fix: Correct newline escaping in normalize script (#750)

scripts/normalize-code-blocks.py

@@ -0,0 +1,138 @@
+#!/usr/bin/env python3
+"""
+normalize-code-blocks.py — Fix inconsistent indentation in training data code blocks.
+
+When code blocks are embedded in JSONL as triple-quoted strings, indentation
+accumulates from the surrounding context. This script normalizes code block
+content using textwrap.dedent and consistent 4-space indentation.
+
+Usage:
+    python3 scripts/normalize-code-blocks.py training/data/preference_pairs.jsonl
+    python3 scripts/normalize-code-blocks.py --dry-run training/data/*.jsonl
+    python3 scripts/normalize-code-blocks.py --check training/data/*.jsonl  # CI mode
+"""
+import argparse
+import json
+import re
+import sys
+import textwrap
+from pathlib import Path
+
+# Matches ```python ... ``` or ``` ... ``` blocks inside string values
+CODE_BLOCK_RE = re.compile(
+    r"(?P<open>```(?:python|py|bash|sh|javascript|js|typescript|ts|go|rust|ruby)?\s*\n)"
+    r"(?P<code>.*?)"
+    r"(?P<close>```)",
+    re.DOTALL,
+)
+
+
+def normalize_code_block(match: re.Match) -> str:
+    """Normalize indentation in a single code block."""
+    open_tag = match.group("open")
+    code = match.group("code")
+    close_tag = match.group("close")
+
+    if not code.strip():
+        return match.group(0)
+
+    dedented = textwrap.dedent(code)
+
+    lines = dedented.split("\n")
+    while lines and not lines[0].strip():
+        lines.pop(0)
+    while lines and not lines[-1].strip():
+        lines.pop()
+
+    normalized = "\n".join(lines)
+    return f"{open_tag}{normalized}\n{close_tag}"
+
+
+def process_line(line: str) -> tuple:
+    """Process a single JSONL line. Returns (new_line, num_fixes)."""
+    try:
+        obj = json.loads(line)
+    except json.JSONDecodeError:
+        return line, 0
+
+    fixes = 0
+
+    def fix_strings(obj):
+        nonlocal fixes
+        if isinstance(obj, str):
+            original = obj
+            fixed = CODE_BLOCK_RE.sub(normalize_code_block, obj)
+            if fixed != original:
+                fixes += 1
+            return fixed
+        elif isinstance(obj, dict):
+            return {k: fix_strings(v) for k, v in obj.items()}
+        elif isinstance(obj, list):
+            return [fix_strings(item) for item in obj]
+        return obj
+
+    fixed_obj = fix_strings(obj)
+    return json.dumps(fixed_obj, ensure_ascii=False) + "\n", fixes
+
+
+def process_file(filepath: str, dry_run: bool = False) -> dict:
+    """Process a single JSONL file. Returns stats dict."""
+    path = Path(filepath)
+    if not path.exists():
+        return {"file": str(filepath), "error": "not found", "fixes": 0, "lines": 0}
+
+    lines = path.read_text(encoding="utf-8").splitlines()
+    fixed_lines = []
+    total_fixes = 0
+
+    for line in lines:
+        if not line.strip():
+            fixed_lines.append(line)
+            continue
+        new_line, fixes = process_line(line)
+        fixed_lines.append(new_line.rstrip("\n"))
+        total_fixes += fixes
+
+    if total_fixes > 0 and not dry_run:
+        path.write_text("\n".join(fixed_lines) + "\n", encoding="utf-8")
+
+    return {
+        "file": str(filepath),
+        "lines": len(lines),
+        "fixes": total_fixes,
+        "changed": total_fixes > 0,
+    }
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Normalize code block indentation in JSONL training data"
+    )
+    parser.add_argument("files", nargs="+", help="JSONL files to process")
+    parser.add_argument("--dry-run", action="store_true", help="Show changes without writing")
+    parser.add_argument("--check", action="store_true", help="CI mode: exit 1 if fixes needed")
+    args = parser.parse_args()
+
+    total_fixes = 0
+    results = []
+
+    for filepath in args.files:
+        result = process_file(filepath, dry_run=args.dry_run or args.check)
+        results.append(result)
+        total_fixes += result["fixes"]
+
+        if result["fixes"] > 0:
+            status = "FIXED" if not args.dry_run and not args.check else "WOULD FIX"
+            print(f"  {status}: {result['file']} — {result['fixes']} code blocks normalized")
+        else:
+            print(f"  OK: {result['file']}")
+
+    print(f"\nTotal: {total_fixes} code blocks normalized across {len(results)} files")
+
+    if args.check and total_fixes > 0:
+        print("FAIL: Code block indentation issues found. Run without --check to fix.")
+        sys.exit(1)
+
+
+if __name__ == "__main__":
+    main()

tests/test_normalize_code_blocks.py

@@ -0,0 +1,151 @@
+#!/usr/bin/env python3
+"""Tests for normalize-code-blocks.py — issue #750"""
+import json
+import sys
+import tempfile
+from pathlib import Path
+
+import pytest
+
+# Import from scripts/
+sys.path.insert(0, str(Path(__file__).resolve().parent.parent / "scripts"))
+from normalize_code_blocks import normalize_code_block, process_line, process_file, CODE_BLOCK_RE
+
+
+class TestCodeBlockRegex:
+    def test_matches_python_block(self):
+        text = "```python\nprint('hi')\n```"
+        assert CODE_BLOCK_RE.search(text)
+
+    def test_matches_plain_block(self):
+        text = "```\nsome code\n```"
+        assert CODE_BLOCK_RE.search(text)
+
+    def test_matches_bash_block(self):
+        text = "```bash\necho hello\n```"
+        assert CODE_BLOCK_RE.search(text)
+
+    def test_ignores_inline_backticks(self):
+        text = "Use `code` inline"
+        assert not CODE_BLOCK_RE.search(text)
+
+    def test_handles_multiline_code(self):
+        text = "```python\ndef foo():\n    return 1\n\ndef bar():\n    return 2\n```"
+        match = CODE_BLOCK_RE.search(text)
+        assert match
+        assert "def foo" in match.group("code")
+
+
+class TestNormalizeCodeBlock:
+    def test_strips_leading_indent(self):
+        match = CODE_BLOCK_RE.search("```python\n    print('hi')\n```")
+        result = normalize_code_block(match)
+        assert "    print" not in result
+        assert "print('hi')" in result
+
+    def test_dedents_mixed_indent(self):
+        code = "```python\n    def foo():\n        return 1\n    def bar():\n        return 2\n```"
+        match = CODE_BLOCK_RE.search(code)
+        result = normalize_code_block(match)
+        lines = result.split("\n")
+        # First non-tag line should have 0 indent
+        code_lines = [l for l in lines if l.strip() and not l.startswith("```")]
+        assert code_lines[0].startswith("def foo")
+
+    def test_strips_trailing_blank_lines(self):
+        match = CODE_BLOCK_RE.search("```python\nprint('hi')\n\n\n```")
+        result = normalize_code_block(match)
+        assert result.endswith("print('hi')\n```")
+
+    def test_preserves_language_tag(self):
+        match = CODE_BLOCK_RE.search("```python\n    x = 1\n```")
+        result = normalize_code_block(match)
+        assert result.startswith("```python")
+
+    def test_empty_block_unchanged(self):
+        match = CODE_BLOCK_RE.search("```python\n   \n```")
+        original = match.group(0)
+        result = normalize_code_block(match)
+        assert result == original
+
+    def test_diff_markers_preserved(self):
+        code = "```\n+def new_func():\n+    return 1\n-def old_func():\n-    return 0\n```"
+        match = CODE_BLOCK_RE.search(code)
+        result = normalize_code_block(match)
+        assert "+def new_func" in result
+        assert "-def old_func" in result
+
+
+class TestProcessLine:
+    def test_valid_json_no_code_blocks(self):
+        line = json.dumps({"prompt": "hello world"})
+        new_line, fixes = process_line(line)
+        assert fixes == 0
+
+    def test_valid_json_with_code_block(self):
+        obj = {"prompt": "Here is code:\n```python\n    x = 1\n```"}
+        line = json.dumps(obj)
+        new_line, fixes = process_line(line)
+        assert fixes == 1
+        parsed = json.loads(new_line)
+        assert "    x = 1" not in parsed["prompt"]
+
+    def test_nested_dict_code_blocks(self):
+        obj = {
+            "prompt": "code: ```python\n    a = 1\n```",
+            "chosen": "```python\n    b = 2\n```",
+        }
+        line = json.dumps(obj)
+        new_line, fixes = process_line(line)
+        assert fixes == 2
+
+    def test_invalid_json_returned_unchanged(self):
+        line = "{broken json"
+        new_line, fixes = process_line(line)
+        assert new_line == line
+        assert fixes == 0
+
+    def test_list_field_code_blocks(self):
+        obj = {"items": ["```python\n    x = 1\n```", "no code here"]}
+        line = json.dumps(obj)
+        new_line, fixes = process_line(line)
+        assert fixes == 1
+
+
+class TestProcessFile:
+    def test_fixes_file_in_place(self, tmp_path):
+        f = tmp_path / "test.jsonl"
+        lines = [
+            json.dumps({"prompt": "```python\n    x = 1\n```"}),
+            json.dumps({"prompt": "no code"}),
+        ]
+        f.write_text("\n".join(lines) + "\n")
+
+        result = process_file(str(f))
+        assert result["fixes"] == 1
+        assert result["lines"] == 2
+
+        # Verify file was actually modified
+        content = f.read_text()
+        assert "    x = 1" not in content
+
+    def test_dry_run_no_write(self, tmp_path):
+        f = tmp_path / "test.jsonl"
+        original = json.dumps({"prompt": "```python\n    x = 1\n```"})
+        f.write_text(original + "\n")
+
+        result = process_file(str(f), dry_run=True)
+        assert result["fixes"] == 1
+
+        # File unchanged
+        assert f.read_text().strip() == original
+
+    def test_missing_file(self, tmp_path):
+        result = process_file(str(tmp_path / "nope.jsonl"))
+        assert "error" in result
+
+    def test_clean_file_no_fixes(self, tmp_path):
+        f = tmp_path / "clean.jsonl"
+        f.write_text(json.dumps({"prompt": "no code blocks here"}) + "\n")
+        result = process_file(str(f))
+        assert result["fixes"] == 0