SLAY2 — Slay the Spire 2 Decision Intelligence
Status as of 2026-10-28 (the source of truth is docs/STATUS.md; nothing below is a claim beyond it):
- Working and tested: exact probability calculators for card-reward rarity with the hidden pity counter, potion drops, "?" rooms, draws, shop prices, card removal and gold (
/odds and POST /api/v1/odds); Route Lab (/routes), which compares up to three typed routes through an act exactly (what the "?" rooms turn into, rare-card and potion odds, gold, relics) and never claims one is safer; and a manual card advisor that is a labelled heuristic, never a win probability (/advisor; it rates a card against the roles the deck lacks and by how it works with the cards you already hold: 385 cards are tagged by what they supply and what they need over 27 mechanics, the page names the engines your deck is building and the cards still waiting for support, and the starting relic counts as a supplier, as does any relic you list that provides a mechanic (that reading of relics is a judgement with no registered test; it was read once, as an exploration, against simulated runs that keep their relics). The strengths of that synergy term are a judgement, checked so far by an exploratory reading of 19,717 simulated runs (/methods#synergy: the term improves the held-out prediction of getting past the Act-1 boss; an engine credit has no support in those data, and a development stage on separate simulated seeds chose to keep none: the page names the engines a deck is building but its index does not count them); the one registered held-out test found that it improves what a simulated pilot picks (33.0 % of runs past the Act-1 boss against 26.0 % without the term, p = 3.8e-60, every class ahead; nothing about later bosses, winning or the live game), and the same test repeated in a simulator whose runs keep their relics gives 40.4 % against 33.3 % (p = 3.0e-58, every class ahead). It covers the draft, not the playing of cards against enemies): you say which Act 1 you are in (Overgrowth or Underdocks, with different elites and bosses) and what the opening Ancient gave you, its effect on your deck is applied as you report it, and the Act 1 version tilts the ranking by at most 30 % as its own labelled line, a judgement and not a measurement. Target: STS2 main v0.107.1; the beta v0.111.0 is refused. The pages are tested on phones, tablets and laptops (320 to 1280 px, enlarged text, no layout shift, no JavaScript) and make no request but to their own origin.
- Deployed and checked: https://slay-part2.vercel.app (Vercel, production alias; verified serving
main at c324e7e on 2026-10-28 at 01:04 UTC, the merge of pull request #18 (00:58 UTC): the seven pages answer 200, /api/v1/health reports that commit, /methods carries the results of the Ascension 9 study (ADR-058), and the whole browser suite (345 tests, desktop and mobile, axe) passed against the hosted site (01:05 to 01:12 UTC, 345 passed, 0 failed, 133 skipped by design); evidence in docs/DEPLOY.md, docs/TEST-REPORT.md and docs/STATUS.md). Everything of pull requests #10 to #18 is live.
- Research harness, with measured results: simulations run on a third-party simulator used locally only (live-game parity is unverified). Card values measured from randomised drafting are shown on
/codex and as a labelled second opinion on /advisor; a rules policy that follows them got past the Act-1 boss in 65.1 % of held-out simulated Ironclad games against 43.3 % for the hand-written ratings, and won 6 of 1,994 (each win re-derived by replay); four more classes were measured and tested the same way, and the Ironclad values were measured again under the stronger tuned pilot (the new values got past the Act-1 boss in 79.4 % of held-out games against 74.8 % for the old ones, and show nothing after it). A threshold-tuned pilot (one pre-registered held-out test) got past the Act-1 boss in 74.8 % of the same games against 65.1 % for the untuned one and did not win more (6 against 5 of about 2,000); a policy that plans every fight on clones (about 2 minutes a game) got past the Act-1 boss in 74.3 % of 300 held-out games against 64.7 % for the rules and won 2 against 1 (no difference in winning can be shown at that size); no neural network beats the rules baseline yet. The opening-relic study (the Ancient Neow's three relics) was registered, run with exact class ids and tested once: a pilot that takes the best-measured relic got past the Act-1 boss in 61.05 % of 9,932 held-out paired games across the five classes against 54.30 % for taking the first helpful relic (exact McNemar p = 9.1e-45); every number is simulated, from one rules pilot, in Overgrowth at Ascension 0 (pull request #7, /methods#opening-ancient once merged). One caveat outweighs the others (gap G14, ADR-052): the pinned simulator drops every relic a run gains after its first fight, so all of those numbers describe a game whose runs hold two relics; with a harness variant that keeps them, the same rules_v2 pilot got past the Act-1 boss in 63.1 % of 2,444 paired development games instead of 52.5 %, went 3.04 floors further and won 34 games against 2 (Ironclad 25 of 482 against none). No registered result is withdrawn, since each compared two arms in the same simulator, and the replications with the relics kept are registered and run (pull requests #11, #12 and #14, all merged): the synergy term still helps (40.4 % against 33.3 % past the Act-1 boss, p = 3.0e-58, ADR-054); card values measured again with the relics kept got past the Act-1 boss in 69.6 % of held-out paired games against 68.2 % for the earlier values (p = 0.001, ADR-053), so Silent, Regent, Necrobinder and Defect now show them and Ironclad, whose own games favour its earlier set, keeps it; the opening-relic study was played again with the relics kept and its rule still helps: 69.2 % of 9,629 held-out paired games against 64.5 % for the first helpful relic (p = 1.7e-23; every class ahead at p below 0.01) (ADR-055; the new relic values replace the first study's on the site, and the first study stays on /methods as a record). The three results were then tested again at Ascension 9, the hardest level the simulator plays (ADR-058, registered before its first game, 80,000 held-out games from a clean checkout of the registration commit, nothing refitted, all 294 wins re-derived, every batch sample-audited, each pooled comparison recomputed from the episode files by code that shares nothing with the harness): each came out ahead, by fewer points than at Ascension 0 because the pilots get past the Act-1 boss far less often there (the synergy term 10.1 % against 7.1 %, p = 6.8e-28; the card values the site shows 24.6 % against 9.6 %, p = 1.3e-247; the opening-relic rule 26.4 % against 22.9 %, p = 1.8e-16). The advisor now tells a run at Ascension 1 to 10 what was and was not tested at its level, and : Silent's own games had the rule ahead at p = 0.019, above the bar of 0.01 set in advance, so the page shows Silent's measurements and says it has no tested advice there. Nothing is tested for Ascension 10 (the second boss is never played) or for 1 to 8 (). Details and caveats on .
- Audited: five red-team rounds in docs/ADVERSARIAL-AUDIT.md (API, full-run truth, UX on phones and tablets, my own study design, my own documents; campaigns 7 and 8 and the trained-network part of 4 are still open). Where the project stands against the acceptance bar of prompt 11, clause by clause: docs/RELEASE-AUDIT.md (an honest partial product).
SLAY2 is an independent, unofficial, Vercel-hosted coaching and decision-intelligence website for players of Mega Crit's Slay the Spire 2. It is not a browser clone or redistribution of the game. The product combines conditional probability calculators, context-sensitive strategy guidance, validated fair-information simulators, search, reproducible machine learning and outstanding premium UI/UX. Gameplay is in the user's legitimately installed copy of STS2. Manual entry works without an account, mod, or subscription.
Quickstart
First-time player (no install): open the site's Odds page, pick Card rewards, say how many non-rare cards you have seen since your last rare (or list the packs you have seen) and read the exact chance of a rare in your next pack; the fraction behind every number is one click away. On Advisor choose your character, enter your deck and the offered cards and read the ranking with its reasons and how stable it is. Your run lives in your browser only; export it as JSON whenever you like.
Advanced player or competitor: /routes (Route Lab) compares up to three routes through an act exactly, in the link, so a route can be shared; POST /api/v1/odds and POST /api/v1/recommend are the same engines as JSON (GET /api/v1/capabilities lists what is supported and what is not; every answer says which build and which assumptions it rests on); /methods shows every rule with its evidence, every simulated batch with its sample and interval, and the known gaps of the simulator. To check a published number yourself, follow docs/REPRODUCIBILITY.md (seeds, hashes, the exact commands, what can and cannot be reproduced without the third-party simulator).
Developer / competitor (local):
node -v # 22.x (see .node-version); pnpm 10.28.0 is pinned by package.json
pnpm install --frozen-lockfile
pnpm --filter @slay2/web dev # http://localhost:3000
pnpm check # typecheck, lint, prettier, all unit tests, production build
pnpm --filter @slay2/web e2e # browsers: PW_CHROMIUM_PATH=/path/to/chromium if Playwright's own is not installed
curl -s localhost:3000/api/v1/capabilities
curl -s -X POST localhost:3000/api/v1/odds -H 'content-type: application/json' \
-d '{"calculator":"draw","pile":10,"targets":1,"draws":5}' # {"n":"1","d":"2"}: exactly 1/2
The simulation harness (sim/, Python) needs a locally fetched, pinned third-party simulator that is never committed; see docs/IP-COMPLIANCE.md and scripts/external-fetch.sh.
Documentation: STATUS (what exists, levels by cohort, next steps) · RELEASE-AUDIT · TEST-REPORT · MODEL-AND-DATA-CARDS · REPRODUCIBILITY · DEPLOY · IP-COMPLIANCE · DECISIONS.
Privacy and IP: no accounts, cookies or analytics; the pages do not send your run anywhere and the public API stores nothing. SLAY2 is unofficial and non-commercial, not affiliated with or endorsed by Mega Crit; it ships no game art, card text or audio (the card catalog holds facts only: ids, names, class, rarity, cost, type and SLAY2's own role tags). The code has no licence file yet (an owner decision): it is visible, not licensed for reuse.
Start Claude Code
First: read CLAUDE.md and START-HERE.md. Then execute prompts/00-MASTER-ORCHESTRATOR.md and phase files 01–11 in numeric order, with tests, evidence and small GitHub commits. These are build instructions, not claims of work already accomplished.
Read docs/RESEARCH-BASELINE.md, docs/SOURCE-REGISTRY.md, docs/ARCHITECTURE-DECISIONS.md, and docs/VALIDATION-PLAN.md before coding.
Critical differences from SLAY (original game)
- STS2 released Early Access March 5, 2026; official game built using Godot with C#, with native mod support. The initial scope is the STS2 official default Steam channel v0.107.1; latest separately recorded beta v0.111.0 as checked October 2026; always re-verify on startup. Never combine main/beta data or training artifacts.
- Five characters: Ironclad, Silent, Regent, Necrobinder, Defect. Class-specific mechanics include Regent Stars/Forge/Sovereign Blade, Necrobinder Osty/Souls/Doom/Graveyard/Ethereal, Silent Sly, Defect orb systems, and Ironclad unique new cards.
- STS2 currently uses three acts, with alternate Act 1 Overgrowth / Underdocks, new Ancients and relic-choice systems. Highest current Ascension A10, not A20, and A10 requires an Act 3 double-boss clear. No STS1 Heart/keys/Act 4 assumptions.
- Solo first, with 2–4 player co-op a separately validated workstream: multiplayer-only cards, ally targeting, synchronized decisions, scaling, seat-level outcomes. Do not promise co-op parity until tested.
- First-ever tutorial/progression runs can have scripted rewards and Neow availability. Ancients/unlocks/Epoch Timeline and reward changes may alter observation and game state.
- Everything changes across Early Access builds: card values, prices, monsters, AI moves, rarities, event eligibility, RNG and even mod protocols. A model that works on the wrong patch should abstain or fall back.
Non-negotiable rules
- No fabricated facts: never claim simulations, trained models, win rates, parity, website deployments, or probability certainty without checked artifacts and explicit version, cohort and sample.
- Honest information: visible-only policy; raw mod/engine RNG state, future draw order, hidden map/rewards, scripted scenario seed and unseen intents MUST NOT leak to player-facing agents.
- Real game experimentation:
1000 combat playouts ≠ 1000 completed games. Validate full-run outcomes against real STS2 before trusting simulated win rates.
- Research-specific rights: Mega Crit retains game IP. Independent original UI art; no bundling extracted art/C# or bulk publishing copyrighted content. Verify data API terms/third-party repository licenses before copying.
- Hybrid intelligent coaching: exact probability calculators when valid; robust rules & search; neural networks only when measured better on fair matched held-out runs.
- No serverless training jobs: Vercel hosts the site and lightweight APIs, intensive training and full-run simulations run on local/isolated workers.
- Incremental resilience: commits after small passing steps, status/evidence update after every stage; resume from real repository state.
Primary research sources
Official: https://www.megacrit.com/press-kits/slay-the-spire-2/ ; https://www.megacrit.com/faq/ ; https://steamcommunity.com/app/2868840/announcements/
Facts: https://slaythespire.wiki.gg/wiki/Slay_the_Spire_2:Ascension ; https://sts.wiki/guides/card-rewards ; https://almanauts.com/slaythespire2/ ; https://simthespire.wordpress.com/
Reference API (terms/rights caution): https://github.com/ptrlrd/spire-codex
Simulator candidates: https://github.com/yizhang-dream/sts2-simulator ; https://github.com/zhiyue/sts2-rl-agent
Agent/bridge: https://github.com/cicero225/STS2FableBot ; https://github.com/Gennadiyev/STS2MCP ; https://github.com/BMingSY/sts2-ai-mcp
This repository is a research-driven prompt program. Claude Code must deliver executable artifacts and measured evidence, not only additional documents.
