@moishy/apb

A solver for the APB method (Athefre's Pair & Block), built on @moishy/cubing-core.

It produces the solution a human following APB would execute — every step labeled, ranked by how ergonomic it is to turn, not by move count.

Try it in your browser →

deno add jsr:@moishy/apb    # Deno — https://jsr.io/@moishy/apb
npm  install @moishy/apb    # Node — https://www.npmjs.com/package/@moishy/apb

Pre-1.0. The solver is verified end to end — every algset is audited against the lookup it is used with, and 540 solves across every replacement and extra all completed. What is not frozen yet is the public API; it settles at 1.0. See the changelog.

Use

import { apb } from "@moishy/apb";

const res = await apb.solve("R U2 F' L D B2 R' U F2 D' L2 B R2 U' F D2 B' L U2 R'");

res.solved; // true
res.solutionString; // "r2 U R2 U F B2 M' D' U r U2 R' U f R' f' ..."
res.solution.length; // 44
res.cost; // 46.45 — ergonomic cost, not move count

for (const seg of res.segments) {
  console.log(seg.unitId, seg.strategyId, seg.moves.length);
}

Nothing beyond the scramble is required; the method ships its own recommended settings.

The Method

APB solves a 2x2x3 block, adds a pair, orients edges, finishes the last slot, then does the last layer in one algorithm:

| Step | id | How It's Solved | | ------------------ | ---------- | --------------------------------------------------------------------------------------------------- | | 2x2x3 | block223 | Search (six strategies available; one enabled by default) | | BR Pair | brPair | 89-case algset | | Edge Orientation | eo | 11-case algset | | Last Extended Slot | lxs | 116-case algset | | Last Layer | zbll | 493 cases in one alg — 472 ZBLL, falling through to the 21 PLLs when the corners are already solved |

Optional replacements cover several steps at once — eoPair (BR pair + EO together), eodrLs, collEpll, ocllPll, backSlotEoLxs — and optional extras insert at a boundary or mid-alg (oll, zbls, winterSummerVariation). All are off by default; in compete mode a replacement is only used when it actually beats the normal route, so enabling one cannot make a solve worse.

2x2x3 Strategies

block223 is where the interesting search happens. Six strategies are registered; only fbDfdb is enabled by default. Mean block length in STM and mean wall-clock per solve, measured over 25 scrambles with dual color neutrality:

| Strategy | Time | STM | | ----------------------------------------------------------- | ------ | ------- | | fbDfdb — Roux first block, then a 527-case DF/DB alg | 0.02s | 9.8 | | direct — one search for the whole 2x2x3 | 1.43s | 7.4 | | cornerFirstFront / cornerFirstBack — 2x2x2, then extend | ~0.15s | 9.7 | | cross1Front / cross1Back — bottom line, then two pairs | ~0.25s | 12.7 |

direct finds the shortest blocks and beats the rest on every scramble tested, but costs ~70× more than the default. Enable it when block quality matters more than latency:

await apb.solve(scramble, {
  stepOptions: { block223: { enabledStrategies: ["fbDfdb", "direct"] } },
});

Worth knowing before you do: widening color neutrality is a cheaper way to get short blocks than changing strategy. fbDfdb at full CN averages 8.9-move blocks in 0.06s, where direct at dual CN averages 7.3 in 1.45s.

await apb.solve(scramble, { colorNeutrality: "full" });

Settings

Common ones:

await apb.solve(scramble, {
  colorNeutrality: "fixed", // "fixed" | "full" | Move[][]; default: dual-CN (8)
  moveCostModel: createDefaultMoveCostModel({ mode: "OH" }),
  lookahead: { depth: 1 }, // choose each step by what it leaves the next
  stepOptions: { block223: { forceStrategy: "fbDfdb" } },
  replacements: { eoPair: { enabled: true } },
}, {
  timeBudgetMs: 10_000,
});

Full reference: Solver Settings.

Also Exported

apbDefinition — the raw MethodDefinition, so a UI can generate its options form from the method itself and stay in sync (this is what the demo page does). Plus the piece groups APB's steps target (BLOCK223, AFTER_BR, F2L, BR_PAIR, LAST_SLOT, EO_EDGE_SLOTS) and the PieceRegion type, for reading a result.

The method-wiring helpers in src/geometry.ts (goal predicates, recognition signatures, lookup builders) are deliberately not exported: they are internals, and some are specific to APB's algsets. To build your own method, read that file as a template — see Adding a Method.

Documentation

• Getting Started
• Solver Settings
• SPEC.md — the method spec, step by step
• Adding a Method — this package is the reference implementation
• CHANGELOG.md — release history

License

MIT © Moshe Rosenberg