Roadmap
What the benchmark plans to do, read from the roadmap in its repository. Finished work moves to the changelog, and Status lists what works today.
Now
In progress
Find the Bugs We Have Not Hit Yet
Run more models for longer, and try in the live game what is so far only tested in code, starting with the 1.6.3 fixes:place_nearplacing a first Armory, purchases counted while the town eats and pays taxes, and the second try at the end-of-run save.Full-Length Runs
Keep long runs reliable. None of the 21 runs in the evaluation was cut short by the game: the reader's watchdog broke 16 garbage-collector stalls in 12 of them. The game now renders on the game machine's NVIDIA GPU, where only a do-nothing run has played so far.A Better Harness
Improve the tools, prompt and context where recorded runs show models getting stuck, make the two machines easier to set up, and play with the text-only menu guide that a fresh checkout uses, which no model has used yet.
Next
After the current work
Tool Fixes from the 1.6.2 Runs
set_workshopstill opened a building drawn in front of the workshop four times on The Sinkhole (three times a tall Armory), after trying the footprint's back corner and roof: try its side corners too.demolish_buildingby tile may hit a building in front of its target the same way, which needs a live test first.Other Maps
Scored benchmarks on more of the game's maps beyond Oasis by the Sea, Large Barren Desert and The Sinkhole, each with a person's reference game and a do-nothing baseline.
Later
Further out
Play Against Enemies
Single-player games against the game's computer-controlled lords, with military and defence scoring. The reader will need to tell the agent's buildings and troops from the opponents' and to read enemy messages. Games against other players stay refused.Other Setups
Game machines on other Linux distributions or Windows, and other window sizes and game builds.
Read from ROADMAP.md at commit cad91f7 (13 October 2026). View it on GitHub