Typed Tibia-rules simulator targeting WGSL for the OT Agent Arena benchmark

The open-source project tibia published a typed Tibia-rules simulator written in Spiral that compiles to C, Rust, and WGSL from a single codebase, designed to train and benchmark autonomous agents against established open-source bot software. As detailed in github.com's report, the system targets the OT Agent Arena benchmark to measure agent performance directly against established implementations like vBot and nExBot on a Canary server harness. Its codebase consists of 27 Spiral modules in its core source directory alongside standards checking tools, an arena probe suite across 15 files, and map extraction utilities for live environments.

This technical report examines the simulator's compiler architecture, its verification loop across GPU and live server runtimes, and the empirical findings gathered from hundreds of logged benchmark matches. The material serves systems researchers, reinforcement learning engineers, and emulator developers analyzing fidelity gaps between simulated game mechanics and live network clients. The data tracks how minute engine quirks like login invulnerability, monster surrounding behavior, and pathfinding delays alter agent win rates, exposing why simple deterministic heuristic rules continue to match or exceed neural policies across complex game states.

Typed Tibia-rules simulator targeting WGSL: what actually changed

The repository updated its full benchmark record through October 2026 after identifying that initial benchmark runs from October 13 inadvertently left vBot's HealBot and TargetBot active alongside test policies. The project withdrew all prior performance claims and substituted fresh measurements across controlled 30-run sets documented in pull request four, verifying every match log to exclude foreign bot interventions. The simulation model underwent iterative corrections against Canary server mechanics to account for a ten-second login protection window, occupancy-aware monster routing that surrounds the knight instead of queuing in single files, and cyclops swing tempos clocked at 2166 milliseconds.

Recent changes consolidated under the mirroring current arena bundle in pull request 17 further aligned client step formulas, knight dance steps, and ground-speed calculations. The maintainers added hunter arena mechanics on October 22, incorporating Canary server timing rules and projectile mechanics where arrows stop exclusively at designated blocker tiles. These adjustments lifted predicted hunter projectile impacts from 14.9 to 21.0 per run to mirror live measurements of 20.45 hits. Additionally, knight retreat steps were aligned to a 650-millisecond execution tempo to narrow structural positioning discrepancies in tight map corridors.

Typed Tibia-rules simulator targeting WGSL: how it works

The simulator's domain core models game rules using advanced functional typing, featuring phantom units for quantity management, generalized algebraic datatype witnesses for action readiness, and higher-kinded types for view projection. Spells index directly by character vocation, and existential spawn catalogs govern creature instantiation over scanned map layouts. Decisions route through policy modules, integer neural networks, or evolutionary search routines. From this single Spiral specification, the compiler generates C to serve as the ground-truth oracle, Rust for the training host harness, WGSL for parallelized GPU training paths, and Lua bindings executing inside OTClient for live evaluation.

Runtime execution models hunting encounters through standalone hunt simulations, multi-creature pack modules, and live client probes. The simulator predicts clear times and resource consumption by simulating vitality, coin economies, combat damage formulas, cooldown periods, and wire protocol state transitions. Live verification runs measure whether the simulated knight matches observed behavior when under threat from multiple spawns. The model also accounts for player perception, forcing the agent to act strictly upon the client's last rendered frame view while emulating Canary's one-second think timer and step follow buffers when monsters retreat out of immediate melee range.

Typed Tibia-rules simulator targeting WGSL: why it matters now

The benchmark reveals the persistent difficulty of transferring learned policies from high-speed game simulators into live server environments without encountering distribution shift. Stepping backward against two cyclopes initially demonstrated a four-potion savings within early iterations of the simulator; however, once monster surrounding logic was accurately implemented, simulated savings dropped to a single potion. When evaluated on live Canary servers, the retreat policy consumed 1.4 more potions than baseline rule execution, confirming that early apparent efficiency gains were artifacts of unmodeled surrounding mechanics and co-active third-party healing probes.

Furthermore, empirical testing demonstrated that neural networks and evolved heuristics failed to outperform tuned deterministic rules across tested configurations, including open fields, tight corridors, potion-limited trials, and respawning arenas. In dynamic hunting grounds featuring a 60-second creature respawn rate, a simple sweep heuristic moving toward the next due spawn outperformed every fixed or evolved routing policy. The findings show that complex reinforcement learning policies produce minimal marginal gains when underlying domain mechanics can be expressed through thresholded rules, challenging standard assumptions regarding the necessity of deep neural models in structured gaming environments.

Typed Tibia-rules simulator targeting WGSL: who is affected

Engineers developing bot architectures and automated testing agents on OpenTibia platforms face immediate practical impacts from the project's data. Benchmark runs show that the Spiral rules policy operates on par with vBot, drinking 14.4 potions per clear compared to 14.8 for vBot when fresh login protection is stripped equally from both runners. Earlier metrics suggesting a wider margin stemmed entirely from Canary terminating login invulnerability upon player speech, item usage, or movement, which penalized vBot by 9.1 opening hits against 1.6 for the Spiral probe before subsequent combat damage stabilized at identical rates.

Developers evaluating rival tools such as nExBot encounter similar parity. Controlled trials under pull request 12 with a 2.5-second post-login buffer showed potion consumption settling at 12.75 for Spiral retreat rules against 13.93 for nExBot, confirming that slight script variations around initial login dialogue heavily distort benchmark outcomes. In addition, cooperative hunting models require caution: live multi-agent runs documented in pull request 24 recorded 5.1 party potion usages against a simulated expectation of 7.56, demonstrating that cooperative duo dynamics diverge from isolated single-agent pack simulations.

Typed Tibia-rules simulator targeting WGSL: what to watch

Researchers must monitor unresolved fidelity gaps in hunter movement dynamics and chase mechanics. While live hunter clearing times sit at 55.9 seconds, the simulator predicts 60.5 seconds because live hunters travel roughly one tile per second when two to three tiles away from the player, outpacing simulated speeds of 0.44 to 0.66 tiles per second. In live play, an initial target leads the knight on an extended chase exceeding 20 tiles in half of all runs, frequently causing trailing hunters to lose contact and compromising the clear, whereas the simulator generates these extended pursuits in only 20 percent of seeds.

Engineers are also tracking the integration of Canary think timers and path recalculation logic under rule D32. Live competitors disable chase mode while adjacent to targets, prompting Canary to stall path replanning until the player's next one-second think cycle, which leaves the knight stationary for an extra step in 98 percent of live cases compared to 91 percent under the simulated rule. Because enabling rule D32 drops simulated clear rates to 53 percent against live success rates reaching 95 percent, the rule remains inactive while maintainers refine spatial pathing and retreat tile selection in corner pockets.