TextCraft for platoon

platoon-textcraft is platoon's first-party TextCraft extension: crafting environments for training tool-using and recursive LLM agents, served as OpenReward (ORS) environments, with platoon's named environments, the CodeAct prompts of platoon's TextCraft agents, and an export of TextCraft-Synth as a Harbor dataset.

• TextCraft comes from ADaPT (Prasad et al., Findings of NAACL 2024): craft Minecraft items by their recipes, with shallow crafting trees.
• TextCraft-Synth is platoon's synthetic variant: generated recipes with deeper crafting trees (up to 12 levels), abstract item names and difficulty tags. The RAO recipe trains on it.

Install

pip install 'platoon-textcraft[platoon]'

The platoon extra installs platoon 0.2 with its ors extra (platoon[ors]>=0.2,<0.3), for what imports platoon: the prompt builders, the named environments and the task generators. To train or evaluate with platoon, add its tinker extra: pip install 'platoon[tinker,ors]' 'platoon-textcraft[platoon]'. The environments themselves need only openreward and pydantic: platoon_textcraft.env imports neither platoon nor IPython, so any ORS client can serve them after pip install platoon-textcraft.

The configs below and the tests are in platoon's repository, not in the package. From a checkout of platoon, in its uv workspace:

uv sync --package platoon-textcraft --extra platoon     # the extension, platoon and the test tools

or into an environment of your own: uv pip install -e "extensions/textcraft[platoon]".

Environments

Task ids are <split>:<index>. The tools are craft(ingredients, target), get_info(items), view_inventory() and the terminal finish(message), which grades the episode: success (reward 1) when every target was crafted on top of the starting inventory and the message is not empty.

Subagents share their parent's inventory: the environment's platoon_fork(task_spec) gives a child its own targets (parsed from its goal, Craft the following items: 2x stick, ...) on the same inventory, and the child's finish grades its own targets (agents.child_grading: env).

Named environments

Installed, the extension registers textcraft@1 and textcraft-synth@1 in the platoon.environments entry-point group (entries.py). Each stands for the environment, the TextCraft prompts for platoon-codeact (platoon_textcraft.prompts:TextCraftPromptBuilder) and the val split for evaluation; textcraft-synth@1 also scores with platoon's RAO reward processor, platoon-delegation. A config names one with env:, picks the arm of the prompts with prompt_builder_options.variant (linear, recursive or depth_aware) and adds the rest:

environments:
  - env: textcraft-synth@1
    harness: platoon-codeact
    harness_options:
      prompt_builder_options: {variant: recursive}
    agents: {max_depth: null, budget: linear, child_grading: env}
    dataset_kwargs: {split_name: train-medium}

platoon env show textcraft-synth prints what it stands for, and platoon extensions check platoon-textcraft checks the extension's metadata. The default episode program, single, applies the rollout config's skip_subagent_reward_computation and propagate_root_success, in that order.

Configs

platoon train extensions/textcraft/configs/textcraft_train.yaml
platoon eval extensions/textcraft/configs/textcraft_synth_eval.yaml inference.model_path=tinker://...

The RAO paper's TextCraft-Synth experiments, its single agent against its recursive agent and its ablation of the subagents' rewards, are a recipe of their own: recipes/rao/.

Data

The task files ship in the package, under src/platoon_textcraft/data/, and data/SHA256SUMS pins them:

data/recipes/ holds the 860 recipes of original TextCraft: the recipe files of Minecraft: Java Edition 1.16.5 (the game's data/minecraft/recipes/), unmodified, as InventivetalentDev/minecraft-assets extracted them from the game. ADaPT's README sends TextCraft's users there for them; ADaPT's repository does not include them. TextCraft-Synth needs no recipe files: the environment builds its world in memory (seed 42, 25 items per domain and tier, 1,452 items).

Every task keeps its gold trajectory (misc.gold_trajectory), a sequence of crafts that reaches the goal. The environments never read it; the Harbor export builds each task's reference solution from it.

The task generators are not deterministic across runs: they draw each task's unrelated base materials from a Python set, so running them again gives the same goals in another order and with other starting inventories, unless they run under the same string-hash seed. The shipped files are the source of truth. To generate others, write them outside the package:

python -m platoon_textcraft.tasks --num_samples 10000 --eval_size 1000 --out /tmp/textcraft-tasks
python -m platoon_textcraft.synth_tasks --num_samples 10000 --eval_size 1000 --seed 42 --out /tmp/textcraft-synth-tasks
python -m platoon_textcraft.synth_recipe_generator --output-dir /tmp/textcraft-synth-recipes --seed 42 --items-per-tier 25

The last command writes the TextCraft-Synth world the environment builds as JSON files, to read.

Export as a Harbor dataset

platoon_textcraft.harbor writes TextCraft-Synth as a Harbor dataset: one task directory per task of the train and val splits, which runners that take Harbor tasks (Harbor itself, platoon's service: harbor) play in a container through a textcraft command:

python -m platoon_textcraft.harbor export --variant synth --out /data/textcraft-synth --version 1.0 \
  --source-revision "$(git rev-parse HEAD)"

The same inputs give the same bytes. The dataset published as apurvaga/textcraft-synth was exported before TextCraft moved here, so its card names the earlier command, python -m platoon.textcraft.harbor export; an export from this package gives the same tasks, byte for byte, and a card that differs only in where the tasks came from. The extension's tests check the first tasks of each split against the published content hashes.

Each task holds:

• instruction.md: the game description the linear agent's prompt opens with (without its strategy tips), the goal, and the four commands, which print what the environment's tools return;
• environment/: a Dockerfile whose build (python:3.12-slim, tmux and the textcraft command) is the same in every task, then copies of the game (harbor/game.py, standard library only), the recipes and the task's task.json;
• tests/test.sh: the verifier. It replays the game's action log (/app/textcraft/actions.jsonl) with its own copies of the game, the recipes and the task, and scores 1 only for an episode the rules allow that ends with finish with every target crafted. It reads the log only as a regular file of at most 32 MiB, reached without symbolic links. task.toml runs it in a separate verifier container (environment_mode = "separate"), which receives the log as an artifact; it also runs in the agent's container.
• solution/solve.sh: the gold trajectory as commands. The export plays every one with the game and fails unless it reaches the goal.

The dataset's root holds registry.json (textcraft-synth-train@1.0, textcraft-synth-val@1.0), task-index.json (each task's split, difficulty, crafting depth, source ids and Harbor content hash, which platoon's select: filters on), splits.json (the environment's split names, such as train-medium, to task names, for platoon's splits:), LICENSE and the dataset card README.md, which says how to run the tasks with Harbor and with platoon. The tasks have no network, so an agent cannot fetch the published solutions.

Modules

• env.py: the ORS environments (TextCraftEnvironment, TextCraftSynthEnvironment); recipe_loader.py, synth_recipe_generator.py, synth_recipe_loader.py: their recipes.
• prompts.py: the system prompts and action spaces of the linear, recursive and depth-aware agents.
• entries.py: the named environments.
• tasks.py, synth_tasks.py: task generation and loading.
• harbor/: the Harbor export (export.py, card.py, templates/) and the in-container game and verifier (game.py).

Test

uv sync --package platoon-textcraft --extra platoon
uv run --package platoon-textcraft pytest extensions/textcraft/tests -n auto --dist loadfile

The tests run on CPU and share platoon's test helpers (tests/). The golden replays of recursive CodeAct rollouts need the Qwen/Qwen3-4B-Instruct-2507 tokenizer in the Hugging Face cache, and skip without it.

License

MIT (LICENSE), platoon's license, for the code and the data: the task files, which platoon generated, and the recipe files in data/recipes/, which are Minecraft's (see Data). Minecraft is a trademark of Mojang; platoon is not affiliated with Mojang or Microsoft.

Citation

TextCraft comes from ADaPT; platoon generated TextCraft-Synth for Recursive Agent Optimization (RAO):

@inproceedings{prasad2024adapt,
  title     = {{ADaPT}: As-Needed Decomposition and Planning with Language Models},
  author    = {Prasad, Archiki and Koller, Alexander and Hartmann, Mareike and Clark, Peter and Sabharwal, Ashish
               and Bansal, Mohit and Khot, Tushar},
  booktitle = {Findings of the Association for Computational Linguistics: NAACL 2024},
  pages     = {4226--4252},
  year      = {2024},
  url       = {https://aclanthology.org/2024.findings-naacl.264/}
}

@article{gandhi2026rao,
  title   = {Recursive Agent Optimization},
  author  = {Gandhi, Apurva and Chakraborty, Satyaki and Wang, Xiangjun and Kumar, Aviral and Neubig, Graham},
  journal = {arXiv preprint arXiv:2605.06639},
  year    = {2026}
}