bot-haiku

Kriegspiel bot that asks an Anthropic Haiku model to choose the next action from the bot's private game state.

What it does

• registers as a listed Kriegspiel bot
• polls assigned games from the live API
• does not create waiting lobby games by default
• can join another bot's waiting lobby game with 0.1% probability while still under its active-game cap
• builds a prompt from the current rule variant, private FEN, legal actions, and private scoresheet
• asks an Anthropic Haiku model for the top ranked next actions in strict JSON
• validates the model output against the server-provided legal actions
• checks Anthropic availability with a tiny cached preflight call before joining a new bot-vs-bot game
• skips the join if Anthropic is unavailable or out of quota
• still falls back safely if the model response itself is malformed

Setup

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env
python bot.py --register
python bot.py

The bot reads the live rules text from the sibling ks-content repository:

• ks-content/rules/berkeley.md
• ks-content/rules/README.md

Set KRIEGSPIEL_CONTENT_RULES_DIR to override that rules directory.

By default the registration email is bot-haiku@kriegspiel.org.

By default the bot does not create open lobby games on its own. That behavior is controlled with:

• KRIEGSPIEL_AUTO_CREATE_LOBBY_GAME=true|false
• KRIEGSPIEL_AUTO_CREATE_RULE_VARIANT=berkeley|berkeley_any
• KRIEGSPIEL_AUTO_CREATE_PLAY_AS=white|black|random
• KRIEGSPIEL_SUPPORTED_RULE_VARIANTS=berkeley,berkeley_any
• KRIEGSPIEL_MAX_ACTIVE_GAMES_BEFORE_CREATE=1

Bot-vs-bot play is also enabled by default:

• the bot samples open waiting games at most once per minute
• it will only consider games created by another bot
• it samples that decision at most once per minute
• it will try to join one with 0.1% probability on that minute check
• it uses the same 1-active-game cap for intentional bot-vs-bot joins
• it keeps the local cooldown even when no join candidate is found, matching backend bot-join limits and avoiding tight lobby scans

Anthropic prompting defaults:

• system prompt carries the rules and overall Kriegspiel scene
• the first user prompt carries private scoresheet history, recent referee items, and legal actions
• later user prompts carry only incremental updates: new referee items, rejected candidates, and refreshed legal options
• the bot keeps per-game conversation history and reuses it on the Anthropic Messages API
• automatic prompt caching is enabled with a 1-hour TTL to maximize cache reuse across retries and later turns
• the bot asks for the top 10 ranked candidate actions by default
• if a batch fails, it asks the model for the next batch of candidates
• defaults can be tuned with:
  • ANTHROPIC_MODEL=claude-haiku-4-5-20251001
  • ANTHROPIC_MODEL_BATCH_SIZE=10
  • ANTHROPIC_MAX_BATCHES_PER_TURN=5
  • ANTHROPIC_PREFLIGHT_SUCCESS_TTL_SECONDS=60
  • ANTHROPIC_PREFLIGHT_FAILURE_TTL_SECONDS=15

Test

python -m unittest discover -s tests

systemd

A production host can run the bot as a service with deploy/kriegspiel-haiku-bot.service.