Repository navigation
v0.18.0
A faster, cheaper, more transparent auto-player
Play speed and prompts
- Auto-play now engages the game's native Fast Mode while it drives and restores your speed on pause/stop. Trivial steps (proceed, single-path maps, reward screens) resolve without a model request.
- Whole-turn plans: one model request can lay out several combat steps; each step is re-checked against the live frame and the plan stops early if the board changed (boss summons, intent flips). New optional "minimum decision interval" under Advanced settings.
- The prompt now splits a static run prefix (deck, relics, ascension modifiers, glossary) from the per-step dynamic state, so cached prefixes cut repeat token cost on long fights.
You can see what it's doing
- In-game HUD: an AI banner, card outline, target line and a reason bubble show each accepted action. F7 pauses/resumes, F8 opens the panel.
- End-of-run report card: decisions and token spend for the run, copyable; a bounded journal (50 runs) feeds a same-character summary back to the model.
MCP and providers
- Both MCP surfaces return slimmed compact state: repeated deck/relic/ascension/glossary blocks and unchanged sections are deduplicated per session;
act(..., then=[...])no longer re-embeds the full state each step. - Providers that truncate streamed and non-streamed tool arguments now fall back to reply-JSON card play instead of stalling; the truncation lesson persists per endpoint/model/key.
Validation boundary (read this)
- The 2026-10-29 isolated live pass covered turbo/autopilot/whole-turn plans/combat math/HUD/report card/the
thenplan on both MCP surfaces — and found four bugs that are now fixed. - Card outline & target line, the F7 hotkey, co-op teammate play and full-run win-rate were not observed live.
- The MCP payload dedup (#233) and prompt-prefix/divergence/pacing work (#234) merged after that live pass and carry offline tests only, per the maintainer's release decision. Details in
CHANGELOG.mdanddocs/live-validation-checklist.md.