Self-play transcripts produced by `pnpm selfplay --transcripts` land in
`tmp/selfplay-runs/<timestamp>/` — operator scratch, not part of source.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The blind-mode CasualBrain heuristic ignored the moderator's
'<own>_in_check' announcement and scored moves on capture/advance/development
signals uncorrelated with check resolution. chess.js rejected every
non-resolving attempt, BotDriver's RETRY_CAP=5 fired, and the bot resigned.
100-game blind self-play: 100% resignations at avg ply 26.
Fix:
- CasualBrain.detectOwnCheck() scans newAnnouncements for the own-color
in_check tag; when set, heuristicPick() applies a +5000 boost to king
moves so they're tried first. Information stays within the public
moderator vocabulary — no oracle access, view-filter invariant intact.
- RETRY_CAP raised 5 -> 25. Vanilla never hits the cap (chess.js verbose
moves are guaranteed legal); blind needs more budget to find a legal
move through pseudo-legal candidates.
- BotDriver.botResign() now logs '[bot resign]' with gameId/color/mode/ply/
reason/detail. Previously silent — operator had no signal in journald.
Verification (100-game blind Casual-vs-Casual self-play):
- avgPly: 26 -> 90 (3.5x)
- Resignations: 100% -> 17%
- Checkmates: 0% -> 42%
- Threefold draws: 0% -> 41%
Vanilla regression check (80 games combined): 0 resigns either way,
strength unchanged (Casual still wins 98% vs random).
78 tests pass (was 75; +2 new check-resolution tests, +1 retry-cap test
updated to match new cap).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- CLAUDE.md: phase line moved to "Phase 1 deployed"; key files lists
the new bot module, game-end extraction, and selfplay harness.
- DECISIONS.md: new "Phase 1 implementation outcomes" subsection records
the CasualBrain-engine reversal, the FEN-vanilla-only invariant, why
blind keeps heuristic, and the bot-slot token randomization. The
earlier "Stockfish deferred" entry is partially superseded.
- .claude/handoffs/: handoff document for the next session.
The hand-rolled scoring heuristic lost to a random-move baseline 7-7 in
self-play — far below the spec's >=80% acceptance bar. Swap in a real
chess engine (js-chess-engine, MIT, ~400KB, no native deps) for vanilla
mode at level 2 with randomness=30 to break threefold cycles.
- BrainInput.fen added; driver populates it ONLY in vanilla mode.
Blind mode omits the FEN so the engine path can't smuggle opponent
positions past the view filter.
- CasualBrain in vanilla: convert FEN -> EngineGame -> ai({level: 2});
validate the engine's move is in legalCandidates; fall back to
heuristic on miss.
- Blind mode unchanged (engine isn't useful when only own pieces are
visible — that's Phase 2 Recon's territory).
Self-play vs RandomBrain (100 games each direction, vanilla):
- Casual(W) vs Random(B): W=97%
- Random(W) vs Casual(B): B=96%
Casual-vs-Casual vanilla balanced, ~5-30ms/move. All 54 tests still pass.
Refresh .secrets.baseline (stale) to allow new pnpm-lock.yaml hashes.
Track aiOpponent in game store; show a pill badge in the topbar for AI
games, update turn label to "<Brain> is moving…" on the bot's turn,
and suppress the disconnected-opponent banner when the opponent is a bot.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- maybeAbandon Promise no longer floats from setTimeout
- broadcastSinceLast loses dead extra parameter
- bot-slot token is randomized so a third party can't hijack the
bot's color by guessing a fixed placeholder
Replace broadcastNewAnnouncements/broadcastUpdate with watermark-based
broadcastSinceLast; add pokeBot helper; make all state-mutating handlers
async; hook pokeBot after every mutation so the CasualBrain fires on
each turn without oracle access.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Extract endGame/finalizeIfEnded to game-end.ts so driver.ts can call
finalizeIfEnded after an applied move (fix: bot checkmate was not
setting game.status='finished'). Wrap entire dispatch() call in
try/catch for exception safety. Move lastSeenAnnouncementCount advance
to after successful dispatch so retry attempts see FSM rejection
announcements. Add checkmate-finalize test; lock retry-cap at 5 calls.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Wires Brain to Game: init/onStateChange/dispose lifecycle, in-flight
mutex, 5-attempt retry loop with attemptHistory, resign-on-cap. Also
adds Game.aiOpponent? field to state.ts for Task 5.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- DECISIONS.md: in-game chat (player↔player and human↔Gemma) deferred
indefinitely. Blind-mode chat is a side channel that defeats the
moderator-vocabulary security boundary; chat with Gemma leaks belief
state mid-game. Resolvable but expensive — revisit only on demand.
- Spec: same deferral noted in "Out of scope".
- New plan: docs/superpowers/plans/2026-04-28-ai-player-phase-1-casual.md
— 13 tasks, 80 sub-steps. Phase 1 only (Casual bot end-to-end). Phase 2
(Recon) gets its own plan once Phase 1 outcomes inform Recon's target.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Updates CLAUDE.md "Current State" + "Key files" to point at the new spec.
Adds DECISIONS.md "AI / computer player" section (11 settled decisions).
Strikes through the prior "Client-side AI / hint generation — out of scope"
row with a "partially superseded" note: the reversal applies only to the
human-vs-AI path. Adds 7 new Deferred/Rejected rows for AI-feature scope.
Handoff at .claude/handoffs/2026-04-28-170713-ai-player-spec.md captures
session state for the next pickup (writing-plans → Phase 1 implementation).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two-phase plan: Casual bot first (algorithmic, ~200 LoC), then Recon bot
(gemma4:26b chat agent) with persistent private chat history per game.
Bots play through the same view filter and FSM as humans — no oracle access.
Endpoint priority: steel141 RTX 3090 Ti primary, pve197 V100 fallback. Mid-game
GPU failover allowed (one-way). Reasoning hidden from user during play, revealed
in collapsible post-game panel.
Reverses the 2026-04-28 DECISIONS.md row "Client-side AI / hint generation —
explicitly out of scope." Reversal is partial: human-vs-human hint generation
remains rejected; this spec only adds AI in the human-vs-AI path.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Svelte 5 $effect tracks every $state read inside its body. The
lifecycle effect that calls game.connect(gameId) implicitly read
state.ws (inside connect()) and then wrote to it, producing an
effect_update_depth_exceeded loop. Symptom in production: the
browser opened ~12 WS connections/sec, none completed the upgrade
handshake, and the lobby flow appeared stuck on 'waiting for
opponent' (the opponent's WS never stabilized long enough for the
server to send 'joined'). untrack() opts the call out of dep tracking.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Svelte 5 runes ($state, $derived, $effect) only run through the
compiler in .svelte and .svelte.ts/js files. A plain .ts file leaves
$state(...) as a literal call at runtime, causing
"ReferenceError: $state is not defined" and a blank page.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
claude (blind_chess)