fix: change default judge_backend from openclaw to api

[ScuttleBot]

Problem

The openclaw backend spins up a full OpenClaw agent with tools for judging. The agent was going rogue — running exec commands and exploring the filesystem instead of just returning JSON scores.

Evidence from live benchmark (Kimi K2.6):

• 128 'Failed to parse judge JSON response' warnings
• Judge transcripts show calls to find, ls, exploring workspace
• Model scored 0.8% (1/123) because judge couldn't parse anything
• Only automated tasks (no judge) passed

Solution

Change the default judge_backend from openclaw to api.

The api backend does direct API calls without spawning an agent with tools. This is more reliable for the simple scoring task.

Files Changed

• scripts/lib_grading.py — Change default parameter in grade_task() and _grade_llm_judge()

Testing

Need to run a benchmark with this fix to verify judge responses are now parseable JSON.

[kilo-code-bot]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

Solid, targeted fix. The change correctly addresses the root cause — switching the default judge_backend from openclaw to api in both grade_task() and _grade_llm_judge() eliminates the rogue agent behavior causing JSON parse failures. The two-line change is minimal and consistent.

Files Reviewed (1 file)

• scripts/lib_grading.py

---

Fix any issues in Kilo Cloud

---

_{Reviewed by claude-4.6-sonnet-20260217 · 73,409 tokens}