snapshot comparison in verification

codeflash/api/aiservice.py

@@ -46,6 +46,8 @@ class AiServiceClient:
         self.llm_call_counter = count(1)
         self.is_local = self.base_url == "http://localhost:8000"
         self.timeout: float | None = 300 if self.is_local else 90
+        # React components are larger (300+ lines of JSX) and need more LLM processing time
+        self.react_timeout: float | None = 300 if self.is_local else 180
 
     def get_next_sequence(self) -> int:
         """Get the next LLM call sequence number."""
@@ -203,7 +205,8 @@ class AiServiceClient:
         logger.debug(f"Sending optimize request: trace_id={trace_id}, n_candidates={payload['n_candidates']}")
 
         try:
-            response = self.make_ai_service_request("/optimize", payload=payload, timeout=self.timeout)
+            timeout = self.react_timeout if is_react_component else self.timeout
+            response = self.make_ai_service_request("/optimize", payload=payload, timeout=timeout)
         except requests.exceptions.RequestException as e:
             logger.exception(f"Error generating optimized candidates: {e}")
             ph("cli-optimize-error-caught", {"error": str(e)})
@@ -806,7 +809,8 @@ class AiServiceClient:
         # DEBUG: Print payload language field
         logger.debug(f"Sending testgen request with language='{payload['language']}', framework='{test_framework}'")
         try:
-            response = self.make_ai_service_request("/testgen", payload=payload, timeout=self.timeout)
+            timeout = self.react_timeout if is_react_component else self.timeout
+            response = self.make_ai_service_request("/testgen", payload=payload, timeout=timeout)
         except requests.exceptions.RequestException as e:
             logger.exception(f"Error generating tests: {e}")
             ph("cli-testgen-error-caught", {"error": str(e)})

codeflash/languages/javascript/comparator.py

@@ -143,6 +143,8 @@ def compare_test_results(
                 scope = TestDiffScope.STDOUT
             elif scope_str == "did_pass":
                 scope = TestDiffScope.DID_PASS
+            elif scope_str == "dom_snapshot":
+                scope = TestDiffScope.DOM_SNAPSHOT
 
             test_info = diff.get("test_info", {})
             # Build a test identifier string for JavaScript tests

codeflash/languages/javascript/edit_tests.py

@@ -210,7 +210,10 @@ def normalize_codeflash_imports(source: str) -> str:
     # Replace CommonJS require
     source = _CODEFLASH_REQUIRE_PATTERN.sub(r"\1 \2 = require('codeflash')", source)
     # Replace ES module import
-    return _CODEFLASH_IMPORT_PATTERN.sub(r"import \1 from 'codeflash'", source)
+    source = _CODEFLASH_IMPORT_PATTERN.sub(r"import \1 from 'codeflash'", source)
+    # Fix outdated @testing-library/jest-dom import paths (v6+ removed /extend-expect subpath)
+    source = source.replace("@testing-library/jest-dom/extend-expect", "@testing-library/jest-dom")
+    return source
 
 
 # Author: ali <mohammed18200118@gmail.com>

codeflash/languages/javascript/frameworks/react/profiler.py

@@ -72,42 +72,37 @@ def instrument_component_with_profiler(source: str, component_name: str, analyze
         if computed is not None:
             replacements.append(computed)
 
+    # Work entirely in bytes to avoid byte-offset vs char-index mismatches
+    # (tree-sitter returns byte offsets; non-ASCII chars cause divergence from Python string indices)
     if replacements:
-        # Reconstruct result in a single pass
+        byte_parts: list[bytes] = []
-        parts: list[str] = []
         prev = 0
         for start, end, wrapped in replacements:
-            # Use original source slices (string indices expected by original logic)
+            byte_parts.append(source_bytes[prev:start])
-            parts.append(source[prev:start])
+            byte_parts.append(wrapped.encode("utf-8"))
-            parts.append(wrapped)
             prev = end
-        parts.append(source[prev:])
+        byte_parts.append(source_bytes[prev:])
-        result = "".join(parts)
+        result_bytes = b"".join(byte_parts)
     else:
-        result = source
+        result_bytes = source_bytes
-
-    # Add render counter code at the top (after imports) using the already-parsed tree
 
     # Add render counter code at the top (after imports)
     counter_code = generate_render_counter_code(component_name)
 
-    # Inline logic similar to _insert_after_imports but reuse existing `tree` to avoid re-parsing
     last_import_end = 0
     for child in tree.root_node.children:
         if child.type == "import_statement":
             last_import_end = child.end_byte
 
     insert_pos = last_import_end
-    while insert_pos < len(result) and result[insert_pos] != "\n":
+    while insert_pos < len(result_bytes) and result_bytes[insert_pos : insert_pos + 1] != b"\n":
         insert_pos += 1
-    if insert_pos < len(result):
+    if insert_pos < len(result_bytes):
         insert_pos += 1  # skip the newline
 
-    result = result[:insert_pos] + "\n" + counter_code + "\n\n" + result[insert_pos:]
+    counter_bytes = ("\n" + counter_code + "\n\n").encode("utf-8")
+    result = (result_bytes[:insert_pos] + counter_bytes + result_bytes[insert_pos:]).decode("utf-8")
 
-    # Ensure React is imported
-
-    # Ensure React is imported
     return _ensure_react_import(result)
 
 
@@ -317,7 +312,8 @@ def _wrap_return_with_profiler(source: str, return_node: Node, profiler_id: str,
         f"</React.Profiler>"
     )
 
-    return source[:jsx_start] + wrapped + source[jsx_end:]
+    # Use byte-level splicing to avoid byte-offset vs char-index mismatches
+    return (source_bytes[:jsx_start] + wrapped.encode("utf-8") + source_bytes[jsx_end:]).decode("utf-8")
 
 
 def _insert_after_imports(source: str, code: str, analyzer: TreeSitterAnalyzer) -> str:
@@ -421,7 +417,7 @@ let _codeflash_render_count_{safe_name} = 0;
 if (typeof beforeEach !== 'undefined') {{
   beforeEach(() => {{ _codeflash_render_count_{safe_name} = 0; }});
 }}
-function _codeflashOnRender_{safe_name}(id, phase, actualDuration, baseDuration) {{
+function _codeflashOnRender_{safe_name}(id: any, phase: any, actualDuration: any, baseDuration: any) {{
   _codeflash_render_count_{safe_name}++;
   console.log(`!######{marker_prefix}:${{id}}:${{phase}}:${{actualDuration}}:${{baseDuration}}:${{_codeflash_render_count_{safe_name}}}######!`);
 }}"""

codeflash/languages/javascript/frameworks/react/testgen.py

@@ -98,6 +98,15 @@ def post_process_react_tests(test_source: str, component_info: ReactComponentInf
     """
     result = test_source
 
+    # Fix outdated @testing-library/jest-dom import paths (v6+ removed /extend-expect subpath)
+    result = result.replace("@testing-library/jest-dom/extend-expect", "@testing-library/jest-dom")
+    # Also fix require() variant
+    result = re.sub(
+        r"""require\s*\(\s*['"]@testing-library/jest-dom/extend-expect['"]\s*\)""",
+        "require('@testing-library/jest-dom')",
+        result,
+    )
+
     # Ensure testing-library import
     if "@testing-library/react" not in result:
         result = "import { render, screen, act } from '@testing-library/react';\n" + result
@@ -168,14 +177,14 @@ def post_process_react_tests(test_source: str, component_info: ReactComponentInf
     # This gives per-interaction A/B signal without the LLM needing to know about it.
     result = inject_interaction_markers(result)
 
-    # Warn if no tests contain interaction calls — mount-phase only markers are
+    # If no tests contain interaction calls, auto-inject a rerender fallback so
-    # not useful for measuring optimization effectiveness.
+    # that EVERY React perf test produces at least one update-phase marker.
     if not has_react_test_interactions(result):
         logger.warning(
-            "[REACT] Generated tests for %s contain no interactions (fireEvent, userEvent, rerender). "
+            "[REACT] Generated tests for %s contain no interactions — auto-injecting rerender fallback.",
-            "Tests will produce only mount-phase markers which cannot measure optimization improvements.",
             component_info.function_name,
         )
+        result = _inject_rerender_fallback(result, component_info.function_name)
 
     # Check interaction density — fewer than MIN_INTERACTION_CALLS total interactions
     # means the test is unlikely to produce enough update-phase renders for reliable measurement.
@@ -221,6 +230,48 @@ def _extract_interaction_label(call_text: str) -> str:
     return m.group(1) if m else "interaction"
 
 
+def inject_dom_snapshot_calls(test_source: str) -> str:
+    """Inject codeflash.snapshotDOM() calls after each user interaction in behavior mode.
+
+    Only active when `captureRender` (not `captureRenderPerf`) is present,
+    meaning the test is running in behavioral verification mode.
+
+    After each fireEvent.*, userEvent.*, or rerender() call, inserts:
+        codeflash.snapshotDOM('after_{label}_{n}');
+    on the next line with matching indentation.
+    """
+    if "captureRender" not in test_source:
+        return test_source
+    if "captureRenderPerf" in test_source:
+        return test_source
+
+    interaction_counter: dict[str, int] = {}
+    lines = test_source.split("\n")
+    new_lines: list[str] = []
+
+    # Also match rerender() calls — use .* to handle nested parens like getByText('Add')
+    snapshot_interaction_pattern = re.compile(
+        r"^(\s*)((?:await\s+)?(?:fireEvent\.\w+|userEvent\.\w+|(?:\w+\.)?rerender)\s*\(.*\))\s*;?\s*$",
+        re.MULTILINE,
+    )
+
+    for line in lines:
+        new_lines.append(line)
+        m = snapshot_interaction_pattern.match(line)
+        if m:
+            indent = m.group(1)
+            call_text = m.group(2)
+            label = _extract_interaction_label(call_text)
+            if label == "interaction":
+                # rerender() call
+                label = "rerender"
+            interaction_counter[label] = interaction_counter.get(label, 0) + 1
+            unique_label = f"{label}_{interaction_counter[label]}"
+            new_lines.append(f"{indent}codeflash.snapshotDOM('after_{unique_label}');")
+
+    return "\n".join(new_lines)
+
+
 def inject_interaction_markers(test_source: str) -> str:
     """Inject _codeflashMarkInteraction() calls before each fireEvent/userEvent call.
 
@@ -325,3 +376,54 @@ def has_high_density_interactions(test_source: str) -> bool:
 
     interaction_calls = _INTERACTION_PATTERNS.findall(test_source)
     return len(interaction_calls) >= _MIN_SEQUENTIAL_INTERACTIONS
+
+
+# Pattern to extract props from the first render(<Component ...props... />) call
+_RENDER_CALL_PROPS_PATTERN = re.compile(
+    r"render\s*\(\s*<(\w+)\s+([^/]*?)\s*/?\s*>",
+)
+
+
+def _extract_render_props(test_source: str, component_name: str) -> str | None:
+    """Extract the props expression from the first render(<Component ...>) call."""
+    for m in _RENDER_CALL_PROPS_PATTERN.finditer(test_source):
+        if m.group(1) == component_name:
+            props_text = m.group(2).strip()
+            if props_text:
+                return props_text
+    return None
+
+
+def _inject_rerender_fallback(test_source: str, component_name: str) -> str:
+    """Inject a rerender efficiency test block when the test has no interactions.
+
+    This ensures every React perf test produces at least one update-phase marker.
+    """
+    # Try to extract props from existing render call
+    props_expr = _extract_render_props(test_source, component_name)
+    if props_expr:
+        jsx_open = f"<{component_name} {props_expr} />"
+    else:
+        jsx_open = f"<{component_name} />"
+
+    rerender_block = f"""
+describe('{component_name} rerender efficiency (auto-generated)', () => {{
+  it('should handle same-props rerenders', () => {{
+    const {{ rerender }} = render({jsx_open});
+    for (let i = 0; i < 10; i++) {{
+      rerender({jsx_open});
+    }}
+  }});
+}});
+"""
+    # Ensure {rerender} is in the @testing-library/react import
+    rtl_import_match = re.search(
+        r"import\s*\{([^}]+)\}\s*from\s*['\"]@testing-library/react['\"]", test_source
+    )
+    if rtl_import_match:
+        imports = rtl_import_match.group(1)
+        if "rerender" not in imports:
+            # Don't add 'rerender' to import — it comes from render() return value, not an import
+            pass
+
+    return test_source.rstrip() + "\n" + rerender_block

codeflash/languages/javascript/instrument.py

@@ -1346,6 +1346,13 @@ def _instrument_js_test_code(
             code=code, function_to_optimize=function_to_optimize, capture_func=capture_func, start_counter=final_counter
         )
 
+    # In behavior mode, inject DOM snapshot calls after each interaction so the
+    # comparator can detect post-interaction DOM divergence between original and candidate.
+    if mode == TestingMode.BEHAVIOR:
+        from codeflash.languages.javascript.frameworks.react.testgen import inject_dom_snapshot_calls  # noqa: PLC0415
+
+        code = inject_dom_snapshot_calls(code)
+
     return code
 
 

codeflash/languages/javascript/support.py

@@ -174,6 +174,14 @@ class JavaScriptSupport:
                     logger.debug(f"Skipping non-exported function: {func.name}")  # noqa: G004
                     continue
 
+                # Skip closures inside React components — they capture hook state
+                # and cannot be tested independently
+                if func.parent_function and func.parent_function in react_component_map:
+                    logger.debug(
+                        f"Skipping closure {func.name} inside React component {func.parent_function}"  # noqa: G004
+                    )
+                    continue
+
                 # Build parents list
                 parents: list[FunctionParent] = []
                 if func.class_name:
@@ -231,8 +239,23 @@ class JavaScriptSupport:
                 source, include_methods=True, include_arrow_functions=True, require_name=True
             )
 
+            # Build React component lookup to filter out closures
+            react_component_names: set[str] = set()
+            try:
+                from codeflash.languages.javascript.frameworks.react.discovery import classify_component  # noqa: PLC0415
+
+                for func in tree_functions:
+                    if classify_component(func, source, analyzer) is not None:
+                        react_component_names.add(func.name)
+            except Exception:
+                pass
+
             functions: list[FunctionToOptimize] = []
             for func in tree_functions:
+                # Skip closures inside React components
+                if func.parent_function and func.parent_function in react_component_names:
+                    continue
+
                 # Build parents list
                 parents: list[FunctionParent] = []
                 if func.class_name:
@@ -1212,6 +1235,47 @@ class JavaScriptSupport:
 
     # === Code Transformation ===
 
+    @staticmethod
+    def _strip_imports(source: str, analyzer: TreeSitterAnalyzer) -> str:
+        """Strip import/require statements from source code.
+
+        Used to remove duplicate imports from optimizer output before inserting
+        into the original file (which already has its own imports).
+        """
+        source_bytes = source.encode("utf8")
+        tree = analyzer.parse(source_bytes)
+
+        # Collect byte ranges of import nodes to remove
+        ranges_to_remove: list[tuple[int, int]] = []
+        for child in tree.root_node.children:
+            if child.type == "import_statement":
+                ranges_to_remove.append((child.start_byte, child.end_byte))
+            # CJS require: const x = require('...')
+            elif child.type in ("lexical_declaration", "variable_declaration"):
+                text = source_bytes[child.start_byte : child.end_byte].decode("utf8")
+                if "require(" in text:
+                    ranges_to_remove.append((child.start_byte, child.end_byte))
+
+        if not ranges_to_remove:
+            return source
+
+        # Build result, skipping removed ranges
+        parts: list[bytes] = []
+        prev_end = 0
+        for start, end in sorted(ranges_to_remove):
+            parts.append(source_bytes[prev_end:start])
+            # Skip trailing newline after import
+            if end < len(source_bytes) and source_bytes[end : end + 1] == b"\n":
+                end += 1
+            prev_end = end
+        parts.append(source_bytes[prev_end:])
+
+        result = b"".join(parts).decode("utf8")
+        # Remove leading blank lines left by stripping imports
+        while result.startswith("\n"):
+            result = result[1:]
+        return result
+
     def replace_function(self, source: str, function: FunctionToOptimize, new_source: str) -> str:
         """Replace a function in source code with new implementation.
 
@@ -1247,6 +1311,15 @@ class JavaScriptSupport:
         else:
             analyzer = TreeSitterAnalyzer(TreeSitterLanguage.JAVASCRIPT)
 
+        # Strip imports from optimizer output — the original file already has them,
+        # and import merging is handled by _add_global_declarations_for_language()
+        new_source = self._strip_imports(new_source, analyzer)
+
+        # If stripping imports left nothing, return original
+        if not new_source.strip():
+            logger.warning("new_source was only imports for %s, returning original", function.function_name)
+            return source
+
         # Check if new_source contains a JSDoc comment - if so, use full replacement
         # to include the updated JSDoc along with the function body
         stripped_new_source = new_source.strip()
@@ -1263,16 +1336,64 @@ class JavaScriptSupport:
             logger.warning(
                 "Could not extract body for %s from optimized code, using full replacement", function.function_name
             )
-            # Verify that new_source contains actual code before falling back to text replacement
+            if self._contains_function_declaration(new_source, function.function_name, analyzer):
-            # This prevents deletion of the original function when new_source is invalid
+                return self._replace_function_text_based(source, function, new_source, analyzer)
-            if not self._contains_function_declaration(new_source, function.function_name, analyzer):
+            # Final fallback: line-range replacement using the function's known line boundaries.
-                logger.warning("new_source does not contain function %s, returning original", function.function_name)
+            # This handles cases where tree-sitter can't parse the optimized output.
-                return source
+            logger.warning(
-            return self._replace_function_text_based(source, function, new_source, analyzer)
+                "Falling back to line-range replacement for %s (lines %d-%d)",
+                function.function_name,
+                function.starting_line,
+                function.ending_line,
+            )
+            return self._replace_function_by_line_range(source, function, new_source)
 
         # Find the original function and replace its body
         return self._replace_function_body(source, function, new_body, analyzer)
 
+    def _replace_function_by_line_range(
+        self, source: str, function: FunctionToOptimize, new_source: str
+    ) -> str:
+        """Last-resort replacement: cut the original function by line range and paste new_source."""
+        if function.starting_line is None or function.ending_line is None:
+            return source
+
+        lines = source.splitlines(keepends=True)
+        if lines and not lines[-1].endswith("\n"):
+            lines[-1] += "\n"
+
+        # Get indentation from original function's first line
+        if function.starting_line <= len(lines):
+            original_first = lines[function.starting_line - 1]
+            original_indent = len(original_first) - len(original_first.lstrip())
+        else:
+            original_indent = 0
+
+        new_lines = new_source.splitlines(keepends=True)
+        if new_lines:
+            first_non_blank = next((l for l in new_lines if l.strip()), new_lines[0])
+            new_indent = len(first_non_blank) - len(first_non_blank.lstrip())
+            indent_diff = original_indent - new_indent
+            if indent_diff != 0:
+                adjusted = []
+                for line in new_lines:
+                    if line.strip():
+                        if indent_diff > 0:
+                            adjusted.append(" " * indent_diff + line)
+                        else:
+                            cur = len(line) - len(line.lstrip())
+                            adjusted.append(line[min(cur, abs(indent_diff)) :])
+                    else:
+                        adjusted.append(line)
+                new_lines = adjusted
+
+        if new_lines and not new_lines[-1].endswith("\n"):
+            new_lines[-1] += "\n"
+
+        before = lines[: function.starting_line - 1]
+        after = lines[function.ending_line :]
+        return "".join(before + new_lines + after)
+
     def _contains_function_declaration(self, source: str, function_name: str, analyzer: TreeSitterAnalyzer) -> bool:
         """Check if source contains a function declaration with the given name.
 
@@ -1912,7 +2033,52 @@ class JavaScriptSupport:
             generated_tests = inject_test_globals(generated_tests, test_framework)
         if is_typescript():
             generated_tests = disable_ts_check(generated_tests)
-        return normalize_generated_tests_imports(generated_tests)
+        generated_tests = normalize_generated_tests_imports(generated_tests)
+
+        # Apply React-specific post-processing (act wrapping, interaction markers, etc.)
+        if self._is_react_source(source_file_path):
+            generated_tests = self._apply_react_postprocessing(generated_tests, source_file_path)
+
+        return generated_tests
+
+    def _is_react_source(self, source_file_path: Path) -> bool:
+        """Check if the source file is a React component."""
+        try:
+            content = source_file_path.read_text("utf-8", errors="replace")
+            return "react" in content.lower() and ("jsx" in content.lower() or source_file_path.suffix in (".tsx", ".jsx"))
+        except OSError:
+            return False
+
+    def _apply_react_postprocessing(self, generated_tests: GeneratedTestsList, source_file_path: Path) -> GeneratedTestsList:
+        """Apply React-specific post-processing to generated tests."""
+        from codeflash.languages.javascript.frameworks.react.testgen import (  # noqa: PLC0415
+            post_process_react_tests,
+        )
+        from codeflash.languages.javascript.frameworks.react.discovery import ComponentType, ReactComponentInfo  # noqa: PLC0415
+        from codeflash.models.models import GeneratedTests, GeneratedTestsList  # noqa: PLC0415
+
+        component_name = source_file_path.stem
+        dummy_info = ReactComponentInfo(
+            function_name=component_name,
+            component_type=ComponentType.FUNCTION,
+        )
+
+        updated_tests = []
+        for test in generated_tests.generated_tests:
+            updated_tests.append(
+                GeneratedTests(
+                    generated_original_test_source=test.generated_original_test_source,
+                    instrumented_behavior_test_source=post_process_react_tests(
+                        test.instrumented_behavior_test_source, dummy_info
+                    ),
+                    instrumented_perf_test_source=post_process_react_tests(
+                        test.instrumented_perf_test_source, dummy_info
+                    ),
+                    behavior_file_path=test.behavior_file_path,
+                    perf_file_path=test.perf_file_path,
+                )
+            )
+        return GeneratedTestsList(generated_tests=updated_tests)
 
     def remove_test_functions_from_generated_tests(
         self, generated_tests: GeneratedTestsList, functions_to_remove: list[str]

codeflash/languages/javascript/treesitter.py

@@ -401,6 +401,29 @@ class TreeSitterAnalyzer:
                     great_grandparent = grandparent.parent
                     if great_grandparent and great_grandparent.type == "export_statement":
                         return True
+            # HOC wrappers: export const Comp = forwardRef/memo(function/arrow)
+            # Tree: function → arguments → call_expression → variable_declarator → lexical_declaration → export_statement
+            if parent and parent.type == "arguments":
+                call_expr = parent.parent
+                if call_expr and call_expr.type == "call_expression":
+                    declarator = call_expr.parent
+                    if declarator and declarator.type == "variable_declarator":
+                        lex_decl = declarator.parent
+                        if lex_decl and lex_decl.type in ("lexical_declaration", "variable_declaration"):
+                            export_stmt = lex_decl.parent
+                            if export_stmt and export_stmt.type == "export_statement":
+                                return True
+                    # Nested HOC: export const X = memo(forwardRef(fn))
+                    if declarator and declarator.type == "arguments":
+                        outer_call = declarator.parent
+                        if outer_call and outer_call.type == "call_expression":
+                            outer_decl = outer_call.parent
+                            if outer_decl and outer_decl.type == "variable_declarator":
+                                lex_decl = outer_decl.parent
+                                if lex_decl and lex_decl.type in ("lexical_declaration", "variable_declaration"):
+                                    export_stmt = lex_decl.parent
+                                    if export_stmt and export_stmt.type == "export_statement":
+                                        return True
 
         # For methods in exported classes
         if node.type == "method_definition":
@@ -609,6 +632,8 @@ class TreeSitterAnalyzer:
 
         return None
 
+    _HOC_WRAPPER_NAMES = frozenset({"forwardRef", "memo", "React.forwardRef", "React.memo"})
+
     def _get_name_from_assignment(self, node: Node, source_bytes: bytes) -> str:
         """Try to extract function name from parent variable declaration or assignment.
 
@@ -617,6 +642,8 @@ class TreeSitterAnalyzer:
         - const foo = function() {}
         - let bar = function() {}
         - obj.method = () => {}
+        - const Tooltip = forwardRef(function TooltipInner(...) {...})
+        - const MemoComp = React.memo((props) => {...})
         """
         parent = node.parent
         if parent is None:
@@ -628,6 +655,30 @@ class TreeSitterAnalyzer:
             if name_node:
                 return self.get_node_text(name_node, source_bytes)
 
+        # Check for HOC wrapper: const Name = forwardRef/memo(function/arrow)
+        # Tree structure: function → arguments → call_expression → variable_declarator
+        if parent.type == "arguments":
+            call_expr = parent.parent
+            if call_expr is not None and call_expr.type == "call_expression":
+                callee = call_expr.child_by_field_name("function")
+                if callee is not None:
+                    callee_text = self.get_node_text(callee, source_bytes)
+                    if callee_text in self._HOC_WRAPPER_NAMES:
+                        declarator = call_expr.parent
+                        if declarator is not None and declarator.type == "variable_declarator":
+                            name_node = declarator.child_by_field_name("name")
+                            if name_node:
+                                return self.get_node_text(name_node, source_bytes)
+                        # Nested HOC: memo(forwardRef(fn)) — call_expr parent is arguments of outer call
+                        if declarator is not None and declarator.type == "arguments":
+                            outer_call = declarator.parent
+                            if outer_call is not None and outer_call.type == "call_expression":
+                                outer_declarator = outer_call.parent
+                                if outer_declarator is not None and outer_declarator.type == "variable_declarator":
+                                    name_node = outer_declarator.child_by_field_name("name")
+                                    if name_node:
+                                        return self.get_node_text(name_node, source_bytes)
+
         # Check for assignment expression: foo = ...
         if parent.type == "assignment_expression":
             left_node = parent.child_by_field_name("left")

codeflash/models/models.py

@@ -83,6 +83,7 @@ class TestDiffScope(str, Enum):
     RETURN_VALUE = "return_value"
     STDOUT = "stdout"
     DID_PASS = "did_pass"  # noqa: S105
+    DOM_SNAPSHOT = "dom_snapshot"
 
 
 @dataclass

codeflash/optimization/function_optimizer.py

@@ -96,6 +96,7 @@ from codeflash.models.models import (
     OptimizedCandidateResult,
     OptimizedCandidateSource,
     OriginalCodeBaseline,
+    TestDiffScope,
     TestFile,
     TestFiles,
     TestingMode,
@@ -2955,6 +2956,13 @@ class FunctionOptimizer:
                 logger.info("h3|Test results matched ✅")
                 console.rule()
             else:
+                dom_snapshot_diffs = [d for d in diffs if d.scope == TestDiffScope.DOM_SNAPSHOT]
+                if dom_snapshot_diffs:
+                    logger.warning(
+                        "[REACT] DOM snapshot divergence detected after %d interaction(s). "
+                        "The optimized component produces different DOM output.",
+                        len(dom_snapshot_diffs),
+                    )
                 self.repair_if_possible(
                     candidate, diffs, eval_ctx, code_context, len(candidate_behavior_results), exp_type
                 )
@@ -3178,8 +3186,7 @@ class FunctionOptimizer:
                 coverage_config_file=coverage_config_file,
                 skip_sqlite_cleanup=skip_cleanup,
             )
-            if testing_type == TestingMode.PERFORMANCE:
+            results.perf_stdout = run_result.stdout
-                results.perf_stdout = run_result.stdout
             return results, coverage_results
         # For LINE_PROFILE mode, Python uses .lprof files while JavaScript uses JSON
         # Return TestResults for JavaScript so _line_profiler_step_javascript can parse the JSON

packages/codeflash/runtime/capture.js

@@ -46,11 +46,11 @@ function _getReact() {
     }
 }
 
-// Try to load better-sqlite3, fall back to JSON if not available
-let useSqlite = false;
-
 // Configuration from environment
 const OUTPUT_FILE = process.env.CODEFLASH_OUTPUT_FILE;
+
+// Enable SQLite when better-sqlite3 is available and an output file is configured
+let useSqlite = !!(Database && OUTPUT_FILE);
 const LOOP_INDEX = parseInt(process.env.CODEFLASH_LOOP_INDEX || '1', 10);
 const TEST_ITERATION = process.env.CODEFLASH_TEST_ITERATION;
 const TEST_MODULE = process.env.CODEFLASH_TEST_MODULE;
@@ -1091,6 +1091,35 @@ function captureRender(funcName, lineId, renderFn, Component, ...createElementAr
     return renderResult;
 }
 
+/**
+ * Capture a DOM snapshot after a user interaction for behavioral verification.
+ *
+ * Writes normalized `document.body.innerHTML` to the SQLite database as an
+ * additional row with function name `__dom_snapshot__`.  The existing
+ * `compare-results.js` comparator compares all rows by invocation ID, so
+ * snapshot rows are compared automatically.
+ *
+ * @param {string} label - Unique label for this snapshot (e.g. 'after_click_1')
+ */
+function snapshotDOM(label) {
+    if (!useSqlite || !db) return;
+    if (typeof document === 'undefined' || !document.body) return;
+
+    // Normalize HTML: collapse whitespace, strip React-internal attributes
+    let html = document.body.innerHTML;
+    html = html.replace(/\s+/g, ' ').trim();
+    html = html.replace(/\s*data-reactroot\s*/g, '');
+
+    const { testModulePath, testClassName, testFunctionName } = _getTestContext();
+    const invocationId = `snapshot_${label}`;
+
+    recordResult(
+        testModulePath, testClassName, testFunctionName,
+        '__dom_snapshot__', invocationId,
+        [label], html, null, 0
+    );
+}
+
 /**
  * Capture a React component render call for PERFORMANCE benchmarking only.
  *
@@ -1334,6 +1363,7 @@ module.exports = {
     capturePerf,       // Performance benchmarking (prints to stdout only)
     captureRender,     // React render behavior verification (writes to SQLite)
     captureRenderPerf, // React render performance benchmarking (prints to stdout only)
+    snapshotDOM,       // DOM snapshot after interaction (behavior verification)
     captureMultiple,
     writeResults,
     clearResults,

packages/codeflash/runtime/compare-results.js

@@ -167,9 +167,13 @@ function compareResults(originalResults, candidateResults) {
 
         if (!isEqual) {
             allEquivalent = false;
+            // Use dom_snapshot scope for __dom_snapshot__ rows
+            const scope = original.functionGettingTested === '__dom_snapshot__'
+                ? 'dom_snapshot'
+                : 'return_value';
             diffs.push({
                 invocation_id: invocationId,
-                scope: 'return_value',
+                scope,
                 original: summarizeValue(originalValue),
                 candidate: summarizeValue(candidateValue),
                 test_info: {

packages/codeflash/runtime/index.d.ts

@@ -36,6 +36,15 @@ export function capturePerf<T extends (...args: any[]) => any>(
     ...args: Parameters<T>
 ): ReturnType<T>;
 
+/**
+ * Capture a DOM snapshot after a user interaction for behavioral verification.
+ * Writes normalized document.body.innerHTML to the SQLite database with
+ * function name '__dom_snapshot__' so the comparator detects DOM divergence.
+ *
+ * @param label - Unique label for this snapshot (e.g. 'after_click_1')
+ */
+export function snapshotDOM(label: string): void;
+
 /**
  * Capture multiple invocations for benchmarking.
  *
@@ -126,6 +135,7 @@ export const TEST_ITERATION: string;
 declare const codeflash: {
     capture: typeof capture;
     capturePerf: typeof capturePerf;
+    snapshotDOM: typeof snapshotDOM;
     captureMultiple: typeof captureMultiple;
     writeResults: typeof writeResults;
     clearResults: typeof clearResults;

packages/codeflash/runtime/index.js

@@ -38,6 +38,7 @@ module.exports = {
 
   captureRender: capture.captureRender,
   captureRenderPerf: capture.captureRenderPerf,
+  snapshotDOM: capture.snapshotDOM,
 
   captureMultiple: capture.captureMultiple,
 

tests/react/test_benchmarking.py

@@ -173,6 +173,8 @@ class TestRenderEfficiencyCritic:
             optimized_render_duration=10.0,
             original_update_render_count=8,
             optimized_update_render_count=8,
+            original_update_duration=80.0,
+            optimized_update_duration=5.0,
         ) is True
 
     def test_rejects_worse_than_best(self):
@@ -448,6 +450,8 @@ class TestRenderEfficiencyCriticTrustDuration:
             optimized_render_duration=10.0,
             original_update_render_count=8,
             optimized_update_render_count=8,
+            original_update_duration=100.0,
+            optimized_update_duration=10.0,
             trust_duration=True,
         ) is True
 

tests/react/test_dom_snapshot.py

@@ -0,0 +1,78 @@
+"""Tests for inject_dom_snapshot_calls in React behavioral verification."""
+
+from __future__ import annotations
+
+from codeflash.languages.javascript.frameworks.react.testgen import inject_dom_snapshot_calls
+
+
+def test_injects_after_fireEvent():
+    source = """\
+import codeflash from 'codeflash';
+const { container } = codeflash.captureRender('Comp', '1', render, Comp);
+fireEvent.click(screen.getByText('Add'));
+fireEvent.change(input, { target: { value: 'hi' } });
+"""
+    result = inject_dom_snapshot_calls(source)
+    assert "codeflash.snapshotDOM('after_click_1');" in result
+    assert "codeflash.snapshotDOM('after_change_1');" in result
+
+
+def test_skips_perf_mode():
+    source = """\
+import codeflash from 'codeflash';
+const result = await codeflash.captureRenderPerf('Comp', '1', render, Comp);
+fireEvent.click(screen.getByText('Add'));
+"""
+    result = inject_dom_snapshot_calls(source)
+    assert "snapshotDOM" not in result
+
+
+def test_skips_without_captureRender():
+    source = """\
+import codeflash from 'codeflash';
+fireEvent.click(screen.getByText('Add'));
+"""
+    result = inject_dom_snapshot_calls(source)
+    assert "snapshotDOM" not in result
+
+
+def test_preserves_indentation():
+    source = """\
+import codeflash from 'codeflash';
+const { container } = codeflash.captureRender('Comp', '1', render, Comp);
+    fireEvent.click(btn);
+        fireEvent.change(input, { target: { value: 'x' } });
+"""
+    result = inject_dom_snapshot_calls(source)
+    lines = result.split("\n")
+    # Find the snapshot lines and check their indentation
+    snapshot_lines = [l for l in lines if "snapshotDOM" in l]
+    assert len(snapshot_lines) == 2
+    assert snapshot_lines[0].startswith("    codeflash.snapshotDOM")
+    assert snapshot_lines[1].startswith("        codeflash.snapshotDOM")
+
+
+def test_no_semicolons():
+    """Real-world projects (e.g. Zustand) don't use semicolons."""
+    source = """\
+const { container } = codeflash.captureRender('Comp', '1', render, Comp)
+    fireEvent.click(screen.getByText('button'))
+    fireEvent.click(screen.getByTestId('test-shallow'))
+"""
+    result = inject_dom_snapshot_calls(source)
+    assert "after_click_1" in result
+    assert "after_click_2" in result
+
+
+def test_sequential_counter():
+    source = """\
+import codeflash from 'codeflash';
+const { container } = codeflash.captureRender('Comp', '1', render, Comp);
+fireEvent.click(btn);
+fireEvent.click(btn);
+fireEvent.click(btn);
+"""
+    result = inject_dom_snapshot_calls(source)
+    assert "after_click_1" in result
+    assert "after_click_2" in result
+    assert "after_click_3" in result