def codeforge_reward_fn(prompts, completions, *, task_meta, **kwargs):
    """Score completions using CodeForge's full pipeline."""
    from codeforge.sandbox.sandbox import run_sandbox
    from codeforge.grounder import ground
    from codeforge.grader import compute_reward

    rewards = []
    for completion, meta in zip(completions, task_meta):
        # Extract code (strip markdown blocks if model adds them despite prompt)
        code = completion
        if "```python" in code:
            code = code.split("```python")[1].split("```")[0].strip()
        elif "```" in code:
            code = code.split("```")[1].split("```")[0].strip()
            
        files = {"main.py": code}

        # 1. Sandbox: ruff + mypy + pytest
        sandbox_files = {**files, **meta["hidden_tests"]}
        try:
            # CodeForge run_sandbox signature:
            result = run_sandbox(files=sandbox_files, tools=meta["tools"])
            sandbox_score = result.composite_score
        except Exception as e:
            print(f"Sandbox error: {e}")
            sandbox_score = 0.0

        # 2. AST Grounding: verify all imports/attrs are real
        try:
            report = ground("\n".join(files.values()))
            groundedness = report.groundedness
        except Exception as e:
            print(f"Grounding error: {e}")
            groundedness = 0.0

        # 3. Brier-calibrated reward
        reward = compute_reward(
            sandbox_score=sandbox_score,
            groundedness=groundedness,
            confidence=0.7,  # Default fixed confidence since model doesn't declare it here
        )
        rewards.append(reward)
        
    return rewards