Pathgen#

The beam search returns one winning landing: a position, an orientation, and the tag from movegen (the spin grade in the shipped engine). The runner cannot act on a landing directly; it needs inputs. Path generation, usually called pathgen, turns the chosen landing into an input string and runs once per decision.

Earlier stages run thousands of times per decision, so they must be cheap. Pathgen runs once and must be exact.

Glossary#

Shared engine vocabulary is defined in the Glossary, and landing-specific terms in Movegen. This page adds:

The contract#

Pathgen takes the winning landing, the piece's current position and orientation (or the swapped-in piece's spawn position if the decision begins with a hold swap), the board, and the runner's movement options. It returns an input path or no path.

Pathgen always has the full set of movement options, even when movegen used a speed restriction. The landing exists on the board, so any valid route to it is acceptable, including a floating route that is shorter than the one used to find it.

Searching for the fastest inputs#

Pathgen searches for the shortest route through movement states. Each state records the position, orientation, and spin so far. Available moves are the runner's inputs:

• Shift one cell left or right, or hold a direction to slide until blocked.
• Rotate clockwise, counterclockwise, or 180 degrees when allowed, using the ruleset's kick tables.
• Soft drop one row, sonic drop to the resting spot, or hard drop.

Each input has a cost, and the route with the lowest total cost wins. A ruleset decides how inputs are priced: charging a held shift or a drop for the distance it covers makes the search prefer the fastest route, while giving every input the same cost makes it prefer the route with the fewest inputs. If pathgen reaches a state by a cheaper route, it replaces the previous route. The search ends when a hard drop reaches the chosen landing. Any path that places a piece therefore ends with a hard drop that locks it exactly where expected. Tie-breaking is deterministic, so the same decision always produces the same path.

Reproducing the promised spin#

The scored spin grade is a promise: the game honors it only if the executed inputs produce the right rotation. Pathgen therefore tracks spin as part of the movement state and recomputes it after every rotation, using the same corner rule as the movegen grading pass. A route that reaches the landing with the wrong spin does not count as a solution. When the promised grade is a spin, the output must end with a rotation that produces it, or pathgen returns no path.

Both layers grade spins by the same corner rule, so the scored grade can be reproduced.

The movement queries for shifts, rotations with kicks, and downward movement come from fast-reachability, the same public library used by movegen.

Hold first#

If the decision begins with a hold swap, the hold input comes first. Movement then starts from the swapped-in piece's entry spot, as returned by the AI (see AI interface). A decision can be "hold, then place," a hold swap on its own, or a plain placement. A hold swap on its own consists only of the hold input.

When there is no path#

If no input sequence can reach the winning landing, pathgen returns no path rather than an incorrect route. The engine then answers with a plain hard drop instead. This should not happen in normal operation: movegen found the landing through valid movement, and pathgen has at least the same options. The promised spin is the one requirement that could still rule out every route.

The inputs the runner receives#

The path is a string of single-character inputs, executed left to right: