inspect-openreward

Run OpenReward environments as Inspect evals.

Provides an Inspect-native Dataset, Scorer, and a session-lifecycle wrapper solver for any OpenReward environment, so you get Inspect's eval harness, transcript viewer, metrics, and model abstraction for free — and OpenReward's tools, tasks, and rewards surface as first-class Inspect primitives. The wrapper takes an arbitrary inner solver chain, so you can keep the default react-style loop or plug in your own scaffolding (system_message, basic_agent, react, custom @solver, …) without touching session management.

Install

uv venv
uv sync

Set OPENREWARD_API_KEY and whichever model provider keys you need (OPENAI_API_KEY, ANTHROPIC_API_KEY, etc.).

Quickstart

See ./src/example/terminal_bench_2_verified.py for a runnable version with both a default, and a customer solver chain, task. ./src/example/terminal_bench_2_verified_example_output.txt contains a sample output from running the example against Kimi K2.6 Reasoning.

Benchmark results

To validate the integration end-to-end, we ran all 87 of the Terminal Bench 2 Verified environment, with Anthropic Sonnet 4.5 driving the Claude Code toolset, scoring 55.0%. This lines up closely with our September 2025 benchmark of the same combination, with the ~5% uplift likely reflecting six months of model and harness improvements.

Custom solver chains

openreward_solver is a wrapper: it owns session open/close, prompt injection, tool conversion + installation, and reward capture, and then runs whatever inner solver chain you hand it. That means any Inspect solver composition slots in — chain together system_message, prompt_template, chain_of_thought, self_critique, use_tools(..., append=True), basic_agent, react, or your own @solver — and the OpenReward session-bound tools remain available throughout:

from inspect_ai.solver import chain, generate, system_message

@task
def terminal_bench_2_verified_custom() -> Task:
    env = OpenReward().environments.get(name="GeneralReasoning/terminal-bench-2-verified")
    return Task(
        dataset=openreward_dataset(env, split="test", limit=10),
        solver=openreward_solver(
            env,
            chain(
                system_message("Think carefully before each tool call."),
                generate(tool_calls="loop"),
            ),
            toolset="claude-code"
        ),
        scorer=openreward_scorer(),
    )

Reward / finished capture is baked into the installed tools, so it keeps working no matter how the inner chain drives tool calls (generate(...), execute_tools(...) inside basic_agent, a custom loop, …).

What each piece does

openreward_dataset(environment, split, limit=None, shuffle=False, seed=None)

Builds an Inspect Dataset from an OpenReward environment split. Each Sample carries the underlying OpenReward Task in metadata; the prompt is resolved lazily by the solver (so image prompts work, and there's no network round-trip at dataset-construction time).

openreward_solver(environment, solver=None, *, toolset=None, tool_choice="auto")

Session-lifecycle wrapper around an arbitrary inner solver chain. Per sample it:

1. Opens environment.session(task=..., toolset=toolset).
2. Fetches session.get_prompt() and injects it as the user message (text and image blocks both supported).
3. Converts session.list_tools() into Inspect tools, auto-detecting the provider from state.model.api and sanitising the JSON schema via openreward.sanitize_tool_schema, and installs them on state.tools / state.tool_choice.
4. Runs the inner solver inside the open session. If solver=None (the default), runs generate(tool_calls="loop") — the react-style loop. Pass a Solver or a list[Solver] (composed via chain(...)) to customise.
5. Captures the terminal reward / finished from tool outputs into state.metadata for the scorer to read — regardless of how the inner chain invokes tools.
6. Closes the session on teardown.

Arguments

• environment: the OpenReward Environment to open sessions against. The dataset should be built from the same environment.
• solver: inner solver (or list[Solver], normalised via inspect_ai.solver.chain). Defaults to generate(tool_calls="loop").
• toolset (keyword-only): optional OpenReward toolset name passed to environment.session(...) — e.g. "claude-code" for a harness-native bash tool surface in addition to the environment's own tools.
• tool_choice (keyword-only): passed through to Inspect's state.tool_choice.

Inner chains can layer on further tools via use_tools(extra_tools, append=True) — the session-bound tools installed by the wrapper remain available.

openreward_scorer()

@scorer that reads the terminal reward captured by openreward_solver and returns an Inspect Score. Metrics: mean() and stderr(). Samples that never produced a reward (e.g. ran out of turns) score 0.0.

openreward_tool_to_inspect(tool_spec, session, provider=None)

Low-level helper for converting a single OpenReward ToolSpec into an Inspect Tool. The openreward_solver wrapper is built on top of this; use it directly if you want to bypass the wrapper and manage the session yourself.

Model provider mapping

The solver maps Inspect's model-provider identifier to OpenReward's Provider enum for schema sanitisation: