{
  "context": "Tests whether the agent correctly handles a large multi-file diff by splitting the review into backend and frontend slices, writing each diff to /tmp/, piping via stdin, passing PROMPT.md, and using --sandbox read-only for each invocation.",
  "type": "weighted_checklist",
  "checklist": [
    {
      "name": "Split into separate passes",
      "description": "Runs at least two separate codex exec invocations — one scoped to the backend directory and one to the frontend directory — rather than a single invocation covering all files",
      "max_score": 20
    },
    {
      "name": "Path-scoped git diff",
      "description": "Uses path arguments on git diff to restrict each pass (e.g., 'git diff main...HEAD -- backend' and '-- frontend')",
      "max_score": 15
    },
    {
      "name": "Write diff to /tmp/",
      "description": "Each diff is written to a file under /tmp/ before being passed to codex",
      "max_score": 10
    },
    {
      "name": "Pipe via stdin not inline",
      "description": "Each codex exec call receives the diff via stdin pipe (cat /tmp/... | codex exec ...), not as an inline argument",
      "max_score": 10
    },
    {
      "name": "PROMPT.md passed each time",
      "description": "Each codex exec invocation includes the contents of .claude/skills/codex-review/resources/PROMPT.md as its prompt",
      "max_score": 15
    },
    {
      "name": "--sandbox read-only flag",
      "description": "Both codex exec invocations include --sandbox read-only",
      "max_score": 10
    },
    {
      "name": "Output captured separately",
      "description": "Each pass captures its output to a separate file (e.g., /tmp/codex-review-backend.out and /tmp/codex-review-frontend.out)",
      "max_score": 10
    },
    {
      "name": "Branch vs main diff",
      "description": "Uses 'git diff main...HEAD' (three-dot syntax) to diff the feature branch against main",
      "max_score": 10
    }
  ]
}