deepresearch-flow 0.5.1__py3-none-any.whl → 0.6.0__py3-none-any.whl

deepresearch_flow/recognize/cli.py

@@ -42,6 +42,10 @@ from deepresearch_flow.recognize.mermaid import (
     extract_mermaid_spans,
     fix_mermaid_text,
     require_mmdc,
+    extract_diagrams_from_text,
+    repair_all_diagrams_global,
+    DiagramTask,
+    apply_replacements,
 )
 from deepresearch_flow.recognize.organize import (
     discover_mineru_dirs,
@@ -73,7 +77,8 @@ def _relative_path(path: Path) -> str:
 
 def _warn_if_not_empty(output_dir: Path) -> None:
     if output_dir.exists() and any(output_dir.iterdir()):
-        logger.warning("Output directory not empty: %s", output_dir)
+        item_count = sum(1 for _ in output_dir.iterdir())
+        logger.warning("Output directory not empty: %s (items=%d)", output_dir, item_count)
 
 
 def _print_summary(title: str, rows: list[tuple[str, str]]) -> None:
@@ -114,6 +119,60 @@ def _map_output_files(
     return mapping
 
 
+RetryKey = tuple[int, str | None, int | None]
+
+
+def _load_retry_targets(report_path: Path) -> dict[Path, set[RetryKey]]:
+    if not report_path.exists():
+        raise click.ClickException(f"Retry report not found: {report_path}")
+    try:
+        payload = json.loads(read_text(report_path))
+    except json.JSONDecodeError as exc:
+        raise click.ClickException(f"Retry report is not valid JSON: {exc}") from exc
+    if not isinstance(payload, list) or not payload:
+        raise click.ClickException(f"Retry report is empty: {report_path}")
+    targets: dict[Path, set[RetryKey]] = {}
+    for entry in payload:
+        if not isinstance(entry, dict):
+            continue
+        path_raw = entry.get("path")
+        line_raw = entry.get("line")
+        if not path_raw or line_raw is None:
+            continue
+        try:
+            line_no = int(line_raw)
+        except (TypeError, ValueError):
+            continue
+        field_path = entry.get("field_path")
+        if not isinstance(field_path, str):
+            field_path = None
+        item_index = entry.get("item_index")
+        if not isinstance(item_index, int):
+            item_index = None
+        key = (line_no, field_path, item_index)
+        targets.setdefault(Path(path_raw).resolve(), set()).add(key)
+    if not targets:
+        raise click.ClickException(f"Retry report has no valid entries: {report_path}")
+    return targets
+
+
+def _filter_retry_spans(
+    spans: list[Any],
+    line_offset: int,
+    field_path: str | None,
+    item_index: int | None,
+    retry_keys: set[RetryKey] | None,
+) -> list[Any]:
+    if not retry_keys:
+        return spans
+    filtered: list[Any] = []
+    for span in spans:
+        line_no = line_offset + span.line - 1
+        if (line_no, field_path, item_index) in retry_keys:
+            filtered.append(span)
+    return filtered
+
+
 def _aggregate_image_counts(paths: Iterable[Path]) -> dict[str, int]:
     totals = {"total": 0, "data": 0, "http": 0, "local": 0}
     for path in paths:
@@ -194,6 +253,8 @@ async def _fix_json_items(
     default_template: str | None,
     fix_level: str,
     format_enabled: bool,
+    progress: tqdm | None = None,
+    progress_lock: asyncio.Lock | None = None,
 ) -> tuple[int, int, int, int]:
     items_total = 0
     items_updated = 0
@@ -218,6 +279,9 @@ async def _fix_json_items(
                 item_updated = True
         if item_updated:
             items_updated += 1
+        if progress and progress_lock:
+            async with progress_lock:
+                progress.update(1)
     return items_total, items_updated, fields_total, fields_updated
 
 
@@ -350,7 +414,12 @@ async def _run_fix_json(
     async def handler(path: Path) -> tuple[int, int, int, int, int]:
         items, payload, template_tag = _load_json_payload(path)
         items_total, items_updated, fields_total, fields_updated = await _fix_json_items(
-            items, template_tag, fix_level, format_enabled
+            items,
+            template_tag,
+            fix_level,
+            format_enabled,
+            progress,
+            progress_lock,
         )
         output_data: Any
         if payload is None:
@@ -367,9 +436,6 @@ async def _run_fix_json(
         async with semaphore:
             result = await handler(path)
             results.append(result)
-            if progress and progress_lock:
-                async with progress_lock:
-                    progress.update(1)
 
     await asyncio.gather(*(runner(path) for path in paths))
     return results
@@ -787,7 +853,16 @@ def recognize_fix(
         _print_summary("recognize fix (dry-run)", rows)
         return
 
-    progress = tqdm(total=len(paths), desc="fix", unit="file")
+    progress_total = len(paths)
+    progress_unit = "file"
+    if json_mode:
+        json_items_total = 0
+        for path in paths:
+            items, _, _ = _load_json_payload(path)
+            json_items_total += sum(1 for item in items if isinstance(item, dict))
+        progress_total = json_items_total
+        progress_unit = "item"
+    progress = tqdm(total=progress_total, desc="fix", unit=progress_unit)
     try:
         if json_mode:
             results = asyncio.run(
@@ -870,6 +945,7 @@ def recognize_fix(
 @click.option("--max-retries", "max_retries", default=3, show_default=True, type=int)
 @click.option("--workers", type=int, default=4, show_default=True, help="Concurrent workers")
 @click.option("--timeout", "timeout", default=120.0, show_default=True, type=float)
+@click.option("--retry-failed", "retry_failed", is_flag=True, help="Retry only failed formulas")
 @click.option(
     "--only-show-error",
     "only_show_error",
@@ -892,6 +968,7 @@ def recognize_fix_math(
     max_retries: int,
     workers: int,
     timeout: float,
+    retry_failed: bool,
     only_show_error: bool,
     report_path: str | None,
     dry_run: bool,
@@ -911,6 +988,8 @@ def recognize_fix_math(
         raise click.ClickException("--max-retries must be non-negative")
     if workers <= 0:
         raise click.ClickException("--workers must be positive")
+    if retry_failed and only_show_error:
+        raise click.ClickException("--retry-failed cannot be used with --only-show-error")
     try:
         require_pylatexenc()
     except RuntimeError as exc:
@@ -954,6 +1033,24 @@ def recognize_fix_math(
         return
 
     output_path = Path(output_dir) if output_dir else None
+    report_target = None
+    if report_path:
+        report_target = Path(report_path)
+    elif not only_show_error:
+        if output_path:
+            report_target = output_path / "fix-math-errors.json"
+        elif in_place:
+            report_target = Path.cwd() / "fix-math-errors.json"
+
+    retry_targets: dict[Path, set[RetryKey]] | None = None
+    if retry_failed:
+        if report_target is None:
+            raise click.ClickException("--retry-failed requires an error report path")
+        retry_targets = _load_retry_targets(report_target)
+        paths = [path for path in paths if path.resolve() in retry_targets]
+        if not paths:
+            raise click.ClickException("No failed formulas matched the provided inputs")
+
     if output_path and not dry_run and not only_show_error:
         output_path = _ensure_output_dir(output_dir)
         _warn_if_not_empty(output_path)
@@ -969,15 +1066,6 @@ def recognize_fix_math(
     else:
         output_map = {path: path for path in paths}
 
-    report_target = None
-    if report_path:
-        report_target = Path(report_path)
-    elif not only_show_error:
-        if output_path:
-            report_target = output_path / "fix-math-errors.json"
-        elif in_place:
-            report_target = Path.cwd() / "fix-math-errors.json"
-
     if dry_run and not only_show_error:
         rows = [
             ("Mode", "json" if json_mode else "markdown"),
@@ -988,6 +1076,7 @@ def recognize_fix_math(
             ("Max retries", str(max_retries)),
             ("Workers", str(workers)),
             ("Timeout", f"{timeout:.1f}s"),
+            ("Retry failed", "yes" if retry_failed else "no"),
             ("Only show error", "yes" if only_show_error else "no"),
             ("In place", "yes" if in_place else "no"),
             ("Output dir", _relative_path(output_path) if output_path else "-"),
@@ -1021,12 +1110,24 @@ def recognize_fix_math(
                             value = item.get(field)
                             if not isinstance(value, str):
                                 continue
-                            spans = extract_math_spans(value, context_chars)
-                            if spans:
-                                formula_progress.total += len(spans)
-                                formula_progress.refresh()
                             line_start, cursor = locate_json_field_start(raw_text, value, cursor)
                             field_path = f"papers[{item_index}].{field}"
+                            spans = extract_math_spans(value, context_chars)
+                            retry_keys = None
+                            if retry_targets is not None:
+                                retry_keys = retry_targets.get(path.resolve(), set())
+                                retry_keys = {
+                                    key
+                                    for key in retry_keys
+                                    if key[1] == field_path and key[2] == item_index
+                                }
+                            spans = _filter_retry_spans(
+                                spans, line_start, field_path, item_index, retry_keys
+                            )
+                            if not spans:
+                                continue
+                            formula_progress.total += len(spans)
+                            formula_progress.refresh()
                             updated, errors = await fix_math_text(
                                 value,
                                 str(path),
@@ -1044,6 +1145,7 @@ def recognize_fix_math(
                                 stats,
                                 repair_enabled=not only_show_error,
                                 spans=spans,
+                                allowed_keys=retry_keys,
                                 progress_cb=lambda: formula_progress.update(1),
                             )
                             if not only_show_error and updated != value:
@@ -1057,6 +1159,12 @@ def recognize_fix_math(
                 else:
                     content = await asyncio.to_thread(read_text, path)
                     spans = extract_math_spans(content, context_chars)
+                    retry_keys = None
+                    if retry_targets is not None:
+                        retry_keys = retry_targets.get(path.resolve(), set())
+                        spans = _filter_retry_spans(spans, 1, None, None, retry_keys)
+                        if not spans:
+                            return stats
                     if spans:
                         formula_progress.total += len(spans)
                         formula_progress.refresh()
@@ -1077,6 +1185,7 @@ def recognize_fix_math(
                         stats,
                         repair_enabled=not only_show_error,
                         spans=spans,
+                        allowed_keys=retry_keys,
                         progress_cb=lambda: formula_progress.update(1),
                     )
                     if not only_show_error:
@@ -1121,6 +1230,7 @@ def recognize_fix_math(
         ("Cleaned", str(stats.formulas_cleaned)),
         ("Repaired", str(stats.formulas_repaired)),
         ("Failed", str(stats.formulas_failed)),
+        ("Retry failed", "yes" if retry_failed else "no"),
         ("Only show error", "yes" if only_show_error else "no"),
         ("Report", _relative_path(report_target) if report_target else "-"),
     ]
@@ -1147,6 +1257,7 @@ def recognize_fix_math(
 @click.option("--max-retries", "max_retries", default=3, show_default=True, type=int)
 @click.option("--workers", type=int, default=4, show_default=True, help="Concurrent workers")
 @click.option("--timeout", "timeout", default=120.0, show_default=True, type=float)
+@click.option("--retry-failed", "retry_failed", is_flag=True, help="Retry only failed diagrams")
 @click.option(
     "--only-show-error",
     "only_show_error",
@@ -1169,6 +1280,7 @@ def recognize_fix_mermaid(
     max_retries: int,
     workers: int,
     timeout: float,
+    retry_failed: bool,
     only_show_error: bool,
     report_path: str | None,
     dry_run: bool,
@@ -1188,6 +1300,8 @@ def recognize_fix_mermaid(
         raise click.ClickException("--max-retries must be non-negative")
     if workers <= 0:
         raise click.ClickException("--workers must be positive")
+    if retry_failed and only_show_error:
+        raise click.ClickException("--retry-failed cannot be used with --only-show-error")
     try:
         require_mmdc()
     except RuntimeError as exc:
@@ -1231,6 +1345,24 @@ def recognize_fix_mermaid(
         return
 
     output_path = Path(output_dir) if output_dir else None
+    report_target = None
+    if report_path:
+        report_target = Path(report_path)
+    elif not only_show_error:
+        if output_path:
+            report_target = output_path / "fix-mermaid-errors.json"
+        elif in_place:
+            report_target = Path.cwd() / "fix-mermaid-errors.json"
+
+    retry_targets: dict[Path, set[RetryKey]] | None = None
+    if retry_failed:
+        if report_target is None:
+            raise click.ClickException("--retry-failed requires an error report path")
+        retry_targets = _load_retry_targets(report_target)
+        paths = [path for path in paths if path.resolve() in retry_targets]
+        if not paths:
+            raise click.ClickException("No failed diagrams matched the provided inputs")
+
     if output_path and not dry_run and not only_show_error:
         output_path = _ensure_output_dir(output_dir)
         _warn_if_not_empty(output_path)
@@ -1246,15 +1378,6 @@ def recognize_fix_mermaid(
     else:
         output_map = {path: path for path in paths}
 
-    report_target = None
-    if report_path:
-        report_target = Path(report_path)
-    elif not only_show_error:
-        if output_path:
-            report_target = output_path / "fix-mermaid-errors.json"
-        elif in_place:
-            report_target = Path.cwd() / "fix-mermaid-errors.json"
-
     if dry_run and not only_show_error:
         rows = [
             ("Mode", "json" if json_mode else "markdown"),
@@ -1265,6 +1388,7 @@ def recognize_fix_mermaid(
             ("Max retries", str(max_retries)),
             ("Workers", str(workers)),
             ("Timeout", f"{timeout:.1f}s"),
+            ("Retry failed", "yes" if retry_failed else "no"),
             ("Only show error", "yes" if only_show_error else "no"),
             ("In place", "yes" if in_place else "no"),
             ("Output dir", _relative_path(output_path) if output_path else "-"),
@@ -1273,112 +1397,260 @@ def recognize_fix_mermaid(
         _print_summary("recognize fix-mermaid (dry-run)", rows)
         return
 
-    progress = tqdm(total=len(paths), desc="fix-mermaid", unit="file")
-    diagram_progress = tqdm(total=0, desc="diagrams", unit="diagram")
+    progress = tqdm(total=len(paths), desc="extract", unit="file")
+    field_progress = tqdm(total=0, desc="extract-field", unit="field", disable=not json_mode, leave=False)
+    diagram_progress = tqdm(total=0, desc="repair", unit="diagram")
     error_records: list[dict[str, Any]] = []
+    
+    # Performance metrics
+    extract_start_time = time.monotonic()
+    repair_start_time = 0.0
+    extract_duration = 0.0
+    repair_duration = 0.0
 
     async def run() -> MermaidFixStats:
-        semaphore = asyncio.Semaphore(workers)
-        progress_lock = asyncio.Lock()
         stats_total = MermaidFixStats()
 
         async with httpx.AsyncClient() as client:
-            async def handle_path(path: Path) -> MermaidFixStats:
-                stats = MermaidFixStats()
+            # Phase 1: Extract all diagrams from all files in parallel (flatten to 1D)
+            progress_lock = asyncio.Lock()
+            field_progress_lock = asyncio.Lock()
+            
+            async def extract_from_file(path: Path) -> list[DiagramTask]:
+                tasks: list[DiagramTask] = []
+                
                 if json_mode:
-                    raw_text = read_text(path)
+                    raw_text = await asyncio.to_thread(read_text, path)
                     items, payload, template_tag = _load_json_payload(path)
+                    
+                    logger.info("Extracting from JSON: %s (%d papers)", _relative_path(path), len(items))
+                    
+                    # Pre-calculate all field positions for parallel extraction
+                    field_locations: list[tuple[int, str, str, str | None, int]] = []
                     cursor = 0
+                    
                     for item_index, item in enumerate(items):
                         if not isinstance(item, dict):
                             continue
                         template = _resolve_item_template(item, template_tag)
                         fields = _template_markdown_fields(template)
+                        
                         for field in fields:
                             value = item.get(field)
                             if not isinstance(value, str):
                                 continue
-                            spans = extract_mermaid_spans(value, context_chars)
-                            if spans:
-                                diagram_progress.total += len(spans)
-                                diagram_progress.refresh()
                             line_start, cursor = locate_json_field_start(raw_text, value, cursor)
                             field_path = f"papers[{item_index}].{field}"
-                            updated, errors = await fix_mermaid_text(
-                                value,
-                                str(path),
-                                line_start,
-                                field_path,
-                                item_index,
-                                provider,
-                                model_name,
-                                api_key,
-                                timeout,
-                                max_retries,
-                                batch_size,
-                                context_chars,
-                                client,
-                                stats,
-                                repair_enabled=not only_show_error,
-                                spans=spans,
-                                progress_cb=lambda: diagram_progress.update(1),
+                            field_locations.append((line_start, value, field_path, None, item_index))
+                    
+                    logger.info("Pre-calculated %d field locations from %s", len(field_locations), _relative_path(path))
+                    
+                    # Apply retry filter to field locations if needed
+                    if retry_targets is not None:
+                        retry_keys = retry_targets.get(path.resolve(), set())
+                        # Prefer filtering by (field_path, item_index) to avoid expensive validation / mmdc calls.
+                        retry_fields = {
+                            (field_path, item_index)
+                            for _, field_path, item_index in retry_keys
+                            if field_path is not None and item_index is not None
+                        }
+                        if retry_fields:
+                            before = len(field_locations)
+                            field_locations = [
+                                loc for loc in field_locations if (loc[2], loc[4]) in retry_fields
+                            ]
+                            logger.info(
+                                "Retry filter: %d/%d fields match (by field_path)",
+                                len(field_locations),
+                                before,
                             )
-                            if not only_show_error and updated != value:
-                                item[field] = updated
-                            error_records.extend(errors)
-                    if not only_show_error:
-                        output_data: Any = items if payload is None else {**payload, "papers": items}
-                        output_path = output_map[path]
-                        serialized = json.dumps(output_data, ensure_ascii=False, indent=2)
-                        await asyncio.to_thread(output_path.write_text, f"{serialized}\n", encoding="utf-8")
+                        else:
+                            # Fallback: filter by line numbers using fast span extraction (no validation).
+                            filtered_locations: list[tuple[int, str, str, str | None, int]] = []
+                            for line_start, value, field_path, _, item_index in field_locations:
+                                spans = extract_mermaid_spans(value, context_chars)
+                                if any(
+                                    (line_start + span.line - 1, field_path, item_index) in retry_keys
+                                    for span in spans
+                                ):
+                                    filtered_locations.append((line_start, value, field_path, None, item_index))
+                            field_locations = filtered_locations
+                            logger.info("Retry filter: %d fields match (by line)", len(field_locations))
+                    
+                    # Parallel extraction from all fields
+                    async def extract_from_field(loc: tuple[int, str, str, str | None, int]) -> list[DiagramTask]:
+                        line_start, value, field_path, _, item_index = loc
+                        field_tasks = extract_diagrams_from_text(
+                            value, path, line_start, field_path, item_index, context_chars,
+                            skip_validation=not only_show_error  # Skip validation unless validating only
+                        )
+                        
+                        # Apply retry filter to individual tasks
+                        if retry_targets is not None:
+                            retry_keys = retry_targets.get(path.resolve(), set())
+                            field_tasks = [
+                                task for task in field_tasks
+                                if (task.file_line_offset + task.span.line - 1, task.field_path, task.item_index) in retry_keys
+                            ]
+                        
+                        return field_tasks
+                    
+                    if field_locations:
+                        logger.info("Extracting diagrams from %d fields in parallel...", len(field_locations))
+
+                        async with field_progress_lock:
+                            field_progress.total += len(field_locations)
+                            field_progress.refresh()
+
+                        # Bounded worker pool (avoid scheduling thousands of coroutines at once).
+                        max_field_workers = 50
+                        field_workers = min(max_field_workers, len(field_locations))
+                        field_queue: asyncio.Queue[tuple[int, str, str, str | None, int] | None] = asyncio.Queue()
+                        for loc in field_locations:
+                            field_queue.put_nowait(loc)
+                        for _ in range(field_workers):
+                            field_queue.put_nowait(None)
+
+                        async def field_worker() -> list[DiagramTask]:
+                            out: list[DiagramTask] = []
+                            while True:
+                                loc = await field_queue.get()
+                                if loc is None:
+                                    break
+                                out.extend(await extract_from_field(loc))
+                                async with field_progress_lock:
+                                    field_progress.update(1)
+                            return out
+
+                        worker_results = await asyncio.gather(*[field_worker() for _ in range(field_workers)])
+                        for batch in worker_results:
+                            tasks.extend(batch)
+                    
+                    logger.info("Extracted %d diagrams from %s", len(tasks), _relative_path(path))
                 else:
                     content = await asyncio.to_thread(read_text, path)
-                    spans = extract_mermaid_spans(content, context_chars)
-                    if spans:
-                        diagram_progress.total += len(spans)
-                        diagram_progress.refresh()
-                    updated, errors = await fix_mermaid_text(
-                        content,
-                        str(path),
-                        1,
-                        None,
-                        None,
-                        provider,
-                        model_name,
-                        api_key,
-                        timeout,
-                        max_retries,
-                        batch_size,
-                        context_chars,
-                        client,
-                        stats,
-                        repair_enabled=not only_show_error,
-                        spans=spans,
-                        progress_cb=lambda: diagram_progress.update(1),
+                    
+                    logger.info("Extracting from markdown: %s", _relative_path(path))
+                    
+                    # Extract diagrams from markdown
+                    file_tasks = extract_diagrams_from_text(
+                        content, path, 1, None, None, context_chars,
+                        skip_validation=not only_show_error  # Skip validation unless validating only
                     )
-                    if not only_show_error:
-                        output_path = output_map[path]
+                    
+                    # Apply retry filter if needed
+                    if retry_targets is not None:
+                        retry_keys = retry_targets.get(path.resolve(), set())
+                        file_tasks = [
+                            task for task in file_tasks
+                            if (task.file_line_offset + task.span.line - 1, task.field_path, task.item_index) in retry_keys
+                        ]
+                    
+                    tasks.extend(file_tasks)
+                    logger.info("Extracted %d diagrams from %s", len(tasks), _relative_path(path))
+                
+                async with progress_lock:
+                    progress.update(1)
+                return tasks
+            
+            # Parallel extraction with progress
+            file_task_lists = await asyncio.gather(*[extract_from_file(path) for path in paths])
+            all_tasks = [task for tasks in file_task_lists for task in tasks]
+            
+            progress.close()
+            field_progress.close()
+            nonlocal extract_duration, repair_start_time
+            extract_duration = time.monotonic() - extract_start_time
+            
+            # Update diagram progress total
+            diagram_progress.total = len(all_tasks)
+            diagram_progress.refresh()
+            
+            if not all_tasks:
+                return stats_total
+            
+            # Phase 2: Global parallel repair (flatten all batches)
+            repair_start_time = time.monotonic()
+            file_replacements, errors = await repair_all_diagrams_global(
+                all_tasks,
+                batch_size,
+                workers,  # Use workers for global batch concurrency
+                provider,
+                model_name,
+                api_key,
+                timeout,
+                max_retries,
+                client,
+                stats_total,
+                progress_cb=lambda: diagram_progress.update(1) if not only_show_error else None,
+            )
+            
+            error_records.extend(errors)
+            diagram_progress.close()
+            nonlocal repair_duration
+            repair_duration = time.monotonic() - repair_start_time
+            
+            # Phase 3: Write back to files
+            if not only_show_error:
+                write_progress = tqdm(total=len(paths), desc="write", unit="file")
+                
+                for path in paths:
+                    replacements = file_replacements.get(path, [])
+                    output_path = output_map[path]
+                    
+                    if json_mode:
+                        # For JSON, apply replacements to fields
+                        raw_text = await asyncio.to_thread(read_text, path)
+                        items, payload, template_tag = _load_json_payload(path)
+                        cursor = 0
+                        
+                        for item_index, item in enumerate(items):
+                            if not isinstance(item, dict):
+                                continue
+                            template = _resolve_item_template(item, template_tag)
+                            fields = _template_markdown_fields(template)
+                            
+                            for field in fields:
+                                value = item.get(field)
+                                if not isinstance(value, str):
+                                    continue
+                                field_path = f"papers[{item_index}].{field}"
+                                
+                                # Find replacements for this specific field
+                                field_replacements = [
+                                    (start, end, repl)
+                                    for start, end, repl in replacements
+                                    if any(
+                                        t.field_path == field_path and t.item_index == item_index and t.span.start == start
+                                        for t in all_tasks
+                                        if t.file_path == path
+                                    )
+                                ]
+                                
+                                if field_replacements:
+                                    updated_value = apply_replacements(value, field_replacements)
+                                    item[field] = updated_value
+                        
+                        output_data: Any = items if payload is None else {**payload, "papers": items}
+                        serialized = json.dumps(output_data, ensure_ascii=False, indent=2)
+                        await asyncio.to_thread(output_path.write_text, f"{serialized}\n", encoding="utf-8")
+                    else:
+                        # For markdown, apply replacements directly
+                        content = await asyncio.to_thread(read_text, path)
+                        updated = apply_replacements(content, replacements)
                         await asyncio.to_thread(output_path.write_text, updated, encoding="utf-8")
-                    error_records.extend(errors)
-                return stats
-
-            async def runner(path: Path) -> None:
-                async with semaphore:
-                    stats = await handle_path(path)
-                    stats_total.diagrams_total += stats.diagrams_total
-                    stats_total.diagrams_invalid += stats.diagrams_invalid
-                    stats_total.diagrams_repaired += stats.diagrams_repaired
-                    stats_total.diagrams_failed += stats.diagrams_failed
-                    async with progress_lock:
-                        progress.update(1)
-
-            await asyncio.gather(*(runner(path) for path in paths))
+                    
+                    write_progress.update(1)
+                
+                write_progress.close()
+        
         return stats_total
 
     try:
         stats = asyncio.run(run())
     finally:
         progress.close()
+        field_progress.close()
         diagram_progress.close()
 
     if report_target and error_records:
@@ -1396,6 +1668,11 @@ def recognize_fix_mermaid(
         ("Invalid", str(stats.diagrams_invalid)),
         ("Repaired", str(stats.diagrams_repaired)),
         ("Failed", str(stats.diagrams_failed)),
+        ("Extract time", _format_duration(extract_duration)),
+        ("Extract avg", f"{extract_duration / stats.diagrams_total:.3f}s/diagram" if stats.diagrams_total > 0 else "-"),
+        ("Repair time", _format_duration(repair_duration)),
+        ("Repair avg", f"{repair_duration / stats.diagrams_invalid:.3f}s/diagram" if stats.diagrams_invalid > 0 else "-"),
+        ("Retry failed", "yes" if retry_failed else "no"),
         ("Only show error", "yes" if only_show_error else "no"),
         ("Report", _relative_path(report_target) if report_target else "-"),
     ]