an over-populated cluster is evicted first

### 10.8 Deployment Pipeline

Promoted bots are automatically containerized and registered on the ladder:

**Build:**
1. Write the bot's source code to a temporary directory
2. Copy the language-appropriate Dockerfile from the starter kit template
3. Build the container image: `acb-evolved-{island}-{generation}-{hash}`
4. Push to container registry

**Register:**
1. Generate a new `bot_id` and `shared_secret`
2. Deploy the container to the always-on strategy bot instance pool
3. Register the bot via the platform API with metadata:
   - `owner`: "evolution-system" (system account)
   - `name`: auto-generated (e.g., `evo-py-g42-7f3a`)
   - `description`: auto-generated from the LLM's strategy summary
   - `lineage`: parent bot IDs + generation number
   - `island`: which island produced it
4. Health check → mark ACTIVE → enters matchmaking

**Lifecycle management:**
- Evolved bots are tagged with `evolved: true` in the database
- The evolution system tracks the **lineage** of every bot (parent IDs,
  generation number, island of origin)
- Evolved bots that drop below rating 800 (bottom 10% of ladder) for 7
  consecutive days are **retired** automatically to prevent population bloat
- Maximum active evolved bots: 50 (configurable). When the cap is reached,
  the lowest-rated evolved bot is retired before a new one is promoted.
- Retired evolved bots remain in the programs database for future sampling
  (their code may still contain useful tactics) but are removed from the
  ladder and their containers are stopped

### 10.9 Evolution Cycle Timing

| Phase | Duration | Notes |
|-------|----------|-------|
| Parent sampling + prompt construction | ~10 seconds | CPU-bound, fast |
| LLM generation (10 candidates) | ~30–60 seconds | Parallel across ensemble |
| Validation (syntax, schema, smoke) | ~2 minutes | Parallel per candidate |
| Evaluation arena (10 matches) | ~10 minutes | Sequential matches, 3s/turn × 500 turns worst case; but most against weak bots end faster |
| Nash computation + promotion | ~5 seconds | Small matrix, fast |
| Container build + deploy | ~2 minutes | Docker build + push |
| **Total cycle time** | **~15 minutes** | |

**Daily output:** ~96 candidates generated, ~10-15 promoted, ~5-10 survive
on the ladder after the 7-day retirement window.

**Throughput is configurable** and depends on match worker capacity. The
ratio of ladder matches to evolution evaluation matches is tunable (default:
70/30 -- 70% of match worker capacity goes to ladder matches, 30% to
evolution evaluation matches). When worker pods are scaled down, the ratio
can be adjusted to prioritize ladder matches over evolution. When excess
capacity is available, evolution throughput increases automatically.

**Container lifecycle:** the evolver Deployment runs as a long-lived
container that intentionally exits after a configurable time period
(default: 4 hours), causing Kubernetes to redeploy the pod. This prevents
memory leaks and stale state accumulation across hundreds of evolution
cycles.

### 10.10 Test Harnesses

Three test harness suites validate correctness across the game engine, bot
protocol, and evolution pipeline. These run as part of CI and as part of
the evolution validation pipeline.

**Game engine test suite:**
- Unit tests for combat resolution (focus fire algorithm), fog of war
  computation, movement and collision, scoring, and endgame condition
  detection
- Property-based tests for determinism: given the same input state and
  moves, the engine must produce the same output state. Random seeds
  generate thousands of input combinations; any nondeterminism is a
  failing test.
- Edge case tests: toroidal wrapping, simultaneous multi-player death,
  contested energy, core capture during spawn phase

**Bot protocol test suite:**
- Schema validation: verify that game state JSON conforms to the
  documented schema (section 4.2) and that move responses are validated
  correctly (section 4.3)
- HMAC verification: test correct signature generation and verification,
  timestamp replay rejection, and constant-time comparison
- Timeout handling: verify that the engine correctly handles bots that
  respond after the 3-second deadline, return non-200 status codes,
  return invalid JSON, or refuse connections
- Malformed response handling: verify graceful degradation for partial
  JSON, missing fields, extra fields, and oversized payloads

**Evolution validation test suite:**
- Syntax checking per language: verify that the validation pipeline
  correctly accepts valid code and rejects invalid code for each
  supported language (Python, Go, Rust, PHP, TypeScript, Java, JavaScript, C#)
- Schema compliance: verify that generated bots correctly implement
  `POST /turn` and `GET /health` with valid HMAC signatures
- Sandbox smoke test: verify that nsjail isolation works correctly
  (no network access, filesystem isolation, resource limits enforced)
- End-to-end: generate a known-good bot from a template, run it through
  the full validation pipeline, and verify it passes all stages

### 10.11 Evolution Dashboard

The web platform includes a dedicated evolution section visible to all visitors:

**Lineage viewer:**
- Interactive tree/graph showing the ancestry of every evolved bot
- Click a node to see the bot's source code, rating history, and match record
- Color-coded by island/language
- Animated timeline showing which bots were active at which point

**Meta tracker:**
- Current Nash equilibrium mixture visualization (pie chart of strategy archetypes)
- How the meta has shifted over time (stacked area chart)
- Which behavioral niches are filled vs empty in the MAP-Elites grid

**Generation log:**
- Stream of recent evolution attempts: generated, validated, evaluated, promoted/rejected
- For each attempt: the prompt summary, the LLM's output, validation results,
  evaluation match results, and promotion decision with reasoning

**Statistics:**
- Total generations run, candidates generated, promotion rate
- Average rating of evolved bots vs human-written bots over time
- Island diversity metrics (how different are the islands from each other)

### 10.12 Separation from Human Ladder

Evolved bots compete on the **same ladder** as human-written bots — there is
no separate tier. This is a deliberate design choice:

**Why mix them:**
- The entire point is to see if LLM-evolved strategies can compete with or
  surpass human-written ones
- Humans can study evolved bot replays and learn new tactics, then write
  better bots that push the meta further — a human-AI co-evolution dynamic
- Separate ladders would remove the competitive pressure that drives evolution

**Identification:**
- Evolved bots are clearly tagged on the leaderboard (`[EVO]` prefix or badge)
- Their lineage and source code are publicly viewable (transparency)
- Human participants can opt to filter the leaderboard to show human-only rankings
- Match history shows whether opponents were evolved or human-written

**Fair play:**
- Evolved bots follow the same rules: same timeout, same schema, same HMAC
- No special treatment in matchmaking — rated and matched identically
- The evolution system is rate-limited (max 50 active evolved bots) to prevent
  flooding the ladder

---

## 11. Artifact Inventory

Every deliverable that gets built, deployed, or published. Grouped by
where it runs.

### 11.1 Monorepo Structure

All platform code lives in a single repository (`ai-code-battle`):

ai-code-battle/ ├── engine/ # Go library — game simulation core │ ├── grid.go # Toroidal grid, tile types, wrapping │ ├── bot.go # Bot state, movement, collision │ ├── combat.go # Focus-fire algorithm │ ├── energy.go # Energy nodes, collection, spawning │ ├── fog.go # Fog of war computation │ ├── capture.go # Core capture/razing │ ├── scoring.go # Score tracking, win conditions │ ├── match.go # Turn loop, phase orchestration │ ├── replay.go # Replay JSON serialization │ ├── winprob.go # Monte Carlo win probability rollout │ └── engine_test.go # Property-based + unit tests │ ├── cmd/ │ ├── acb-local/ # CLI: run a match locally (stdin/stdout bots) │ ├── acb-mapgen/ # CLI: generate symmetric maps │ ├── acb-worker/ # Container: match execution worker │ ├── acb-evolver/ # Container: LLM evolution pipeline │ ├── acb-index-builder/ # Container: PostgreSQL -> JSON -> Pages + R2 pruning │ # replay pruning is handled by acb-index-builder (weekly cycle) │ ├── cmd/acb-api/ # Go API service (replaces Cloudflare Worker) │ ├── main.go # HTTP server + ticker scheduler │ ├── routes/ │ │ ├── register.go # POST /api/register, /api/rotate-key │ │ ├── jobs.go # POST /api/jobs/{id}/result │ │ ├── predict.go # POST /api/predict │ │ ├── feedback.go # POST /api/feedback │ │ └── status.go # GET /api/status/{bot_id} │ ├── tickers/ │ │ ├── matchmaker.go # Every 1 min: create match jobs, enqueue in Valkey │ │ ├── health.go # Every 15 min: ping bot endpoints │ │ └── reaper.go # Every 5 min: reclaim stale jobs │ ├── lib/ │ │ ├── hmac.go # HMAC-SHA256 signing/verification │ │ ├── glicko2.go # Glicko-2 rating computation │ │ └── schema.go # Request/response JSON schema validators │ ├── db/ │ │ ├── postgres.go # PostgreSQL connection + queries │ │ └── valkey.go # Valkey (Redis) connection + queue ops │ └── Dockerfile │ ├── migrations/ # PostgreSQL schema migrations │ └── 0001_initial.sql │ ├── web/ # Static site (TypeScript + Vite) │ ├── src/ │ │ ├── app.ts # SPA router, data fetching, layout │ │ ├── pages/ │ │ │ ├── home.ts # Homepage: hero replay, leaderboard, playlists │ │ │ ├── watch.ts # /watch: replay browser, playlists │ │ │ ├── replay.ts # /watch/replay/{id}: full viewer │ │ │ ├── series.ts # /watch/series/{id}: series page │ │ │ ├── compete.ts # /compete: sandbox, registration, docs │ │ │ ├── sandbox.ts # /compete/sandbox: WASM sandbox │ │ │ ├── leaderboard.ts # /leaderboard │ │ │ ├── bot-profile.ts # /bot/{id}: public bot profile │ │ │ ├── evolution.ts # /evolution: live observatory │ │ │ ├── blog.ts # /blog: meta reports + chronicles │ │ │ ├── season.ts # /season/{id}: season archive │ │ │ ├── predictions.ts # /watch/predictions │ │ │ └── embed.ts # /embed/{id}: lightweight embed player │ │ ├── components/ │ │ │ ├── replay-canvas.ts # Canvas renderer: bots, grid, animations │ │ │ ├── territory.ts # Voronoi + influence overlay renderers │ │ │ ├── particles.ts # Particle pool + death/energy animations │ │ │ ├── follow-camera.ts # Bounding box tracking + lerp viewport │ │ │ ├── pip.ts # Picture-in-picture manager │ │ │ ├── director.ts # Adaptive auto-speed controller │ │ │ ├── win-prob.ts # Win probability sparkline graph │ │ │ ├── event-timeline.ts# Event icon ribbon │ │ │ ├── clip-export.ts # GIF/MP4 export via MediaRecorder │ │ │ ├── annotation.ts # Spatial + text replay annotations │ │ │ ├── leaderboard-table.ts │ │ │ ├── bot-card.ts # Bot profile card renderer (Canvas PNG) │ │ │ ├── match-card.ts # Match summary card │ │ │ ├── playlist-row.ts # Horizontal scrollable playlist │ │ │ ├── prediction-widget.ts │ │ │ ├── observatory-feed.ts # Live evolution status │ │ │ ├── blog-post.ts # Markdown renderer for blog content │ │ │ └── skeleton.ts # Per-page skeleton screens │ │ ├── lib/ │ │ │ ├── data.ts # Data fetching from Pages (same origin) + R2 (cross-origin), caching │ │ │ ├── preload.ts # Hover preload + route cache │ │ │ ├── disclosure.ts # Progressive feature revelation (XP) │ │ │ ├── accessibility.ts # Color palettes, keyboard shortcuts │ │ │ ├── ambient.ts # Favicon badges, tab titles, haptic │ │ │ └── season-theme.ts # Background hue shift per season │ │ └── styles/ │ │ ├── base.css # Dark theme, typography, reset │ │ ├── components.css # Component styles │ │ └── mobile.css # Responsive breakpoints, bottom tab bar │ ├── public/ │ │ ├── docs/ # Static documentation pages │ │ └── img/ # Logos, icons, UI assets │ ├── vite.config.ts │ └── tsconfig.json │ ├── wasm/ # WASM builds for the browser sandbox │ ├── engine/ # Go game engine → WASM │ │ ├── main_wasm.go # WASM exports: runMatch, loadState, step │ │ └── build.sh # GOOS=js GOARCH=wasm go build │ └── bots/ # Built-in bot WASM builds │ ├── gatherer/ # Go → WASM │ ├── rusher/ # Rust → WASM (wasm32-unknown-unknown) │ ├── swarm/ # TypeScript → WASM (AssemblyScript) │ ├── random/ # Go → WASM (lightweight reimpl) │ ├── guardian/ # Go → WASM (reimpl from PHP) │ └── hunter/ # Go → WASM (reimpl from Java) │ ├── bots/ # Production bot HTTP servers (6 languages) │ ├── random/ # Python — Flask, ~50 lines strategy │ │ ├── Dockerfile │ │ ├── main.py │ │ ├── game.py │ │ └── requirements.txt │ ├── gatherer/ # Go — net/http, BFS pathfinding │ │ ├── Dockerfile │ │ ├── main.go │ │ ├── strategy.go │ │ └── game/ │ ├── rusher/ # Rust — axum, BFS to enemy core │ │ ├── Dockerfile │ │ ├── Cargo.toml │ │ └── src/ │ ├── guardian/ # PHP — built-in server, perimeter defense │ │ ├── Dockerfile │ │ ├── index.php │ │ ├── strategy.php │ │ └── game.php │ ├── swarm/ # TypeScript — Fastify, formation advance │ │ ├── Dockerfile │ │ ├── package.json │ │ └── src/ │ └── hunter/ # Java — Javalin, target isolation │ ├── Dockerfile │ ├── pom.xml │ └── src/ │ ├── starters/ # Forkable starter kit template repos │ ├── python/ │ ├── go/ │ ├── rust/ │ ├── php/ │ ├── typescript/ │ ├── javascript/ │ ├── java/ │ └── csharp/ │ ├── docs/ # Project documentation │ ├── plan/ │ │ └── plan.md # This document │ ├── research/ │ │ ├── ants-ai-challenge.md │ │ └── llm-bot-evolution.md │ └── notes/ │ └── requirements.md │ ├── CLAUDE.md └── README.md


### 11.2 Deployable Artifacts

**Container images (Forgejo registry: `forgejo.ardenone.com/ai-code-battle/`):**

| Image | Source | Base | Purpose | K8s Resource |
|-------|--------|------|---------|--------------|
| `acb-api` | `cmd/acb-api/` | Go on Alpine | API + scheduling | Deployment (1 replica) |
| `acb-worker` | `cmd/acb-worker/` | Go on Alpine | Match execution | Deployment (2-10 replicas) |
| `acb-evolver` | `cmd/acb-evolver/` | Go on Alpine | LLM evolution pipeline | Deployment (1 replica, self-restarts every 4h) |
| `acb-index-builder` | `cmd/acb-index-builder/` | Go on Alpine (includes wrangler) | PostgreSQL -> JSON -> Cloudflare Pages + weekly R2 replay pruning | Deployment (sleep-loop) |
| `acb-strategy-random` | `bots/random/` | Python 3.13 slim | RandomBot | Deployment (1 replica) |
| `acb-strategy-gatherer` | `bots/gatherer/` | Go on Alpine | GathererBot | Deployment (1 replica) |
| `acb-strategy-rusher` | `bots/rusher/` | Rust on Alpine | RusherBot | Deployment (1 replica) |
| `acb-strategy-guardian` | `bots/guardian/` | PHP 8.4 CLI Alpine | GuardianBot | Deployment (1 replica) |
| `acb-strategy-swarm` | `bots/swarm/` | Node 22 Alpine | SwarmBot | Deployment (1 replica) |
| `acb-strategy-hunter` | `bots/hunter/` | Temurin 21 JRE Alpine | HunterBot | Deployment (1 replica) |
| `acb-evolved-*` | Generated by evolver | Varies | LLM-evolved bots | Deployments (0-50) |

**K8s manifests (ArgoCD-managed):**

| Resource | Source | Namespace |
|----------|--------|-----------|
| All Deployments, Services, Deployments, PVCs, IngressRoutes, SealedSecrets | `declarative-config/k8s/apexalgo-iad/ai-code-battle/` | `ai-code-battle` |
| ArgoCD Application | `declarative-config/k8s/apexalgo-iad/ai-code-battle/argocd-application.yaml` | `argocd` |
| PostgreSQL database `acb` | Created in existing CNPG cluster `cnpg-apexalgo` | `cnpg` |

**WASM artifacts (built at compile time, deployed to Cloudflare Pages):**

| Artifact | Source | Target | Size |
|----------|--------|--------|------|
| `engine.wasm` | `wasm/engine/` | `GOOS=js GOARCH=wasm` | ~15 MB |
| `gatherer.wasm` | `wasm/bots/gatherer/` | Go → WASM | ~12 MB |
| `rusher.wasm` | `wasm/bots/rusher/` | Rust → `wasm32-unknown-unknown` | ~3 MB |
| `swarm.wasm` | `wasm/bots/swarm/` | AssemblyScript → WASM | ~5 MB |
| `random.wasm` | `wasm/bots/random/` | Go → WASM | ~10 MB |
| `guardian.wasm` | `wasm/bots/guardian/` | Go → WASM (reimpl) | ~12 MB |
| `hunter.wasm` | `wasm/bots/hunter/` | Go → WASM (reimpl) | ~12 MB |

**Published template repos (one per language):**

| Repo | Language | Contents |
|------|----------|----------|
| `acb-starter-python` | Python | Flask server, HMAC, game types, stub strategy, Dockerfile |
| `acb-starter-go` | Go | net/http, shared game/ package, Dockerfile |
| `acb-starter-rust` | Rust | axum + serde, HMAC crate, Dockerfile |
| `acb-starter-php` | PHP | Built-in server, hash_hmac, Dockerfile |
| `acb-starter-typescript` | TypeScript | Fastify, typed interfaces, Dockerfile |
| `acb-starter-javascript` | JavaScript | Node.js built-in http, HMAC, zero dependencies, Dockerfile |
| `acb-starter-java` | Java | Javalin, javax.crypto.Mac, Maven, Dockerfile |
| `acb-starter-csharp` | C# | ASP.NET Core minimal API, System.Security.Cryptography HMAC, Dockerfile |

**CLI tools (built from monorepo, used locally):**

| Tool | Source | Purpose |
|------|--------|---------|
| `acb-local` | `cmd/acb-local/` | Run a match between two local bots (stdin/stdout), output replay JSON |
| `acb-mapgen` | `cmd/acb-mapgen/` | Generate symmetric maps with configurable parameters |

### 11.3 Build & Deploy Pipeline

Source (git push to main) │ ├──► Argo Events sensor (GitHub webhook) │ │ │ ├──► Argo Workflow: build-images │ │ ├── go test ./engine/... (game engine tests) │ │ ├── go test ./cmd/... (API/worker/evolver tests) │ │ ├── npm test (web) (SPA tests) │ │ ├── Kaniko image builds (all container images) │ │ └── Push to Forgejo registry │ │ │ └──► Argo Workflow: build-site │ ├── npm ci && npm run build (Vite build) │ ├── WASM builds (engine + 6 bots) │ └── Push site build artifact to Forgejo registry │ ├──► ArgoCD (watches declarative-config repo) │ └── Syncs K8s manifests -> ai-code-battle namespace │ └──► PostgreSQL migrations └── Run via init container or migration Job on deploy

Index builder Deployment (every ~15 min build, ~90 min deploy): │ ├── Query PostgreSQL directly ├── Generate JSON index files └── Deploy to Cloudflare Pages via wrangler pages deploy


### 11.4 Shared Libraries & Code Reuse

Several components share code. The monorepo structure avoids duplication:

| Shared Code | Used By | Language |
|-------------|---------|---------|
| `engine/` | `acb-worker`, `acb-evolver`, `acb-local`, `acb-mapgen`, WASM engine | Go |
| `engine/replay.go` | `acb-worker` (write), `acb-index-builder` (read for stats) | Go |
| `engine/winprob.go` | `acb-worker` (post-match computation) | Go |
| `engine/auth.go` | Match workers, matchmaker (HMAC verification) | Go |
| `engine/glicko2.go` | Match workers (rating updates on result write) | Go |
| `web/src/components/replay-canvas.ts` | Full viewer, embed, sandbox, homepage | TypeScript |
| `web/src/lib/data.ts` | All pages (data fetching + caching) | TypeScript |

The game engine is the foundational shared artifact — it compiles to:
1. A Go library (imported by worker, evolver, CLI tools)
2. A WASM module (loaded by the browser sandbox and embed viewer)
3. A test binary (run in CI)

---

## 12. Implementation Phases

### Phase 1: Core Engine (Foundation)

Build the game simulation as a standalone Go library with a CLI runner.

**Deliverables:**
- `engine/` package: grid, bots, energy, combat, fog of war, turn execution
- `cmd/acb-local/` CLI: run a match between two local bot processes
  (stdin/stdout for dev convenience) and output a replay JSON file
- Replay JSON writer
- Comprehensive unit tests for combat resolution, fog of war, wrapping,
  collision, scoring, endgame conditions
- Map generation tool: `cmd/acb-mapgen/`

**Exit criteria:** can run a complete 500-turn match between two bots locally
and produce a valid replay file.

### Phase 2: HTTP Protocol & Strategy Bots

**Deliverables:**
- HTTP bot interface in the engine (replaces stdin/stdout for production)
- HMAC signing and verification library (Go, reusable by GathererBot)
- GathererBot (Go) and RandomBot (Python) — validate the protocol works
  across languages before building the remaining four
- RusherBot (Rust), GuardianBot (PHP), SwarmBot (TypeScript), HunterBot (Java)
- All 6 bots containerized with language-appropriate Dockerfiles
- Starter kit template repos for each language (fork-ready)
- Integration test: engine runs a full match between bots in different
  languages over HTTP

**Exit criteria:** can run a complete match between any two strategy bot
containers (in different languages) over HTTP, with HMAC authentication,
producing a valid replay.

### Phase 3: Replay Viewer

**Deliverables:**
- TypeScript Canvas-based replay viewer
- Play/pause, scrub, speed control
- Fog of war perspective toggle
- Score overlay
- Loads replay JSON from local file or URL

**Exit criteria:** can open a replay file in a browser and watch a complete
match with all visual elements rendering correctly.

### Phase 4: Match Orchestration

**Deliverables:**
- Matchmaker Deployment (`acb-matchmaker`): internal tickers for pairing
  bots (1 min), health checking (15 min), stale job reaping (5 min).
  Enqueues job IDs into Valkey. No external exposure.
- PostgreSQL schema (CNPG): `bots`, `matches`, `match_participants`, `jobs`,
  `rating_history` tables in the `acb` database
- Index builder Deployment (`acb-index-builder`): reads PostgreSQL directly,
  generates index JSON files every ~15 min, deploys to Cloudflare Pages
  every ~90 min, manages R2 warm cache (promote from B2, prune old)
- Match worker Deployment (`acb-worker`): BRPOPs jobs from Valkey, runs
  matches, uploads replays to B2, writes results + Glicko-2 ratings to
  PostgreSQL
- Glicko-2 rating update logic in the match worker (runs on result write)

**Exit criteria:** matchmaker creates jobs and enqueues them in Valkey,
worker pods dequeue and execute them, replays land on B2, results flow
into PostgreSQL, ratings update, and leaderboard.json rebuilds automatically.
System recovers from worker pod failure via the stale job reaper.

### Phase 5: Web Platform

**Deliverables:**
- Cloudflare Pages static SPA (`ai-code-battle.pages.dev`): leaderboard,
  match history, bot profiles, replay viewer, docs/getting-started page
- SPA fetches replay files from R2 warm cache with B2 cold archive fallback
- Index builder deploying leaderboard, bot profiles, playlists to Pages
- Match workers uploading replays and per-match metadata to B2
- Index builder promoting recent replays from B2 to R2 warm cache

**Exit criteria:** anyone can browse matches, view leaderboards, and watch
replays — SPA from Pages, recent replays from R2, old replays from B2.
All data is static and pre-computed.

### Phase 6: Deployment & Production

**Deliverables:**
- K8s manifests committed to `declarative-config/k8s/apexalgo-iad/ai-code-battle/`:
  namespace, Deployments, Services, SealedSecrets (flat directory structure)
- Cloudflare Pages project (`ai-code-battle`, already exists at
  `ai-code-battle.pages.dev`)
- Cloudflare R2 bucket (warm replay cache, free tier)
- Backblaze B2 bucket (cold replay archive)
- ArgoCD Application syncing the manifests directory
- Argo Events sensor: GitHub webhook triggers on push to `ai-code-battle` repo
- Argo Workflows: image build (Kaniko -> Forgejo registry), site build
  (npm build -> artifact for Pages deploy)
- SealedSecrets for PostgreSQL, Valkey, B2, R2, Cloudflare API token
  (most already provisioned in the namespace)
- Monitoring: matchmaker metrics endpoint + Discord/Slack alerting webhooks

**Exit criteria:** platform is publicly accessible — SPA from Pages, recent
replays from R2, old replays from B2. All K8s manifests are GitOps-managed
by ArgoCD, CI pipelines rebuild images and site on git push, matches run
autonomously, and the leaderboard updates every ~90 minutes.

### Phase 7: LLM-Driven Evolution

**Deliverables:**
- Programs database with island model (4 islands, MAP-Elites behavior grid)
- Prompt builder: parent sampling, replay analysis, meta description
- LLM ensemble integration (fast + strong model tiers)
- Validation pipeline: syntax → schema → sandbox smoke test (nsjail)
- Evaluation arena: 10-match mini-tournament per candidate
- Promotion gate: Nash equilibrium computation (PSRO) + MAP-Elites niche fill
- Automated container build + deploy + register pipeline for promoted bots
- Retirement policy: auto-retire low-rated evolved bots, enforce population cap
- Evolution dashboard: lineage viewer, meta tracker, generation log
- Seed the programs database with the 6 built-in strategy bots as initial
  population

**Exit criteria:** evolution system runs autonomously — generates candidates,
validates, evaluates, promotes, deploys, and retires bots without human
intervention. At least one evolved bot reaches the top 50% of the ladder
within the first week of operation.

### Phase 8: Enhanced Features

**Deliverables:**
- WASM game engine build (`GOOS=js GOARCH=wasm`) with `loadState()`,
  `step()`, and `runMatch()` API for browser use
- WASM bot interface spec: `init()`, `compute_moves()`, `free_result()`
  exports for bot-to-engine communication
- Pre-compiled WASM builds of the 6 built-in strategy bots (Go/Rust/TS
  natively; PHP/Java reimplemented in Go for WASM)
- In-browser sandbox: Monaco editor (TS quick-start) + WASM upload mode +
  opponent selector + replay viewer integration
- Win probability computation in the match worker (Monte Carlo rollout) +
  critical moments detector + replay viewer sparkline graph
- Replay enrichment pipeline: selective AI commentary for featured matches
- Clip maker: GIF + MP4 export with 5 social media format presets
  (landscape, square, portrait, compact GIF, square GIF)
- Rival detection query + rivalry pages with template-generated narratives
- Community replay feedback system: tagged annotations feeding evolution
- PostgreSQL schema additions: `replay_feedback` table
- Go API addition: `POST /api/feedback` for submitting replay annotations

**Exit criteria:** users can write and test bots in the browser (TS
quick-start or uploaded WASM) without deploying anything, watch enriched
replays with commentary and win probability, export clips for social
sharing, view rivalries, and submit tactical feedback that influences the
evolution pipeline.

### Phase 9: Platform Depth

**Deliverables:**
- Bot debug telemetry: optional `debug` field in move response schema,
  stored in replay, rendered in viewer side panel + grid overlays
- Replay view modes: dots (default), Voronoi territory, influence gradient
  — all computed client-side, toggled via viewer toolbar
- Embeddable replay widget: `/embed/{match_id}` route on the static site, minimal
  Chrome, auto-play, ~50KB, Open Graph tags
- Replay playlists: auto-curated collections rebuilt by index builder
  Deployment, deployed to Pages, browsable on the static site
- Prediction system: PostgreSQL `predictions` table, Go API endpoints for
  submit + resolve, prediction leaderboard JSON deployed to Pages
- Map evolution pipeline: engagement scoring, breeding/mutation, symmetry
  validation, positional fairness monitoring, user map voting
- Multi-game series: PostgreSQL `series` table, series scheduler, unified
  replay presentation, spoiler toggle
- Match event timeline: client-side event extraction, icon ribbon in
  replay viewer, click-to-jump
- Seasonal system: PostgreSQL `seasons` table, ladder reset logic, season
  archive pages, versioned game rules with backward compatibility
- Bot profile cards: Canvas-rendered PNG, shareable URL with OG tags

**Exit criteria:** the platform supports seasonal competition with map
evolution, multi-game series, predictions for non-coders, embeddable
replays, curated playlists, three replay view modes, bot debug telemetry,
event timelines, and shareable bot profile cards.

### Phase 10: Ecosystem & Polish

**Deliverables:**
- Weekly meta report: auto-generated blog post deployed to Pages,
  rendered on `/blog` with LLM-enhanced narrative sections
- Public match data: documented static JSON file paths on Pages and R2,
  OpenAPI-style documentation at `/docs/api`, versioned replay format spec
- Accessibility suite: Tol color-blind palette + shape-per-player, keyboard
  shortcuts for replay viewer, high contrast mode, reduced motion, screen
  reader transcript, focus indicators
- Live evolution observatory: evolver writes `live.json` to R2
  every cycle, observatory page polls and renders live feed + lineage tree
  + meta shift chart
- Narrative engine: weekly story arc detection, LLM-generated 200-word
  chronicles, published alongside meta reports on `/blog`

**Exit criteria:** the platform publishes weekly editorial content (meta
report + story arcs) as blog posts, exposes all match data as documented
static JSON, meets WCAG accessibility standards for color and keyboard
navigation, and streams the evolution process as a live observatory.

---

## 13. Enhanced Features

### 13.1 In-Browser WASM Game Sandbox

The game engine and bots compile to WebAssembly, enabling users to develop
and test bots entirely in the browser against real opponents — zero
deployment, zero server setup.

**Architecture — WASM per module, not JS functions:**

A meaningful bot needs pathfinding, state tracking across turns, spatial
data structures, and threat assessment. That's a real program, not a
20-line JavaScript function. Limiting bots to JS callbacks would undermine
the platform's multi-language premise.

Instead, the sandbox loads **separate WASM modules** for the game engine
and each bot:

Browser ├── Game Engine (Go → WASM, ~15 MB) │ ├── loadState(json) → set engine to a specific turn state │ ├── step(moves[]) → advance one turn, return events │ └── runMatch(config, map) → run full match coordinating bot WASMs │ ├── Bot WASMs (pre-compiled, loaded on demand) │ ├── gatherer.wasm (Go → WASM, ~12 MB) │ ├── rusher.wasm (Rust → WASM, ~3 MB) │ ├── swarm.wasm (TypeScript → WASM via wasm-pack, ~5 MB) │ ├── random.wasm (Go → WASM, ~10 MB) -- lightweight reimpl │ ├── guardian.wasm (Go → WASM, ~12 MB) -- reimpl from PHP │ └── hunter.wasm (Go → WASM, ~12 MB) -- reimpl from Java │ ├── User's Bot WASM (compiled locally, uploaded as .wasm file) │ └── or: user writes Go/Rust/TS, compiles in-browser via toolchain │ ├── Monaco Editor (code editing for quick-start JS/TS mode) └── Replay Viewer (Canvas, renders result)


**WASM communication interface:**

Each bot WASM exports a standard interface:

// Exported by every bot WASM module fn init(config_json: *const u8, config_len: u32) fn compute_moves(state_json: *const u8, state_len: u32) -> *const u8 fn free_result(ptr: *const u8)


The engine WASM orchestrates the match: each turn, it serializes the
fog-filtered game state as JSON, calls each bot WASM's `compute_moves`,
deserializes the moves, and advances the simulation. Bots maintain their
own internal state across turns inside their WASM linear memory.

**Language support in the sandbox:**

| Language | WASM Compilation | Sandbox Support |
|----------|-----------------|-----------------|
| Go | `GOOS=js GOARCH=wasm` (native) | Full |
| Rust | `wasm32-unknown-unknown` (native) | Full |
| TypeScript | AssemblyScript or wasm-pack | Full |
| Python | Pyodide (~20 MB runtime) | Heavy but feasible |
| PHP | Not practical for WASM | HTTP ladder only |
| Java | Not practical for WASM | HTTP ladder only |

For the built-in opponents, GuardianBot (PHP) and HunterBot (Java) are
**reimplemented in Go** as sandbox-only WASM builds. They are behaviorally
equivalent — same BFS, same combat logic, same heuristics — not identical
code.

**Memory budget:**

| Configuration | Memory |
|--------------|--------|
| Engine + 1 user bot + 1 opponent | ~30–40 MB |
| Engine + 1 user bot + 3 opponents (4-player) | ~55–75 MB |
| Engine + 1 user bot + 5 opponents (6-player) | ~75–105 MB |
| With Pyodide (Python user bot) | Add ~20 MB |

Desktop browsers typically have 2–4 GB available. Even the heaviest
configuration is <5% of available memory. Mobile is tighter but the
sandbox is a desktop-first dev tool.

A 500-turn 2-player match simulates in ~2–3 seconds (WASM-to-WASM calls
have overhead vs native, but each turn's computation is trivial).

**User flows (two modes):**

*Quick-start mode (JS/TS in Monaco):*

1. User visits `/sandbox`
2. Monaco editor pre-loaded with a TypeScript starter bot
3. User writes strategy code with full type hints and autocomplete
4. Code compiles to WASM in-browser via AssemblyScript
5. Selects opponent and map, clicks "Run Match"
6. Engine orchestrates match between user WASM and opponent WASM (~2–3s)
7. Replay viewer renders result inline
8. Modify, re-run — instant feedback loop

*Full mode (upload compiled WASM):*

1. User develops a bot locally in Go, Rust, or any WASM-targeting language
2. Compiles to `.wasm` using their own toolchain
3. Uploads the `.wasm` file to the sandbox page
4. Sandbox validates the exported interface (`init`, `compute_moves`)
5. Runs match against selected opponents
6. When ready for the real ladder, deploys the same bot logic as an HTTP
   server using a starter kit

**Why this matters:** The sandbox preserves the platform's multi-language
strength while eliminating the deployment barrier. Users can develop
substantial, stateful bots in real languages — not toy JS functions —
and iterate locally before committing to the HTTP ladder.

### 13.2 Win Probability Graph + Critical Moments

Every match replay includes a **win probability curve** — a per-turn estimate
of each player's chance of winning — and a set of **critical moments** where
the game's outcome shifted decisively.

**Win probability computation:**

After each match, the worker computes win probability using Monte Carlo
rollout:

for each turn T in the match: state = game_state_at_turn_T wins = [0, 0, ..., 0] // per player for i in 1..100: result = simulate_random_play(state, remaining_turns) wins[result.winner] += 1 win_prob[T] = wins / 100


`simulate_random_play` runs the game engine with random valid moves for all
players from the given state to completion. 100 rollouts × 500 turns is
~50,000 engine steps — the Go engine handles this in <1 second.

The result is stored in the replay JSON as a `win_prob` array:
```json
"win_prob": [
    [0.50, 0.50],
    [0.51, 0.49],
    [0.48, 0.52],
    ...
]

Size: ~4 KB for a 500-turn, 2-player match. Negligible.

Critical moments:

A critical moment is any turn where |Δwin_prob| exceeds 0.15 (15%) for any player. Typically 3–5 per match. Stored in the replay JSON:

"critical_moments": [
    { "turn": 87, "delta": 0.22, "description": "SwarmBot loses 6 units in eastern engagement" },
    { "turn": 203, "delta": -0.31, "description": "GathererBot's core captured" }
]

The description is auto-generated from the turn's events (deaths, captures, large position changes). No LLM needed — template-based.

Replay viewer integration:

Sparkline graph below the main canvas: one line per player, color-coded
Horizontal axis: turns. Vertical axis: 0%–100% win probability
Critical moment markers: vertical dashed lines on the graph with labels
Click to jump: clicking any point on the graph scrubs to that turn
Quick nav buttons: "Next critical moment" / "Previous critical moment" to skip between turning points

This transforms replay viewing from "press play and wait 5 minutes" to "click the 3 interesting moments and watch 30 seconds of decisive action."

13.3 Replay Enrichment (Selective AI Commentary)

Select replays receive AI-generated natural language commentary — a play-by-play narration that makes matches accessible to casual viewers.

Not all replays are enriched. Commentary is generated selectively for:

Featured matches: matches flagged by the system as particularly interesting (high win probability variance, close finishes, upsets where a lower-rated bot wins)
Rivalry matches: matches between detected rivals (§13.5)
Evolution milestones: first match of a newly promoted evolved bot, or matches where an evolved bot breaks into the top 10
User-requested: participants can request enrichment for their own matches (rate-limited: 5 per day per bot)

Selection criteria (automatic):

enrich if:
  - win_prob crossed 0.5 at least 3 times (back-and-forth match)
  - final score difference ≤ 2
  - winner's rating was ≥100 lower than loser's (upset)
  - match involved a newly promoted evolved bot
  - match is between detected rivals

At ~60 matches/hour, roughly 10–15% qualify — about 6–9 enriched replays per hour.

Commentary generation:

Enrichment is performed by a coding agent in the cluster that takes the replay JSON + match metadata as input and generates a Markdown-formatted play-by-play output. The agent uses an LLM to analyze the replay data and produce structured commentary. This runs as a post-processing step on a dedicated container (not on the match worker itself).

Agent input:

Full replay JSON (turn-by-turn game state)
Match metadata (players, ratings, map size, win condition)
Win probability curve (sampled every 10 turns)
Critical moments array

Agent output: a Markdown file ({match_id}-commentary.md) stored alongside the replay on the PV:

# SwarmBot vs GathererBot — 60x60 Grid

## Turn 1: Opening
Both bots spawn at opposite corners of a 60x60 grid with heavy wall
cover in the center. SwarmBot immediately sends all units east in a
tight cluster.

## Turn 42: First Contact
GathererBot's scout stumbles into SwarmBot's formation near the central
energy cluster. The scout is outnumbered 8-to-1 and eliminated instantly.

## Turn 87: The Turning Point
SwarmBot pushes through the eastern corridor but GathererBot has quietly
amassed 14 units behind the western wall line — a force SwarmBot doesn't
know exists.

Cost: ~$0.01-0.03 per enriched match at Haiku-class pricing. At 9 enriched matches/hour: ~$2-6/day, ~$60-180/month. Reasonable.

Replay viewer integration:

The replay viewer fetches the companion .md file from Nginx and renders the Markdown as commentary subtitles below the canvas, synchronized to turn playback
Commentary sections are keyed to turn numbers; the viewer displays the relevant section as playback progresses
Toggle on/off via a "Commentary" button
Enriched replays are badged on the match list ("Featured" / "Narrated")

13.4 Shareable Replay Clips

One-click export of a replay segment as a GIF or video, formatted for major social media platforms. This is the viral growth engine.

Export formats:

Preset	Resolution	Aspect	Format	Target
Landscape	1920×1080	16:9	MP4	YouTube, Twitter, Discord
Square	1080×1080	1:1	MP4	Twitter, Instagram feed
Portrait	1080×1920	9:16	MP4	TikTok, YouTube Shorts, IG Stories
GIF (compact)	640×360	16:9	GIF	Discord embeds, forums
GIF (square)	480×480	1:1	GIF	Twitter, Slack

User flow:

While watching a replay, click "Clip" (scissors icon)
Drag handles on the turn scrubber to select a segment (default: 20 turns centered on the current turn, or the nearest critical moment)
Select format preset from dropdown
Optional: toggle overlays (score, win probability, commentary subtitles)
Click "Export"
Browser records the Canvas replay segment using OffscreenCanvas + MediaRecorder API (MP4/WebM) or gif.js (GIF)
Processing happens entirely client-side — no server cost
Download button appears, plus "Share" buttons:
- Twitter/X: opens compose with the clip attached + auto-generated text ("SwarmBot pulls off a comeback against HunterBot! 🎮 aicodebattle.com/replay/{id}")
- Reddit: copies a markdown link with embedded video
- Discord: downloads the file (under Discord's 25MB upload limit)
- Copy link: shareable URL to the replay at the specific turn range

Clip overlay: the exported clip includes:

Player names + colors in a header bar
Score overlay (bottom-left)
Win probability mini-graph (bottom strip, if enabled)
"aicodebattle.com" watermark (small, bottom-right)

GIF optimization: GIFs are limited to 256 colors and can be large. The clip maker uses:

Reduced frame rate (10 fps for GIF vs 30 fps for MP4)
Color quantization optimized for the grid art style
Max 10-second duration for GIFs (longer clips → MP4 only)
Target size: <5 MB for GIFs, <15 MB for MP4

Implementation: ~200 lines of TypeScript. MediaRecorder for MP4, gif.js for GIF, OffscreenCanvas for headless rendering. All runs in the browser. The share buttons use Web Share API where available, fallback to window.open() with pre-composed URLs.

13.5 Automatic Rival Detection

The platform automatically identifies rivalries — pairs of bots that frequently play each other with close results — and surfaces them as narrative-driven content.

Detection algorithm:

-- Run by the index builder Deployment
SELECT
    a.bot_id AS bot_a,
    b.bot_id AS bot_b,
    COUNT(*) AS matches,
    SUM(CASE WHEN winner = a.player_slot THEN 1 ELSE 0 END) AS a_wins,
    SUM(CASE WHEN winner = b.player_slot THEN 1 ELSE 0 END) AS b_wins
FROM match_participants a
JOIN match_participants b ON a.match_id = b.match_id AND a.bot_id < b.bot_id
JOIN matches m ON m.match_id = a.match_id
WHERE m.status = 'completed'
GROUP BY a.bot_id, b.bot_id
HAVING COUNT(*) >= 10
ORDER BY COUNT(*) * (1.0 - ABS(CAST(a_wins - b_wins AS REAL) / COUNT(*))) DESC
LIMIT 20

The ranking formula: matches_played × (1 - |win_rate_imbalance|). High-scoring pairs have many matches with near-50/50 results — the definition of a rivalry.

Rivalry page (/rivalry/{bot_a_id}/{bot_b_id}):

{
    "bot_a": { "id": "b_4e8c1d2f", "name": "SwarmBot", "owner": "alice" },
    "bot_b": { "id": "b_9a1b3c4d", "name": "HunterBot", "owner": "bob" },
    "matches": 23,
    "record": { "a_wins": 11, "b_wins": 11, "draws": 1 },
    "closest_match": "m_abc123",
    "longest_streak": { "holder": "b_4e8c1d2f", "length": 4 },
    "recent_matches": ["m_abc123", "m_def456", ...],
    "narrative": "SwarmBot and HunterBot have met 23 times — the series is dead even at 11-11-1. SwarmBot held a 4-match winning streak from Mar 15-18, but HunterBot answered with 3 straight victories. Their last match was decided by a single point."
}

The narrative is template-generated from the stats (no LLM needed):

"{bot_a} and {bot_b} have met {n} times — {record_description}.
{streak_description}. {recent_trend_description}."

Platform integration:

Rivalry widget on the landing page: "Top Rivalries" with head-to-head records and links to key matches
Bot profile pages show "Rivals" section listing detected rivalries
Rivalry matches are auto-flagged for replay enrichment (§13.3)
Leaderboard can show "rivalry mode" — filter to matches between two specific bots

13.6 Community Replay Feedback

Users can leave tagged feedback on specific moments in replays. Feedback is anchored to a (replay_id, turn) pair and is visible to other viewers. High-signal feedback is fed into the evolution pipeline as strategic hints.

Feedback types:

Type	Icon	Purpose
Tactical insight	💡	"This flanking move was brilliant because..."
Mistake spotted	⚠️	"Bot should have retreated here — outnumbered 3:1"
Strategy idea	🧪	"What if a bot used this wall corridor as a chokepoint?"
Highlight	⭐	"Amazing play" (lightweight, like a star/upvote)

PostgreSQL schema:

CREATE TABLE replay_feedback (
    feedback_id   TEXT PRIMARY KEY,
    match_id      TEXT NOT NULL,
    turn          INTEGER NOT NULL,
    type          TEXT NOT NULL,  -- 'insight', 'mistake', 'idea', 'highlight'
    body          TEXT NOT NULL,
    author        TEXT NOT NULL,  -- free text (no accounts, like registration)
    upvotes       INTEGER NOT NULL DEFAULT 0,
    created_at    TEXT NOT NULL
);

CREATE INDEX idx_feedback_match ON replay_feedback(match_id, turn);

Replay viewer integration:

Small markers appear on the turn scrubber at turns with feedback
Hovering shows a preview count: "3 comments at turn 87"
Clicking opens a side panel showing all feedback for that turn
Users can add their own feedback via a form in the side panel
Upvote button on each feedback item (1 per visitor via localStorage)

Feeding into evolution:

The evolution pipeline's prompt builder (§10.3) consumes community feedback as an additional signal:

The index builder Deployment aggregates high-upvote feedback of type idea and mistake into data/evolution/community_hints.json (written to the Nginx PV)
The evolver reads this file and includes the top-voted recent hints in the prompt:

## Community Tactical Insights (from replay annotations)

Replay m_abc123, Turn 87 (12 upvotes):
"The bot should have used the narrow wall corridor at (30,42)-(30,48)
as a chokepoint instead of engaging in the open. A defensive line of
3 units there could have held off the 8-unit swarm."

Replay m_def456, Turn 203 (8 upvotes):
"When outnumbered 3:1, retreating toward the nearest energy cluster
and spawning reinforcements is better than fighting — the focus combat
system guarantees you lose the 3v1."

If a resulting evolved bot performs well, the feedback items that contributed to its prompt are credited on the evolution dashboard: "Feedback from user 'tactician42' on replay m_abc123 contributed to evo-py-g42-7f3a (rating: 1720)"

Moderation:

Feedback is plain text, max 500 characters
No accounts means no banning — but feedback is public and upvote-ranked, so low-quality content sinks
A simple word filter catches obvious spam
The evolution pipeline only consumes feedback with ≥3 upvotes, filtering noise automatically
Admin can delete feedback via a PostgreSQL query (no UI needed initially)

Why this matters: It creates a human-AI collaboration loop. Spectators contribute strategic insight, the AI translates it into code, the platform evaluates it, and successful feedback is credited. This gives non-coders a way to participate meaningfully in the competition.

14. Platform Depth Features

14.1 Bot Debug Telemetry + Reasoning Visualization

Bots can optionally include a debug field in their move response. The engine stores it in the replay without interpreting it. The replay viewer renders it.

Extended move response schema:

{
  "moves": [
    { "row": 10, "col": 15, "direction": "N" }
  ],
  "debug": {
    "reasoning": "3 energy within 5 tiles east; enemy cluster north — avoiding",
    "targets": [
      { "row": 20, "col": 25, "label": "energy", "priority": 0.9 },
      { "row": 8, "col": 30, "label": "threat", "priority": 0.7 }
    ],
    "values": {
      "energy_reserves": 7,
      "threat_level": "medium",
      "mode": "gathering"
    },
    "heatmap": {
      "name": "threat",
      "data": [[0, 0, 0.2, 0.8], [0, 0.1, 0.5, 0.9]]
    }
  }
}

Schema rules for debug:

Entirely optional — bots that omit it behave identically
Max size: 10 KB per turn (prevents replay bloat; excess is truncated)
The engine never reads or acts on debug data — it's pass-through to replay
No fields inside debug are validated beyond size — bots can put anything
Only the bot's owner sees debug data by default; owners can toggle public visibility per-bot in their bot profile

Replay viewer rendering:

Debug field	Rendering
`reasoning`	Text in a collapsible side panel, one entry per turn
`targets`	Colored markers on the grid (green = high priority, red = low) with labels
`values`	Key-value table in the side panel, updates each turn
`heatmap`	Semi-transparent color overlay on the grid (blue→red gradient)

All debug rendering is toggled via a "Debug" button in the viewer toolbar. When off, no debug data is shown (default for spectators). When on, the viewer shows the selected player's debug output.

Replay size impact:

A bot sending 5 KB of debug data per turn across 500 turns adds 2.5 MB to the replay. With gzip compression (~90% on structured JSON), that's ~250 KB. Acceptable alongside the ~50 KB base replay.

Why it matters: This is a visual debugger for distributed bot code. Instead of reading logs, developers watch their bot's thought process alongside its actions. For spectators who opt in, seeing "the bot is scared of the northern cluster" while watching it move south creates narrative that no commentary system can match.

14.2 Territory Control Heatmap Overlay

The replay viewer supports three visualization modes, toggled via a toolbar dropdown. All computed client-side from bot positions — no server cost.

Mode 1: Dots (default)

The current view — bots as colored circles on the grid. Minimal, clean, fast.

Mode 2: Voronoi Territory

Each tile on the grid is colored by which player's nearest bot is closest. Creates clean territorial borders that shift each turn.

Computation per turn:
  for each visible tile (row, col):
    min_dist = infinity
    owner = none
    for each bot on the grid:
      d = toroidal_distance_squared(tile, bot)
      if d < min_dist:
        min_dist = d
        owner = bot.owner
    tile_color = player_colors[owner] at 30% opacity

For a 60×60 grid with 50 bots, that's 3,600 × 50 = 180,000 distance calculations per turn — trivial for modern JS (~1ms). The result is a per-tile color array rendered as a single full-grid Canvas fillRect pass underneath the bot sprites.

Mode 3: Influence Gradient

Force projection based on bot count and distance. Each player's influence at a tile is the sum of 1 / (1 + distance) across all their bots. Rendered as a smooth gradient:

for each visible tile:
    influence = [0, 0, ..., 0]  // per player
    for each bot:
        d = toroidal_distance(tile, bot)
        influence[bot.owner] += 1.0 / (1.0 + d)
    dominant = argmax(influence)
    strength = influence[dominant] / sum(influence)
    tile_color = player_colors[dominant] at (strength × 50%) opacity

The gradient creates a softer, more organic visualization than Voronoi — you can see where influence is strong (dense, saturated) vs weak (faint, contested). Frontlines appear as narrow bands where no player dominates.

Performance: both modes compute in <5ms per turn on a 60×60 grid. The replay viewer caches the overlay bitmap per turn and only recomputes on turn change. At 32 turns/second (16× speed), this stays within frame budget.

Toolbar UI:

View: [Dots ▼]  [Dots | Territory | Influence]

Switching modes is instant — the underlying replay data doesn't change, only the rendering pipeline.

A lightweight, standalone replay player that works in an iframe anywhere.

URL format:

https://aicodebattle.com/embed/{match_id}
https://aicodebattle.com/embed/{match_id}?start=87&speed=4&mode=territory

Query parameters:

Param	Default	Description
`start`	0	Starting turn
`speed`	2	Playback speed (1, 2, 4, 8, 16)
`mode`	dots	Visualization mode (dots, territory, influence)
`autoplay`	true	Start playing on load
`controls`	true	Show play/pause and speed controls

Widget design:

Stripped-down replay viewer: canvas + minimal controls bar. No scrubber, no side panel, no fog-of-war toggle. Just the match playing.

┌──────────────────────────────┐
│                              │
│        [Canvas]              │
│                              │
├──────────────────────────────┤
│ ▶ 2x  SwarmBot 3 — 1 Hunter │
│            Watch full ↗      │
└──────────────────────────────┘

"Watch full" links to the main replay page on aicodebattle.com.

Implementation:

Separate route served by Nginx: /embed/{match_id}
Loads the same replay JSON from Nginx
Renders with the same Canvas engine, minus chrome
Total bundle: ~50 KB (JS + CSS)

Open Graph tags for rich previews when pasting the URL:

<meta property="og:title" content="SwarmBot vs HunterBot — AI Code Battle" />
<meta property="og:description" content="SwarmBot wins 3-1 in 342 turns" />
<meta property="og:image" content="https://aicodebattle.com/thumbnails/m_7f3a9b2c.png" />

Thumbnail: auto-generated PNG of the final turn state, created by the index builder Deployment or pre-rendered by the match worker

Infrastructure impact: embed loads are static Nginx requests (cached by Cloudflare at the edge). Negligible additional load.

14.4 Replay Playlists + Auto-Curation

Automatically curated collections of replays, browsable from the static site's landing page.

Playlist definitions:

Playlist	Query Criteria	Rebuild Frequency
"Closest Finishes"	`final_score_diff <= 1` sorted by `win_prob_crossings DESC`	Every ~15 min (index builder deploy)
"Biggest Upsets"	`winner_rating - loser_rating <= -150`	Every ~15 min
"Best Comebacks"	`min(win_prob) < 0.2 AND winner = underdog`	Every ~15 min
"Evolution Breakthroughs"	Evolved bot's first win against a top-10 bot	Every ~15 min
"Rivalry Classics"	Matches between detected rivals, sorted by closeness	Every ~15 min
"This Week's Highlights"	Top 10 by community upvote count (from §13.6)	Every ~15 min
"New Bot Debuts"	First match of each newly registered bot	Every ~15 min
"Season Highlights"	Top 20 matches of the current season by engagement	Every ~15 min

PV storage: data/playlists/{slug}.json

{
  "name": "Closest Finishes",
  "description": "Matches decided by a single point or less",
  "updated_at": "2026-03-23T14:35:00Z",
  "matches": [
    {
      "match_id": "m_7f3a9b2c",
      "players": ["SwarmBot", "HunterBot"],
      "scores": [3, 2],
      "date": "2026-03-23T14:30:00Z",
      "thumbnail_url": "https://aicodebattle.com/thumbnails/m_7f3a9b2c.png",
      "enriched": true
    }
  ]
}

Static site UI: landing page shows playlists as horizontal scrollable rows (Netflix-style). Each card shows a thumbnail, player names, and score. Click opens the replay.

Infrastructure impact: playlist JSONs are tiny (<50 KB each). They're rebuilt by the index builder Deployment and written to the Nginx PV -- just additional PostgreSQL queries within the existing index build cycle.

14.5 Prediction System

Visitors predict outcomes of upcoming notable matches. Correct predictions earn reputation. A prediction leaderboard tracks the best analysts.

Which matches get predictions:

The matchmaker flags a match as "predictable" when:

Both bots are in the top 20
It's a rivalry match
It's a series match (§14.7)
An evolved bot faces a top-10 human-written bot

At ~60 matches/hour, roughly 5–10% are flagged — about 3–6 per hour.

Flow:

Scheduler creates a match job with predictable: true
Go API writes the match to a predictions_open state in PostgreSQL
Static site shows "Upcoming Matches" with a predict button
Visitor clicks a player to predict (stored via POST /api/predict)
Prediction window: open from job creation until the match starts executing (typically 1–5 minutes)
Match executes normally
On result submission, Go API resolves predictions in PostgreSQL
Index builder Deployment updates the prediction leaderboard JSON on the PV (next ~90-min deploy cycle)

PostgreSQL schema: (see §8.3 for the consolidated schema)

CREATE TABLE predictions (
    prediction_id   TEXT PRIMARY KEY,
    match_id        TEXT NOT NULL,
    predictor_id    TEXT NOT NULL,  -- localStorage-generated UUID
    predictor_name  TEXT,           -- optional display name
    predicted_bot_id TEXT NOT NULL, -- bot_id of the predicted winner
    correct         INTEGER,       -- null until resolved
    created_at      TEXT NOT NULL
);

CREATE TABLE predictor_stats (
    predictor_id    TEXT PRIMARY KEY,
    predictor_name  TEXT,
    correct         INTEGER NOT NULL DEFAULT 0,
    incorrect       INTEGER NOT NULL DEFAULT 0,
    streak          INTEGER NOT NULL DEFAULT 0,
    best_streak     INTEGER NOT NULL DEFAULT 0,
    rating          REAL NOT NULL DEFAULT 1000.0
);

Predictions are tied to bot identity (predicted_bot_id), not player slot. Resolution matches the winning bot's bot_id against the predicted bot_id. This avoids ambiguity when the same bot appears in different player slots across matches.

Predictor rating uses a simplified Elo: correct prediction on a balanced match (close ratings) = small gain; correct prediction on a heavy underdog = large gain.

Resource usage:

Metric	Usage	Notes
PostgreSQL writes	~6 predictions/match x 6 matches/hour x 24h = ~864/day	Negligible
PostgreSQL reads	~50 leaderboard reads/day	Negligible
Go API requests	`POST /api/predict` ~864/day	Negligible

Comfortably within cluster capacity at any realistic scale.

Static site UI:

"Predictions" page showing upcoming predictable matches with bot profiles and head-to-head records
One-click predict button (no login required — UUID from localStorage)
After match: result shown with "You were right/wrong" + points earned
Prediction leaderboard: top 50 analysts ranked by prediction rating

14.6 Map Evolution

Maps evolve alongside bots. High-engagement maps breed to produce new maps. Low-engagement maps retire. User feedback and positional fairness monitoring ensure quality.

Engagement scoring:

After each match, the map receives an engagement score:

engagement = (
    win_prob_crossings × 3.0 +
    critical_moments × 2.0 +
    map_coverage_pct × 1.0 +
    closeness × 2.0 +
    avg_turn_count / max_turns × 1.0
)

where:
    closeness = 1.0 - (abs(score_diff) / max(total_score, 1))
    map_coverage_pct = tiles_visited_by_any_bot / total_open_tiles

The map's engagement score is the rolling average across its last 20 matches.

Positional fairness monitoring:

A map is positionally fair if no starting position has a systematic advantage. Monitored by tracking win rate per player slot:

SELECT
    map_id,
    player_slot,
    COUNT(*) AS games,
    AVG(CASE WHEN winner = player_slot THEN 1.0 ELSE 0.0 END) AS win_rate
FROM match_participants mp
JOIN matches m ON m.match_id = mp.match_id
GROUP BY map_id, player_slot
HAVING COUNT(*) >= 80

If any player slot's win rate deviates from the expected rate (1/N for N-player maps) by more than 10 percentage points across 80+ matches, the map is flagged as unfair and removed from the competitive pool. At 80 matches with a 10pp threshold, the false positive rate from random variance drops to ~2% (compared to ~15% at 20 matches).

Example: on a 2-player map, if player slot 0 wins 58% of the time after 80 matches, the map is flagged (58% - 50% = 8% -- close to threshold, monitored). At 60%, it is flagged and removed.

User map voting:

After watching a replay, visitors can upvote or downvote the map (not the match -- the map). Stored in PostgreSQL:

CREATE TABLE map_votes (
    vote_id     TEXT PRIMARY KEY,
    map_id      TEXT NOT NULL,
    voter_id    TEXT NOT NULL,  -- localStorage UUID
    vote        INTEGER NOT NULL,  -- +1 or -1
    created_at  TEXT NOT NULL,
    UNIQUE(map_id, voter_id)
);

Map voting influences the evolution system:

Maps with net negative votes get a 0.5× engagement multiplier (less likely to breed)
Maps with >10 net positive votes get a 1.5× multiplier
Maps with >20 net negative votes are force-retired regardless of engagement

The replay viewer shows a simple 👍/👎 widget for the map (not the bots) alongside map metadata (name, dimensions, wall density, energy count).

Breeding algorithm:

Runs weekly on the evolver Deployment. Produces ~5 new maps per player-count tier.

1. Select parents:
   - Top 5 maps by engagement × vote_multiplier for this player count
   - Weighted random: higher engagement = more likely to be selected

2. Crossover:
   - Divide parent maps into quadrants (or thirds for 3/6-player)
   - Randomly select quadrants from each parent
   - Compose into a new map

3. Apply symmetry:
   - Generate one sector from the composed quadrants
   - Mirror/rotate to fill the full map for the target player count
   - This guarantees positional fairness by construction

4. Mutate:
   - Randomly flip 5-10% of tiles (wall ↔ open)
   - Shift 1-3 energy node positions by 1-3 tiles
   - Apply cellular automata smoothing (2 iterations) to avoid
     jagged walls

5. Validate:
   - BFS from every core must reach every other core
   - BFS from every core must reach ≥3 energy nodes
   - Open area per player must be between 900 and 5000 tiles
   - Wall density must be between 5% and 30%

6. Smoke-test:
   - Run 3 matches with built-in bots on the candidate map
   - Engagement score must exceed 50th percentile of current pool
   - If failed: discard and retry (max 3 attempts per candidate)

7. Add to pool:
   - Store map JSON on the Nginx PV (`maps/{map_id}.json`)
   - Insert into PostgreSQL maps table with `status: 'active'`
   - Available for matchmaking in the next scheduler cycle

Lifecycle:

Status	Meaning
`active`	In the matchmaking pool, eligible for competitive play
`probation`	Fairness flag triggered — under review, still playable
`retired`	Removed from pool (low engagement, unfair, or force-retired)
`classic`	Top 5 all-time maps, immune from retirement

Active pool: 50 maps per player count (2, 3, 4, 6)
New maps: ~5 per week per player count
Retirement: bottom 10% by engagement score pruned monthly
Classic promotion: maps that sustain top-5 engagement for 3+ months

14.7 Multi-Game Series

Best-of-N matches between two bots across different maps. Series produce more meaningful ratings than single matches and create narrative arcs.

Series types:

Type	Games	Trigger
Best-of-3	3	Auto-scheduled for top-20 bots, 1 per day per bot
Best-of-5	5	Weekly featured series between top rivalries
Best-of-7	7	Season championship bracket (§14.9)

Map selection for series:

Each game in a series uses a different map, selected to test different strategic dimensions:

Game 1: Map with highest engagement score (the "classic")
Game 2: Map with highest wall density in pool (corridors/chokepoints)
Game 3: Map with lowest wall density in pool (open field)
Game 4: Most recently evolved map (untested terrain)
Game 5+: Random from remaining pool

This ensures series test bot adaptability, not just performance on one map type.

PostgreSQL schema:

CREATE TABLE series (
    series_id     TEXT PRIMARY KEY,
    bot_a_id      TEXT NOT NULL,
    bot_b_id      TEXT NOT NULL,
    format        INTEGER NOT NULL,  -- 3, 5, or 7
    status        TEXT NOT NULL DEFAULT 'pending',
    a_wins        INTEGER NOT NULL DEFAULT 0,
    b_wins        INTEGER NOT NULL DEFAULT 0,
    season_id     TEXT,
    created_at    TEXT NOT NULL,
    completed_at  TEXT
);

CREATE TABLE series_games (
    series_id     TEXT NOT NULL,
    game_number   INTEGER NOT NULL,
    match_id      TEXT,  -- null until played
    map_id        TEXT NOT NULL,
    winner        INTEGER,
    PRIMARY KEY (series_id, game_number)
);

Execution:

The scheduler creates all games in a series as pending jobs with sequential ordering. Workers execute them in order (game 2 doesn't start until game 1 completes). If either bot reaches the winning threshold (2 for bo3, 3 for bo5, 4 for bo7), remaining games are skipped.

Rating impact:

Series results contribute to Glicko-2 ratings as follows:

Each individual game in the series contributes to the pairwise rating update (same as a single match)
The series winner gets a bonus rating adjustment of +10 mu (small but meaningful — rewards series consistency)

Replay presentation:

The series page (/series/{series_id}) shows all games as a unified experience:

SwarmBot vs HunterBot — Best of 5 (Season 4 Semifinals)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Game 1  ✓ SwarmBot    3-1  Map: The Labyrinth      [Watch]
Game 2  ✓ HunterBot   2-4  Map: Open Expanse       [Watch]
Game 3  ✓ HunterBot   1-3  Map: Coral Reef         [Watch]
Game 4    ???                                       [Reveal]
Game 5    ???                                       [Reveal]

Series: HunterBot leads 2-1

Spoiler toggle: by default, future games are hidden ("???"). Viewers click "Reveal" to show the result — or "Watch All" to experience the series sequentially with auto-advancing between games.

14.8 Match Event Timeline

A horizontal event ribbon below the replay canvas showing significant events as colored, clickable icons.

Event types:

Icon	Event	Trigger
⚔️	Combat	2+ bots died this turn
🏰	Core captured	A core was razed
💎	Energy milestone	Player collected 3+ energy in one turn
💀	Mass death	5+ bots died this turn
📈	Momentum shift	Win probability crossed 50%
🌟	Critical moment	Win probability shifted >15%
🐣	Spawn wave	3+ bots spawned this turn

Implementation:

Events are extracted client-side from the replay data on load. For each turn, check the events array (deaths, captures, spawns, energy_collected) against the trigger thresholds. Win probability events come from the win_prob and critical_moments arrays already in the replay.

Rendering:

┌──────────────────────────────────────────────────┐
│                  [Canvas]                         │
├──────────────────────────────────────────────────┤
│ Win Prob: ~~~~~~~~~/\~~~~~/\~~~~/\~~~~~~         │  ← sparkline
├──────────────────────────────────────────────────┤
│ Events: ·💎·····⚔️··💎···🏰⚔️···💎···⚔️💀··🏰🌟│  ← timeline
├──────────────────────────────────────────────────┤
│ ◄ ▶ ⏸  Turn 203/500   Speed: 4x   View: [Dots]│  ← controls
└──────────────────────────────────────────────────┘

Icons are positioned proportionally along the timeline by turn number
Hovering an icon shows a tooltip: "Turn 87: 3 bots killed in eastern corridor"
Clicking an icon scrubs the replay to that turn
Dense clusters of icons indicate "hot zones" of activity — visually obvious even at a glance
The timeline is rendered as an HTML element overlaid on the viewer (not Canvas) for accessibility and hover interactions

The event timeline and win probability graph work together: the graph shows the trend, the timeline shows the moments. A viewer can scan the timeline for icon clusters, then check the win probability graph to see if those moments mattered.

14.9 Seasonal Rotations

The platform runs in seasons — 4-week competitive periods with a fresh map pool, a new ladder, and a theme. Seasons provide urgency, freshness, and a reason to come back.

Season structure:

Week	Phase	Description
1	Discovery	New map pool + theme released. All bots start at default rating. Exploration matches.
2–3	Competition	Main ladder. Matchmaking intensifies. Mid-season stats published.
4	Championship	Top 8 bots by rating enter a best-of-7 bracket. Season champion crowned.
Between	Break (3 days)	New maps bred via map evolution. Season archive published.

What resets each season:

Glicko-2 ratings (mu/phi/sigma reset to defaults)
Map pool (evolved maps from previous season + new generated maps)
Prediction standings
Playlist contents

What persists:

Bot registrations and endpoints (bots don't re-register)
All-time records and historical season archives (browsable)
Evolution population (continues across seasons, adapts to new maps)
Community feedback and replay annotations

PostgreSQL schema:

CREATE TABLE seasons (
    season_id     TEXT PRIMARY KEY,
    name          TEXT NOT NULL,
    theme         TEXT NOT NULL,
    rules_version INTEGER NOT NULL,
    started_at    TEXT NOT NULL,
    ended_at      TEXT,
    champion_id   TEXT,
    status        TEXT NOT NULL DEFAULT 'active'
);

Season themes and game rule versioning:

Each season can introduce minor rule variations that keep the meta fresh. The critical constraint: existing bots must continue to work without modification. This is achieved through additive, optional changes only.

Backward compatibility rules:

ALLOWED per-season changes (additive, non-breaking):
  ✓ New tile types that bots can ignore (treated as open by old bots)
  ✓ New optional fields in the game state JSON (old bots ignore them)
  ✓ Adjusted numeric parameters within the existing schema:
    - vision_radius2, attack_radius2, spawn_cost, energy_interval
    - These are sent in the config object each match — bots that read
      config adapt automatically; bots that hardcode values still work
      but may be suboptimal
  ✓ New scoring bonuses (additive to existing scoring)
  ✓ Map pool changes (different maps, not different map format)

FORBIDDEN (would break existing bots):
  ✗ Removing or renaming existing fields in game state / move schema
  ✗ Changing the meaning of existing fields
  ✗ New required fields in the move response
  ✗ Changing the coordinate system or grid topology
  ✗ Removing movement directions (N/E/S/W)
  ✗ Changing the turn structure (phases must remain in the same order)

Example seasonal themes:

Season	Theme	Rule Variation
1	"The Labyrinth"	High wall density maps, `vision_radius2: 36` (reduced from 49)
2	"Energy Rush"	`energy_interval: 5` (doubled production), `spawn_cost: 2` (cheaper bots)
3	"Fog of War"	`vision_radius2: 25` (heavily reduced), new optional `sonar` field in game state showing approximate enemy count per quadrant
4	"The Colosseum"	`attack_radius2: 8` (extended range), open maps, aggressive meta
5	"Shifting Sands"	New tile type `quicksand` in game state (bots that don't handle it treat it as open — they can enter but movement costs 2 turns)

For season 5's quicksand example: the game state sends { "row": 15, "col": 20, "type": "quicksand" } in a new terrain array. Old bots that don't read terrain still function — they walk through quicksand unknowingly (and get slowed). New bots that parse terrain can avoid quicksand tiles, gaining a strategic edge. This creates an incentive to update bots each season without forcing anyone to.

Season config in the match protocol:

The game state's config object already includes all tunable parameters. Seasonal changes are just different values:

{
  "config": {
    "season_id": "s4",
    "season_name": "The Colosseum",
    "rules_version": 4,
    "rows": 60,
    "cols": 60,
    "max_turns": 500,
    "vision_radius2": 49,
    "attack_radius2": 8,
    "spawn_cost": 3,
    "energy_interval": 10,
    "special_tiles": ["quicksand"]
  }
}

Bots that read config.attack_radius2 adapt automatically. Bots that hardcode attack_radius2 = 12 still work but use stale assumptions. special_tiles is a new array listing any non-standard tile types in play — old bots that don't read it are unaffected.

Season archive:

Each completed season gets an archive page (/season/{season_id}):

Champion + top 10 + bracket results
Most improved bot (biggest rating gain)
Best newcomer (highest-rated bot registered this season)
Most watched match (by replay view count)
Evolution highlights (best evolved bot, most creative strategy)
Map of the season (highest engagement score)
All replays preserved and browsable

Season championship bracket:

In week 4, the top 8 bots enter a single-elimination bracket of best-of-7 series (§14.7). The bracket is published on the season page with live updates as series complete.

Quarterfinals:
  #1 SwarmBot vs #8 NewBot         → SwarmBot (4-1)
  #4 GathererBot vs #5 RusherBot   → RusherBot (4-3)
  #3 HunterBot vs #6 evo-go-g12    → HunterBot (4-2)
  #2 GuardianBot vs #7 evo-py-g8   → GuardianBot (4-0)

Semifinals:
  SwarmBot vs RusherBot             → SwarmBot (4-2)
  HunterBot vs GuardianBot          → HunterBot (4-3)

Finals:
  SwarmBot vs HunterBot             → ???

14.10 Bot Profile Cards

Auto-generated visual cards summarizing a bot's identity, stats, and character in a single shareable image.

Card generation:

The card is rendered as a PNG via OffscreenCanvas (in the browser on demand, or pre-rendered by the index builder Deployment for top-50 bots).

Card content:

┌─────────────────────────────────┐
│                                 │
│   SwarmBot              #3      │
│   by alice         Rating: 1820 │
│                                 │
│   ┌─────────────────────────┐   │
│   │ Archetype:              │   │
│   │   FORMATION SWARM       │   │
│   │                         │   │
│   │ Season 4 · 142 games    │   │
│   └─────────────────────────┘   │
│                                 │
│   Win Rate     69%  ████████░░  │
│   vs Rushers   82%  █████████░  │
│   vs Turtles   45%  ████░░░░░░  │
│                                 │
│   Signature: Eastern corridor   │
│   push on 4-player maps         │
│                                 │
│   Rival: HunterBot (11-11-1)   │
│                                 │
│   ⚔️ 847 kills  💎 2.1k energy  │
│   🏰 23 captures  📈 +320 Elo   │
│                                 │
│   aicodebattle.com              │
└─────────────────────────────────┘

Data sources (all from existing bot profile JSON):

Field	Source
Rating, rank	Leaderboard
Archetype	Strategy classifier from behavioral features (§10.2 MAP-Elites behavior grid)
Win rate breakdown	PostgreSQL query: wins vs each archetype cluster
Signature	Most statistically distinctive behavior vs population average
Rival	From rival detection (§13.5)
Kill/energy/capture stats	Aggregate from match_participants

"Signature" computation:

For each bot, compare its behavioral features (aggression, economy, exploration, formation) to the population mean. The dimension where the bot deviates most is its signature. Combined with map-type analysis:

if bot.aggression is 2σ above mean AND best_map_type == "4-player":
    signature = "Aggressive multi-front warfare on 4-player maps"
if bot.economy is 1.5σ above mean AND bot.exploration > 80%:
    signature = "Full-map economic dominance"

Template-generated from ~20 signature patterns.

Sharing:

"Share Card" button on the bot profile page generates a PNG download
Direct URL: https://aicodebattle.com/card/{bot_id}.png
- Served as a static PNG from Nginx PV (pre-rendered for top-50 bots)
- Or rendered on-demand by the Go API endpoint that reads the bot profile from PostgreSQL, draws to Canvas (using Go image libraries or a pre-built image template), and returns the PNG

Open Graph tags on the URL so pasting it into Twitter/Discord/Slack shows the card as a rich preview:

<meta property="og:image" content="https://aicodebattle.com/card/b_4e8c1d2f.png" />
<meta property="og:title" content="SwarmBot — #3 Rated — AI Code Battle" />

The card image includes the platform URL as a watermark, driving traffic

15. Ecosystem & Polish

15.1 Weekly Meta Report (Blog Posts)

Every Monday, the platform publishes a "State of the Game" blog post — an auto-generated analysis of the competitive landscape for the current season.

Published to: /blog/meta-week-{N}-season-{S} on the static site.

Blog infrastructure:

Blog posts are JSON files on Cloudflare Pages (data/blog/posts/{slug}.json), each containing:

{
  "slug": "meta-week-12-season-4",
  "title": "Week 12 Meta Report — Season 4: The Colosseum",
  "date": "2026-03-23",
  "type": "meta-report",
  "content_md": "# Week 12 Meta Report\n\n## Dominant Strategies\n...",
  "summary": "Swarm tactics dominate as attack_radius2 increase favors formations...",
  "tags": ["meta-report", "season-4"]
}

The static site's /blog page fetches data/blog/index.json (list of all posts) and renders them client-side with a Markdown renderer.

Report contents:

Section	Data Source	Generation
Dominant Strategies	Archetype distribution of top-20 bots	PostgreSQL query -> template
Rising / Falling Bots	Biggest rating movers (+/-) this week	PostgreSQL query -> template
Counter-Strategy Spotlight	Under-represented archetypes in top 20	PostgreSQL query -> LLM narrative
Map of the Week	Highest engagement map	PostgreSQL query -> template
Evolution Highlights	Promotion count, best evolved bot, most novel attempt	PostgreSQL query -> LLM narrative
Prediction Standings	Top 5 predictors, accuracy rates	PostgreSQL query -> template
Season Progress	Weeks remaining, championship seedings	PostgreSQL query -> template

Generation pipeline:

The index builder Deployment runs a weekly blog generation pass (triggered when dayOfWeek == Monday during its regular cycle)
Queries PostgreSQL directly for all data points above
Template-fills the structured sections (strategy distribution, ratings, maps, predictions)
Sends the free-text sections (counter-strategy spotlight, evolution highlights) to a cheap LLM with the data context + a journalism-style prompt
Assembles the full Markdown post
Writes the blog JSON file to the staging directory (data/blog/posts/{slug}.json)
Updates data/blog/index.json — deployed to Pages on the next cycle

Cost: one LLM call per week (~$0.05). Negligible.

Why blog posts: Blog posts are indexable by search engines (driving organic traffic), shareable as URLs, and accumulate into a historical record of the platform's competitive evolution. They also give the platform a human-feeling editorial voice even though the content is auto-generated.

15.2 Public Match Data (Static JSON)

All platform data is pre-computed and stored as static JSON files, split between Cloudflare Pages (indexes) and Cloudflare R2 (replays and per-match data). Index files are rebuilt every ~15 min by the index builder and deployed to Pages every ~90 min. Replays and per-match data are uploaded to R2 in real time by match workers. The "API" is simply documented file paths -- no dynamic endpoints, no query parameters, no rate limiting needed.

Documented data paths:

PAGES = https://aicodebattle.com        (Cloudflare Pages)
R2    = https://r2.aicodebattle.com     (Cloudflare R2)
API   = https://api.aicodebattle.com    (K8s Go API, dynamic only)

--- Index files on Pages (deployed every ~90 min by index builder) ---

Leaderboard:
  GET {PAGES}/data/leaderboard.json

Bot directory:
  GET {PAGES}/data/bots/index.json
  GET {PAGES}/data/bots/{bot_id}.json

Match index:
  GET {PAGES}/data/matches/index.json
  GET {PAGES}/data/matches/index-{page}.json   (older pages)

Series:
  GET {PAGES}/data/series/index.json
  GET {PAGES}/data/series/{series_id}.json

Seasons:
  GET {PAGES}/data/seasons/index.json
  GET {PAGES}/data/seasons/{season_id}.json

Playlists:
  GET {PAGES}/data/playlists/{slug}.json

Meta:
  GET {PAGES}/data/meta/archetypes.json
  GET {PAGES}/data/meta/rivalries.json

Evolution (indexes):
  GET {PAGES}/data/evolution/lineage.json
  GET {PAGES}/data/evolution/meta.json
  GET {PAGES}/data/evolution/community_hints.json

Blog:
  GET {PAGES}/data/blog/index.json
  GET {PAGES}/data/blog/posts/{slug}.json

Predictions:
  GET {PAGES}/data/predictions/leaderboard.json
  GET {PAGES}/data/predictions/open.json

Maps:
  GET {PAGES}/maps/index.json
  GET {PAGES}/maps/{map_id}.json

--- Real-time data on R2 (written by workers/evolver) ---

Individual match metadata:
  GET {R2}/matches/{match_id}.json

Replays:
  GET {R2}/replays/{match_id}.json.gz

Evolution (live feed):
  GET {R2}/evolution/live.json

Thumbnails:
  GET {R2}/thumbnails/{match_id}.png

Bot cards:
  GET {R2}/cards/{bot_id}.png

Replay format specification:

Published at /docs/replay-format on the static site. Contains:

JSON Schema file (replay-schema-v{N}.json) served by Pages -- third-party tools can validate replays programmatically
Field-by-field documentation with types, semantics, and examples
Versioning policy: additive changes only, matching the seasonal backward compatibility rules (§14.9). New fields may appear in future versions; old fields are never removed or renamed.
Example replays for each version (downloadable)
Changelog of schema changes per season

Documentation page (/docs/data):

A static page listing every data path above with descriptions, update frequency, and example curl commands. No authentication, no API keys, no rate limiting -- it's just static files served by Cloudflare.

Why static JSON, not a dynamic API:

All this data already exists as static files on Pages and R2. The index builder Deployment already produces leaderboard.json, bot profiles, match indexes, playlists, etc. Adding a dynamic API layer would add complexity for data that's already pre-computed and publicly readable. Cloudflare serves these globally with automatic CDN caching.

Third-party tools just fetch() the URLs. If they need to poll for updates, they check the updated_at field in each JSON file. Index files on Pages refresh every ~90 minutes. R2 cache headers guide freshness for replays (immutable) and the evolution live feed (10s).

15.3 Accessibility Suite

Color-blind safe palettes:

The platform ships with two palette options. Users toggle between them via a dropdown in the replay viewer toolbar. Preference persists in localStorage.

Players	Default	Color-Blind Safe (Tol)
Player 1	Blue (#2196F3)	Blue (#0077BB)
Player 2	Red (#F44336)	Orange (#EE7733)
Player 3	Green (#4CAF50)	Cyan (#009988)
Player 4	Yellow (#FFEB3B)	Magenta (#EE3377)
Player 5	Purple (#9C27B0)	Grey (#BBBBBB)
Player 6	Teal (#009688)	Black (#000000)

The Tol palette is designed by Paul Tol for maximum distinguishability under protanopia, deuteranopia, and tritanopia.

Shape-per-player (redundant encoding):

Each player's bots are rendered with a distinct shape in addition to color, ensuring identification without color vision:

Player	Shape
1	Circle ●
2	Square ■
3	Triangle ▲
4	Diamond ◆
5	Pentagon ⬠
6	Hexagon ⬡

Shapes are visible in all three view modes (dots, territory, influence). In territory/influence mode, bot sprites retain their shapes on top of the colored overlay.

Keyboard shortcuts:

Key	Action
`Space`	Play / Pause
`←` / `→`	Step back / forward one turn
`Shift+←` / `Shift+→`	Jump 10 turns
`[` / `]`	Previous / Next critical moment
`1`–`5`	Speed preset (1×, 2×, 4×, 8×, 16×)
`V`	Cycle view mode (dots → territory → influence)
`F`	Cycle fog of war perspective
`T`	Toggle debug telemetry panel
`E`	Toggle event timeline
`C`	Toggle commentary subtitles
`?`	Show keyboard shortcuts overlay

A "⌨️" icon in the toolbar opens the shortcuts reference as an overlay.

High contrast mode:

Toggled via toolbar or H key. Changes:

Grid lines: thin grey → bold white
Background: dark grey → pure black
Bot sprites: add 2px white outline
Territory/influence overlays: increase opacity from 30% to 50%
Energy nodes: yellow → bright white with yellow border
Walls: dark grey → medium grey with white border
Dead bots: fading red → solid white X

Reduced motion:

Respects the prefers-reduced-motion CSS media query automatically. When active:

Energy node pulse animation → static icon
Dead bot fade effect → instant removal
Bot movement trails → disabled
Combat flash → static highlight for one turn
Replay speed presets remain (this is user-controlled motion, not decorative)

Screen reader transcript:

A "Transcript" button in the toolbar opens a text panel showing a turn-by-turn summary generated from replay events:

Turn 87: Player 1 (SwarmBot) moved 8 bots east. Player 2 (HunterBot)
moved 3 bots south. Combat at (30,42): 2 SwarmBot units and 1 HunterBot
unit killed. SwarmBot collected energy at (25,38). Win probability:
SwarmBot 62%, HunterBot 38%.

Generated client-side from the replay data. ARIA live region announces each turn's summary during auto-playback.

Focus management:

All interactive elements have visible focus indicators (2px blue outline, offset by 2px for contrast)
Tab order follows a logical flow: toolbar → canvas (focusable for keyboard shortcuts) → scrubber → controls
Canvas receives focus on click; keyboard shortcuts only activate when canvas is focused (prevents conflicts with page-level shortcuts)
Skip-to-content link at page top for screen reader users

15.4 Live Evolution Observatory

The evolution dashboard becomes a real-time observatory where visitors watch the AI evolution system work — candidates being generated, tested, rejected, and promoted.

Data flow:

The evolver Deployment writes a status file to R2 at each stage of every evolution cycle:

Upload to R2: evolution/live.json
Served as:    https://r2.aicodebattle.com/evolution/live.json

Updated at every state transition: generation start, validation complete, each evaluation match result, promotion decision. At ~15 minutes per cycle with ~5 state transitions, that's ~20 R2 writes per hour. R2 serves the file with Cache-Control: max-age=10.

live.json schema:

{
  "updated_at": "2026-03-23T14:32:15Z",
  "cycle": {
    "generation": 847,
    "started_at": "2026-03-23T14:20:00Z",
    "phase": "evaluating",
    "candidate": {
      "id": "go-847-3",
      "island": "go",
      "language": "Go",
      "parents": [
        { "id": "go-831-1", "rating": 1580 },
        { "id": "go-839-2", "rating": 1540 }
      ],
      "community_hint": "try retreating when outnumbered 3:1",
      "validation": {
        "syntax": { "passed": true, "time_ms": 120 },
        "schema": { "passed": true, "time_ms": 450 },
        "smoke": { "passed": true, "time_ms": 3200 }
      },
      "evaluation": {
        "matches_total": 10,
        "matches_played": 4,
        "results": [
          { "opponent": "strategy-random", "won": true, "score": "5-1" },
          { "opponent": "strategy-swarm", "won": false, "score": "2-3" },
          { "opponent": "evo-go-g840", "won": true, "score": "4-2" },
          { "opponent": "strategy-hunter", "won": true, "score": "3-1" }
        ]
      }
    }
  },
  "recent_activity": [
    {
      "time": "2026-03-23T14:32:00Z",
      "generation": 847,
      "candidate": "go-847-2",
      "island": "go",
      "result": "rejected",
      "reason": "Nash gate: expected payoff -0.12 vs Nash mixture",
      "stage": "promotion"
    },
    {
      "time": "2026-03-23T14:28:00Z",
      "generation": 846,
      "candidate": "py-846-5",
      "island": "python",
      "result": "rejected",
      "reason": "Smoke test: crashed on turn 12",
      "stage": "validation"
    },
    {
      "time": "2026-03-23T14:25:00Z",
      "generation": 846,
      "candidate": "rs-846-1",
      "island": "rust",
      "result": "promoted",
      "bot_id": "evo-rs-g846",
      "initial_rating": 1500,
      "stage": "deployment"
    }
  ],
  "islands": {
    "python": { "population": 18, "best_rating": 1580, "best_bot": "evo-py-g820" },
    "go": { "population": 20, "best_rating": 1650, "best_bot": "evo-go-g831" },
    "rust": { "population": 17, "best_rating": 1520, "best_bot": "evo-rs-g846" },
    "mixed": { "population": 20, "best_rating": 1710, "best_bot": "evo-mx-g802" }
  },
  "totals": {
    "generations_total": 847,
    "candidates_today": 96,
    "promoted_today": 12,
    "promotion_rate_7d": 0.12,
    "highest_evolved_rating": 1710,
    "evolved_in_top_10": 3
  }
}

Observatory page (/evolution):

The static site polls /evolution/live.json every 10 seconds and renders:

Top bar: island overview

┌────────────┬────────────┬────────────┬────────────┐
│  🐍 Python  │  🔵 Go      │  🦀 Rust    │  🔀 Mixed   │
│  pop: 18   │  pop: 20   │  pop: 17   │  pop: 20   │
│  best: 1580│  best: 1650│  best: 1520│  best: 1710│
└────────────┴────────────┴────────────┴────────────┘

Center: current cycle status

Shows the current candidate's progress through the pipeline as a step indicator: [Generate] → [✓ Syntax] → [✓ Schema] → [✓ Smoke] → [Evaluating 4/10] → [Promotion?]

Below that, a mini-results table showing the candidate's evaluation matches as they complete: opponent, result, score.

If a community hint influenced this candidate's prompt, it's shown: 💡 Community hint: "try retreating when outnumbered 3:1" (by tactician42)

Bottom: activity feed

A scrolling log of recent evolution events, color-coded:

🟢 Promoted (green)
🔴 Rejected at validation (red)
🟡 Rejected at Nash gate (yellow)

Each entry shows the candidate ID, island, result, and reason.

Tabs: lineage tree + meta chart

Lineage tree: interactive d3.js force-directed graph. Each node is a bot (evolved or built-in). Edges connect parents to children. Nodes are colored by island. Size proportional to rating. Click a node to see the bot's profile. The tree grows as new bots are promoted.
Meta shift chart: stacked area chart (d3.js or Chart.js) showing the archetype distribution of the evolved population over generations. X-axis: generation number. Y-axis: percentage. Each archetype is a colored band. Watch strategies emerge, dominate, and get countered over time.

Both visualizations are built from data/evolution/lineage.json and data/evolution/meta.json (served from Pages, produced by the index builder Deployment). The live feed overlay is the only component that polls evolution/live.json (written by the evolver to R2).

15.5 Narrative Engine (Chronicles)

Auto-generated storylines from match data, published alongside the weekly meta report as blog posts on /blog.

Story arc detection:

The weekly index builder pass (same as the meta report, §15.1) scans PostgreSQL for active story arcs:

Arc Type	PostgreSQL Query Trigger
Rise	Bot gained >=200 rating in the last 7 days
Fall	Bot lost >=200 rating in the last 7 days
Rivalry Intensifies	Rivalry pair played 5+ matches this week with alternating wins
Upset of the Week	Biggest single-match rating gap where the underdog won
Evolution Milestone	Evolved bot reached a new all-time-high rating or entered top 5
Comeback	Bot recovered >=150 rating after a decline
Season Narrative	End of season (championship results, final standings)

Generation pipeline:

Detect 3-5 active arcs from PostgreSQL queries
For each arc, compile context: bot profiles, rating history, key match IDs with scores, archetype data, rival relationships
Prompt a cheap LLM (Haiku-class):

Write a 200-word sports-journalism narrative about this event in the
AI Code Battle platform. Be dramatic but factual. Reference specific
matches. Write in present tense. Do not use emojis.

Arc type: Rise
Bot: evo-go-g31
Season: 4 (The Colosseum)
Rating: 1320 → 1580 over 7 days
Key matches:
  - Beat SwarmBot (#1, 1820) on "The Labyrinth" — score 4-2, turn 287
  - Won bo3 series vs HunterBot (#4, 1650) 2-1
  - Lost to GuardianBot (#2, 1720) by 1 point on "Open Expanse"
Archetype: hybrid swarm-gatherer
Origin: evolved, generation 31, Go island
Parents: evo-go-g28 (gatherer archetype) × evo-go-g25 (swarm archetype)
Community hint that influenced it: "combine tight formations with
energy-first opening"

Assemble output as a blog post JSON file with:
- Headline (generated by LLM)
- 200-word narrative
- Embedded replay links for key matches
- Bot profile card image (§14.10)
- Rating chart (data for client-side rendering)
Write to staging directory: data/blog/posts/{slug}.json
Update data/blog/index.json — deployed to Pages on the next cycle

Blog page (/blog):

Lists all posts reverse-chronologically
Post types: meta-report and chronicle (story arcs)
Each post renders as a full page with embedded replay widgets (§13.3) at key moments
Tags for filtering: meta-report, rise, fall, rivalry, upset, evolution, comeback, season-recap

Weekly output: 1 meta report + 3–5 chronicles = 4–6 blog posts/week.

Cost: ~$0.05 per LLM call × 6 posts/week = ~$0.30/week, ~$1.30/month.

Why it matters: Chronicles transform raw match data into stories that people share, discuss, and follow. "The Rise of evo-go-g31" is a headline someone posts on Hacker News. "GathererBot's Decline" is a cautionary tale that sparks strategy discussion. The narrative engine gives the platform a voice — it feels alive, with characters and plot arcs, not just numbers on a leaderboard.

16. User Experience Design

The platform serves three distinct audiences with different needs. The UX must be simple enough that a first-time visitor understands the platform in 10 seconds, and deep enough that a regular user can access all features without friction. Mobile and desktop are both first-class.

16.1 Audiences

Audience	What They Want	Frequency
Spectator	Watch cool bot battles, browse leaderboard, follow stories	Daily, 5–15 min sessions
Participant	Build and improve bots, track performance, iterate	Several times/week, 30–60 min sessions
Visitor	Understand what this is, see something impressive, maybe come back	Once, 1–3 minutes

The default experience is optimized for spectators — the largest audience. Participants have dedicated sections. First-time visitors get a clear value proposition immediately.

16.2 Information Architecture

/                       Home (hero + featured replay + highlights)
├── /watch              Spectator hub
│   ├── /watch/replays  Browse all replays (playlists, search, filters)
│   ├── /watch/replay/{id}  Full replay viewer
│   ├── /watch/series/{id}  Series replay page
│   └── /watch/predictions  Predict upcoming matches
├── /compete            Participant hub
│   ├── /compete/sandbox    In-browser WASM sandbox
│   ├── /compete/register   Register a bot
│   ├── /compete/bot/{id}   Your bot's dashboard (owner view)
│   └── /compete/docs       Protocol spec, starter kits, guides
├── /leaderboard        Rankings (current season)
├── /evolution          Evolution observatory (live feed + lineage)
├── /blog               Meta reports + chronicles
├── /season/{id}        Season archive (past) or current season status
└── /bot/{id}           Bot public profile (anyone can view)

Three entry points, three audiences:

Spectators enter through /watch or the homepage highlights
Participants enter through /compete or /compete/sandbox
Visitors land on / and are guided to one of the above

16.3 Homepage

The homepage answers three questions in 10 seconds:

What is this? (headline)
What does it look like? (auto-playing featured replay)
What can I do? (two clear CTAs)

Layout:

┌──────────────────────────────────────────────────┐
│  AI Code Battle                                   │
│  Bots compete. Strategies evolve. You watch.     │
│                                                   │
│  [Watch Battles]          [Build a Bot]           │
├──────────────────────────────────────────────────┤
│                                                   │
│  ┌──────────────────────────────────────────┐    │
│  │  Featured Replay (auto-playing, muted)    │    │
│  │  Territory view, win probability graph    │    │
│  │  Commentary subtitles if enriched         │    │
│  └──────────────────────────────────────────┘    │
│  "SwarmBot vs HunterBot — Season 4 Semifinals"  │
│                                                   │
├──────────────────────────────────────────────────┤
│  Top 5 Leaderboard     │  Latest Stories          │
│  #1 SwarmBot    1820   │  "The Rise of evo-go-g31"│
│  #2 GuardianBot 1720   │  "Week 12 Meta Report"   │
│  #3 evo-mx-g802 1710   │  "Rivalry: Swarm v Hunt" │
│  #4 HunterBot   1650   │                          │
│  #5 evo-go-g831 1650   │                          │
│  [Full leaderboard →]  │  [All stories →]         │
├─────────────────────────┴─────────────────────────┤
│  Playlists                                        │
│  [Closest Finishes] [Biggest Upsets] [Comebacks]  │
│  ← scrollable cards with thumbnails →             │
├──────────────────────────────────────────────────┤
│  Season 4: The Colosseum — Week 3 of 4           │
│  Championship bracket starts in 8 days            │
│  [Predictions open →]                             │
├──────────────────────────────────────────────────┤
│  Evolution Observatory (mini)                     │
│  Gen #847 · 12 bots promoted today · 3 in top 10 │
│  [Watch evolution live →]                         │
└──────────────────────────────────────────────────┘

Key decisions:

The featured replay auto-plays in territory view (most visually impressive mode) — the visitor sees something immediately interesting without clicking anything
Two CTAs, not five — "Watch" for spectators, "Build" for participants
The leaderboard is a compact summary, not the whole page — the platform is about watching, not reading tables
Playlists are below the fold but above the season/evolution sections
Everything above the fold fits on a 1080p screen

Desktop: persistent top bar

┌──────────────────────────────────────────────────┐
│ ⚔️ AI Code Battle   Watch  Compete  Leaderboard  │
│                      Evolution  Blog   Season 4   │
└──────────────────────────────────────────────────┘

Logo + name (links to home)
Primary nav: Watch, Compete, Leaderboard
Secondary nav: Evolution, Blog, current Season

Mobile: bottom tab bar + hamburger

┌──────────────────────────────────────────┐
│  [content area]                           │
│                                           │
│                                           │
├──────────────────────────────────────────┤
│  🏠 Home  👀 Watch  ⚔️ Compete  🏆 Board │
└──────────────────────────────────────────┘

Four bottom tabs for the primary actions. Evolution, Blog, Season, and secondary pages are accessed via the hamburger menu (☰) in the top bar.

The bottom tab bar is the standard mobile pattern (iOS tab bar, Android bottom nav). Thumb-reachable, always visible, no scrolling needed to navigate.

16.5 Responsive Design

The platform is designed mobile-first for spectating and desktop-first for participating (writing bots).

Breakpoints:

Breakpoint	Target	Layout
<640px	Phone	Single column, bottom tab bar, touch-optimized
640–1024px	Tablet	Two column where useful, top nav
>1024px	Desktop	Full layout, sidebar where appropriate

Replay viewer on mobile:

The replay viewer is the most complex component. On mobile:

┌────────────────────┐
│                    │
│   [Canvas]         │  ← full width, 1:1 aspect ratio
│                    │
├────────────────────┤
│ ▶ 4x  Score: 3-1  │  ← compact controls bar
├────────────────────┤
│ ··⚔️··💎··🏰··⚔️·· │  ← event timeline (scrollable)
├────────────────────┤
│ Win Prob ~~~~~~~~~ │  ← sparkline (tap to scrub)
├────────────────────┤
│ Commentary text... │  ← if enriched, scrollable
└────────────────────┘

Canvas renders at full phone width with 1:1 aspect ratio (square grid maps fit naturally)
Pinch-to-zoom on the canvas (native gesture handling)
Tap to play/pause (large touch target — the entire canvas)
Swipe left/right on the timeline to scrub turns
View mode toggle (dots/territory/influence) via a floating button
Debug telemetry panel is a slide-up sheet (collapsed by default)

Leaderboard on mobile:

Simplified table with just rank, name, rating, and trend arrow (↑/↓). Tap a row to expand inline with games played, win rate, archetype. "Full stats" link goes to the bot profile page.

Sandbox on mobile:

The WASM sandbox is desktop-only. On mobile, the /compete/sandbox page shows a clear message: "The sandbox requires a desktop browser. On mobile, you can watch replays, browse the leaderboard, make predictions, and read the docs." With a QR code or "Send to desktop" link.

Bot registration (/compete/register) works on mobile — it's just a form.

16.6 First-Time Visitor Flow

A visitor who has never seen the platform:

1. Lands on homepage
   → Sees headline + auto-playing featured replay (territory view)
   → Understands "this is a platform where bots fight on a grid"

2. Watches for 10–30 seconds
   → Sees win probability shift, commentary subtitles, territory change
   → Understands "this is dynamic and strategic, not random"

3. Clicks "Watch Battles" or a playlist card
   → Enters /watch with curated playlists
   → Picks a replay that sounds interesting ("Biggest Upsets")

4. Watches a full replay (1–3 minutes at 4x speed)
   → Uses event timeline to skip to the action
   → Maybe makes a prediction on an upcoming match

5. Returns later
   → Checks predictions results
   → Browses new blog posts / chronicles
   → Eventually clicks "Build a Bot" → sandbox → starter kit → ladder

The funnel is: see (homepage) → watch (replays) → engage (predictions, feedback) → build (sandbox, compete). Each step has a lower barrier than the next.

16.7 Page Load Performance

The SPA should feel instant. Performance budget:

Metric	Target	How
First Contentful Paint	<1s	Cloudflare CDN, minimal critical CSS
Largest Contentful Paint	<2s	Defer replay loading, hero image as CSS gradient
Time to Interactive	<2s	Small JS bundle (<200KB gzipped), code-split per route
Replay load	<3s	Replay JSON gzipped (~50KB), streamed parse
WASM engine load	<5s	Lazy-loaded only on sandbox/embed pages

Code splitting strategy:

app.js        (~30KB gz)   Core SPA router, leaderboard, blog, nav
replay.js     (~80KB gz)   Replay viewer + territory renderer + charts
sandbox.js    (~20KB gz)   Sandbox orchestrator (loads WASM on demand)
engine.wasm   (~3MB)       Game engine (loaded only in sandbox/embed)
bot-*.wasm    (~3–12MB ea) Built-in bots (loaded only in sandbox)

The homepage loads only app.js. Replay viewer code is loaded when a user navigates to a replay. WASM is loaded only when the sandbox is opened. A visitor who just browses the homepage and leaderboard never downloads replay viewer or WASM code.

Data fetching:

All data files are served by Cloudflare (Pages for indexes, R2 for replays) with appropriate cache behavior. Subsequent visits hit the edge cache. The SPA uses stale-while-revalidate pattern: show cached data immediately, fetch fresh data in the background, update the UI when it arrives. The leaderboard never shows a loading spinner on repeat visits — it shows the cached version instantly and refreshes within seconds.

16.8 Design Language

Visual principles:

Dark theme by default — grid-based games look better on dark backgrounds. White text, subtle grid lines, colored elements pop.
Minimal chrome — let the replay canvas, leaderboard data, and content speak. No decorative borders, shadows, or gradients.
Color is information — player colors, archetype colors, win probability (green/red), and status indicators use color purposefully. Everything else is grey/white on dark.
Typography — monospace for data (ratings, scores, turn counts), sans-serif for prose (blog, commentary, descriptions). Two typefaces maximum.
Animation is functional — replay playback, win probability graph updates, evolution observatory feed. No decorative hover effects or page transitions.

Component library:

A small set of reusable components covers the entire site:

Component	Used In
`ReplayCanvas`	Replay viewer, embed, sandbox, homepage hero
`WinProbGraph`	Replay viewer, series page, bot profile
`EventTimeline`	Replay viewer
`LeaderboardTable`	Leaderboard page, homepage summary
`BotCard`	Bot profile, leaderboard, series, playlists
`PlaylistRow`	Homepage, /watch
`MatchCard`	Playlists, match history, series
`PredictionWidget`	Predictions page, homepage
`ObservatoryFeed`	Evolution page, homepage mini
`BlogPostCard`	Blog page, homepage
`AnnotationOverlay`	Replay viewer (spatial + text feedback)

Each component works at all breakpoints. The replay canvas adapts its render resolution to the container size. Tables collapse to cards on mobile. Graphs switch from detailed to sparkline on narrow screens.

16.9 Replay Canvas Micro-Animations

The replay renderer decouples game tick rate from render frame rate. Game state updates at the turn rate (2–32 ticks/second depending on speed setting). The Canvas renders at 60fps via requestAnimationFrame, interpolating positions and animation states between ticks. Bots slide smoothly between grid positions instead of teleporting.

Animation inventory:

Event	Animation	Duration
Bot idle	Subtle 2% scale pulse, 2s cycle. Stops on movement.	Continuous
Bot movement	1-tile motion trail fading behind the bot, indicates direction of travel. At high speed, trails create visible flow patterns.	150ms fade
Combat threat	Thin dashed line between bots within attack range (red). Shows who threatens whom.	1 turn
Attack event	Directed arrows from each attacker to the dying bot's tile on the turn a combat kill lands. The Go engine's `executeCombat()` needs to emit a `combat_death` event (type constant already exists: `EventCombatDeath`) alongside each `bot_died`, listing `killers: [{bot_id, owner, position}]` — the enemies within attack radius that triggered the outnumbering condition. The web viewer reads `combat_death` events and draws a solid line (attacker player color, arrowhead) from each killer's tile center to the defender's tile center, fading over 300ms. Old replays lacking `combat_death` events keep the existing proximity-inference lines. Distinct from the ambient threat dashes — these fire only on actual kills and encode exact participants rather than spatial proximity.	300ms fade
Bot death	Burst of 6–8 particles scattering outward from death position, fading to transparent.	400ms
Energy collection	4-line starburst radiating from the energy node + small "+1" text floating upward.	200ms
Core capture	Radial shockwave ring expanding from the core. Core color transitions from loser to capturer.	500ms
Bot spawn	New bot scales up from 0% to 100% with a soft glow matching the player color.	200ms

Particle system:

A pool of 100 reusable particle objects (prevents GC pressure). Each particle has: x, y, vx, vy, alpha, lifetime. On bot death, 6–8 particles are activated with random velocities. Each frame updates position and fades alpha. When alpha reaches 0, the particle returns to the pool.

interface Particle {
  x: number; y: number
  vx: number; vy: number
  alpha: number
  color: string
  lifetime: number  // frames remaining
}

const POOL_SIZE = 100
const pool: Particle[] = Array.from({ length: POOL_SIZE }, () => ({
  x: 0, y: 0, vx: 0, vy: 0, alpha: 0, color: '', lifetime: 0
}))

Performance: All animations are simple Canvas draw calls (arcs, lines, globalAlpha fades). With 50 bots and 20 active particles, frame time is <1ms on any modern device. The interpolation between ticks uses a simple lerp:

// In render loop:
const t = (now - lastTick) / tickInterval  // 0..1 between ticks
for (const bot of bots) {
  const renderX = bot.prevX + (bot.x - bot.prevX) * t
  const renderY = bot.prevY + (bot.y - bot.prevY) * t
  drawBot(renderX, renderY, bot.color, bot.shape)
}

16.10 Adaptive Auto-Speed Playback (Director Mode)

A playback mode where the replay automatically speeds through uneventful turns and slows for combat and critical moments.

Action density per turn:

action_density(turn) = (
    deaths × 3.0 +
    captures × 5.0 +
    energy_collected × 0.5 +
    spawns × 1.0 +
    abs(delta_win_prob) × 10.0
)

Speed mapping:

Action Density	Speed	Effect
0 (nothing)	16×	Boring turns fly by
0.1–1.0 (minor)	8×	Light scouting
1.0–3.0 (moderate)	4×	Engagement starting
3.0–5.0 (significant)	2×	Active combat
5.0+ (critical)	1×	Dramatic slowdown

Speed transitions are eased over 0.5 seconds (not instant) — the viewer feels the tempo shift.

User controls:

Activated via a "Director" option in the speed selector (alongside manual 1×, 2×, 4×, 8×, 16×)
Speed indicator shows live: ▶ Director 8× → 2× with smooth transition
Target duration slider: 30s / 1min / 2min / 5min — scales all speeds proportionally to approximate the target total playback time
Manual scrubbing pauses Director Mode until the scrubber is released

Computation: action density for all turns is pre-computed on replay load (single pass through the turns array, <1ms). A speed schedule array maps each turn to its target speed. During playback, the tick interval adjusts smoothly.

16.11 Smooth View Mode Morphing

Switching between dots, Voronoi territory, and influence gradient uses a 300ms animated cross-fade morph instead of a hard cut.

Transition technique:

The renderer maintains two off-screen Canvas buffers — one for the current view mode's overlay, one for the target. During a transition, a blend parameter t eases from 0 to 1 over 300ms (ease-in-out).

function renderOverlayTransition(ctx: CanvasRenderingContext2D, t: number) {
  // Draw outgoing mode, fading out
  ctx.globalAlpha = 1 - t
  renderOverlay(currentMode)

  // Draw incoming mode, fading in
  ctx.globalAlpha = t
  renderOverlay(targetMode)

  // Bots always on top at full opacity
  ctx.globalAlpha = 1
  renderBots()
}

Mode-specific transitions:

Dots → Territory: Color expands outward from each bot's position to fill Voronoi cells — like paint spilling from unit positions
Territory → Influence: Sharp Voronoi borders soften and bleed into gradients at contested zones
Influence → Dots: Colored overlay fades uniformly to transparent

Cost: two overlay renders per frame during the 300ms transition (~18 frames at 60fps). Each overlay is a single-pass pixel fill — no performance concern.

16.12 "Follow Bot" Camera Mode

Lock the viewport to track one player's units. The camera pans and zooms dynamically to keep the followed player's bots centered and visible.

Camera algorithm:

Each frame:
  1. Compute bounding box of tracked player's living bots
  2. Add 8-tile margin on each side (reveals approaching enemies)
  3. Lerp viewport center toward bounding box center (factor: 0.15)
  4. Lerp zoom level to fit bounding box + margin (factor: 0.10)
  5. Clamp: never zoom closer than 15×15 tiles, never wider than full grid

The slower lerp on zoom (0.10 vs 0.15) ensures zoom changes are gentler than panning — prevents disorientation from rapid zoom oscillation.

Split group handling: if the tracked player's bots split into groups

20 tiles apart, the camera briefly zooms out to show both groups, then follows the larger group once they're too far apart to fit.

Activation:

Click a player name/color in the score overlay
Press 1–6 to follow that player number
Press 0 or Esc to return to full-grid view
Mobile: tap a player's color swatch in the score header

Pairs with fog of war: Follow mode + fog-of-war perspective on the followed player creates a first-person experience — you see what they see, the camera goes where they go. Enemies emerge from darkness at the edge of vision.

16.13 Picture-in-Picture Replay

Navigate away from a replay and it minimizes to a floating mini-player in the bottom corner. The replay keeps playing while you browse.

Behavior:

User is watching a replay on /watch/replay/m_7f3a
User clicks a link to any other page
SPA router's beforeNavigate hook detects the replay viewer is active

Instead of unmounting, the Canvas element is reparented into a fixed-position container:

.pip-container {
  position: fixed;
  bottom: 16px;
  right: 16px;
  width: 200px;
  height: 200px;
  z-index: 1000;
  border-radius: 8px;
  box-shadow: 0 4px 20px rgba(0,0,0,0.5);
  cursor: pointer;
}

Canvas continues rendering at 200×200 resolution. Same requestAnimationFrame loop, same playback speed. Animations and game state are uninterrupted.
Mini-player shows: tiny canvas, current score, play/pause icon
Click mini-player → navigates back to replay page at the current turn
Click ✕ → closes the mini-player

Transitions:

Minimize: canvas scales from full-size to 200×200 with ease-out (300ms), sliding to bottom-right. Page content fades in underneath.
Expand: reverse. Uses CSS transform: scale() translate() with will-change: transform for GPU-accelerated animation. Zero layout reflows.

Mobile: mini-player is 120×120, positioned above the bottom tab bar.

16.14 Performance Trifecta

Three techniques that together make the site feel like a native app.

1. Route preloading on hover:

When the user hovers a link for 100ms (desktop) or fires touchstart (mobile), fetch the target page's primary data file:

function preloadOnHover(link: HTMLAnchorElement) {
  let timer: number
  link.addEventListener('mouseenter', () => {
    timer = setTimeout(() => {
      const dataUrl = routeToDataUrl(link.pathname)
      if (dataUrl && !preloadCache.has(dataUrl)) {
        preloadCache.add(dataUrl)
        fetch(dataUrl)  // browser HTTP cache stores the response
      }
    }, 100)
  })
  link.addEventListener('mouseleave', () => clearTimeout(timer))
}

By the time the user clicks, the data is already in the browser cache.

2. Skeleton screens:

Every page has a skeleton that exactly matches its content layout:

Leaderboard: rows of grey bars matching rank/name/rating column widths
Bot profile: grey circle (avatar area), bars for name/rating/stats
Replay page: grey rectangle (canvas area) + thin bar (scrubber)

Skeletons use a shimmer animation: a light gradient (linear-gradient) sweeps left-to-right at 1.5s intervals. Skeleton → real content: opacity fade-in over 150ms, zero layout shift (skeleton and content occupy identical space).

3. Instant back navigation:

const routeCache = new Map<string, {
  html: string,
  scrollY: number,
  data: any
}>()

// On navigate away:
routeCache.set(currentPath, {
  html: contentEl.innerHTML,
  scrollY: window.scrollY,
  data: currentPageData
})

// On back navigation:
const cached = routeCache.get(targetPath)
if (cached) {
  contentEl.innerHTML = cached.html
  window.scrollTo(0, cached.scrollY)
  // Optionally: background-refresh stale data
}

Cache holds the last 5 pages. Back navigation is 0ms.

Combined result:

Navigation	Time
Forward (hovered)	0ms — data preloaded
Forward (not hovered)	200ms skeleton → 300ms data
Back	0ms — cached
First visit	200ms skeleton → 500ms data

16.15 Progressive Disclosure

The replay viewer reveals features gradually based on user engagement.

Experience tracking:

// localStorage
let viewerXP = parseInt(localStorage.getItem('viewer_xp') || '0')

// Increment when user watches a replay for >30 seconds
if (watchDuration > 30_000) {
  viewerXP++
  localStorage.setItem('viewer_xp', String(viewerXP))
}

Feature revelation schedule:

XP Level	Controls Visible	New This Level
0	Play/pause, speed, scrubber	— (first visit overlay: "Space = play/pause, ←/→ = step, 1-5 = speed")
2	+ Event timeline	"New: Event Timeline — key moments at a glance"
5	+ View mode toggle, critical moment nav	"New: Territory View — see who controls the map"
10	+ Follow mode, fog perspective, clip export	"New: Follow a bot — camera tracks one player"
20	+ Debug telemetry, annotations, Director Mode	All controls visible

Revelation animation: new controls slide in from below (200ms ease-out) with a brief golden border pulse and a tooltip that fades after 3 seconds.

Manual override: ☰ menu → "Show all controls" reveals everything immediately. Power users are never gated.

16.16 Swipe-Through Playlists (Mobile)

On mobile, playlists become full-screen, auto-playing cards. Swipe up to advance.

Layout per card:

┌────────────────────┐
│  Closest Finishes   │  ← playlist name (sticky, translucent)
│  3 of 12            │
├────────────────────┤
│                    │
│  [Replay Canvas]   │  ← full-width, auto-plays in Director Mode
│                    │
├────────────────────┤
│ SwarmBot 3-2 Hunt  │  ← compact score bar
│ ⚔️💎🏰  ~45s left  │  ← event icons + estimated remaining time
├────────────────────┤
│  ↑ swipe for next  │  ← hint (fades after first swipe)
└────────────────────┘

Gestures:

Gesture	Action
Swipe up	Cross-fade to next replay (300ms transition)
Swipe down	Go back to previous replay
Tap canvas	Play/pause
Tap score bar	Expand win probability graph + commentary
Swipe right / tap ✕	Exit playlist mode

Auto-advance: when a replay ends, pause 3 seconds showing final score with a countdown ring animation, then advance to next replay. Tap to cancel.

Preloading: while the current replay plays, the next replay in the playlist is fetched in the background. Swipe transitions are instant.

16.17 Theater Mode

Fullscreen, chrome-free replay viewing.

Desktop: press F or click the fullscreen icon.

┌──────────────────────────────────────────────────────┐
│                                                       │
│                                                       │
│           [Full-screen Canvas at native res]           │
│          Vignette effect at edges (subtle)             │
│                                                       │
│                                                       │
├───────────────────────────────────────────────────────┤
│ ▶ Director  SwarmBot 3 · HunterBot 1   Turn 203/500 │  ← semi-transparent,
│ ··⚔️··💎··🏰⚔️··💎···⚔️💀··🏰🌟··                    │    fades after 3s
└───────────────────────────────────────────────────────┘

Background: pure black
Controls bar: semi-transparent, auto-hides after 3 seconds of inactivity. Mouse move / tap to reveal.
Win probability: two thin colored bars at the top edge of the screen (proportional width, no chart — ambient information)
Critical moment vignette pulse: edges darken briefly when win probability shifts >15%, creating a cinematic "something happened" cue

Mobile: triggered by rotating to landscape on a replay page. Same layout. Touch controls:

Gesture	Action
Tap	Play/pause
Swipe left/right	Step turns
Pinch	Zoom
Two-finger tap	Cycle view mode
Swipe up from bottom	Reveal controls bar

Exit: Esc (desktop) or rotate to portrait (mobile).

16.18 Ambient Activity Awareness

Subtle, non-intrusive signals that keep users aware of platform activity.

Favicon badge:

Dynamic favicon updated via Canvas + <link rel="icon"> swap:

State	Favicon	Trigger
Normal	⚔️	Default
Match result	⚔️🔴	Your bot finished a match (detected via data poll)
Prediction resolved	⚔️🟡	A prediction you made was resolved
Season event	⚔️🟢	Championship bracket update, season milestone

Badge clears when the user focuses the tab.

Tab title updates:

Default:         "AI Code Battle"
Unread result:   "(1) AI Code Battle"
Bot won:         "✓ SwarmBot won! — AI Code Battle"
Prediction:      "You were right! — AI Code Battle"

Resets on document.visibilitychange → visible.

Mobile haptic:

Brief 50ms vibration pulse at each critical moment during replay playback (opt-in toggle in viewer controls):

if ('vibrate' in navigator && prefs.haptic) {
  navigator.vibrate(50)
}

Seasonal background color shift:

The page background subtly shifts hue per season. Not a different color — a subtle tint on the base dark grey:

Season	Base `#1a1a2e` Shifts To	Mood
The Labyrinth	`#1e1a2e` (hint of purple)	Mysterious
Energy Rush	`#1a2e1e` (hint of green)	Abundant
Fog of War	`#1a1a3e` (cooler blue)	Uncertain
The Colosseum	`#2e1a1a` (hint of red)	Aggressive

Most users won't consciously notice. The platform just feels different each season.

Live match counter (homepage):

⚔️ 1,847 matches today · 23 bots active · Gen #852 evolving

Updated every 30 seconds from evolution/live.json. Shows the platform is alive.

229 KiB Raw Blame History Unescape Escape

AI Code Battle — Implementation Plan

1. Overview

2. System Architecture

Cloudflare (Static Tier)

Backblaze B2 (Cold Archive)

apexalgo-iad (Compute Tier)

Go API (deferred)

Data Architecture

Component Summary

3. Game Mechanics

3.1 Map & Grid

3.2 Units (Bots)

3.3 Energy & Economy

3.4 Combat

3.5 Fog of War

3.6 Scoring & Win Conditions

3.7 Turn Structure

3.7.1 Zone Parameters

3.8 Map Generation

4. Communication Protocol

4.1 HTTP Interface

4.2 Game State Schema (Engine → Bot)

4.3 Move Schema (Bot -> Engine)

4.4 Authentication (HMAC Shared Secret)

4.5 Timeout & Error Handling

5. Strategy Bots

5.1 RandomBot — Python

5.2 GathererBot — Go

5.3 RusherBot — Rust

5.4 GuardianBot — PHP

5.5 SwarmBot — TypeScript

5.6 HunterBot — Java

5.7 Container Templates

5.8 Starter Kit & SDK Libraries

6. Tournament System

6.1 Matchmaking

6.2 Rating System

6.3 Continuous Tournament

7. Replay System

7.1 Replay Data Format

7.2 Storage

7.3 Browser Replay Viewer

8. Web Platform

8.1 Cloudflare Pages (Static Site)

8.2 Go API Service

8.3 PostgreSQL (CNPG)

8.4 Index Builder (Deployment)

8.5 Bot Registration

8.6 Leaderboard

8.7 Match History & Profiles

9. Deployment & Infrastructure

9.1 Design Principles

9.2 Kubernetes Namespace Layout

9.3 Container Images

9.4 Match Job Coordination

9.5 Resilience & Pod Failure

9.6 Networking & Security

9.7 ArgoCD & GitOps

9.8 CI/CD Pipelines (Argo Workflows + Events)

9.9 Monitoring

10. LLM-Driven Bot Evolution

10.1 Architecture Overview

10.2 Programs Database (Island Model)

10.3 Prompt Construction

Parent B — Rating: 1580, Style: defensive-swarm

Parent A's Worst Loss (Replay Summary)

Current Meta

Your Task

13.3 Replay Enrichment (Selective AI Commentary)

13.4 Shareable Replay Clips

13.5 Automatic Rival Detection

13.6 Community Replay Feedback

14. Platform Depth Features

14.1 Bot Debug Telemetry + Reasoning Visualization

14.2 Territory Control Heatmap Overlay

14.3 Embeddable Replay Widget

14.4 Replay Playlists + Auto-Curation

14.5 Prediction System

14.6 Map Evolution

229 KiB

Raw Blame History