Workload · Agentic coding · 3D western action · Revision 1

Wild West Showdown 3D

A staged, long-horizon challenge to build a complete browser game while preserving a plan and project state across fresh sessions. The game combines an explorable western town, touch controls, DRAW-based duels, progression and a final boss.

Compare 1 attempt Task prompt

1 recorded attempt

Recorded attempt

Wild West Showdown 3D gameplay in the town, with the player HUD and a gunslinger nearby — Wild West Showdown 3D · TensorFold Native · gameplay capture

09–12 October 2026

Qwen 3.8 27B

TensorFold native 1.0.4 · Pi 1.1.0

Session partial Local run

The playable game was completed after manual continuation. Pi's final session reported 65/65 smoke checks; the deterministic runner stopped at M03 and its planned independent final review did not run.

Play the build Run sheet

Three.js licence notice

Task requirements & full brief

This is a staged challenge: the requested project files preserve plans and state across fresh Pi sessions, while a deterministic runner is intended to sequence Architect, Builder, Reviewer and Fixer roles. The retained run did not complete that autonomous protocol. The runner stopped during M03 after two unsuccessful sessions; the user continued manually. The final Pi session reported 65/65 smoke checks, but the runner's final review gate did not run. Session metrics are deliberately limited to the final manual completion session and are not pooled with earlier phases. No transcripts or private runner prompts are published.

Maintain GAME_SPEC.md, PLAN.md, STATE.md, DECISIONS.md and REVIEW.md as the project source of truth.
Build an explorable 3D western town with blocky characters and a responsive third-person camera.
Support desktop controls and first-class iPad/tablet multi-touch controls.
Implement fair DRAW-based duels, meaningful aiming, early-shot penalties and distinct opponents.
Add money, reputation, upgrades, a final gunslinger, victory state and restart.
Review the complete playthrough and record remaining real-device verification limits.

How to read this comparison

Compare the generated builds and the checks recorded for each attempt. There is no normalised quality score, and the cloud task’s cumulative token totals and end-to-end rate include repeated context, reasoning, tool calls and waiting — they are not comparable with local per-request decode or prefill rates. The method page sets out which figures can and cannot be set against each other.