{
    "kind": "meta",
    "name": "meta-codereview-current-diff",
    "always": false,
    "triggers": [
        "multi-reviewer diff",
        "codereview my diff",
        "三路 review",
        "审查当前 diff"
    ],
    "provenance": {
        "origin": "opensquilla-original",
        "license": "Apache-2.0"
    },
    "composition": {
        "steps": [
            {
                "id": "read_diff",
                "kind": "skill_exec",
                "with": {
                    "cwd": "{{ inputs.workspace_dir | default('.') }}",
                    "mode": "cached_fallback_worktree"
                },
                "skill": "git-diff"
            },
            {
                "id": "review_safety",
                "kind": "llm_chat",
                "with": {
                    "task": "You are the *safety reviewer* for a 3-reviewer codereview bundle.\nApply ONLY the rules below; do not invent additional concerns.\n\nDiff under review:\n---\n{{ outputs.read_diff | truncate(8000) }}\n---\n\nRules (in priority order):\n  1. CRITICAL if the diff introduces SQL injection (string\n     concatenation into SQL, unparameterized queries via .format\n     or f-strings into execute\/query).\n  2. CRITICAL if the diff introduces shell-injection\n     (subprocess with shell=True + user input, or os.system on\n     unvalidated strings).\n  3. CRITICAL if the diff hardcodes a credential (sk-…, ghp_…,\n     AKIA…, private key, password=… string literal).\n  4. WARNING if the diff templates user-controlled strings into\n     tool args \/ prompts WITHOUT `xml_escape` (G1.6 contract).\n  5. CLEAR if none of the above apply.\n\nReply with EXACTLY one line, no preamble:\n  CRITICAL: <one-sentence reason>\n  WARNING: <one-sentence reason>\n  CLEAR: no safety concerns found\n",
                    "system": "You are a precise code safety reviewer. Reply with exactly one requested verdict line."
                },
                "depends_on": [
                    "read_diff"
                ]
            },
            {
                "id": "review_tests",
                "kind": "llm_chat",
                "with": {
                    "task": "You are the *test-coverage reviewer* for a 3-reviewer codereview\nbundle. Apply ONLY the rules below.\n\nDiff under review:\n---\n{{ outputs.read_diff | truncate(8000) }}\n---\n\nRules:\n  1. MISSING_TESTS if the diff adds a new public function \/ class \/\n     protocol under `src\/` and does NOT add a corresponding test\n     under `tests\/`.\n  2. MISSING_TESTS if the diff is a bug fix (commit message style\n     `fix(...):`) and does NOT add a regression test.\n  3. PASS if refactor-only (no behaviour change), even without\n     new tests.\n  4. PASS if the diff is doc-only or config-only.\n\nReply with EXACTLY one line, no preamble:\n  MISSING_TESTS: <which functions\/classes lack tests>\n  PASS: tests adequate (or n\/a)\n",
                    "system": "You are a precise test-coverage reviewer. Reply with exactly one requested verdict line."
                },
                "depends_on": [
                    "read_diff"
                ]
            },
            {
                "id": "review_style",
                "kind": "llm_chat",
                "with": {
                    "task": "You are the *style \/ idiom reviewer* for a 3-reviewer codereview\nbundle. Look for project-specific anti-patterns.\n\nDiff under review:\n---\n{{ outputs.read_diff | truncate(8000) }}\n---\n\nAnti-patterns to flag (do not invent others):\n  - Functions > 80 source lines\n  - Conditional nesting > 4 deep\n  - Magic numeric literals without a named constant\n  - Stale TODO\/FIXME comments left in shipped code\n  - Bare `except:` clauses (no exception class)\n  - `print()` calls instead of structlog `log.info(...)`\n\nReply with EXACTLY one line, no preamble:\n  ANTIPATTERNS: <one-line list with line refs>\n  CLEAN: no style issues found\n",
                    "system": "You are a precise style reviewer. Reply with exactly one requested verdict line."
                },
                "depends_on": [
                    "read_diff"
                ]
            },
            {
                "id": "arbitrate",
                "kind": "llm_chat",
                "with": {
                    "task": "Three reviewers ran on the diff:\n  - Safety: {{ outputs.review_safety }}\n  - Tests:  {{ outputs.review_tests }}\n  - Style:  {{ outputs.review_style }}\n\nApply STRICTLY (higher rule wins; do NOT mix or soften):\n  1. If Safety starts with \"CRITICAL\" → final verdict BLOCK.\n     Pass through the safety reviewer's reason verbatim.\n  2. Else if Safety starts with \"WARNING\" OR Tests starts with\n     \"MISSING_TESTS\" → final verdict BLOCK_WITH_OVERRIDE\n     (user may proceed with explicit acknowledgement; pass the\n     warning(s) verbatim).\n  3. Else → PASS_WITH_NOTES. If Style starts with \"ANTIPATTERNS\",\n     append the style notes; if all three are clean, write\n     \"clean\" instead.\n\nReply with EXACTLY this structure on the first line, then\nadditional lines as needed:\n  BLOCK: <safety critical reason>\n  BLOCK_WITH_OVERRIDE: <warning(s); user must confirm>\n  PASS_WITH_NOTES: <style notes; or \"clean\">",
                    "system": "You arbitrate code review verdicts. Follow the priority rules exactly."
                },
                "depends_on": [
                    "review_safety",
                    "review_tests",
                    "review_style"
                ]
            }
        ]
    },
    "description": "Read the current uncommitted diff, run three independent reviewers (safety + tests-coverage + style) in parallel, then arbitrate a single BLOCK \/ BLOCK_WITH_OVERRIDE \/ PASS_WITH_NOTES verdict. Use before commit when you want a multi-perspective second-opinion instead of a single-reviewer agent loop.",
    "meta_priority": 65,
    "final_text_mode": "step:arbitrate"
}