fix(textarena): honor reset(seed) so episodes are reproducible

[adithya-s-k]

What

TextArenaEnvironment.reset() accepts a seed argument but never uses it. It calls self._ta_env.reset(num_players=...) without seeding, so the episode (for Wordle, the secret word, chosen via the global random module) is non-deterministic on every reset. The client -> HTTP -> server plumbing already forwards seed into environment.reset(seed=...); only the wrapper was dropping it.

This makes the environment non-reproducible, and it specifically breaks GRPO-style training: all rollouts in a group must share the same episode (same word) for the group-relative advantage baseline to be meaningful. Today each rollout in a group draws a different word.

Repro (before)

from textarena_env.server.environment import TextArenaEnvironment
env = TextArenaEnvironment(env_id="Wordle-v0", num_players=1)
env.reset(seed=123); w1 = env._ta_env.state.game_state["secret_word"]
env.reset(seed=123); w2 = env._ta_env.state.game_state["secret_word"]
assert w1 == w2   # fails on main

Fix

TextArena's own reset(seed=...) does not apply the seed to word selection (Wordle uses random.choice(word_list)), so the wrapper seeds the process-global random (and numpy if present) immediately before the underlying reset:

• atomic under a process-wide lock so seed + selection can't interleave across concurrent sessions,
• prior RNG state restored afterwards so unseeded sessions sharing the process are undisturbed,
• seed also forwarded to the underlying reset for games/versions that consume it,
• seed=None keeps the existing random behaviour.

Tests

tests/envs/test_textarena_seed.py (skipped if textarena isn't installed):

• same seed -> same word, across repeated resets and across separate instances
• seed actually drives word selection
• unseeded reset still works
• a seeded reset restores the global RNG stream

Validated locally with textarena 0.7.4: 5 passed. ruff format/ruff check clean.

---

Note

Medium Risk
Touches reset/concurrency behavior for a shared server env used in training; logic is localized but global RNG save/restore must stay correct under concurrent reset calls.

Overview
TextArenaEnvironment.reset(seed=...) now drives episode selection (e.g. Wordle’s secret word) instead of ignoring the seed. TextArena picks episodes via the process-global random (and sometimes numpy) RNGs without applying reset(seed=...), so the wrapper seeds those RNGs immediately before the underlying reset, forwards seed when the game supports it, and leaves seed=None behavior unchanged.

Seeding runs under a process-wide lock so concurrent sessions can’t interleave seed + selection, and the prior RNG state is restored afterward so unseeded resets in the same process aren’t perturbed. This enables reproducible rollouts (e.g. GRPO groups sharing the same word).

Adds tests/envs/test_textarena_seed.py (skipped without textarena) for same-seed determinism across resets/instances, seed-driven variety, unseeded resets, and global RNG preservation.

^{Reviewed by Cursor Bugbot for commit 58723a9. Bugbot is set up for automated code reviews on this repo. Configure here.}

[bot-ci-comment]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

Want higher recall? High effort reviews run extra passes and find more bugs. A team admin can switch effort levels in the Cursor dashboard.

^{Reviewed by Cursor Bugbot for commit 58723a9. Configure here.}

[cursor]

Unseeded reset skips seed lock

Medium Severity

_seeded_reset only acquires _SEED_LOCK when seed is set, so an unseeded reset (and __init__'s direct _ta_env.reset) can run while another session has temporarily reseeded the process-global RNGs. With SUPPORTS_CONCURRENT_SESSIONS and a thread-pool server, that interleaving can steal draws from a seeded episode or break same-seed reproducibility.

Additional Locations (1)

• envs/textarena_env/server/environment.py#L138-L139

 

^{Reviewed by Cursor Bugbot for commit 58723a9. Configure here.}