fix(grader): support 'arguments' key for tool calls in task_workflow

[mgoulart]

Problem

The grader in task_workflow.md checks whether the agent read config.json by inspecting tool call parameters:

params = item.get("params", {})

This only works for agents that serialize tool parameters under the "params" key (Cursor, Windsurf). OpenClaw and Claude Code use "arguments" instead — matching the OpenAI tool call spec. As a result, read_config always scores 0 for those agents even when they correctly read the file.

Fix

Fall back to "arguments" when "params" is absent:

# Support both "params" (Cursor/Windsurf) and "arguments" (OpenClaw/Claude Code)
params = item.get("params", item.get("arguments", {}))

Verified

Confirmed by running task_10_workflow (pre-rename) against kimi-k2p5 on Fireworks. The agent correctly read config.json — visible in the transcript — but scored 0 on read_config before this fix and 1.0 after.

🤖 Generated with Claude Code

[kilo-code-bot]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

The fix correctly handles both "params" and "arguments" key formats for tool call parameters, matching the OpenAI spec used by Claude Code/OpenClaw while preserving compatibility with Cursor/Windsurf. The fallback chain item.get("params", item.get("arguments", {})) is safe and idiomatic.

Files Reviewed (1 file)

• tasks/task_workflow.md

---

_{Reviewed by claude-4.6-sonnet-20260217 · 58,447 tokens}

[olearycrew]

@mgoulart thanks for this fix!