@deepseek-ai/dsh-client-ui-game-holdem

@deepseek-ai/dsh-client-ui-game-holdem

English | 中文

Optional browser table for holdem-rules-v1. It registers the durable holdem key in game-center.table and renders two-to-eight compact local-avatar seats on one elliptical anchor system, stack and dealer/blind labels, five fixed board slots, the caller's hole cards, finite amount-bearing actions, and settlement playback. The stage always uses a deep neutral surround and green felt so cards, names, actions, and chips keep the same contrast in every Harness theme. Avatar, name, stack, hole-card, wager, and showdown anchors derive from the same seat angle, so adding seats does not require another absolute-position matrix. Wagers remain on an inner radial track; bet and raise text does not persist because the chip amount is authoritative, while folded seats desaturate and only paused, out, or all-in state keeps a badge. The current actor receives a ring. The current pot display exposes main and side pots with eligible seats, while a pending-order trace and highlighted flow segment describe the actual remaining action order rather than a decorative fixed arrow.

The renderer applies the shared zero-Cordis game UI scope from ui-primitives at its root and composes its dismissible disclosure, chevron, and semantic status badge. Hold'em-specific seat geometry, wager tracks, board, action dock, coach content, and settlement playback remain local to this package, so another game can reuse interaction and visual tokens without inheriting poker behavior.

Settlement playback is gated from the live next hand. It reconstructs board, revealed holes, contributions, pot winners, ranks, best five cards, kickers, and post-hand stacks from one settlement snapshot; runs remaining board cards in street order; reveals showdown hands; and flies each pot to its own winners before presenting lightweight YOU WIN, YOU LOSE, or PUSH feedback. Winner treatment follows net hand change, so collecting a small side pot while losing more overall cannot produce a win label or winning-hand glow. The green confirmation control acknowledges that settlement before exposing the next hand, while terminal settlements continue to final standings instead of claiming another hand exists. Returning to an already advanced room does not replay a historical settlement.

Keyboard users reach fold, check, call, quick sizing, and submission as native controls. The compact amount slider uses actual chip totals across its track and snaps to finite Host-generated raise candidates. Preflop shortcuts select 2.5×, 3×, 4×, or all-in targets; postflop shortcuts select one-third-pot, half-pot, three-quarter-pot, pot, or all-in targets. A short stack that cannot complete a call receives one passive call-all-in action and no raise controls. Submission state and conflicts stay beside the action dock, a visible countdown describes the unattended-human deadline, and an optional local sound plays once for each new owned action revision. After that deadline, a funded human sees a compact paused-table notice and one Return to table action; action and coach controls remain hidden until it resumes the same match. The caller's two hole cards fan below the hero seat without sharing the avatar or wager anchor. An incompatible rules projection fails closed instead of rendering unknown private fields.

The table coach is a local deterministic probability engine, not an Agent or language-model call. Its collapsed control sits on the table baseline and its expanded panel moves above the action dock instead of covering a decision. The live tab shows the hand number, street, authoritative revision, Host-computed showdown equity, win/tie/loss frequencies, method and sample count, active unknown-opponent count, pot odds, and street guidance. Two offline tabs provide a bundled hand-ranking chart and basic Hold'em concepts without a runtime network image. It recomputes after visible board or public-action changes; a stable match revision produces a stable result. Heads-up rivers add exact uniform-range enumeration with its assumptions and non-GTO boundary visible beside the action. VPIP, PFR, aggression, Wilson intervals, and confidence come only from public action history. Fewer than 20 observed hands are explicitly insufficient and fewer than 100 remain emerging. The UI never presents EV loss or a GTO verdict without qualifying solver output. The coach, replay-download menu, and room menu all support outside-click and Escape dismissal with stateful disclosure indicators.

When the owned stack reaches zero, the table removes action and coach controls and labels the seat as an out spectator while the remaining table continues. A completed match displays hands completed and sorted final stacks, with the rematch as a centered primary action and replay download and lobby navigation in a secondary row. The component receives a caller-safe MatchView, never reads Cordis Context, and submits only an action id present in match.legalActions.

Model Experience

None, as this browser renderer contributes no prompt, tool, observation, or model request.

KV Cache effect

None; the Host Hold'em profile and isolated Agent seat controller own all model-visible content.

Known Limitations and Deferred Work

• Version 1 does not accept arbitrary chip amounts. The slider and quick sizing controls select only the authoritative candidates shared by humans and Agents.
• Historical hand analysis is not embedded in the live table. Authoritative event replay and identity-aware private projections remain on the Host, while the room exposes bounded download actions.
• The public-action style snapshot currently covers the selected match rather than a cross-match player profile. Position splits, persisted hand selection, solver-backed EV review, and opponent timing indicators are deferred.
• The renderer displays tournament chips only and has no cash-value, wallet, purchase, or payout integration.