feat: add ART MCP-RL taskset adapter

[frvade]

Description

Adds a Verifiers v1 taskset adapter for OpenPipe ART MCP-RL scenario files.

The adapter:

• loads ART JSON/JSONL scenarios as typed Verifiers v1 tasks;
• preserves scenario metadata and difficulty;
• exports tasks back to ART-compatible rows;
• validates malformed tasks and unsupported input shapes.

This targets the two-way ART ↔ Verifiers portability track in the ART-E / ART bounty.

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Documentation update
• Test improvement

Testing

• All existing tests pass when running uv run pytest locally.
• New tests have been added to cover the changes

Validation performed:

• uvx ruff check verifiers/v1/tasksets/art_mcp — passed
• uvx ruff format --check verifiers/v1/tasksets/art_mcp — passed
• python -m compileall for the adapter — passed
• Round-trip smoke test using ART’s checked-in mcp_balldontlie/scenarios/val.jsonl fixture — all 8 scenarios loaded and exported without losing task or difficulty

The full test suite was not run because the local environment cannot build the existing pycosat dependency without a C compiler. No unit tests were added because the repository’s AGENTS.md explicitly asks contributors not to add unit tests for v1 changes and recommends temporary validation scripts instead.

Checklist

• My code follows the style guidelines of this project as outlined in AGENTS.md
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• Any dependent changes have been merged and published

Additional Notes

No dependencies or configuration files are changed. Documentation was not modified because the new adapter is self-contained and does not alter existing APIs.

---

Note

Low Risk
Additive, isolated taskset adapter with no changes to existing APIs, auth, or runtime paths; failures are limited to explicit validation errors on malformed scenario files.

Overview
Adds a new verifiers/v1/tasksets/art_mcp package so OpenPipe ART MCP-RL scenario files can be used as native Verifiers v1 tasks.

ArtMCPTaskset reads a configured file path and builds ArtMCPTask instances whose prompts come from each row’s task, with optional system_prompt on the taskset config. ArtMCPTaskData keeps source_task, difficulty, and extra fields in art_metadata for round-trip export.

load_art_rows / _read_rows accept JSONL, a top-level JSON array, or a JSON object with scenarios or tasks, and reject bad shapes or rows (missing non-empty task, invalid difficulty). art_rows_from_tasks writes ART-compatible dicts back from loaded tasks.

^{Reviewed by Cursor Bugbot for commit 9e6ade2. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Add ART MCP-RL taskset adapter with JSONL and JSON file support

• Adds a new ArtMCPTaskset under verifiers/v1/tasksets/art_mcp/ that loads ART scenario files and constructs ArtMCPTask instances from them.
• Supports .jsonl files (one JSON object per line) and .json files (top-level list or dict with a scenarios/tasks key); raises ValueError for unsupported shapes or invalid rows.
• Each task carries the original ART task string as its prompt, an integer difficulty (default 1), an optional system prompt from config, and arbitrary ART metadata preserved on ArtMCPTaskData.
• Includes art_rows_from_tasks to convert in-memory ArtMCPTask instances back to ART-compatible dicts for export.

^{Macroscope summarized 9e6ade2.}

[cursor]

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

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

^{Reviewed by Cursor Bugbot for commit a1c92d0. Configure here.}