uv workspace for paper-grounded D&D simulation and long-horizon campaign layer.
Implements the framework from Setting the DC: Tool-Grounded D&D Simulations to Test LLM Agents (NeurIPS 2025 Workshop). Paper © original authors; code in this repo is MIT (implementation only).
| Package | Path | Purpose |
|---|---|---|
dnd-tools |
packages/dnd-tools/ |
Paper implementation (frozen) — 30+ validated tools, seeded simulation, 6 metrics. See packages/dnd-tools/README.md for full reproduction docs. |
dnd-campaign |
packages/dnd-campaign/ |
Long-horizon campaign layer — CampaignState (snapshots/persistence/rest), CampaignTools (delegation), CampaignSession (multi-encounter), memory compaction. Depends on dnd-tools via workspace. |
packages/dnd-tools/src/dnd_tools/ # paper
packages/dnd-campaign/src/dnd_campaign/ # campaign
uv sync
uv run dnd-tools demo --seed 42 --turns 10
uv run dnd-campaign demo --seed 42 --turns 10uv run pytest
uv run ruff check . && uv run ruff format --check . && uv run ty check .All package tests live in packages/<name>/tests/. Paper reproduction (27 scenarios, heuristic baseline O 0.795 | A 0.552 | func_err 2.1%) is documented in packages/dnd-tools/README.md — not repeated here.
MIT — implementation only; paper © authors.
