feat(model-specific-tools): Add model-specific tool selection from TOML config

[gspeter-max]

Summary

Add model-specific tool selection from TOML config files. Enables different models to use different tool sets from the same environment.

Changes

• verifiers/utils/tool_registry.py (new): Tool registry with @register_tool(env_id, tool_name) decorator
• verifiers/utils/env_utils.py: Add tools parameter to load_environment()
• packages/verifiers-rl/verifiers_rl/scripts/train.py: Config tool loading and validation
• environments/tool_test/tool_test.py: Backward compatibility with tools parameter
• tests/test_env_utils_tools.py (new): Integration tests
• tests/test_tool_registry.py (new): Unit tests

Usage

[[env]]
id = "tool-test"
tools = ["tool_A", "tool_B"]  # Use only these tools

Testing

• ✅ All 10 tests passing
• ✅ Backward compatible (uses DEFAULT_TOOL_LIST when tools=None)
• ✅ String tool resolution via registry
• ✅ Proper error handling

---

Note

Medium Risk
Changes the environment loading path and introduces dynamic tool resolution based on strings, which could impact training/runtime behavior if configs or registrations are incorrect, though covered by new tests.

Overview
Adds a global, thread-safe tool registry (@register_tool, get_tools, etc.) and extends load_environment to accept tools as either callables or tool-name strings that are resolved after importing the environment module.

Updates vf-train to read optional env.tools from TOML (supporting both [[env]] and [env] formats) and pass it through, and updates the tool-test environment to register its tools, honor a provided tool subset (including empty list), and generate prompts based on the selected tools. Adds unit/integration tests covering tool registration, resolution, and error cases.

^{Written by Cursor Bugbot for commit f1d9215. This will update automatically on new commits. Configure here.}

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[cursor]

Cursor Bugbot has reviewed your changes and found 3 potential issues.

^{Bugbot Autofix is OFF. To automatically fix reported issues with Cloud Agents, enable Autofix in the Cursor dashboard.}

[cursor]

Registry lookup uses full env_id, mismatching namespace-prefixed IDs

Medium Severity

The get_tools call on line 99 passes the raw env_id (e.g., "primeintellect/tool-test") to the registry, but module import on line 78 strips the namespace prefix via .split("/")[-1]. The @register_tool decorators in environment files use the simple ID (e.g., "tool-test"), so the registry key won't match when a prefixed env_id is used. Existing RL configs conventionally use prefixed IDs like "primeintellect/wiki-search", so any config following that convention with tools specified will fail at tool resolution with a KeyError.

Additional Locations (2)

• verifiers/utils/env_utils.py#L77-L78
• environments/tool_test/tool_test.py#L9-L10

 

[cursor]

Documentation not updated for new tools parameter

Low Severity

The load_environment function signature changed to include a new tools parameter, but docs/reference.md (line 771) still shows the old signature vf.load_environment(env_id: str, **kwargs) -> Environment. This is a core user-facing API change that the project's documentation rules require to be reflected in the relevant docs.

 

^{Triggered by project rule: BugBot Instructions}

[cursor]

Broad except KeyError swallows non-tool KeyErrors from environments

Medium Severity

The outer except KeyError: raise at line 163 is intended to let tool-resolution KeyErrors propagate, but it catches all KeyErrors raised anywhere inside the try block — including from env_load_func(**env_args) on line 148. Before this PR, a KeyError from the environment's own load_environment() would be caught by except Exception and wrapped in a RuntimeError with helpful context (which environment failed). Now it bypasses that wrapping and propagates as a raw KeyError, losing the environment-identification context and degrading the debugging experience.