Add AgentBoard BabyAI environment (#32)

## Summary

Add a self-contained AgentBoard BabyAI environment using the native
Verifiers 0.4 taskset and environment APIs.

- Bundle the text-game runtime, prompt, and 112 annotated episodes; no
AgentBoard checkout is needed.
- Preserve the existing deterministic 84/28 train/eval split across 28
task families, progress reward, action parsing, and grounding/success
metrics.
- Include a portable Prime Inference evaluation configuration and
concise setup instructions.
- Include upstream attribution and the separate runtime/prompt (Apache
2.0) and annotation (GPL 2.0) licenses.

Published environment: [bhoy1/agentboard-babyai, version
0.3.0](https://app.primeintellect.ai/dashboard/environments/bhoy1/agentboard-babyai).

## Verification

- Scoped pre-commit checks: Ruff lint/format, Markdown formatting/lint,
and Git LFS checks all passed.
- Repository package checks: 2 passed, including package build.
- Existing external regression suite: 23 passed and 1 optional upstream
comparison skipped. Includes 38 golden trajectory replays, all 112
initial observations, turn-limit behavior, concurrent rollout isolation,
and cleanup.
- Real Prime Inference evaluation from this contribution: Qwen3.5-9B, 1
held-out task with 2 rollouts, thinking disabled, 20 turns, 1,024-token
output budget. Both completed without harness errors, with reward 0.25
and stop condition `max_turns`.

The contribution's [evaluation
run](https://app.primeintellect.ai/dashboard/evaluations/savi3ix7zcjlua3x26nxxh7u)
was created.


<!-- Macroscope's pull request summary starts here -->
<!-- Macroscope will only edit the content between these invisible
markers, and the markers themselves will not be visible in the GitHub
rendered markdown. -->
<!-- If you delete either of the start / end markers from your PR's
description, Macroscope will append its summary at the bottom of the
description. -->
> [!NOTE]
> ### Add AgentBoard BabyAI environment package with text simulator and
eval config
> - Adds a new `agentboard-babyai` package with a
Gymnasium/MiniGrid-based text simulator (`BabyAI` in
[agentboard_babyai_runtime.py](https://github.com/PrimeIntellect-ai/residency-environments/pull/32/files#diff-be425817e801ab23a50f90676c99504ff1476095da3d7a460d535ff253c99181))
that validates high-level actions, tracks progress reward, and reports
success, failure, or step-limit termination
> - Adds a Verifiers integration
([agentboard_babyai.py](https://github.com/PrimeIntellect-ai/residency-environments/pull/32/files#diff-cfd10d39b442bf92beb1f2b3a6c64379b762914f1dfb8a6f2b87f70de52d522b))
that builds prompts from the packaged JSON asset, parses model replies
into single normalized actions, and drives multi-turn rollouts with
metrics for success, grounding, turns, and invalid actions
> - Supports `all`, `train_3_per_subtask`, and `eval_1_per_subtask`
splits with a configurable base seed (default 1234)
> - Includes an eval config
([eval.toml](https://github.com/PrimeIntellect-ai/residency-environments/pull/32/files#diff-650fda196fe27f64843882f04187d9520ba76efb843a7f3d84c2f6099ba6e808))
selecting Qwen/Qwen3.5-9B with thinking disabled, and packages task
data, prompt assets, and upstream AgentBoard provenance docs
> - Risk: the simulator reads MiniGrid internals via `env.unwrapped` and
expresses the 2D cross-product test as a scalar determinant; the
ordered-door verifiers (`_BeforeDoors`, `_AfterDoors`) changed failure
handling, so check those if port parity matters
>
> <!-- Macroscope's review summary starts here -->
>
> <sup><a href="https://app.macroscope.com">Macroscope</a> summarized
7a8d4a3.</sup>
> <!-- Macroscope's review summary ends here -->
>
<!-- Macroscope's pull request summary ends here -->

Author: Bhoy1

configs/agentboard-babyai/eval.toml

@@ -0,0 +1,37 @@
+model = "Qwen/Qwen3.5-9B"
+num_rollouts = 1
+max_concurrent = 2
+push = true
+output_dir = "outputs/agentboard-babyai"
+
+[env.taskset]
+id = "agentboard-babyai"
+split = "eval_1_per_subtask"
+base_seed = 1234
+
+[env.agent]
+max_turns = 20
+
+[env.agent.runtime]
+type = "subprocess"
+
+[env.agent.timeout]
+setup = 120
+rollout = 300
+
+[client]
+type = "eval"
+base_url = "https://api.pinference.ai/api/v1"
+api_key_var = "PRIME_API_KEY"
+
+[client.timeout]
+read = 120
+
+[sampling]
+temperature = 0.7
+max_tokens = 1024
+top_k = -1
+min_p = 0.0
+
+[sampling.chat_template_kwargs]
+enable_thinking = false

environments/agentboard-babyai/NOTICE.md

@@ -0,0 +1,21 @@
+# Third-Party Notices
+
+This environment contains material adapted from AgentBoard:
+
+- Project: AgentBoard
+- Repository: https://github.com/hkust-nlp/AgentBoard
+- Commit: `bb7255e2daf1989069a186dad9e53f70680961db`
+- Authors: Chang Ma, Junlei Zhang, Zhihao Zhu, Cheng Yang, Yujiu Yang,
+  Yaohui Jin, Zhenzhong Lan, Lingpeng Kong, and Junxian He
+
+`agentboard_babyai_runtime.py` and
+`assets/babyai_vanilla_prompt.json` are adapted from
+AgentBoard code/configuration distributed under Apache License 2.0. The runtime
+was extracted from the larger AgentBoard framework; registry, rendering, and
+unrelated dependency hooks were removed. See `licenses/Apache-2.0.txt`.
+
+`assets/test.jsonl` is AgentBoard dataset material
+distributed under GNU General Public License version 2. See
+`licenses/GPL-2.0.txt`.
+
+The Verifiers integration consists of project-specific additions.

environments/agentboard-babyai/README.md

@@ -0,0 +1,20 @@
+# AgentBoard BabyAI
+
+A self-contained, text-based BabyAI environment for navigation and object manipulation, adapted from AgentBoard for Verifiers.
+
+Requires Python 3.11+ and Verifiers 0.4.x. The runtime, prompt, and 112 task annotations are bundled; an AgentBoard checkout is not required. The split is 84 training episodes and 28 held-out episodes across 28 task families.
+
+## Evaluation
+
+From this repository's root, install the package and use the supplied Prime Inference configuration. Authenticate first with `prime login` or `PRIME_API_KEY`.
+
+```bash
+uv pip install -e ./environments/agentboard-babyai
+uv run --no-sync vf-eval @ configs/agentboard-babyai/eval.toml -n 1 -r 2
+```
+
+Omit `-n 1 -r 2` to evaluate all 28 held-out episodes once. The configuration uses Qwen3.5-9B with thinking disabled, a 20-turn limit, and the tool-free `null` harness.
+
+## Licensing
+
+The AgentBoard runtime and prompt are under Apache 2.0; the task annotations are under GPL 2.0. See `NOTICE.md`, `UPSTREAM_AGENTBOARD.md`, and `licenses/` for attribution and license details.

environments/agentboard-babyai/UPSTREAM_AGENTBOARD.md

@@ -0,0 +1,26 @@
+# AgentBoard Provenance
+
+The self-contained BabyAI runtime and assets were frozen from AgentBoard commit:
+
+```text
+bb7255e2daf1989069a186dad9e53f70680961db
+```
+
+Source checksums before extraction:
+
+| AgentBoard source | SHA-256 |
+| --- | --- |
+| `agentboard/environment/babyai_env.py` | `9c6ed31d492772ff902d8bddcb4096199fb8a17d866b5d1841f173b1ebbf6b33` |
+| `data/babyai/test.jsonl` | `d63d3dec56a623ea1c381a0815b31d6cfc3f707c5693c6d66e49b47c053c25bc` |
+| `agentboard/prompts/VanillaAgent/babyai_vanilla_prompt.json` | `1a7e040e67617c8556beab948c78e4ca4d78494fa96f4fe9b6b02c118a208068` |
+
+The port intentionally preserves AgentBoard's textual observations, high-level
+action expansion, action validation, progress reward, completion behavior,
+task annotations, prompt, and seed mapping. Its parity suite compares all 112
+episodes against the frozen AgentBoard checkout.
+
+Two compatibility adaptations are localized in the packaged runtime. MiniGrid
+internal attributes are read through Gymnasium's `env.unwrapped`, while
+`reset()` and `step()` continue to use the wrapped environment. AgentBoard's
+2D `numpy.cross` test is expressed directly as the equivalent scalar
+determinant because newer NumPy releases reject two-element vector inputs.