{
  "context": "Tests whether the agent produces a correct greenfield design document for a coding agent harness, covering architecture selection, loop design, tool constraints, error handling, evaluation separation, and output completeness according to the skill's prescribed structure and benchmarks.",
  "type": "weighted_checklist",
  "checklist": [
    {
      "name": "Architecture level stated",
      "description": "design.md explicitly states an architecture level (e.g. Level 0-4 or equivalent single/multi-agent decision) with a rationale",
      "max_score": 8
    },
    {
      "name": "Separate evaluator",
      "description": "design.md specifies that evaluation is architecturally separated from generation (a distinct evaluator component, not the generator self-evaluating)",
      "max_score": 10
    },
    {
      "name": "Loop pattern matched to duration",
      "description": "design.md selects a loop pattern appropriate for a minutes-to-hours quality-critical coding task (Generator-Critic or Build-Verify-Fix, not just ReAct)",
      "max_score": 8
    },
    {
      "name": "Generator-Critic iteration cap",
      "description": "design.md specifies an iteration cap of 3 (or 2-3) for the generator-critic or refinement loop — not an uncapped loop",
      "max_score": 8
    },
    {
      "name": "Tool count within budget",
      "description": "design.md specifies a tool count that falls within the range of 8-12 tools per agent (or explicitly states tools must stay under 20% of context budget)",
      "max_score": 8
    },
    {
      "name": "Content-based file editing",
      "description": "design.md specifies file editing using content-based or string-replacement strategy (str_replace or equivalent) — NOT line-number-based patches",
      "max_score": 8
    },
    {
      "name": "Absolute filepaths required",
      "description": "design.md specifies that agents must use absolute filepaths (not relative paths) for file operations",
      "max_score": 8
    },
    {
      "name": "Architectural loop detection",
      "description": "design.md includes a loop detection or LoopGuard mechanism implemented architecturally (not relying solely on the agent's own judgment)",
      "max_score": 10
    },
    {
      "name": "Layered termination mechanisms",
      "description": "design.md specifies at least two termination mechanisms beyond the task completing (e.g. hard iteration cap AND wall-clock timeout, or objective verification)",
      "max_score": 8
    },
    {
      "name": "Context management strategy",
      "description": "design.md specifies a context management strategy matched to the expected task duration (e.g. FIC at phase boundaries for 30min-3h tasks)",
      "max_score": 8
    },
    {
      "name": "Output sections complete",
      "description": "design.md contains at least 8 of these sections: Requirements Summary, Architecture Decision, Loop Design, Action Space, Observation Formatting Strategy, Context Management Strategy, Evaluation Design, Error Handling & Recovery, Prompt Architecture, Decomposition Strategy, Complexity Budget / Simplification Plan, Key Metrics to Track, Open Questions",
      "max_score": 8
    },
    {
      "name": "Reasoning budget differentiated",
      "description": "design.md recommends differentiated reasoning budget (higher reasoning for planning/verification phases, lower for implementation) rather than uniform reasoning level throughout",
      "max_score": 8
    }
  ]
}