V0.2 liveDesktop assistant · Game AI

Hearthstone
Standard Agent

A desktop assistant that follows the public game state, reads the legal actions Hearthstone exposes to its client, and explains the strongest visible move without controlling the game.

View source on GitHub Run it locally

1,184

collectible cards in the current seven-set Standard catalogue

regression tests covering live actions, privacy, data validation and game flow

60.6%

average structured-effect coverage; uncertainty remains visible

The loop

Legal moves come from the client, not guesswork.

01Validated Standard catalogue

02Public-state tracker

03Client legal actions

04Effect-aware scoring

05Lethal search

06Desktop overlay

Implemented

The V0.2 live-advice pipeline

1,184 collectible cards across the seven current Standard sets in the validated snapshot
Authoritative legal actions and targets parsed from Power.log OPTIONS packets
Card names, text, cost, type, mechanics and Standard-set membership available to the advisor
Visible one-step and multi-action lethal detection with an explainable action sequence
Effect-aware scoring for damage, healing, draw, summon, buffs, removal, keywords and board pressure
Confidence automatically reduced when a complex card effect is not fully structured
Opponent hand identities hidden behind a public Observation interface
Validated HearthstoneJSON refresh pipeline with rotation-review detection and a coverage report
Top-most desktop chat overlay with read-only Power.log event tracking
Installable desktop control center with explicit start, stop, reopen and exit controls
Practice and friendly-match advice; ladder is restricted to tracking and post-game review
Authentication provider boundary and visible sign-in/register placeholders for a future service

Try it

Run an automated match

git clone https://github.com/nieruiyuanpdskd2007-commits/my-website.git
cd my-website/projects/hearthstone-standard-agent
python3 main.py --agent-a rule --agent-b random --verbose
python3 desktop_main.py
python3 -m live.main --mode practice --demo

The simulator and live parser are intentionally dependency-free. Generate JSONL training samples with--replayor run batch evaluation with--games 100.

Roadmap

V0.1Playable research loop and desktop shellComplete

V0.2Current Standard knowledge and authoritative live-action advisorAvailable now

V0.3Replay evaluation and deeper per-card effect modelsNext

V0.5Entity Transformer, Policy + Value and information-set searchPlanned

V1.0Multiple classes, decks and patch-aware opponent modelingGoal