Semantic Lint Rules for Pull Request Diffs

{
  "model": "jev-latest",
  "state": {
    "pr_title": "{{pr-title}}",
    "pr_description": "{{pr-description}}",
    "diff_hunk": "{{diff-hunk}}",
    "team_conventions": "{{team-conventions}}"
  },
  "questions": {
    "matches_description": {
      "type": "noul",
      "instructions": "The change in this hunk is something the PR description says the PR does"
    },
    "silent_behaviour_change": {
      "type": "noul",
      "instructions": "This hunk changes runtime behaviour for existing callers in a way the description does not mention"
    },
    "secret_or_credential": {
      "type": "noul",
      "instructions": "This hunk adds a value that looks like a password, API key, token, or private URL"
    },
    "convention_break": {
      "type": "choice",
      "instructions": "Which team convention from the state does this hunk break, if any?",
      "criteria": {
        "{{convention-a}}": "{{convention-a-description}}",
        "{{convention-b}}": "{{convention-b-description}}",
        "none": "Breaks none of the listed conventions"
      }
    },
    "needs_test": {
      "type": "noul",
      "instructions": "A careful reviewer would expect a test to accompany this change and none is present in the diff"
    },
    "review_priority": {
      "type": "score",
      "instructions": "How much human attention does this hunk deserve?",
      "criteria": [
        "Mechanical, rename or formatting",
        "Routine logic, low blast radius",
        "Touches shared code or data paths",
        "Touches auth, payments, migrations, or deletion"
      ]
    }
  }
}

Run one request per hunk, in parallel, on every PR. Block the merge when secret_or_credential is above 0.6. Post a comment when silent_behaviour_change is above 0.7 or matches_description is below 0.3. Order the human review queue by review_priority, highest first. This is a staging filter before the LLM reviewer, not a replacement for it; the hunks that score as mechanical can skip the expensive pass.