Add bot to play against.

CLAUDE.md

@@ -0,0 +1,38 @@
+# Super Auto Pets: The Board Game
+
+Go backend (pure rules engine in `internal/game`, WebSocket rooms in
+`internal/server`, computer opponent in `internal/ai`) + React frontend in
+`web/`. See README.md for the full layout and rules summary.
+
+- `mise run test` — Go tests · `mise run check` — go vet + frontend tsc
+
+## Keep the AI in sync with game behavior
+
+**Whenever game behavior changes — rules, cards, effects, phases, log
+entries, or views — make sure the computer opponent accounts for it.** The
+AI plays from the same information a human sees, so changes ripple into it
+in specific ways:
+
+- **Battle rules**: `internal/ai` evaluates moves by running the real
+  resolver (`game.SimulateBattle`), so battle changes are picked up
+  automatically — but re-check the hand-written heuristics that summarize
+  battle wisdom: `leadScore`, `keepValue`, `deckValue` (eval.go) and the
+  food-placement strategies / synergy pet list (arrange.go).
+- **New or changed cards/effects**: shop-time deck effects are mirrored in
+  `applyTemplateShopEffects` (shop.go); a new shop-time trigger or action
+  must be added there or the bot will misvalue it.
+- **New public actions or log changes**: the bot tracks the opponent via
+  structured tags on public log entries (`LogBuy`, `LogSell`, `LogTrade`,
+  `LogTradePick`, spawn counts — see log.go). New public actions need tags
+  plus handling in `Observe` (memory.go). Tags must only ever duplicate
+  facts the entry's text already states publicly.
+- **View changes**: the AI decides from `game.View` only — never hand it
+  the `Game`. If a field is added to the view, confirm it doesn't leak
+  hidden information (deck order, trade options, shop decks), because the
+  AI (and any client) would legitimately see it.
+- **Phase/flow changes**: the server's bot driver (`internal/server/bots.go`)
+  must know when a bot owes an action (`ai.Pending`) and have a legal
+  fallback (`botFallback`) for any new phase or forced decision.
+
+After any such change, run `go test ./internal/ai/` — it plays complete
+bot-vs-bot games and fails on any illegal or missing bot action.

README.md

@@ -1,7 +1,8 @@
 # Super Auto Pets: The Board Game — Online
 
 A web app for playing the Super Auto Pets board game remotely. 1v1 for now;
-the engine is built to grow to more players.
+the engine is built to grow to more players. Play a friend by room code, or
+play solo against a computer opponent (easy / medium / hard).
 
 ## Stack
 
@@ -94,8 +95,22 @@ All six tiers use the real card data:
 ```
 cmd/server/          entrypoint
 internal/game/       rules engine (pure, fully tested)
-internal/server/     HTTP + WebSocket rooms
+internal/ai/         computer opponent (decides from a player View only)
+internal/server/     HTTP + WebSocket rooms; drives bot turns
 internal/store/      SQLite persistence
 internal/env/        .env loading
 web/                 React frontend
 ```
+
+## The computer opponent
+
+Any seat can be a bot (`Player.IsBot`); humans and bots are interchangeable
+to the engine, which is what will let future >2-player games mix them
+freely. The AI in `internal/ai` never touches the `Game` — it decides from a
+`game.View`, the same per-player state a human client is sent, plus a
+persisted memory of public observations (battle lineups, the event log, shop
+row changes). It cannot see your deck order, hidden trade picks, or the
+shuffled shop decks. It scores candidate moves by Monte-Carlo battle
+rollouts (`game.SimulateBattle`) against sampled guesses of your deck and
+ordering, blended with a long-term deck-value heuristic; difficulty tunes a
+softmax over the scored moves plus the rollout budget.

internal/ai/ai.go

@@ -0,0 +1,173 @@
+// Package ai implements a computer-controlled player.
+//
+// The bot is strictly information-hygienic: every decision is made from a
+// game.View — the exact same state the server would send a human sitting in
+// that seat — plus a Memory built purely from past public observations
+// (battle lineups, the shared event log, and shop-row changes). The bot never
+// touches the Game struct, so it cannot read the opponent's secret deck
+// order, hidden trade picks, or upcoming shop cards even by accident.
+//
+// Decisions are made by generating candidate moves and scoring each one as a
+// blend of two signals:
+//
+//   - immediate: the estimated probability of winning the next battle,
+//     measured by Monte-Carlo rollouts (game.SimulateBattle) against sampled
+//     guesses of the opponent's deck and ordering;
+//   - future: a heuristic value of the resulting deck (power, tiers, suit
+//     synergy toward Triples) that only pays off in later rounds.
+//
+// The blend shifts toward "immediate" as the game nears its end and when the
+// bot trails on trophies, which is what lets it deliberately take a weak
+// round early to set up a stronger one later.
+//
+// Difficulty is a single level in [0, 1]: it sets the softmax temperature
+// used to choose among scored candidates (a perfect bot always takes the top
+// move; an easy bot often takes merely decent ones) and scales the rollout
+// budget (an easy bot estimates win chances more noisily).
+package ai
+
+import (
+	"math"
+	"math/rand/v2"
+
+	"github.com/greyson/super-auto-pets-board-game/internal/game"
+)
+
+// Action is one move the bot wants to make, mirroring the client protocol.
+type Action struct {
+	Type  string   // "buy" | "sell" | "trade" | "tradeChoose" | "pass" | "arrange" | "ready"
+	Row   int      // buy
+	Cards []string // sell / trade
+	Pick  int      // tradeChoose
+	Order []string // arrange
+}
+
+// Bot is a computer player at a fixed difficulty level.
+type Bot struct {
+	level float64
+}
+
+// New creates a bot with the given skill level in [0, 1].
+func New(level float64) *Bot {
+	return &Bot{level: min(max(level, 0), 1)}
+}
+
+// Act computes the bot's next move from its view of the game, or nil when no
+// input is owed. It does not modify the memory.
+func (b *Bot) Act(v *game.View, mem *Memory) *Action {
+	if v.YouSeat < 0 || v.YouSeat >= len(v.Players) {
+		return nil
+	}
+	me := &v.Players[v.YouSeat]
+	switch v.Phase {
+	case game.PhaseShop:
+		if v.Pending != nil {
+			if v.Pending.PlayerID == me.ID {
+				return b.decideTradeChoose(v, mem)
+			}
+			return nil
+		}
+		if v.Turn == v.YouSeat && me.Coins > 0 {
+			return b.decideShop(v, mem)
+		}
+	case game.PhaseCleanup:
+		if !me.Ready {
+			return b.decideCleanup(v, mem)
+		}
+	case game.PhaseArrange:
+		if !me.Ready {
+			return b.decideArrange(v, mem)
+		}
+	case game.PhaseBattle:
+		if !me.Ready {
+			return &Action{Type: "ready"}
+		}
+	}
+	return nil
+}
+
+// Pending reports whether the seat owes the game an action right now — the
+// server uses it to decide when to schedule a bot move.
+func Pending(v *game.View) bool {
+	if v.YouSeat < 0 || v.YouSeat >= len(v.Players) {
+		return false
+	}
+	me := &v.Players[v.YouSeat]
+	switch v.Phase {
+	case game.PhaseShop:
+		if v.Pending != nil {
+			return v.Pending.PlayerID == me.ID
+		}
+		return v.Turn == v.YouSeat && me.Coins > 0
+	case game.PhaseCleanup, game.PhaseArrange, game.PhaseBattle:
+		return !me.Ready
+	}
+	return false
+}
+
+// candidate is one scored move option. Most candidates map to a single
+// hypothetical deck; a trade maps to several (one per sampled reward card)
+// whose scores are averaged.
+type candidate struct {
+	act   *Action
+	decks [][]game.Card
+	bias  float64 // small nudge applied on top of the evaluated score
+	score float64
+}
+
+// pick chooses among candidates with a softmax over their scores. The
+// difficulty level sets the temperature: near 0 the bot always takes the
+// best move; higher temperatures make it increasingly willing to take
+// second-best (or worse) options.
+func (b *Bot) pick(cands []candidate) candidate {
+	if len(cands) == 1 {
+		return cands[0]
+	}
+	temp := 0.02 + 0.30*(1-b.level)
+	best := math.Inf(-1)
+	for _, c := range cands {
+		best = max(best, c.score)
+	}
+	weights := make([]float64, len(cands))
+	total := 0.0
+	for i, c := range cands {
+		weights[i] = math.Exp((c.score - best) / temp)
+		total += weights[i]
+	}
+	r := rand.Float64() * total
+	for i, w := range weights {
+		r -= w
+		if r <= 0 {
+			return cands[i]
+		}
+	}
+	return cands[len(cands)-1]
+}
+
+// budget returns the rollout counts for this difficulty: how many opponent
+// deck/order guesses to test against, and how many dice-randomized battle
+// simulations to run per guess. Fewer samples means noisier estimates, which
+// is itself part of what makes an easy bot easy.
+func (b *Bot) budget() (oppSamples, simsPer int) {
+	oppSamples = 6 + int(b.level*8) // 6 .. 14
+	simsPer = 1 + int(b.level*2)    // 1 .. 3
+	return
+}
+
+// immediateWeight is how much of a move's score comes from the next battle
+// versus long-term deck value. Later rounds shift weight toward "win now"
+// (round 6 is worth double and there is no later); trailing on trophies
+// pushes the same way, while a comfortable lead frees the bot to invest.
+func immediateWeight(v *game.View) float64 {
+	w := 0.40
+	if v.MaxRounds > 1 {
+		w += 0.60 * float64(v.Round-1) / float64(v.MaxRounds-1)
+	}
+	me := v.Players[v.YouSeat]
+	for _, p := range v.Players {
+		if p.Seat != v.YouSeat {
+			w += 0.08 * float64(p.Trophies-me.Trophies)
+		}
+	}
+	return min(max(w, 0.25), 1)
+}

internal/ai/ai_test.go

@@ -0,0 +1,204 @@
+package ai
+
+import (
+	"testing"
+
+	"github.com/greyson/super-auto-pets-board-game/internal/game"
+)
+
+// playBotGame drives a full game with bots in both seats, the same way the
+// server would: observe on every state change, then act when input is owed.
+// It fails the test if a bot ever produces an illegal action or the game
+// stops making progress.
+func playBotGame(t *testing.T, levelA, levelB float64) *game.Game {
+	t.Helper()
+	g := game.New()
+	pa, err := g.AddBot("Bot A", levelA)
+	if err != nil {
+		t.Fatalf("AddBot A: %v", err)
+	}
+	pb, err := g.AddBot("Bot B", levelB)
+	if err != nil {
+		t.Fatalf("AddBot B: %v", err)
+	}
+	bots := map[string]*Bot{pa.ID: New(levelA), pb.ID: New(levelB)}
+	mems := map[string]*Memory{pa.ID: {}, pb.ID: {}}
+
+	observe := func() {
+		for _, p := range g.Players {
+			v := g.ViewFor(p.ID)
+			Observe(&v, mems[p.ID])
+		}
+	}
+	observe()
+
+	for steps := 0; g.Phase != game.PhaseGameOver; steps++ {
+		if steps > 2000 {
+			t.Fatalf("game made no progress; stuck in phase %s round %d", g.Phase, g.Round)
+		}
+		acted := false
+		for _, p := range g.Players {
+			v := g.ViewFor(p.ID)
+			if !Pending(&v) {
+				continue
+			}
+			act := bots[p.ID].Act(&v, mems[p.ID])
+			if act == nil {
+				t.Fatalf("bot %s owes an action in phase %s but returned none", p.Name, g.Phase)
+			}
+			if err := applyAction(g, p.ID, act); err != nil {
+				t.Fatalf("bot %s illegal action %q in phase %s round %d: %v", p.Name, act.Type, g.Phase, g.Round, err)
+			}
+			observe()
+			acted = true
+			break // one action per iteration, like one message per broadcast
+		}
+		if !acted {
+			t.Fatalf("no bot owes an action but the game is not over (phase %s)", g.Phase)
+		}
+	}
+	return g
+}
+
+// applyAction mirrors the server's dispatch of bot actions onto the engine.
+func applyAction(g *game.Game, playerID string, a *Action) error {
+	switch a.Type {
+	case "buy":
+		return g.Buy(playerID, a.Row)
+	case "sell":
+		if g.Phase == game.PhaseCleanup {
+			return g.CleanupSell(playerID, a.Cards)
+		}
+		return g.Sell(playerID, a.Cards)
+	case "trade":
+		return g.TradeStart(playerID, a.Cards)
+	case "tradeChoose":
+		return g.TradeChoose(playerID, a.Pick)
+	case "pass":
+		return g.Pass(playerID)
+	case "arrange":
+		return g.SubmitOrder(playerID, a.Order)
+	case "ready":
+		return g.AcknowledgeBattle(playerID)
+	}
+	return game.ErrInvalidAction
+}
+
+// TestBotsFinishGames plays complete games at each difficulty pairing. This
+// is the main safety net: every phase, every action type, every round, with
+// two independent AIs generating whatever situations they generate.
+func TestBotsFinishGames(t *testing.T) {
+	for _, levels := range [][2]float64{{1, 1}, {0.25, 1}, {0, 0}, {0.6, 0.25}} {
+		for range 3 {
+			g := playBotGame(t, levels[0], levels[1])
+			if g.Round != game.MaxRounds {
+				t.Errorf("game ended on round %d, want %d", g.Round, game.MaxRounds)
+			}
+		}
+	}
+}
+
+// TestObserveTracksOpponentDeck checks the memory's opponent model against
+// the opponent's real deck after known public actions. The model may only
+// contain information a human spectator would have.
+func TestObserveTracksOpponentDeck(t *testing.T) {
+	g := game.New()
+	pa, _ := g.AddBot("Bot A", 1)
+	pb, _ := g.AddBot("Bot B", 1)
+	mem := &Memory{}
+	obs := func() {
+		v := g.ViewFor(pa.ID)
+		Observe(&v, mem)
+	}
+	obs()
+
+	// Whoever holds priority shops first; walk both players through buys.
+	first, second := g.Players[g.PrioritySeat], g.Players[1-g.PrioritySeat]
+	for range 3 { // 3 coins each, alternating
+		for _, p := range []*game.Player{first, second} {
+			if err := g.Buy(p.ID, 0); err != nil {
+				t.Fatalf("buy: %v", err)
+			}
+			obs()
+		}
+	}
+
+	// The model of B's deck must now match B's real deck card-for-card:
+	// every buy was public (and buy effects like Otter's apple are printed
+	// on the card).
+	assertModelMatches(t, mem, pb)
+
+	// Play out the round; the battle lineup resync must also match.
+	for g.Phase == game.PhaseCleanup {
+		t.Fatal("unexpected cleanup with 3 buys")
+	}
+	for _, p := range g.Players {
+		ids := make([]string, len(p.Deck))
+		for i, c := range p.Deck {
+			ids[i] = c.ID
+		}
+		if err := g.SubmitOrder(p.ID, ids); err != nil {
+			t.Fatalf("submit: %v", err)
+		}
+		obs()
+	}
+	if g.Phase != game.PhaseBattle {
+		t.Fatalf("phase = %s, want battle", g.Phase)
+	}
+	obs()
+	for _, p := range g.Players {
+		if err := g.AcknowledgeBattle(p.ID); err != nil {
+			t.Fatalf("ack: %v", err)
+		}
+		obs()
+	}
+	// Round 2 shop: temporaries expired; model must match B's real deck.
+	assertModelMatches(t, mem, pb)
+}
+
+// assertModelMatches requires the opponent model to agree with the real deck
+// as a multiset of card names (IDs can legitimately differ for cards the bot
+// reconstructed from public information).
+func assertModelMatches(t *testing.T, mem *Memory, opp *game.Player) {
+	t.Helper()
+	want := map[string]int{}
+	for _, c := range opp.Deck {
+		want[c.Name]++
+	}
+	got := map[string]int{}
+	for _, c := range mem.Opp.Known {
+		got[c.Name]++
+	}
+	if len(mem.Opp.Hidden) != 0 {
+		t.Errorf("model has %d hidden cards, want 0 (everything was public)", len(mem.Opp.Hidden))
+	}
+	for name, n := range want {
+		if got[name] != n {
+			t.Errorf("model has %d × %s, real deck has %d", got[name], name, n)
+		}
+	}
+	for name, n := range got {
+		if want[name] == 0 {
+			t.Errorf("model claims %d × %s that the real deck lacks", n, name)
+		}
+	}
+}
+
+// TestSimulateBattleIsPure verifies rollouts don't corrupt anything the
+// caller hands in.
+func TestSimulateBattleIsPure(t *testing.T) {
+	deckA := []game.Card{
+		{ID: "a1", Kind: game.KindPet, Name: "Ant", Power: 1,
+			Effects: []game.Effect{{Trigger: game.TriggerFaint, Action: game.ActionSummonTop, Card: "apple"}}},
+	}
+	deckB := []game.Card{
+		{ID: "b1", Kind: game.KindPet, Name: "Duck", Power: 2},
+	}
+	res := game.SimulateBattle(1, 0, deckA, deckB, nil)
+	if res == nil || res.WinnerSeat != 1 {
+		t.Fatalf("expected seat 1 (Duck) to win, got %+v", res)
+	}
+	if len(deckA) != 1 || len(deckB) != 1 || deckA[0].ID != "a1" || deckB[0].ID != "b1" {
+		t.Error("SimulateBattle mutated its input decks")
+	}
+}

internal/ai/arrange.go

@@ -0,0 +1,195 @@
+package ai
+
+import (
+	"slices"
+	"strings"
+
+	"github.com/greyson/super-auto-pets-board-game/internal/game"
+)
+
+// decideArrange searches for the best secret battle ordering of the bot's
+// deck. This is where "guess what the opponent will do" matters most: every
+// candidate ordering is judged by simulated battles against a spread of
+// sampled opponent decks and orderings, never against the opponent's real
+// (hidden) choice.
+//
+// The search runs in two stages to stay cheap:
+//  1. every permutation of the pets (≤ 5! = 120), each with a default food
+//     placement, gets a quick screening score;
+//  2. the best few permutations are re-scored precisely, each trying several
+//     food-placement variants (front-loaded, on the strongest pet, spread,
+//     on an apple-synergy pet like Rooster or Leopard).
+func (b *Bot) decideArrange(v *game.View, mem *Memory) *Action {
+	cx := newCtx(v, mem)
+	deck := cx.me.Deck
+
+	var pets, foods []game.Card
+	for _, c := range deck {
+		if c.IsPet() {
+			pets = append(pets, c)
+		} else {
+			foods = append(foods, c)
+		}
+	}
+	if len(pets) == 0 {
+		// Nothing can fight; any order loses identically.
+		return &Action{Type: "arrange", Order: cardIDs(deck)}
+	}
+
+	oppSamples, simsPer := b.budget()
+	oppDecks := cx.oppArrangements(oppSamples)
+
+	// Stage 1: screen every pet permutation with the default food placement
+	// against a subset of the opponent guesses.
+	perms := permutations(len(pets), 200)
+	screen := oppDecks[:min(4+int(b.level*4), len(oppDecks))]
+	type scored struct {
+		perm  []int
+		score float64
+	}
+	ranked := make([]scored, 0, len(perms))
+	for _, perm := range perms {
+		arr := buildArrangement(pets, perm, foods, placeFront)
+		ranked = append(ranked, scored{perm, cx.winProb(arr, screen, 1)})
+	}
+	slices.SortStableFunc(ranked, func(a, b scored) int {
+		switch {
+		case a.score > b.score:
+			return -1
+		case a.score < b.score:
+			return 1
+		}
+		return 0
+	})
+
+	// Stage 2: refine the leaders with every food-placement variant and the
+	// full opponent sample set.
+	var cands []candidate
+	seen := map[string]bool{}
+	for _, r := range ranked[:min(5, len(ranked))] {
+		for _, place := range []foodPlacement{placeFront, placeStrongest, placeSpread, placeSynergy} {
+			arr := buildArrangement(pets, r.perm, foods, place)
+			key := fingerprint(arr)
+			if seen[key] {
+				continue
+			}
+			seen[key] = true
+			cands = append(cands, candidate{
+				act:   &Action{Type: "arrange", Order: cardIDs(arr)},
+				score: cx.winProb(arr, oppDecks, simsPer),
+			})
+		}
+	}
+	return b.pick(cands).act
+}
+
+// foodPlacement decides which pet slot (index into the pet order) each food
+// card sits in front of.
+type foodPlacement func(pets []game.Card, foodIdx int, food game.Card) int
+
+// placeFront stacks everything on the leading pet: it fights the most
+// clashes, so buffs there see the most use.
+func placeFront([]game.Card, int, game.Card) int { return 0 }
+
+// placeStrongest feeds the biggest pet — apples on a heavy hitter compound,
+// and perks protect the pet that fights longest.
+func placeStrongest(pets []game.Card, _ int, _ game.Card) int {
+	best := 0
+	for i, p := range pets {
+		if p.Power > pets[best].Power {
+			best = i
+		}
+	}
+	return best
+}
+
+// placeSpread deals foods round-robin so one Skunk or Wolverine can't strip
+// the whole stockpile at once.
+func placeSpread(pets []game.Card, foodIdx int, _ game.Card) int {
+	return foodIdx % len(pets)
+}
+
+// placeSynergy targets pets whose abilities key off attached apples
+// (Rooster's bees, Dodo's recycling, Leopard's per-power rocks, Peacock and
+// Scorpion wanting to survive); falls back to the strongest pet.
+func placeSynergy(pets []game.Card, foodIdx int, food game.Card) int {
+	for i, p := range pets {
+		switch p.Name {
+		case "Rooster", "Dodo", "Leopard", "Peacock", "Scorpion":
+			return i
+		}
+	}
+	return placeStrongest(pets, foodIdx, food)
+}
+
+// buildArrangement lays out the deck: foods assigned to a pet slot appear
+// directly above that pet, and no food ever trails uselessly at the bottom.
+// Perks assigned to the same pet keep only the last one applied, so extras
+// are pushed to later pets.
+func buildArrangement(pets []game.Card, perm []int, foods []game.Card, place foodPlacement) []game.Card {
+	ordered := make([]game.Card, len(perm))
+	for i, pi := range perm {
+		ordered[i] = pets[pi]
+	}
+	assign := make([][]game.Card, len(ordered))
+	perkUsed := make([]bool, len(ordered))
+	for fi, f := range foods {
+		at := place(ordered, fi, f)
+		if f.Perk {
+			// Slide duplicate perks onto the next unperked pet.
+			for at < len(ordered) && perkUsed[at] {
+				at++
+			}
+			if at >= len(ordered) {
+				at = len(ordered) - 1
+			}
+			perkUsed[at] = true
+		}
+		assign[at] = append(assign[at], f)
+	}
+	out := make([]game.Card, 0, len(pets)+len(foods))
+	for i, p := range ordered {
+		out = append(out, assign[i]...)
+		out = append(out, p)
+	}
+	return out
+}
+
+// permutations enumerates permutations of n indices, up to limit (5 pets is
+// 120, so the limit only guards hypothetical future rule changes).
+func permutations(n, limit int) [][]int {
+	idx := make([]int, n)
+	for i := range idx {
+		idx[i] = i
+	}
+	var out [][]int
+	var rec func(k int)
+	rec = func(k int) {
+		if len(out) >= limit {
+			return
+		}
+		if k == n {
+			out = append(out, slices.Clone(idx))
+			return
+		}
+		for i := k; i < n; i++ {
+			idx[k], idx[i] = idx[i], idx[k]
+			rec(k + 1)
+			idx[k], idx[i] = idx[i], idx[k]
+		}
+	}
+	rec(0)
+	return out
+}
+
+func cardIDs(cards []game.Card) []string {
+	ids := make([]string, len(cards))
+	for i, c := range cards {
+		ids[i] = c.ID
+	}
+	return ids
+}
+
+func fingerprint(cards []game.Card) string {
+	return strings.Join(cardIDs(cards), "|")
+}

internal/ai/eval.go

@@ -0,0 +1,201 @@
+package ai
+
+import (
+	"math"
+	"math/rand/v2"
+	"slices"
+
+	"github.com/greyson/super-auto-pets-board-game/internal/game"
+)
+
+// winScore converts a simulated battle outcome to a utility for mySeat:
+// win 1, draw 0.5 (nobody gains ground), loss 0.
+func winScore(res *game.BattleResult, mySeat int) float64 {
+	switch res.WinnerSeat {
+	case mySeat:
+		return 1
+	case -1:
+		return 0.5
+	default:
+		return 0
+	}
+}
+
+// winProb estimates the chance the arranged deck wins the upcoming battle by
+// simulating it against every sampled opponent arrangement, simsPer times
+// each (dice rerolled every time). All candidates in one decision share the
+// same opponent samples, so comparisons between them are paired and fair.
+func (cx *ctx) winProb(myDeck []game.Card, oppDecks [][]game.Card, simsPer int) float64 {
+	if len(oppDecks) == 0 || simsPer <= 0 {
+		return 0.5
+	}
+	total, n := 0.0, 0
+	for _, opp := range oppDecks {
+		for range simsPer {
+			var res *game.BattleResult
+			if cx.me.Seat == 0 {
+				res = game.SimulateBattle(cx.v.Round, cx.v.PrioritySeat, myDeck, opp, nil)
+			} else {
+				res = game.SimulateBattle(cx.v.Round, cx.v.PrioritySeat, opp, myDeck, nil)
+			}
+			total += winScore(res, cx.me.Seat)
+			n++
+		}
+	}
+	return total / float64(n)
+}
+
+// keepValue ranks a single card's worth to the bot's future: what it loses
+// by selling or trading it away. Temporary cards (apples) are nearly free to
+// lose — they vanish after the next battle anyway.
+func keepValue(c game.Card) float64 {
+	if c.Temporary {
+		return 0.15
+	}
+	if c.IsFood() {
+		if c.Perk {
+			return 1.6
+		}
+		return 0.3
+	}
+	val := float64(c.Power)*0.55 + float64(c.Tier)*0.8
+	if len(c.Effects) > 0 {
+		val += 0.6
+	}
+	return val
+}
+
+// deckValue is the future-facing worth of a deck: card quality plus suit
+// synergy (pairs and triples enable the Triple trade-in, the only path to
+// higher-tier cards than the current round offers). Temporary cards count
+// for nothing here — their value shows up in the battle rollouts instead.
+func deckValue(deck []game.Card, round, maxRounds int) float64 {
+	total := 0.0
+	suits := map[game.Suit]int{}
+	for _, c := range deck {
+		if c.Temporary {
+			continue
+		}
+		total += keepValue(c)
+		if c.IsPet() {
+			suits[c.Suit]++
+		}
+	}
+	if round < maxRounds {
+		for _, k := range suits {
+			switch {
+			case k >= 3:
+				total += 1.5
+			case k == 2:
+				total += 0.6
+			}
+		}
+	}
+	return total
+}
+
+// normFuture squashes an unbounded deck value into (0, 1) so it can be
+// blended with a win probability.
+func normFuture(val float64) float64 {
+	return val / (val + 15)
+}
+
+// leadScore ranks pets for early battle positions: raw power fights longest,
+// faint effects want to actually faint (early), play effects fire on entry
+// wherever they are but are worth protecting slightly less.
+func leadScore(c game.Card) float64 {
+	s := float64(c.Power)
+	for _, e := range c.Effects {
+		switch e.Trigger {
+		case game.TriggerFaint:
+			s += 1.5
+		case game.TriggerPlay:
+			s += 0.8
+		case game.TriggerHurt:
+			s += 0.5
+		}
+	}
+	return s
+}
+
+// heuristicOrder arranges a deck the way a reasonable player might: pets
+// sorted by leadScore, apples front-loaded, perks spread across the
+// strongest pets, never a food trailing at the bottom. temp adds Gumbel
+// noise to every placement — 0 gives the deterministic "book" order, higher
+// values give increasingly scrambled-but-plausible alternatives (used to
+// model the range of orders an opponent might pick).
+func heuristicOrder(deck []game.Card, temp float64) []game.Card {
+	var pets, apples, perks, otherFood []game.Card
+	for _, c := range deck {
+		switch {
+		case c.IsPet():
+			pets = append(pets, c)
+		case c.Perk:
+			perks = append(perks, c)
+		case c.Food == game.FoodApple:
+			apples = append(apples, c)
+		default:
+			otherFood = append(otherFood, c)
+		}
+	}
+	noisy := func(base float64) float64 {
+		if temp <= 0 {
+			return base
+		}
+		// Gumbel-perturbed scores turn a sort into a plausibility-weighted
+		// random ranking.
+		return base - temp*math.Log(-math.Log(rand.Float64()))
+	}
+	slices.SortStableFunc(pets, func(a, b game.Card) int {
+		av, bv := noisy(leadScore(a)), noisy(leadScore(b))
+		switch {
+		case av > bv:
+			return -1
+		case av < bv:
+			return 1
+		}
+		return 0
+	})
+	if len(pets) == 0 {
+		// No pets means an immediate loss; foods are wasted regardless.
+		return append(append(append(apples, perks...), otherFood...), pets...)
+	}
+	// Assign foods to pet indices, then interleave.
+	assign := make([][]game.Card, len(pets))
+	strongest := 0
+	for i, p := range pets {
+		if p.Power > pets[strongest].Power {
+			strongest = i
+		}
+	}
+	for _, a := range apples {
+		at := 0
+		if temp > 0 && rand.Float64() < 0.4 {
+			at = rand.IntN(len(pets))
+		}
+		assign[at] = append(assign[at], a)
+	}
+	// Perks one per pet, best pets first (a pet only keeps its last perk).
+	perkOrder := []int{strongest}
+	for i := range pets {
+		if i != strongest {
+			perkOrder = append(perkOrder, i)
+		}
+	}
+	for i, p := range perks {
+		at := perkOrder[min(i, len(perkOrder)-1)]
+		if temp > 0 && rand.Float64() < 0.3 {
+			at = rand.IntN(len(pets))
+		}
+		assign[at] = append(assign[at], p)
+	}
+	for i, f := range otherFood {
+		assign[i%len(pets)] = append(assign[i%len(pets)], f)
+	}
+	out := make([]game.Card, 0, len(deck))
+	for i, p := range pets {
+		out = append(out, assign[i]...)
+		out = append(out, p)
+	}
+	return out
+}

internal/ai/memory.go

@@ -0,0 +1,209 @@
+package ai
+
+import (
+	"encoding/json"
+	"fmt"
+	"slices"
+
+	"github.com/greyson/super-auto-pets-board-game/internal/game"
+)
+
+// Memory is the bot's private notebook: everything it has legitimately
+// learned from public information, carried between turns (and, serialized
+// into the game state, across server restarts). It is the bot's substitute
+// for a human player's attention — nothing in here is unavailable to a human
+// watching the same screen.
+type Memory struct {
+	LastSeq         int         `json:"lastSeq"`         // last event-log entry processed
+	LastBattleRound int         `json:"lastBattleRound"` // last battle lineup ingested
+	PrevShopRow     []game.Card `json:"prevShopRow"`     // shop row at the previous observation
+	Opp             OppModel    `json:"opp"`
+}
+
+// OppModel is the bot's belief about one opponent's deck. Known holds cards
+// it has actually seen there (battle lineups reveal entire decks each round;
+// shop buys are public); Hidden counts cards it knows exist but has never
+// seen — trade-in picks, whose tier is public but whose identity is not.
+type OppModel struct {
+	Seat   int          `json:"seat"`
+	Known  []game.Card  `json:"known"`
+	Hidden []HiddenCard `json:"hidden,omitempty"`
+}
+
+// HiddenCard is a card the opponent holds that the bot has not seen. Name is
+// set when the card was later named publicly (e.g. a trade pick revealed by
+// its buy ability) — the suit still isn't known, but the stats are.
+type HiddenCard struct {
+	Tier int    `json:"tier"`
+	Name string `json:"name,omitempty"`
+}
+
+// LoadMemory decodes a bot's stored memory; a nil or corrupt blob yields a
+// fresh one (the model self-heals from the next battle lineup anyway).
+func LoadMemory(raw json.RawMessage) *Memory {
+	m := &Memory{}
+	if len(raw) > 0 {
+		_ = json.Unmarshal(raw, m)
+	}
+	return m
+}
+
+// Marshal encodes the memory for storage on the bot's Player.
+func (m *Memory) Marshal() json.RawMessage {
+	raw, err := json.Marshal(m)
+	if err != nil {
+		return nil
+	}
+	return raw
+}
+
+// Observe updates the memory from the bot's latest view. The server calls
+// this on every state change, so consecutive observations are one action
+// apart. It reads three public sources, in order:
+//
+//  1. new event-log entries, whose structured tags describe opponent shop
+//     actions (buys name the card, sells name what left, trades list the
+//     discarded trio, spawn entries count apples gained);
+//  2. the latest battle's lineups, which reveal both decks in full and reset
+//     the model to ground truth every round (so any drift lasts one round);
+//  3. the opponent's public deck size, as a reconciliation safety net.
+func Observe(v *game.View, m *Memory) {
+	if v.YouSeat < 0 {
+		return
+	}
+	oppSeat := -1
+	for _, p := range v.Players {
+		if p.Seat != v.YouSeat {
+			oppSeat = p.Seat
+			break
+		}
+	}
+	if oppSeat < 0 {
+		return
+	}
+	m.Opp.Seat = oppSeat
+
+	for _, e := range v.Log {
+		if e.Seq <= m.LastSeq {
+			continue
+		}
+		m.LastSeq = e.Seq
+		if e.Seat != oppSeat {
+			continue
+		}
+		switch {
+		case e.Kind == game.LogBuy:
+			if c, ok := cardByID(m.PrevShopRow, e.Source); ok {
+				m.Opp.Known = append(m.Opp.Known, c)
+			} else if c, ok := templateByName(e.CardName); ok {
+				m.Opp.Known = append(m.Opp.Known, c)
+			}
+		case e.Kind == game.LogSell:
+			m.removeOppCard(e.Source, e.CardName)
+			m.Opp.Known = append(m.Opp.Known, memApple(len(m.Opp.Known)))
+		case e.Kind == game.LogTrade:
+			for _, id := range e.Cards {
+				m.removeOppCard(id, "")
+			}
+		case e.Kind == game.LogTradePick:
+			m.Opp.Hidden = append(m.Opp.Hidden,
+				HiddenCard{Tier: min(e.Round+1, game.MaxRounds), Name: e.CardName})
+		case e.Spawn == "apple" && e.Kind == "":
+			n := max(e.Count, 1)
+			for range n {
+				m.Opp.Known = append(m.Opp.Known, memApple(len(m.Opp.Known)))
+			}
+		}
+	}
+
+	// Battle lineups are ground truth: rebuild the model from the opponent's
+	// revealed deck, minus temporary cards (they expire with the battle).
+	if v.Battle != nil && v.Battle.Round > m.LastBattleRound && oppSeat < len(v.Battle.Lineups) {
+		m.LastBattleRound = v.Battle.Round
+		m.Opp.Known = m.Opp.Known[:0]
+		m.Opp.Hidden = nil
+		for _, c := range v.Battle.Lineups[oppSeat] {
+			if !c.Temporary {
+				m.Opp.Known = append(m.Opp.Known, c)
+			}
+		}
+	}
+
+	// Reconcile with the public deck size. Skipped during the battle phase,
+	// where the live deck still holds temporaries the model excludes.
+	if v.Phase == game.PhaseShop || v.Phase == game.PhaseCleanup || v.Phase == game.PhaseArrange {
+		size := v.Players[slices.IndexFunc(v.Players, func(p game.PlayerView) bool { return p.Seat == oppSeat })].DeckSize
+		for len(m.Opp.Known)+len(m.Opp.Hidden) < size {
+			m.Opp.Hidden = append(m.Opp.Hidden, HiddenCard{Tier: v.Round})
+		}
+		for len(m.Opp.Known)+len(m.Opp.Hidden) > size {
+			if len(m.Opp.Hidden) > 0 {
+				m.Opp.Hidden = m.Opp.Hidden[:len(m.Opp.Hidden)-1]
+			} else {
+				m.Opp.Known = m.Opp.Known[:len(m.Opp.Known)-1]
+			}
+		}
+	}
+
+	m.PrevShopRow = append(m.PrevShopRow[:0], v.ShopRow...)
+}
+
+// removeOppCard drops one card from the model: by exact ID when we tracked
+// it, by name as a fallback (model-minted apples have synthetic IDs), and
+// failing both, one hidden card — something we didn't know they had left.
+func (m *Memory) removeOppCard(id, name string) {
+	if i := slices.IndexFunc(m.Opp.Known, func(c game.Card) bool { return c.ID == id }); i >= 0 {
+		m.Opp.Known = slices.Delete(m.Opp.Known, i, i+1)
+		return
+	}
+	if name != "" {
+		if i := slices.IndexFunc(m.Opp.Known, func(c game.Card) bool { return c.Name == name }); i >= 0 {
+			m.Opp.Known = slices.Delete(m.Opp.Known, i, i+1)
+			return
+		}
+	}
+	if len(m.Opp.Hidden) > 0 {
+		m.Opp.Hidden = m.Opp.Hidden[:len(m.Opp.Hidden)-1]
+	}
+}
+
+func cardByID(cards []game.Card, id string) (game.Card, bool) {
+	if id == "" {
+		return game.Card{}, false
+	}
+	for _, c := range cards {
+		if c.ID == id {
+			return c, true
+		}
+	}
+	return game.Card{}, false
+}
+
+// templateByName mints a reference copy of a named card from the printed
+// tier contents. The suit is whatever the first printed copy has — callers
+// only rely on stats and effects.
+func templateByName(name string) (game.Card, bool) {
+	if name == "" {
+		return game.Card{}, false
+	}
+	for tier := 1; tier <= game.MaxRounds; tier++ {
+		for _, c := range game.TierContents(tier) {
+			if c.Name == name {
+				return c, true
+			}
+		}
+	}
+	return game.Card{}, false
+}
+
+// memApple mints an apple for the opponent model. The ID is synthetic — it
+// only needs to not collide with real card IDs.
+func memApple(n int) game.Card {
+	return game.Card{
+		ID:        fmt.Sprintf("mem-apple-%d", n),
+		Kind:      game.KindFood,
+		Name:      "Apple",
+		Food:      game.FoodApple,
+		Temporary: true,
+	}
+}

internal/ai/predict.go

@@ -0,0 +1,109 @@
+package ai
+
+import (
+	"fmt"
+	"math/rand/v2"
+
+	"github.com/greyson/super-auto-pets-board-game/internal/game"
+)
+
+// ctx is per-decision scratch state: the view, the memory, and caches shared
+// by every candidate evaluated in the same decision.
+type ctx struct {
+	v       *game.View
+	m       *Memory
+	me      *game.PlayerView
+	oppSeat int
+	pools   map[int][]game.Card // unseen cards per tier, for hidden-card guesses
+	simID   int
+}
+
+func newCtx(v *game.View, m *Memory) *ctx {
+	cx := &ctx{v: v, m: m, me: &v.Players[v.YouSeat], oppSeat: m.Opp.Seat, pools: map[int][]game.Card{}}
+	if cx.oppSeat == cx.me.Seat || cx.oppSeat < 0 {
+		// Memory hasn't observed yet (shouldn't happen in practice).
+		for _, p := range v.Players {
+			if p.Seat != v.YouSeat {
+				cx.oppSeat = p.Seat
+			}
+		}
+	}
+	return cx
+}
+
+func (cx *ctx) nextSimID() string {
+	cx.simID++
+	return fmt.Sprintf("sim-%d", cx.simID)
+}
+
+// unseenPool lists the printed cards of a tier that the bot cannot account
+// for anywhere it can see — its own deck, the opponent model, the shop row.
+// Hidden opponent cards are drawn from this pool, so the bot's guesses
+// respect card counting without peeking at the real decks.
+func (cx *ctx) unseenPool(tier int) []game.Card {
+	if pool, ok := cx.pools[tier]; ok {
+		return pool
+	}
+	seen := map[string]int{}
+	note := func(c game.Card) {
+		if c.Tier == tier {
+			seen[c.Name]++
+		}
+	}
+	for _, c := range cx.me.Deck {
+		note(c)
+	}
+	for _, c := range cx.m.Opp.Known {
+		note(c)
+	}
+	for _, c := range cx.v.ShopRow {
+		note(c)
+	}
+	var pool []game.Card
+	for _, c := range game.TierContents(tier) {
+		if seen[c.Name] > 0 {
+			seen[c.Name]--
+			continue
+		}
+		pool = append(pool, c)
+	}
+	cx.pools[tier] = pool
+	return pool
+}
+
+// sampleOppDeck instantiates one concrete guess at the opponent's deck:
+// known cards as-is, hidden cards drawn from the unseen pool of their tier
+// (or their named template, when a pick was later revealed).
+func (cx *ctx) sampleOppDeck() []game.Card {
+	deck := append([]game.Card(nil), cx.m.Opp.Known...)
+	for _, h := range cx.m.Opp.Hidden {
+		var c game.Card
+		if t, ok := templateByName(h.Name); ok {
+			c = t
+		} else if pool := cx.unseenPool(h.Tier); len(pool) > 0 {
+			c = pool[rand.IntN(len(pool))]
+		} else {
+			continue // tier exhausted and fully accounted for; nothing to guess
+		}
+		c.ID = cx.nextSimID()
+		deck = append(deck, c)
+	}
+	return deck
+}
+
+// oppArrangements produces n independent guesses of what the opponent will
+// field: each is a sampled deck put in a plausible order by the same
+// heuristic the bot itself uses, with enough noise that the bot prepares for
+// a range of opponent plans rather than assuming one. The first guess is the
+// noise-free "book" ordering.
+func (cx *ctx) oppArrangements(n int) [][]game.Card {
+	out := make([][]game.Card, 0, n)
+	for i := range n {
+		temp := 0.8
+		if i == 0 {
+			temp = 0
+		}
+		out = append(out, heuristicOrder(cx.sampleOppDeck(), temp))
+	}
+	return out
+}

internal/ai/shop.go

@@ -0,0 +1,260 @@
+package ai
+
+import (
+	"math/rand/v2"
+	"slices"
+
+	"github.com/greyson/super-auto-pets-board-game/internal/game"
+)
+
+// applyTemplateShopEffects mirrors the engine's shop-time triggers on a
+// hypothetical deck: buying an Otter really does come with an apple, and the
+// bot should value that. Only deck-changing effects matter here (coin
+// refunds don't alter the deck being scored).
+func (cx *ctx) applyTemplateShopEffects(deck []game.Card, c game.Card, trigger game.EffectTrigger) []game.Card {
+	for _, e := range c.Effects {
+		if e.Trigger != trigger || cx.v.Round < e.MinRound {
+			continue
+		}
+		switch e.Action {
+		case game.ActionGainApple:
+			for range max(e.Count, 1) {
+				deck = append(deck, cx.simApple())
+			}
+		case game.ActionDoubleApples:
+			apples := 0
+			for _, dc := range deck {
+				if dc.Food == game.FoodApple {
+					apples++
+				}
+			}
+			for range apples {
+				deck = append(deck, cx.simApple())
+			}
+		}
+	}
+	return deck
+}
+
+func (cx *ctx) simApple() game.Card {
+	return game.Card{
+		ID:        cx.nextSimID(),
+		Kind:      game.KindFood,
+		Name:      "Apple",
+		Food:      game.FoodApple,
+		Temporary: true,
+	}
+}
+
+// previewCleanup applies the forced end-of-shop sale to a hypothetical deck:
+// while over the pet limit, the lowest-value pet is sold for an apple. This
+// lets the bot buy a sixth pet on purpose, knowing what it will cost.
+func (cx *ctx) previewCleanup(deck []game.Card) []game.Card {
+	for {
+		pets := 0
+		worst, worstVal := -1, 0.0
+		for i, c := range deck {
+			if !c.IsPet() {
+				continue
+			}
+			pets++
+			if v := keepValue(c); worst < 0 || v < worstVal {
+				worst, worstVal = i, v
+			}
+		}
+		if pets <= cx.v.MaxPets || worst < 0 {
+			return deck
+		}
+		sold := deck[worst]
+		deck = slices.Delete(deck, worst, worst+1)
+		deck = append(deck, cx.simApple())
+		deck = cx.applyTemplateShopEffects(deck, sold, game.TriggerSell)
+	}
+}
+
+// score fills in every candidate's score: a weighted blend of the estimated
+// next-battle win chance (deck arranged by the book ordering — the full
+// ordering search happens later, at arrange time) and the deck's future
+// value. All candidates face the same opponent guesses.
+func (b *Bot) score(cx *ctx, cands []candidate) {
+	oppSamples, simsPer := b.budget()
+	oppDecks := cx.oppArrangements(oppSamples)
+	alpha := immediateWeight(cx.v)
+	for i := range cands {
+		total := 0.0
+		for _, deck := range cands[i].decks {
+			imm := cx.winProb(heuristicOrder(deck, 0), oppDecks, simsPer)
+			fut := normFuture(deckValue(deck, cx.v.Round, cx.v.MaxRounds))
+			total += alpha*imm + (1-alpha)*fut
+		}
+		cands[i].score = total/float64(len(cands[i].decks)) + cands[i].bias
+	}
+}
+
+// decideShop picks one shop action: buy a row card, sell some own cards,
+// trade in a suit triple, or pass.
+func (b *Bot) decideShop(v *game.View, mem *Memory) *Action {
+	cx := newCtx(v, mem)
+	deck := cx.me.Deck
+	var cands []candidate
+
+	// Passing forfeits the bot's remaining coins; it is the baseline every
+	// other option must beat, with a nudge because spending is usually right.
+	cands = append(cands, candidate{
+		act:   &Action{Type: "pass"},
+		decks: [][]game.Card{slices.Clone(deck)},
+		bias:  -0.02,
+	})
+
+	for i, c := range v.ShopRow {
+		if c.ID == "" {
+			continue
+		}
+		nd := append(slices.Clone(deck), c)
+		nd = cx.applyTemplateShopEffects(nd, c, game.TriggerBuy)
+		nd = cx.previewCleanup(nd)
+		cands = append(cands, candidate{
+			act:   &Action{Type: "buy", Row: i},
+			decks: [][]game.Card{nd},
+		})
+	}
+
+	// Sell candidates: the worst 1, 2, or 3 keepers. One gold sells any
+	// number of cards, so bulk-dumping junk before a battle is one action.
+	// Temporary cards are excluded — selling an apple for an apple is a pure
+	// waste of gold.
+	sellable := slices.Clone(deck)
+	sellable = slices.DeleteFunc(sellable, func(c game.Card) bool { return c.Temporary })
+	slices.SortStableFunc(sellable, func(a, b game.Card) int {
+		av, bv := keepValue(a), keepValue(b)
+		switch {
+		case av < bv:
+			return -1
+		case av > bv:
+			return 1
+		}
+		return 0
+	})
+	for k := 1; k <= min(3, len(sellable)); k++ {
+		ids := make([]string, 0, k)
+		nd := slices.Clone(deck)
+		for _, s := range sellable[:k] {
+			ids = append(ids, s.ID)
+			idx := slices.IndexFunc(nd, func(c game.Card) bool { return c.ID == s.ID })
+			nd = slices.Delete(nd, idx, idx+1)
+			nd = append(nd, cx.simApple())
+			nd = cx.applyTemplateShopEffects(nd, s, game.TriggerSell)
+		}
+		cands = append(cands, candidate{
+			act:   &Action{Type: "sell", Cards: ids},
+			decks: [][]game.Card{nd},
+		})
+	}
+
+	// Trade candidates: for each suit with three or more pets, trade the
+	// three lowest-value ones. The reward card is unknown (top two of the
+	// next tier's deck), so each trade is scored across several sampled
+	// rewards.
+	if v.Round < v.MaxRounds && v.Round < len(v.DeckCounts) && v.DeckCounts[v.Round] >= 2 {
+		bySuit := map[game.Suit][]game.Card{}
+		for _, c := range deck {
+			if c.IsPet() && c.Suit != "" {
+				bySuit[c.Suit] = append(bySuit[c.Suit], c)
+			}
+		}
+		for _, pets := range bySuit {
+			if len(pets) < game.TradeInCount {
+				continue
+			}
+			slices.SortStableFunc(pets, func(a, b game.Card) int {
+				av, bv := keepValue(a), keepValue(b)
+				switch {
+				case av < bv:
+					return -1
+				case av > bv:
+					return 1
+				}
+				return 0
+			})
+			trio := pets[:game.TradeInCount]
+			base := slices.Clone(deck)
+			ids := make([]string, 0, game.TradeInCount)
+			for _, t := range trio {
+				ids = append(ids, t.ID)
+				idx := slices.IndexFunc(base, func(c game.Card) bool { return c.ID == t.ID })
+				base = slices.Delete(base, idx, idx+1)
+				base = cx.applyTemplateShopEffects(base, t, game.TriggerTriple)
+			}
+			pool := cx.unseenPool(v.Round + 1)
+			if len(pool) == 0 {
+				pool = game.TierContents(v.Round + 1)
+			}
+			var decks [][]game.Card
+			for range 3 {
+				reward := pool[rand.IntN(len(pool))]
+				reward.ID = cx.nextSimID()
+				nd := append(slices.Clone(base), reward)
+				nd = cx.applyTemplateShopEffects(nd, reward, game.TriggerBuy)
+				nd = cx.previewCleanup(nd)
+				decks = append(decks, nd)
+			}
+			cands = append(cands, candidate{
+				act:   &Action{Type: "trade", Cards: ids},
+				decks: decks,
+			})
+		}
+	}
+
+	b.score(cx, cands)
+	return b.pick(cands).act
+}
+
+// decideTradeChoose resolves the bot's own pending trade: score keeping
+// either revealed card and pick.
+func (b *Bot) decideTradeChoose(v *game.View, mem *Memory) *Action {
+	cx := newCtx(v, mem)
+	var cands []candidate
+	for pick, c := range v.Pending.Options {
+		nd := append(slices.Clone(cx.me.Deck), c)
+		nd = cx.applyTemplateShopEffects(nd, c, game.TriggerBuy)
+		nd = cx.previewCleanup(nd)
+		cands = append(cands, candidate{
+			act:   &Action{Type: "tradeChoose", Pick: pick},
+			decks: [][]game.Card{nd},
+		})
+	}
+	b.score(cx, cands)
+	return b.pick(cands).act
+}
+
+// decideCleanup performs the forced sale down to the pet limit, dumping the
+// lowest-value pets. This one is deterministic at every difficulty — even a
+// weak player doesn't discard their best pet by accident.
+func (b *Bot) decideCleanup(v *game.View, mem *Memory) *Action {
+	cx := newCtx(v, mem)
+	excess := cx.me.PetCount - v.MaxPets
+	if excess <= 0 {
+		return nil
+	}
+	pets := make([]game.Card, 0, cx.me.PetCount)
+	for _, c := range cx.me.Deck {
+		if c.IsPet() {
+			pets = append(pets, c)
+		}
+	}
+	slices.SortStableFunc(pets, func(a, b game.Card) int {
+		av, bv := keepValue(a), keepValue(b)
+		switch {
+		case av < bv:
+			return -1
+		case av > bv:
+			return 1
+		}
+		return 0
+	})
+	ids := make([]string, 0, excess)
+	for _, p := range pets[:excess] {
+		ids = append(ids, p.ID)
+	}
+	return &Action{Type: "sell", Cards: ids}
+}

internal/game/battle.go

@@ -874,10 +874,10 @@ func (g *Game) resolveBattle() {
 	// Tagged "result" so the client can hold it back until the replay finishes
 	// (the outcome is known now, but showing it early would spoil the battle).
 	if winner < 0 {
-		g.addLog(LogEntry{Seat: -1, Icon: "⚔️", Kind: "result",
+		g.addLog(LogEntry{Seat: -1, Icon: "⚔️", Kind: LogResult,
 			Text: fmt.Sprintf("Round %d battle ends in a draw.", g.Round)})
 	} else {
-		g.addLog(LogEntry{Seat: winner, Icon: "⚔️", Kind: "result",
+		g.addLog(LogEntry{Seat: winner, Icon: "⚔️", Kind: LogResult,
 			Text: fmt.Sprintf("%s wins the round %d battle (+%d🏆).", pname(winner), g.Round, res.Trophies)})
 	}
 	for _, p := range g.Players {

internal/game/game.go

@@ -3,6 +3,7 @@ package game
 import (
 	"crypto/rand"
 	"encoding/hex"
+	"encoding/json"
 	"errors"
 	"fmt"
 	"math/big"
@@ -49,6 +50,13 @@ type Player struct {
 	// TripledThisRound records whether the player used the Triple (trade-in)
 	// action during the current round's shop (Bison's Battle Prep).
 	TripledThisRound bool `json:"tripledThisRound"`
+	// IsBot marks a computer-controlled seat. The engine treats bots exactly
+	// like humans; the server drives their actions. BotLevel is the bot's
+	// skill in [0, 1]; BotMemory is the bot's private notebook, opaque to the
+	// engine and persisted with the game so knowledge survives restarts.
+	IsBot     bool            `json:"isBot,omitempty"`
+	BotLevel  float64         `json:"botLevel,omitempty"`
+	BotMemory json.RawMessage `json:"botMemory,omitempty"`
 }
 
 // PetCount counts pet cards in the player's deck.
@@ -196,6 +204,19 @@ func (g *Game) AddPlayer(name string) (*Player, error) {
 	return p, nil
 }
 
+// AddBot seats a computer-controlled player. Bots count as connected from
+// the start; the server is responsible for driving their actions.
+func (g *Game) AddBot(name string, level float64) (*Player, error) {
+	p, err := g.AddPlayer(name)
+	if err != nil {
+		return nil, err
+	}
+	p.IsBot = true
+	p.BotLevel = min(max(level, 0), 1)
+	p.Connected = true
+	return p, nil
+}
+
 // PlayerByID returns the player, or nil.
 func (g *Game) PlayerByID(id string) *Player {
 	for _, p := range g.Players {
@@ -278,7 +299,8 @@ func (g *Game) Buy(playerID string, rowIdx int) error {
 	p.Coins--
 	bought := g.ShopRow[rowIdx]
 	p.Deck = append(p.Deck, bought)
-	g.logf(p.Seat, "🛒", "%s bought %s %s.", p.Name, article(bought.Name), bought.Name)
+	g.addLog(LogEntry{Seat: p.Seat, Icon: "🛒", Kind: LogBuy, Source: bought.ID, CardName: bought.Name,
+		Text: fmt.Sprintf("%s bought %s %s.", p.Name, article(bought.Name), bought.Name)})
 	g.ShopRow[rowIdx] = g.drawFromTier(g.Round)
 	g.applyShopTrigger(p, bought, TriggerBuy)
 	g.advanceShopTurn()
@@ -320,7 +342,7 @@ func (g *Game) applyShopTrigger(p *Player, c Card, trigger EffectTrigger) {
 			for range n {
 				p.Deck = append(p.Deck, g.newApple())
 			}
-			g.addLog(LogEntry{Seat: p.Seat, Icon: "🍎", Source: c.ID, Spawn: "apple",
+			g.addLog(LogEntry{Seat: p.Seat, Icon: "🍎", Source: c.ID, Spawn: "apple", Count: n,
 				Text: fmt.Sprintf("%s adds %d apple%s to %s's deck.", c.Name, n, plural(n), p.Name)})
 		case ActionRefreshGold:
 			p.Coins = min(p.Coins+e.count(), CoinsPerRound)
@@ -336,7 +358,7 @@ func (g *Game) applyShopTrigger(p *Player, c Card, trigger EffectTrigger) {
 				p.Deck = append(p.Deck, g.newApple())
 			}
 			if apples > 0 {
-				g.addLog(LogEntry{Seat: p.Seat, Icon: "🍎", Source: c.ID, Spawn: "apple",
+				g.addLog(LogEntry{Seat: p.Seat, Icon: "🍎", Source: c.ID, Spawn: "apple", Count: apples,
 					Text: fmt.Sprintf("%s doubles %s's apples (+%d).", c.Name, p.Name, apples)})
 			}
 		}
@@ -363,7 +385,7 @@ func (g *Game) sellCards(p *Player, cardIDs []string) error {
 	}
 	for _, c := range sold {
 		p.Deck = append(p.Deck, g.newApple())
-		g.addLog(LogEntry{Seat: p.Seat, Icon: "🍎", Source: c.ID, Spawn: "apple",
+		g.addLog(LogEntry{Seat: p.Seat, Icon: "🍎", Kind: LogSell, Source: c.ID, CardName: c.Name, Spawn: "apple",
 			Text: fmt.Sprintf("%s sold %s — it becomes an apple.", p.Name, c.Name)})
 		g.applyShopTrigger(p, c, TriggerSell)
 	}
@@ -427,11 +449,14 @@ func (g *Game) TradeStart(playerID string, cardIDs []string) error {
 	// The discarded trio is public — everyone sees what was given up — even
 	// though the pet ultimately chosen stays secret (see TradeChoose).
 	names := make([]string, len(traded))
+	ids := make([]string, len(traded))
 	for i, c := range traded {
 		names[i] = c.Name
+		ids[i] = c.ID
 	}
-	g.logf(p.Seat, "🔄", "%s traded in %s (%s) for a tier %d pick.",
+	g.addLog(LogEntry{Seat: p.Seat, Icon: "🔄", Kind: LogTrade, Cards: ids,
-		p.Name, strings.Join(names, ", "), suit, nextTier)
+		Text: fmt.Sprintf("%s traded in %s (%s) for a tier %d pick.",
+			p.Name, strings.Join(names, ", "), suit, nextTier)})
 	g.Pending = &PendingTrade{
 		PlayerID: playerID,
 		Tier:     nextTier,
@@ -463,10 +488,12 @@ func (g *Game) TradeChoose(playerID string, pick int) error {
 	// can't see the deck. But a pet with a Buy ability performs it publicly, so
 	// we have to reveal that pet (its effect log names it anyway).
 	if hasBuyEffect(chosen) {
-		g.logf(p.Seat, "🔄", "%s's trade pick is %s %s — its buy ability triggers.",
+		g.addLog(LogEntry{Seat: p.Seat, Icon: "🔄", Kind: LogTradePick, CardName: chosen.Name,
-			p.Name, article(chosen.Name), chosen.Name)
+			Text: fmt.Sprintf("%s's trade pick is %s %s — its buy ability triggers.",
+				p.Name, article(chosen.Name), chosen.Name)})
 	} else {
-		g.logf(p.Seat, "🔄", "%s keeps their trade pick hidden.", p.Name)
+		g.addLog(LogEntry{Seat: p.Seat, Icon: "🔄", Kind: LogTradePick,
+			Text: fmt.Sprintf("%s keeps their trade pick hidden.", p.Name)})
 	}
 	// Pets obtained via the Triple action trigger their Buy effects.
 	g.applyShopTrigger(p, chosen, TriggerBuy)

internal/game/log.go

@@ -14,12 +14,28 @@ type LogEntry struct {
 	Phase  Phase  `json:"phase"`
 	Seat   int    `json:"seat"`             // acting seat, or -1 when none
 	Icon   string `json:"icon,omitempty"`   // leading emoji
-	Kind   string `json:"kind,omitempty"`   // e.g. "result" (battle outcome)
+	Kind   string `json:"kind,omitempty"`   // structured tag; see constants below
 	Text   string `json:"text"`             // the sentence itself
 	Source string `json:"source,omitempty"` // card id that caused a spawn
 	Spawn  string `json:"spawn,omitempty"`  // "apple" | "bee" for spawn entries
+	// The fields below add machine-readable copies of facts the Text already
+	// states publicly, so observers (the AI player included) don't have to
+	// parse English. They must never carry information the text doesn't.
+	Count    int      `json:"count,omitempty"`    // e.g. apples gained
+	CardName string   `json:"cardName,omitempty"` // named card, when public
+	Cards    []string `json:"cards,omitempty"`    // card ids involved, when public
 }
 
+// Structured LogEntry.Kind tags. Only "result" affects the client; the rest
+// exist so observers can follow the public action stream structurally.
+const (
+	LogResult    = "result"    // battle outcome (client holds it until the replay ends)
+	LogBuy       = "buy"       // Seat bought Source/CardName from the shop row
+	LogSell      = "sell"      // Seat sold Source/CardName (it became an apple)
+	LogTrade     = "trade"     // Seat traded in Cards for a next-tier pick
+	LogTradePick = "tradePick" // Seat took their pick; CardName set when revealed
+)
+
 // addLog appends an entry, stamping it with the next sequence number and the
 // current round/phase. Callers set Seat/Icon/Text (and Source/Spawn when the
 // entry represents something spawning off a card).

internal/game/sim.go

@@ -0,0 +1,35 @@
+package game
+
+// SimulateBattle resolves a hypothetical two-player battle between the given
+// arranged decks (top of deck first) and returns the result. It runs on a
+// scratch game, so it never touches real state — callers (notably the AI
+// player) can roll out as many what-if battles as they like. Dice rolls are
+// random unless rollDie is non-nil.
+func SimulateBattle(round, prioritySeat int, deckA, deckB []Card, rollDie func() int) *BattleResult {
+	g := &Game{
+		Round:        round,
+		PrioritySeat: prioritySeat,
+		RollDie:      rollDie,
+		// Cards minted during the simulation (apples, bees) get IDs far away
+		// from real ones, purely to avoid confusion when reading results.
+		NextCardID: 1_000_000,
+		Players: []*Player{
+			{Name: "A", Seat: 0, Deck: append([]Card(nil), deckA...)},
+			{Name: "B", Seat: 1, Deck: append([]Card(nil), deckB...)},
+		},
+	}
+	g.resolveBattle()
+	return g.Battle
+}
+
+// TierContents returns the full printed contents of a tier's shop deck —
+// public information from the box. Cards carry placeholder IDs; they are
+// reference data, not live instances.
+func TierContents(tier int) []Card {
+	scratch := &Game{}
+	scratch.buildShopDecks()
+	if tier < 1 || tier > len(scratch.ShopDecks) {
+		return nil
+	}
+	return scratch.ShopDecks[tier-1]
+}

internal/game/view.go

@@ -10,6 +10,7 @@ type PlayerView struct {
 	Trophies  int    `json:"trophies"`
 	Ready     bool   `json:"ready"`
 	Connected bool   `json:"connected"`
+	IsBot     bool   `json:"isBot,omitempty"`
 	DeckSize  int    `json:"deckSize"`
 	PetCount  int    `json:"petCount"`
 	Deck      []Card `json:"deck,omitempty"` // self only
@@ -70,6 +71,7 @@ func (g *Game) ViewFor(playerID string) View {
 			Trophies:  p.Trophies,
 			Ready:     p.Ready,
 			Connected: p.Connected,
+			IsBot:     p.IsBot,
 			DeckSize:  len(p.Deck),
 			PetCount:  p.PetCount(),
 		}

internal/server/bot_e2e_test.go

@@ -0,0 +1,83 @@
+package server
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/coder/websocket"
+
+	"github.com/greyson/super-auto-pets-board-game/internal/game"
+	"github.com/greyson/super-auto-pets-board-game/internal/store"
+)
+
+// TestE2EBotGame creates a vs-computer game over the API, plays the human's
+// shop turns over the wire, and verifies the scheduled bot actually takes
+// its own turns (spends coins) without any second client connected.
+func TestE2EBotGame(t *testing.T) {
+	st, err := store.Open(t.TempDir())
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer st.Close()
+	srv := New(st, "", false)
+	ts := httptest.NewServer(srv.Handler())
+	defer ts.Close()
+
+	ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
+	defer cancel()
+
+	resp, err := http.Post(ts.URL+"/api/games", "application/json",
+		bytes.NewBufferString(`{"name":"Human","bot":"easy"}`))
+	if err != nil {
+		t.Fatal(err)
+	}
+	var join joinResponse
+	if err := json.NewDecoder(resp.Body).Decode(&join); err != nil {
+		t.Fatal(err)
+	}
+	resp.Body.Close()
+
+	wsBase := "ws" + strings.TrimPrefix(ts.URL, "http")
+	ws, _, err := websocket.Dial(ctx,
+		wsBase+"/api/ws?game="+join.GameID+"&player="+join.PlayerID+"&token="+join.Token, nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer ws.Close(websocket.StatusNormalClosure, "")
+
+	v := readState(t, ctx, ws)
+	if v.Phase != game.PhaseShop {
+		t.Fatalf("phase = %s, want shop (bot fills the lobby instantly)", v.Phase)
+	}
+	var bot *game.PlayerView
+	for i := range v.Players {
+		if v.Players[i].IsBot {
+			bot = &v.Players[i]
+		}
+	}
+	if bot == nil {
+		t.Fatal("no bot seat in the game")
+	}
+
+	// Play the human side: pass whenever it's our turn. The game can only
+	// reach the arrange phase if the bot spends its own three coins too.
+	deadline := time.Now().Add(25 * time.Second)
+	for v.Phase == game.PhaseShop && time.Now().Before(deadline) {
+		if v.Turn == v.YouSeat && v.Players[v.YouSeat].Coins > 0 && v.Pending == nil {
+			send(t, ctx, ws, map[string]any{"type": "pass"})
+		}
+		rctx, rcancel := context.WithTimeout(ctx, 10*time.Second)
+		v = readState(t, rctx, ws)
+		rcancel()
+	}
+	if v.Phase == game.PhaseShop {
+		t.Fatalf("shop never ended; bot coins=%d", v.Players[bot.Seat].Coins)
+	}
+	t.Logf("reached phase %s; bot played its shop turns", v.Phase)
+}

internal/server/bots.go

@@ -0,0 +1,181 @@
+package server
+
+import (
+	"log/slog"
+	"math/rand/v2"
+	"time"
+
+	"github.com/greyson/super-auto-pets-board-game/internal/ai"
+	"github.com/greyson/super-auto-pets-board-game/internal/game"
+)
+
+// botDifficulty maps the API's difficulty names to a skill level and a
+// table name for the bot.
+var botDifficulty = map[string]struct {
+	level float64
+	name  string
+}{
+	"easy":   {0.25, "Robo Rookie"},
+	"medium": {0.60, "Robo Rival"},
+	"hard":   {1.00, "Robo Ace"},
+}
+
+// commitLocked is the one path every game mutation goes through: bots update
+// their memories from the new public state, the game is persisted, every
+// client gets its view, and the next bot move (if any) is scheduled. Callers
+// must hold r.mu.
+func (s *Server) commitLocked(r *room) {
+	observeBotsLocked(r.game)
+	s.persist(r)
+	r.broadcastLocked()
+	s.scheduleBotsLocked(r)
+}
+
+// observeBotsLocked gives each bot a look at the current state through its
+// own player view — the same information a human in that seat would see.
+func observeBotsLocked(g *game.Game) {
+	for _, p := range g.Players {
+		if !p.IsBot {
+			continue
+		}
+		mem := ai.LoadMemory(p.BotMemory)
+		view := g.ViewFor(p.ID)
+		ai.Observe(&view, mem)
+		p.BotMemory = mem.Marshal()
+	}
+}
+
+// scheduleBotsLocked arms a delayed move for the first bot that owes the
+// game an action. The delay is there purely for feel — instant replies make
+// the opponent seem like a vending machine. Only one timer runs per room;
+// each fired move re-schedules the next.
+func (s *Server) scheduleBotsLocked(r *room) {
+	if r.botArmed {
+		return
+	}
+	for _, p := range r.game.Players {
+		if !p.IsBot {
+			continue
+		}
+		view := r.game.ViewFor(p.ID)
+		if !ai.Pending(&view) {
+			continue
+		}
+		r.botArmed = true
+		playerID := p.ID
+		time.AfterFunc(botDelay(r.game.Phase), func() { s.runBot(r, playerID) })
+		return
+	}
+}
+
+// botDelay picks a humanlike pause before a bot move.
+func botDelay(phase game.Phase) time.Duration {
+	ms := func(base, jitter int) time.Duration {
+		return time.Duration(base+rand.IntN(jitter+1)) * time.Millisecond
+	}
+	switch phase {
+	case game.PhaseShop:
+		return ms(700, 900)
+	case game.PhaseCleanup:
+		return ms(900, 600)
+	case game.PhaseArrange:
+		return ms(1600, 1600)
+	default: // battle acknowledgement
+		return ms(500, 300)
+	}
+}
+
+// runBot fires one scheduled bot move. The state may have changed while the
+// timer ran, so everything is revalidated under the lock.
+func (s *Server) runBot(r *room, playerID string) {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	r.botArmed = false
+	p := r.game.PlayerByID(playerID)
+	if p == nil || !p.IsBot {
+		return
+	}
+	view := r.game.ViewFor(playerID)
+	if !ai.Pending(&view) {
+		// Someone else moved the game on; check the other seats.
+		s.scheduleBotsLocked(r)
+		return
+	}
+	act := ai.New(p.BotLevel).Act(&view, ai.LoadMemory(p.BotMemory))
+	var err error
+	if act == nil {
+		err = game.ErrInvalidAction
+	} else {
+		err = applyBotAction(r.game, playerID, act)
+	}
+	if err != nil {
+		// A bot must never wedge the game: fall back to the simplest legal
+		// move for the phase.
+		slog.Warn("bot action failed; using fallback", "game", r.game.ID, "player", playerID, "err", err)
+		if err := botFallback(r.game, playerID); err != nil {
+			slog.Error("bot fallback failed", "game", r.game.ID, "player", playerID, "err", err)
+			return
+		}
+	}
+	s.commitLocked(r)
+}
+
+// applyBotAction maps a bot decision onto the engine, mirroring the client
+// message dispatch in apply().
+func applyBotAction(g *game.Game, playerID string, a *ai.Action) error {
+	switch a.Type {
+	case "buy":
+		return g.Buy(playerID, a.Row)
+	case "sell":
+		if g.Phase == game.PhaseCleanup {
+			return g.CleanupSell(playerID, a.Cards)
+		}
+		return g.Sell(playerID, a.Cards)
+	case "trade":
+		return g.TradeStart(playerID, a.Cards)
+	case "tradeChoose":
+		return g.TradeChoose(playerID, a.Pick)
+	case "pass":
+		return g.Pass(playerID)
+	case "arrange":
+		return g.SubmitOrder(playerID, a.Order)
+	case "ready":
+		return g.AcknowledgeBattle(playerID)
+	}
+	return game.ErrInvalidAction
+}
+
+// botFallback makes the trivially legal move for whatever the game is
+// waiting on: pass the shop turn, take the first trade option, sell the
+// first excess pets, submit the deck as-is, or acknowledge the battle.
+func botFallback(g *game.Game, playerID string) error {
+	p := g.PlayerByID(playerID)
+	if p == nil {
+		return game.ErrInvalidAction
+	}
+	switch g.Phase {
+	case game.PhaseShop:
+		if g.Pending != nil && g.Pending.PlayerID == playerID {
+			return g.TradeChoose(playerID, 0)
+		}
+		return g.Pass(playerID)
+	case game.PhaseCleanup:
+		excess := p.PetCount() - game.MaxPets
+		ids := make([]string, 0, excess)
+		for _, c := range p.Deck {
+			if c.IsPet() && len(ids) < excess {
+				ids = append(ids, c.ID)
+			}
+		}
+		return g.CleanupSell(playerID, ids)
+	case game.PhaseArrange:
+		ids := make([]string, len(p.Deck))
+		for i, c := range p.Deck {
+			ids[i] = c.ID
+		}
+		return g.SubmitOrder(playerID, ids)
+	case game.PhaseBattle:
+		return g.AcknowledgeBattle(playerID)
+	}
+	return game.ErrInvalidAction
+}

internal/server/server.go

@@ -62,6 +62,9 @@ type room struct {
 	game  *game.Game
 	conns map[*client]struct{}
 	debug bool // mirrors Server.debug, for broadcastLocked
+	// botArmed is set while a delayed bot move is scheduled, so only one
+	// timer exists per room at a time.
+	botArmed bool
 }
 
 // getRoom returns the room for a game ID, loading it from the store if it
@@ -78,6 +81,11 @@ func (s *Server) getRoom(gameID string) (*room, error) {
 	}
 	r := &room{game: g, conns: make(map[*client]struct{}), debug: s.debug}
 	s.rooms[gameID] = r
+	// If the game was persisted mid-bot-turn (e.g. across a server restart),
+	// get the bot moving again.
+	r.mu.Lock()
+	s.scheduleBotsLocked(r)
+	r.mu.Unlock()
 	return r, nil
 }
 
@@ -116,24 +124,45 @@ type joinResponse struct {
 func (s *Server) handleCreate(w http.ResponseWriter, req *http.Request) {
 	var body struct {
 		Name string `json:"name"`
+		// Bot, when set, fills the other seat with a computer player:
+		// "easy" | "medium" | "hard".
+		Bot string `json:"bot"`
 	}
 	if err := json.NewDecoder(req.Body).Decode(&body); err != nil {
 		httpError(w, http.StatusBadRequest, "invalid JSON body")
 		return
 	}
+	var bot struct {
+		level float64
+		name  string
+	}
+	if body.Bot != "" {
+		var ok bool
+		bot, ok = botDifficulty[body.Bot]
+		if !ok {
+			httpError(w, http.StatusBadRequest, "unknown bot difficulty")
+			return
+		}
+	}
 	g := game.New()
 	p, err := g.AddPlayer(strings.TrimSpace(body.Name))
 	if err != nil {
 		httpError(w, http.StatusBadRequest, err.Error())
 		return
 	}
+	if body.Bot != "" {
+		if _, err := g.AddBot(bot.name, bot.level); err != nil {
+			httpError(w, http.StatusBadRequest, err.Error())
+			return
+		}
+	}
 	r := &room{game: g, conns: make(map[*client]struct{}), debug: s.debug}
 	s.mu.Lock()
 	s.rooms[g.ID] = r
 	s.mu.Unlock()
 
 	r.mu.Lock()
-	s.persist(r)
+	s.commitLocked(r)
 	r.mu.Unlock()
 	writeJSON(w, joinResponse{GameID: g.ID, Code: g.Code, PlayerID: p.ID, Token: p.Token})
 }
@@ -163,9 +192,8 @@ func (s *Server) handleJoin(w http.ResponseWriter, req *http.Request) {
 		httpError(w, http.StatusConflict, err.Error())
 		return
 	}
-	s.persist(r)
 	resp := joinResponse{GameID: r.game.ID, Code: r.game.Code, PlayerID: p.ID, Token: p.Token}
-	r.broadcastLocked()
+	s.commitLocked(r)
 	r.mu.Unlock()
 	writeJSON(w, resp)
 }

internal/server/ws.go

@@ -73,6 +73,9 @@ func (s *Server) handleWS(w http.ResponseWriter, req *http.Request) {
 	r.conns[c] = struct{}{}
 	p.Connected = true
 	r.broadcastLocked()
+	// Safety net: if a scheduled bot move was ever lost (crash between
+	// persist and timer), a player connecting re-arms it.
+	s.scheduleBotsLocked(r)
 	r.mu.Unlock()
 
 	defer func() {
@@ -147,8 +150,7 @@ func (s *Server) apply(r *room, c *client, msg clientMessage) {
 		c.sendError(err.Error())
 		return
 	}
-	s.persist(r)
+	s.commitLocked(r)
-	r.broadcastLocked()
 }
 
 // broadcastLocked sends each connected client its own view of the game.

web/src/api.ts

@@ -13,8 +13,10 @@ async function post(path: string, body: unknown): Promise<Session> {
   return data as Session
 }
 
-export function createGame(name: string): Promise<Session> {
+export type BotDifficulty = 'easy' | 'medium' | 'hard'
-  return post('/api/games', { name })
+
+export function createGame(name: string, bot?: BotDifficulty): Promise<Session> {
+  return post('/api/games', bot ? { name, bot } : { name })
 }
 
 export function joinGame(code: string, name: string): Promise<Session> {

web/src/components/Home.tsx

@@ -1,7 +1,14 @@
 import { useState } from 'react'
 import { createGame, joinGame } from '../api'
+import type { BotDifficulty } from '../api'
 import type { Session } from '../types'
 
+const BOT_LEVELS: { value: BotDifficulty; label: string; blurb: string }[] = [
+  { value: 'easy', label: '🐣 Easy', blurb: 'Learns you the ropes' },
+  { value: 'medium', label: '🐺 Medium', blurb: 'Puts up a fight' },
+  { value: 'hard', label: '🦁 Hard', blurb: 'Shows no mercy' },
+]
+
 // Home is the create/join screen shown when there's no active session.
 export function Home({ onSession }: { onSession: (s: Session) => void }) {
   const [name, setName] = useState('')
@@ -50,6 +57,24 @@ export function Home({ onSession }: { onSession: (s: Session) => void }) {
           Host a new game
         </button>
 
+        <div className="home-divider">
+          <span>or challenge the computer</span>
+        </div>
+
+        <div className="home-bots">
+          {BOT_LEVELS.map((b) => (
+            <button
+              key={b.value}
+              className="btn btn-secondary"
+              disabled={busy}
+              title={b.blurb}
+              onClick={() => run(() => createGame(name, b.value))}
+            >
+              {b.label}
+            </button>
+          ))}
+        </div>
+
         <div className="home-divider">
           <span>or join a friend</span>
         </div>

web/src/components/Table.tsx

@@ -93,7 +93,13 @@ export function Table({ session, onLeave }: { session: Session; onLeave: () => v
               key={p.id}
               className={`topbar-player ${p.seat === view.youSeat ? 'is-you' : ''}`}
             >
-              <span className={`conn-dot ${p.connected ? 'on' : 'off'}`} />
+              {p.isBot ? (
+                <span className="bot-dot" title="Computer player">
+                  🤖
+                </span>
+              ) : (
+                <span className={`conn-dot ${p.connected ? 'on' : 'off'}`} />
+              )}
               <span className="topbar-name">{p.name}</span>
               <span className="chip">🏆 {p.trophies}</span>
               {(view.phase === 'shop' || view.phase === 'cleanup') && (
@@ -127,7 +133,7 @@ export function Table({ session, onLeave }: { session: Session; onLeave: () => v
         )}
       </div>
 
-      {opponents.some((p) => !p.connected) && view.phase !== 'lobby' && (
+      {opponents.some((p) => !p.connected && !p.isBot) && view.phase !== 'lobby' && (
         <div className="banner banner-warn">An opponent is disconnected…</div>
       )}
       {error && <div className="toast">{error}</div>}

web/src/styles.css

@@ -371,6 +371,17 @@ h3 {
   gap: 10px;
 }
 
+/* The three computer-opponent difficulty buttons share the row evenly. */
+.home-bots {
+  display: flex;
+  gap: 10px;
+}
+
+.home-bots .btn {
+  flex: 1;
+  white-space: nowrap;
+}
+
 .home-error {
   color: var(--red-soft);
   font-weight: 700;
@@ -479,6 +490,12 @@ h3 {
   background: var(--red);
 }
 
+/* Bots swap the connection dot for a little robot face. */
+.bot-dot {
+  font-size: 0.85rem;
+  line-height: 1;
+}
+
 /* The room code, stamped like a label on a card tray. */
 .topbar-code {
   font-family: var(--font-display);

web/src/types.ts

@@ -26,6 +26,7 @@ export interface PlayerView {
   trophies: number
   ready: boolean
   connected: boolean
+  isBot?: boolean
   deckSize: number
   petCount: number
   deck?: Card[]